Merge tag 'bcachefs-2025-01-20.2' of git://evilpiepirate.org/bcachefs

Pull bcachefs updates from Kent Overstreet:
 "Lots of scalability work, another big on-disk format change. On-disk
  format version goes from 1.13 to 1.20.

  Like 6.11, this is another big and expensive automatic/required on
  disk format upgrade. This is planned to be the last big on disk format
  upgrade before the experimental label comes off. There will be one
  more minor on disk format update for a few things that couldn't make
  this release.

  Headline improvements:

   - Self healing work:

     Allocator and reflink now run the exact same check/repair code that
     fsck does at runtime, where applicable.

     The long term goal here is to remove inconsistent() errors (that
     cause us to go emergency read only) by lifting fsck code up to
     normal runtime paths; we should only go emergency read-only if we
     detect an inconsistency that was due to a runtime bug - or truly
     catastrophic damage (corrupted btree roots/interior nodes).

   - Reflink repair no longer deletes reflink pointers:

     Instead we flip an error bit and log the error, and they can still
     be deleted by file deletion. This means a temporary failure to find
     an indirect extent (perhaps repaired later by btree node scan)
     won't result in unnecessary data loss

   - Improvements to rebalance data path option handling:

     We can now correctly apply changed filesystem-level io path options
     to pending rebalance work, and soon we'll be able to apply
     file-level io path option changes to indirect extents

   - Fix mount time regression that some users encountered post the 6.11
     disk accounting rewrite.

     Accounting keys were encoded little endian (typetag in the low
     bits) - which didn't anticipate adding accounting keys for every
     inode, which aren't stored in memory and we don't want to scan at
     mount time.

   - fsck time on large filesystems is improved by multiple orders of
     magnitude. Previously, 100TB was about the practical max filesystem
     size, where users were reporting fsck times of a day+. With the new
     changes (which nearly eliminate backpointers fsck overhead), we
     fsck'd a filesystem with 10PB of data in 1.5 hours.

     The problematic fsck passes were walking every extent and checking
     for missing backpointers, and walking every backpointer to check
     for dangling backpointers. As we've been adding more and more
     runtime self healing there was no reason to keep around the
     backpointers -> extents pass; dangling backpointers are just
     deleted, and we can do that when using them - thus, backpointers ->
     extents is now only run in debug mode.

     extents -> backpointers does need to exist, since missing
     backpointers would mean we can't find data to move it (for e.g.
     copygc, device evacuate, scrub). But the new on disk format version
     makes possible a new strategy where we sum up backpointers within a
     bucket and check it against the bucket sector counts, and then only
     scan for missing backpointers if the counts are off (and then, only
     for specific buckets).

  Full list of on disk format changes:

   - 1.14: backpointer_bucket_gen

     Backpointers now have a field for the bucket generation number,
     replacing the obsolete bucket_offset field. This is needed for the
     new "sum up backpointers within a bucket" code, since backpointers
     use the btree write buffer - meaning we will see stale reads, and
     this runs online, with the filesystem in full rw mode.

   - 1.15: disk_accounting_big_endian

     As previously described, fix the endianness of accounting keys so
     that accounting keys with the same typetag sort together, and
     accounting read can skip types it's not interested in.

   - 1.16: reflink_p_may_update_opts:

     This version indicates that a new reflink pointer field is
     understood and may be used; the field indicates whether the reflink
     pointer has permissions to update IO path options (e.g.
     compression, replicas) may be updated on the indirect extent it
     points to.

     This completes the rebalance/reflink data path option handling from
     the 6.13 pull request.

   - 1.17: inode_depth

     Add a new inode field, bi_depth, to accelerate the
     check_directory_structure fsck path, which checks for loops in the
     filesystem heirarchy.

     check_inodes and check_dirents check connectivity, so
     check_directory_structure only has to check for loops - by walking
     back up to the root from every directory.

     But a path can't be a loop if it has a counter that increases
     monotonically from root to leaf - adding a depth counter means that
     we can check for loops with only local (parent -> child) checks. We
     might need to occasionally renumber the depth field in fsck if
     directories have been moved around, but then future fsck runs will
     be much faster.

   - 1.18: persistent_inode_cursors

     Previously, the cursor used for inode allocation was only kept in
     memory, which meant that users with large filesystems and lots of
     files were reporting that the first create after mounting would
     take awhile - since it had to scan from the start.

     Inode allocation cursors are now persistent, and also include a
     generation field (incremented on wraparound, which will only happen
     if inode allocation is restricted to 32 bit inodes), so that we
     don't have to leave inode_generation keys around after a delete.

     The option for 32 bit inode numbers may now also be set on
     individual directories, and non-32 bit inode allocations are
     disallowed from allocating from the 32 bit part of the inode number
     space.

   - 1.19: autofix_errors

     Runtime self healing is now the default.o

   - 1.20: directory size (from Hongbo)

     directory i_size is now meaningful, and not 0"

* tag 'bcachefs-2025-01-20.2' of git://evilpiepirate.org/bcachefs: (268 commits)
  bcachefs: Fix check_inode_hash_info_matches_root()
  bcachefs: Document issue with bch_stripe layout
  bcachefs: Fix self healing on read error
  bcachefs: Pop all the transactions from the abort one
  bcachefs: Only abort the transactions in the cycle
  bcachefs: Introduce lock_graph_pop_from
  bcachefs: Convert open-coded lock_graph_pop_all to helper
  bcachefs: Do not allow no fail lock request to fail
  bcachefs: Merge the condition to avoid additional invocation
  Revert "bcachefs: Fix bch2_btree_node_upgrade()"
  bcachefs: bcachefs_metadata_version_directory_size
  bcachefs: make directory i_size meaningful
  bcachefs: check_unreachable_inodes is not actually PASS_ONLINE yet
  bcachefs: Don't use BTREE_ITER_cached when walking alloc btree during fsck
  bcachefs: Check for dirents to overwritten inodes
  bcachefs: bch2_btree_iter_peek_slot() handles navigating to nonexistent depth
  bcachefs: Don't set btree_path to updtodate if we don't fill
  bcachefs: __bch2_btree_pos_to_text()
  bcachefs: printbuf_reset() handles tabstops
  bcachefs: Silence read-only errors when deleting snapshots
  ...

141 files changed, 7175 insertions, 4639 deletions

diff --git a/Documentation/filesystems/bcachefs/CodingStyle.rst b/Documentation/filesystems/bcachefs/CodingStyle.rst
index 01de555e21d8..b29562a6bf55 100644
--- a/Documentation/filesystems/bcachefs/CodingStyle.rst
+++ b/Documentation/filesystems/bcachefs/CodingStyle.rst
@@ -183,4 +183,4 @@ even better as a code comment.
 A good code comment is wonderful, but even better is the comment that didn't
 need to exist because the code was so straightforward as to be obvious;
 organized into small clean and tidy modules, with clear and descriptive names
-for functions and variable, where every line of code has a clear purpose.
+for functions and variables, where every line of code has a clear purpose.
diff --git a/fs/bcachefs/Kconfig b/fs/bcachefs/Kconfig
index ab6c95b895b3..464b927e4fff 100644
--- a/fs/bcachefs/Kconfig
+++ b/fs/bcachefs/Kconfig
@@ -90,7 +90,7 @@ config BCACHEFS_SIX_OPTIMISTIC_SPIN
 
 config BCACHEFS_PATH_TRACEPOINTS
 	bool "Extra btree_path tracepoints"
-	depends on BCACHEFS_FS
+	depends on BCACHEFS_FS && TRACING
 	help
 	Enable extra tracepoints for debugging btree_path operations; we don't
 	normally want these enabled because they happen at very high rates.
diff --git a/fs/bcachefs/Makefile b/fs/bcachefs/Makefile
index 56d20e219f59..d2689388d5e8 100644
--- a/fs/bcachefs/Makefile
+++ b/fs/bcachefs/Makefile
@@ -82,6 +82,7 @@ bcachefs-y		:=	\
 	siphash.o		\
 	six.o			\
 	snapshot.o		\
+	str_hash.o		\
 	subvolume.o		\
 	super.o			\
 	super-io.o		\
diff --git a/fs/bcachefs/acl.c b/fs/bcachefs/acl.c
index 87f1be9d4db4..99487727ae64 100644
--- a/fs/bcachefs/acl.c
+++ b/fs/bcachefs/acl.c
@@ -184,11 +184,6 @@ invalid:
 	return ERR_PTR(-EINVAL);
 }
 
-#define acl_for_each_entry(acl, acl_e)			\
-	for (acl_e = acl->a_entries;			\
-	     acl_e < acl->a_entries + acl->a_count;	\
-	     acl_e++)
-
 /*
  * Convert from in-memory to filesystem representation.
  */
@@ -199,11 +194,11 @@ bch2_acl_to_xattr(struct btree_trans *trans,
 {
 	struct bkey_i_xattr *xattr;
 	bch_acl_header *acl_header;
-	const struct posix_acl_entry *acl_e;
+	const struct posix_acl_entry *acl_e, *pe;
 	void *outptr;
 	unsigned nr_short = 0, nr_long = 0, acl_len, u64s;
 
-	acl_for_each_entry(acl, acl_e) {
+	FOREACH_ACL_ENTRY(acl_e, acl, pe) {
 		switch (acl_e->e_tag) {
 		case ACL_USER:
 		case ACL_GROUP:
@@ -241,7 +236,7 @@ bch2_acl_to_xattr(struct btree_trans *trans,
 
 	outptr = (void *) acl_header + sizeof(*acl_header);
 
-	acl_for_each_entry(acl, acl_e) {
+	FOREACH_ACL_ENTRY(acl_e, acl, pe) {
 		bch_acl_entry *entry = outptr;
 
 		entry->e_tag = cpu_to_le16(acl_e->e_tag);
diff --git a/fs/bcachefs/alloc_background.c b/fs/bcachefs/alloc_background.c
index c84a91572a1d..fc2ef33b67b3 100644
--- a/fs/bcachefs/alloc_background.c
+++ b/fs/bcachefs/alloc_background.c
@@ -198,7 +198,7 @@ static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 }
 
 int bch2_alloc_v1_validate(struct bch_fs *c, struct bkey_s_c k,
-			   enum bch_validate_flags flags)
+			   struct bkey_validate_context from)
 {
 	struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 	int ret = 0;
@@ -213,7 +213,7 @@ fsck_err:
 }
 
 int bch2_alloc_v2_validate(struct bch_fs *c, struct bkey_s_c k,
-			   enum bch_validate_flags flags)
+			   struct bkey_validate_context from)
 {
 	struct bkey_alloc_unpacked u;
 	int ret = 0;
@@ -226,7 +226,7 @@ fsck_err:
 }
 
 int bch2_alloc_v3_validate(struct bch_fs *c, struct bkey_s_c k,
-			   enum bch_validate_flags flags)
+			   struct bkey_validate_context from)
 {
 	struct bkey_alloc_unpacked u;
 	int ret = 0;
@@ -239,7 +239,7 @@ fsck_err:
 }
 
 int bch2_alloc_v4_validate(struct bch_fs *c, struct bkey_s_c k,
-			   enum bch_validate_flags flags)
+			   struct bkey_validate_context from)
 {
 	struct bch_alloc_v4 a;
 	int ret = 0;
@@ -322,9 +322,9 @@ fsck_err:
 void bch2_alloc_v4_swab(struct bkey_s k)
 {
 	struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
-	struct bch_backpointer *bp, *bps;
 
-	a->journal_seq		= swab64(a->journal_seq);
+	a->journal_seq_nonempty	= swab64(a->journal_seq_nonempty);
+	a->journal_seq_empty	= swab64(a->journal_seq_empty);
 	a->flags		= swab32(a->flags);
 	a->dirty_sectors	= swab32(a->dirty_sectors);
 	a->cached_sectors	= swab32(a->cached_sectors);
@@ -333,13 +333,6 @@ void bch2_alloc_v4_swab(struct bkey_s k)
 	a->stripe		= swab32(a->stripe);
 	a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 	a->stripe_sectors	= swab32(a->stripe_sectors);
-
-	bps = alloc_v4_backpointers(a);
-	for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
-		bp->bucket_offset	= swab40(bp->bucket_offset);
-		bp->bucket_len		= swab32(bp->bucket_len);
-		bch2_bpos_swab(&bp->pos);
-	}
 }
 
 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
@@ -354,16 +347,17 @@ void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c
 	prt_printf(out, "gen %u oldest_gen %u data_type ", a->gen, a->oldest_gen);
 	bch2_prt_data_type(out, a->data_type);
 	prt_newline(out);
-	prt_printf(out, "journal_seq       %llu\n",	a->journal_seq);
-	prt_printf(out, "need_discard      %llu\n",	BCH_ALLOC_V4_NEED_DISCARD(a));
-	prt_printf(out, "need_inc_gen      %llu\n",	BCH_ALLOC_V4_NEED_INC_GEN(a));
-	prt_printf(out, "dirty_sectors     %u\n",	a->dirty_sectors);
-	prt_printf(out, "stripe_sectors    %u\n",	a->stripe_sectors);
-	prt_printf(out, "cached_sectors    %u\n",	a->cached_sectors);
-	prt_printf(out, "stripe            %u\n",	a->stripe);
-	prt_printf(out, "stripe_redundancy %u\n",	a->stripe_redundancy);
-	prt_printf(out, "io_time[READ]     %llu\n",	a->io_time[READ]);
-	prt_printf(out, "io_time[WRITE]    %llu\n",	a->io_time[WRITE]);
+	prt_printf(out, "journal_seq_nonempty %llu\n",	a->journal_seq_nonempty);
+	prt_printf(out, "journal_seq_empty    %llu\n",	a->journal_seq_empty);
+	prt_printf(out, "need_discard         %llu\n",	BCH_ALLOC_V4_NEED_DISCARD(a));
+	prt_printf(out, "need_inc_gen         %llu\n",	BCH_ALLOC_V4_NEED_INC_GEN(a));
+	prt_printf(out, "dirty_sectors        %u\n",	a->dirty_sectors);
+	prt_printf(out, "stripe_sectors       %u\n",	a->stripe_sectors);
+	prt_printf(out, "cached_sectors       %u\n",	a->cached_sectors);
+	prt_printf(out, "stripe               %u\n",	a->stripe);
+	prt_printf(out, "stripe_redundancy    %u\n",	a->stripe_redundancy);
+	prt_printf(out, "io_time[READ]        %llu\n",	a->io_time[READ]);
+	prt_printf(out, "io_time[WRITE]       %llu\n",	a->io_time[WRITE]);
 
 	if (ca)
 		prt_printf(out, "fragmentation     %llu\n",	alloc_lru_idx_fragmentation(*a, ca));
@@ -392,7 +386,7 @@ void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 		struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 
 		*out = (struct bch_alloc_v4) {
-			.journal_seq		= u.journal_seq,
+			.journal_seq_nonempty	= u.journal_seq,
 			.flags			= u.need_discard,
 			.gen			= u.gen,
 			.oldest_gen		= u.oldest_gen,
@@ -517,7 +511,7 @@ static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 }
 
 int bch2_bucket_gens_validate(struct bch_fs *c, struct bkey_s_c k,
-			     enum bch_validate_flags flags)
+			      struct bkey_validate_context from)
 {
 	int ret = 0;
 
@@ -664,74 +658,80 @@ int bch2_alloc_read(struct bch_fs *c)
 
 /* Free space/discard btree: */
 
+static int __need_discard_or_freespace_err(struct btree_trans *trans,
+					   struct bkey_s_c alloc_k,
+					   bool set, bool discard, bool repair)
+{
+	struct bch_fs *c = trans->c;
+	enum bch_fsck_flags flags = FSCK_CAN_IGNORE|(repair ? FSCK_CAN_FIX : 0);
+	enum bch_sb_error_id err_id = discard
+		? BCH_FSCK_ERR_need_discard_key_wrong
+		: BCH_FSCK_ERR_freespace_key_wrong;
+	enum btree_id btree = discard ? BTREE_ID_need_discard : BTREE_ID_freespace;
+	struct printbuf buf = PRINTBUF;
+
+	bch2_bkey_val_to_text(&buf, c, alloc_k);
+
+	int ret = __bch2_fsck_err(NULL, trans, flags, err_id,
+				  "bucket incorrectly %sset in %s btree\n"
+				  "  %s",
+				  set ? "" : "un",
+				  bch2_btree_id_str(btree),
+				  buf.buf);
+	if (ret == -BCH_ERR_fsck_ignore ||
+	    ret == -BCH_ERR_fsck_errors_not_fixed)
+		ret = 0;
+
+	printbuf_exit(&buf);
+	return ret;
+}
+
+#define need_discard_or_freespace_err(...)		\
+	fsck_err_wrap(__need_discard_or_freespace_err(__VA_ARGS__))
+
+#define need_discard_or_freespace_err_on(cond, ...)		\
+	(unlikely(cond) ?  need_discard_or_freespace_err(__VA_ARGS__) : false)
+
 static int bch2_bucket_do_index(struct btree_trans *trans,
 				struct bch_dev *ca,
 				struct bkey_s_c alloc_k,
 				const struct bch_alloc_v4 *a,
 				bool set)
 {
-	struct bch_fs *c = trans->c;
-	struct btree_iter iter;
-	struct bkey_s_c old;
-	struct bkey_i *k;
 	enum btree_id btree;
-	enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
-	enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
-	struct printbuf buf = PRINTBUF;
-	int ret;
+	struct bpos pos;
 
 	if (a->data_type != BCH_DATA_free &&
 	    a->data_type != BCH_DATA_need_discard)
 		return 0;
 
-	k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
-	if (IS_ERR(k))
-		return PTR_ERR(k);
-
-	bkey_init(&k->k);
-	k->k.type = new_type;
-
 	switch (a->data_type) {
 	case BCH_DATA_free:
 		btree = BTREE_ID_freespace;
-		k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
-		bch2_key_resize(&k->k, 1);
+		pos = alloc_freespace_pos(alloc_k.k->p, *a);
 		break;
 	case BCH_DATA_need_discard:
 		btree = BTREE_ID_need_discard;
-		k->k.p = alloc_k.k->p;
+		pos = alloc_k.k->p;
 		break;
 	default:
 		return 0;
 	}
 
-	old = bch2_bkey_get_iter(trans, &iter, btree,
-			     bkey_start_pos(&k->k),
-			     BTREE_ITER_intent);
-	ret = bkey_err(old);
+	struct btree_iter iter;
+	struct bkey_s_c old = bch2_bkey_get_iter(trans, &iter, btree, pos, BTREE_ITER_intent);
+	int ret = bkey_err(old);
 	if (ret)
 		return ret;
 
-	if (ca->mi.freespace_initialized &&
-	    c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info &&
-	    bch2_trans_inconsistent_on(old.k->type != old_type, trans,
-			"incorrect key when %s %s:%llu:%llu:0 (got %s should be %s)\n"
-			"  for %s",
-			set ? "setting" : "clearing",
-			bch2_btree_id_str(btree),
-			iter.pos.inode,
-			iter.pos.offset,
-			bch2_bkey_types[old.k->type],
-			bch2_bkey_types[old_type],
-			(bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
-		ret = -EIO;
-		goto err;
-	}
+	need_discard_or_freespace_err_on(ca->mi.freespace_initialized &&
+					 !old.k->type != set,
+					 trans, alloc_k, set,
+					 btree == BTREE_ID_need_discard, false);
 
-	ret = bch2_trans_update(trans, &iter, k, 0);
-err:
+	ret = bch2_btree_bit_mod_iter(trans, &iter, set);
+fsck_err:
 	bch2_trans_iter_exit(trans, &iter);
-	printbuf_exit(&buf);
 	return ret;
 }
 
@@ -858,7 +858,10 @@ int bch2_trigger_alloc(struct btree_trans *trans,
 	if (flags & BTREE_TRIGGER_transactional) {
 		alloc_data_type_set(new_a, new_a->data_type);
 
-		if (bch2_bucket_sectors_total(*new_a) > bch2_bucket_sectors_total(*old_a)) {
+		int is_empty_delta = (int) data_type_is_empty(new_a->data_type) -
+				     (int) data_type_is_empty(old_a->data_type);
+
+		if (is_empty_delta < 0) {
 			new_a->io_time[READ] = bch2_current_io_time(c, READ);
 			new_a->io_time[WRITE]= bch2_current_io_time(c, WRITE);
 			SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
@@ -928,37 +931,55 @@ int bch2_trigger_alloc(struct btree_trans *trans,
 	}
 
 	if ((flags & BTREE_TRIGGER_atomic) && (flags & BTREE_TRIGGER_insert)) {
-		u64 journal_seq = trans->journal_res.seq;
-		u64 bucket_journal_seq = new_a->journal_seq;
+		u64 transaction_seq = trans->journal_res.seq;
+		BUG_ON(!transaction_seq);
 
-		if ((flags & BTREE_TRIGGER_insert) &&
-		    data_type_is_empty(old_a->data_type) !=
-		    data_type_is_empty(new_a->data_type) &&
-		    new.k->type == KEY_TYPE_alloc_v4) {
-			struct bch_alloc_v4 *v = bkey_s_to_alloc_v4(new).v;
+		if (log_fsck_err_on(transaction_seq && new_a->journal_seq_nonempty > transaction_seq,
+				    trans, alloc_key_journal_seq_in_future,
+				    "bucket journal seq in future (currently at %llu)\n%s",
+				    journal_cur_seq(&c->journal),
+				    (bch2_bkey_val_to_text(&buf, c, new.s_c), buf.buf)))
+			new_a->journal_seq_nonempty = transaction_seq;
 
-			/*
-			 * If the btree updates referring to a bucket weren't flushed
-			 * before the bucket became empty again, then the we don't have
-			 * to wait on a journal flush before we can reuse the bucket:
-			 */
-			v->journal_seq = bucket_journal_seq =
-				data_type_is_empty(new_a->data_type) &&
-				(journal_seq == v->journal_seq ||
-				 bch2_journal_noflush_seq(&c->journal, v->journal_seq))
-				? 0 : journal_seq;
+		int is_empty_delta = (int) data_type_is_empty(new_a->data_type) -
+				     (int) data_type_is_empty(old_a->data_type);
+
+		/*
+		 * Record journal sequence number of empty -> nonempty transition:
+		 * Note that there may be multiple empty -> nonempty
+		 * transitions, data in a bucket may be overwritten while we're
+		 * still writing to it - so be careful to only record the first:
+		 * */
+		if (is_empty_delta < 0 &&
+		    new_a->journal_seq_empty <= c->journal.flushed_seq_ondisk) {
+			new_a->journal_seq_nonempty	= transaction_seq;
+			new_a->journal_seq_empty	= 0;
 		}
 
-		if (!data_type_is_empty(old_a->data_type) &&
-		    data_type_is_empty(new_a->data_type) &&
-		    bucket_journal_seq) {
-			ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
-					c->journal.flushed_seq_ondisk,
-					new.k->p.inode, new.k->p.offset,
-					bucket_journal_seq);
-			if (bch2_fs_fatal_err_on(ret, c,
-					"setting bucket_needs_journal_commit: %s", bch2_err_str(ret)))
-				goto err;
+		/*
+		 * Bucket becomes empty: mark it as waiting for a journal flush,
+		 * unless updates since empty -> nonempty transition were never
+		 * flushed - we may need to ask the journal not to flush
+		 * intermediate sequence numbers:
+		 */
+		if (is_empty_delta > 0) {
+			if (new_a->journal_seq_nonempty == transaction_seq ||
+			    bch2_journal_noflush_seq(&c->journal,
+						     new_a->journal_seq_nonempty,
+						     transaction_seq)) {
+				new_a->journal_seq_nonempty = new_a->journal_seq_empty = 0;
+			} else {
+				new_a->journal_seq_empty = transaction_seq;
+
+				ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
+									   c->journal.flushed_seq_ondisk,
+									   new.k->p.inode, new.k->p.offset,
+									   transaction_seq);
+				if (bch2_fs_fatal_err_on(ret, c,
+						"setting bucket_needs_journal_commit: %s",
+						bch2_err_str(ret)))
+					goto err;
+			}
 		}
 
 		if (new_a->gen != old_a->gen) {
@@ -974,7 +995,7 @@ int bch2_trigger_alloc(struct btree_trans *trans,
 
 #define eval_state(_a, expr)		({ const struct bch_alloc_v4 *a = _a; expr; })
 #define statechange(expr)		!eval_state(old_a, expr) && eval_state(new_a, expr)
-#define bucket_flushed(a)		(!a->journal_seq || a->journal_seq <= c->journal.flushed_seq_ondisk)
+#define bucket_flushed(a)		(a->journal_seq_empty <= c->journal.flushed_seq_ondisk)
 
 		if (statechange(a->data_type == BCH_DATA_free) &&
 		    bucket_flushed(new_a))
@@ -1006,6 +1027,7 @@ int bch2_trigger_alloc(struct btree_trans *trans,
 		rcu_read_unlock();
 	}
 err:
+fsck_err:
 	printbuf_exit(&buf);
 	bch2_dev_put(ca);
 	return ret;
@@ -1045,7 +1067,7 @@ static struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos
 		 * btree node min/max is a closed interval, upto takes a half
 		 * open interval:
 		 */
-		k = bch2_btree_iter_peek_upto(&iter2, end);
+		k = bch2_btree_iter_peek_max(&iter2, end);
 		next = iter2.pos;
 		bch2_trans_iter_exit(iter->trans, &iter2);
 
@@ -1129,7 +1151,6 @@ int bch2_check_alloc_key(struct btree_trans *trans,
 	struct bch_fs *c = trans->c;
 	struct bch_alloc_v4 a_convert;
 	const struct bch_alloc_v4 *a;
-	unsigned discard_key_type, freespace_key_type;
 	unsigned gens_offset;
 	struct bkey_s_c k;
 	struct printbuf buf = PRINTBUF;
@@ -1149,64 +1170,30 @@ int bch2_check_alloc_key(struct btree_trans *trans,
 
 	a = bch2_alloc_to_v4(alloc_k, &a_convert);
 
-	discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
 	bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
 	k = bch2_btree_iter_peek_slot(discard_iter);
 	ret = bkey_err(k);
 	if (ret)
 		goto err;
 
-	if (fsck_err_on(k.k->type != discard_key_type,
-			trans, need_discard_key_wrong,
-			"incorrect key in need_discard btree (got %s should be %s)\n"
-			"  %s",
-			bch2_bkey_types[k.k->type],
-			bch2_bkey_types[discard_key_type],
-			(bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
-		struct bkey_i *update =
-			bch2_trans_kmalloc(trans, sizeof(*update));
-
-		ret = PTR_ERR_OR_ZERO(update);
-		if (ret)
-			goto err;
-
-		bkey_init(&update->k);
-		update->k.type	= discard_key_type;
-		update->k.p	= discard_iter->pos;
-
-		ret = bch2_trans_update(trans, discard_iter, update, 0);
+	bool is_discarded = a->data_type == BCH_DATA_need_discard;
+	if (need_discard_or_freespace_err_on(!!k.k->type != is_discarded,
+					     trans, alloc_k, !is_discarded, true, true)) {
+		ret = bch2_btree_bit_mod_iter(trans, discard_iter, is_discarded);
 		if (ret)
 			goto err;
 	}
 
-	freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
 	bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
 	k = bch2_btree_iter_peek_slot(freespace_iter);
 	ret = bkey_err(k);
 	if (ret)
 		goto err;
 
-	if (fsck_err_on(k.k->type != freespace_key_type,
-			trans, freespace_key_wrong,
-			"incorrect key in freespace btree (got %s should be %s)\n"
-			"  %s",
-			bch2_bkey_types[k.k->type],
-			bch2_bkey_types[freespace_key_type],
-			(printbuf_reset(&buf),
-			 bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
-		struct bkey_i *update =
-			bch2_trans_kmalloc(trans, sizeof(*update));
-
-		ret = PTR_ERR_OR_ZERO(update);
-		if (ret)
-			goto err;
-
-		bkey_init(&update->k);
-		update->k.type	= freespace_key_type;
-		update->k.p	= freespace_iter->pos;
-		bch2_key_resize(&update->k, 1);
-
-		ret = bch2_trans_update(trans, freespace_iter, update, 0);
+	bool is_free = a->data_type == BCH_DATA_free;
+	if (need_discard_or_freespace_err_on(!!k.k->type != is_free,
+					     trans, alloc_k, !is_free, false, true)) {
+		ret = bch2_btree_bit_mod_iter(trans, freespace_iter, is_free);
 		if (ret)
 			goto err;
 	}
@@ -1368,51 +1355,88 @@ fsck_err:
 	return ret;
 }
 
-static noinline_for_stack int bch2_check_discard_freespace_key(struct btree_trans *trans,
-					      struct btree_iter *iter)
+struct check_discard_freespace_key_async {
+	struct work_struct	work;
+	struct bch_fs		*c;
+	struct bbpos		pos;
+};
+
+static int bch2_recheck_discard_freespace_key(struct btree_trans *trans, struct bbpos pos)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, pos.btree, pos.pos, 0);
+	int ret = bkey_err(k);
+	if (ret)
+		return ret;
+
+	u8 gen;
+	ret = k.k->type != KEY_TYPE_set
+		? bch2_check_discard_freespace_key(trans, &iter, &gen, false)
+		: 0;
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+static void check_discard_freespace_key_work(struct work_struct *work)
+{
+	struct check_discard_freespace_key_async *w =
+		container_of(work, struct check_discard_freespace_key_async, work);
+
+	bch2_trans_do(w->c, bch2_recheck_discard_freespace_key(trans, w->pos));
+	bch2_write_ref_put(w->c, BCH_WRITE_REF_check_discard_freespace_key);
+	kfree(w);
+}
+
+int bch2_check_discard_freespace_key(struct btree_trans *trans, struct btree_iter *iter, u8 *gen,
+				     bool async_repair)
 {
 	struct bch_fs *c = trans->c;
-	struct btree_iter alloc_iter;
-	struct bkey_s_c alloc_k;
-	struct bch_alloc_v4 a_convert;
-	const struct bch_alloc_v4 *a;
-	u64 genbits;
-	struct bpos pos;
 	enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
 		? BCH_DATA_need_discard
 		: BCH_DATA_free;
 	struct printbuf buf = PRINTBUF;
-	int ret;
 
-	pos = iter->pos;
-	pos.offset &= ~(~0ULL << 56);
-	genbits = iter->pos.offset & (~0ULL << 56);
+	struct bpos bucket = iter->pos;
+	bucket.offset &= ~(~0ULL << 56);
+	u64 genbits = iter->pos.offset & (~0ULL << 56);
 
-	alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
-	ret = bkey_err(alloc_k);
+	struct btree_iter alloc_iter;
+	struct bkey_s_c alloc_k = bch2_bkey_get_iter(trans, &alloc_iter,
+						     BTREE_ID_alloc, bucket,
+						     async_repair ? BTREE_ITER_cached : 0);
+	int ret = bkey_err(alloc_k);
 	if (ret)
 		return ret;
 
-	if (fsck_err_on(!bch2_dev_bucket_exists(c, pos),
-			trans, need_discard_freespace_key_to_invalid_dev_bucket,
-			"entry in %s btree for nonexistant dev:bucket %llu:%llu",
-			bch2_btree_id_str(iter->btree_id), pos.inode, pos.offset))
-		goto delete;
+	if (!bch2_dev_bucket_exists(c, bucket)) {
+		if (fsck_err(trans, need_discard_freespace_key_to_invalid_dev_bucket,
+			     "entry in %s btree for nonexistant dev:bucket %llu:%llu",
+			     bch2_btree_id_str(iter->btree_id), bucket.inode, bucket.offset))
+			goto delete;
+		ret = 1;
+		goto out;
+	}
 
-	a = bch2_alloc_to_v4(alloc_k, &a_convert);
+	struct bch_alloc_v4 a_convert;
+	const struct bch_alloc_v4 *a = bch2_alloc_to_v4(alloc_k, &a_convert);
+
+	if (a->data_type != state ||
+	    (state == BCH_DATA_free &&
+	     genbits != alloc_freespace_genbits(*a))) {
+		if (fsck_err(trans, need_discard_freespace_key_bad,
+			     "%s\n  incorrectly set at %s:%llu:%llu:0 (free %u, genbits %llu should be %llu)",
+			     (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
+			     bch2_btree_id_str(iter->btree_id),
+			     iter->pos.inode,
+			     iter->pos.offset,
+			     a->data_type == state,
+			     genbits >> 56, alloc_freespace_genbits(*a) >> 56))
+			goto delete;
+		ret = 1;
+		goto out;
+	}
 
-	if (fsck_err_on(a->data_type != state ||
-			(state == BCH_DATA_free &&
-			 genbits != alloc_freespace_genbits(*a)),
-			trans, need_discard_freespace_key_bad,
-			"%s\n  incorrectly set at %s:%llu:%llu:0 (free %u, genbits %llu should be %llu)",
-			(bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
-			bch2_btree_id_str(iter->btree_id),
-			iter->pos.inode,
-			iter->pos.offset,
-			a->data_type == state,
-			genbits >> 56, alloc_freespace_genbits(*a) >> 56))
-		goto delete;
+	*gen = a->gen;
 out:
 fsck_err:
 	bch2_set_btree_iter_dontneed(&alloc_iter);
@@ -1420,11 +1444,40 @@ fsck_err:
 	printbuf_exit(&buf);
 	return ret;
 delete:
-	ret =   bch2_btree_delete_extent_at(trans, iter,
-			iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0) ?:
-		bch2_trans_commit(trans, NULL, NULL,
-			BCH_TRANS_COMMIT_no_enospc);
-	goto out;
+	if (!async_repair) {
+		ret =   bch2_btree_bit_mod_iter(trans, iter, false) ?:
+			bch2_trans_commit(trans, NULL, NULL,
+				BCH_TRANS_COMMIT_no_enospc) ?:
+			-BCH_ERR_transaction_restart_commit;
+		goto out;
+	} else {
+		/*
+		 * We can't repair here when called from the allocator path: the
+		 * commit will recurse back into the allocator
+		 */
+		struct check_discard_freespace_key_async *w =
+			kzalloc(sizeof(*w), GFP_KERNEL);
+		if (!w)
+			goto out;
+
+		if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_check_discard_freespace_key)) {
+			kfree(w);
+			goto out;
+		}
+
+		INIT_WORK(&w->work, check_discard_freespace_key_work);
+		w->c = c;
+		w->pos = BBPOS(iter->btree_id, iter->pos);
+		queue_work(c->write_ref_wq, &w->work);
+		goto out;
+	}
+}
+
+static int bch2_check_discard_freespace_key_fsck(struct btree_trans *trans, struct btree_iter *iter)
+{
+	u8 gen;
+	int ret = bch2_check_discard_freespace_key(trans, iter, &gen, false);
+	return ret < 0 ? ret : 0;
 }
 
 /*
@@ -1581,7 +1634,7 @@ bkey_err:
 	ret = for_each_btree_key(trans, iter,
 			BTREE_ID_need_discard, POS_MIN,
 			BTREE_ITER_prefetch, k,
-		bch2_check_discard_freespace_key(trans, &iter));
+		bch2_check_discard_freespace_key_fsck(trans, &iter));
 	if (ret)
 		goto err;
 
@@ -1594,7 +1647,7 @@ bkey_err:
 			break;
 
 		ret = bkey_err(k) ?:
-			bch2_check_discard_freespace_key(trans, &iter);
+			bch2_check_discard_freespace_key_fsck(trans, &iter);
 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 			ret = 0;
 			continue;
@@ -1757,7 +1810,8 @@ static int bch2_discard_one_bucket(struct btree_trans *trans,
 				   struct bch_dev *ca,
 				   struct btree_iter *need_discard_iter,
 				   struct bpos *discard_pos_done,
-				   struct discard_buckets_state *s)
+				   struct discard_buckets_state *s,
+				   bool fastpath)
 {
 	struct bch_fs *c = trans->c;
 	struct bpos pos = need_discard_iter->pos;
@@ -1793,44 +1847,23 @@ static int bch2_discard_one_bucket(struct btree_trans *trans,
 	if (ret)
 		goto out;
 
-	if (bch2_bucket_sectors_total(a->v)) {
-		if (bch2_trans_inconsistent_on(c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info,
-					       trans, "attempting to discard bucket with dirty data\n%s",
-					       (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
-			ret = -EIO;
-		goto out;
-	}
-
 	if (a->v.data_type != BCH_DATA_need_discard) {
-		if (data_type_is_empty(a->v.data_type) &&
-		    BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
-			a->v.gen++;
-			SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
-			goto write;
+		if (need_discard_or_freespace_err(trans, k, true, true, true)) {
+			ret = bch2_btree_bit_mod_iter(trans, need_discard_iter, false);
+			if (ret)
+				goto out;
+			goto commit;
 		}
 
-		if (bch2_trans_inconsistent_on(c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info,
-					       trans, "bucket incorrectly set in need_discard btree\n"
-					       "%s",
-					       (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
-			ret = -EIO;
 		goto out;
 	}
 
-	if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
-		if (bch2_trans_inconsistent_on(c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info,
-					       trans, "clearing need_discard but journal_seq %llu > flushed_seq %llu\n%s",
-					       a->v.journal_seq,
-					       c->journal.flushed_seq_ondisk,
-					       (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
-			ret = -EIO;
-		goto out;
-	}
-
-	if (discard_in_flight_add(ca, iter.pos.offset, true))
-		goto out;
+	if (!fastpath) {
+		if (discard_in_flight_add(ca, iter.pos.offset, true))
+			goto out;
 
-	discard_locked = true;
+		discard_locked = true;
+	}
 
 	if (!bkey_eq(*discard_pos_done, iter.pos) &&
 	    ca->mi.discard && !c->opts.nochanges) {
@@ -1844,6 +1877,7 @@ static int bch2_discard_one_bucket(struct btree_trans *trans,
 				     ca->mi.bucket_size,
 				     GFP_KERNEL);
 		*discard_pos_done = iter.pos;
+		s->discarded++;
 
 		ret = bch2_trans_relock_notrace(trans);
 		if (ret)
@@ -1851,22 +1885,25 @@ static int bch2_discard_one_bucket(struct btree_trans *trans,
 	}
 
 	SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
-write:
 	alloc_data_type_set(&a->v, a->v.data_type);
 
-	ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
-		bch2_trans_commit(trans, NULL, NULL,
-				  BCH_WATERMARK_btree|
-				  BCH_TRANS_COMMIT_no_enospc);
+	ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
+	if (ret)
+		goto out;
+commit:
+	ret = bch2_trans_commit(trans, NULL, NULL,
+				BCH_WATERMARK_btree|
+				BCH_TRANS_COMMIT_no_enospc);
 	if (ret)
 		goto out;
 
 	count_event(c, bucket_discard);
-	s->discarded++;
 out:
+fsck_err:
 	if (discard_locked)
 		discard_in_flight_remove(ca, iter.pos.offset);
-	s->seen++;
+	if (!ret)
+		s->seen++;
 	bch2_trans_iter_exit(trans, &iter);
 	printbuf_exit(&buf);
 	return ret;
@@ -1886,11 +1923,11 @@ static void bch2_do_discards_work(struct work_struct *work)
 	 * successful commit:
 	 */
 	ret = bch2_trans_run(c,
-		for_each_btree_key_upto(trans, iter,
+		for_each_btree_key_max(trans, iter,
 				   BTREE_ID_need_discard,
 				   POS(ca->dev_idx, 0),
 				   POS(ca->dev_idx, U64_MAX), 0, k,
-			bch2_discard_one_bucket(trans, ca, &iter, &discard_pos_done, &s)));
+			bch2_discard_one_bucket(trans, ca, &iter, &discard_pos_done, &s, false)));
 
 	trace_discard_buckets(c, s.seen, s.open, s.need_journal_commit, s.discarded,
 			      bch2_err_str(ret));
@@ -1923,27 +1960,29 @@ void bch2_do_discards(struct bch_fs *c)
 		bch2_dev_do_discards(ca);
 }
 
-static int bch2_clear_bucket_needs_discard(struct btree_trans *trans, struct bpos bucket)
+static int bch2_do_discards_fast_one(struct btree_trans *trans,
+				     struct bch_dev *ca,
+				     u64 bucket,
+				     struct bpos *discard_pos_done,
+				     struct discard_buckets_state *s)
 {
-	struct btree_iter iter;
-	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, bucket, BTREE_ITER_intent);
-	struct bkey_s_c k = bch2_btree_iter_peek_slot(&iter);
-	int ret = bkey_err(k);
+	struct btree_iter need_discard_iter;
+	struct bkey_s_c discard_k = bch2_bkey_get_iter(trans, &need_discard_iter,
+					BTREE_ID_need_discard, POS(ca->dev_idx, bucket), 0);
+	int ret = bkey_err(discard_k);
 	if (ret)
-		goto err;
-
-	struct bkey_i_alloc_v4 *a = bch2_alloc_to_v4_mut(trans, k);
-	ret = PTR_ERR_OR_ZERO(a);
-	if (ret)
-		goto err;
+		return ret;
 
-	BUG_ON(a->v.dirty_sectors);
-	SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
-	alloc_data_type_set(&a->v, a->v.data_type);
+	if (log_fsck_err_on(discard_k.k->type != KEY_TYPE_set,
+			    trans, discarding_bucket_not_in_need_discard_btree,
+			    "attempting to discard bucket %u:%llu not in need_discard btree",
+			    ca->dev_idx, bucket))
+		goto out;
 
-	ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
-err:
-	bch2_trans_iter_exit(trans, &iter);
+	ret = bch2_discard_one_bucket(trans, ca, &need_discard_iter, discard_pos_done, s, true);
+out:
+fsck_err:
+	bch2_trans_iter_exit(trans, &need_discard_iter);
 	return ret;
 }
 
@@ -1951,6 +1990,10 @@ static void bch2_do_discards_fast_work(struct work_struct *work)
 {
 	struct bch_dev *ca = container_of(work, struct bch_dev, discard_fast_work);
 	struct bch_fs *c = ca->fs;
+	struct discard_buckets_state s = {};
+	struct bpos discard_pos_done = POS_MAX;
+	struct btree_trans *trans = bch2_trans_get(c);
+	int ret = 0;
 
 	while (1) {
 		bool got_bucket = false;
@@ -1971,16 +2014,8 @@ static void bch2_do_discards_fast_work(struct work_struct *work)
 		if (!got_bucket)
 			break;
 
-		if (ca->mi.discard && !c->opts.nochanges)
-			blkdev_issue_discard(ca->disk_sb.bdev,
-					     bucket_to_sector(ca, bucket),
-					     ca->mi.bucket_size,
-					     GFP_KERNEL);
-
-		int ret = bch2_trans_commit_do(c, NULL, NULL,
-			BCH_WATERMARK_btree|
-			BCH_TRANS_COMMIT_no_enospc,
-			bch2_clear_bucket_needs_discard(trans, POS(ca->dev_idx, bucket)));
+		ret = lockrestart_do(trans,
+			bch2_do_discards_fast_one(trans, ca, bucket, &discard_pos_done, &s));
 		bch_err_fn(c, ret);
 
 		discard_in_flight_remove(ca, bucket);
@@ -1989,6 +2024,9 @@ static void bch2_do_discards_fast_work(struct work_struct *work)
 			break;
 	}
 
+	trace_discard_buckets(c, s.seen, s.open, s.need_journal_commit, s.discarded, bch2_err_str(ret));
+
+	bch2_trans_put(trans);
 	percpu_ref_put(&ca->io_ref);
 	bch2_write_ref_put(c, BCH_WRITE_REF_discard_fast);
 }
@@ -2030,8 +2068,11 @@ static int invalidate_one_bucket(struct btree_trans *trans,
 		return 1;
 
 	if (!bch2_dev_bucket_exists(c, bucket)) {
-		prt_str(&buf, "lru entry points to invalid bucket");
-		goto err;
+		if (fsck_err(trans, lru_entry_to_invalid_bucket,
+			     "lru key points to nonexistent device:bucket %llu:%llu",
+			     bucket.inode, bucket.offset))
+			return bch2_btree_bit_mod_buffered(trans, BTREE_ID_lru, lru_iter->pos, false);
+		goto out;
 	}
 
 	if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
@@ -2072,28 +2113,9 @@ static int invalidate_one_bucket(struct btree_trans *trans,
 	trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
 	--*nr_to_invalidate;
 out:
+fsck_err:
 	printbuf_exit(&buf);
 	return ret;
-err:
-	prt_str(&buf, "\n  lru key: ");
-	bch2_bkey_val_to_text(&buf, c, lru_k);
-
-	prt_str(&buf, "\n  lru entry: ");
-	bch2_lru_pos_to_text(&buf, lru_iter->pos);
-
-	prt_str(&buf, "\n  alloc key: ");
-	if (!a)
-		bch2_bpos_to_text(&buf, bucket);
-	else
-		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
-
-	bch_err(c, "%s", buf.buf);
-	if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_lrus) {
-		bch2_inconsistent_error(c);
-		ret = -EINVAL;
-	}
-
-	goto out;
 }
 
 static struct bkey_s_c next_lru_key(struct btree_trans *trans, struct btree_iter *iter,
@@ -2101,7 +2123,7 @@ static struct bkey_s_c next_lru_key(struct btree_trans *trans, struct btree_iter
 {
 	struct bkey_s_c k;
 again:
-	k = bch2_btree_iter_peek_upto(iter, lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX));
+	k = bch2_btree_iter_peek_max(iter, lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX));
 	if (!k.k && !*wrapped) {
 		bch2_btree_iter_set_pos(iter, lru_pos(ca->dev_idx, 0, 0));
 		*wrapped = true;
diff --git a/fs/bcachefs/alloc_background.h b/fs/bcachefs/alloc_background.h
index 163a67b97a40..de25ba4ee94b 100644
--- a/fs/bcachefs/alloc_background.h
+++ b/fs/bcachefs/alloc_background.h
@@ -8,8 +8,6 @@
 #include "debug.h"
 #include "super.h"
 
-enum bch_validate_flags;
-
 /* How out of date a pointer gen is allowed to be: */
 #define BUCKET_GC_GEN_MAX	96U
 
@@ -245,10 +243,14 @@ struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *, struct bkey_s
 
 int bch2_bucket_io_time_reset(struct btree_trans *, unsigned, size_t, int);
 
-int bch2_alloc_v1_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
-int bch2_alloc_v2_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
-int bch2_alloc_v3_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
-int bch2_alloc_v4_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_alloc_v1_validate(struct bch_fs *, struct bkey_s_c,
+			   struct bkey_validate_context);
+int bch2_alloc_v2_validate(struct bch_fs *, struct bkey_s_c,
+			   struct bkey_validate_context);
+int bch2_alloc_v3_validate(struct bch_fs *, struct bkey_s_c,
+			   struct bkey_validate_context);
+int bch2_alloc_v4_validate(struct bch_fs *, struct bkey_s_c,
+			   struct bkey_validate_context);
 void bch2_alloc_v4_swab(struct bkey_s);
 void bch2_alloc_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
@@ -282,7 +284,7 @@ void bch2_alloc_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 })
 
 int bch2_bucket_gens_validate(struct bch_fs *, struct bkey_s_c,
-			     enum bch_validate_flags);
+			      struct bkey_validate_context);
 void bch2_bucket_gens_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
 #define bch2_bkey_ops_bucket_gens ((struct bkey_ops) {	\
@@ -307,6 +309,8 @@ int bch2_alloc_key_to_dev_counters(struct btree_trans *, struct bch_dev *,
 int bch2_trigger_alloc(struct btree_trans *, enum btree_id, unsigned,
 		       struct bkey_s_c, struct bkey_s,
 		       enum btree_iter_update_trigger_flags);
+
+int bch2_check_discard_freespace_key(struct btree_trans *, struct btree_iter *, u8 *, bool);
 int bch2_check_alloc_info(struct bch_fs *);
 int bch2_check_alloc_to_lru_refs(struct bch_fs *);
 void bch2_dev_do_discards(struct bch_dev *);
diff --git a/fs/bcachefs/alloc_background_format.h b/fs/bcachefs/alloc_background_format.h
index befdaa95c515..740238369a5a 100644
--- a/fs/bcachefs/alloc_background_format.h
+++ b/fs/bcachefs/alloc_background_format.h
@@ -58,7 +58,7 @@ LE32_BITMASK(BCH_ALLOC_V3_NEED_INC_GEN,struct bch_alloc_v3, flags,  1,  2)
 
 struct bch_alloc_v4 {
 	struct bch_val		v;
-	__u64			journal_seq;
+	__u64			journal_seq_nonempty;
 	__u32			flags;
 	__u8			gen;
 	__u8			oldest_gen;
@@ -70,7 +70,7 @@ struct bch_alloc_v4 {
 	__u32			stripe;
 	__u32			nr_external_backpointers;
 	/* end of fields in original version of alloc_v4 */
-	__u64			_fragmentation_lru; /* obsolete */
+	__u64			journal_seq_empty;
 	__u32			stripe_sectors;
 	__u32			pad;
 } __packed __aligned(8);
diff --git a/fs/bcachefs/alloc_foreground.c b/fs/bcachefs/alloc_foreground.c
index 372178c8d416..6df41c331a52 100644
--- a/fs/bcachefs/alloc_foreground.c
+++ b/fs/bcachefs/alloc_foreground.c
@@ -107,14 +107,10 @@ void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
 		return;
 	}
 
-	percpu_down_read(&c->mark_lock);
 	spin_lock(&ob->lock);
-
 	ob->valid = false;
 	ob->data_type = 0;
-
 	spin_unlock(&ob->lock);
-	percpu_up_read(&c->mark_lock);
 
 	spin_lock(&c->freelist_lock);
 	bch2_open_bucket_hash_remove(c, ob);
@@ -156,6 +152,14 @@ static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 	return ob;
 }
 
+static inline bool is_superblock_bucket(struct bch_fs *c, struct bch_dev *ca, u64 b)
+{
+	if (c->curr_recovery_pass > BCH_RECOVERY_PASS_trans_mark_dev_sbs)
+		return false;
+
+	return bch2_is_superblock_bucket(ca, b);
+}
+
 static void open_bucket_free_unused(struct bch_fs *c, struct open_bucket *ob)
 {
 	BUG_ON(c->open_buckets_partial_nr >=
@@ -175,20 +179,6 @@ static void open_bucket_free_unused(struct bch_fs *c, struct open_bucket *ob)
 	closure_wake_up(&c->freelist_wait);
 }
 
-/* _only_ for allocating the journal on a new device: */
-long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
-{
-	while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
-		u64 b = ca->new_fs_bucket_idx++;
-
-		if (!is_superblock_bucket(ca, b) &&
-		    (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
-			return b;
-	}
-
-	return -1;
-}
-
 static inline unsigned open_buckets_reserved(enum bch_watermark watermark)
 {
 	switch (watermark) {
@@ -206,33 +196,40 @@ static inline unsigned open_buckets_reserved(enum bch_watermark watermark)
 	}
 }
 
-static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
-					      u64 bucket,
-					      enum bch_watermark watermark,
-					      const struct bch_alloc_v4 *a,
-					      struct bucket_alloc_state *s,
-					      struct closure *cl)
+static inline bool may_alloc_bucket(struct bch_fs *c,
+				    struct bpos bucket,
+				    struct bucket_alloc_state *s)
 {
-	struct open_bucket *ob;
-
-	if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
-		s->skipped_nouse++;
-		return NULL;
-	}
-
-	if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
+	if (bch2_bucket_is_open(c, bucket.inode, bucket.offset)) {
 		s->skipped_open++;
-		return NULL;
+		return false;
 	}
 
 	if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
-			c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
+			c->journal.flushed_seq_ondisk, bucket.inode, bucket.offset)) {
 		s->skipped_need_journal_commit++;
-		return NULL;
+		return false;
 	}
 
-	if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
+	if (bch2_bucket_nocow_is_locked(&c->nocow_locks, bucket)) {
 		s->skipped_nocow++;
+		return false;
+	}
+
+	return true;
+}
+
+static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
+					      u64 bucket, u8 gen,
+					      enum bch_watermark watermark,
+					      struct bucket_alloc_state *s,
+					      struct closure *cl)
+{
+	if (unlikely(is_superblock_bucket(c, ca, bucket)))
+		return NULL;
+
+	if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
+		s->skipped_nouse++;
 		return NULL;
 	}
 
@@ -254,14 +251,13 @@ static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *
 		return NULL;
 	}
 
-	ob = bch2_open_bucket_alloc(c);
+	struct open_bucket *ob = bch2_open_bucket_alloc(c);
 
 	spin_lock(&ob->lock);
-
 	ob->valid	= true;
 	ob->sectors_free = ca->mi.bucket_size;
 	ob->dev		= ca->dev_idx;
-	ob->gen		= a->gen;
+	ob->gen		= gen;
 	ob->bucket	= bucket;
 	spin_unlock(&ob->lock);
 
@@ -276,111 +272,29 @@ static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *
 }
 
 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
-					    enum bch_watermark watermark, u64 free_entry,
+					    enum bch_watermark watermark,
 					    struct bucket_alloc_state *s,
-					    struct bkey_s_c freespace_k,
+					    struct btree_iter *freespace_iter,
 					    struct closure *cl)
 {
 	struct bch_fs *c = trans->c;
-	struct btree_iter iter = { NULL };
-	struct bkey_s_c k;
-	struct open_bucket *ob;
-	struct bch_alloc_v4 a_convert;
-	const struct bch_alloc_v4 *a;
-	u64 b = free_entry & ~(~0ULL << 56);
-	unsigned genbits = free_entry >> 56;
-	struct printbuf buf = PRINTBUF;
-	int ret;
-
-	if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
-		prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
-		       "  freespace key ",
-			ca->mi.first_bucket, ca->mi.nbuckets);
-		bch2_bkey_val_to_text(&buf, c, freespace_k);
-		bch2_trans_inconsistent(trans, "%s", buf.buf);
-		ob = ERR_PTR(-EIO);
-		goto err;
-	}
+	u64 b = freespace_iter->pos.offset & ~(~0ULL << 56);
 
-	k = bch2_bkey_get_iter(trans, &iter,
-			       BTREE_ID_alloc, POS(ca->dev_idx, b),
-			       BTREE_ITER_cached);
-	ret = bkey_err(k);
-	if (ret) {
-		ob = ERR_PTR(ret);
-		goto err;
-	}
-
-	a = bch2_alloc_to_v4(k, &a_convert);
-
-	if (a->data_type != BCH_DATA_free) {
-		if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
-			ob = NULL;
-			goto err;
-		}
-
-		prt_printf(&buf, "non free bucket in freespace btree\n"
-		       "  freespace key ");
-		bch2_bkey_val_to_text(&buf, c, freespace_k);
-		prt_printf(&buf, "\n  ");
-		bch2_bkey_val_to_text(&buf, c, k);
-		bch2_trans_inconsistent(trans, "%s", buf.buf);
-		ob = ERR_PTR(-EIO);
-		goto err;
-	}
-
-	if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
-	    c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
-		prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
-		       "  freespace key ",
-		       genbits, alloc_freespace_genbits(*a) >> 56);
-		bch2_bkey_val_to_text(&buf, c, freespace_k);
-		prt_printf(&buf, "\n  ");
-		bch2_bkey_val_to_text(&buf, c, k);
-		bch2_trans_inconsistent(trans, "%s", buf.buf);
-		ob = ERR_PTR(-EIO);
-		goto err;
-	}
-
-	if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_extents_to_backpointers) {
-		struct bch_backpointer bp;
-		struct bpos bp_pos = POS_MIN;
-
-		ret = bch2_get_next_backpointer(trans, ca, POS(ca->dev_idx, b), -1,
-						&bp_pos, &bp,
-						BTREE_ITER_nopreserve);
-		if (ret) {
-			ob = ERR_PTR(ret);
-			goto err;
-		}
+	if (!may_alloc_bucket(c, POS(ca->dev_idx, b), s))
+		return NULL;
 
-		if (!bkey_eq(bp_pos, POS_MAX)) {
-			/*
-			 * Bucket may have data in it - we don't call
-			 * bc2h_trans_inconnsistent() because fsck hasn't
-			 * finished yet
-			 */
-			ob = NULL;
-			goto err;
-		}
-	}
+	u8 gen;
+	int ret = bch2_check_discard_freespace_key(trans, freespace_iter, &gen, true);
+	if (ret < 0)
+		return ERR_PTR(ret);
+	if (ret)
+		return NULL;
 
-	ob = __try_alloc_bucket(c, ca, b, watermark, a, s, cl);
-	if (!ob)
-		bch2_set_btree_iter_dontneed(&iter);
-err:
-	if (iter.path)
-		bch2_set_btree_iter_dontneed(&iter);
-	bch2_trans_iter_exit(trans, &iter);
-	printbuf_exit(&buf);
-	return ob;
+	return __try_alloc_bucket(c, ca, b, gen, watermark, s, cl);
 }
 
 /*
  * This path is for before the freespace btree is initialized:
- *
- * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
- * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
  */
 static noinline struct open_bucket *
 bch2_bucket_alloc_early(struct btree_trans *trans,
@@ -389,10 +303,11 @@ bch2_bucket_alloc_early(struct btree_trans *trans,
 			struct bucket_alloc_state *s,
 			struct closure *cl)
 {
+	struct bch_fs *c = trans->c;
 	struct btree_iter iter, citer;
 	struct bkey_s_c k, ck;
 	struct open_bucket *ob = NULL;
-	u64 first_bucket = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
+	u64 first_bucket = ca->mi.first_bucket;
 	u64 *dev_alloc_cursor = &ca->alloc_cursor[s->btree_bitmap];
 	u64 alloc_start = max(first_bucket, *dev_alloc_cursor);
 	u64 alloc_cursor = alloc_start;
@@ -415,10 +330,6 @@ again:
 		if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
 			break;
 
-		if (ca->new_fs_bucket_idx &&
-		    is_superblock_bucket(ca, k.k->p.offset))
-			continue;
-
 		if (s->btree_bitmap != BTREE_BITMAP_ANY &&
 		    s->btree_bitmap != bch2_dev_btree_bitmap_marked_sectors(ca,
 				bucket_to_sector(ca, bucket), ca->mi.bucket_size)) {
@@ -452,7 +363,10 @@ again:
 
 		s->buckets_seen++;
 
-		ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, watermark, a, s, cl);
+		ob = may_alloc_bucket(c, k.k->p, s)
+			? __try_alloc_bucket(c, ca, k.k->p.offset, a->gen,
+					     watermark, s, cl)
+			: NULL;
 next:
 		bch2_set_btree_iter_dontneed(&citer);
 		bch2_trans_iter_exit(trans, &citer);
@@ -489,20 +403,21 @@ static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
 	u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(*dev_alloc_cursor));
 	u64 alloc_cursor = alloc_start;
 	int ret;
-
-	BUG_ON(ca->new_fs_bucket_idx);
 again:
-	for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
-				     POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
-		if (k.k->p.inode != ca->dev_idx)
-			break;
+	for_each_btree_key_max_norestart(trans, iter, BTREE_ID_freespace,
+					 POS(ca->dev_idx, alloc_cursor),
+					 POS(ca->dev_idx, U64_MAX),
+					 0, k, ret) {
+		/*
+		 * peek normally dosen't trim extents - they can span iter.pos,
+		 * which is not what we want here:
+		 */
+		iter.k.size = iter.k.p.offset - iter.pos.offset;
 
-		for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
-		     alloc_cursor < k.k->p.offset;
-		     alloc_cursor++) {
+		while (iter.k.size) {
 			s->buckets_seen++;
 
-			u64 bucket = alloc_cursor & ~(~0ULL << 56);
+			u64 bucket = iter.pos.offset & ~(~0ULL << 56);
 			if (s->btree_bitmap != BTREE_BITMAP_ANY &&
 			    s->btree_bitmap != bch2_dev_btree_bitmap_marked_sectors(ca,
 					bucket_to_sector(ca, bucket), ca->mi.bucket_size)) {
@@ -511,32 +426,36 @@ again:
 					goto fail;
 
 				bucket = sector_to_bucket(ca,
-						round_up(bucket_to_sector(ca, bucket) + 1,
+						round_up(bucket_to_sector(ca, bucket + 1),
 							 1ULL << ca->mi.btree_bitmap_shift));
-				u64 genbits = alloc_cursor >> 56;
-				alloc_cursor = bucket | (genbits << 56);
+				alloc_cursor = bucket|(iter.pos.offset & (~0ULL << 56));
 
-				if (alloc_cursor > k.k->p.offset)
-					bch2_btree_iter_set_pos(&iter, POS(ca->dev_idx, alloc_cursor));
+				bch2_btree_iter_set_pos(&iter, POS(ca->dev_idx, alloc_cursor));
 				s->skipped_mi_btree_bitmap++;
-				continue;
+				goto next;
 			}
 
-			ob = try_alloc_bucket(trans, ca, watermark,
-					      alloc_cursor, s, k, cl);
+			ob = try_alloc_bucket(trans, ca, watermark, s, &iter, cl);
 			if (ob) {
+				if (!IS_ERR(ob))
+					*dev_alloc_cursor = iter.pos.offset;
 				bch2_set_btree_iter_dontneed(&iter);
 				break;
 			}
-		}
 
+			iter.k.size--;
+			iter.pos.offset++;
+		}
+next:
 		if (ob || ret)
 			break;
 	}
 fail:
 	bch2_trans_iter_exit(trans, &iter);
 
-	if (!ob && ret)
+	BUG_ON(ob && ret);
+
+	if (ret)
 		ob = ERR_PTR(ret);
 
 	if (!ob && alloc_start > ca->mi.first_bucket) {
@@ -544,8 +463,6 @@ fail:
 		goto again;
 	}
 
-	*dev_alloc_cursor = alloc_cursor;
-
 	return ob;
 }
 
@@ -595,6 +512,7 @@ static noinline void trace_bucket_alloc2(struct bch_fs *c, struct bch_dev *ca,
  * @watermark:	how important is this allocation?
  * @data_type:	BCH_DATA_journal, btree, user...
  * @cl:		if not NULL, closure to be used to wait if buckets not available
+ * @nowait:	if true, do not wait for buckets to become available
  * @usage:	for secondarily also returning the current device usage
  *
  * Returns:	an open_bucket on success, or an ERR_PTR() on failure.
@@ -629,6 +547,10 @@ again:
 		bch2_dev_do_invalidates(ca);
 
 	if (!avail) {
+		if (watermark > BCH_WATERMARK_normal &&
+		    c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_allocations)
+			goto alloc;
+
 		if (cl && !waiting) {
 			closure_wait(&c->freelist_wait, cl);
 			waiting = true;
@@ -711,9 +633,9 @@ struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 	unsigned i;
 
 	for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
-		ret.devs[ret.nr++] = i;
+		ret.data[ret.nr++] = i;
 
-	bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
+	bubble_sort(ret.data, ret.nr, dev_stripe_cmp);
 	return ret;
 }
 
@@ -785,18 +707,13 @@ int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 		      struct closure *cl)
 {
 	struct bch_fs *c = trans->c;
-	struct dev_alloc_list devs_sorted =
-		bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 	int ret = -BCH_ERR_insufficient_devices;
 
 	BUG_ON(*nr_effective >= nr_replicas);
 
-	for (unsigned i = 0; i < devs_sorted.nr; i++) {
-		struct bch_dev_usage usage;
-		struct open_bucket *ob;
-
-		unsigned dev = devs_sorted.devs[i];
-		struct bch_dev *ca = bch2_dev_tryget_noerror(c, dev);
+	struct dev_alloc_list devs_sorted = bch2_dev_alloc_list(c, stripe, devs_may_alloc);
+	darray_for_each(devs_sorted, i) {
+		struct bch_dev *ca = bch2_dev_tryget_noerror(c, *i);
 		if (!ca)
 			continue;
 
@@ -805,8 +722,9 @@ int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 			continue;
 		}
 
-		ob = bch2_bucket_alloc_trans(trans, ca, watermark, data_type,
-					     cl, flags & BCH_WRITE_ALLOC_NOWAIT, &usage);
+		struct bch_dev_usage usage;
+		struct open_bucket *ob = bch2_bucket_alloc_trans(trans, ca, watermark, data_type,
+						     cl, flags & BCH_WRITE_ALLOC_NOWAIT, &usage);
 		if (!IS_ERR(ob))
 			bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 		bch2_dev_put(ca);
@@ -850,10 +768,6 @@ static int bucket_alloc_from_stripe(struct btree_trans *trans,
 			 struct closure *cl)
 {
 	struct bch_fs *c = trans->c;
-	struct dev_alloc_list devs_sorted;
-	struct ec_stripe_head *h;
-	struct open_bucket *ob;
-	unsigned i, ec_idx;
 	int ret = 0;
 
 	if (nr_replicas < 2)
@@ -862,34 +776,32 @@ static int bucket_alloc_from_stripe(struct btree_trans *trans,
 	if (ec_open_bucket(c, ptrs))
 		return 0;
 
-	h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
+	struct ec_stripe_head *h =
+		bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
 	if (IS_ERR(h))
 		return PTR_ERR(h);
 	if (!h)
 		return 0;
 
-	devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
-
-	for (i = 0; i < devs_sorted.nr; i++)
-		for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
+	struct dev_alloc_list devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
+	darray_for_each(devs_sorted, i)
+		for (unsigned ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 			if (!h->s->blocks[ec_idx])
 				continue;
 
-			ob = c->open_buckets + h->s->blocks[ec_idx];
-			if (ob->dev == devs_sorted.devs[i] &&
-			    !test_and_set_bit(ec_idx, h->s->blocks_allocated))
-				goto got_bucket;
+			struct open_bucket *ob = c->open_buckets + h->s->blocks[ec_idx];
+			if (ob->dev == *i && !test_and_set_bit(ec_idx, h->s->blocks_allocated)) {
+				ob->ec_idx	= ec_idx;
+				ob->ec		= h->s;
+				ec_stripe_new_get(h->s, STRIPE_REF_io);
+
+				ret = add_new_bucket(c, ptrs, devs_may_alloc,
+						     nr_replicas, nr_effective,
+						     have_cache, ob);
+				goto out;
+			}
 		}
-	goto out_put_head;
-got_bucket:
-	ob->ec_idx	= ec_idx;
-	ob->ec		= h->s;
-	ec_stripe_new_get(h->s, STRIPE_REF_io);
-
-	ret = add_new_bucket(c, ptrs, devs_may_alloc,
-			     nr_replicas, nr_effective,
-			     have_cache, ob);
-out_put_head:
+out:
 	bch2_ec_stripe_head_put(c, h);
 	return ret;
 }
diff --git a/fs/bcachefs/alloc_foreground.h b/fs/bcachefs/alloc_foreground.h
index 1a16fd5bd4f8..f25481a0d1a0 100644
--- a/fs/bcachefs/alloc_foreground.h
+++ b/fs/bcachefs/alloc_foreground.h
@@ -20,7 +20,7 @@ void bch2_reset_alloc_cursors(struct bch_fs *);
 
 struct dev_alloc_list {
 	unsigned	nr;
-	u8		devs[BCH_SB_MEMBERS_MAX];
+	u8		data[BCH_SB_MEMBERS_MAX];
 };
 
 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *,
@@ -28,8 +28,6 @@ struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *,
 					  struct bch_devs_mask *);
 void bch2_dev_stripe_increment(struct bch_dev *, struct dev_stripe_state *);
 
-long bch2_bucket_alloc_new_fs(struct bch_dev *);
-
 static inline struct bch_dev *ob_dev(struct bch_fs *c, struct open_bucket *ob)
 {
 	return bch2_dev_have_ref(c, ob->dev);
diff --git a/fs/bcachefs/backpointers.c b/fs/bcachefs/backpointers.c
index 654a58132a4d..ebeb6a5ff9d2 100644
--- a/fs/bcachefs/backpointers.c
+++ b/fs/bcachefs/backpointers.c
@@ -14,42 +14,8 @@
 
 #include <linux/mm.h>
 
-static bool extent_matches_bp(struct bch_fs *c,
-			      enum btree_id btree_id, unsigned level,
-			      struct bkey_s_c k,
-			      struct bpos bucket,
-			      struct bch_backpointer bp)
-{
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const union bch_extent_entry *entry;
-	struct extent_ptr_decoded p;
-
-	rcu_read_lock();
-	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-		struct bpos bucket2;
-		struct bch_backpointer bp2;
-
-		if (p.ptr.cached)
-			continue;
-
-		struct bch_dev *ca = bch2_dev_rcu(c, p.ptr.dev);
-		if (!ca)
-			continue;
-
-		bch2_extent_ptr_to_bp(c, ca, btree_id, level, k, p, entry, &bucket2, &bp2);
-		if (bpos_eq(bucket, bucket2) &&
-		    !memcmp(&bp, &bp2, sizeof(bp))) {
-			rcu_read_unlock();
-			return true;
-		}
-	}
-	rcu_read_unlock();
-
-	return false;
-}
-
 int bch2_backpointer_validate(struct bch_fs *c, struct bkey_s_c k,
-			      enum bch_validate_flags flags)
+			      struct bkey_validate_context from)
 {
 	struct bkey_s_c_backpointer bp = bkey_s_c_to_backpointer(k);
 	int ret = 0;
@@ -59,67 +25,70 @@ int bch2_backpointer_validate(struct bch_fs *c, struct bkey_s_c k,
 			 "backpointer level bad: %u >= %u",
 			 bp.v->level, BTREE_MAX_DEPTH);
 
-	rcu_read_lock();
-	struct bch_dev *ca = bch2_dev_rcu_noerror(c, bp.k->p.inode);
-	if (!ca) {
-		/* these will be caught by fsck */
-		rcu_read_unlock();
-		return 0;
-	}
-
-	struct bpos bucket = bp_pos_to_bucket(ca, bp.k->p);
-	struct bpos bp_pos = bucket_pos_to_bp_noerror(ca, bucket, bp.v->bucket_offset);
-	rcu_read_unlock();
-
-	bkey_fsck_err_on((bp.v->bucket_offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT) >= ca->mi.bucket_size ||
-			 !bpos_eq(bp.k->p, bp_pos),
-			 c, backpointer_bucket_offset_wrong,
-			 "backpointer bucket_offset wrong");
+	bkey_fsck_err_on(bp.k->p.inode == BCH_SB_MEMBER_INVALID,
+			 c, backpointer_dev_bad,
+			 "backpointer for BCH_SB_MEMBER_INVALID");
 fsck_err:
 	return ret;
 }
 
-void bch2_backpointer_to_text(struct printbuf *out, const struct bch_backpointer *bp)
+void bch2_backpointer_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 {
-	prt_printf(out, "btree=%s l=%u offset=%llu:%u len=%u pos=",
-	       bch2_btree_id_str(bp->btree_id),
-	       bp->level,
-	       (u64) (bp->bucket_offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT),
-	       (u32) bp->bucket_offset & ~(~0U << MAX_EXTENT_COMPRESS_RATIO_SHIFT),
-	       bp->bucket_len);
-	bch2_bpos_to_text(out, bp->pos);
-}
+	struct bkey_s_c_backpointer bp = bkey_s_c_to_backpointer(k);
 
-void bch2_backpointer_k_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
-{
 	rcu_read_lock();
-	struct bch_dev *ca = bch2_dev_rcu_noerror(c, k.k->p.inode);
+	struct bch_dev *ca = bch2_dev_rcu_noerror(c, bp.k->p.inode);
 	if (ca) {
-		struct bpos bucket = bp_pos_to_bucket(ca, k.k->p);
+		u32 bucket_offset;
+		struct bpos bucket = bp_pos_to_bucket_and_offset(ca, bp.k->p, &bucket_offset);
 		rcu_read_unlock();
-		prt_str(out, "bucket=");
-		bch2_bpos_to_text(out, bucket);
-		prt_str(out, " ");
+		prt_printf(out, "bucket=%llu:%llu:%u ", bucket.inode, bucket.offset, bucket_offset);
 	} else {
 		rcu_read_unlock();
+		prt_printf(out, "sector=%llu:%llu ", bp.k->p.inode, bp.k->p.offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT);
 	}
 
-	bch2_backpointer_to_text(out, bkey_s_c_to_backpointer(k).v);
+	bch2_btree_id_level_to_text(out, bp.v->btree_id, bp.v->level);
+	prt_printf(out, " suboffset=%u len=%u gen=%u pos=",
+		   (u32) bp.k->p.offset & ~(~0U << MAX_EXTENT_COMPRESS_RATIO_SHIFT),
+		   bp.v->bucket_len,
+		   bp.v->bucket_gen);
+	bch2_bpos_to_text(out, bp.v->pos);
 }
 
 void bch2_backpointer_swab(struct bkey_s k)
 {
 	struct bkey_s_backpointer bp = bkey_s_to_backpointer(k);
 
-	bp.v->bucket_offset	= swab40(bp.v->bucket_offset);
 	bp.v->bucket_len	= swab32(bp.v->bucket_len);
 	bch2_bpos_swab(&bp.v->pos);
 }
 
+static bool extent_matches_bp(struct bch_fs *c,
+			      enum btree_id btree_id, unsigned level,
+			      struct bkey_s_c k,
+			      struct bkey_s_c_backpointer bp)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
+
+	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+		struct bkey_i_backpointer bp2;
+		bch2_extent_ptr_to_bp(c, btree_id, level, k, p, entry, &bp2);
+
+		if (bpos_eq(bp.k->p, bp2.k.p) &&
+		    !memcmp(bp.v, &bp2.v, sizeof(bp2.v)))
+			return true;
+	}
+
+	return false;
+}
+
 static noinline int backpointer_mod_err(struct btree_trans *trans,
-					struct bch_backpointer bp,
-					struct bkey_s_c bp_k,
 					struct bkey_s_c orig_k,
+					struct bkey_i_backpointer *new_bp,
+					struct bkey_s_c found_bp,
 					bool insert)
 {
 	struct bch_fs *c = trans->c;
@@ -127,12 +96,12 @@ static noinline int backpointer_mod_err(struct btree_trans *trans,
 
 	if (insert) {
 		prt_printf(&buf, "existing backpointer found when inserting ");
-		bch2_backpointer_to_text(&buf, &bp);
+		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&new_bp->k_i));
 		prt_newline(&buf);
 		printbuf_indent_add(&buf, 2);
 
 		prt_printf(&buf, "found ");
-		bch2_bkey_val_to_text(&buf, c, bp_k);
+		bch2_bkey_val_to_text(&buf, c, found_bp);
 		prt_newline(&buf);
 
 		prt_printf(&buf, "for ");
@@ -144,11 +113,11 @@ static noinline int backpointer_mod_err(struct btree_trans *trans,
 		printbuf_indent_add(&buf, 2);
 
 		prt_printf(&buf, "searching for ");
-		bch2_backpointer_to_text(&buf, &bp);
+		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&new_bp->k_i));
 		prt_newline(&buf);
 
 		prt_printf(&buf, "got ");
-		bch2_bkey_val_to_text(&buf, c, bp_k);
+		bch2_bkey_val_to_text(&buf, c, found_bp);
 		prt_newline(&buf);
 
 		prt_printf(&buf, "for ");
@@ -167,161 +136,118 @@ static noinline int backpointer_mod_err(struct btree_trans *trans,
 }
 
 int bch2_bucket_backpointer_mod_nowritebuffer(struct btree_trans *trans,
-				struct bch_dev *ca,
-				struct bpos bucket,
-				struct bch_backpointer bp,
 				struct bkey_s_c orig_k,
+				struct bkey_i_backpointer *bp,
 				bool insert)
 {
 	struct btree_iter bp_iter;
-	struct bkey_s_c k;
-	struct bkey_i_backpointer *bp_k;
-	int ret;
-
-	bp_k = bch2_trans_kmalloc_nomemzero(trans, sizeof(struct bkey_i_backpointer));
-	ret = PTR_ERR_OR_ZERO(bp_k);
-	if (ret)
-		return ret;
-
-	bkey_backpointer_init(&bp_k->k_i);
-	bp_k->k.p = bucket_pos_to_bp(ca, bucket, bp.bucket_offset);
-	bp_k->v = bp;
-
-	if (!insert) {
-		bp_k->k.type = KEY_TYPE_deleted;
-		set_bkey_val_u64s(&bp_k->k, 0);
-	}
-
-	k = bch2_bkey_get_iter(trans, &bp_iter, BTREE_ID_backpointers,
-			       bp_k->k.p,
+	struct bkey_s_c k = bch2_bkey_get_iter(trans, &bp_iter, BTREE_ID_backpointers,
+			       bp->k.p,
 			       BTREE_ITER_intent|
 			       BTREE_ITER_slots|
 			       BTREE_ITER_with_updates);
-	ret = bkey_err(k);
+	int ret = bkey_err(k);
 	if (ret)
-		goto err;
+		return ret;
 
 	if (insert
 	    ? k.k->type
 	    : (k.k->type != KEY_TYPE_backpointer ||
-	       memcmp(bkey_s_c_to_backpointer(k).v, &bp, sizeof(bp)))) {
-		ret = backpointer_mod_err(trans, bp, k, orig_k, insert);
+	       memcmp(bkey_s_c_to_backpointer(k).v, &bp->v, sizeof(bp->v)))) {
+		ret = backpointer_mod_err(trans, orig_k, bp, k, insert);
 		if (ret)
 			goto err;
 	}
 
-	ret = bch2_trans_update(trans, &bp_iter, &bp_k->k_i, 0);
+	if (!insert) {
+		bp->k.type = KEY_TYPE_deleted;
+		set_bkey_val_u64s(&bp->k, 0);
+	}
+
+	ret = bch2_trans_update(trans, &bp_iter, &bp->k_i, 0);
 err:
 	bch2_trans_iter_exit(trans, &bp_iter);
 	return ret;
 }
 
-/*
- * Find the next backpointer >= *bp_offset:
- */
-int bch2_get_next_backpointer(struct btree_trans *trans,
-			      struct bch_dev *ca,
-			      struct bpos bucket, int gen,
-			      struct bpos *bp_pos,
-			      struct bch_backpointer *bp,
-			      unsigned iter_flags)
+static int bch2_backpointer_del(struct btree_trans *trans, struct bpos pos)
 {
-	struct bpos bp_end_pos = bucket_pos_to_bp(ca, bpos_nosnap_successor(bucket), 0);
-	struct btree_iter alloc_iter = { NULL }, bp_iter = { NULL };
-	struct bkey_s_c k;
-	int ret = 0;
-
-	if (bpos_ge(*bp_pos, bp_end_pos))
-		goto done;
-
-	if (gen >= 0) {
-		k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc,
-				       bucket, BTREE_ITER_cached|iter_flags);
-		ret = bkey_err(k);
-		if (ret)
-			goto out;
-
-		if (k.k->type != KEY_TYPE_alloc_v4 ||
-		    bkey_s_c_to_alloc_v4(k).v->gen != gen)
-			goto done;
-	}
-
-	*bp_pos = bpos_max(*bp_pos, bucket_pos_to_bp(ca, bucket, 0));
-
-	for_each_btree_key_norestart(trans, bp_iter, BTREE_ID_backpointers,
-				     *bp_pos, iter_flags, k, ret) {
-		if (bpos_ge(k.k->p, bp_end_pos))
-			break;
+	return (likely(!bch2_backpointers_no_use_write_buffer)
+		? bch2_btree_delete_at_buffered(trans, BTREE_ID_backpointers, pos)
+		: bch2_btree_delete(trans, BTREE_ID_backpointers, pos, 0)) ?:
+		 bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
+}
 
-		*bp_pos = k.k->p;
-		*bp = *bkey_s_c_to_backpointer(k).v;
-		goto out;
-	}
-done:
-	*bp_pos = SPOS_MAX;
-out:
-	bch2_trans_iter_exit(trans, &bp_iter);
-	bch2_trans_iter_exit(trans, &alloc_iter);
-	return ret;
+static inline int bch2_backpointers_maybe_flush(struct btree_trans *trans,
+					 struct bkey_s_c visiting_k,
+					 struct bkey_buf *last_flushed)
+{
+	return likely(!bch2_backpointers_no_use_write_buffer)
+		? bch2_btree_write_buffer_maybe_flush(trans, visiting_k, last_flushed)
+		: 0;
 }
 
-static void backpointer_not_found(struct btree_trans *trans,
-				  struct bpos bp_pos,
-				  struct bch_backpointer bp,
-				  struct bkey_s_c k)
+static int backpointer_target_not_found(struct btree_trans *trans,
+				  struct bkey_s_c_backpointer bp,
+				  struct bkey_s_c target_k,
+				  struct bkey_buf *last_flushed)
 {
 	struct bch_fs *c = trans->c;
 	struct printbuf buf = PRINTBUF;
+	int ret = 0;
 
 	/*
 	 * If we're using the btree write buffer, the backpointer we were
 	 * looking at may have already been deleted - failure to find what it
 	 * pointed to is not an error:
 	 */
-	if (likely(!bch2_backpointers_no_use_write_buffer))
-		return;
-
-	struct bpos bucket;
-	if (!bp_pos_to_bucket_nodev(c, bp_pos, &bucket))
-		return;
+	ret = last_flushed
+		? bch2_backpointers_maybe_flush(trans, bp.s_c, last_flushed)
+		: 0;
+	if (ret)
+		return ret;
 
 	prt_printf(&buf, "backpointer doesn't match %s it points to:\n  ",
-		   bp.level ? "btree node" : "extent");
-	prt_printf(&buf, "bucket: ");
-	bch2_bpos_to_text(&buf, bucket);
-	prt_printf(&buf, "\n  ");
+		   bp.v->level ? "btree node" : "extent");
+	bch2_bkey_val_to_text(&buf, c, bp.s_c);
 
-	prt_printf(&buf, "backpointer pos: ");
-	bch2_bpos_to_text(&buf, bp_pos);
 	prt_printf(&buf, "\n  ");
+	bch2_bkey_val_to_text(&buf, c, target_k);
 
-	bch2_backpointer_to_text(&buf, &bp);
-	prt_printf(&buf, "\n  ");
-	bch2_bkey_val_to_text(&buf, c, k);
-	if (c->curr_recovery_pass >= BCH_RECOVERY_PASS_check_extents_to_backpointers)
-		bch_err_ratelimited(c, "%s", buf.buf);
-	else
-		bch2_trans_inconsistent(trans, "%s", buf.buf);
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(target_k);
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
+	bkey_for_each_ptr_decode(target_k.k, ptrs, p, entry)
+		if (p.ptr.dev == bp.k->p.inode) {
+			prt_printf(&buf, "\n  ");
+			struct bkey_i_backpointer bp2;
+			bch2_extent_ptr_to_bp(c, bp.v->btree_id, bp.v->level, target_k, p, entry, &bp2);
+			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&bp2.k_i));
+		}
 
+	if (fsck_err(trans, backpointer_to_missing_ptr,
+		     "%s", buf.buf))
+		ret = bch2_backpointer_del(trans, bp.k->p);
+fsck_err:
 	printbuf_exit(&buf);
+	return ret;
 }
 
 struct bkey_s_c bch2_backpointer_get_key(struct btree_trans *trans,
+					 struct bkey_s_c_backpointer bp,
 					 struct btree_iter *iter,
-					 struct bpos bp_pos,
-					 struct bch_backpointer bp,
-					 unsigned iter_flags)
+					 unsigned iter_flags,
+					 struct bkey_buf *last_flushed)
 {
-	if (likely(!bp.level)) {
-		struct bch_fs *c = trans->c;
+	struct bch_fs *c = trans->c;
 
-		struct bpos bucket;
-		if (!bp_pos_to_bucket_nodev(c, bp_pos, &bucket))
-			return bkey_s_c_err(-EIO);
+	if (unlikely(bp.v->btree_id >= btree_id_nr_alive(c)))
+		return bkey_s_c_null;
 
+	if (likely(!bp.v->level)) {
 		bch2_trans_node_iter_init(trans, iter,
-					  bp.btree_id,
-					  bp.pos,
+					  bp.v->btree_id,
+					  bp.v->pos,
 					  0, 0,
 					  iter_flags);
 		struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
@@ -330,67 +256,64 @@ struct bkey_s_c bch2_backpointer_get_key(struct btree_trans *trans,
 			return k;
 		}
 
-		if (k.k && extent_matches_bp(c, bp.btree_id, bp.level, k, bucket, bp))
+		if (k.k &&
+		    extent_matches_bp(c, bp.v->btree_id, bp.v->level, k, bp))
 			return k;
 
 		bch2_trans_iter_exit(trans, iter);
-		backpointer_not_found(trans, bp_pos, bp, k);
-		return bkey_s_c_null;
+		int ret = backpointer_target_not_found(trans, bp, k, last_flushed);
+		return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 	} else {
-		struct btree *b = bch2_backpointer_get_node(trans, iter, bp_pos, bp);
+		struct btree *b = bch2_backpointer_get_node(trans, bp, iter, last_flushed);
+		if (IS_ERR_OR_NULL(b))
+			return ((struct bkey_s_c) { .k = ERR_CAST(b) });
 
-		if (IS_ERR_OR_NULL(b)) {
-			bch2_trans_iter_exit(trans, iter);
-			return IS_ERR(b) ? bkey_s_c_err(PTR_ERR(b)) : bkey_s_c_null;
-		}
 		return bkey_i_to_s_c(&b->key);
 	}
 }
 
 struct btree *bch2_backpointer_get_node(struct btree_trans *trans,
+					struct bkey_s_c_backpointer bp,
 					struct btree_iter *iter,
-					struct bpos bp_pos,
-					struct bch_backpointer bp)
+					struct bkey_buf *last_flushed)
 {
 	struct bch_fs *c = trans->c;
 
-	BUG_ON(!bp.level);
-
-	struct bpos bucket;
-	if (!bp_pos_to_bucket_nodev(c, bp_pos, &bucket))
-		return ERR_PTR(-EIO);
+	BUG_ON(!bp.v->level);
 
 	bch2_trans_node_iter_init(trans, iter,
-				  bp.btree_id,
-				  bp.pos,
+				  bp.v->btree_id,
+				  bp.v->pos,
 				  0,
-				  bp.level - 1,
+				  bp.v->level - 1,
 				  0);
 	struct btree *b = bch2_btree_iter_peek_node(iter);
 	if (IS_ERR_OR_NULL(b))
 		goto err;
 
-	BUG_ON(b->c.level != bp.level - 1);
+	BUG_ON(b->c.level != bp.v->level - 1);
 
-	if (extent_matches_bp(c, bp.btree_id, bp.level,
-			      bkey_i_to_s_c(&b->key),
-			      bucket, bp))
+	if (extent_matches_bp(c, bp.v->btree_id, bp.v->level,
+			      bkey_i_to_s_c(&b->key), bp))
 		return b;
 
 	if (btree_node_will_make_reachable(b)) {
 		b = ERR_PTR(-BCH_ERR_backpointer_to_overwritten_btree_node);
 	} else {
-		backpointer_not_found(trans, bp_pos, bp, bkey_i_to_s_c(&b->key));
-		b = NULL;
+		int ret = backpointer_target_not_found(trans, bp, bkey_i_to_s_c(&b->key), last_flushed);
+		b = ret ? ERR_PTR(ret) : NULL;
 	}
 err:
 	bch2_trans_iter_exit(trans, iter);
 	return b;
 }
 
-static int bch2_check_btree_backpointer(struct btree_trans *trans, struct btree_iter *bp_iter,
-					struct bkey_s_c k)
+static int bch2_check_backpointer_has_valid_bucket(struct btree_trans *trans, struct bkey_s_c k,
+						   struct bkey_buf *last_flushed)
 {
+	if (k.k->type != KEY_TYPE_backpointer)
+		return 0;
+
 	struct bch_fs *c = trans->c;
 	struct btree_iter alloc_iter = { NULL };
 	struct bkey_s_c alloc_k;
@@ -399,10 +322,14 @@ static int bch2_check_btree_backpointer(struct btree_trans *trans, struct btree_
 
 	struct bpos bucket;
 	if (!bp_pos_to_bucket_nodev_noerror(c, k.k->p, &bucket)) {
+		ret = bch2_backpointers_maybe_flush(trans, k, last_flushed);
+		if (ret)
+			goto out;
+
 		if (fsck_err(trans, backpointer_to_missing_device,
 			     "backpointer for missing device:\n%s",
 			     (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
-			ret = bch2_btree_delete_at(trans, bp_iter, 0);
+			ret = bch2_backpointer_del(trans, k.k->p);
 		goto out;
 	}
 
@@ -411,13 +338,16 @@ static int bch2_check_btree_backpointer(struct btree_trans *trans, struct btree_
 	if (ret)
 		goto out;
 
-	if (fsck_err_on(alloc_k.k->type != KEY_TYPE_alloc_v4,
-			trans, backpointer_to_missing_alloc,
-			"backpointer for nonexistent alloc key: %llu:%llu:0\n%s",
-			alloc_iter.pos.inode, alloc_iter.pos.offset,
-			(bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
-		ret = bch2_btree_delete_at(trans, bp_iter, 0);
-		goto out;
+	if (alloc_k.k->type != KEY_TYPE_alloc_v4) {
+		ret = bch2_backpointers_maybe_flush(trans, k, last_flushed);
+		if (ret)
+			goto out;
+
+		if (fsck_err(trans, backpointer_to_missing_alloc,
+			     "backpointer for nonexistent alloc key: %llu:%llu:0\n%s",
+			     alloc_iter.pos.inode, alloc_iter.pos.offset,
+			     (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
+			ret = bch2_backpointer_del(trans, k.k->p);
 	}
 out:
 fsck_err:
@@ -429,18 +359,24 @@ fsck_err:
 /* verify that every backpointer has a corresponding alloc key */
 int bch2_check_btree_backpointers(struct bch_fs *c)
 {
+	struct bkey_buf last_flushed;
+	bch2_bkey_buf_init(&last_flushed);
+	bkey_init(&last_flushed.k->k);
+
 	int ret = bch2_trans_run(c,
 		for_each_btree_key_commit(trans, iter,
 			BTREE_ID_backpointers, POS_MIN, 0, k,
 			NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
-		  bch2_check_btree_backpointer(trans, &iter, k)));
+		  bch2_check_backpointer_has_valid_bucket(trans, k, &last_flushed)));
+
+	bch2_bkey_buf_exit(&last_flushed, c);
 	bch_err_fn(c, ret);
 	return ret;
 }
 
 struct extents_to_bp_state {
-	struct bpos	bucket_start;
-	struct bpos	bucket_end;
+	struct bpos	bp_start;
+	struct bpos	bp_end;
 	struct bkey_buf last_flushed;
 };
 
@@ -501,9 +437,13 @@ found:
 		goto err;
 
 	prt_str(&buf, "extents pointing to same space, but first extent checksum bad:");
-	prt_printf(&buf, "\n  %s ", bch2_btree_id_str(btree));
+	prt_printf(&buf, "\n  ");
+	bch2_btree_id_to_text(&buf, btree);
+	prt_str(&buf, " ");
 	bch2_bkey_val_to_text(&buf, c, extent);
-	prt_printf(&buf, "\n  %s ", bch2_btree_id_str(o_btree));
+	prt_printf(&buf, "\n  ");
+	bch2_btree_id_to_text(&buf, o_btree);
+	prt_str(&buf, " ");
 	bch2_bkey_val_to_text(&buf, c, extent2);
 
 	struct nonce nonce = extent_nonce(extent.k->bversion, p.crc);
@@ -524,41 +464,25 @@ err:
 
 static int check_bp_exists(struct btree_trans *trans,
 			   struct extents_to_bp_state *s,
-			   struct bpos bucket,
-			   struct bch_backpointer bp,
+			   struct bkey_i_backpointer *bp,
 			   struct bkey_s_c orig_k)
 {
 	struct bch_fs *c = trans->c;
-	struct btree_iter bp_iter = {};
 	struct btree_iter other_extent_iter = {};
 	struct printbuf buf = PRINTBUF;
-	struct bkey_s_c bp_k;
-	int ret = 0;
 
-	struct bch_dev *ca = bch2_dev_bucket_tryget(c, bucket);
-	if (!ca) {
-		prt_str(&buf, "extent for nonexistent device:bucket ");
-		bch2_bpos_to_text(&buf, bucket);
-		prt_str(&buf, "\n  ");
-		bch2_bkey_val_to_text(&buf, c, orig_k);
-		bch_err(c, "%s", buf.buf);
-		ret = -BCH_ERR_fsck_repair_unimplemented;
-		goto err;
-	}
-
-	if (bpos_lt(bucket, s->bucket_start) ||
-	    bpos_gt(bucket, s->bucket_end))
-		goto out;
+	if (bpos_lt(bp->k.p, s->bp_start) ||
+	    bpos_gt(bp->k.p, s->bp_end))
+		return 0;
 
-	bp_k = bch2_bkey_get_iter(trans, &bp_iter, BTREE_ID_backpointers,
-				  bucket_pos_to_bp(ca, bucket, bp.bucket_offset),
-				  0);
-	ret = bkey_err(bp_k);
+	struct btree_iter bp_iter;
+	struct bkey_s_c bp_k = bch2_bkey_get_iter(trans, &bp_iter, BTREE_ID_backpointers, bp->k.p, 0);
+	int ret = bkey_err(bp_k);
 	if (ret)
 		goto err;
 
 	if (bp_k.k->type != KEY_TYPE_backpointer ||
-	    memcmp(bkey_s_c_to_backpointer(bp_k).v, &bp, sizeof(bp))) {
+	    memcmp(bkey_s_c_to_backpointer(bp_k).v, &bp->v, sizeof(bp->v))) {
 		ret = bch2_btree_write_buffer_maybe_flush(trans, orig_k, &s->last_flushed);
 		if (ret)
 			goto err;
@@ -570,7 +494,6 @@ err:
 fsck_err:
 	bch2_trans_iter_exit(trans, &other_extent_iter);
 	bch2_trans_iter_exit(trans, &bp_iter);
-	bch2_dev_put(ca);
 	printbuf_exit(&buf);
 	return ret;
 check_existing_bp:
@@ -578,10 +501,10 @@ check_existing_bp:
 	if (bp_k.k->type != KEY_TYPE_backpointer)
 		goto missing;
 
-	struct bch_backpointer other_bp = *bkey_s_c_to_backpointer(bp_k).v;
+	struct bkey_s_c_backpointer other_bp = bkey_s_c_to_backpointer(bp_k);
 
 	struct bkey_s_c other_extent =
-		bch2_backpointer_get_key(trans, &other_extent_iter, bp_k.k->p, other_bp, 0);
+		bch2_backpointer_get_key(trans, other_bp, &other_extent_iter, 0, NULL);
 	ret = bkey_err(other_extent);
 	if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 		ret = 0;
@@ -600,19 +523,23 @@ check_existing_bp:
 		bch_err(c, "%s", buf.buf);
 
 		if (other_extent.k->size <= orig_k.k->size) {
-			ret = drop_dev_and_update(trans, other_bp.btree_id, other_extent, bucket.inode);
+			ret = drop_dev_and_update(trans, other_bp.v->btree_id,
+						  other_extent, bp->k.p.inode);
 			if (ret)
 				goto err;
 			goto out;
 		} else {
-			ret = drop_dev_and_update(trans, bp.btree_id, orig_k, bucket.inode);
+			ret = drop_dev_and_update(trans, bp->v.btree_id, orig_k, bp->k.p.inode);
 			if (ret)
 				goto err;
 			goto missing;
 		}
 	}
 
-	ret = check_extent_checksum(trans, other_bp.btree_id, other_extent, bp.btree_id, orig_k, bucket.inode);
+	ret = check_extent_checksum(trans,
+				    other_bp.v->btree_id, other_extent,
+				    bp->v.btree_id, orig_k,
+				    bp->k.p.inode);
 	if (ret < 0)
 		goto err;
 	if (ret) {
@@ -620,7 +547,8 @@ check_existing_bp:
 		goto missing;
 	}
 
-	ret = check_extent_checksum(trans, bp.btree_id, orig_k, other_bp.btree_id, other_extent, bucket.inode);
+	ret = check_extent_checksum(trans, bp->v.btree_id, orig_k,
+				    other_bp.v->btree_id, other_extent, bp->k.p.inode);
 	if (ret < 0)
 		goto err;
 	if (ret) {
@@ -629,7 +557,7 @@ check_existing_bp:
 	}
 
 	printbuf_reset(&buf);
-	prt_printf(&buf, "duplicate extents pointing to same space on dev %llu\n  ", bucket.inode);
+	prt_printf(&buf, "duplicate extents pointing to same space on dev %llu\n  ", bp->k.p.inode);
 	bch2_bkey_val_to_text(&buf, c, orig_k);
 	prt_str(&buf, "\n  ");
 	bch2_bkey_val_to_text(&buf, c, other_extent);
@@ -638,21 +566,15 @@ check_existing_bp:
 	goto err;
 missing:
 	printbuf_reset(&buf);
-	prt_printf(&buf, "missing backpointer for btree=%s l=%u ",
-	       bch2_btree_id_str(bp.btree_id), bp.level);
+	prt_str(&buf, "missing backpointer\n  for:  ");
 	bch2_bkey_val_to_text(&buf, c, orig_k);
-	prt_printf(&buf, "\n  got:   ");
+	prt_printf(&buf, "\n  want: ");
+	bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&bp->k_i));
+	prt_printf(&buf, "\n  got:  ");
 	bch2_bkey_val_to_text(&buf, c, bp_k);
 
-	struct bkey_i_backpointer n_bp_k;
-	bkey_backpointer_init(&n_bp_k.k_i);
-	n_bp_k.k.p = bucket_pos_to_bp(ca, bucket, bp.bucket_offset);
-	n_bp_k.v = bp;
-	prt_printf(&buf, "\n  want:  ");
-	bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&n_bp_k.k_i));
-
 	if (fsck_err(trans, ptr_to_missing_backpointer, "%s", buf.buf))
-		ret = bch2_bucket_backpointer_mod(trans, ca, bucket, bp, orig_k, true);
+		ret = bch2_bucket_backpointer_mod(trans, orig_k, bp, true);
 
 	goto out;
 }
@@ -663,31 +585,33 @@ static int check_extent_to_backpointers(struct btree_trans *trans,
 					struct bkey_s_c k)
 {
 	struct bch_fs *c = trans->c;
-	struct bkey_ptrs_c ptrs;
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 	const union bch_extent_entry *entry;
 	struct extent_ptr_decoded p;
-	int ret;
 
-	ptrs = bch2_bkey_ptrs_c(k);
 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-		struct bpos bucket_pos = POS_MIN;
-		struct bch_backpointer bp;
-
 		if (p.ptr.cached)
 			continue;
 
+		if (p.ptr.dev == BCH_SB_MEMBER_INVALID)
+			continue;
+
 		rcu_read_lock();
 		struct bch_dev *ca = bch2_dev_rcu_noerror(c, p.ptr.dev);
-		if (ca)
-			bch2_extent_ptr_to_bp(c, ca, btree, level, k, p, entry, &bucket_pos, &bp);
+		bool check = ca && test_bit(PTR_BUCKET_NR(ca, &p.ptr), ca->bucket_backpointer_mismatches);
+		bool empty = ca && test_bit(PTR_BUCKET_NR(ca, &p.ptr), ca->bucket_backpointer_empty);
 		rcu_read_unlock();
 
-		if (!ca)
-			continue;
+		if (check || empty) {
+			struct bkey_i_backpointer bp;
+			bch2_extent_ptr_to_bp(c, btree, level, k, p, entry, &bp);
 
-		ret = check_bp_exists(trans, s, bucket_pos, bp, k);
-		if (ret)
-			return ret;
+			int ret = check
+				? check_bp_exists(trans, s, &bp, k)
+				: bch2_bucket_backpointer_mod(trans, k, &bp, true);
+			if (ret)
+				return ret;
+		}
 	}
 
 	return 0;
@@ -896,54 +820,330 @@ static int bch2_check_extents_to_backpointers_pass(struct btree_trans *trans,
 	return 0;
 }
 
+enum alloc_sector_counter {
+	ALLOC_dirty,
+	ALLOC_cached,
+	ALLOC_stripe,
+	ALLOC_SECTORS_NR
+};
+
+static enum alloc_sector_counter data_type_to_alloc_counter(enum bch_data_type t)
+{
+	switch (t) {
+	case BCH_DATA_btree:
+	case BCH_DATA_user:
+		return ALLOC_dirty;
+	case BCH_DATA_cached:
+		return ALLOC_cached;
+	case BCH_DATA_stripe:
+		return ALLOC_stripe;
+	default:
+		BUG();
+	}
+}
+
+static int check_bucket_backpointers_to_extents(struct btree_trans *, struct bch_dev *, struct bpos);
+
+static int check_bucket_backpointer_mismatch(struct btree_trans *trans, struct bkey_s_c alloc_k,
+					     struct bkey_buf *last_flushed)
+{
+	struct bch_fs *c = trans->c;
+	struct bch_alloc_v4 a_convert;
+	const struct bch_alloc_v4 *a = bch2_alloc_to_v4(alloc_k, &a_convert);
+	bool need_commit = false;
+
+	if (a->data_type == BCH_DATA_sb ||
+	    a->data_type == BCH_DATA_journal ||
+	    a->data_type == BCH_DATA_parity)
+		return 0;
+
+	u32 sectors[ALLOC_SECTORS_NR];
+	memset(sectors, 0, sizeof(sectors));
+
+	struct bch_dev *ca = bch2_dev_bucket_tryget_noerror(trans->c, alloc_k.k->p);
+	if (!ca)
+		return 0;
+
+	struct btree_iter iter;
+	struct bkey_s_c bp_k;
+	int ret = 0;
+	for_each_btree_key_max_norestart(trans, iter, BTREE_ID_backpointers,
+				bucket_pos_to_bp_start(ca, alloc_k.k->p),
+				bucket_pos_to_bp_end(ca, alloc_k.k->p), 0, bp_k, ret) {
+		if (bp_k.k->type != KEY_TYPE_backpointer)
+			continue;
+
+		struct bkey_s_c_backpointer bp = bkey_s_c_to_backpointer(bp_k);
+
+		if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_backpointer_bucket_gen &&
+		    (bp.v->bucket_gen != a->gen ||
+		     bp.v->pad)) {
+			ret = bch2_backpointer_del(trans, bp_k.k->p);
+			if (ret)
+				break;
+
+			need_commit = true;
+			continue;
+		}
+
+		if (bp.v->bucket_gen != a->gen)
+			continue;
+
+		sectors[data_type_to_alloc_counter(bp.v->data_type)] += bp.v->bucket_len;
+	};
+	bch2_trans_iter_exit(trans, &iter);
+	if (ret)
+		goto err;
+
+	if (need_commit) {
+		ret = bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
+		if (ret)
+			goto err;
+	}
+
+	/* Cached pointers don't have backpointers: */
+
+	if (sectors[ALLOC_dirty]  != a->dirty_sectors ||
+	    sectors[ALLOC_stripe] != a->stripe_sectors) {
+		if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_backpointer_bucket_gen) {
+			ret = bch2_backpointers_maybe_flush(trans, alloc_k, last_flushed);
+			if (ret)
+				goto err;
+		}
+
+		if (sectors[ALLOC_dirty]  > a->dirty_sectors ||
+		    sectors[ALLOC_stripe] > a->stripe_sectors) {
+			ret = check_bucket_backpointers_to_extents(trans, ca, alloc_k.k->p) ?:
+				-BCH_ERR_transaction_restart_nested;
+			goto err;
+		}
+
+		if (!sectors[ALLOC_dirty] &&
+		    !sectors[ALLOC_stripe])
+			__set_bit(alloc_k.k->p.offset, ca->bucket_backpointer_empty);
+		else
+			__set_bit(alloc_k.k->p.offset, ca->bucket_backpointer_mismatches);
+	}
+err:
+	bch2_dev_put(ca);
+	return ret;
+}
+
+static bool backpointer_node_has_missing(struct bch_fs *c, struct bkey_s_c k)
+{
+	switch (k.k->type) {
+	case KEY_TYPE_btree_ptr_v2: {
+		bool ret = false;
+
+		rcu_read_lock();
+		struct bpos pos = bkey_s_c_to_btree_ptr_v2(k).v->min_key;
+		while (pos.inode <= k.k->p.inode) {
+			if (pos.inode >= c->sb.nr_devices)
+				break;
+
+			struct bch_dev *ca = bch2_dev_rcu_noerror(c, pos.inode);
+			if (!ca)
+				goto next;
+
+			struct bpos bucket = bp_pos_to_bucket(ca, pos);
+			bucket.offset = find_next_bit(ca->bucket_backpointer_mismatches,
+						      ca->mi.nbuckets, bucket.offset);
+			if (bucket.offset == ca->mi.nbuckets)
+				goto next;
+
+			ret = bpos_le(bucket_pos_to_bp_end(ca, bucket), k.k->p);
+			if (ret)
+				break;
+next:
+			pos = SPOS(pos.inode + 1, 0, 0);
+		}
+		rcu_read_unlock();
+
+		return ret;
+	}
+	case KEY_TYPE_btree_ptr:
+		return true;
+	default:
+		return false;
+	}
+}
+
+static int btree_node_get_and_pin(struct btree_trans *trans, struct bkey_i *k,
+				  enum btree_id btree, unsigned level)
+{
+	struct btree_iter iter;
+	bch2_trans_node_iter_init(trans, &iter, btree, k->k.p, 0, level, 0);
+	struct btree *b = bch2_btree_iter_peek_node(&iter);
+	int ret = PTR_ERR_OR_ZERO(b);
+	if (ret)
+		goto err;
+
+	if (b)
+		bch2_node_pin(trans->c, b);
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+static int bch2_pin_backpointer_nodes_with_missing(struct btree_trans *trans,
+						   struct bpos start, struct bpos *end)
+{
+	struct bch_fs *c = trans->c;
+	int ret = 0;
+
+	struct bkey_buf tmp;
+	bch2_bkey_buf_init(&tmp);
+
+	bch2_btree_cache_unpin(c);
+
+	*end = SPOS_MAX;
+
+	s64 mem_may_pin = mem_may_pin_bytes(c);
+	struct btree_iter iter;
+	bch2_trans_node_iter_init(trans, &iter, BTREE_ID_backpointers, start,
+				  0, 1, BTREE_ITER_prefetch);
+	ret = for_each_btree_key_continue(trans, iter, 0, k, ({
+		if (!backpointer_node_has_missing(c, k))
+			continue;
+
+		mem_may_pin -= c->opts.btree_node_size;
+		if (mem_may_pin <= 0)
+			break;
+
+		bch2_bkey_buf_reassemble(&tmp, c, k);
+		struct btree_path *path = btree_iter_path(trans, &iter);
+
+		BUG_ON(path->level != 1);
+
+		bch2_btree_node_prefetch(trans, path, tmp.k, path->btree_id, path->level - 1);
+	}));
+	if (ret)
+		return ret;
+
+	struct bpos pinned = SPOS_MAX;
+	mem_may_pin = mem_may_pin_bytes(c);
+	bch2_trans_node_iter_init(trans, &iter, BTREE_ID_backpointers, start,
+				  0, 1, BTREE_ITER_prefetch);
+	ret = for_each_btree_key_continue(trans, iter, 0, k, ({
+		if (!backpointer_node_has_missing(c, k))
+			continue;
+
+		mem_may_pin -= c->opts.btree_node_size;
+		if (mem_may_pin <= 0) {
+			*end = pinned;
+			break;
+		}
+
+		bch2_bkey_buf_reassemble(&tmp, c, k);
+		struct btree_path *path = btree_iter_path(trans, &iter);
+
+		BUG_ON(path->level != 1);
+
+		int ret2 = btree_node_get_and_pin(trans, tmp.k, path->btree_id, path->level - 1);
+
+		if (!ret2)
+			pinned = tmp.k->k.p;
+
+		ret;
+	}));
+	if (ret)
+		return ret;
+
+	return ret;
+}
+
 int bch2_check_extents_to_backpointers(struct bch_fs *c)
 {
+	int ret = 0;
+
+	/*
+	 * Can't allow devices to come/go/resize while we have bucket bitmaps
+	 * allocated
+	 */
+	lockdep_assert_held(&c->state_lock);
+
+	for_each_member_device(c, ca) {
+		BUG_ON(ca->bucket_backpointer_mismatches);
+		ca->bucket_backpointer_mismatches = kvcalloc(BITS_TO_LONGS(ca->mi.nbuckets),
+							     sizeof(unsigned long),
+							     GFP_KERNEL);
+		ca->bucket_backpointer_empty = kvcalloc(BITS_TO_LONGS(ca->mi.nbuckets),
+							sizeof(unsigned long),
+							GFP_KERNEL);
+		if (!ca->bucket_backpointer_mismatches ||
+		    !ca->bucket_backpointer_empty) {
+			bch2_dev_put(ca);
+			ret = -BCH_ERR_ENOMEM_backpointer_mismatches_bitmap;
+			goto err_free_bitmaps;
+		}
+	}
+
 	struct btree_trans *trans = bch2_trans_get(c);
-	struct extents_to_bp_state s = { .bucket_start = POS_MIN };
-	int ret;
+	struct extents_to_bp_state s = { .bp_start = POS_MIN };
 
 	bch2_bkey_buf_init(&s.last_flushed);
 	bkey_init(&s.last_flushed.k->k);
 
+	ret = for_each_btree_key(trans, iter, BTREE_ID_alloc,
+				 POS_MIN, BTREE_ITER_prefetch, k, ({
+		check_bucket_backpointer_mismatch(trans, k, &s.last_flushed);
+	}));
+	if (ret)
+		goto err;
+
+	u64 nr_buckets = 0, nr_mismatches = 0, nr_empty = 0;
+	for_each_member_device(c, ca) {
+		nr_buckets	+= ca->mi.nbuckets;
+		nr_mismatches	+= bitmap_weight(ca->bucket_backpointer_mismatches, ca->mi.nbuckets);
+		nr_empty	+= bitmap_weight(ca->bucket_backpointer_empty, ca->mi.nbuckets);
+	}
+
+	if (!nr_mismatches && !nr_empty)
+		goto err;
+
+	bch_info(c, "scanning for missing backpointers in %llu/%llu buckets",
+		 nr_mismatches + nr_empty, nr_buckets);
+
 	while (1) {
-		struct bbpos end;
-		ret = bch2_get_btree_in_memory_pos(trans,
-				BIT_ULL(BTREE_ID_backpointers),
-				BIT_ULL(BTREE_ID_backpointers),
-				BBPOS(BTREE_ID_backpointers, s.bucket_start), &end);
+		ret = bch2_pin_backpointer_nodes_with_missing(trans, s.bp_start, &s.bp_end);
 		if (ret)
 			break;
 
-		s.bucket_end = end.pos;
-
-		if ( bpos_eq(s.bucket_start, POS_MIN) &&
-		    !bpos_eq(s.bucket_end, SPOS_MAX))
+		if ( bpos_eq(s.bp_start, POS_MIN) &&
+		    !bpos_eq(s.bp_end, SPOS_MAX))
 			bch_verbose(c, "%s(): alloc info does not fit in ram, running in multiple passes with %zu nodes per pass",
 				    __func__, btree_nodes_fit_in_ram(c));
 
-		if (!bpos_eq(s.bucket_start, POS_MIN) ||
-		    !bpos_eq(s.bucket_end, SPOS_MAX)) {
+		if (!bpos_eq(s.bp_start, POS_MIN) ||
+		    !bpos_eq(s.bp_end, SPOS_MAX)) {
 			struct printbuf buf = PRINTBUF;
 
 			prt_str(&buf, "check_extents_to_backpointers(): ");
-			bch2_bpos_to_text(&buf, s.bucket_start);
+			bch2_bpos_to_text(&buf, s.bp_start);
 			prt_str(&buf, "-");
-			bch2_bpos_to_text(&buf, s.bucket_end);
+			bch2_bpos_to_text(&buf, s.bp_end);
 
 			bch_verbose(c, "%s", buf.buf);
 			printbuf_exit(&buf);
 		}
 
 		ret = bch2_check_extents_to_backpointers_pass(trans, &s);
-		if (ret || bpos_eq(s.bucket_end, SPOS_MAX))
+		if (ret || bpos_eq(s.bp_end, SPOS_MAX))
 			break;
 
-		s.bucket_start = bpos_successor(s.bucket_end);
+		s.bp_start = bpos_successor(s.bp_end);
 	}
+err:
 	bch2_trans_put(trans);
 	bch2_bkey_buf_exit(&s.last_flushed, c);
-
 	bch2_btree_cache_unpin(c);
+err_free_bitmaps:
+	for_each_member_device(c, ca) {
+		kvfree(ca->bucket_backpointer_empty);
+		ca->bucket_backpointer_empty = NULL;
+		kvfree(ca->bucket_backpointer_mismatches);
+		ca->bucket_backpointer_mismatches = NULL;
+	}
 
 	bch_err_fn(c, ret);
 	return ret;
@@ -959,44 +1159,43 @@ static int check_one_backpointer(struct btree_trans *trans,
 		return 0;
 
 	struct bkey_s_c_backpointer bp = bkey_s_c_to_backpointer(bp_k);
-	struct bch_fs *c = trans->c;
-	struct btree_iter iter;
 	struct bbpos pos = bp_to_bbpos(*bp.v);
-	struct bkey_s_c k;
-	struct printbuf buf = PRINTBUF;
-	int ret;
 
 	if (bbpos_cmp(pos, start) < 0 ||
 	    bbpos_cmp(pos, end) > 0)
 		return 0;
 
-	k = bch2_backpointer_get_key(trans, &iter, bp.k->p, *bp.v, 0);
-	ret = bkey_err(k);
+	struct btree_iter iter;
+	struct bkey_s_c k = bch2_backpointer_get_key(trans, bp, &iter, 0, last_flushed);
+	int ret = bkey_err(k);
 	if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 		return 0;
 	if (ret)
 		return ret;
 
-	if (!k.k) {
-		ret = bch2_btree_write_buffer_maybe_flush(trans, bp.s_c, last_flushed);
-		if (ret)
-			goto out;
-
-		if (fsck_err(trans, backpointer_to_missing_ptr,
-			     "backpointer for missing %s\n  %s",
-			     bp.v->level ? "btree node" : "extent",
-			     (bch2_bkey_val_to_text(&buf, c, bp.s_c), buf.buf))) {
-			ret = bch2_btree_delete_at_buffered(trans, BTREE_ID_backpointers, bp.k->p);
-			goto out;
-		}
-	}
-out:
-fsck_err:
 	bch2_trans_iter_exit(trans, &iter);
-	printbuf_exit(&buf);
 	return ret;
 }
 
+static int check_bucket_backpointers_to_extents(struct btree_trans *trans,
+						struct bch_dev *ca, struct bpos bucket)
+{
+	u32 restart_count = trans->restart_count;
+	struct bkey_buf last_flushed;
+	bch2_bkey_buf_init(&last_flushed);
+	bkey_init(&last_flushed.k->k);
+
+	int ret = for_each_btree_key_max(trans, iter, BTREE_ID_backpointers,
+				      bucket_pos_to_bp_start(ca, bucket),
+				      bucket_pos_to_bp_end(ca, bucket),
+				      0, k,
+		check_one_backpointer(trans, BBPOS_MIN, BBPOS_MAX, k, &last_flushed)
+	);
+
+	bch2_bkey_buf_exit(&last_flushed, trans->c);
+	return ret ?: trans_was_restarted(trans, restart_count);
+}
+
 static int bch2_check_backpointers_to_extents_pass(struct btree_trans *trans,
 						   struct bbpos start,
 						   struct bbpos end)
@@ -1009,9 +1208,8 @@ static int bch2_check_backpointers_to_extents_pass(struct btree_trans *trans,
 	bkey_init(&last_flushed.k->k);
 	progress_init(&progress, trans->c, BIT_ULL(BTREE_ID_backpointers));
 
-	int ret = for_each_btree_key_commit(trans, iter, BTREE_ID_backpointers,
-				  POS_MIN, BTREE_ITER_prefetch, k,
-				  NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
+	int ret = for_each_btree_key(trans, iter, BTREE_ID_backpointers,
+				     POS_MIN, BTREE_ITER_prefetch, k, ({
 			progress_update_iter(trans, &progress, &iter, "backpointers_to_extents");
 			check_one_backpointer(trans, start, end, k, &last_flushed);
 	}));
diff --git a/fs/bcachefs/backpointers.h b/fs/bcachefs/backpointers.h
index 3b29fdf519dd..060dad1521ee 100644
--- a/fs/bcachefs/backpointers.h
+++ b/fs/bcachefs/backpointers.h
@@ -18,14 +18,14 @@ static inline u64 swab40(u64 x)
 		((x & 0xff00000000ULL) >> 32));
 }
 
-int bch2_backpointer_validate(struct bch_fs *, struct bkey_s_c k, enum bch_validate_flags);
-void bch2_backpointer_to_text(struct printbuf *, const struct bch_backpointer *);
-void bch2_backpointer_k_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
+int bch2_backpointer_validate(struct bch_fs *, struct bkey_s_c k,
+			      struct bkey_validate_context);
+void bch2_backpointer_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 void bch2_backpointer_swab(struct bkey_s);
 
 #define bch2_bkey_ops_backpointer ((struct bkey_ops) {	\
 	.key_validate	= bch2_backpointer_validate,	\
-	.val_to_text	= bch2_backpointer_k_to_text,	\
+	.val_to_text	= bch2_backpointer_to_text,	\
 	.swab		= bch2_backpointer_swab,	\
 	.min_val_size	= 32,				\
 })
@@ -43,22 +43,24 @@ static inline struct bpos bp_pos_to_bucket(const struct bch_dev *ca, struct bpos
 	return POS(bp_pos.inode, sector_to_bucket(ca, bucket_sector));
 }
 
+static inline struct bpos bp_pos_to_bucket_and_offset(const struct bch_dev *ca, struct bpos bp_pos,
+						      u32 *bucket_offset)
+{
+	u64 bucket_sector = bp_pos.offset >> MAX_EXTENT_COMPRESS_RATIO_SHIFT;
+
+	return POS(bp_pos.inode, sector_to_bucket_and_offset(ca, bucket_sector, bucket_offset));
+}
+
 static inline bool bp_pos_to_bucket_nodev_noerror(struct bch_fs *c, struct bpos bp_pos, struct bpos *bucket)
 {
 	rcu_read_lock();
-	struct bch_dev *ca = bch2_dev_rcu(c, bp_pos.inode);
+	struct bch_dev *ca = bch2_dev_rcu_noerror(c, bp_pos.inode);
 	if (ca)
 		*bucket = bp_pos_to_bucket(ca, bp_pos);
 	rcu_read_unlock();
 	return ca != NULL;
 }
 
-static inline bool bp_pos_to_bucket_nodev(struct bch_fs *c, struct bpos bp_pos, struct bpos *bucket)
-{
-	return !bch2_fs_inconsistent_on(!bp_pos_to_bucket_nodev_noerror(c, bp_pos, bucket),
-					c, "backpointer for missing device %llu", bp_pos.inode);
-}
-
 static inline struct bpos bucket_pos_to_bp_noerror(const struct bch_dev *ca,
 						   struct bpos bucket,
 						   u64 bucket_offset)
@@ -80,31 +82,35 @@ static inline struct bpos bucket_pos_to_bp(const struct bch_dev *ca,
 	return ret;
 }
 
-int bch2_bucket_backpointer_mod_nowritebuffer(struct btree_trans *, struct bch_dev *,
-				struct bpos bucket, struct bch_backpointer, struct bkey_s_c, bool);
+static inline struct bpos bucket_pos_to_bp_start(const struct bch_dev *ca, struct bpos bucket)
+{
+	return bucket_pos_to_bp(ca, bucket, 0);
+}
+
+static inline struct bpos bucket_pos_to_bp_end(const struct bch_dev *ca, struct bpos bucket)
+{
+	return bpos_nosnap_predecessor(bucket_pos_to_bp(ca, bpos_nosnap_successor(bucket), 0));
+}
+
+int bch2_bucket_backpointer_mod_nowritebuffer(struct btree_trans *,
+				struct bkey_s_c,
+				struct bkey_i_backpointer *,
+				bool);
 
 static inline int bch2_bucket_backpointer_mod(struct btree_trans *trans,
-				struct bch_dev *ca,
-				struct bpos bucket,
-				struct bch_backpointer bp,
 				struct bkey_s_c orig_k,
+				struct bkey_i_backpointer *bp,
 				bool insert)
 {
 	if (unlikely(bch2_backpointers_no_use_write_buffer))
-		return bch2_bucket_backpointer_mod_nowritebuffer(trans, ca, bucket, bp, orig_k, insert);
-
-	struct bkey_i_backpointer bp_k;
-
-	bkey_backpointer_init(&bp_k.k_i);
-	bp_k.k.p = bucket_pos_to_bp(ca, bucket, bp.bucket_offset);
-	bp_k.v = bp;
+		return bch2_bucket_backpointer_mod_nowritebuffer(trans, orig_k, bp, insert);
 
 	if (!insert) {
-		bp_k.k.type = KEY_TYPE_deleted;
-		set_bkey_val_u64s(&bp_k.k, 0);
+		bp->k.type = KEY_TYPE_deleted;
+		set_bkey_val_u64s(&bp->k, 0);
 	}
 
-	return bch2_trans_update_buffered(trans, BTREE_ID_backpointers, &bp_k.k_i);
+	return bch2_trans_update_buffered(trans, BTREE_ID_backpointers, &bp->k_i);
 }
 
 static inline enum bch_data_type bch2_bkey_ptr_data_type(struct bkey_s_c k,
@@ -134,44 +140,29 @@ static inline enum bch_data_type bch2_bkey_ptr_data_type(struct bkey_s_c k,
 	}
 }
 
-static inline void __bch2_extent_ptr_to_bp(struct bch_fs *c, struct bch_dev *ca,
+static inline void bch2_extent_ptr_to_bp(struct bch_fs *c,
 			   enum btree_id btree_id, unsigned level,
 			   struct bkey_s_c k, struct extent_ptr_decoded p,
 			   const union bch_extent_entry *entry,
-			   struct bpos *bucket_pos, struct bch_backpointer *bp,
-			   u64 sectors)
+			   struct bkey_i_backpointer *bp)
 {
-	u32 bucket_offset;
-	*bucket_pos = PTR_BUCKET_POS_OFFSET(ca, &p.ptr, &bucket_offset);
-	*bp = (struct bch_backpointer) {
+	bkey_backpointer_init(&bp->k_i);
+	bp->k.p = POS(p.ptr.dev, ((u64) p.ptr.offset << MAX_EXTENT_COMPRESS_RATIO_SHIFT) + p.crc.offset);
+	bp->v	= (struct bch_backpointer) {
 		.btree_id	= btree_id,
 		.level		= level,
 		.data_type	= bch2_bkey_ptr_data_type(k, p, entry),
-		.bucket_offset	= ((u64) bucket_offset << MAX_EXTENT_COMPRESS_RATIO_SHIFT) +
-			p.crc.offset,
-		.bucket_len	= sectors,
+		.bucket_gen	= p.ptr.gen,
+		.bucket_len	= ptr_disk_sectors(level ? btree_sectors(c) : k.k->size, p),
 		.pos		= k.k->p,
 	};
 }
 
-static inline void bch2_extent_ptr_to_bp(struct bch_fs *c, struct bch_dev *ca,
-			   enum btree_id btree_id, unsigned level,
-			   struct bkey_s_c k, struct extent_ptr_decoded p,
-			   const union bch_extent_entry *entry,
-			   struct bpos *bucket_pos, struct bch_backpointer *bp)
-{
-	u64 sectors = ptr_disk_sectors(level ? btree_sectors(c) : k.k->size, p);
-
-	__bch2_extent_ptr_to_bp(c, ca, btree_id, level, k, p, entry, bucket_pos, bp, sectors);
-}
-
-int bch2_get_next_backpointer(struct btree_trans *, struct bch_dev *ca, struct bpos, int,
-			      struct bpos *, struct bch_backpointer *, unsigned);
-struct bkey_s_c bch2_backpointer_get_key(struct btree_trans *, struct btree_iter *,
-					 struct bpos, struct bch_backpointer,
-					 unsigned);
-struct btree *bch2_backpointer_get_node(struct btree_trans *, struct btree_iter *,
-					struct bpos, struct bch_backpointer);
+struct bkey_buf;
+struct bkey_s_c bch2_backpointer_get_key(struct btree_trans *, struct bkey_s_c_backpointer,
+					 struct btree_iter *, unsigned, struct bkey_buf *);
+struct btree *bch2_backpointer_get_node(struct btree_trans *, struct bkey_s_c_backpointer,
+					struct btree_iter *, struct bkey_buf *);
 
 int bch2_check_btree_backpointers(struct bch_fs *);
 int bch2_check_extents_to_backpointers(struct bch_fs *);
diff --git a/fs/bcachefs/bbpos.h b/fs/bcachefs/bbpos.h
index be2edced5213..63abe17f35ea 100644
--- a/fs/bcachefs/bbpos.h
+++ b/fs/bcachefs/bbpos.h
@@ -29,7 +29,7 @@ static inline struct bbpos bbpos_successor(struct bbpos pos)
 
 static inline void bch2_bbpos_to_text(struct printbuf *out, struct bbpos pos)
 {
-	prt_str(out, bch2_btree_id_str(pos.btree));
+	bch2_btree_id_to_text(out, pos.btree);
 	prt_char(out, ':');
 	bch2_bpos_to_text(out, pos.pos);
 }
diff --git a/fs/bcachefs/bcachefs.h b/fs/bcachefs/bcachefs.h
index e94a83b8113e..161cf2f05d2a 100644
--- a/fs/bcachefs/bcachefs.h
+++ b/fs/bcachefs/bcachefs.h
@@ -205,6 +205,7 @@
 #include <linux/zstd.h>
 
 #include "bcachefs_format.h"
+#include "btree_journal_iter_types.h"
 #include "disk_accounting_types.h"
 #include "errcode.h"
 #include "fifo.h"
@@ -293,6 +294,8 @@ do {									\
 
 #define bch_info(c, fmt, ...) \
 	bch2_print(c, KERN_INFO bch2_fmt(c, fmt), ##__VA_ARGS__)
+#define bch_info_ratelimited(c, fmt, ...) \
+	bch2_print_ratelimited(c, KERN_INFO bch2_fmt(c, fmt), ##__VA_ARGS__)
 #define bch_notice(c, fmt, ...) \
 	bch2_print(c, KERN_NOTICE bch2_fmt(c, fmt), ##__VA_ARGS__)
 #define bch_warn(c, fmt, ...) \
@@ -352,6 +355,12 @@ do {									\
 		bch_info(c, fmt, ##__VA_ARGS__);			\
 } while (0)
 
+#define bch_verbose_ratelimited(c, fmt, ...)				\
+do {									\
+	if ((c)->opts.verbose)						\
+		bch_info_ratelimited(c, fmt, ##__VA_ARGS__);		\
+} while (0)
+
 #define pr_verbose_init(opts, fmt, ...)					\
 do {									\
 	if (opt_get(opts, verbose))					\
@@ -538,20 +547,20 @@ struct bch_dev {
 
 	/*
 	 * Buckets:
-	 * Per-bucket arrays are protected by c->mark_lock, bucket_lock and
-	 * gc_gens_lock, for device resize - holding any is sufficient for
-	 * access: Or rcu_read_lock(), but only for dev_ptr_stale():
+	 * Per-bucket arrays are protected by either rcu_read_lock or
+	 * state_lock, for device resize.
 	 */
 	GENRADIX(struct bucket)	buckets_gc;
 	struct bucket_gens __rcu *bucket_gens;
 	u8			*oldest_gen;
 	unsigned long		*buckets_nouse;
-	struct rw_semaphore	bucket_lock;
+
+	unsigned long		*bucket_backpointer_mismatches;
+	unsigned long		*bucket_backpointer_empty;
 
 	struct bch_dev_usage __percpu	*usage;
 
 	/* Allocator: */
-	u64			new_fs_bucket_idx;
 	u64			alloc_cursor[3];
 
 	unsigned		nr_open_buckets;
@@ -606,6 +615,7 @@ struct bch_dev {
 	x(going_ro)			\
 	x(write_disable_complete)	\
 	x(clean_shutdown)		\
+	x(recovery_running)		\
 	x(fsck_running)			\
 	x(initial_gc_unfixed)		\
 	x(need_delete_dead_snapshots)	\
@@ -650,28 +660,6 @@ struct journal_seq_blacklist_table {
 	}			entries[];
 };
 
-struct journal_keys {
-	/* must match layout in darray_types.h */
-	size_t			nr, size;
-	struct journal_key {
-		u64		journal_seq;
-		u32		journal_offset;
-		enum btree_id	btree_id:8;
-		unsigned	level:8;
-		bool		allocated;
-		bool		overwritten;
-		struct bkey_i	*k;
-	}			*data;
-	/*
-	 * Gap buffer: instead of all the empty space in the array being at the
-	 * end of the buffer - from @nr to @size - the empty space is at @gap.
-	 * This means that sequential insertions are O(n) instead of O(n^2).
-	 */
-	size_t			gap;
-	atomic_t		ref;
-	bool			initial_ref_held;
-};
-
 struct btree_trans_buf {
 	struct btree_trans	*trans;
 };
@@ -680,6 +668,7 @@ struct btree_trans_buf {
 	((subvol_inum) { BCACHEFS_ROOT_SUBVOL,	BCACHEFS_ROOT_INO })
 
 #define BCH_WRITE_REFS()						\
+	x(journal)							\
 	x(trans)							\
 	x(write)							\
 	x(promote)							\
@@ -692,6 +681,7 @@ struct btree_trans_buf {
 	x(dio_write)							\
 	x(discard)							\
 	x(discard_fast)							\
+	x(check_discard_freespace_key)					\
 	x(invalidate)							\
 	x(delete_dead_snapshots)					\
 	x(gc_gens)							\
@@ -735,6 +725,12 @@ struct bch_fs {
 	struct percpu_ref	writes;
 #endif
 	/*
+	 * Certain operations are only allowed in single threaded mode, during
+	 * recovery, and we want to assert that this is the case:
+	 */
+	struct task_struct	*recovery_task;
+
+	/*
 	 * Analagous to c->writes, for asynchronous ops that don't necessarily
 	 * need fs to be read-write
 	 */
@@ -764,6 +760,8 @@ struct bch_fs {
 		__uuid_t	user_uuid;
 
 		u16		version;
+		u16		version_incompat;
+		u16		version_incompat_allowed;
 		u16		version_min;
 		u16		version_upgrade_complete;
 
@@ -834,9 +832,10 @@ struct bch_fs {
 	struct work_struct	btree_interior_update_work;
 
 	struct workqueue_struct	*btree_node_rewrite_worker;
-
-	struct list_head	pending_node_rewrites;
-	struct mutex		pending_node_rewrites_lock;
+	struct list_head	btree_node_rewrites;
+	struct list_head	btree_node_rewrites_pending;
+	spinlock_t		btree_node_rewrites_lock;
+	struct closure_waitlist	btree_node_rewrites_wait;
 
 	/* btree_io.c: */
 	spinlock_t		btree_write_error_lock;
@@ -967,8 +966,7 @@ struct bch_fs {
 	struct rhashtable	promote_table;
 
 	mempool_t		compression_bounce[2];
-	mempool_t		compress_workspace[BCH_COMPRESSION_TYPE_NR];
-	mempool_t		decompress_workspace;
+	mempool_t		compress_workspace[BCH_COMPRESSION_OPT_NR];
 	size_t			zstd_workspace_size;
 
 	struct crypto_shash	*sha256;
@@ -1027,6 +1025,7 @@ struct bch_fs {
 	struct list_head	vfs_inodes_list;
 	struct mutex		vfs_inodes_lock;
 	struct rhashtable	vfs_inodes_table;
+	struct rhltable		vfs_inodes_by_inum_table;
 
 	/* VFS IO PATH - fs-io.c */
 	struct bio_set		writepage_bioset;
@@ -1048,10 +1047,12 @@ struct bch_fs {
 	 * for signaling to the toplevel code which pass we want to run now.
 	 */
 	enum bch_recovery_pass	curr_recovery_pass;
+	enum bch_recovery_pass	next_recovery_pass;
 	/* bitmask of recovery passes that we actually ran */
 	u64			recovery_passes_complete;
 	/* never rewinds version of curr_recovery_pass */
 	enum bch_recovery_pass	recovery_pass_done;
+	spinlock_t		recovery_pass_lock;
 	struct semaphore	online_fsck_mutex;
 
 	/* DEBUG JUNK */
@@ -1062,9 +1063,6 @@ struct bch_fs {
 	struct btree_node	*verify_ondisk;
 	struct mutex		verify_lock;
 
-	u64			*unused_inode_hints;
-	unsigned		inode_shard_bits;
-
 	/*
 	 * A btree node on disk could have too many bsets for an iterator to fit
 	 * on the stack - have to dynamically allocate them
@@ -1086,8 +1084,6 @@ struct bch_fs {
 	u64			counters_on_mount[BCH_COUNTER_NR];
 	u64 __percpu		*counters;
 
-	unsigned		copy_gc_enabled:1;
-
 	struct bch2_time_stats	times[BCH_TIME_STAT_NR];
 
 	struct btree_transaction_stats btree_transaction_stats[BCH_TRANSACTIONS_NR];
diff --git a/fs/bcachefs/bcachefs_format.h b/fs/bcachefs/bcachefs_format.h
index 5004f6ba997c..f70f0108401f 100644
--- a/fs/bcachefs/bcachefs_format.h
+++ b/fs/bcachefs/bcachefs_format.h
@@ -418,7 +418,8 @@ static inline void bkey_init(struct bkey *k)
 	x(snapshot_tree,	31)			\
 	x(logged_op_truncate,	32)			\
 	x(logged_op_finsert,	33)			\
-	x(accounting,		34)
+	x(accounting,		34)			\
+	x(inode_alloc_cursor,	35)
 
 enum bch_bkey_type {
 #define x(name, nr) KEY_TYPE_##name	= nr,
@@ -463,7 +464,8 @@ struct bch_backpointer {
 	__u8			btree_id;
 	__u8			level;
 	__u8			data_type;
-	__u64			bucket_offset:40;
+	__u8			bucket_gen;
+	__u32			pad;
 	__u32			bucket_len;
 	struct bpos		pos;
 } __packed __aligned(8);
@@ -499,8 +501,6 @@ struct bch_sb_field {
 #include "disk_groups_format.h"
 #include "extents_format.h"
 #include "ec_format.h"
-#include "dirent_format.h"
-#include "disk_groups_format.h"
 #include "inode_format.h"
 #include "journal_seq_blacklist_format.h"
 #include "logged_ops_format.h"
@@ -679,7 +679,14 @@ struct bch_sb_field_ext {
 	x(disk_accounting_v3,		BCH_VERSION(1, 10))		\
 	x(disk_accounting_inum,		BCH_VERSION(1, 11))		\
 	x(rebalance_work_acct_fix,	BCH_VERSION(1, 12))		\
-	x(inode_has_child_snapshots,	BCH_VERSION(1, 13))
+	x(inode_has_child_snapshots,	BCH_VERSION(1, 13))		\
+	x(backpointer_bucket_gen,	BCH_VERSION(1, 14))		\
+	x(disk_accounting_big_endian,	BCH_VERSION(1, 15))		\
+	x(reflink_p_may_update_opts,	BCH_VERSION(1, 16))		\
+	x(inode_depth,			BCH_VERSION(1, 17))		\
+	x(persistent_inode_cursors,	BCH_VERSION(1, 18))		\
+	x(autofix_errors,		BCH_VERSION(1, 19))		\
+	x(directory_size,		BCH_VERSION(1, 20))
 
 enum bcachefs_metadata_version {
 	bcachefs_metadata_version_min = 9,
@@ -844,6 +851,10 @@ LE64_BITMASK(BCH_SB_VERSION_UPGRADE_COMPLETE,
 					struct bch_sb, flags[5],  0, 16);
 LE64_BITMASK(BCH_SB_ALLOCATOR_STUCK_TIMEOUT,
 					struct bch_sb, flags[5], 16, 32);
+LE64_BITMASK(BCH_SB_VERSION_INCOMPAT,	struct bch_sb, flags[5], 32, 48);
+LE64_BITMASK(BCH_SB_VERSION_INCOMPAT_ALLOWED,
+					struct bch_sb, flags[5], 48, 64);
+LE64_BITMASK(BCH_SB_SHARD_INUMS_NBITS,	struct bch_sb, flags[6],  0,  4);
 
 static inline __u64 BCH_SB_COMPRESSION_TYPE(const struct bch_sb *sb)
 {
@@ -896,21 +907,22 @@ static inline void SET_BCH_SB_BACKGROUND_COMPRESSION_TYPE(struct bch_sb *sb, __u
 	x(new_varint,			15)	\
 	x(journal_no_flush,		16)	\
 	x(alloc_v2,			17)	\
-	x(extents_across_btree_nodes,	18)
+	x(extents_across_btree_nodes,	18)	\
+	x(incompat_version_field,	19)
 
 #define BCH_SB_FEATURES_ALWAYS				\
-	((1ULL << BCH_FEATURE_new_extent_overwrite)|	\
-	 (1ULL << BCH_FEATURE_extents_above_btree_updates)|\
-	 (1ULL << BCH_FEATURE_btree_updates_journalled)|\
-	 (1ULL << BCH_FEATURE_alloc_v2)|\
-	 (1ULL << BCH_FEATURE_extents_across_btree_nodes))
+	(BIT_ULL(BCH_FEATURE_new_extent_overwrite)|	\
+	 BIT_ULL(BCH_FEATURE_extents_above_btree_updates)|\
+	 BIT_ULL(BCH_FEATURE_btree_updates_journalled)|\
+	 BIT_ULL(BCH_FEATURE_alloc_v2)|\
+	 BIT_ULL(BCH_FEATURE_extents_across_btree_nodes))
 
 #define BCH_SB_FEATURES_ALL				\
 	(BCH_SB_FEATURES_ALWAYS|			\
-	 (1ULL << BCH_FEATURE_new_siphash)|		\
-	 (1ULL << BCH_FEATURE_btree_ptr_v2)|		\
-	 (1ULL << BCH_FEATURE_new_varint)|		\
-	 (1ULL << BCH_FEATURE_journal_no_flush))
+	 BIT_ULL(BCH_FEATURE_new_siphash)|		\
+	 BIT_ULL(BCH_FEATURE_btree_ptr_v2)|		\
+	 BIT_ULL(BCH_FEATURE_new_varint)|		\
+	 BIT_ULL(BCH_FEATURE_journal_no_flush))
 
 enum bch_sb_feature {
 #define x(f, n) BCH_FEATURE_##f,
@@ -1032,7 +1044,7 @@ static inline _Bool bch2_csum_type_is_encryption(enum bch_csum_type type)
 	x(crc64,		2)	\
 	x(xxhash,		3)
 
-enum bch_csum_opts {
+enum bch_csum_opt {
 #define x(t, n) BCH_CSUM_OPT_##t = n,
 	BCH_CSUM_OPTS()
 #undef x
@@ -1221,6 +1233,15 @@ struct jset_entry_log {
 	u8			d[];
 } __packed __aligned(8);
 
+static inline unsigned jset_entry_log_msg_bytes(struct jset_entry_log *l)
+{
+	unsigned b = vstruct_bytes(&l->entry) - offsetof(struct jset_entry_log, d);
+
+	while (b && !l->d[b - 1])
+		--b;
+	return b;
+}
+
 struct jset_entry_datetime {
 	struct jset_entry	entry;
 	__le64			seconds;
@@ -1268,14 +1289,18 @@ LE32_BITMASK(JSET_NO_FLUSH,	struct jset, flags, 5, 6);
 /* Btree: */
 
 enum btree_id_flags {
-	BTREE_ID_EXTENTS	= BIT(0),
-	BTREE_ID_SNAPSHOTS	= BIT(1),
-	BTREE_ID_SNAPSHOT_FIELD	= BIT(2),
-	BTREE_ID_DATA		= BIT(3),
+	BTREE_IS_extents	= BIT(0),
+	BTREE_IS_snapshots	= BIT(1),
+	BTREE_IS_snapshot_field	= BIT(2),
+	BTREE_IS_data		= BIT(3),
+	BTREE_IS_write_buffer	= BIT(4),
 };
 
 #define BCH_BTREE_IDS()								\
-	x(extents,		0,	BTREE_ID_EXTENTS|BTREE_ID_SNAPSHOTS|BTREE_ID_DATA,\
+	x(extents,		0,						\
+	  BTREE_IS_extents|							\
+	  BTREE_IS_snapshots|							\
+	  BTREE_IS_data,							\
 	  BIT_ULL(KEY_TYPE_whiteout)|						\
 	  BIT_ULL(KEY_TYPE_error)|						\
 	  BIT_ULL(KEY_TYPE_cookie)|						\
@@ -1283,17 +1308,20 @@ enum btree_id_flags {
 	  BIT_ULL(KEY_TYPE_reservation)|					\
 	  BIT_ULL(KEY_TYPE_reflink_p)|						\
 	  BIT_ULL(KEY_TYPE_inline_data))					\
-	x(inodes,		1,	BTREE_ID_SNAPSHOTS,			\
+	x(inodes,		1,						\
+	  BTREE_IS_snapshots,							\
 	  BIT_ULL(KEY_TYPE_whiteout)|						\
 	  BIT_ULL(KEY_TYPE_inode)|						\
 	  BIT_ULL(KEY_TYPE_inode_v2)|						\
 	  BIT_ULL(KEY_TYPE_inode_v3)|						\
 	  BIT_ULL(KEY_TYPE_inode_generation))					\
-	x(dirents,		2,	BTREE_ID_SNAPSHOTS,			\
+	x(dirents,		2,						\
+	  BTREE_IS_snapshots,							\
 	  BIT_ULL(KEY_TYPE_whiteout)|						\
 	  BIT_ULL(KEY_TYPE_hash_whiteout)|					\
 	  BIT_ULL(KEY_TYPE_dirent))						\
-	x(xattrs,		3,	BTREE_ID_SNAPSHOTS,			\
+	x(xattrs,		3,						\
+	  BTREE_IS_snapshots,							\
 	  BIT_ULL(KEY_TYPE_whiteout)|						\
 	  BIT_ULL(KEY_TYPE_cookie)|						\
 	  BIT_ULL(KEY_TYPE_hash_whiteout)|					\
@@ -1307,7 +1335,9 @@ enum btree_id_flags {
 	  BIT_ULL(KEY_TYPE_quota))						\
 	x(stripes,		6,	0,					\
 	  BIT_ULL(KEY_TYPE_stripe))						\
-	x(reflink,		7,	BTREE_ID_EXTENTS|BTREE_ID_DATA,		\
+	x(reflink,		7,						\
+	  BTREE_IS_extents|							\
+	  BTREE_IS_data,							\
 	  BIT_ULL(KEY_TYPE_reflink_v)|						\
 	  BIT_ULL(KEY_TYPE_indirect_inline_data)|				\
 	  BIT_ULL(KEY_TYPE_error))						\
@@ -1315,28 +1345,38 @@ enum btree_id_flags {
 	  BIT_ULL(KEY_TYPE_subvolume))						\
 	x(snapshots,		9,	0,					\
 	  BIT_ULL(KEY_TYPE_snapshot))						\
-	x(lru,			10,	0,					\
+	x(lru,			10,						\
+	  BTREE_IS_write_buffer,						\
 	  BIT_ULL(KEY_TYPE_set))						\
-	x(freespace,		11,	BTREE_ID_EXTENTS,			\
+	x(freespace,		11,						\
+	  BTREE_IS_extents,							\
 	  BIT_ULL(KEY_TYPE_set))						\
 	x(need_discard,		12,	0,					\
 	  BIT_ULL(KEY_TYPE_set))						\
-	x(backpointers,		13,	0,					\
+	x(backpointers,		13,						\
+	  BTREE_IS_write_buffer,						\
 	  BIT_ULL(KEY_TYPE_backpointer))					\
 	x(bucket_gens,		14,	0,					\
 	  BIT_ULL(KEY_TYPE_bucket_gens))					\
 	x(snapshot_trees,	15,	0,					\
 	  BIT_ULL(KEY_TYPE_snapshot_tree))					\
-	x(deleted_inodes,	16,	BTREE_ID_SNAPSHOT_FIELD,		\
+	x(deleted_inodes,	16,						\
+	  BTREE_IS_snapshot_field|						\
+	  BTREE_IS_write_buffer,						\
 	  BIT_ULL(KEY_TYPE_set))						\
 	x(logged_ops,		17,	0,					\
 	  BIT_ULL(KEY_TYPE_logged_op_truncate)|					\
-	  BIT_ULL(KEY_TYPE_logged_op_finsert))					\
-	x(rebalance_work,	18,	BTREE_ID_SNAPSHOT_FIELD,		\
+	  BIT_ULL(KEY_TYPE_logged_op_finsert)|					\
+	  BIT_ULL(KEY_TYPE_inode_alloc_cursor))					\
+	x(rebalance_work,	18,						\
+	  BTREE_IS_snapshot_field|						\
+	  BTREE_IS_write_buffer,						\
 	  BIT_ULL(KEY_TYPE_set)|BIT_ULL(KEY_TYPE_cookie))			\
 	x(subvolume_children,	19,	0,					\
 	  BIT_ULL(KEY_TYPE_set))						\
-	x(accounting,		20,	BTREE_ID_SNAPSHOT_FIELD,		\
+	x(accounting,		20,						\
+	  BTREE_IS_snapshot_field|						\
+	  BTREE_IS_write_buffer,						\
 	  BIT_ULL(KEY_TYPE_accounting))						\
 
 enum btree_id {
@@ -1361,6 +1401,8 @@ static inline bool btree_id_is_alloc(enum btree_id id)
 	case BTREE_ID_need_discard:
 	case BTREE_ID_freespace:
 	case BTREE_ID_bucket_gens:
+	case BTREE_ID_lru:
+	case BTREE_ID_accounting:
 		return true;
 	default:
 		return false;
diff --git a/fs/bcachefs/bkey.h b/fs/bcachefs/bkey.h
index 41df24a53d97..054e2d5e8448 100644
--- a/fs/bcachefs/bkey.h
+++ b/fs/bcachefs/bkey.h
@@ -9,13 +9,6 @@
 #include "util.h"
 #include "vstructs.h"
 
-enum bch_validate_flags {
-	BCH_VALIDATE_write		= BIT(0),
-	BCH_VALIDATE_commit		= BIT(1),
-	BCH_VALIDATE_journal		= BIT(2),
-	BCH_VALIDATE_silent		= BIT(3),
-};
-
 #if 0
 
 /*
diff --git a/fs/bcachefs/bkey_methods.c b/fs/bcachefs/bkey_methods.c
index e7ac227ba7e8..15c93576b5c2 100644
--- a/fs/bcachefs/bkey_methods.c
+++ b/fs/bcachefs/bkey_methods.c
@@ -28,7 +28,7 @@ const char * const bch2_bkey_types[] = {
 };
 
 static int deleted_key_validate(struct bch_fs *c, struct bkey_s_c k,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	return 0;
 }
@@ -42,7 +42,7 @@ static int deleted_key_validate(struct bch_fs *c, struct bkey_s_c k,
 })
 
 static int empty_val_key_validate(struct bch_fs *c, struct bkey_s_c k,
-				 enum bch_validate_flags flags)
+				  struct bkey_validate_context from)
 {
 	int ret = 0;
 
@@ -59,7 +59,7 @@ fsck_err:
 })
 
 static int key_type_cookie_validate(struct bch_fs *c, struct bkey_s_c k,
-				    enum bch_validate_flags flags)
+				    struct bkey_validate_context from)
 {
 	return 0;
 }
@@ -83,7 +83,7 @@ static void key_type_cookie_to_text(struct printbuf *out, struct bch_fs *c,
 })
 
 static int key_type_inline_data_validate(struct bch_fs *c, struct bkey_s_c k,
-					 enum bch_validate_flags flags)
+					 struct bkey_validate_context from)
 {
 	return 0;
 }
@@ -124,7 +124,7 @@ const struct bkey_ops bch2_bkey_null_ops = {
 };
 
 int bch2_bkey_val_validate(struct bch_fs *c, struct bkey_s_c k,
-			   enum bch_validate_flags flags)
+			   struct bkey_validate_context from)
 {
 	if (test_bit(BCH_FS_no_invalid_checks, &c->flags))
 		return 0;
@@ -140,7 +140,7 @@ int bch2_bkey_val_validate(struct bch_fs *c, struct bkey_s_c k,
 	if (!ops->key_validate)
 		return 0;
 
-	ret = ops->key_validate(c, k, flags);
+	ret = ops->key_validate(c, k, from);
 fsck_err:
 	return ret;
 }
@@ -161,9 +161,10 @@ const char *bch2_btree_node_type_str(enum btree_node_type type)
 }
 
 int __bch2_bkey_validate(struct bch_fs *c, struct bkey_s_c k,
-			 enum btree_node_type type,
-			 enum bch_validate_flags flags)
+			 struct bkey_validate_context from)
 {
+	enum btree_node_type type = __btree_node_type(from.level, from.btree);
+
 	if (test_bit(BCH_FS_no_invalid_checks, &c->flags))
 		return 0;
 
@@ -177,7 +178,7 @@ int __bch2_bkey_validate(struct bch_fs *c, struct bkey_s_c k,
 		return 0;
 
 	bkey_fsck_err_on(k.k->type < KEY_TYPE_MAX &&
-			 (type == BKEY_TYPE_btree || (flags & BCH_VALIDATE_commit)) &&
+			 (type == BKEY_TYPE_btree || (from.flags & BCH_VALIDATE_commit)) &&
 			 !(bch2_key_types_allowed[type] & BIT_ULL(k.k->type)),
 			 c, bkey_invalid_type_for_btree,
 			 "invalid key type for btree %s (%s)",
@@ -228,15 +229,15 @@ fsck_err:
 }
 
 int bch2_bkey_validate(struct bch_fs *c, struct bkey_s_c k,
-		      enum btree_node_type type,
-		      enum bch_validate_flags flags)
+		       struct bkey_validate_context from)
 {
-	return __bch2_bkey_validate(c, k, type, flags) ?:
-		bch2_bkey_val_validate(c, k, flags);
+	return __bch2_bkey_validate(c, k, from) ?:
+		bch2_bkey_val_validate(c, k, from);
 }
 
 int bch2_bkey_in_btree_node(struct bch_fs *c, struct btree *b,
-			    struct bkey_s_c k, enum bch_validate_flags flags)
+			    struct bkey_s_c k,
+			    struct bkey_validate_context from)
 {
 	int ret = 0;
 
diff --git a/fs/bcachefs/bkey_methods.h b/fs/bcachefs/bkey_methods.h
index 018fb72e32d3..bf34111cdf00 100644
--- a/fs/bcachefs/bkey_methods.h
+++ b/fs/bcachefs/bkey_methods.h
@@ -22,7 +22,7 @@ extern const struct bkey_ops bch2_bkey_null_ops;
  */
 struct bkey_ops {
 	int		(*key_validate)(struct bch_fs *c, struct bkey_s_c k,
-					enum bch_validate_flags flags);
+					struct bkey_validate_context from);
 	void		(*val_to_text)(struct printbuf *, struct bch_fs *,
 				       struct bkey_s_c);
 	void		(*swab)(struct bkey_s);
@@ -48,13 +48,14 @@ static inline const struct bkey_ops *bch2_bkey_type_ops(enum bch_bkey_type type)
 		: &bch2_bkey_null_ops;
 }
 
-int bch2_bkey_val_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
-int __bch2_bkey_validate(struct bch_fs *, struct bkey_s_c, enum btree_node_type,
-			 enum bch_validate_flags);
-int bch2_bkey_validate(struct bch_fs *, struct bkey_s_c, enum btree_node_type,
-		       enum bch_validate_flags);
+int bch2_bkey_val_validate(struct bch_fs *, struct bkey_s_c,
+			   struct bkey_validate_context);
+int __bch2_bkey_validate(struct bch_fs *, struct bkey_s_c,
+			 struct bkey_validate_context);
+int bch2_bkey_validate(struct bch_fs *, struct bkey_s_c,
+		       struct bkey_validate_context);
 int bch2_bkey_in_btree_node(struct bch_fs *, struct btree *, struct bkey_s_c,
-			    enum bch_validate_flags);
+			    struct bkey_validate_context from);
 
 void bch2_bpos_to_text(struct printbuf *, struct bpos);
 void bch2_bkey_to_text(struct printbuf *, const struct bkey *);
diff --git a/fs/bcachefs/bkey_types.h b/fs/bcachefs/bkey_types.h
index c9ae9e42b385..b4f328f9853c 100644
--- a/fs/bcachefs/bkey_types.h
+++ b/fs/bcachefs/bkey_types.h
@@ -210,4 +210,32 @@ static inline struct bkey_i_##name *bkey_##name##_init(struct bkey_i *_k)\
 BCH_BKEY_TYPES();
 #undef x
 
+enum bch_validate_flags {
+	BCH_VALIDATE_write		= BIT(0),
+	BCH_VALIDATE_commit		= BIT(1),
+	BCH_VALIDATE_silent		= BIT(2),
+};
+
+#define BKEY_VALIDATE_CONTEXTS()	\
+	x(unknown)			\
+	x(superblock)			\
+	x(journal)			\
+	x(btree_root)			\
+	x(btree_node)			\
+	x(commit)
+
+struct bkey_validate_context {
+	enum {
+#define x(n)	BKEY_VALIDATE_##n,
+	BKEY_VALIDATE_CONTEXTS()
+#undef x
+	}			from:8;
+	enum bch_validate_flags	flags:8;
+	u8			level;
+	enum btree_id		btree;
+	bool			root:1;
+	unsigned		journal_offset;
+	u64			journal_seq;
+};
+
 #endif /* _BCACHEFS_BKEY_TYPES_H */
diff --git a/fs/bcachefs/btree_cache.c b/fs/bcachefs/btree_cache.c
index 7123019ab3bc..672ca2c1d37d 100644
--- a/fs/bcachefs/btree_cache.c
+++ b/fs/bcachefs/btree_cache.c
@@ -222,7 +222,6 @@ void bch2_node_pin(struct bch_fs *c, struct btree *b)
 	struct btree_cache *bc = &c->btree_cache;
 
 	mutex_lock(&bc->lock);
-	BUG_ON(!__btree_node_pinned(bc, b));
 	if (b != btree_node_root(c, b) && !btree_node_pinned(b)) {
 		set_btree_node_pinned(b);
 		list_move(&b->list, &bc->live[1].list);
@@ -326,7 +325,7 @@ void bch2_btree_node_update_key_early(struct btree_trans *trans,
 	if (!IS_ERR_OR_NULL(b)) {
 		mutex_lock(&c->btree_cache.lock);
 
-		bch2_btree_node_hash_remove(&c->btree_cache, b);
+		__bch2_btree_node_hash_remove(&c->btree_cache, b);
 
 		bkey_copy(&b->key, new);
 		ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
@@ -1004,16 +1003,14 @@ static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
 		return;
 
 	prt_printf(&buf,
-	       "btree node header doesn't match ptr\n"
-	       "btree %s level %u\n"
-	       "ptr: ",
-	       bch2_btree_id_str(b->c.btree_id), b->c.level);
+		   "btree node header doesn't match ptr: ");
+	bch2_btree_id_level_to_text(&buf, b->c.btree_id, b->c.level);
+	prt_str(&buf, "\nptr: ");
 	bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 
-	prt_printf(&buf, "\nheader: btree %s level %llu\n"
-	       "min ",
-	       bch2_btree_id_str(BTREE_NODE_ID(b->data)),
-	       BTREE_NODE_LEVEL(b->data));
+	prt_str(&buf, "\nheader: ");
+	bch2_btree_id_level_to_text(&buf, BTREE_NODE_ID(b->data), BTREE_NODE_LEVEL(b->data));
+	prt_str(&buf, "\nmin ");
 	bch2_bpos_to_text(&buf, b->data->min_key);
 
 	prt_printf(&buf, "\nmax ");
@@ -1133,7 +1130,7 @@ retry:
 
 	if (unlikely(btree_node_read_error(b))) {
 		six_unlock_type(&b->c.lock, lock_type);
-		return ERR_PTR(-BCH_ERR_btree_node_read_error);
+		return ERR_PTR(-BCH_ERR_btree_node_read_err_cached);
 	}
 
 	EBUG_ON(b->c.btree_id != path->btree_id);
@@ -1223,7 +1220,7 @@ struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *
 
 	if (unlikely(btree_node_read_error(b))) {
 		six_unlock_type(&b->c.lock, lock_type);
-		return ERR_PTR(-BCH_ERR_btree_node_read_error);
+		return ERR_PTR(-BCH_ERR_btree_node_read_err_cached);
 	}
 
 	EBUG_ON(b->c.btree_id != path->btree_id);
@@ -1305,7 +1302,7 @@ lock_node:
 
 	if (unlikely(btree_node_read_error(b))) {
 		six_unlock_read(&b->c.lock);
-		b = ERR_PTR(-BCH_ERR_btree_node_read_error);
+		b = ERR_PTR(-BCH_ERR_btree_node_read_err_cached);
 		goto out;
 	}
 
@@ -1398,13 +1395,31 @@ void bch2_btree_id_to_text(struct printbuf *out, enum btree_id btree)
 		prt_printf(out, "(unknown btree %u)", btree);
 }
 
+void bch2_btree_id_level_to_text(struct printbuf *out, enum btree_id btree, unsigned level)
+{
+	prt_str(out, "btree=");
+	bch2_btree_id_to_text(out, btree);
+	prt_printf(out, " level=%u", level);
+}
+
+void __bch2_btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
+			      enum btree_id btree, unsigned level, struct bkey_s_c k)
+{
+	bch2_btree_id_to_text(out, btree);
+	prt_printf(out, " level %u/", level);
+	struct btree_root *r = bch2_btree_id_root(c, btree);
+	if (r)
+		prt_printf(out, "%u", r->level);
+	else
+		prt_printf(out, "(unknown)");
+	prt_printf(out, "\n  ");
+
+	bch2_bkey_val_to_text(out, c, k);
+}
+
 void bch2_btree_pos_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
 {
-	prt_printf(out, "%s level %u/%u\n  ",
-	       bch2_btree_id_str(b->c.btree_id),
-	       b->c.level,
-	       bch2_btree_id_root(c, b->c.btree_id)->level);
-	bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
+	__bch2_btree_pos_to_text(out, c, b->c.btree_id, b->c.level, bkey_i_to_s_c(&b->key));
 }
 
 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
@@ -1478,8 +1493,12 @@ void bch2_btree_cache_to_text(struct printbuf *out, const struct btree_cache *bc
 	prt_printf(out, "cannibalize lock:\t%p\n",	bc->alloc_lock);
 	prt_newline(out);
 
-	for (unsigned i = 0; i < ARRAY_SIZE(bc->nr_by_btree); i++)
-		prt_btree_cache_line(out, c, bch2_btree_id_str(i), bc->nr_by_btree[i]);
+	for (unsigned i = 0; i < ARRAY_SIZE(bc->nr_by_btree); i++) {
+		bch2_btree_id_to_text(out, i);
+		prt_printf(out, "\t");
+		prt_human_readable_u64(out, bc->nr_by_btree[i] * c->opts.btree_node_size);
+		prt_printf(out, " (%zu)\n", bc->nr_by_btree[i]);
+	}
 
 	prt_newline(out);
 	prt_printf(out, "freed:\t%zu\n", bc->nr_freed);
diff --git a/fs/bcachefs/btree_cache.h b/fs/bcachefs/btree_cache.h
index 66e86d1a178d..ca3c1b145330 100644
--- a/fs/bcachefs/btree_cache.h
+++ b/fs/bcachefs/btree_cache.h
@@ -128,19 +128,27 @@ static inline struct btree_root *bch2_btree_id_root(struct bch_fs *c, unsigned i
 	} else {
 		unsigned idx = id - BTREE_ID_NR;
 
-		EBUG_ON(idx >= c->btree_roots_extra.nr);
+		/* This can happen when we're called from btree_node_scan */
+		if (idx >= c->btree_roots_extra.nr)
+			return NULL;
+
 		return &c->btree_roots_extra.data[idx];
 	}
 }
 
 static inline struct btree *btree_node_root(struct bch_fs *c, struct btree *b)
 {
-	return bch2_btree_id_root(c, b->c.btree_id)->b;
+	struct btree_root *r = bch2_btree_id_root(c, b->c.btree_id);
+
+	return r ? r->b : NULL;
 }
 
-const char *bch2_btree_id_str(enum btree_id);
+const char *bch2_btree_id_str(enum btree_id);	/* avoid */
 void bch2_btree_id_to_text(struct printbuf *, enum btree_id);
+void bch2_btree_id_level_to_text(struct printbuf *, enum btree_id, unsigned);
 
+void __bch2_btree_pos_to_text(struct printbuf *, struct bch_fs *,
+			      enum btree_id, unsigned, struct bkey_s_c);
 void bch2_btree_pos_to_text(struct printbuf *, struct bch_fs *, const struct btree *);
 void bch2_btree_node_to_text(struct printbuf *, struct bch_fs *, const struct btree *);
 void bch2_btree_cache_to_text(struct printbuf *, const struct btree_cache *);
diff --git a/fs/bcachefs/btree_gc.c b/fs/bcachefs/btree_gc.c
index 81dcf9e512c0..dd1d9b74076e 100644
--- a/fs/bcachefs/btree_gc.c
+++ b/fs/bcachefs/btree_gc.c
@@ -29,6 +29,7 @@
 #include "move.h"
 #include "recovery_passes.h"
 #include "reflink.h"
+#include "recovery.h"
 #include "replicas.h"
 #include "super-io.h"
 #include "trace.h"
@@ -56,8 +57,8 @@ void bch2_gc_pos_to_text(struct printbuf *out, struct gc_pos *p)
 {
 	prt_str(out, bch2_gc_phase_strs[p->phase]);
 	prt_char(out, ' ');
-	bch2_btree_id_to_text(out, p->btree);
-	prt_printf(out, " l=%u ", p->level);
+	bch2_btree_id_level_to_text(out, p->btree, p->level);
+	prt_char(out, ' ');
 	bch2_bpos_to_text(out, p->pos);
 }
 
@@ -209,8 +210,9 @@ static int btree_check_node_boundaries(struct btree_trans *trans, struct btree *
 	if (bpos_eq(expected_start, cur->data->min_key))
 		return 0;
 
-	prt_printf(&buf, "  at btree %s level %u:\n  parent: ",
-		   bch2_btree_id_str(b->c.btree_id), b->c.level);
+	prt_printf(&buf, "  at ");
+	bch2_btree_id_level_to_text(&buf, b->c.btree_id, b->c.level);
+	prt_printf(&buf, ":\n  parent: ");
 	bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 
 	if (prev) {
@@ -277,8 +279,9 @@ static int btree_repair_node_end(struct btree_trans *trans, struct btree *b,
 	if (bpos_eq(child->key.k.p, b->key.k.p))
 		return 0;
 
-	prt_printf(&buf, "at btree %s level %u:\n  parent: ",
-		   bch2_btree_id_str(b->c.btree_id), b->c.level);
+	prt_printf(&buf, "  at ");
+	bch2_btree_id_level_to_text(&buf, b->c.btree_id, b->c.level);
+	prt_printf(&buf, ":\n  parent: ");
 	bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 
 	prt_str(&buf, "\n  child: ");
@@ -341,14 +344,14 @@ again:
 		ret = PTR_ERR_OR_ZERO(cur);
 
 		printbuf_reset(&buf);
+		bch2_btree_id_level_to_text(&buf, b->c.btree_id, b->c.level - 1);
+		prt_char(&buf, ' ');
 		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(cur_k.k));
 
 		if (mustfix_fsck_err_on(bch2_err_matches(ret, EIO),
-				trans, btree_node_unreadable,
-				"Topology repair: unreadable btree node at btree %s level %u:\n"
+				trans, btree_node_read_error,
+				"Topology repair: unreadable btree node at\n"
 				"  %s",
-				bch2_btree_id_str(b->c.btree_id),
-				b->c.level - 1,
 				buf.buf)) {
 			bch2_btree_node_evict(trans, cur_k.k);
 			cur = NULL;
@@ -357,11 +360,9 @@ again:
 			if (ret)
 				break;
 
-			if (!btree_id_is_alloc(b->c.btree_id)) {
-				ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_scan_for_btree_nodes);
-				if (ret)
-					break;
-			}
+			ret = bch2_btree_lost_data(c, b->c.btree_id);
+			if (ret)
+				break;
 			continue;
 		}
 
@@ -370,7 +371,7 @@ again:
 			break;
 
 		if (bch2_btree_node_is_stale(c, cur)) {
-			bch_info(c, "btree node %s older than nodes found by scanning", buf.buf);
+			bch_info(c, "btree node older than nodes found by scanning\n  %s", buf.buf);
 			six_unlock_read(&cur->c.lock);
 			bch2_btree_node_evict(trans, cur_k.k);
 			ret = bch2_journal_key_delete(c, b->c.btree_id,
@@ -478,14 +479,13 @@ again:
 	}
 
 	printbuf_reset(&buf);
+	bch2_btree_id_level_to_text(&buf, b->c.btree_id, b->c.level);
+	prt_newline(&buf);
 	bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 
 	if (mustfix_fsck_err_on(!have_child,
 			trans, btree_node_topology_interior_node_empty,
-			"empty interior btree node at btree %s level %u\n"
-			"  %s",
-			bch2_btree_id_str(b->c.btree_id),
-			b->c.level, buf.buf))
+			"empty interior btree node at %s", buf.buf))
 		ret = DROP_THIS_NODE;
 err:
 fsck_err:
@@ -511,6 +511,7 @@ int bch2_check_topology(struct bch_fs *c)
 {
 	struct btree_trans *trans = bch2_trans_get(c);
 	struct bpos pulled_from_scan = POS_MIN;
+	struct printbuf buf = PRINTBUF;
 	int ret = 0;
 
 	bch2_trans_srcu_unlock(trans);
@@ -519,19 +520,22 @@ int bch2_check_topology(struct bch_fs *c)
 		struct btree_root *r = bch2_btree_id_root(c, i);
 		bool reconstructed_root = false;
 
+		printbuf_reset(&buf);
+		bch2_btree_id_to_text(&buf, i);
+
 		if (r->error) {
-			ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_scan_for_btree_nodes);
+			ret = bch2_btree_lost_data(c, i);
 			if (ret)
 				break;
 reconstruct_root:
-			bch_info(c, "btree root %s unreadable, must recover from scan", bch2_btree_id_str(i));
+			bch_info(c, "btree root %s unreadable, must recover from scan", buf.buf);
 
 			r->alive = false;
 			r->error = 0;
 
 			if (!bch2_btree_has_scanned_nodes(c, i)) {
 				mustfix_fsck_err(trans, btree_root_unreadable_and_scan_found_nothing,
-						 "no nodes found for btree %s, continue?", bch2_btree_id_str(i));
+						 "no nodes found for btree %s, continue?", buf.buf);
 				bch2_btree_root_alloc_fake_trans(trans, i, 0);
 			} else {
 				bch2_btree_root_alloc_fake_trans(trans, i, 1);
@@ -560,13 +564,14 @@ reconstruct_root:
 			if (!reconstructed_root)
 				goto reconstruct_root;
 
-			bch_err(c, "empty btree root %s", bch2_btree_id_str(i));
+			bch_err(c, "empty btree root %s", buf.buf);
 			bch2_btree_root_alloc_fake_trans(trans, i, 0);
 			r->alive = false;
 			ret = 0;
 		}
 	}
 fsck_err:
+	printbuf_exit(&buf);
 	bch2_trans_put(trans);
 	return ret;
 }
@@ -713,6 +718,7 @@ static int bch2_gc_btrees(struct bch_fs *c)
 {
 	struct btree_trans *trans = bch2_trans_get(c);
 	enum btree_id ids[BTREE_ID_NR];
+	struct printbuf buf = PRINTBUF;
 	unsigned i;
 	int ret = 0;
 
@@ -727,14 +733,9 @@ static int bch2_gc_btrees(struct bch_fs *c)
 			continue;
 
 		ret = bch2_gc_btree(trans, btree, true);
-
-		if (mustfix_fsck_err_on(bch2_err_matches(ret, EIO),
-					trans, btree_node_read_error,
-			       "btree node read error for %s",
-			       bch2_btree_id_str(btree)))
-			ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
 	}
-fsck_err:
+
+	printbuf_exit(&buf);
 	bch2_trans_put(trans);
 	bch_err_fn(c, ret);
 	return ret;
@@ -802,7 +803,6 @@ static int bch2_alloc_write_key(struct btree_trans *trans,
 	old = bch2_alloc_to_v4(k, &old_convert);
 	gc = new = *old;
 
-	percpu_down_read(&c->mark_lock);
 	__bucket_m_to_alloc(&gc, *gc_bucket(ca, iter->pos.offset));
 
 	old_gc = gc;
@@ -813,7 +813,6 @@ static int bch2_alloc_write_key(struct btree_trans *trans,
 		gc.data_type = old->data_type;
 		gc.dirty_sectors = old->dirty_sectors;
 	}
-	percpu_up_read(&c->mark_lock);
 
 	/*
 	 * gc.data_type doesn't yet include need_discard & need_gc_gen states -
@@ -831,11 +830,9 @@ static int bch2_alloc_write_key(struct btree_trans *trans,
 		 * safe w.r.t. transaction restarts, so fixup the gc_bucket so
 		 * we don't run it twice:
 		 */
-		percpu_down_read(&c->mark_lock);
 		struct bucket *gc_m = gc_bucket(ca, iter->pos.offset);
 		gc_m->data_type = gc.data_type;
 		gc_m->dirty_sectors = gc.dirty_sectors;
-		percpu_up_read(&c->mark_lock);
 	}
 
 	if (fsck_err_on(new.data_type != gc.data_type,
@@ -895,11 +892,11 @@ static int bch2_gc_alloc_done(struct bch_fs *c)
 
 	for_each_member_device(c, ca) {
 		ret = bch2_trans_run(c,
-			for_each_btree_key_upto_commit(trans, iter, BTREE_ID_alloc,
+			for_each_btree_key_max_commit(trans, iter, BTREE_ID_alloc,
 					POS(ca->dev_idx, ca->mi.first_bucket),
 					POS(ca->dev_idx, ca->mi.nbuckets - 1),
 					BTREE_ITER_slots|BTREE_ITER_prefetch, k,
-					NULL, NULL, BCH_TRANS_COMMIT_lazy_rw,
+					NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 				bch2_alloc_write_key(trans, &iter, ca, k)));
 		if (ret) {
 			bch2_dev_put(ca);
@@ -928,98 +925,6 @@ static int bch2_gc_alloc_start(struct bch_fs *c)
 	return ret;
 }
 
-static int bch2_gc_write_reflink_key(struct btree_trans *trans,
-				     struct btree_iter *iter,
-				     struct bkey_s_c k,
-				     size_t *idx)
-{
-	struct bch_fs *c = trans->c;
-	const __le64 *refcount = bkey_refcount_c(k);
-	struct printbuf buf = PRINTBUF;
-	struct reflink_gc *r;
-	int ret = 0;
-
-	if (!refcount)
-		return 0;
-
-	while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
-	       r->offset < k.k->p.offset)
-		++*idx;
-
-	if (!r ||
-	    r->offset != k.k->p.offset ||
-	    r->size != k.k->size) {
-		bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
-		return -EINVAL;
-	}
-
-	if (fsck_err_on(r->refcount != le64_to_cpu(*refcount),
-			trans, reflink_v_refcount_wrong,
-			"reflink key has wrong refcount:\n"
-			"  %s\n"
-			"  should be %u",
-			(bch2_bkey_val_to_text(&buf, c, k), buf.buf),
-			r->refcount)) {
-		struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
-		ret = PTR_ERR_OR_ZERO(new);
-		if (ret)
-			goto out;
-
-		if (!r->refcount)
-			new->k.type = KEY_TYPE_deleted;
-		else
-			*bkey_refcount(bkey_i_to_s(new)) = cpu_to_le64(r->refcount);
-		ret = bch2_trans_update(trans, iter, new, 0);
-	}
-out:
-fsck_err:
-	printbuf_exit(&buf);
-	return ret;
-}
-
-static int bch2_gc_reflink_done(struct bch_fs *c)
-{
-	size_t idx = 0;
-
-	int ret = bch2_trans_run(c,
-		for_each_btree_key_commit(trans, iter,
-				BTREE_ID_reflink, POS_MIN,
-				BTREE_ITER_prefetch, k,
-				NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
-			bch2_gc_write_reflink_key(trans, &iter, k, &idx)));
-	c->reflink_gc_nr = 0;
-	return ret;
-}
-
-static int bch2_gc_reflink_start(struct bch_fs *c)
-{
-	c->reflink_gc_nr = 0;
-
-	int ret = bch2_trans_run(c,
-		for_each_btree_key(trans, iter, BTREE_ID_reflink, POS_MIN,
-				   BTREE_ITER_prefetch, k, ({
-			const __le64 *refcount = bkey_refcount_c(k);
-
-			if (!refcount)
-				continue;
-
-			struct reflink_gc *r = genradix_ptr_alloc(&c->reflink_gc_table,
-							c->reflink_gc_nr++, GFP_KERNEL);
-			if (!r) {
-				ret = -BCH_ERR_ENOMEM_gc_reflink_start;
-				break;
-			}
-
-			r->offset	= k.k->p.offset;
-			r->size		= k.k->size;
-			r->refcount	= 0;
-			0;
-		})));
-
-	bch_err_fn(c, ret);
-	return ret;
-}
-
 static int bch2_gc_write_stripes_key(struct btree_trans *trans,
 				     struct btree_iter *iter,
 				     struct bkey_s_c k)
@@ -1171,7 +1076,6 @@ static int gc_btree_gens_key(struct btree_trans *trans,
 	if (unlikely(test_bit(BCH_FS_going_ro, &c->flags)))
 		return -EROFS;
 
-	percpu_down_read(&c->mark_lock);
 	rcu_read_lock();
 	bkey_for_each_ptr(ptrs, ptr) {
 		struct bch_dev *ca = bch2_dev_rcu(c, ptr->dev);
@@ -1180,7 +1084,6 @@ static int gc_btree_gens_key(struct btree_trans *trans,
 
 		if (dev_ptr_stale(ca, ptr) > 16) {
 			rcu_read_unlock();
-			percpu_up_read(&c->mark_lock);
 			goto update;
 		}
 	}
@@ -1195,7 +1098,6 @@ static int gc_btree_gens_key(struct btree_trans *trans,
 			*gen = ptr->gen;
 	}
 	rcu_read_unlock();
-	percpu_up_read(&c->mark_lock);
 	return 0;
 update:
 	u = bch2_bkey_make_mut(trans, iter, &k, 0);
@@ -1224,7 +1126,6 @@ static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct bch_dev
 		return ret;
 
 	a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
-	alloc_data_type_set(&a_mut->v, a_mut->v.data_type);
 
 	return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
 }
@@ -1337,9 +1238,16 @@ void bch2_gc_gens_async(struct bch_fs *c)
 		bch2_write_ref_put(c, BCH_WRITE_REF_gc_gens);
 }
 
-void bch2_fs_gc_init(struct bch_fs *c)
+void bch2_fs_btree_gc_exit(struct bch_fs *c)
 {
-	seqcount_init(&c->gc_pos_lock);
+}
 
+int bch2_fs_btree_gc_init(struct bch_fs *c)
+{
+	seqcount_init(&c->gc_pos_lock);
 	INIT_WORK(&c->gc_gens_work, bch2_gc_gens_work);
+
+	init_rwsem(&c->gc_lock);
+	mutex_init(&c->gc_gens_lock);
+	return 0;
 }
diff --git a/fs/bcachefs/btree_gc.h b/fs/bcachefs/btree_gc.h
index 8a47e8bd0791..9693a90a48a2 100644
--- a/fs/bcachefs/btree_gc.h
+++ b/fs/bcachefs/btree_gc.h
@@ -82,6 +82,8 @@ void bch2_gc_pos_to_text(struct printbuf *, struct gc_pos *);
 
 int bch2_gc_gens(struct bch_fs *);
 void bch2_gc_gens_async(struct bch_fs *);
-void bch2_fs_gc_init(struct bch_fs *);
+
+void bch2_fs_btree_gc_exit(struct bch_fs *);
+int bch2_fs_btree_gc_init(struct bch_fs *);
 
 #endif /* _BCACHEFS_BTREE_GC_H */
diff --git a/fs/bcachefs/btree_io.c b/fs/bcachefs/btree_io.c
index 839d68802e42..e371e60e3133 100644
--- a/fs/bcachefs/btree_io.c
+++ b/fs/bcachefs/btree_io.c
@@ -25,9 +25,8 @@
 
 static void bch2_btree_node_header_to_text(struct printbuf *out, struct btree_node *bn)
 {
-	prt_printf(out, "btree=%s l=%u seq %llux\n",
-		   bch2_btree_id_str(BTREE_NODE_ID(bn)),
-		   (unsigned) BTREE_NODE_LEVEL(bn), bn->keys.seq);
+	bch2_btree_id_level_to_text(out, BTREE_NODE_ID(bn), BTREE_NODE_LEVEL(bn));
+	prt_printf(out, " seq %llx %llu\n", bn->keys.seq, BTREE_NODE_SEQ(bn));
 	prt_str(out, "min: ");
 	bch2_bpos_to_text(out, bn->min_key);
 	prt_newline(out);
@@ -490,8 +489,8 @@ void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 	if (b->nsets == MAX_BSETS &&
 	    !btree_node_write_in_flight(b) &&
 	    should_compact_all(c, b)) {
-		bch2_btree_node_write(c, b, SIX_LOCK_write,
-				      BTREE_WRITE_init_next_bset);
+		bch2_btree_node_write_trans(trans, b, SIX_LOCK_write,
+					    BTREE_WRITE_init_next_bset);
 		reinit_iter = true;
 	}
 
@@ -832,13 +831,32 @@ fsck_err:
 	return ret;
 }
 
+static int btree_node_bkey_val_validate(struct bch_fs *c, struct btree *b,
+					struct bkey_s_c k,
+					enum bch_validate_flags flags)
+{
+	return bch2_bkey_val_validate(c, k, (struct bkey_validate_context) {
+		.from	= BKEY_VALIDATE_btree_node,
+		.level	= b->c.level,
+		.btree	= b->c.btree_id,
+		.flags	= flags
+	});
+}
+
 static int bset_key_validate(struct bch_fs *c, struct btree *b,
 			     struct bkey_s_c k,
-			     bool updated_range, int rw)
+			     bool updated_range,
+			     enum bch_validate_flags flags)
 {
-	return __bch2_bkey_validate(c, k, btree_node_type(b), 0) ?:
-		(!updated_range ? bch2_bkey_in_btree_node(c, b, k, 0) : 0) ?:
-		(rw == WRITE ? bch2_bkey_val_validate(c, k, 0) : 0);
+	struct bkey_validate_context from = (struct bkey_validate_context) {
+		.from	= BKEY_VALIDATE_btree_node,
+		.level	= b->c.level,
+		.btree	= b->c.btree_id,
+		.flags	= flags,
+	};
+	return __bch2_bkey_validate(c, k, from) ?:
+		(!updated_range ? bch2_bkey_in_btree_node(c, b, k, from) : 0) ?:
+		(flags & BCH_VALIDATE_write ? btree_node_bkey_val_validate(c, b, k, flags) : 0);
 }
 
 static bool bkey_packed_valid(struct bch_fs *c, struct btree *b,
@@ -855,7 +873,21 @@ static bool bkey_packed_valid(struct bch_fs *c, struct btree *b,
 
 	struct bkey tmp;
 	struct bkey_s u = __bkey_disassemble(b, k, &tmp);
-	return !__bch2_bkey_validate(c, u.s_c, btree_node_type(b), BCH_VALIDATE_silent);
+	return !__bch2_bkey_validate(c, u.s_c,
+				     (struct bkey_validate_context) {
+					.from	= BKEY_VALIDATE_btree_node,
+					.level	= b->c.level,
+					.btree	= b->c.btree_id,
+					.flags	= BCH_VALIDATE_silent
+				     });
+}
+
+static inline int btree_node_read_bkey_cmp(const struct btree *b,
+				const struct bkey_packed *l,
+				const struct bkey_packed *r)
+{
+	return bch2_bkey_cmp_packed(b, l, r)
+		?: (int) bkey_deleted(r) - (int) bkey_deleted(l);
 }
 
 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
@@ -918,7 +950,7 @@ static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 				    BSET_BIG_ENDIAN(i), write,
 				    &b->format, k);
 
-		if (prev && bkey_iter_cmp(b, prev, k) > 0) {
+		if (prev && btree_node_read_bkey_cmp(b, prev, k) >= 0) {
 			struct bkey up = bkey_unpack_key(b, prev);
 
 			printbuf_reset(&buf);
@@ -965,6 +997,7 @@ drop_this_key:
 got_good_key:
 		le16_add_cpu(&i->u64s, -next_good_key);
 		memmove_u64s_down(k, bkey_p_next(k), (u64 *) vstruct_end(i) - (u64 *) k);
+		set_btree_node_need_rewrite(b);
 	}
 fsck_err:
 	printbuf_exit(&buf);
@@ -1038,39 +1071,51 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 
 	while (b->written < (ptr_written ?: btree_sectors(c))) {
 		unsigned sectors;
-		struct nonce nonce;
 		bool first = !b->written;
-		bool csum_bad;
 
-		if (!b->written) {
+		if (first) {
+			bne = NULL;
 			i = &b->data->keys;
+		} else {
+			bne = write_block(b);
+			i = &bne->keys;
 
-			btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
-				     -BCH_ERR_btree_node_read_err_want_retry,
-				     c, ca, b, i, NULL,
-				     bset_unknown_csum,
-				     "unknown checksum type %llu", BSET_CSUM_TYPE(i));
-
-			nonce = btree_nonce(i, b->written << 9);
+			if (i->seq != b->data->keys.seq)
+				break;
+		}
 
-			struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
-			csum_bad = bch2_crc_cmp(b->data->csum, csum);
-			if (csum_bad)
-				bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
+		struct nonce nonce = btree_nonce(i, b->written << 9);
+		bool good_csum_type = bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i));
 
-			btree_err_on(csum_bad,
-				     -BCH_ERR_btree_node_read_err_want_retry,
-				     c, ca, b, i, NULL,
-				     bset_bad_csum,
-				     "%s",
-				     (printbuf_reset(&buf),
-				      bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), b->data->csum, csum),
-				      buf.buf));
-
-			ret = bset_encrypt(c, i, b->written << 9);
-			if (bch2_fs_fatal_err_on(ret, c,
-					"decrypting btree node: %s", bch2_err_str(ret)))
-				goto fsck_err;
+		btree_err_on(!good_csum_type,
+			     bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i))
+			     ? -BCH_ERR_btree_node_read_err_must_retry
+			     : -BCH_ERR_btree_node_read_err_want_retry,
+			     c, ca, b, i, NULL,
+			     bset_unknown_csum,
+			     "unknown checksum type %llu", BSET_CSUM_TYPE(i));
+
+		if (first) {
+			if (good_csum_type) {
+				struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
+				bool csum_bad = bch2_crc_cmp(b->data->csum, csum);
+				if (csum_bad)
+					bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
+
+				btree_err_on(csum_bad,
+					     -BCH_ERR_btree_node_read_err_want_retry,
+					     c, ca, b, i, NULL,
+					     bset_bad_csum,
+					     "%s",
+					     (printbuf_reset(&buf),
+					      bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), b->data->csum, csum),
+					      buf.buf));
+
+				ret = bset_encrypt(c, i, b->written << 9);
+				if (bch2_fs_fatal_err_on(ret, c,
+							 "decrypting btree node: %s", bch2_err_str(ret)))
+					goto fsck_err;
+			}
 
 			btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
 				     !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
@@ -1081,37 +1126,26 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 
 			sectors = vstruct_sectors(b->data, c->block_bits);
 		} else {
-			bne = write_block(b);
-			i = &bne->keys;
-
-			if (i->seq != b->data->keys.seq)
-				break;
-
-			btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
-				     -BCH_ERR_btree_node_read_err_want_retry,
-				     c, ca, b, i, NULL,
-				     bset_unknown_csum,
-				     "unknown checksum type %llu", BSET_CSUM_TYPE(i));
-
-			nonce = btree_nonce(i, b->written << 9);
-			struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
-			csum_bad = bch2_crc_cmp(bne->csum, csum);
-			if (ca && csum_bad)
-				bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
-
-			btree_err_on(csum_bad,
-				     -BCH_ERR_btree_node_read_err_want_retry,
-				     c, ca, b, i, NULL,
-				     bset_bad_csum,
-				     "%s",
-				     (printbuf_reset(&buf),
-				      bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), bne->csum, csum),
-				      buf.buf));
-
-			ret = bset_encrypt(c, i, b->written << 9);
-			if (bch2_fs_fatal_err_on(ret, c,
-					"decrypting btree node: %s", bch2_err_str(ret)))
-				goto fsck_err;
+			if (good_csum_type) {
+				struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
+				bool csum_bad = bch2_crc_cmp(bne->csum, csum);
+				if (ca && csum_bad)
+					bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
+
+				btree_err_on(csum_bad,
+					     -BCH_ERR_btree_node_read_err_want_retry,
+					     c, ca, b, i, NULL,
+					     bset_bad_csum,
+					     "%s",
+					     (printbuf_reset(&buf),
+					      bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), bne->csum, csum),
+					      buf.buf));
+
+				ret = bset_encrypt(c, i, b->written << 9);
+				if (bch2_fs_fatal_err_on(ret, c,
+						"decrypting btree node: %s", bch2_err_str(ret)))
+					goto fsck_err;
+			}
 
 			sectors = vstruct_sectors(bne, c->block_bits);
 		}
@@ -1216,7 +1250,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 		struct bkey tmp;
 		struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 
-		ret = bch2_bkey_val_validate(c, u.s_c, READ);
+		ret = btree_node_bkey_val_validate(c, b, u.s_c, READ);
 		if (ret == -BCH_ERR_fsck_delete_bkey ||
 		    (bch2_inject_invalid_keys &&
 		     !bversion_cmp(u.k->bversion, MAX_VERSION))) {
@@ -1226,6 +1260,7 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 			memmove_u64s_down(k, bkey_p_next(k),
 					  (u64 *) vstruct_end(i) - (u64 *) k);
 			set_btree_bset_end(b, b->set);
+			set_btree_node_need_rewrite(b);
 			continue;
 		}
 		if (ret)
@@ -1339,13 +1374,18 @@ start:
 			       rb->start_time);
 	bio_put(&rb->bio);
 
-	if (saw_error &&
+	if ((saw_error ||
+	     btree_node_need_rewrite(b)) &&
 	    !btree_node_read_error(b) &&
 	    c->curr_recovery_pass != BCH_RECOVERY_PASS_scan_for_btree_nodes) {
-		printbuf_reset(&buf);
-		bch2_bpos_to_text(&buf, b->key.k.p);
-		bch_err_ratelimited(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
-			 __func__, bch2_btree_id_str(b->c.btree_id), b->c.level, buf.buf);
+		if (saw_error) {
+			printbuf_reset(&buf);
+			bch2_btree_id_level_to_text(&buf, b->c.btree_id, b->c.level);
+			prt_str(&buf, " ");
+			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
+			bch_err_ratelimited(c, "%s: rewriting btree node at due to error\n  %s",
+					    __func__, buf.buf);
+		}
 
 		bch2_btree_node_rewrite_async(c, b);
 	}
@@ -1933,7 +1973,12 @@ static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
 	bool saw_error;
 
 	int ret = bch2_bkey_validate(c, bkey_i_to_s_c(&b->key),
-				     BKEY_TYPE_btree, WRITE);
+				     (struct bkey_validate_context) {
+					.from	= BKEY_VALIDATE_btree_node,
+					.level	= b->c.level + 1,
+					.btree	= b->c.btree_id,
+					.flags	= BCH_VALIDATE_write,
+				     });
 	if (ret) {
 		bch2_fs_inconsistent(c, "invalid btree node key before write");
 		return ret;
@@ -2300,6 +2345,34 @@ void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
 	}
 }
 
+void bch2_btree_node_write_trans(struct btree_trans *trans, struct btree *b,
+				 enum six_lock_type lock_type_held,
+				 unsigned flags)
+{
+	struct bch_fs *c = trans->c;
+
+	if (lock_type_held == SIX_LOCK_intent ||
+	    (lock_type_held == SIX_LOCK_read &&
+	     six_lock_tryupgrade(&b->c.lock))) {
+		__bch2_btree_node_write(c, b, flags);
+
+		/* don't cycle lock unnecessarily: */
+		if (btree_node_just_written(b) &&
+		    six_trylock_write(&b->c.lock)) {
+			bch2_btree_post_write_cleanup(c, b);
+			__bch2_btree_node_unlock_write(trans, b);
+		}
+
+		if (lock_type_held == SIX_LOCK_read)
+			six_lock_downgrade(&b->c.lock);
+	} else {
+		__bch2_btree_node_write(c, b, flags);
+		if (lock_type_held == SIX_LOCK_write &&
+		    btree_node_just_written(b))
+			bch2_btree_post_write_cleanup(c, b);
+	}
+}
+
 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
 {
 	struct bucket_table *tbl;
diff --git a/fs/bcachefs/btree_io.h b/fs/bcachefs/btree_io.h
index 9b01ca3de907..6f9e4a6dacf7 100644
--- a/fs/bcachefs/btree_io.h
+++ b/fs/bcachefs/btree_io.h
@@ -144,11 +144,13 @@ enum btree_write_flags {
 void __bch2_btree_node_write(struct bch_fs *, struct btree *, unsigned);
 void bch2_btree_node_write(struct bch_fs *, struct btree *,
 			   enum six_lock_type, unsigned);
+void bch2_btree_node_write_trans(struct btree_trans *, struct btree *,
+				 enum six_lock_type, unsigned);
 
-static inline void btree_node_write_if_need(struct bch_fs *c, struct btree *b,
+static inline void btree_node_write_if_need(struct btree_trans *trans, struct btree *b,
 					    enum six_lock_type lock_held)
 {
-	bch2_btree_node_write(c, b, lock_held, BTREE_WRITE_ONLY_IF_NEED);
+	bch2_btree_node_write_trans(trans, b, lock_held, BTREE_WRITE_ONLY_IF_NEED);
 }
 
 bool bch2_btree_flush_all_reads(struct bch_fs *);
diff --git a/fs/bcachefs/btree_iter.c b/fs/bcachefs/btree_iter.c
index eef9b89c561d..367231ab1980 100644
--- a/fs/bcachefs/btree_iter.c
+++ b/fs/bcachefs/btree_iter.c
@@ -270,8 +270,10 @@ static void bch2_btree_iter_verify_entry_exit(struct btree_iter *iter)
 	BUG_ON(!(iter->flags & BTREE_ITER_all_snapshots) &&
 	       iter->pos.snapshot != iter->snapshot);
 
-	BUG_ON(bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
-	       bkey_gt(iter->pos, iter->k.p));
+	BUG_ON(iter->flags & BTREE_ITER_all_snapshots	? !bpos_eq(iter->pos, iter->k.p) :
+	       !(iter->flags & BTREE_ITER_is_extents)	? !bkey_eq(iter->pos, iter->k.p) :
+	       (bkey_lt(iter->pos, bkey_start_pos(&iter->k)) ||
+		bkey_gt(iter->pos, iter->k.p)));
 }
 
 static int bch2_btree_iter_verify_ret(struct btree_iter *iter, struct bkey_s_c k)
@@ -327,7 +329,7 @@ out:
 void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 			    struct bpos pos)
 {
-	bch2_trans_verify_not_unlocked(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 
 	struct btree_path *path;
 	struct trans_for_each_path_inorder_iter iter;
@@ -697,6 +699,19 @@ void bch2_trans_node_add(struct btree_trans *trans,
 	bch2_trans_revalidate_updates_in_node(trans, b);
 }
 
+void bch2_trans_node_drop(struct btree_trans *trans,
+			  struct btree *b)
+{
+	struct btree_path *path;
+	unsigned i, level = b->c.level;
+
+	trans_for_each_path(trans, path, i)
+		if (path->l[level].b == b) {
+			btree_node_unlock(trans, path, level);
+			path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
+		}
+}
+
 /*
  * A btree node has been modified in such a way as to invalidate iterators - fix
  * them:
@@ -720,7 +735,7 @@ static inline int btree_path_lock_root(struct btree_trans *trans,
 				       unsigned long trace_ip)
 {
 	struct bch_fs *c = trans->c;
-	struct btree *b, **rootp = &bch2_btree_id_root(c, path->btree_id)->b;
+	struct btree_root *r = bch2_btree_id_root(c, path->btree_id);
 	enum six_lock_type lock_type;
 	unsigned i;
 	int ret;
@@ -728,7 +743,12 @@ static inline int btree_path_lock_root(struct btree_trans *trans,
 	EBUG_ON(path->nodes_locked);
 
 	while (1) {
-		b = READ_ONCE(*rootp);
+		struct btree *b = READ_ONCE(r->b);
+		if (unlikely(!b)) {
+			BUG_ON(!r->error);
+			return r->error;
+		}
+
 		path->level = READ_ONCE(b->c.level);
 
 		if (unlikely(path->level < depth_want)) {
@@ -748,14 +768,12 @@ static inline int btree_path_lock_root(struct btree_trans *trans,
 		ret = btree_node_lock(trans, path, &b->c,
 				      path->level, lock_type, trace_ip);
 		if (unlikely(ret)) {
-			if (bch2_err_matches(ret, BCH_ERR_lock_fail_root_changed))
-				continue;
 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 				return ret;
 			BUG();
 		}
 
-		if (likely(b == READ_ONCE(*rootp) &&
+		if (likely(b == READ_ONCE(r->b) &&
 			   b->c.level == path->level &&
 			   !race_fault())) {
 			for (i = 0; i < path->level; i++)
@@ -825,6 +843,8 @@ static int btree_path_prefetch_j(struct btree_trans *trans, struct btree_path *p
 
 	bch2_bkey_buf_init(&tmp);
 
+	jiter->fail_if_too_many_whiteouts = true;
+
 	while (nr-- && !ret) {
 		if (!bch2_btree_node_relock(trans, path, path->level))
 			break;
@@ -1000,7 +1020,7 @@ retry_all:
 
 	bch2_trans_unlock(trans);
 	cond_resched();
-	trans_set_locked(trans);
+	trans_set_locked(trans, false);
 
 	if (unlikely(trans->memory_allocation_failure)) {
 		struct closure cl;
@@ -1267,7 +1287,7 @@ __bch2_btree_path_set_pos(struct btree_trans *trans,
 {
 	int cmp = bpos_cmp(new_pos, trans->paths[path_idx].pos);
 
-	bch2_trans_verify_not_in_restart(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 	EBUG_ON(!trans->paths[path_idx].ref);
 
 	trace_btree_path_set_pos(trans, trans->paths + path_idx, &new_pos);
@@ -1427,17 +1447,31 @@ void __noreturn bch2_trans_restart_error(struct btree_trans *trans, u32 restart_
 	      (void *) trans->last_begin_ip);
 }
 
-void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
+static void __noreturn bch2_trans_in_restart_error(struct btree_trans *trans)
 {
+#ifdef CONFIG_BCACHEFS_DEBUG
+	struct printbuf buf = PRINTBUF;
+	bch2_prt_backtrace(&buf, &trans->last_restarted_trace);
+	panic("in transaction restart: %s, last restarted by\n%s",
+	      bch2_err_str(trans->restarted),
+	      buf.buf);
+#else
 	panic("in transaction restart: %s, last restarted by %pS\n",
 	      bch2_err_str(trans->restarted),
 	      (void *) trans->last_restarted_ip);
+#endif
 }
 
-void __noreturn bch2_trans_unlocked_error(struct btree_trans *trans)
+void __noreturn bch2_trans_unlocked_or_in_restart_error(struct btree_trans *trans)
 {
-	panic("trans should be locked, unlocked by %pS\n",
-	      (void *) trans->last_unlock_ip);
+	if (trans->restarted)
+		bch2_trans_in_restart_error(trans);
+
+	if (!trans->locked)
+		panic("trans should be locked, unlocked by %pS\n",
+		      (void *) trans->last_unlock_ip);
+
+	BUG();
 }
 
 noinline __cold
@@ -1450,10 +1484,11 @@ void bch2_trans_updates_to_text(struct printbuf *buf, struct btree_trans *trans)
 	trans_for_each_update(trans, i) {
 		struct bkey_s_c old = { &i->old_k, i->old_v };
 
-		prt_printf(buf, "update: btree=%s cached=%u %pS\n",
-		       bch2_btree_id_str(i->btree_id),
-		       i->cached,
-		       (void *) i->ip_allocated);
+		prt_str(buf, "update: btree=");
+		bch2_btree_id_to_text(buf, i->btree_id);
+		prt_printf(buf, " cached=%u %pS\n",
+			   i->cached,
+			   (void *) i->ip_allocated);
 
 		prt_printf(buf, "  old ");
 		bch2_bkey_val_to_text(buf, trans->c, old);
@@ -1486,13 +1521,13 @@ static void bch2_btree_path_to_text_short(struct printbuf *out, struct btree_tra
 {
 	struct btree_path *path = trans->paths + path_idx;
 
-	prt_printf(out, "path: idx %3u ref %u:%u %c %c %c btree=%s l=%u pos ",
+	prt_printf(out, "path: idx %3u ref %u:%u %c %c %c ",
 		   path_idx, path->ref, path->intent_ref,
 		   path->preserve ? 'P' : ' ',
 		   path->should_be_locked ? 'S' : ' ',
-		   path->cached ? 'C' : 'B',
-		   bch2_btree_id_str(path->btree_id),
-		   path->level);
+		   path->cached ? 'C' : 'B');
+	bch2_btree_id_level_to_text(out, path->btree_id, path->level);
+	prt_str(out, " pos ");
 	bch2_bpos_to_text(out, path->pos);
 
 	if (!path->cached && btree_node_locked(path, path->level)) {
@@ -1717,8 +1752,7 @@ btree_path_idx_t bch2_path_get(struct btree_trans *trans,
 	struct trans_for_each_path_inorder_iter iter;
 	btree_path_idx_t path_pos = 0, path_idx;
 
-	bch2_trans_verify_not_unlocked(trans);
-	bch2_trans_verify_not_in_restart(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 	bch2_trans_verify_locks(trans);
 
 	btree_trans_sort_paths(trans);
@@ -1833,7 +1867,7 @@ struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *
 			!bkey_eq(path->pos, ck->key.pos));
 
 		*u = ck->k->k;
-		k = bkey_i_to_s_c(ck->k);
+		k = (struct bkey_s_c) { u, &ck->k->v };
 	}
 
 	return k;
@@ -1843,7 +1877,6 @@ hole:
 	return (struct bkey_s_c) { u, NULL };
 }
 
-
 void bch2_set_btree_iter_dontneed(struct btree_iter *iter)
 {
 	struct btree_trans *trans = iter->trans;
@@ -1870,7 +1903,7 @@ bch2_btree_iter_traverse(struct btree_iter *iter)
 	struct btree_trans *trans = iter->trans;
 	int ret;
 
-	bch2_trans_verify_not_unlocked(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 
 	iter->path = bch2_btree_path_set_pos(trans, iter->path,
 					btree_iter_search_key(iter),
@@ -1945,7 +1978,7 @@ struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
 	int ret;
 
 	EBUG_ON(trans->paths[iter->path].cached);
-	bch2_trans_verify_not_in_restart(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 	bch2_btree_iter_verify(iter);
 
 	ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
@@ -2101,7 +2134,7 @@ static struct bkey_i *bch2_btree_journal_peek(struct btree_trans *trans,
 {
 	struct btree_path *path = btree_iter_path(trans, iter);
 
-	return bch2_journal_keys_peek_upto(trans->c, iter->btree_id,
+	return bch2_journal_keys_peek_max(trans->c, iter->btree_id,
 					   path->level,
 					   path->pos,
 					   end_pos,
@@ -2124,21 +2157,47 @@ struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
 }
 
 static noinline
-struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
-					 struct btree_iter *iter,
-					 struct bkey_s_c k)
+void btree_trans_peek_journal(struct btree_trans *trans,
+			      struct btree_iter *iter,
+			      struct bkey_s_c *k)
 {
 	struct btree_path *path = btree_iter_path(trans, iter);
 	struct bkey_i *next_journal =
 		bch2_btree_journal_peek(trans, iter,
-				k.k ? k.k->p : path_l(path)->b->key.k.p);
-
+				k->k ? k->k->p : path_l(path)->b->key.k.p);
 	if (next_journal) {
 		iter->k = next_journal->k;
-		k = bkey_i_to_s_c(next_journal);
+		*k = bkey_i_to_s_c(next_journal);
 	}
+}
 
-	return k;
+static struct bkey_i *bch2_btree_journal_peek_prev(struct btree_trans *trans,
+					      struct btree_iter *iter,
+					      struct bpos end_pos)
+{
+	struct btree_path *path = btree_iter_path(trans, iter);
+
+	return bch2_journal_keys_peek_prev_min(trans->c, iter->btree_id,
+					   path->level,
+					   path->pos,
+					   end_pos,
+					   &iter->journal_idx);
+}
+
+static noinline
+void btree_trans_peek_prev_journal(struct btree_trans *trans,
+				   struct btree_iter *iter,
+				   struct bkey_s_c *k)
+{
+	struct btree_path *path = btree_iter_path(trans, iter);
+	struct bkey_i *next_journal =
+		bch2_btree_journal_peek_prev(trans, iter,
+				k->k ? k->k->p : path_l(path)->b->key.k.p);
+
+	if (next_journal) {
+		iter->k = next_journal->k;
+		*k = bkey_i_to_s_c(next_journal);
+	}
 }
 
 /*
@@ -2154,8 +2213,7 @@ struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos
 	struct bkey_s_c k;
 	int ret;
 
-	bch2_trans_verify_not_in_restart(trans);
-	bch2_trans_verify_not_unlocked(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 
 	if ((iter->flags & BTREE_ITER_key_cache_fill) &&
 	    bpos_eq(iter->pos, pos))
@@ -2184,10 +2242,15 @@ struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos
 	btree_path_set_should_be_locked(trans, trans->paths + iter->key_cache_path);
 
 	k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u);
-	if (k.k && !bkey_err(k)) {
-		iter->k = u;
-		k.k = &iter->k;
-	}
+	if (!k.k)
+		return k;
+
+	if ((iter->flags & BTREE_ITER_all_snapshots) &&
+	    !bpos_eq(pos, k.k->p))
+		return bkey_s_c_null;
+
+	iter->k = u;
+	k.k = &iter->k;
 	return k;
 }
 
@@ -2201,8 +2264,6 @@ static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bp
 	bch2_btree_iter_verify(iter);
 
 	while (1) {
-		struct btree_path_level *l;
-
 		iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
 					iter->flags & BTREE_ITER_intent,
 					btree_iter_ip_allocated(iter));
@@ -2212,17 +2273,17 @@ static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bp
 			/* ensure that iter->k is consistent with iter->pos: */
 			bch2_btree_iter_set_pos(iter, iter->pos);
 			k = bkey_s_c_err(ret);
-			goto out;
+			break;
 		}
 
 		struct btree_path *path = btree_iter_path(trans, iter);
-		l = path_l(path);
+		struct btree_path_level *l = path_l(path);
 
 		if (unlikely(!l->b)) {
 			/* No btree nodes at requested level: */
 			bch2_btree_iter_set_pos(iter, SPOS_MAX);
 			k = bkey_s_c_null;
-			goto out;
+			break;
 		}
 
 		btree_path_set_should_be_locked(trans, path);
@@ -2233,15 +2294,14 @@ static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bp
 		    k.k &&
 		    (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
 			k = k2;
-			ret = bkey_err(k);
-			if (ret) {
+			if (bkey_err(k)) {
 				bch2_btree_iter_set_pos(iter, iter->pos);
-				goto out;
+				break;
 			}
 		}
 
 		if (unlikely(iter->flags & BTREE_ITER_with_journal))
-			k = btree_trans_peek_journal(trans, iter, k);
+			btree_trans_peek_journal(trans, iter, &k);
 
 		if (unlikely((iter->flags & BTREE_ITER_with_updates) &&
 			     trans->nr_updates))
@@ -2270,32 +2330,32 @@ static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bp
 			/* End of btree: */
 			bch2_btree_iter_set_pos(iter, SPOS_MAX);
 			k = bkey_s_c_null;
-			goto out;
+			break;
 		}
 	}
-out:
-	bch2_btree_iter_verify(iter);
 
+	bch2_btree_iter_verify(iter);
 	return k;
 }
 
 /**
- * bch2_btree_iter_peek_upto() - returns first key greater than or equal to
+ * bch2_btree_iter_peek_max() - returns first key greater than or equal to
  * iterator's current position
  * @iter:	iterator to peek from
  * @end:	search limit: returns keys less than or equal to @end
  *
  * Returns:	key if found, or an error extractable with bkey_err().
  */
-struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos end)
+struct bkey_s_c bch2_btree_iter_peek_max(struct btree_iter *iter, struct bpos end)
 {
 	struct btree_trans *trans = iter->trans;
 	struct bpos search_key = btree_iter_search_key(iter);
 	struct bkey_s_c k;
-	struct bpos iter_pos;
+	struct bpos iter_pos = iter->pos;
 	int ret;
 
-	bch2_trans_verify_not_unlocked(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
+	bch2_btree_iter_verify_entry_exit(iter);
 	EBUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && bkey_eq(end, POS_MAX));
 
 	if (iter->update_path) {
@@ -2304,8 +2364,6 @@ struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos e
 		iter->update_path = 0;
 	}
 
-	bch2_btree_iter_verify_entry_exit(iter);
-
 	while (1) {
 		k = __bch2_btree_iter_peek(iter, search_key);
 		if (unlikely(!k.k))
@@ -2313,75 +2371,75 @@ struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *iter, struct bpos e
 		if (unlikely(bkey_err(k)))
 			goto out_no_locked;
 
-		/*
-		 * We need to check against @end before FILTER_SNAPSHOTS because
-		 * if we get to a different inode that requested we might be
-		 * seeing keys for a different snapshot tree that will all be
-		 * filtered out.
-		 *
-		 * But we can't do the full check here, because bkey_start_pos()
-		 * isn't monotonically increasing before FILTER_SNAPSHOTS, and
-		 * that's what we check against in extents mode:
-		 */
-		if (unlikely(!(iter->flags & BTREE_ITER_is_extents)
-			     ? bkey_gt(k.k->p, end)
-			     : k.k->p.inode > end.inode))
-			goto end;
+		if (iter->flags & BTREE_ITER_filter_snapshots) {
+			/*
+			 * We need to check against @end before FILTER_SNAPSHOTS because
+			 * if we get to a different inode that requested we might be
+			 * seeing keys for a different snapshot tree that will all be
+			 * filtered out.
+			 *
+			 * But we can't do the full check here, because bkey_start_pos()
+			 * isn't monotonically increasing before FILTER_SNAPSHOTS, and
+			 * that's what we check against in extents mode:
+			 */
+			if (unlikely(!(iter->flags & BTREE_ITER_is_extents)
+				     ? bkey_gt(k.k->p, end)
+				     : k.k->p.inode > end.inode))
+				goto end;
+
+			if (iter->update_path &&
+			    !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) {
+				bch2_path_put_nokeep(trans, iter->update_path,
+						     iter->flags & BTREE_ITER_intent);
+				iter->update_path = 0;
+			}
 
-		if (iter->update_path &&
-		    !bkey_eq(trans->paths[iter->update_path].pos, k.k->p)) {
-			bch2_path_put_nokeep(trans, iter->update_path,
-					     iter->flags & BTREE_ITER_intent);
-			iter->update_path = 0;
-		}
+			if ((iter->flags & BTREE_ITER_intent) &&
+			    !(iter->flags & BTREE_ITER_is_extents) &&
+			    !iter->update_path) {
+				struct bpos pos = k.k->p;
 
-		if ((iter->flags & BTREE_ITER_filter_snapshots) &&
-		    (iter->flags & BTREE_ITER_intent) &&
-		    !(iter->flags & BTREE_ITER_is_extents) &&
-		    !iter->update_path) {
-			struct bpos pos = k.k->p;
+				if (pos.snapshot < iter->snapshot) {
+					search_key = bpos_successor(k.k->p);
+					continue;
+				}
 
-			if (pos.snapshot < iter->snapshot) {
-				search_key = bpos_successor(k.k->p);
-				continue;
-			}
+				pos.snapshot = iter->snapshot;
 
-			pos.snapshot = iter->snapshot;
+				/*
+				 * advance, same as on exit for iter->path, but only up
+				 * to snapshot
+				 */
+				__btree_path_get(trans, trans->paths + iter->path, iter->flags & BTREE_ITER_intent);
+				iter->update_path = iter->path;
+
+				iter->update_path = bch2_btree_path_set_pos(trans,
+							iter->update_path, pos,
+							iter->flags & BTREE_ITER_intent,
+							_THIS_IP_);
+				ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
+				if (unlikely(ret)) {
+					k = bkey_s_c_err(ret);
+					goto out_no_locked;
+				}
+			}
 
 			/*
-			 * advance, same as on exit for iter->path, but only up
-			 * to snapshot
+			 * We can never have a key in a leaf node at POS_MAX, so
+			 * we don't have to check these successor() calls:
 			 */
-			__btree_path_get(trans, trans->paths + iter->path, iter->flags & BTREE_ITER_intent);
-			iter->update_path = iter->path;
-
-			iter->update_path = bch2_btree_path_set_pos(trans,
-						iter->update_path, pos,
-						iter->flags & BTREE_ITER_intent,
-						_THIS_IP_);
-			ret = bch2_btree_path_traverse(trans, iter->update_path, iter->flags);
-			if (unlikely(ret)) {
-				k = bkey_s_c_err(ret);
-				goto out_no_locked;
+			if (!bch2_snapshot_is_ancestor(trans->c,
+						       iter->snapshot,
+						       k.k->p.snapshot)) {
+				search_key = bpos_successor(k.k->p);
+				continue;
 			}
-		}
-
-		/*
-		 * We can never have a key in a leaf node at POS_MAX, so
-		 * we don't have to check these successor() calls:
-		 */
-		if ((iter->flags & BTREE_ITER_filter_snapshots) &&
-		    !bch2_snapshot_is_ancestor(trans->c,
-					       iter->snapshot,
-					       k.k->p.snapshot)) {
-			search_key = bpos_successor(k.k->p);
-			continue;
-		}
 
-		if (bkey_whiteout(k.k) &&
-		    !(iter->flags & BTREE_ITER_all_snapshots)) {
-			search_key = bkey_successor(iter, k.k->p);
-			continue;
+			if (bkey_whiteout(k.k) &&
+			    !(iter->flags & BTREE_ITER_key_cache_fill)) {
+				search_key = bkey_successor(iter, k.k->p);
+				continue;
+			}
 		}
 
 		/*
@@ -2451,127 +2509,204 @@ struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
 	return bch2_btree_iter_peek(iter);
 }
 
-/**
- * bch2_btree_iter_peek_prev() - returns first key less than or equal to
- * iterator's current position
- * @iter:	iterator to peek from
- *
- * Returns:	key if found, or an error extractable with bkey_err().
- */
-struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
+static struct bkey_s_c __bch2_btree_iter_peek_prev(struct btree_iter *iter, struct bpos search_key)
 {
 	struct btree_trans *trans = iter->trans;
-	struct bpos search_key = iter->pos;
-	struct bkey_s_c k;
-	struct bkey saved_k;
-	const struct bch_val *saved_v;
-	btree_path_idx_t saved_path = 0;
-	int ret;
-
-	bch2_trans_verify_not_unlocked(trans);
-	EBUG_ON(btree_iter_path(trans, iter)->cached ||
-		btree_iter_path(trans, iter)->level);
-
-	if (iter->flags & BTREE_ITER_with_journal)
-		return bkey_s_c_err(-BCH_ERR_btree_iter_with_journal_not_supported);
+	struct bkey_s_c k, k2;
 
 	bch2_btree_iter_verify(iter);
-	bch2_btree_iter_verify_entry_exit(iter);
-
-	if (iter->flags & BTREE_ITER_filter_snapshots)
-		search_key.snapshot = U32_MAX;
 
 	while (1) {
 		iter->path = bch2_btree_path_set_pos(trans, iter->path, search_key,
-						iter->flags & BTREE_ITER_intent,
-						btree_iter_ip_allocated(iter));
+					iter->flags & BTREE_ITER_intent,
+					btree_iter_ip_allocated(iter));
 
-		ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
+		int ret = bch2_btree_path_traverse(trans, iter->path, iter->flags);
 		if (unlikely(ret)) {
 			/* ensure that iter->k is consistent with iter->pos: */
 			bch2_btree_iter_set_pos(iter, iter->pos);
 			k = bkey_s_c_err(ret);
-			goto out_no_locked;
+			break;
 		}
 
 		struct btree_path *path = btree_iter_path(trans, iter);
+		struct btree_path_level *l = path_l(path);
+
+		if (unlikely(!l->b)) {
+			/* No btree nodes at requested level: */
+			bch2_btree_iter_set_pos(iter, SPOS_MAX);
+			k = bkey_s_c_null;
+			break;
+		}
+
+		btree_path_set_should_be_locked(trans, path);
 
-		k = btree_path_level_peek(trans, path, &path->l[0], &iter->k);
-		if (!k.k ||
-		    ((iter->flags & BTREE_ITER_is_extents)
-		     ? bpos_ge(bkey_start_pos(k.k), search_key)
-		     : bpos_gt(k.k->p, search_key)))
-			k = btree_path_level_prev(trans, path, &path->l[0], &iter->k);
+		k = btree_path_level_peek_all(trans->c, l, &iter->k);
+		if (!k.k || bpos_gt(k.k->p, search_key)) {
+			k = btree_path_level_prev(trans, path, l, &iter->k);
+
+			BUG_ON(k.k && bpos_gt(k.k->p, search_key));
+		}
+
+		if (unlikely(iter->flags & BTREE_ITER_with_key_cache) &&
+		    k.k &&
+		    (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
+			k = k2;
+			if (bkey_err(k2)) {
+				bch2_btree_iter_set_pos(iter, iter->pos);
+				break;
+			}
+		}
+
+		if (unlikely(iter->flags & BTREE_ITER_with_journal))
+			btree_trans_peek_prev_journal(trans, iter, &k);
 
 		if (unlikely((iter->flags & BTREE_ITER_with_updates) &&
 			     trans->nr_updates))
 			bch2_btree_trans_peek_prev_updates(trans, iter, &k);
 
-		if (likely(k.k)) {
-			if (iter->flags & BTREE_ITER_filter_snapshots) {
-				if (k.k->p.snapshot == iter->snapshot)
-					goto got_key;
+		if (likely(k.k && !bkey_deleted(k.k))) {
+			break;
+		} else if (k.k) {
+			search_key = bpos_predecessor(k.k->p);
+		} else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) {
+			/* Advance to previous leaf node: */
+			search_key = bpos_predecessor(path->l[0].b->data->min_key);
+		} else {
+			/* Start of btree: */
+			bch2_btree_iter_set_pos(iter, POS_MIN);
+			k = bkey_s_c_null;
+			break;
+		}
+	}
+
+	bch2_btree_iter_verify(iter);
+	return k;
+}
+
+/**
+ * bch2_btree_iter_peek_prev_min() - returns first key less than or equal to
+ * iterator's current position
+ * @iter:	iterator to peek from
+ * @end:	search limit: returns keys greater than or equal to @end
+ *
+ * Returns:	key if found, or an error extractable with bkey_err().
+ */
+struct bkey_s_c bch2_btree_iter_peek_prev_min(struct btree_iter *iter, struct bpos end)
+{
+	if ((iter->flags & (BTREE_ITER_is_extents|BTREE_ITER_filter_snapshots)) &&
+	   !bkey_eq(iter->pos, POS_MAX)) {
+		/*
+		 * bkey_start_pos(), for extents, is not monotonically
+		 * increasing until after filtering for snapshots:
+		 *
+		 * Thus, for extents we need to search forward until we find a
+		 * real visible extents - easiest to just use peek_slot() (which
+		 * internally uses peek() for extents)
+		 */
+		struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
+		if (bkey_err(k))
+			return k;
+
+		if (!bkey_deleted(k.k) &&
+		    (!(iter->flags & BTREE_ITER_is_extents) ||
+		     bkey_lt(bkey_start_pos(k.k), iter->pos)))
+			return k;
+	}
+
+	struct btree_trans *trans = iter->trans;
+	struct bpos search_key = iter->pos;
+	struct bkey_s_c k;
+	btree_path_idx_t saved_path = 0;
+
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
+	bch2_btree_iter_verify_entry_exit(iter);
+	EBUG_ON((iter->flags & BTREE_ITER_filter_snapshots) && bpos_eq(end, POS_MIN));
+
+	while (1) {
+		k = __bch2_btree_iter_peek_prev(iter, search_key);
+		if (unlikely(!k.k))
+			goto end;
+		if (unlikely(bkey_err(k)))
+			goto out_no_locked;
 
+		if (iter->flags & BTREE_ITER_filter_snapshots) {
+			struct btree_path *s = saved_path ? trans->paths + saved_path : NULL;
+			if (s && bpos_lt(k.k->p, SPOS(s->pos.inode, s->pos.offset, iter->snapshot))) {
 				/*
-				 * If we have a saved candidate, and we're no
-				 * longer at the same _key_ (not pos), return
-				 * that candidate
+				 * If we have a saved candidate, and we're past
+				 * the last possible snapshot overwrite, return
+				 * it:
 				 */
-				if (saved_path && !bkey_eq(k.k->p, saved_k.p)) {
-					bch2_path_put_nokeep(trans, iter->path,
-						      iter->flags & BTREE_ITER_intent);
-					iter->path = saved_path;
+				bch2_path_put_nokeep(trans, iter->path,
+					      iter->flags & BTREE_ITER_intent);
+				iter->path = saved_path;
+				saved_path = 0;
+				k = bch2_btree_path_peek_slot(btree_iter_path(trans, iter), &iter->k);
+				break;
+			}
+
+			/*
+			 * We need to check against @end before FILTER_SNAPSHOTS because
+			 * if we get to a different inode that requested we might be
+			 * seeing keys for a different snapshot tree that will all be
+			 * filtered out.
+			 */
+			if (unlikely(bkey_lt(k.k->p, end)))
+				goto end;
+
+			if (!bch2_snapshot_is_ancestor(trans->c, iter->snapshot, k.k->p.snapshot)) {
+				search_key = bpos_predecessor(k.k->p);
+				continue;
+			}
+
+			if (k.k->p.snapshot != iter->snapshot) {
+				/*
+				 * Have a key visible in iter->snapshot, but
+				 * might have overwrites: - save it and keep
+				 * searching. Unless it's a whiteout - then drop
+				 * our previous saved candidate:
+				 */
+				if (saved_path) {
+					bch2_path_put_nokeep(trans, saved_path,
+					      iter->flags & BTREE_ITER_intent);
 					saved_path = 0;
-					iter->k	= saved_k;
-					k.v	= saved_v;
-					goto got_key;
 				}
 
-				if (bch2_snapshot_is_ancestor(trans->c,
-							      iter->snapshot,
-							      k.k->p.snapshot)) {
-					if (saved_path)
-						bch2_path_put_nokeep(trans, saved_path,
-						      iter->flags & BTREE_ITER_intent);
+				if (!bkey_whiteout(k.k)) {
 					saved_path = btree_path_clone(trans, iter->path,
 								iter->flags & BTREE_ITER_intent,
 								_THIS_IP_);
-					path = btree_iter_path(trans, iter);
-					trace_btree_path_save_pos(trans, path, trans->paths + saved_path);
-					saved_k = *k.k;
-					saved_v = k.v;
+					trace_btree_path_save_pos(trans,
+								  trans->paths + iter->path,
+								  trans->paths + saved_path);
 				}
 
 				search_key = bpos_predecessor(k.k->p);
 				continue;
 			}
-got_key:
-			if (bkey_whiteout(k.k) &&
-			    !(iter->flags & BTREE_ITER_all_snapshots)) {
+
+			if (bkey_whiteout(k.k)) {
 				search_key = bkey_predecessor(iter, k.k->p);
-				if (iter->flags & BTREE_ITER_filter_snapshots)
-					search_key.snapshot = U32_MAX;
+				search_key.snapshot = U32_MAX;
 				continue;
 			}
-
-			btree_path_set_should_be_locked(trans, path);
-			break;
-		} else if (likely(!bpos_eq(path->l[0].b->data->min_key, POS_MIN))) {
-			/* Advance to previous leaf node: */
-			search_key = bpos_predecessor(path->l[0].b->data->min_key);
-		} else {
-			/* Start of btree: */
-			bch2_btree_iter_set_pos(iter, POS_MIN);
-			k = bkey_s_c_null;
-			goto out_no_locked;
 		}
-	}
 
-	EBUG_ON(bkey_gt(bkey_start_pos(k.k), iter->pos));
+		EBUG_ON(iter->flags & BTREE_ITER_all_snapshots		? bpos_gt(k.k->p, iter->pos) :
+			iter->flags & BTREE_ITER_is_extents		? bkey_ge(bkey_start_pos(k.k), iter->pos) :
+									  bkey_gt(k.k->p, iter->pos));
+
+		if (unlikely(iter->flags & BTREE_ITER_all_snapshots	? bpos_lt(k.k->p, end) :
+			     iter->flags & BTREE_ITER_is_extents	? bkey_le(k.k->p, end) :
+									  bkey_lt(k.k->p, end)))
+			goto end;
+
+		break;
+	}
 
 	/* Extents can straddle iter->pos: */
-	if (bkey_lt(k.k->p, iter->pos))
-		iter->pos = k.k->p;
+	iter->pos = bpos_min(iter->pos, k.k->p);;
 
 	if (iter->flags & BTREE_ITER_filter_snapshots)
 		iter->pos.snapshot = iter->snapshot;
@@ -2581,8 +2716,11 @@ out_no_locked:
 
 	bch2_btree_iter_verify_entry_exit(iter);
 	bch2_btree_iter_verify(iter);
-
 	return k;
+end:
+	bch2_btree_iter_set_pos(iter, end);
+	k = bkey_s_c_null;
+	goto out_no_locked;
 }
 
 /**
@@ -2607,7 +2745,7 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
 	struct bkey_s_c k;
 	int ret;
 
-	bch2_trans_verify_not_unlocked(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 	bch2_btree_iter_verify(iter);
 	bch2_btree_iter_verify_entry_exit(iter);
 	EBUG_ON(btree_iter_path(trans, iter)->level && (iter->flags & BTREE_ITER_with_key_cache));
@@ -2632,6 +2770,10 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
 		goto out_no_locked;
 	}
 
+	struct btree_path *path = btree_iter_path(trans, iter);
+	if (unlikely(!btree_path_node(path, path->level)))
+		return bkey_s_c_null;
+
 	if ((iter->flags & BTREE_ITER_cached) ||
 	    !(iter->flags & (BTREE_ITER_is_extents|BTREE_ITER_filter_snapshots))) {
 		k = bkey_s_c_null;
@@ -2658,6 +2800,11 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
 		k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k);
 		if (unlikely(!k.k))
 			goto out_no_locked;
+
+		if (unlikely(k.k->type == KEY_TYPE_whiteout &&
+			     (iter->flags & BTREE_ITER_filter_snapshots) &&
+			     !(iter->flags & BTREE_ITER_key_cache_fill)))
+			iter->k.type = KEY_TYPE_deleted;
 	} else {
 		struct bpos next;
 		struct bpos end = iter->pos;
@@ -2671,7 +2818,7 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
 			struct btree_iter iter2;
 
 			bch2_trans_copy_iter(&iter2, iter);
-			k = bch2_btree_iter_peek_upto(&iter2, end);
+			k = bch2_btree_iter_peek_max(&iter2, end);
 
 			if (k.k && !bkey_err(k)) {
 				swap(iter->key_cache_path, iter2.key_cache_path);
@@ -2682,7 +2829,7 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
 		} else {
 			struct bpos pos = iter->pos;
 
-			k = bch2_btree_iter_peek_upto(iter, end);
+			k = bch2_btree_iter_peek_max(iter, end);
 			if (unlikely(bkey_err(k)))
 				bch2_btree_iter_set_pos(iter, pos);
 			else
@@ -2902,7 +3049,7 @@ void bch2_trans_iter_init_outlined(struct btree_trans *trans,
 			  unsigned flags)
 {
 	bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
-			       bch2_btree_iter_flags(trans, btree_id, flags),
+			       bch2_btree_iter_flags(trans, btree_id, 0, flags),
 			       _RET_IP_);
 }
 
@@ -2918,8 +3065,11 @@ void bch2_trans_node_iter_init(struct btree_trans *trans,
 	flags |= BTREE_ITER_snapshot_field;
 	flags |= BTREE_ITER_all_snapshots;
 
+	if (!depth && btree_id_cached(trans->c, btree_id))
+		flags |= BTREE_ITER_with_key_cache;
+
 	bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
-			       __bch2_btree_iter_flags(trans, btree_id, flags),
+			       bch2_btree_iter_flags(trans, btree_id, depth, flags),
 			       _RET_IP_);
 
 	iter->min_depth	= depth;
@@ -3122,14 +3272,14 @@ u32 bch2_trans_begin(struct btree_trans *trans)
 
 	trans->last_begin_ip = _RET_IP_;
 
-	trans_set_locked(trans);
+	trans_set_locked(trans, false);
 
 	if (trans->restarted) {
 		bch2_btree_path_traverse_all(trans);
 		trans->notrace_relock_fail = false;
 	}
 
-	bch2_trans_verify_not_unlocked(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 	return trans->restart_count;
 }
 
@@ -3228,7 +3378,7 @@ got_trans:
 	trans->srcu_idx		= srcu_read_lock(&c->btree_trans_barrier);
 	trans->srcu_lock_time	= jiffies;
 	trans->srcu_held	= true;
-	trans_set_locked(trans);
+	trans_set_locked(trans, false);
 
 	closure_init_stack_release(&trans->ref);
 	return trans;
@@ -3262,6 +3412,9 @@ void bch2_trans_put(struct btree_trans *trans)
 {
 	struct bch_fs *c = trans->c;
 
+	if (trans->restarted)
+		bch2_trans_in_restart_error(trans);
+
 	bch2_trans_unlock(trans);
 
 	trans_for_each_update(trans, i)
@@ -3285,6 +3438,10 @@ void bch2_trans_put(struct btree_trans *trans)
 	closure_return_sync(&trans->ref);
 	trans->locking_wait.task = NULL;
 
+#ifdef CONFIG_BCACHEFS_DEBUG
+	darray_exit(&trans->last_restarted_trace);
+#endif
+
 	unsigned long *paths_allocated = trans->paths_allocated;
 	trans->paths_allocated	= NULL;
 	trans->paths		= NULL;
@@ -3338,8 +3495,9 @@ bch2_btree_bkey_cached_common_to_text(struct printbuf *out,
 	pid = owner ? owner->pid : 0;
 	rcu_read_unlock();
 
-	prt_printf(out, "\t%px %c l=%u %s:", b, b->cached ? 'c' : 'b',
-		   b->level, bch2_btree_id_str(b->btree_id));
+	prt_printf(out, "\t%px %c ", b, b->cached ? 'c' : 'b');
+	bch2_btree_id_to_text(out, b->btree_id);
+	prt_printf(out, " l=%u:", b->level);
 	bch2_bpos_to_text(out, btree_node_pos(b));
 
 	prt_printf(out, "\t locks %u:%u:%u held by pid %u",
@@ -3378,11 +3536,11 @@ void bch2_btree_trans_to_text(struct printbuf *out, struct btree_trans *trans)
 		if (!path->nodes_locked)
 			continue;
 
-		prt_printf(out, "  path %u %c l=%u %s:",
-		       idx,
-		       path->cached ? 'c' : 'b',
-		       path->level,
-		       bch2_btree_id_str(path->btree_id));
+		prt_printf(out, "  path %u %c ",
+			   idx,
+			   path->cached ? 'c' : 'b');
+		bch2_btree_id_to_text(out, path->btree_id);
+		prt_printf(out, " l=%u:", path->level);
 		bch2_bpos_to_text(out, path->pos);
 		prt_newline(out);
 
@@ -3488,7 +3646,7 @@ int bch2_fs_btree_iter_init(struct bch_fs *c)
 #ifdef CONFIG_LOCKDEP
 	fs_reclaim_acquire(GFP_KERNEL);
 	struct btree_trans *trans = bch2_trans_get(c);
-	trans_set_locked(trans);
+	trans_set_locked(trans, false);
 	bch2_trans_put(trans);
 	fs_reclaim_release(GFP_KERNEL);
 #endif
diff --git a/fs/bcachefs/btree_iter.h b/fs/bcachefs/btree_iter.h
index 0bda054f80d7..b9538e6e6d65 100644
--- a/fs/bcachefs/btree_iter.h
+++ b/fs/bcachefs/btree_iter.h
@@ -23,6 +23,7 @@ static inline void __btree_path_get(struct btree_trans *trans, struct btree_path
 {
 	unsigned idx = path - trans->paths;
 
+	EBUG_ON(idx >= trans->nr_paths);
 	EBUG_ON(!test_bit(idx, trans->paths_allocated));
 	if (unlikely(path->ref == U8_MAX)) {
 		bch2_dump_trans_paths_updates(trans);
@@ -36,6 +37,7 @@ static inline void __btree_path_get(struct btree_trans *trans, struct btree_path
 
 static inline bool __btree_path_put(struct btree_trans *trans, struct btree_path *path, bool intent)
 {
+	EBUG_ON(path - trans->paths >= trans->nr_paths);
 	EBUG_ON(!test_bit(path - trans->paths, trans->paths_allocated));
 	EBUG_ON(!path->ref);
 	EBUG_ON(!path->intent_ref && intent);
@@ -234,12 +236,12 @@ int __must_check bch2_btree_path_traverse_one(struct btree_trans *,
 					      btree_path_idx_t,
 					      unsigned, unsigned long);
 
-static inline void bch2_trans_verify_not_unlocked(struct btree_trans *);
+static inline void bch2_trans_verify_not_unlocked_or_in_restart(struct btree_trans *);
 
 static inline int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
 					  btree_path_idx_t path, unsigned flags)
 {
-	bch2_trans_verify_not_unlocked(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 
 	if (trans->paths[path].uptodate < BTREE_ITER_NEED_RELOCK)
 		return 0;
@@ -324,38 +326,33 @@ static inline void bch2_trans_verify_not_restarted(struct btree_trans *trans,
 		bch2_trans_restart_error(trans, restart_count);
 }
 
-void __noreturn bch2_trans_in_restart_error(struct btree_trans *);
+void __noreturn bch2_trans_unlocked_or_in_restart_error(struct btree_trans *);
 
-static inline void bch2_trans_verify_not_in_restart(struct btree_trans *trans)
+static inline void bch2_trans_verify_not_unlocked_or_in_restart(struct btree_trans *trans)
 {
-	if (trans->restarted)
-		bch2_trans_in_restart_error(trans);
-}
-
-void __noreturn bch2_trans_unlocked_error(struct btree_trans *);
-
-static inline void bch2_trans_verify_not_unlocked(struct btree_trans *trans)
-{
-	if (!trans->locked)
-		bch2_trans_unlocked_error(trans);
+	if (trans->restarted || !trans->locked)
+		bch2_trans_unlocked_or_in_restart_error(trans);
 }
 
 __always_inline
-static int btree_trans_restart_nounlock(struct btree_trans *trans, int err)
+static int btree_trans_restart_ip(struct btree_trans *trans, int err, unsigned long ip)
 {
 	BUG_ON(err <= 0);
 	BUG_ON(!bch2_err_matches(-err, BCH_ERR_transaction_restart));
 
 	trans->restarted = err;
-	trans->last_restarted_ip = _THIS_IP_;
+	trans->last_restarted_ip = ip;
+#ifdef CONFIG_BCACHEFS_DEBUG
+	darray_exit(&trans->last_restarted_trace);
+	bch2_save_backtrace(&trans->last_restarted_trace, current, 0, GFP_NOWAIT);
+#endif
 	return -err;
 }
 
 __always_inline
 static int btree_trans_restart(struct btree_trans *trans, int err)
 {
-	btree_trans_restart_nounlock(trans, err);
-	return -err;
+	return btree_trans_restart_ip(trans, err, _THIS_IP_);
 }
 
 bool bch2_btree_node_upgrade(struct btree_trans *,
@@ -375,6 +372,7 @@ static inline void bch2_btree_path_downgrade(struct btree_trans *trans,
 void bch2_trans_downgrade(struct btree_trans *);
 
 void bch2_trans_node_add(struct btree_trans *trans, struct btree_path *, struct btree *);
+void bch2_trans_node_drop(struct btree_trans *trans, struct btree *);
 void bch2_trans_node_reinit_iter(struct btree_trans *, struct btree *);
 
 int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter);
@@ -384,15 +382,21 @@ struct btree *bch2_btree_iter_peek_node(struct btree_iter *);
 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *);
 struct btree *bch2_btree_iter_next_node(struct btree_iter *);
 
-struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *, struct bpos);
+struct bkey_s_c bch2_btree_iter_peek_max(struct btree_iter *, struct bpos);
 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *);
 
 static inline struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
 {
-	return bch2_btree_iter_peek_upto(iter, SPOS_MAX);
+	return bch2_btree_iter_peek_max(iter, SPOS_MAX);
+}
+
+struct bkey_s_c bch2_btree_iter_peek_prev_min(struct btree_iter *, struct bpos);
+
+static inline struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
+{
+	return bch2_btree_iter_peek_prev_min(iter, POS_MIN);
 }
 
-struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *);
 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *);
 
 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *);
@@ -443,10 +447,17 @@ static inline void bch2_btree_iter_set_snapshot(struct btree_iter *iter, u32 sna
 
 void bch2_trans_iter_exit(struct btree_trans *, struct btree_iter *);
 
-static inline unsigned __bch2_btree_iter_flags(struct btree_trans *trans,
-					       unsigned btree_id,
-					       unsigned flags)
+static inline unsigned bch2_btree_iter_flags(struct btree_trans *trans,
+					     unsigned btree_id,
+					     unsigned level,
+					     unsigned flags)
 {
+	if (level || !btree_id_cached(trans->c, btree_id)) {
+		flags &= ~BTREE_ITER_cached;
+		flags &= ~BTREE_ITER_with_key_cache;
+	} else if (!(flags & BTREE_ITER_cached))
+		flags |= BTREE_ITER_with_key_cache;
+
 	if (!(flags & (BTREE_ITER_all_snapshots|BTREE_ITER_not_extents)) &&
 	    btree_id_is_extents(btree_id))
 		flags |= BTREE_ITER_is_extents;
@@ -465,19 +476,6 @@ static inline unsigned __bch2_btree_iter_flags(struct btree_trans *trans,
 	return flags;
 }
 
-static inline unsigned bch2_btree_iter_flags(struct btree_trans *trans,
-					     unsigned btree_id,
-					     unsigned flags)
-{
-	if (!btree_id_cached(trans->c, btree_id)) {
-		flags &= ~BTREE_ITER_cached;
-		flags &= ~BTREE_ITER_with_key_cache;
-	} else if (!(flags & BTREE_ITER_cached))
-		flags |= BTREE_ITER_with_key_cache;
-
-	return __bch2_btree_iter_flags(trans, btree_id, flags);
-}
-
 static inline void bch2_trans_iter_init_common(struct btree_trans *trans,
 					  struct btree_iter *iter,
 					  unsigned btree_id, struct bpos pos,
@@ -514,7 +512,7 @@ static inline void bch2_trans_iter_init(struct btree_trans *trans,
 	if (__builtin_constant_p(btree_id) &&
 	    __builtin_constant_p(flags))
 		bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
-				bch2_btree_iter_flags(trans, btree_id, flags),
+				bch2_btree_iter_flags(trans, btree_id, 0, flags),
 				_THIS_IP_);
 	else
 		bch2_trans_iter_init_outlined(trans, iter, btree_id, pos, flags);
@@ -593,13 +591,18 @@ static inline struct bkey_s_c bch2_bkey_get_iter(struct btree_trans *trans,
 	bkey_s_c_to_##_type(__bch2_bkey_get_iter(_trans, _iter,			\
 				       _btree_id, _pos, _flags, KEY_TYPE_##_type))
 
+static inline void __bkey_val_copy(void *dst_v, unsigned dst_size, struct bkey_s_c src_k)
+{
+	unsigned b = min_t(unsigned, dst_size, bkey_val_bytes(src_k.k));
+	memcpy(dst_v, src_k.v, b);
+	if (unlikely(b < dst_size))
+		memset(dst_v + b, 0, dst_size - b);
+}
+
 #define bkey_val_copy(_dst_v, _src_k)					\
 do {									\
-	unsigned b = min_t(unsigned, sizeof(*_dst_v),			\
-			   bkey_val_bytes(_src_k.k));			\
-	memcpy(_dst_v, _src_k.v, b);					\
-	if (b < sizeof(*_dst_v))					\
-		memset((void *) (_dst_v) + b, 0, sizeof(*_dst_v) - b);	\
+	BUILD_BUG_ON(!__typecheck(*_dst_v, *_src_k.v));			\
+	__bkey_val_copy(_dst_v, sizeof(*_dst_v), _src_k.s_c);		\
 } while (0)
 
 static inline int __bch2_bkey_get_val_typed(struct btree_trans *trans,
@@ -608,17 +611,10 @@ static inline int __bch2_bkey_get_val_typed(struct btree_trans *trans,
 				unsigned val_size, void *val)
 {
 	struct btree_iter iter;
-	struct bkey_s_c k;
-	int ret;
-
-	k = __bch2_bkey_get_iter(trans, &iter, btree_id, pos, flags, type);
-	ret = bkey_err(k);
+	struct bkey_s_c k = __bch2_bkey_get_iter(trans, &iter, btree_id, pos, flags, type);
+	int ret = bkey_err(k);
 	if (!ret) {
-		unsigned b = min_t(unsigned, bkey_val_bytes(k.k), val_size);
-
-		memcpy(val, k.v, b);
-		if (unlikely(b < sizeof(*val)))
-			memset((void *) val + b, 0, sizeof(*val) - b);
+		__bkey_val_copy(val, val_size, k);
 		bch2_trans_iter_exit(trans, &iter);
 	}
 
@@ -677,12 +673,12 @@ static inline struct bkey_s_c bch2_btree_iter_peek_type(struct btree_iter *iter,
 						bch2_btree_iter_peek(iter);
 }
 
-static inline struct bkey_s_c bch2_btree_iter_peek_upto_type(struct btree_iter *iter,
+static inline struct bkey_s_c bch2_btree_iter_peek_max_type(struct btree_iter *iter,
 							     struct bpos end,
 							     unsigned flags)
 {
 	if (!(flags & BTREE_ITER_slots))
-		return bch2_btree_iter_peek_upto(iter, end);
+		return bch2_btree_iter_peek_max(iter, end);
 
 	if (bkey_gt(iter->pos, end))
 		return bkey_s_c_null;
@@ -746,7 +742,7 @@ transaction_restart:							\
 	_ret2 ?: trans_was_restarted(_trans, _restart_count);		\
 })
 
-#define for_each_btree_key_upto_continue(_trans, _iter,			\
+#define for_each_btree_key_max_continue(_trans, _iter,			\
 					 _end, _flags, _k, _do)		\
 ({									\
 	struct bkey_s_c _k;						\
@@ -754,7 +750,7 @@ transaction_restart:							\
 									\
 	do {								\
 		_ret3 = lockrestart_do(_trans, ({			\
-			(_k) = bch2_btree_iter_peek_upto_type(&(_iter),	\
+			(_k) = bch2_btree_iter_peek_max_type(&(_iter),	\
 						_end, (_flags));	\
 			if (!(_k).k)					\
 				break;					\
@@ -768,9 +764,9 @@ transaction_restart:							\
 })
 
 #define for_each_btree_key_continue(_trans, _iter, _flags, _k, _do)	\
-	for_each_btree_key_upto_continue(_trans, _iter, SPOS_MAX, _flags, _k, _do)
+	for_each_btree_key_max_continue(_trans, _iter, SPOS_MAX, _flags, _k, _do)
 
-#define for_each_btree_key_upto(_trans, _iter, _btree_id,		\
+#define for_each_btree_key_max(_trans, _iter, _btree_id,		\
 				_start, _end, _flags, _k, _do)		\
 ({									\
 	bch2_trans_begin(trans);					\
@@ -779,12 +775,12 @@ transaction_restart:							\
 	bch2_trans_iter_init((_trans), &(_iter), (_btree_id),		\
 			     (_start), (_flags));			\
 									\
-	for_each_btree_key_upto_continue(_trans, _iter, _end, _flags, _k, _do);\
+	for_each_btree_key_max_continue(_trans, _iter, _end, _flags, _k, _do);\
 })
 
 #define for_each_btree_key(_trans, _iter, _btree_id,			\
 			   _start, _flags, _k, _do)			\
-	for_each_btree_key_upto(_trans, _iter, _btree_id, _start,	\
+	for_each_btree_key_max(_trans, _iter, _btree_id, _start,	\
 				 SPOS_MAX, _flags, _k, _do)
 
 #define for_each_btree_key_reverse(_trans, _iter, _btree_id,		\
@@ -828,33 +824,33 @@ transaction_restart:							\
 			    (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 					(_journal_seq), (_commit_flags)))
 
-#define for_each_btree_key_upto_commit(_trans, _iter, _btree_id,	\
+#define for_each_btree_key_max_commit(_trans, _iter, _btree_id,	\
 				  _start, _end, _iter_flags, _k,	\
 				  _disk_res, _journal_seq, _commit_flags,\
 				  _do)					\
-	for_each_btree_key_upto(_trans, _iter, _btree_id, _start, _end, _iter_flags, _k,\
+	for_each_btree_key_max(_trans, _iter, _btree_id, _start, _end, _iter_flags, _k,\
 			    (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 					(_journal_seq), (_commit_flags)))
 
 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *);
 
-#define for_each_btree_key_upto_norestart(_trans, _iter, _btree_id,	\
+#define for_each_btree_key_max_norestart(_trans, _iter, _btree_id,	\
 			   _start, _end, _flags, _k, _ret)		\
 	for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),	\
 				  (_start), (_flags));			\
-	     (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),\
+	     (_k) = bch2_btree_iter_peek_max_type(&(_iter), _end, _flags),\
 	     !((_ret) = bkey_err(_k)) && (_k).k;			\
 	     bch2_btree_iter_advance(&(_iter)))
 
-#define for_each_btree_key_upto_continue_norestart(_iter, _end, _flags, _k, _ret)\
+#define for_each_btree_key_max_continue_norestart(_iter, _end, _flags, _k, _ret)\
 	for (;									\
-	     (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),	\
+	     (_k) = bch2_btree_iter_peek_max_type(&(_iter), _end, _flags),	\
 	     !((_ret) = bkey_err(_k)) && (_k).k;				\
 	     bch2_btree_iter_advance(&(_iter)))
 
 #define for_each_btree_key_norestart(_trans, _iter, _btree_id,		\
 			   _start, _flags, _k, _ret)			\
-	for_each_btree_key_upto_norestart(_trans, _iter, _btree_id, _start,\
+	for_each_btree_key_max_norestart(_trans, _iter, _btree_id, _start,\
 					  SPOS_MAX, _flags, _k, _ret)
 
 #define for_each_btree_key_reverse_norestart(_trans, _iter, _btree_id,	\
@@ -866,7 +862,7 @@ struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *);
 	     bch2_btree_iter_rewind(&(_iter)))
 
 #define for_each_btree_key_continue_norestart(_iter, _flags, _k, _ret)	\
-	for_each_btree_key_upto_continue_norestart(_iter, SPOS_MAX, _flags, _k, _ret)
+	for_each_btree_key_max_continue_norestart(_iter, SPOS_MAX, _flags, _k, _ret)
 
 /*
  * This should not be used in a fastpath, without first trying _do in
diff --git a/fs/bcachefs/btree_journal_iter.c b/fs/bcachefs/btree_journal_iter.c
index c1657182c275..6d25e3f85ce8 100644
--- a/fs/bcachefs/btree_journal_iter.c
+++ b/fs/bcachefs/btree_journal_iter.c
@@ -16,6 +16,17 @@
  * operations for the regular btree iter code to use:
  */
 
+static inline size_t pos_to_idx(struct journal_keys *keys, size_t pos)
+{
+	size_t gap_size = keys->size - keys->nr;
+
+	BUG_ON(pos >= keys->gap && pos < keys->gap + gap_size);
+
+	if (pos >= keys->gap)
+		pos -= gap_size;
+	return pos;
+}
+
 static inline size_t idx_to_pos(struct journal_keys *keys, size_t idx)
 {
 	size_t gap_size = keys->size - keys->nr;
@@ -61,7 +72,7 @@ static size_t bch2_journal_key_search(struct journal_keys *keys,
 }
 
 /* Returns first non-overwritten key >= search key: */
-struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *c, enum btree_id btree_id,
+struct bkey_i *bch2_journal_keys_peek_max(struct bch_fs *c, enum btree_id btree_id,
 					   unsigned level, struct bpos pos,
 					   struct bpos end_pos, size_t *idx)
 {
@@ -84,27 +95,92 @@ search:
 		}
 	}
 
+	struct bkey_i *ret = NULL;
+	rcu_read_lock(); /* for overwritten_ranges */
+
 	while ((k = *idx < keys->nr ? idx_to_key(keys, *idx) : NULL)) {
 		if (__journal_key_cmp(btree_id, level, end_pos, k) < 0)
-			return NULL;
+			break;
 
 		if (k->overwritten) {
-			(*idx)++;
+			if (k->overwritten_range)
+				*idx = rcu_dereference(k->overwritten_range)->end;
+			else
+				*idx += 1;
 			continue;
 		}
 
-		if (__journal_key_cmp(btree_id, level, pos, k) <= 0)
-			return k->k;
+		if (__journal_key_cmp(btree_id, level, pos, k) <= 0) {
+			ret = k->k;
+			break;
+		}
 
 		(*idx)++;
 		iters++;
 		if (iters == 10) {
 			*idx = 0;
+			rcu_read_unlock();
 			goto search;
 		}
 	}
 
-	return NULL;
+	rcu_read_unlock();
+	return ret;
+}
+
+struct bkey_i *bch2_journal_keys_peek_prev_min(struct bch_fs *c, enum btree_id btree_id,
+					   unsigned level, struct bpos pos,
+					   struct bpos end_pos, size_t *idx)
+{
+	struct journal_keys *keys = &c->journal_keys;
+	unsigned iters = 0;
+	struct journal_key *k;
+
+	BUG_ON(*idx > keys->nr);
+search:
+	if (!*idx)
+		*idx = __bch2_journal_key_search(keys, btree_id, level, pos);
+
+	while (*idx &&
+	       __journal_key_cmp(btree_id, level, end_pos, idx_to_key(keys, *idx - 1)) <= 0) {
+		(*idx)++;
+		iters++;
+		if (iters == 10) {
+			*idx = 0;
+			goto search;
+		}
+	}
+
+	struct bkey_i *ret = NULL;
+	rcu_read_lock(); /* for overwritten_ranges */
+
+	while ((k = *idx < keys->nr ? idx_to_key(keys, *idx) : NULL)) {
+		if (__journal_key_cmp(btree_id, level, end_pos, k) > 0)
+			break;
+
+		if (k->overwritten) {
+			if (k->overwritten_range)
+				*idx = rcu_dereference(k->overwritten_range)->start - 1;
+			else
+				*idx -= 1;
+			continue;
+		}
+
+		if (__journal_key_cmp(btree_id, level, pos, k) >= 0) {
+			ret = k->k;
+			break;
+		}
+
+		--(*idx);
+		iters++;
+		if (iters == 10) {
+			*idx = 0;
+			goto search;
+		}
+	}
+
+	rcu_read_unlock();
+	return ret;
 }
 
 struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *c, enum btree_id btree_id,
@@ -112,11 +188,12 @@ struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *c, enum btree_id btree
 {
 	size_t idx = 0;
 
-	return bch2_journal_keys_peek_upto(c, btree_id, level, pos, pos, &idx);
+	return bch2_journal_keys_peek_max(c, btree_id, level, pos, pos, &idx);
 }
 
 static void journal_iter_verify(struct journal_iter *iter)
 {
+#ifdef CONFIG_BCACHEFS_DEBUG
 	struct journal_keys *keys = iter->keys;
 	size_t gap_size = keys->size - keys->nr;
 
@@ -126,10 +203,10 @@ static void journal_iter_verify(struct journal_iter *iter)
 	if (iter->idx < keys->size) {
 		struct journal_key *k = keys->data + iter->idx;
 
-		int cmp = cmp_int(k->btree_id,	iter->btree_id) ?:
-			  cmp_int(k->level,	iter->level);
-		BUG_ON(cmp < 0);
+		int cmp = __journal_key_btree_cmp(iter->btree_id, iter->level, k);
+		BUG_ON(cmp > 0);
 	}
+#endif
 }
 
 static void journal_iters_fix(struct bch_fs *c)
@@ -182,7 +259,7 @@ int bch2_journal_key_insert_take(struct bch_fs *c, enum btree_id id,
 		 * Ensure these keys are done last by journal replay, to unblock
 		 * journal reclaim:
 		 */
-		.journal_seq	= U32_MAX,
+		.journal_seq	= U64_MAX,
 	};
 	struct journal_keys *keys = &c->journal_keys;
 	size_t idx = bch2_journal_key_search(keys, id, level, k->k.p);
@@ -290,6 +367,68 @@ bool bch2_key_deleted_in_journal(struct btree_trans *trans, enum btree_id btree,
 		bkey_deleted(&keys->data[idx].k->k));
 }
 
+static void __bch2_journal_key_overwritten(struct journal_keys *keys, size_t pos)
+{
+	struct journal_key *k = keys->data + pos;
+	size_t idx = pos_to_idx(keys, pos);
+
+	k->overwritten = true;
+
+	struct journal_key *prev = idx > 0 ? keys->data + idx_to_pos(keys, idx - 1) : NULL;
+	struct journal_key *next = idx + 1 < keys->nr ? keys->data + idx_to_pos(keys, idx + 1) : NULL;
+
+	bool prev_overwritten = prev && prev->overwritten;
+	bool next_overwritten = next && next->overwritten;
+
+	struct journal_key_range_overwritten *prev_range =
+		prev_overwritten ? prev->overwritten_range : NULL;
+	struct journal_key_range_overwritten *next_range =
+		next_overwritten ? next->overwritten_range : NULL;
+
+	BUG_ON(prev_range && prev_range->end != idx);
+	BUG_ON(next_range && next_range->start != idx + 1);
+
+	if (prev_range && next_range) {
+		prev_range->end = next_range->end;
+
+		keys->data[pos].overwritten_range = prev_range;
+		for (size_t i = next_range->start; i < next_range->end; i++) {
+			struct journal_key *ip = keys->data + idx_to_pos(keys, i);
+			BUG_ON(ip->overwritten_range != next_range);
+			ip->overwritten_range = prev_range;
+		}
+
+		kfree_rcu_mightsleep(next_range);
+	} else if (prev_range) {
+		prev_range->end++;
+		k->overwritten_range = prev_range;
+		if (next_overwritten) {
+			prev_range->end++;
+			next->overwritten_range = prev_range;
+		}
+	} else if (next_range) {
+		next_range->start--;
+		k->overwritten_range = next_range;
+		if (prev_overwritten) {
+			next_range->start--;
+			prev->overwritten_range = next_range;
+		}
+	} else if (prev_overwritten || next_overwritten) {
+		struct journal_key_range_overwritten *r = kmalloc(sizeof(*r), GFP_KERNEL);
+		if (!r)
+			return;
+
+		r->start = idx - (size_t) prev_overwritten;
+		r->end = idx + 1 + (size_t) next_overwritten;
+
+		rcu_assign_pointer(k->overwritten_range, r);
+		if (prev_overwritten)
+			prev->overwritten_range = r;
+		if (next_overwritten)
+			next->overwritten_range = r;
+	}
+}
+
 void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
 				  unsigned level, struct bpos pos)
 {
@@ -299,8 +438,12 @@ void bch2_journal_key_overwritten(struct bch_fs *c, enum btree_id btree,
 	if (idx < keys->size &&
 	    keys->data[idx].btree_id	== btree &&
 	    keys->data[idx].level	== level &&
-	    bpos_eq(keys->data[idx].k->k.p, pos))
-		keys->data[idx].overwritten = true;
+	    bpos_eq(keys->data[idx].k->k.p, pos) &&
+	    !keys->data[idx].overwritten) {
+		mutex_lock(&keys->overwrite_lock);
+		__bch2_journal_key_overwritten(keys, idx);
+		mutex_unlock(&keys->overwrite_lock);
+	}
 }
 
 static void bch2_journal_iter_advance(struct journal_iter *iter)
@@ -314,24 +457,32 @@ static void bch2_journal_iter_advance(struct journal_iter *iter)
 
 static struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
 {
+	struct bkey_s_c ret = bkey_s_c_null;
+
 	journal_iter_verify(iter);
 
+	rcu_read_lock();
 	while (iter->idx < iter->keys->size) {
 		struct journal_key *k = iter->keys->data + iter->idx;
 
-		int cmp = cmp_int(k->btree_id,	iter->btree_id) ?:
-			  cmp_int(k->level,	iter->level);
-		if (cmp > 0)
+		int cmp = __journal_key_btree_cmp(iter->btree_id, iter->level, k);
+		if (cmp < 0)
 			break;
 		BUG_ON(cmp);
 
-		if (!k->overwritten)
-			return bkey_i_to_s_c(k->k);
+		if (!k->overwritten) {
+			ret = bkey_i_to_s_c(k->k);
+			break;
+		}
 
-		bch2_journal_iter_advance(iter);
+		if (k->overwritten_range)
+			iter->idx = idx_to_pos(iter->keys, rcu_dereference(k->overwritten_range)->end);
+		else
+			bch2_journal_iter_advance(iter);
 	}
+	rcu_read_unlock();
 
-	return bkey_s_c_null;
+	return ret;
 }
 
 static void bch2_journal_iter_exit(struct journal_iter *iter)
@@ -382,6 +533,7 @@ static void btree_and_journal_iter_prefetch(struct btree_and_journal_iter *_iter
 		: (level > 1 ? 1 : 16);
 
 	iter.prefetch = false;
+	iter.fail_if_too_many_whiteouts = true;
 	bch2_bkey_buf_init(&tmp);
 
 	while (nr--) {
@@ -400,6 +552,7 @@ static void btree_and_journal_iter_prefetch(struct btree_and_journal_iter *_iter
 struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *iter)
 {
 	struct bkey_s_c btree_k, journal_k = bkey_s_c_null, ret;
+	size_t iters = 0;
 
 	if (iter->prefetch && iter->journal.level)
 		btree_and_journal_iter_prefetch(iter);
@@ -407,6 +560,11 @@ again:
 	if (iter->at_end)
 		return bkey_s_c_null;
 
+	iters++;
+
+	if (iters > 20 && iter->fail_if_too_many_whiteouts)
+		return bkey_s_c_null;
+
 	while ((btree_k = bch2_journal_iter_peek_btree(iter)).k &&
 	       bpos_lt(btree_k.k->p, iter->pos))
 		bch2_journal_iter_advance_btree(iter);
@@ -481,16 +639,6 @@ void bch2_btree_and_journal_iter_init_node_iter(struct btree_trans *trans,
 
 /* sort and dedup all keys in the journal: */
 
-void bch2_journal_entries_free(struct bch_fs *c)
-{
-	struct journal_replay **i;
-	struct genradix_iter iter;
-
-	genradix_for_each(&c->journal_entries, iter, i)
-		kvfree(*i);
-	genradix_free(&c->journal_entries);
-}
-
 /*
  * When keys compare equal, oldest compares first:
  */
@@ -515,15 +663,26 @@ void bch2_journal_keys_put(struct bch_fs *c)
 
 	move_gap(keys, keys->nr);
 
-	darray_for_each(*keys, i)
+	darray_for_each(*keys, i) {
+		if (i->overwritten_range &&
+		    (i == &darray_last(*keys) ||
+		     i->overwritten_range != i[1].overwritten_range))
+			kfree(i->overwritten_range);
+
 		if (i->allocated)
 			kfree(i->k);
+	}
 
 	kvfree(keys->data);
 	keys->data = NULL;
 	keys->nr = keys->gap = keys->size = 0;
 
-	bch2_journal_entries_free(c);
+	struct journal_replay **i;
+	struct genradix_iter iter;
+
+	genradix_for_each(&c->journal_entries, iter, i)
+		kvfree(*i);
+	genradix_free(&c->journal_entries);
 }
 
 static void __journal_keys_sort(struct journal_keys *keys)
@@ -628,8 +787,20 @@ void bch2_journal_keys_dump(struct bch_fs *c)
 
 	darray_for_each(*keys, i) {
 		printbuf_reset(&buf);
+		prt_printf(&buf, "btree=");
+		bch2_btree_id_to_text(&buf, i->btree_id);
+		prt_printf(&buf, " l=%u ", i->level);
 		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
-		pr_err("%s l=%u %s", bch2_btree_id_str(i->btree_id), i->level, buf.buf);
+		pr_err("%s", buf.buf);
 	}
 	printbuf_exit(&buf);
 }
+
+void bch2_fs_journal_keys_init(struct bch_fs *c)
+{
+	struct journal_keys *keys = &c->journal_keys;
+
+	atomic_set(&keys->ref, 1);
+	keys->initial_ref_held = true;
+	mutex_init(&keys->overwrite_lock);
+}
diff --git a/fs/bcachefs/btree_journal_iter.h b/fs/bcachefs/btree_journal_iter.h
index 1653de9d609b..2a3082919b8d 100644
--- a/fs/bcachefs/btree_journal_iter.h
+++ b/fs/bcachefs/btree_journal_iter.h
@@ -26,16 +26,24 @@ struct btree_and_journal_iter {
 	struct bpos		pos;
 	bool			at_end;
 	bool			prefetch;
+	bool			fail_if_too_many_whiteouts;
 };
 
+static inline int __journal_key_btree_cmp(enum btree_id	l_btree_id,
+					  unsigned	l_level,
+					  const struct journal_key *r)
+{
+	return -cmp_int(l_level,	r->level) ?:
+		cmp_int(l_btree_id,	r->btree_id);
+}
+
 static inline int __journal_key_cmp(enum btree_id	l_btree_id,
 				    unsigned		l_level,
 				    struct bpos	l_pos,
 				    const struct journal_key *r)
 {
-	return (cmp_int(l_btree_id,	r->btree_id) ?:
-		cmp_int(l_level,	r->level) ?:
-		bpos_cmp(l_pos,	r->k->k.p));
+	return __journal_key_btree_cmp(l_btree_id, l_level, r) ?:
+		bpos_cmp(l_pos,	r->k->k.p);
 }
 
 static inline int journal_key_cmp(const struct journal_key *l, const struct journal_key *r)
@@ -43,7 +51,9 @@ static inline int journal_key_cmp(const struct journal_key *l, const struct jour
 	return __journal_key_cmp(l->btree_id, l->level, l->k->k.p, r);
 }
 
-struct bkey_i *bch2_journal_keys_peek_upto(struct bch_fs *, enum btree_id,
+struct bkey_i *bch2_journal_keys_peek_max(struct bch_fs *, enum btree_id,
+				unsigned, struct bpos, struct bpos, size_t *);
+struct bkey_i *bch2_journal_keys_peek_prev_min(struct bch_fs *, enum btree_id,
 				unsigned, struct bpos, struct bpos, size_t *);
 struct bkey_i *bch2_journal_keys_peek_slot(struct bch_fs *, enum btree_id,
 					   unsigned, struct bpos);
@@ -79,8 +89,6 @@ static inline void bch2_journal_keys_put_initial(struct bch_fs *c)
 	c->journal_keys.initial_ref_held = false;
 }
 
-void bch2_journal_entries_free(struct bch_fs *);
-
 int bch2_journal_keys_sort(struct bch_fs *);
 
 void bch2_shoot_down_journal_keys(struct bch_fs *, enum btree_id,
@@ -89,4 +97,6 @@ void bch2_shoot_down_journal_keys(struct bch_fs *, enum btree_id,
 
 void bch2_journal_keys_dump(struct bch_fs *);
 
+void bch2_fs_journal_keys_init(struct bch_fs *);
+
 #endif /* _BCACHEFS_BTREE_JOURNAL_ITER_H */
diff --git a/fs/bcachefs/btree_journal_iter_types.h b/fs/bcachefs/btree_journal_iter_types.h
new file mode 100644
index 000000000000..8b773823704f
--- /dev/null
+++ b/fs/bcachefs/btree_journal_iter_types.h
@@ -0,0 +1,36 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_BTREE_JOURNAL_ITER_TYPES_H
+#define _BCACHEFS_BTREE_JOURNAL_ITER_TYPES_H
+
+struct journal_key_range_overwritten {
+	size_t			start, end;
+};
+
+struct journal_key {
+	u64			journal_seq;
+	u32			journal_offset;
+	enum btree_id		btree_id:8;
+	unsigned		level:8;
+	bool			allocated;
+	bool			overwritten;
+	struct journal_key_range_overwritten __rcu *
+				overwritten_range;
+	struct bkey_i		*k;
+};
+
+struct journal_keys {
+	/* must match layout in darray_types.h */
+	size_t			nr, size;
+	struct journal_key	*data;
+	/*
+	 * Gap buffer: instead of all the empty space in the array being at the
+	 * end of the buffer - from @nr to @size - the empty space is at @gap.
+	 * This means that sequential insertions are O(n) instead of O(n^2).
+	 */
+	size_t			gap;
+	atomic_t		ref;
+	bool			initial_ref_held;
+	struct mutex		overwrite_lock;
+};
+
+#endif /* _BCACHEFS_BTREE_JOURNAL_ITER_TYPES_H */
diff --git a/fs/bcachefs/btree_key_cache.c b/fs/bcachefs/btree_key_cache.c
index 244610b1d0b5..3b62296c3100 100644
--- a/fs/bcachefs/btree_key_cache.c
+++ b/fs/bcachefs/btree_key_cache.c
@@ -197,7 +197,9 @@ out:
 	return ck;
 }
 
-static int btree_key_cache_create(struct btree_trans *trans, struct btree_path *path,
+static int btree_key_cache_create(struct btree_trans *trans,
+				  struct btree_path *path,
+				  struct btree_path *ck_path,
 				  struct bkey_s_c k)
 {
 	struct bch_fs *c = trans->c;
@@ -217,7 +219,7 @@ static int btree_key_cache_create(struct btree_trans *trans, struct btree_path *
 	key_u64s = min(256U, (key_u64s * 3) / 2);
 	key_u64s = roundup_pow_of_two(key_u64s);
 
-	struct bkey_cached *ck = bkey_cached_alloc(trans, path, key_u64s);
+	struct bkey_cached *ck = bkey_cached_alloc(trans, ck_path, key_u64s);
 	int ret = PTR_ERR_OR_ZERO(ck);
 	if (ret)
 		return ret;
@@ -226,19 +228,19 @@ static int btree_key_cache_create(struct btree_trans *trans, struct btree_path *
 		ck = bkey_cached_reuse(bc);
 		if (unlikely(!ck)) {
 			bch_err(c, "error allocating memory for key cache item, btree %s",
-				bch2_btree_id_str(path->btree_id));
+				bch2_btree_id_str(ck_path->btree_id));
 			return -BCH_ERR_ENOMEM_btree_key_cache_create;
 		}
 	}
 
 	ck->c.level		= 0;
-	ck->c.btree_id		= path->btree_id;
-	ck->key.btree_id	= path->btree_id;
-	ck->key.pos		= path->pos;
+	ck->c.btree_id		= ck_path->btree_id;
+	ck->key.btree_id	= ck_path->btree_id;
+	ck->key.pos		= ck_path->pos;
 	ck->flags		= 1U << BKEY_CACHED_ACCESSED;
 
 	if (unlikely(key_u64s > ck->u64s)) {
-		mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
+		mark_btree_node_locked_noreset(ck_path, 0, BTREE_NODE_UNLOCKED);
 
 		struct bkey_i *new_k = allocate_dropping_locks(trans, ret,
 				kmalloc(key_u64s * sizeof(u64), _gfp));
@@ -258,22 +260,29 @@ static int btree_key_cache_create(struct btree_trans *trans, struct btree_path *
 
 	bkey_reassemble(ck->k, k);
 
+	ret = bch2_btree_node_lock_write(trans, path, &path_l(path)->b->c);
+	if (unlikely(ret))
+		goto err;
+
 	ret = rhashtable_lookup_insert_fast(&bc->table, &ck->hash, bch2_btree_key_cache_params);
+
+	bch2_btree_node_unlock_write(trans, path, path_l(path)->b);
+
 	if (unlikely(ret)) /* raced with another fill? */
 		goto err;
 
 	atomic_long_inc(&bc->nr_keys);
 	six_unlock_write(&ck->c.lock);
 
-	enum six_lock_type lock_want = __btree_lock_want(path, 0);
+	enum six_lock_type lock_want = __btree_lock_want(ck_path, 0);
 	if (lock_want == SIX_LOCK_read)
 		six_lock_downgrade(&ck->c.lock);
-	btree_path_cached_set(trans, path, ck, (enum btree_node_locked_type) lock_want);
-	path->uptodate = BTREE_ITER_UPTODATE;
+	btree_path_cached_set(trans, ck_path, ck, (enum btree_node_locked_type) lock_want);
+	ck_path->uptodate = BTREE_ITER_UPTODATE;
 	return 0;
 err:
 	bkey_cached_free(bc, ck);
-	mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
+	mark_btree_node_locked_noreset(ck_path, 0, BTREE_NODE_UNLOCKED);
 
 	return ret;
 }
@@ -282,10 +291,8 @@ static noinline int btree_key_cache_fill(struct btree_trans *trans,
 					 struct btree_path *ck_path,
 					 unsigned flags)
 {
-	if (flags & BTREE_ITER_cached_nofill) {
-		ck_path->uptodate = BTREE_ITER_UPTODATE;
+	if (flags & BTREE_ITER_cached_nofill)
 		return 0;
-	}
 
 	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
@@ -293,6 +300,7 @@ static noinline int btree_key_cache_fill(struct btree_trans *trans,
 	int ret;
 
 	bch2_trans_iter_init(trans, &iter, ck_path->btree_id, ck_path->pos,
+			     BTREE_ITER_intent|
 			     BTREE_ITER_key_cache_fill|
 			     BTREE_ITER_cached_nofill);
 	iter.flags &= ~BTREE_ITER_with_journal;
@@ -306,9 +314,19 @@ static noinline int btree_key_cache_fill(struct btree_trans *trans,
 	if (unlikely(ret))
 		goto out;
 
-	ret = btree_key_cache_create(trans, ck_path, k);
+	ret = btree_key_cache_create(trans, btree_iter_path(trans, &iter), ck_path, k);
 	if (ret)
 		goto err;
+
+	if (trace_key_cache_fill_enabled()) {
+		struct printbuf buf = PRINTBUF;
+
+		bch2_bpos_to_text(&buf, ck_path->pos);
+		prt_char(&buf, ' ');
+		bch2_bkey_val_to_text(&buf, trans->c, k);
+		trace_key_cache_fill(trans, buf.buf);
+		printbuf_exit(&buf);
+	}
 out:
 	/* We're not likely to need this iterator again: */
 	bch2_set_btree_iter_dontneed(&iter);
@@ -424,8 +442,15 @@ static int btree_key_cache_flush_pos(struct btree_trans *trans,
 	    !test_bit(JOURNAL_space_low, &c->journal.flags))
 		commit_flags |= BCH_TRANS_COMMIT_no_journal_res;
 
-	ret   = bch2_btree_iter_traverse(&b_iter) ?:
-		bch2_trans_update(trans, &b_iter, ck->k,
+	struct bkey_s_c btree_k = bch2_btree_iter_peek_slot(&b_iter);
+	ret = bkey_err(btree_k);
+	if (ret)
+		goto err;
+
+	/* * Check that we're not violating cache coherency rules: */
+	BUG_ON(bkey_deleted(btree_k.k));
+
+	ret   = bch2_trans_update(trans, &b_iter, ck->k,
 				  BTREE_UPDATE_key_cache_reclaim|
 				  BTREE_UPDATE_internal_snapshot_node|
 				  BTREE_TRIGGER_norun) ?:
@@ -433,7 +458,7 @@ static int btree_key_cache_flush_pos(struct btree_trans *trans,
 				  BCH_TRANS_COMMIT_no_check_rw|
 				  BCH_TRANS_COMMIT_no_enospc|
 				  commit_flags);
-
+err:
 	bch2_fs_fatal_err_on(ret &&
 			     !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 			     !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
@@ -586,8 +611,18 @@ void bch2_btree_key_cache_drop(struct btree_trans *trans,
 	bkey_cached_free(bc, ck);
 
 	mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
-	btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
-	path->should_be_locked = false;
+
+	struct btree_path *path2;
+	unsigned i;
+	trans_for_each_path(trans, path2, i)
+		if (path2->l[0].b == (void *) ck) {
+			__bch2_btree_path_unlock(trans, path2);
+			path2->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_drop);
+			path2->should_be_locked = false;
+			btree_path_set_dirty(path2, BTREE_ITER_NEED_TRAVERSE);
+		}
+
+	bch2_trans_verify_locks(trans);
 }
 
 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
diff --git a/fs/bcachefs/btree_locking.c b/fs/bcachefs/btree_locking.c
index efe2a007b482..10b805a60f52 100644
--- a/fs/bcachefs/btree_locking.c
+++ b/fs/bcachefs/btree_locking.c
@@ -109,6 +109,12 @@ static noinline void lock_graph_pop_all(struct lock_graph *g)
 		lock_graph_up(g);
 }
 
+static noinline void lock_graph_pop_from(struct lock_graph *g, struct trans_waiting_for_lock *i)
+{
+	while (g->g + g->nr > i)
+		lock_graph_up(g);
+}
+
 static void __lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
 {
 	g->g[g->nr++] = (struct trans_waiting_for_lock) {
@@ -124,15 +130,20 @@ static void lock_graph_down(struct lock_graph *g, struct btree_trans *trans)
 	__lock_graph_down(g, trans);
 }
 
-static bool lock_graph_remove_non_waiters(struct lock_graph *g)
+static bool lock_graph_remove_non_waiters(struct lock_graph *g,
+					  struct trans_waiting_for_lock *from)
 {
 	struct trans_waiting_for_lock *i;
 
-	for (i = g->g + 1; i < g->g + g->nr; i++)
+	if (from->trans->locking != from->node_want) {
+		lock_graph_pop_from(g, from);
+		return true;
+	}
+
+	for (i = from + 1; i < g->g + g->nr; i++)
 		if (i->trans->locking != i->node_want ||
 		    i->trans->locking_wait.start_time != i[-1].lock_start_time) {
-			while (g->g + g->nr > i)
-				lock_graph_up(g);
+			lock_graph_pop_from(g, i);
 			return true;
 		}
 
@@ -179,13 +190,14 @@ static int btree_trans_abort_preference(struct btree_trans *trans)
 	return 3;
 }
 
-static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
+static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle,
+				struct trans_waiting_for_lock *from)
 {
 	struct trans_waiting_for_lock *i, *abort = NULL;
 	unsigned best = 0, pref;
 	int ret;
 
-	if (lock_graph_remove_non_waiters(g))
+	if (lock_graph_remove_non_waiters(g, from))
 		return 0;
 
 	/* Only checking, for debugfs: */
@@ -195,7 +207,7 @@ static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
 		goto out;
 	}
 
-	for (i = g->g; i < g->g + g->nr; i++) {
+	for (i = from; i < g->g + g->nr; i++) {
 		pref = btree_trans_abort_preference(i->trans);
 		if (pref > best) {
 			abort = i;
@@ -229,8 +241,9 @@ static noinline int break_cycle(struct lock_graph *g, struct printbuf *cycle)
 	ret = abort_lock(g, abort);
 out:
 	if (ret)
-		while (g->nr)
-			lock_graph_up(g);
+		lock_graph_pop_all(g);
+	else
+		lock_graph_pop_from(g, abort);
 	return ret;
 }
 
@@ -243,7 +256,7 @@ static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
 	for (i = g->g; i < g->g + g->nr; i++)
 		if (i->trans == trans) {
 			closure_put(&trans->ref);
-			return break_cycle(g, cycle);
+			return break_cycle(g, cycle, i);
 		}
 
 	if (g->nr == ARRAY_SIZE(g->g)) {
@@ -252,8 +265,7 @@ static int lock_graph_descend(struct lock_graph *g, struct btree_trans *trans,
 		if (orig_trans->lock_may_not_fail)
 			return 0;
 
-		while (g->nr)
-			lock_graph_up(g);
+		lock_graph_pop_all(g);
 
 		if (cycle)
 			return 0;
@@ -281,7 +293,7 @@ int bch2_check_for_deadlock(struct btree_trans *trans, struct printbuf *cycle)
 
 	g.nr = 0;
 
-	if (trans->lock_must_abort) {
+	if (trans->lock_must_abort && !trans->lock_may_not_fail) {
 		if (cycle)
 			return -1;
 
@@ -336,7 +348,7 @@ next:
 				 * structures - which means it can't be blocked
 				 * waiting on a lock:
 				 */
-				if (!lock_graph_remove_non_waiters(&g)) {
+				if (!lock_graph_remove_non_waiters(&g, g.g)) {
 					/*
 					 * If lock_graph_remove_non_waiters()
 					 * didn't do anything, it must be
@@ -512,7 +524,6 @@ bool bch2_btree_node_upgrade(struct btree_trans *trans,
 			     struct btree_path *path, unsigned level)
 {
 	struct btree *b = path->l[level].b;
-	struct six_lock_count count = bch2_btree_node_lock_counts(trans, path, &b->c, level);
 
 	if (!is_btree_node(path, level))
 		return false;
@@ -536,24 +547,11 @@ bool bch2_btree_node_upgrade(struct btree_trans *trans,
 	if (race_fault())
 		return false;
 
-	if (btree_node_locked(path, level)) {
-		bool ret;
-
-		six_lock_readers_add(&b->c.lock, -count.n[SIX_LOCK_read]);
-		ret = six_lock_tryupgrade(&b->c.lock);
-		six_lock_readers_add(&b->c.lock, count.n[SIX_LOCK_read]);
-
-		if (ret)
-			goto success;
-	} else {
-		if (six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
-			goto success;
-	}
+	if (btree_node_locked(path, level)
+	    ? six_lock_tryupgrade(&b->c.lock)
+	    : six_relock_type(&b->c.lock, SIX_LOCK_intent, path->l[level].lock_seq))
+		goto success;
 
-	/*
-	 * Do we already have an intent lock via another path? If so, just bump
-	 * lock count:
-	 */
 	if (btree_node_lock_seq_matches(path, b, level) &&
 	    btree_node_lock_increment(trans, &b->c, level, BTREE_NODE_INTENT_LOCKED)) {
 		btree_node_unlock(trans, path, level);
@@ -782,7 +780,7 @@ static inline int __bch2_trans_relock(struct btree_trans *trans, bool trace)
 			return bch2_trans_relock_fail(trans, path, &f, trace);
 	}
 
-	trans_set_locked(trans);
+	trans_set_locked(trans, true);
 out:
 	bch2_trans_verify_locks(trans);
 	return 0;
@@ -818,6 +816,17 @@ void bch2_trans_unlock_long(struct btree_trans *trans)
 	bch2_trans_srcu_unlock(trans);
 }
 
+void bch2_trans_unlock_write(struct btree_trans *trans)
+{
+	struct btree_path *path;
+	unsigned i;
+
+	trans_for_each_path(trans, path, i)
+		for (unsigned l = 0; l < BTREE_MAX_DEPTH; l++)
+			if (btree_node_write_locked(path, l))
+				bch2_btree_node_unlock_write(trans, path, path->l[l].b);
+}
+
 int __bch2_trans_mutex_lock(struct btree_trans *trans,
 			    struct mutex *lock)
 {
@@ -856,6 +865,9 @@ void bch2_btree_path_verify_locks(struct btree_path *path)
 		       (want == BTREE_NODE_UNLOCKED ||
 			have != BTREE_NODE_WRITE_LOCKED) &&
 		       want != have);
+
+		BUG_ON(btree_node_locked(path, l) &&
+		       path->l[l].lock_seq != six_lock_seq(&path->l[l].b->c.lock));
 	}
 }
 
diff --git a/fs/bcachefs/btree_locking.h b/fs/bcachefs/btree_locking.h
index 7c07f9fa9add..b54ef48eb8cc 100644
--- a/fs/bcachefs/btree_locking.h
+++ b/fs/bcachefs/btree_locking.h
@@ -16,6 +16,7 @@
 void bch2_btree_lock_init(struct btree_bkey_cached_common *, enum six_lock_init_flags);
 
 void bch2_trans_unlock_noassert(struct btree_trans *);
+void bch2_trans_unlock_write(struct btree_trans *);
 
 static inline bool is_btree_node(struct btree_path *path, unsigned l)
 {
@@ -75,13 +76,6 @@ static inline void mark_btree_node_locked_noreset(struct btree_path *path,
 	path->nodes_locked |= (type + 1) << (level << 1);
 }
 
-static inline void mark_btree_node_unlocked(struct btree_path *path,
-					    unsigned level)
-{
-	EBUG_ON(btree_node_write_locked(path, level));
-	mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
-}
-
 static inline void mark_btree_node_locked(struct btree_trans *trans,
 					  struct btree_path *path,
 					  unsigned level,
@@ -124,19 +118,25 @@ static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
 
 /* unlock: */
 
+void bch2_btree_node_unlock_write(struct btree_trans *,
+			struct btree_path *, struct btree *);
+
 static inline void btree_node_unlock(struct btree_trans *trans,
 				     struct btree_path *path, unsigned level)
 {
 	int lock_type = btree_node_locked_type(path, level);
 
 	EBUG_ON(level >= BTREE_MAX_DEPTH);
-	EBUG_ON(lock_type == BTREE_NODE_WRITE_LOCKED);
 
 	if (lock_type != BTREE_NODE_UNLOCKED) {
+		if (unlikely(lock_type == BTREE_NODE_WRITE_LOCKED)) {
+			bch2_btree_node_unlock_write(trans, path, path->l[level].b);
+			lock_type = BTREE_NODE_INTENT_LOCKED;
+		}
 		six_unlock_type(&path->l[level].b->c.lock, lock_type);
 		btree_trans_lock_hold_time_update(trans, path, level);
+		mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
 	}
-	mark_btree_node_unlocked(path, level);
 }
 
 static inline int btree_path_lowest_level_locked(struct btree_path *path)
@@ -163,35 +163,39 @@ static inline void __bch2_btree_path_unlock(struct btree_trans *trans,
  * succeed:
  */
 static inline void
+__bch2_btree_node_unlock_write(struct btree_trans *trans, struct btree *b)
+{
+	if (!b->c.lock.write_lock_recurse) {
+		struct btree_path *linked;
+		unsigned i;
+
+		trans_for_each_path_with_node(trans, b, linked, i)
+			linked->l[b->c.level].lock_seq++;
+	}
+
+	six_unlock_write(&b->c.lock);
+}
+
+static inline void
 bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
 				     struct btree *b)
 {
-	struct btree_path *linked;
-	unsigned i;
-
 	EBUG_ON(path->l[b->c.level].b != b);
 	EBUG_ON(path->l[b->c.level].lock_seq != six_lock_seq(&b->c.lock));
 	EBUG_ON(btree_node_locked_type(path, b->c.level) != SIX_LOCK_write);
 
 	mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
-
-	trans_for_each_path_with_node(trans, b, linked, i)
-		linked->l[b->c.level].lock_seq++;
-
-	six_unlock_write(&b->c.lock);
+	__bch2_btree_node_unlock_write(trans, b);
 }
 
-void bch2_btree_node_unlock_write(struct btree_trans *,
-			struct btree_path *, struct btree *);
-
 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p);
 
 /* lock: */
 
-static inline void trans_set_locked(struct btree_trans *trans)
+static inline void trans_set_locked(struct btree_trans *trans, bool try)
 {
 	if (!trans->locked) {
-		lock_acquire_exclusive(&trans->dep_map, 0, 0, NULL, _THIS_IP_);
+		lock_acquire_exclusive(&trans->dep_map, 0, try, NULL, _THIS_IP_);
 		trans->locked = true;
 		trans->last_unlock_ip = 0;
 
@@ -282,7 +286,7 @@ static inline int btree_node_lock(struct btree_trans *trans,
 	int ret = 0;
 
 	EBUG_ON(level >= BTREE_MAX_DEPTH);
-	bch2_trans_verify_not_unlocked(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 
 	if (likely(six_trylock_type(&b->lock, type)) ||
 	    btree_node_lock_increment(trans, b, level, (enum btree_node_locked_type) type) ||
diff --git a/fs/bcachefs/btree_node_scan.c b/fs/bcachefs/btree_node_scan.c
index 30131c3bdd97..a7f06deee13c 100644
--- a/fs/bcachefs/btree_node_scan.c
+++ b/fs/bcachefs/btree_node_scan.c
@@ -12,6 +12,7 @@
 #include "recovery_passes.h"
 
 #include <linux/kthread.h>
+#include <linux/min_heap.h>
 #include <linux/sort.h>
 
 struct find_btree_nodes_worker {
@@ -22,17 +23,15 @@ struct find_btree_nodes_worker {
 
 static void found_btree_node_to_text(struct printbuf *out, struct bch_fs *c, const struct found_btree_node *n)
 {
-	prt_printf(out, "%s l=%u seq=%u journal_seq=%llu cookie=%llx ",
-		   bch2_btree_id_str(n->btree_id), n->level, n->seq,
-		   n->journal_seq, n->cookie);
+	bch2_btree_id_level_to_text(out, n->btree_id, n->level);
+	prt_printf(out, " seq=%u journal_seq=%llu cookie=%llx ",
+		   n->seq, n->journal_seq, n->cookie);
 	bch2_bpos_to_text(out, n->min_key);
 	prt_str(out, "-");
 	bch2_bpos_to_text(out, n->max_key);
 
 	if (n->range_updated)
 		prt_str(out, " range updated");
-	if (n->overwritten)
-		prt_str(out, " overwritten");
 
 	for (unsigned i = 0; i < n->nr_ptrs; i++) {
 		prt_char(out, ' ');
@@ -140,6 +139,24 @@ static int found_btree_node_cmp_pos(const void *_l, const void *_r)
 	       -found_btree_node_cmp_time(l, r);
 }
 
+static inline bool found_btree_node_cmp_pos_less(const void *l, const void *r, void *arg)
+{
+	return found_btree_node_cmp_pos(l, r) < 0;
+}
+
+static inline void found_btree_node_swap(void *_l, void *_r, void *arg)
+{
+	struct found_btree_node *l = _l;
+	struct found_btree_node *r = _r;
+
+	swap(*l, *r);
+}
+
+static const struct min_heap_callbacks found_btree_node_heap_cbs = {
+	.less = found_btree_node_cmp_pos_less,
+	.swp = found_btree_node_swap,
+};
+
 static void try_read_btree_node(struct find_btree_nodes *f, struct bch_dev *ca,
 				struct bio *bio, struct btree_node *bn, u64 offset)
 {
@@ -159,6 +176,9 @@ static void try_read_btree_node(struct find_btree_nodes *f, struct bch_dev *ca,
 		return;
 
 	if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(&bn->keys))) {
+		if (!c->chacha20)
+			return;
+
 		struct nonce nonce = btree_nonce(&bn->keys, 0);
 		unsigned bytes = (void *) &bn->keys - (void *) &bn->flags;
 
@@ -292,55 +312,48 @@ err:
 	return f->ret ?: ret;
 }
 
-static void bubble_up(struct found_btree_node *n, struct found_btree_node *end)
+static bool nodes_overlap(const struct found_btree_node *l,
+			  const struct found_btree_node *r)
 {
-	while (n + 1 < end &&
-	       found_btree_node_cmp_pos(n, n + 1) > 0) {
-		swap(n[0], n[1]);
-		n++;
-	}
+	return (l->btree_id	== r->btree_id &&
+		l->level	== r->level &&
+		bpos_gt(l->max_key, r->min_key));
 }
 
 static int handle_overwrites(struct bch_fs *c,
-			     struct found_btree_node *start,
-			     struct found_btree_node *end)
+			     struct found_btree_node *l,
+			     found_btree_nodes *nodes_heap)
 {
-	struct found_btree_node *n;
-again:
-	for (n = start + 1;
-	     n < end &&
-	     n->btree_id	== start->btree_id &&
-	     n->level		== start->level &&
-	     bpos_lt(n->min_key, start->max_key);
-	     n++)  {
-		int cmp = found_btree_node_cmp_time(start, n);
+	struct found_btree_node *r;
+
+	while ((r = min_heap_peek(nodes_heap)) &&
+	       nodes_overlap(l, r)) {
+		int cmp = found_btree_node_cmp_time(l, r);
 
 		if (cmp > 0) {
-			if (bpos_cmp(start->max_key, n->max_key) >= 0)
-				n->overwritten = true;
+			if (bpos_cmp(l->max_key, r->max_key) >= 0)
+				min_heap_pop(nodes_heap, &found_btree_node_heap_cbs, NULL);
 			else {
-				n->range_updated = true;
-				n->min_key = bpos_successor(start->max_key);
-				n->range_updated = true;
-				bubble_up(n, end);
-				goto again;
+				r->range_updated = true;
+				r->min_key = bpos_successor(l->max_key);
+				r->range_updated = true;
+				min_heap_sift_down(nodes_heap, 0, &found_btree_node_heap_cbs, NULL);
 			}
 		} else if (cmp < 0) {
-			BUG_ON(bpos_cmp(n->min_key, start->min_key) <= 0);
+			BUG_ON(bpos_eq(l->min_key, r->min_key));
 
-			start->max_key = bpos_predecessor(n->min_key);
-			start->range_updated = true;
-		} else if (n->level) {
-			n->overwritten = true;
+			l->max_key = bpos_predecessor(r->min_key);
+			l->range_updated = true;
+		} else if (r->level) {
+			min_heap_pop(nodes_heap, &found_btree_node_heap_cbs, NULL);
 		} else {
-			if (bpos_cmp(start->max_key, n->max_key) >= 0)
-				n->overwritten = true;
+			if (bpos_cmp(l->max_key, r->max_key) >= 0)
+				min_heap_pop(nodes_heap, &found_btree_node_heap_cbs, NULL);
 			else {
-				n->range_updated = true;
-				n->min_key = bpos_successor(start->max_key);
-				n->range_updated = true;
-				bubble_up(n, end);
-				goto again;
+				r->range_updated = true;
+				r->min_key = bpos_successor(l->max_key);
+				r->range_updated = true;
+				min_heap_sift_down(nodes_heap, 0, &found_btree_node_heap_cbs, NULL);
 			}
 		}
 	}
@@ -352,6 +365,7 @@ int bch2_scan_for_btree_nodes(struct bch_fs *c)
 {
 	struct find_btree_nodes *f = &c->found_btree_nodes;
 	struct printbuf buf = PRINTBUF;
+	found_btree_nodes nodes_heap = {};
 	size_t dst;
 	int ret = 0;
 
@@ -406,29 +420,57 @@ int bch2_scan_for_btree_nodes(struct bch_fs *c)
 		bch2_print_string_as_lines(KERN_INFO, buf.buf);
 	}
 
-	dst = 0;
-	darray_for_each(f->nodes, i) {
-		if (i->overwritten)
-			continue;
+	swap(nodes_heap, f->nodes);
+
+	{
+		/* darray must have same layout as a heap */
+		min_heap_char real_heap;
+		BUILD_BUG_ON(sizeof(nodes_heap.nr)	!= sizeof(real_heap.nr));
+		BUILD_BUG_ON(sizeof(nodes_heap.size)	!= sizeof(real_heap.size));
+		BUILD_BUG_ON(offsetof(found_btree_nodes, nr)	!= offsetof(min_heap_char, nr));
+		BUILD_BUG_ON(offsetof(found_btree_nodes, size)	!= offsetof(min_heap_char, size));
+	}
+
+	min_heapify_all(&nodes_heap, &found_btree_node_heap_cbs, NULL);
 
-		ret = handle_overwrites(c, i, &darray_top(f->nodes));
+	if (nodes_heap.nr) {
+		ret = darray_push(&f->nodes, *min_heap_peek(&nodes_heap));
 		if (ret)
 			goto err;
 
-		BUG_ON(i->overwritten);
-		f->nodes.data[dst++] = *i;
+		min_heap_pop(&nodes_heap, &found_btree_node_heap_cbs, NULL);
 	}
-	f->nodes.nr = dst;
 
-	if (c->opts.verbose) {
+	while (true) {
+		ret = handle_overwrites(c, &darray_last(f->nodes), &nodes_heap);
+		if (ret)
+			goto err;
+
+		if (!nodes_heap.nr)
+			break;
+
+		ret = darray_push(&f->nodes, *min_heap_peek(&nodes_heap));
+		if (ret)
+			goto err;
+
+		min_heap_pop(&nodes_heap, &found_btree_node_heap_cbs, NULL);
+	}
+
+	for (struct found_btree_node *n = f->nodes.data; n < &darray_last(f->nodes); n++)
+		BUG_ON(nodes_overlap(n, n + 1));
+
+	if (0 && c->opts.verbose) {
 		printbuf_reset(&buf);
 		prt_printf(&buf, "%s: nodes found after overwrites:\n", __func__);
 		found_btree_nodes_to_text(&buf, c, f->nodes);
 		bch2_print_string_as_lines(KERN_INFO, buf.buf);
+	} else {
+		bch_info(c, "btree node scan found %zu nodes after overwrites", f->nodes.nr);
 	}
 
 	eytzinger0_sort(f->nodes.data, f->nodes.nr, sizeof(f->nodes.data[0]), found_btree_node_cmp_pos, NULL);
 err:
+	darray_exit(&nodes_heap);
 	printbuf_exit(&buf);
 	return ret;
 }
@@ -499,7 +541,9 @@ int bch2_get_scanned_nodes(struct bch_fs *c, enum btree_id btree,
 	if (c->opts.verbose) {
 		struct printbuf buf = PRINTBUF;
 
-		prt_printf(&buf, "recovering %s l=%u ", bch2_btree_id_str(btree), level);
+		prt_str(&buf, "recovery ");
+		bch2_btree_id_level_to_text(&buf, btree, level);
+		prt_str(&buf, " ");
 		bch2_bpos_to_text(&buf, node_min);
 		prt_str(&buf, " - ");
 		bch2_bpos_to_text(&buf, node_max);
@@ -533,7 +577,12 @@ int bch2_get_scanned_nodes(struct bch_fs *c, enum btree_id btree,
 		bch_verbose(c, "%s(): recovering %s", __func__, buf.buf);
 		printbuf_exit(&buf);
 
-		BUG_ON(bch2_bkey_validate(c, bkey_i_to_s_c(&tmp.k), BKEY_TYPE_btree, 0));
+		BUG_ON(bch2_bkey_validate(c, bkey_i_to_s_c(&tmp.k),
+					  (struct bkey_validate_context) {
+						.from	= BKEY_VALIDATE_btree_node,
+						.level	= level + 1,
+						.btree	= btree,
+					  }));
 
 		ret = bch2_journal_key_insert(c, btree, level + 1, &tmp.k);
 		if (ret)
diff --git a/fs/bcachefs/btree_node_scan_types.h b/fs/bcachefs/btree_node_scan_types.h
index b6c36c45d0be..2811b6857c97 100644
--- a/fs/bcachefs/btree_node_scan_types.h
+++ b/fs/bcachefs/btree_node_scan_types.h
@@ -6,7 +6,6 @@
 
 struct found_btree_node {
 	bool			range_updated:1;
-	bool			overwritten:1;
 	u8			btree_id;
 	u8			level;
 	unsigned		sectors_written;
diff --git a/fs/bcachefs/btree_trans_commit.c b/fs/bcachefs/btree_trans_commit.c
index 9bf471fa4361..6b79b672e0b1 100644
--- a/fs/bcachefs/btree_trans_commit.c
+++ b/fs/bcachefs/btree_trans_commit.c
@@ -133,7 +133,7 @@ static inline int bch2_trans_lock_write(struct btree_trans *trans)
 	return 0;
 }
 
-static inline void bch2_trans_unlock_write(struct btree_trans *trans)
+static inline void bch2_trans_unlock_updates_write(struct btree_trans *trans)
 {
 	if (likely(trans->write_locked)) {
 		trans_for_each_update(trans, i)
@@ -249,7 +249,7 @@ static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 		new |= 1 << BTREE_NODE_need_write;
 	} while (!try_cmpxchg(&b->flags, &old, new));
 
-	btree_node_write_if_need(c, b, SIX_LOCK_read);
+	btree_node_write_if_need(trans, b, SIX_LOCK_read);
 	six_unlock_read(&b->c.lock);
 
 	bch2_trans_put(trans);
@@ -384,7 +384,7 @@ btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
 	struct bkey_i *new_k;
 	int ret;
 
-	bch2_trans_unlock_write(trans);
+	bch2_trans_unlock_updates_write(trans);
 	bch2_trans_unlock(trans);
 
 	new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
@@ -479,8 +479,7 @@ static int run_one_mem_trigger(struct btree_trans *trans,
 				old, flags);
 }
 
-static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
-				 bool overwrite)
+static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i)
 {
 	verify_update_old_key(trans, i);
 
@@ -507,10 +506,10 @@ static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_
 		return bch2_key_trigger(trans, i->btree_id, i->level, old, bkey_i_to_s(i->k),
 					BTREE_TRIGGER_insert|
 					BTREE_TRIGGER_overwrite|flags) ?: 1;
-	} else if (overwrite && !i->overwrite_trigger_run) {
+	} else if (!i->overwrite_trigger_run) {
 		i->overwrite_trigger_run = true;
 		return bch2_key_trigger_old(trans, i->btree_id, i->level, old, flags) ?: 1;
-	} else if (!overwrite && !i->insert_trigger_run) {
+	} else if (!i->insert_trigger_run) {
 		i->insert_trigger_run = true;
 		return bch2_key_trigger_new(trans, i->btree_id, i->level, bkey_i_to_s(i->k), flags) ?: 1;
 	} else {
@@ -519,39 +518,45 @@ static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_
 }
 
 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
-			      unsigned btree_id_start)
+			      unsigned *btree_id_updates_start)
 {
-	for (int overwrite = 1; overwrite >= 0; --overwrite) {
-		bool trans_trigger_run;
+	bool trans_trigger_run;
 
-		/*
-		 * Running triggers will append more updates to the list of updates as
-		 * we're walking it:
-		 */
-		do {
-			trans_trigger_run = false;
-
-			for (unsigned i = btree_id_start;
-			     i < trans->nr_updates && trans->updates[i].btree_id <= btree_id;
-			     i++) {
-				if (trans->updates[i].btree_id != btree_id)
-					continue;
+	/*
+	 * Running triggers will append more updates to the list of updates as
+	 * we're walking it:
+	 */
+	do {
+		trans_trigger_run = false;
 
-				int ret = run_one_trans_trigger(trans, trans->updates + i, overwrite);
-				if (ret < 0)
-					return ret;
-				if (ret)
-					trans_trigger_run = true;
+		for (unsigned i = *btree_id_updates_start;
+		     i < trans->nr_updates && trans->updates[i].btree_id <= btree_id;
+		     i++) {
+			if (trans->updates[i].btree_id < btree_id) {
+				*btree_id_updates_start = i;
+				continue;
 			}
-		} while (trans_trigger_run);
-	}
+
+			int ret = run_one_trans_trigger(trans, trans->updates + i);
+			if (ret < 0)
+				return ret;
+			if (ret)
+				trans_trigger_run = true;
+		}
+	} while (trans_trigger_run);
+
+	trans_for_each_update(trans, i)
+		BUG_ON(!(i->flags & BTREE_TRIGGER_norun) &&
+		       i->btree_id == btree_id &&
+		       btree_node_type_has_trans_triggers(i->bkey_type) &&
+		       (!i->insert_trigger_run || !i->overwrite_trigger_run));
 
 	return 0;
 }
 
 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 {
-	unsigned btree_id = 0, btree_id_start = 0;
+	unsigned btree_id = 0, btree_id_updates_start = 0;
 	int ret = 0;
 
 	/*
@@ -565,27 +570,15 @@ static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 		if (btree_id == BTREE_ID_alloc)
 			continue;
 
-		while (btree_id_start < trans->nr_updates &&
-		       trans->updates[btree_id_start].btree_id < btree_id)
-			btree_id_start++;
-
-		ret = run_btree_triggers(trans, btree_id, btree_id_start);
+		ret = run_btree_triggers(trans, btree_id, &btree_id_updates_start);
 		if (ret)
 			return ret;
 	}
 
-	for (unsigned idx = 0; idx < trans->nr_updates; idx++) {
-		struct btree_insert_entry *i = trans->updates + idx;
-
-		if (i->btree_id > BTREE_ID_alloc)
-			break;
-		if (i->btree_id == BTREE_ID_alloc) {
-			ret = run_btree_triggers(trans, BTREE_ID_alloc, idx);
-			if (ret)
-				return ret;
-			break;
-		}
-	}
+	btree_id_updates_start = 0;
+	ret = run_btree_triggers(trans, BTREE_ID_alloc, &btree_id_updates_start);
+	if (ret)
+		return ret;
 
 #ifdef CONFIG_BCACHEFS_DEBUG
 	trans_for_each_update(trans, i)
@@ -609,14 +602,6 @@ static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 	return 0;
 }
 
-static struct bversion journal_pos_to_bversion(struct journal_res *res, unsigned offset)
-{
-	return (struct bversion) {
-		.hi = res->seq >> 32,
-		.lo = (res->seq << 32) | (res->offset + offset),
-	};
-}
-
 static inline int
 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 			       struct btree_insert_entry **stopped_at,
@@ -627,12 +612,11 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 	unsigned u64s = 0;
 	int ret = 0;
 
-	bch2_trans_verify_not_unlocked(trans);
-	bch2_trans_verify_not_in_restart(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 
 	if (race_fault()) {
 		trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
-		return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
+		return btree_trans_restart(trans, BCH_ERR_transaction_restart_fault_inject);
 	}
 
 	/*
@@ -701,25 +685,14 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 	struct jset_entry *entry = trans->journal_entries;
 
 	percpu_down_read(&c->mark_lock);
-
 	for (entry = trans->journal_entries;
 	     entry != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
 	     entry = vstruct_next(entry))
 		if (entry->type == BCH_JSET_ENTRY_write_buffer_keys &&
 		    entry->start->k.type == KEY_TYPE_accounting) {
-			BUG_ON(!trans->journal_res.ref);
-
-			struct bkey_i_accounting *a = bkey_i_to_accounting(entry->start);
-
-			a->k.bversion = journal_pos_to_bversion(&trans->journal_res,
-							(u64 *) entry - (u64 *) trans->journal_entries);
-			BUG_ON(bversion_zero(a->k.bversion));
-
-			if (likely(!(flags & BCH_TRANS_COMMIT_skip_accounting_apply))) {
-				ret = bch2_accounting_mem_mod_locked(trans, accounting_i_to_s_c(a), BCH_ACCOUNTING_normal);
-				if (ret)
-					goto revert_fs_usage;
-			}
+			ret = bch2_accounting_trans_commit_hook(trans, bkey_i_to_accounting(entry->start), flags);
+			if (ret)
+				goto revert_fs_usage;
 		}
 	percpu_up_read(&c->mark_lock);
 
@@ -739,33 +712,17 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 			goto fatal_err;
 	}
 
-	trans_for_each_update(trans, i) {
-		enum bch_validate_flags invalid_flags = 0;
+	struct bkey_validate_context validate_context = { .from	= BKEY_VALIDATE_commit };
 
-		if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
-			invalid_flags |= BCH_VALIDATE_write|BCH_VALIDATE_commit;
-
-		ret = bch2_bkey_validate(c, bkey_i_to_s_c(i->k),
-					 i->bkey_type, invalid_flags);
-		if (unlikely(ret)){
-			bch2_trans_inconsistent(trans, "invalid bkey on insert from %s -> %ps\n",
-						trans->fn, (void *) i->ip_allocated);
-			goto fatal_err;
-		}
-		btree_insert_entry_checks(trans, i);
-	}
+	if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
+		validate_context.flags = BCH_VALIDATE_write|BCH_VALIDATE_commit;
 
 	for (struct jset_entry *i = trans->journal_entries;
 	     i != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
 	     i = vstruct_next(i)) {
-		enum bch_validate_flags invalid_flags = 0;
-
-		if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
-			invalid_flags |= BCH_VALIDATE_write|BCH_VALIDATE_commit;
-
 		ret = bch2_journal_entry_validate(c, NULL, i,
 						  bcachefs_metadata_version_current,
-						  CPU_BIG_ENDIAN, invalid_flags);
+						  CPU_BIG_ENDIAN, validate_context);
 		if (unlikely(ret)) {
 			bch2_trans_inconsistent(trans, "invalid journal entry on insert from %s\n",
 						trans->fn);
@@ -773,6 +730,19 @@ bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 		}
 	}
 
+	trans_for_each_update(trans, i) {
+		validate_context.level	= i->level;
+		validate_context.btree	= i->btree_id;
+
+		ret = bch2_bkey_validate(c, bkey_i_to_s_c(i->k), validate_context);
+		if (unlikely(ret)){
+			bch2_trans_inconsistent(trans, "invalid bkey on insert from %s -> %ps\n",
+						trans->fn, (void *) i->ip_allocated);
+			goto fatal_err;
+		}
+		btree_insert_entry_checks(trans, i);
+	}
+
 	if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 		struct journal *j = &c->journal;
 		struct jset_entry *entry;
@@ -833,13 +803,9 @@ revert_fs_usage:
 	     entry2 != entry;
 	     entry2 = vstruct_next(entry2))
 		if (entry2->type == BCH_JSET_ENTRY_write_buffer_keys &&
-		    entry2->start->k.type == KEY_TYPE_accounting) {
-			struct bkey_s_accounting a = bkey_i_to_s_accounting(entry2->start);
-
-			bch2_accounting_neg(a);
-			bch2_accounting_mem_mod_locked(trans, a.c, BCH_ACCOUNTING_normal);
-			bch2_accounting_neg(a);
-		}
+		    entry2->start->k.type == KEY_TYPE_accounting)
+			bch2_accounting_trans_commit_revert(trans,
+					bkey_i_to_accounting(entry2->start), flags);
 	percpu_up_read(&c->mark_lock);
 	return ret;
 }
@@ -902,7 +868,7 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags
 	if (!ret && unlikely(trans->journal_replay_not_finished))
 		bch2_drop_overwrites_from_journal(trans);
 
-	bch2_trans_unlock_write(trans);
+	bch2_trans_unlock_updates_write(trans);
 
 	if (!ret && trans->journal_pin)
 		bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
@@ -994,24 +960,6 @@ int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 	return ret;
 }
 
-static noinline int
-bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
-{
-	struct bch_fs *c = trans->c;
-	int ret;
-
-	if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
-	    test_bit(BCH_FS_started, &c->flags))
-		return -BCH_ERR_erofs_trans_commit;
-
-	ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
-	if (ret)
-		return ret;
-
-	bch2_write_ref_get(c, BCH_WRITE_REF_trans);
-	return 0;
-}
-
 /*
  * This is for updates done in the early part of fsck - btree_gc - before we've
  * gone RW. we only add the new key to the list of keys for journal replay to
@@ -1022,6 +970,8 @@ do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 {
 	struct bch_fs *c = trans->c;
 
+	BUG_ON(current != c->recovery_task);
+
 	trans_for_each_update(trans, i) {
 		int ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 		if (ret)
@@ -1047,8 +997,7 @@ int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 	struct bch_fs *c = trans->c;
 	int ret = 0;
 
-	bch2_trans_verify_not_unlocked(trans);
-	bch2_trans_verify_not_in_restart(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 
 	if (!trans->nr_updates &&
 	    !trans->journal_entries_u64s)
@@ -1058,16 +1007,13 @@ int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 	if (ret)
 		goto out_reset;
 
-	if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
-		ret = do_bch2_trans_commit_to_journal_replay(trans);
-		goto out_reset;
-	}
-
 	if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
 	    unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
-		ret = bch2_trans_commit_get_rw_cold(trans, flags);
-		if (ret)
-			goto out_reset;
+		if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags)))
+			ret = do_bch2_trans_commit_to_journal_replay(trans);
+		else
+			ret = -BCH_ERR_erofs_trans_commit;
+		goto out_reset;
 	}
 
 	EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
@@ -1112,8 +1058,7 @@ int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 	}
 retry:
 	errored_at = NULL;
-	bch2_trans_verify_not_unlocked(trans);
-	bch2_trans_verify_not_in_restart(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 	if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
 		memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 	memset(&trans->fs_usage_delta, 0, sizeof(trans->fs_usage_delta));
diff --git a/fs/bcachefs/btree_types.h b/fs/bcachefs/btree_types.h
index 4568a41fefaf..a6f251eb4164 100644
--- a/fs/bcachefs/btree_types.h
+++ b/fs/bcachefs/btree_types.h
@@ -513,6 +513,9 @@ struct btree_trans {
 	u64			last_begin_time;
 	unsigned long		last_begin_ip;
 	unsigned long		last_restarted_ip;
+#ifdef CONFIG_BCACHEFS_DEBUG
+	bch_stacktrace		last_restarted_trace;
+#endif
 	unsigned long		last_unlock_ip;
 	unsigned long		srcu_lock_time;
 
@@ -787,53 +790,64 @@ static inline bool btree_node_type_has_triggers(enum btree_node_type type)
 	return BIT_ULL(type) & BTREE_NODE_TYPE_HAS_TRIGGERS;
 }
 
-static inline bool btree_node_type_is_extents(enum btree_node_type type)
+static inline bool btree_id_is_extents(enum btree_id btree)
 {
 	const u64 mask = 0
-#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_ID_EXTENTS)) << (nr + 1))
+#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_IS_extents)) << nr)
 	BCH_BTREE_IDS()
 #undef x
 	;
 
-	return BIT_ULL(type) & mask;
+	return BIT_ULL(btree) & mask;
 }
 
-static inline bool btree_id_is_extents(enum btree_id btree)
+static inline bool btree_node_type_is_extents(enum btree_node_type type)
 {
-	return btree_node_type_is_extents(__btree_node_type(0, btree));
+	return type != BKEY_TYPE_btree && btree_id_is_extents(type - 1);
+}
+
+static inline bool btree_type_has_snapshots(enum btree_id btree)
+{
+	const u64 mask = 0
+#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_IS_snapshots)) << nr)
+	BCH_BTREE_IDS()
+#undef x
+	;
+
+	return BIT_ULL(btree) & mask;
 }
 
-static inline bool btree_type_has_snapshots(enum btree_id id)
+static inline bool btree_type_has_snapshot_field(enum btree_id btree)
 {
 	const u64 mask = 0
-#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_ID_SNAPSHOTS)) << nr)
+#define x(name, nr, flags, ...)	|((!!((flags) & (BTREE_IS_snapshot_field|BTREE_IS_snapshots))) << nr)
 	BCH_BTREE_IDS()
 #undef x
 	;
 
-	return BIT_ULL(id) & mask;
+	return BIT_ULL(btree) & mask;
 }
 
-static inline bool btree_type_has_snapshot_field(enum btree_id id)
+static inline bool btree_type_has_ptrs(enum btree_id btree)
 {
 	const u64 mask = 0
-#define x(name, nr, flags, ...)	|((!!((flags) & (BTREE_ID_SNAPSHOT_FIELD|BTREE_ID_SNAPSHOTS))) << nr)
+#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_IS_data)) << nr)
 	BCH_BTREE_IDS()
 #undef x
 	;
 
-	return BIT_ULL(id) & mask;
+	return BIT_ULL(btree) & mask;
 }
 
-static inline bool btree_type_has_ptrs(enum btree_id id)
+static inline bool btree_type_uses_write_buffer(enum btree_id btree)
 {
 	const u64 mask = 0
-#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_ID_DATA)) << nr)
+#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_IS_write_buffer)) << nr)
 	BCH_BTREE_IDS()
 #undef x
 	;
 
-	return BIT_ULL(id) & mask;
+	return BIT_ULL(btree) & mask;
 }
 
 struct btree_root {
diff --git a/fs/bcachefs/btree_update.c b/fs/bcachefs/btree_update.c
index 5d809e8bd170..13d794f201a5 100644
--- a/fs/bcachefs/btree_update.c
+++ b/fs/bcachefs/btree_update.c
@@ -144,7 +144,7 @@ int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
 	       !(ret = bkey_err(old_k)) &&
 	       bkey_eq(old_pos, old_k.k->p)) {
 		struct bpos whiteout_pos =
-			SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
+			SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);
 
 		if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
 		    snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
@@ -296,7 +296,7 @@ static int bch2_trans_update_extent(struct btree_trans *trans,
 			     BTREE_ITER_intent|
 			     BTREE_ITER_with_updates|
 			     BTREE_ITER_not_extents);
-	k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
+	k = bch2_btree_iter_peek_max(&iter, POS(insert->k.p.inode, U64_MAX));
 	if ((ret = bkey_err(k)))
 		goto err;
 	if (!k.k)
@@ -323,7 +323,7 @@ static int bch2_trans_update_extent(struct btree_trans *trans,
 			goto out;
 next:
 		bch2_btree_iter_advance(&iter);
-		k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
+		k = bch2_btree_iter_peek_max(&iter, POS(insert->k.p.inode, U64_MAX));
 		if ((ret = bkey_err(k)))
 			goto err;
 		if (!k.k)
@@ -588,12 +588,9 @@ struct jset_entry *__bch2_trans_jset_entry_alloc(struct btree_trans *trans, unsi
 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
 			     enum btree_id btree, struct bpos end)
 {
-	struct bkey_s_c k;
-	int ret = 0;
-
-	bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_intent);
-	k = bch2_btree_iter_prev(iter);
-	ret = bkey_err(k);
+	bch2_trans_iter_init(trans, iter, btree, end, BTREE_ITER_intent);
+	struct bkey_s_c k = bch2_btree_iter_peek_prev(iter);
+	int ret = bkey_err(k);
 	if (ret)
 		goto err;
 
@@ -672,27 +669,19 @@ int bch2_btree_insert(struct bch_fs *c, enum btree_id id, struct bkey_i *k,
 			     bch2_btree_insert_trans(trans, id, k, iter_flags));
 }
 
-int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
-				unsigned len, unsigned update_flags)
+int bch2_btree_delete_at(struct btree_trans *trans,
+			 struct btree_iter *iter, unsigned update_flags)
 {
-	struct bkey_i *k;
-
-	k = bch2_trans_kmalloc(trans, sizeof(*k));
-	if (IS_ERR(k))
-		return PTR_ERR(k);
+	struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(*k));
+	int ret = PTR_ERR_OR_ZERO(k);
+	if (ret)
+		return ret;
 
 	bkey_init(&k->k);
 	k->k.p = iter->pos;
-	bch2_key_resize(&k->k, len);
 	return bch2_trans_update(trans, iter, k, update_flags);
 }
 
-int bch2_btree_delete_at(struct btree_trans *trans,
-			 struct btree_iter *iter, unsigned update_flags)
-{
-	return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
-}
-
 int bch2_btree_delete(struct btree_trans *trans,
 		      enum btree_id btree, struct bpos pos,
 		      unsigned update_flags)
@@ -721,7 +710,7 @@ int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
 	int ret = 0;
 
 	bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_intent);
-	while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
+	while ((k = bch2_btree_iter_peek_max(&iter, end)).k) {
 		struct disk_reservation disk_res =
 			bch2_disk_reservation_init(trans->c, 0);
 		struct bkey_i delete;
@@ -794,8 +783,7 @@ int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 	return ret;
 }
 
-int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
-		       struct bpos pos, bool set)
+int bch2_btree_bit_mod_iter(struct btree_trans *trans, struct btree_iter *iter, bool set)
 {
 	struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(*k));
 	int ret = PTR_ERR_OR_ZERO(k);
@@ -804,13 +792,21 @@ int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
 
 	bkey_init(&k->k);
 	k->k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
-	k->k.p = pos;
+	k->k.p = iter->pos;
+	if (iter->flags & BTREE_ITER_is_extents)
+		bch2_key_resize(&k->k, 1);
 
+	return bch2_trans_update(trans, iter, k, 0);
+}
+
+int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
+		       struct bpos pos, bool set)
+{
 	struct btree_iter iter;
 	bch2_trans_iter_init(trans, &iter, btree, pos, BTREE_ITER_intent);
 
-	ret   = bch2_btree_iter_traverse(&iter) ?:
-		bch2_trans_update(trans, &iter, k, 0);
+	int ret = bch2_btree_iter_traverse(&iter) ?:
+		  bch2_btree_bit_mod_iter(trans, &iter, set);
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
@@ -827,10 +823,17 @@ int bch2_btree_bit_mod_buffered(struct btree_trans *trans, enum btree_id btree,
 	return bch2_trans_update_buffered(trans, btree, &k);
 }
 
-static int __bch2_trans_log_msg(struct btree_trans *trans, struct printbuf *buf, unsigned u64s)
+int bch2_trans_log_msg(struct btree_trans *trans, struct printbuf *buf)
 {
+	unsigned u64s = DIV_ROUND_UP(buf->pos, sizeof(u64));
+	prt_chars(buf, '\0', u64s * sizeof(u64) - buf->pos);
+
+	int ret = buf->allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
+	if (ret)
+		return ret;
+
 	struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, jset_u64s(u64s));
-	int ret = PTR_ERR_OR_ZERO(e);
+	ret = PTR_ERR_OR_ZERO(e);
 	if (ret)
 		return ret;
 
@@ -865,9 +868,8 @@ __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
 		memcpy(l->d, buf.buf, buf.pos);
 		c->journal.early_journal_entries.nr += jset_u64s(u64s);
 	} else {
-		ret = bch2_trans_commit_do(c, NULL, NULL,
-			BCH_TRANS_COMMIT_lazy_rw|commit_flags,
-			__bch2_trans_log_msg(trans, &buf, u64s));
+		ret = bch2_trans_commit_do(c, NULL, NULL, commit_flags,
+			bch2_trans_log_msg(trans, &buf));
 	}
 err:
 	printbuf_exit(&buf);
diff --git a/fs/bcachefs/btree_update.h b/fs/bcachefs/btree_update.h
index 70b3c989fac2..8f22ef9a7651 100644
--- a/fs/bcachefs/btree_update.h
+++ b/fs/bcachefs/btree_update.h
@@ -24,7 +24,6 @@ void bch2_btree_insert_key_leaf(struct btree_trans *, struct btree_path *,
 #define BCH_TRANS_COMMIT_FLAGS()							\
 	x(no_enospc,	"don't check for enospc")					\
 	x(no_check_rw,	"don't attempt to take a ref on c->writes")			\
-	x(lazy_rw,	"go read-write if we haven't yet - only for use in recovery")	\
 	x(no_journal_res, "don't take a journal reservation, instead "			\
 			"pin journal entry referred to by trans->journal_res.seq")	\
 	x(journal_reclaim, "operation required for journal reclaim; may return error"	\
@@ -47,8 +46,6 @@ enum bch_trans_commit_flags {
 
 void bch2_trans_commit_flags_to_text(struct printbuf *, enum bch_trans_commit_flags);
 
-int bch2_btree_delete_extent_at(struct btree_trans *, struct btree_iter *,
-				unsigned, unsigned);
 int bch2_btree_delete_at(struct btree_trans *, struct btree_iter *, unsigned);
 int bch2_btree_delete(struct btree_trans *, enum btree_id, struct bpos, unsigned);
 
@@ -66,6 +63,7 @@ int bch2_btree_delete_range_trans(struct btree_trans *, enum btree_id,
 int bch2_btree_delete_range(struct bch_fs *, enum btree_id,
 			    struct bpos, struct bpos, unsigned, u64 *);
 
+int bch2_btree_bit_mod_iter(struct btree_trans *, struct btree_iter *, bool);
 int bch2_btree_bit_mod(struct btree_trans *, enum btree_id, struct bpos, bool);
 int bch2_btree_bit_mod_buffered(struct btree_trans *, enum btree_id, struct bpos, bool);
 
@@ -161,6 +159,7 @@ void bch2_trans_commit_hook(struct btree_trans *,
 			    struct btree_trans_commit_hook *);
 int __bch2_trans_commit(struct btree_trans *, unsigned);
 
+int bch2_trans_log_msg(struct btree_trans *, struct printbuf *);
 __printf(2, 3) int bch2_fs_log_msg(struct bch_fs *, const char *, ...);
 __printf(2, 3) int bch2_journal_log_msg(struct bch_fs *, const char *, ...);
 
@@ -244,7 +243,8 @@ static inline struct bkey_i *bch2_bkey_make_mut_noupdate(struct btree_trans *tra
 				KEY_TYPE_##_type, sizeof(struct bkey_i_##_type)))
 
 static inline struct bkey_i *__bch2_bkey_make_mut(struct btree_trans *trans, struct btree_iter *iter,
-					struct bkey_s_c *k, unsigned flags,
+					struct bkey_s_c *k,
+					enum btree_iter_update_trigger_flags flags,
 					unsigned type, unsigned min_bytes)
 {
 	struct bkey_i *mut = __bch2_bkey_make_mut_noupdate(trans, *k, type, min_bytes);
@@ -261,8 +261,9 @@ static inline struct bkey_i *__bch2_bkey_make_mut(struct btree_trans *trans, str
 	return mut;
 }
 
-static inline struct bkey_i *bch2_bkey_make_mut(struct btree_trans *trans, struct btree_iter *iter,
-						struct bkey_s_c *k, unsigned flags)
+static inline struct bkey_i *bch2_bkey_make_mut(struct btree_trans *trans,
+						struct btree_iter *iter, struct bkey_s_c *k,
+						enum btree_iter_update_trigger_flags flags)
 {
 	return __bch2_bkey_make_mut(trans, iter, k, flags, 0, 0);
 }
@@ -274,7 +275,8 @@ static inline struct bkey_i *bch2_bkey_make_mut(struct btree_trans *trans, struc
 static inline struct bkey_i *__bch2_bkey_get_mut_noupdate(struct btree_trans *trans,
 					 struct btree_iter *iter,
 					 unsigned btree_id, struct bpos pos,
-					 unsigned flags, unsigned type, unsigned min_bytes)
+					 enum btree_iter_update_trigger_flags flags,
+					 unsigned type, unsigned min_bytes)
 {
 	struct bkey_s_c k = __bch2_bkey_get_iter(trans, iter,
 				btree_id, pos, flags|BTREE_ITER_intent, type);
@@ -289,7 +291,7 @@ static inline struct bkey_i *__bch2_bkey_get_mut_noupdate(struct btree_trans *tr
 static inline struct bkey_i *bch2_bkey_get_mut_noupdate(struct btree_trans *trans,
 					       struct btree_iter *iter,
 					       unsigned btree_id, struct bpos pos,
-					       unsigned flags)
+					       enum btree_iter_update_trigger_flags flags)
 {
 	return __bch2_bkey_get_mut_noupdate(trans, iter, btree_id, pos, flags, 0, 0);
 }
@@ -297,7 +299,8 @@ static inline struct bkey_i *bch2_bkey_get_mut_noupdate(struct btree_trans *tran
 static inline struct bkey_i *__bch2_bkey_get_mut(struct btree_trans *trans,
 					 struct btree_iter *iter,
 					 unsigned btree_id, struct bpos pos,
-					 unsigned flags, unsigned type, unsigned min_bytes)
+					 enum btree_iter_update_trigger_flags flags,
+					 unsigned type, unsigned min_bytes)
 {
 	struct bkey_i *mut = __bch2_bkey_get_mut_noupdate(trans, iter,
 				btree_id, pos, flags|BTREE_ITER_intent, type, min_bytes);
@@ -318,7 +321,8 @@ static inline struct bkey_i *__bch2_bkey_get_mut(struct btree_trans *trans,
 static inline struct bkey_i *bch2_bkey_get_mut_minsize(struct btree_trans *trans,
 						       struct btree_iter *iter,
 						       unsigned btree_id, struct bpos pos,
-						       unsigned flags, unsigned min_bytes)
+						       enum btree_iter_update_trigger_flags flags,
+						       unsigned min_bytes)
 {
 	return __bch2_bkey_get_mut(trans, iter, btree_id, pos, flags, 0, min_bytes);
 }
@@ -326,7 +330,7 @@ static inline struct bkey_i *bch2_bkey_get_mut_minsize(struct btree_trans *trans
 static inline struct bkey_i *bch2_bkey_get_mut(struct btree_trans *trans,
 					       struct btree_iter *iter,
 					       unsigned btree_id, struct bpos pos,
-					       unsigned flags)
+					       enum btree_iter_update_trigger_flags flags)
 {
 	return __bch2_bkey_get_mut(trans, iter, btree_id, pos, flags, 0, 0);
 }
@@ -337,7 +341,8 @@ static inline struct bkey_i *bch2_bkey_get_mut(struct btree_trans *trans,
 			KEY_TYPE_##_type, sizeof(struct bkey_i_##_type)))
 
 static inline struct bkey_i *__bch2_bkey_alloc(struct btree_trans *trans, struct btree_iter *iter,
-					       unsigned flags, unsigned type, unsigned val_size)
+					       enum btree_iter_update_trigger_flags flags,
+					       unsigned type, unsigned val_size)
 {
 	struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(*k) + val_size);
 	int ret;
diff --git a/fs/bcachefs/btree_update_interior.c b/fs/bcachefs/btree_update_interior.c
index d596ef93239f..f4aeadbe53c1 100644
--- a/fs/bcachefs/btree_update_interior.c
+++ b/fs/bcachefs/btree_update_interior.c
@@ -58,11 +58,15 @@ int bch2_btree_node_check_topology(struct btree_trans *trans, struct btree *b)
 	       !bpos_eq(bkey_i_to_btree_ptr_v2(&b->key)->v.min_key,
 			b->data->min_key));
 
+	bch2_bkey_buf_init(&prev);
+	bkey_init(&prev.k->k);
+	bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
+
 	if (b == btree_node_root(c, b)) {
 		if (!bpos_eq(b->data->min_key, POS_MIN)) {
 			printbuf_reset(&buf);
 			bch2_bpos_to_text(&buf, b->data->min_key);
-			need_fsck_err(trans, btree_root_bad_min_key,
+			log_fsck_err(trans, btree_root_bad_min_key,
 				      "btree root with incorrect min_key: %s", buf.buf);
 			goto topology_repair;
 		}
@@ -70,18 +74,14 @@ int bch2_btree_node_check_topology(struct btree_trans *trans, struct btree *b)
 		if (!bpos_eq(b->data->max_key, SPOS_MAX)) {
 			printbuf_reset(&buf);
 			bch2_bpos_to_text(&buf, b->data->max_key);
-			need_fsck_err(trans, btree_root_bad_max_key,
+			log_fsck_err(trans, btree_root_bad_max_key,
 				      "btree root with incorrect max_key: %s", buf.buf);
 			goto topology_repair;
 		}
 	}
 
 	if (!b->c.level)
-		return 0;
-
-	bch2_bkey_buf_init(&prev);
-	bkey_init(&prev.k->k);
-	bch2_btree_and_journal_iter_init_node_iter(trans, &iter, b);
+		goto out;
 
 	while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 		if (k.k->type != KEY_TYPE_btree_ptr_v2)
@@ -97,16 +97,16 @@ int bch2_btree_node_check_topology(struct btree_trans *trans, struct btree *b)
 			bch2_topology_error(c);
 
 			printbuf_reset(&buf);
-			prt_str(&buf, "end of prev node doesn't match start of next node\n"),
-			prt_printf(&buf, "  in btree %s level %u node ",
-				   bch2_btree_id_str(b->c.btree_id), b->c.level);
+			prt_str(&buf, "end of prev node doesn't match start of next node\n  in ");
+			bch2_btree_id_level_to_text(&buf, b->c.btree_id, b->c.level);
+			prt_str(&buf, " node ");
 			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 			prt_str(&buf, "\n  prev ");
 			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(prev.k));
 			prt_str(&buf, "\n  next ");
 			bch2_bkey_val_to_text(&buf, c, k);
 
-			need_fsck_err(trans, btree_node_topology_bad_min_key, "%s", buf.buf);
+			log_fsck_err(trans, btree_node_topology_bad_min_key, "%s", buf.buf);
 			goto topology_repair;
 		}
 
@@ -118,25 +118,25 @@ int bch2_btree_node_check_topology(struct btree_trans *trans, struct btree *b)
 		bch2_topology_error(c);
 
 		printbuf_reset(&buf);
-		prt_str(&buf, "empty interior node\n");
-		prt_printf(&buf, "  in btree %s level %u node ",
-			   bch2_btree_id_str(b->c.btree_id), b->c.level);
+		prt_str(&buf, "empty interior node\n  in ");
+		bch2_btree_id_level_to_text(&buf, b->c.btree_id, b->c.level);
+		prt_str(&buf, " node ");
 		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 
-		need_fsck_err(trans, btree_node_topology_empty_interior_node, "%s", buf.buf);
+		log_fsck_err(trans, btree_node_topology_empty_interior_node, "%s", buf.buf);
 		goto topology_repair;
 	} else if (!bpos_eq(prev.k->k.p, b->key.k.p)) {
 		bch2_topology_error(c);
 
 		printbuf_reset(&buf);
-		prt_str(&buf, "last child node doesn't end at end of parent node\n");
-		prt_printf(&buf, "  in btree %s level %u node ",
-			   bch2_btree_id_str(b->c.btree_id), b->c.level);
+		prt_str(&buf, "last child node doesn't end at end of parent node\n  in ");
+		bch2_btree_id_level_to_text(&buf, b->c.btree_id, b->c.level);
+		prt_str(&buf, " node ");
 		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
 		prt_str(&buf, "\n  last key ");
 		bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(prev.k));
 
-		need_fsck_err(trans, btree_node_topology_bad_max_key, "%s", buf.buf);
+		log_fsck_err(trans, btree_node_topology_bad_max_key, "%s", buf.buf);
 		goto topology_repair;
 	}
 out:
@@ -146,13 +146,7 @@ fsck_err:
 	printbuf_exit(&buf);
 	return ret;
 topology_repair:
-	if ((c->opts.recovery_passes & BIT_ULL(BCH_RECOVERY_PASS_check_topology)) &&
-	    c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology) {
-		bch2_inconsistent_error(c);
-		ret = -BCH_ERR_btree_need_topology_repair;
-	} else {
-		ret = bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_topology);
-	}
+	ret = bch2_topology_error(c);
 	goto out;
 }
 
@@ -244,7 +238,6 @@ static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 				       struct btree *b)
 {
 	struct bch_fs *c = trans->c;
-	unsigned i, level = b->c.level;
 
 	bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 
@@ -255,13 +248,9 @@ static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 	mutex_unlock(&c->btree_cache.lock);
 
 	six_unlock_write(&b->c.lock);
-	mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
+	mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 
-	trans_for_each_path(trans, path, i)
-		if (path->l[level].b == b) {
-			btree_node_unlock(trans, path, level);
-			path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
-		}
+	bch2_trans_node_drop(trans, b);
 }
 
 static void bch2_btree_node_free_never_used(struct btree_update *as,
@@ -270,8 +259,6 @@ static void bch2_btree_node_free_never_used(struct btree_update *as,
 {
 	struct bch_fs *c = as->c;
 	struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
-	struct btree_path *path;
-	unsigned i, level = b->c.level;
 
 	BUG_ON(!list_empty(&b->write_blocked));
 	BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
@@ -293,11 +280,7 @@ static void bch2_btree_node_free_never_used(struct btree_update *as,
 
 	six_unlock_intent(&b->c.lock);
 
-	trans_for_each_path(trans, path, i)
-		if (path->l[level].b == b) {
-			btree_node_unlock(trans, path, level);
-			path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
-		}
+	bch2_trans_node_drop(trans, b);
 }
 
 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
@@ -809,7 +792,7 @@ err:
 		mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 		six_unlock_write(&b->c.lock);
 
-		btree_node_write_if_need(c, b, SIX_LOCK_intent);
+		btree_node_write_if_need(trans, b, SIX_LOCK_intent);
 		btree_node_unlock(trans, path, b->c.level);
 		bch2_path_put(trans, path_idx, true);
 	}
@@ -830,7 +813,7 @@ err:
 		b = as->new_nodes[i];
 
 		btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
-		btree_node_write_if_need(c, b, SIX_LOCK_read);
+		btree_node_write_if_need(trans, b, SIX_LOCK_read);
 		six_unlock_read(&b->c.lock);
 	}
 
@@ -1366,9 +1349,14 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
 	if (unlikely(!test_bit(JOURNAL_replay_done, &c->journal.flags)))
 		bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
 
-	if (bch2_bkey_validate(c, bkey_i_to_s_c(insert),
-			      btree_node_type(b), BCH_VALIDATE_write) ?:
-	    bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), BCH_VALIDATE_write)) {
+	struct bkey_validate_context from = (struct bkey_validate_context) {
+		.from	= BKEY_VALIDATE_btree_node,
+		.level	= b->c.level,
+		.btree	= b->c.btree_id,
+		.flags	= BCH_VALIDATE_commit,
+	};
+	if (bch2_bkey_validate(c, bkey_i_to_s_c(insert), from) ?:
+	    bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), from)) {
 		bch2_fs_inconsistent(c, "%s: inserting invalid bkey", __func__);
 		dump_stack();
 	}
@@ -1418,15 +1406,26 @@ bch2_btree_insert_keys_interior(struct btree_update *as,
 	       (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
 		;
 
-	while (!bch2_keylist_empty(keys)) {
-		insert = bch2_keylist_front(keys);
+	for (;
+	     insert != keys->top && bpos_le(insert->k.p, b->key.k.p);
+	     insert = bkey_next(insert))
+		bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
 
-		if (bpos_gt(insert->k.p, b->key.k.p))
-			break;
+	if (bch2_btree_node_check_topology(trans, b)) {
+		struct printbuf buf = PRINTBUF;
 
-		bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
-		bch2_keylist_pop_front(keys);
+		for (struct bkey_i *k = keys->keys;
+		     k != insert;
+		     k = bkey_next(k)) {
+			bch2_bkey_val_to_text(&buf, trans->c, bkey_i_to_s_c(k));
+			prt_newline(&buf);
+		}
+
+		panic("%s(): check_topology error: inserted keys\n%s", __func__, buf.buf);
 	}
+
+	memmove_u64s_down(keys->keys, insert, keys->top_p - insert->_data);
+	keys->top_p -= insert->_data - keys->keys_p;
 }
 
 static bool key_deleted_in_insert(struct keylist *insert_keys, struct bpos pos)
@@ -1575,8 +1574,6 @@ static void btree_split_insert_keys(struct btree_update *as,
 		bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
 
 		bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
-
-		BUG_ON(bch2_btree_node_check_topology(trans, b));
 	}
 }
 
@@ -1599,8 +1596,6 @@ static int btree_split(struct btree_update *as, struct btree_trans *trans,
 	if (ret)
 		return ret;
 
-	bch2_btree_interior_update_will_free_node(as, b);
-
 	if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
 		struct btree *n[2];
 
@@ -1699,16 +1694,18 @@ static int btree_split(struct btree_update *as, struct btree_trans *trans,
 	if (ret)
 		goto err;
 
+	bch2_btree_interior_update_will_free_node(as, b);
+
 	if (n3) {
 		bch2_btree_update_get_open_buckets(as, n3);
-		bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
+		bch2_btree_node_write_trans(trans, n3, SIX_LOCK_intent, 0);
 	}
 	if (n2) {
 		bch2_btree_update_get_open_buckets(as, n2);
-		bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
+		bch2_btree_node_write_trans(trans, n2, SIX_LOCK_intent, 0);
 	}
 	bch2_btree_update_get_open_buckets(as, n1);
-	bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
+	bch2_btree_node_write_trans(trans, n1, SIX_LOCK_intent, 0);
 
 	/*
 	 * The old node must be freed (in memory) _before_ unlocking the new
@@ -1827,8 +1824,6 @@ static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *t
 
 	btree_update_updated_node(as, b);
 	bch2_btree_node_unlock_write(trans, path, b);
-
-	BUG_ON(bch2_btree_node_check_topology(trans, b));
 	return 0;
 split:
 	/*
@@ -1905,7 +1900,7 @@ static void __btree_increase_depth(struct btree_update *as, struct btree_trans *
 	BUG_ON(ret);
 
 	bch2_btree_update_get_open_buckets(as, n);
-	bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
+	bch2_btree_node_write_trans(trans, n, SIX_LOCK_intent, 0);
 	bch2_trans_node_add(trans, path, n);
 	six_unlock_intent(&n->c.lock);
 
@@ -1953,8 +1948,7 @@ int __bch2_foreground_maybe_merge(struct btree_trans *trans,
 	u64 start_time = local_clock();
 	int ret = 0;
 
-	bch2_trans_verify_not_in_restart(trans);
-	bch2_trans_verify_not_unlocked(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 	BUG_ON(!trans->paths[path].should_be_locked);
 	BUG_ON(!btree_node_locked(&trans->paths[path], level));
 
@@ -2058,9 +2052,6 @@ int __bch2_foreground_maybe_merge(struct btree_trans *trans,
 
 	trace_and_count(c, btree_node_merge, trans, b);
 
-	bch2_btree_interior_update_will_free_node(as, b);
-	bch2_btree_interior_update_will_free_node(as, m);
-
 	n = bch2_btree_node_alloc(as, trans, b->c.level);
 
 	SET_BTREE_NODE_SEQ(n->data,
@@ -2096,10 +2087,13 @@ int __bch2_foreground_maybe_merge(struct btree_trans *trans,
 	if (ret)
 		goto err_free_update;
 
+	bch2_btree_interior_update_will_free_node(as, b);
+	bch2_btree_interior_update_will_free_node(as, m);
+
 	bch2_trans_verify_paths(trans);
 
 	bch2_btree_update_get_open_buckets(as, n);
-	bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
+	bch2_btree_node_write_trans(trans, n, SIX_LOCK_intent, 0);
 
 	bch2_btree_node_free_inmem(trans, trans->paths + path, b);
 	bch2_btree_node_free_inmem(trans, trans->paths + sib_path, m);
@@ -2150,8 +2144,6 @@ int bch2_btree_node_rewrite(struct btree_trans *trans,
 	if (ret)
 		goto out;
 
-	bch2_btree_interior_update_will_free_node(as, b);
-
 	n = bch2_btree_node_alloc_replacement(as, trans, b);
 
 	bch2_btree_build_aux_trees(n);
@@ -2175,8 +2167,10 @@ int bch2_btree_node_rewrite(struct btree_trans *trans,
 	if (ret)
 		goto err;
 
+	bch2_btree_interior_update_will_free_node(as, b);
+
 	bch2_btree_update_get_open_buckets(as, n);
-	bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
+	bch2_btree_node_write_trans(trans, n, SIX_LOCK_intent, 0);
 
 	bch2_btree_node_free_inmem(trans, btree_iter_path(trans, iter), b);
 
@@ -2201,42 +2195,50 @@ struct async_btree_rewrite {
 	struct list_head	list;
 	enum btree_id		btree_id;
 	unsigned		level;
-	struct bpos		pos;
-	__le64			seq;
+	struct bkey_buf		key;
 };
 
 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
 					  struct async_btree_rewrite *a)
 {
-	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
-	struct btree *b;
-	int ret;
-
-	bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
+	bch2_trans_node_iter_init(trans, &iter,
+				  a->btree_id, a->key.k->k.p,
 				  BTREE_MAX_DEPTH, a->level, 0);
-	b = bch2_btree_iter_peek_node(&iter);
-	ret = PTR_ERR_OR_ZERO(b);
+	struct btree *b = bch2_btree_iter_peek_node(&iter);
+	int ret = PTR_ERR_OR_ZERO(b);
 	if (ret)
 		goto out;
 
-	if (!b || b->data->keys.seq != a->seq) {
+	bool found = b && btree_ptr_hash_val(&b->key) == btree_ptr_hash_val(a->key.k);
+	ret = found
+		? bch2_btree_node_rewrite(trans, &iter, b, 0)
+		: -ENOENT;
+
+#if 0
+	/* Tracepoint... */
+	if (!ret || ret == -ENOENT) {
+		struct bch_fs *c = trans->c;
 		struct printbuf buf = PRINTBUF;
 
-		if (b)
-			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
-		else
-			prt_str(&buf, "(null");
-		bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
-			 __func__, a->seq, buf.buf);
+		if (!ret) {
+			prt_printf(&buf, "rewrite node:\n  ");
+			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(a->key.k));
+		} else {
+			prt_printf(&buf, "node to rewrite not found:\n  want: ");
+			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(a->key.k));
+			prt_printf(&buf, "\n  got:  ");
+			if (b)
+				bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
+			else
+				prt_str(&buf, "(null)");
+		}
+		bch_info(c, "%s", buf.buf);
 		printbuf_exit(&buf);
-		goto out;
 	}
-
-	ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
+#endif
 out:
 	bch2_trans_iter_exit(trans, &iter);
-
 	return ret;
 }
 
@@ -2247,81 +2249,97 @@ static void async_btree_node_rewrite_work(struct work_struct *work)
 	struct bch_fs *c = a->c;
 
 	int ret = bch2_trans_do(c, async_btree_node_rewrite_trans(trans, a));
-	bch_err_fn_ratelimited(c, ret);
+	if (ret != -ENOENT)
+		bch_err_fn_ratelimited(c, ret);
+
+	spin_lock(&c->btree_node_rewrites_lock);
+	list_del(&a->list);
+	spin_unlock(&c->btree_node_rewrites_lock);
+
+	closure_wake_up(&c->btree_node_rewrites_wait);
+
+	bch2_bkey_buf_exit(&a->key, c);
 	bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
 	kfree(a);
 }
 
 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
 {
-	struct async_btree_rewrite *a;
-	int ret;
-
-	a = kmalloc(sizeof(*a), GFP_NOFS);
-	if (!a) {
-		bch_err(c, "%s: error allocating memory", __func__);
+	struct async_btree_rewrite *a = kmalloc(sizeof(*a), GFP_NOFS);
+	if (!a)
 		return;
-	}
 
 	a->c		= c;
 	a->btree_id	= b->c.btree_id;
 	a->level	= b->c.level;
-	a->pos		= b->key.k.p;
-	a->seq		= b->data->keys.seq;
 	INIT_WORK(&a->work, async_btree_node_rewrite_work);
 
-	if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
-		mutex_lock(&c->pending_node_rewrites_lock);
-		list_add(&a->list, &c->pending_node_rewrites);
-		mutex_unlock(&c->pending_node_rewrites_lock);
-		return;
-	}
+	bch2_bkey_buf_init(&a->key);
+	bch2_bkey_buf_copy(&a->key, c, &b->key);
 
-	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
-		if (test_bit(BCH_FS_started, &c->flags)) {
-			bch_err(c, "%s: error getting c->writes ref", __func__);
-			kfree(a);
-			return;
-		}
+	bool now = false, pending = false;
 
-		ret = bch2_fs_read_write_early(c);
-		bch_err_msg(c, ret, "going read-write");
-		if (ret) {
-			kfree(a);
-			return;
-		}
+	spin_lock(&c->btree_node_rewrites_lock);
+	if (c->curr_recovery_pass > BCH_RECOVERY_PASS_journal_replay &&
+	    bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
+		list_add(&a->list, &c->btree_node_rewrites);
+		now = true;
+	} else if (!test_bit(BCH_FS_may_go_rw, &c->flags)) {
+		list_add(&a->list, &c->btree_node_rewrites_pending);
+		pending = true;
+	}
+	spin_unlock(&c->btree_node_rewrites_lock);
 
-		bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
+	if (now) {
+		queue_work(c->btree_node_rewrite_worker, &a->work);
+	} else if (pending) {
+		/* bch2_do_pending_node_rewrites will execute */
+	} else {
+		bch2_bkey_buf_exit(&a->key, c);
+		kfree(a);
 	}
+}
 
-	queue_work(c->btree_node_rewrite_worker, &a->work);
+void bch2_async_btree_node_rewrites_flush(struct bch_fs *c)
+{
+	closure_wait_event(&c->btree_node_rewrites_wait,
+			   list_empty(&c->btree_node_rewrites));
 }
 
 void bch2_do_pending_node_rewrites(struct bch_fs *c)
 {
-	struct async_btree_rewrite *a, *n;
-
-	mutex_lock(&c->pending_node_rewrites_lock);
-	list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
-		list_del(&a->list);
+	while (1) {
+		spin_lock(&c->btree_node_rewrites_lock);
+		struct async_btree_rewrite *a =
+			list_pop_entry(&c->btree_node_rewrites_pending,
+				       struct async_btree_rewrite, list);
+		if (a)
+			list_add(&a->list, &c->btree_node_rewrites);
+		spin_unlock(&c->btree_node_rewrites_lock);
+
+		if (!a)
+			break;
 
 		bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
 		queue_work(c->btree_node_rewrite_worker, &a->work);
 	}
-	mutex_unlock(&c->pending_node_rewrites_lock);
 }
 
 void bch2_free_pending_node_rewrites(struct bch_fs *c)
 {
-	struct async_btree_rewrite *a, *n;
+	while (1) {
+		spin_lock(&c->btree_node_rewrites_lock);
+		struct async_btree_rewrite *a =
+			list_pop_entry(&c->btree_node_rewrites_pending,
+				       struct async_btree_rewrite, list);
+		spin_unlock(&c->btree_node_rewrites_lock);
 
-	mutex_lock(&c->pending_node_rewrites_lock);
-	list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
-		list_del(&a->list);
+		if (!a)
+			break;
 
+		bch2_bkey_buf_exit(&a->key, c);
 		kfree(a);
 	}
-	mutex_unlock(&c->pending_node_rewrites_lock);
 }
 
 static int __bch2_btree_node_update_key(struct btree_trans *trans,
@@ -2575,8 +2593,9 @@ static void bch2_btree_update_to_text(struct printbuf *out, struct btree_update
 	prt_printf(out, "%ps: ", (void *) as->ip_started);
 	bch2_trans_commit_flags_to_text(out, as->flags);
 
-	prt_printf(out, " btree=%s l=%u-%u mode=%s nodes_written=%u cl.remaining=%u journal_seq=%llu\n",
-		   bch2_btree_id_str(as->btree_id),
+	prt_str(out, " ");
+	bch2_btree_id_to_text(out, as->btree_id);
+	prt_printf(out, " l=%u-%u mode=%s nodes_written=%u cl.remaining=%u journal_seq=%llu\n",
 		   as->update_level_start,
 		   as->update_level_end,
 		   bch2_btree_update_modes[as->mode],
@@ -2677,6 +2696,9 @@ void bch2_btree_reserve_cache_to_text(struct printbuf *out, struct bch_fs *c)
 
 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
 {
+	WARN_ON(!list_empty(&c->btree_node_rewrites));
+	WARN_ON(!list_empty(&c->btree_node_rewrites_pending));
+
 	if (c->btree_node_rewrite_worker)
 		destroy_workqueue(c->btree_node_rewrite_worker);
 	if (c->btree_interior_update_worker)
@@ -2692,8 +2714,9 @@ void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
 	mutex_init(&c->btree_interior_update_lock);
 	INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
 
-	INIT_LIST_HEAD(&c->pending_node_rewrites);
-	mutex_init(&c->pending_node_rewrites_lock);
+	INIT_LIST_HEAD(&c->btree_node_rewrites);
+	INIT_LIST_HEAD(&c->btree_node_rewrites_pending);
+	spin_lock_init(&c->btree_node_rewrites_lock);
 }
 
 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
diff --git a/fs/bcachefs/btree_update_interior.h b/fs/bcachefs/btree_update_interior.h
index 10f400957f21..7930ffea3075 100644
--- a/fs/bcachefs/btree_update_interior.h
+++ b/fs/bcachefs/btree_update_interior.h
@@ -159,7 +159,7 @@ static inline int bch2_foreground_maybe_merge(struct btree_trans *trans,
 					      unsigned level,
 					      unsigned flags)
 {
-	bch2_trans_verify_not_unlocked(trans);
+	bch2_trans_verify_not_unlocked_or_in_restart(trans);
 
 	return  bch2_foreground_maybe_merge_sibling(trans, path, level, flags,
 						    btree_prev_sib) ?:
@@ -334,6 +334,7 @@ void bch2_journal_entry_to_btree_root(struct bch_fs *, struct jset_entry *);
 struct jset_entry *bch2_btree_roots_to_journal_entries(struct bch_fs *,
 					struct jset_entry *, unsigned long);
 
+void bch2_async_btree_node_rewrites_flush(struct bch_fs *);
 void bch2_do_pending_node_rewrites(struct bch_fs *);
 void bch2_free_pending_node_rewrites(struct bch_fs *);
 
diff --git a/fs/bcachefs/btree_write_buffer.c b/fs/bcachefs/btree_write_buffer.c
index 1639c60dffa0..b56c4987b8c9 100644
--- a/fs/bcachefs/btree_write_buffer.c
+++ b/fs/bcachefs/btree_write_buffer.c
@@ -19,8 +19,6 @@
 static int bch2_btree_write_buffer_journal_flush(struct journal *,
 				struct journal_entry_pin *, u64);
 
-static int bch2_journal_keys_to_write_buffer(struct bch_fs *, struct journal_buf *);
-
 static inline bool __wb_key_ref_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
 {
 	return (cmp_int(l->hi, r->hi) ?:
@@ -314,6 +312,8 @@ static int bch2_btree_write_buffer_flush_locked(struct btree_trans *trans)
 	darray_for_each(wb->sorted, i) {
 		struct btree_write_buffered_key *k = &wb->flushing.keys.data[i->idx];
 
+		BUG_ON(!btree_type_uses_write_buffer(k->btree));
+
 		for (struct wb_key_ref *n = i + 1; n < min(i + 4, &darray_top(wb->sorted)); n++)
 			prefetch(&wb->flushing.keys.data[n->idx]);
 
@@ -481,21 +481,55 @@ err:
 	return ret;
 }
 
-static int fetch_wb_keys_from_journal(struct bch_fs *c, u64 seq)
+static int bch2_journal_keys_to_write_buffer(struct bch_fs *c, struct journal_buf *buf)
+{
+	struct journal_keys_to_wb dst;
+	int ret = 0;
+
+	bch2_journal_keys_to_write_buffer_start(c, &dst, le64_to_cpu(buf->data->seq));
+
+	for_each_jset_entry_type(entry, buf->data, BCH_JSET_ENTRY_write_buffer_keys) {
+		jset_entry_for_each_key(entry, k) {
+			ret = bch2_journal_key_to_wb(c, &dst, entry->btree_id, k);
+			if (ret)
+				goto out;
+		}
+
+		entry->type = BCH_JSET_ENTRY_btree_keys;
+	}
+out:
+	ret = bch2_journal_keys_to_write_buffer_end(c, &dst) ?: ret;
+	return ret;
+}
+
+static int fetch_wb_keys_from_journal(struct bch_fs *c, u64 max_seq)
 {
 	struct journal *j = &c->journal;
 	struct journal_buf *buf;
+	bool blocked;
 	int ret = 0;
 
-	while (!ret && (buf = bch2_next_write_buffer_flush_journal_buf(j, seq))) {
+	while (!ret && (buf = bch2_next_write_buffer_flush_journal_buf(j, max_seq, &blocked))) {
 		ret = bch2_journal_keys_to_write_buffer(c, buf);
+
+		if (!blocked && !ret) {
+			spin_lock(&j->lock);
+			buf->need_flush_to_write_buffer = false;
+			spin_unlock(&j->lock);
+		}
+
 		mutex_unlock(&j->buf_lock);
+
+		if (blocked) {
+			bch2_journal_unblock(j);
+			break;
+		}
 	}
 
 	return ret;
 }
 
-static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq,
+static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 max_seq,
 					bool *did_work)
 {
 	struct bch_fs *c = trans->c;
@@ -505,7 +539,7 @@ static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq,
 	do {
 		bch2_trans_unlock(trans);
 
-		fetch_from_journal_err = fetch_wb_keys_from_journal(c, seq);
+		fetch_from_journal_err = fetch_wb_keys_from_journal(c, max_seq);
 
 		*did_work |= wb->inc.keys.nr || wb->flushing.keys.nr;
 
@@ -518,8 +552,8 @@ static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq,
 		mutex_unlock(&wb->flushing.lock);
 	} while (!ret &&
 		 (fetch_from_journal_err ||
-		  (wb->inc.pin.seq && wb->inc.pin.seq <= seq) ||
-		  (wb->flushing.pin.seq && wb->flushing.pin.seq <= seq)));
+		  (wb->inc.pin.seq && wb->inc.pin.seq <= max_seq) ||
+		  (wb->flushing.pin.seq && wb->flushing.pin.seq <= max_seq)));
 
 	return ret;
 }
@@ -600,6 +634,14 @@ int bch2_btree_write_buffer_maybe_flush(struct btree_trans *trans,
 	bch2_bkey_buf_init(&tmp);
 
 	if (!bkey_and_val_eq(referring_k, bkey_i_to_s_c(last_flushed->k))) {
+		if (trace_write_buffer_maybe_flush_enabled()) {
+			struct printbuf buf = PRINTBUF;
+
+			bch2_bkey_val_to_text(&buf, c, referring_k);
+			trace_write_buffer_maybe_flush(trans, _RET_IP_, buf.buf);
+			printbuf_exit(&buf);
+		}
+
 		bch2_bkey_buf_reassemble(&tmp, c, referring_k);
 
 		if (bkey_is_btree_ptr(referring_k.k)) {
@@ -771,31 +813,6 @@ int bch2_journal_keys_to_write_buffer_end(struct bch_fs *c, struct journal_keys_
 	return ret;
 }
 
-static int bch2_journal_keys_to_write_buffer(struct bch_fs *c, struct journal_buf *buf)
-{
-	struct journal_keys_to_wb dst;
-	int ret = 0;
-
-	bch2_journal_keys_to_write_buffer_start(c, &dst, le64_to_cpu(buf->data->seq));
-
-	for_each_jset_entry_type(entry, buf->data, BCH_JSET_ENTRY_write_buffer_keys) {
-		jset_entry_for_each_key(entry, k) {
-			ret = bch2_journal_key_to_wb(c, &dst, entry->btree_id, k);
-			if (ret)
-				goto out;
-		}
-
-		entry->type = BCH_JSET_ENTRY_btree_keys;
-	}
-
-	spin_lock(&c->journal.lock);
-	buf->need_flush_to_write_buffer = false;
-	spin_unlock(&c->journal.lock);
-out:
-	ret = bch2_journal_keys_to_write_buffer_end(c, &dst) ?: ret;
-	return ret;
-}
-
 static int wb_keys_resize(struct btree_write_buffer_keys *wb, size_t new_size)
 {
 	if (wb->keys.size >= new_size)
diff --git a/fs/bcachefs/buckets.c b/fs/bcachefs/buckets.c
index ec7d9a59bea9..345b117a4a4a 100644
--- a/fs/bcachefs/buckets.c
+++ b/fs/bcachefs/buckets.c
@@ -18,7 +18,9 @@
 #include "error.h"
 #include "inode.h"
 #include "movinggc.h"
+#include "rebalance.h"
 #include "recovery.h"
+#include "recovery_passes.h"
 #include "reflink.h"
 #include "replicas.h"
 #include "subvolume.h"
@@ -260,8 +262,6 @@ int bch2_check_fix_ptrs(struct btree_trans *trans,
 	struct printbuf buf = PRINTBUF;
 	int ret = 0;
 
-	percpu_down_read(&c->mark_lock);
-
 	bkey_for_each_ptr_decode(k.k, ptrs_c, p, entry_c) {
 		ret = bch2_check_fix_ptr(trans, k, p, entry_c, &do_update);
 		if (ret)
@@ -362,7 +362,6 @@ found:
 			bch_info(c, "new key %s", buf.buf);
 		}
 
-		percpu_up_read(&c->mark_lock);
 		struct btree_iter iter;
 		bch2_trans_node_iter_init(trans, &iter, btree, new->k.p, 0, level,
 					  BTREE_ITER_intent|BTREE_ITER_all_snapshots);
@@ -371,8 +370,6 @@ found:
 					  BTREE_UPDATE_internal_snapshot_node|
 					  BTREE_TRIGGER_norun);
 		bch2_trans_iter_exit(trans, &iter);
-		percpu_down_read(&c->mark_lock);
-
 		if (ret)
 			goto err;
 
@@ -380,7 +377,6 @@ found:
 			bch2_btree_node_update_key_early(trans, btree, level - 1, k, new);
 	}
 err:
-	percpu_up_read(&c->mark_lock);
 	printbuf_exit(&buf);
 	return ret;
 }
@@ -401,8 +397,8 @@ int bch2_bucket_ref_update(struct btree_trans *trans, struct bch_dev *ca,
 	BUG_ON(!sectors);
 
 	if (gen_after(ptr->gen, b_gen)) {
-		bch2_fsck_err(trans, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
-			      ptr_gen_newer_than_bucket_gen,
+		bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_allocations);
+		log_fsck_err(trans, ptr_gen_newer_than_bucket_gen,
 			"bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 			"while marking %s",
 			ptr->dev, bucket_nr, b_gen,
@@ -415,8 +411,8 @@ int bch2_bucket_ref_update(struct btree_trans *trans, struct bch_dev *ca,
 	}
 
 	if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
-		bch2_fsck_err(trans, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
-			      ptr_too_stale,
+		bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_allocations);
+		log_fsck_err(trans, ptr_too_stale,
 			"bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 			"while marking %s",
 			ptr->dev, bucket_nr, b_gen,
@@ -435,8 +431,8 @@ int bch2_bucket_ref_update(struct btree_trans *trans, struct bch_dev *ca,
 	}
 
 	if (b_gen != ptr->gen) {
-		bch2_fsck_err(trans, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
-			      stale_dirty_ptr,
+		bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_allocations);
+		log_fsck_err(trans, stale_dirty_ptr,
 			"bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 			"while marking %s",
 			ptr->dev, bucket_nr, b_gen,
@@ -451,8 +447,8 @@ int bch2_bucket_ref_update(struct btree_trans *trans, struct bch_dev *ca,
 	}
 
 	if (bucket_data_type_mismatch(bucket_data_type, ptr_data_type)) {
-		bch2_fsck_err(trans, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
-			      ptr_bucket_data_type_mismatch,
+		bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_allocations);
+		log_fsck_err(trans, ptr_bucket_data_type_mismatch,
 			"bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 			"while marking %s",
 			ptr->dev, bucket_nr, b_gen,
@@ -466,8 +462,8 @@ int bch2_bucket_ref_update(struct btree_trans *trans, struct bch_dev *ca,
 	}
 
 	if ((u64) *bucket_sectors + sectors > U32_MAX) {
-		bch2_fsck_err(trans, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
-			      bucket_sector_count_overflow,
+		bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_allocations);
+		log_fsck_err(trans, bucket_sector_count_overflow,
 			"bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U32_MAX\n"
 			"while marking %s",
 			ptr->dev, bucket_nr, b_gen,
@@ -485,7 +481,9 @@ out:
 	printbuf_exit(&buf);
 	return ret;
 err:
+fsck_err:
 	bch2_dump_trans_updates(trans);
+	bch2_inconsistent_error(c);
 	ret = -BCH_ERR_bucket_ref_update;
 	goto out;
 }
@@ -543,7 +541,8 @@ static int __mark_pointer(struct btree_trans *trans, struct bch_dev *ca,
 			  struct bkey_s_c k,
 			  const struct extent_ptr_decoded *p,
 			  s64 sectors, enum bch_data_type ptr_data_type,
-			  struct bch_alloc_v4 *a)
+			  struct bch_alloc_v4 *a,
+			  bool insert)
 {
 	u32 *dst_sectors = p->has_ec	? &a->stripe_sectors :
 		!p->ptr.cached		? &a->dirty_sectors :
@@ -553,8 +552,8 @@ static int __mark_pointer(struct btree_trans *trans, struct bch_dev *ca,
 
 	if (ret)
 		return ret;
-
-	alloc_data_type_set(a, ptr_data_type);
+	if (insert)
+		alloc_data_type_set(a, ptr_data_type);
 	return 0;
 }
 
@@ -570,8 +569,10 @@ static int bch2_trigger_pointer(struct btree_trans *trans,
 	struct printbuf buf = PRINTBUF;
 	int ret = 0;
 
-	u64 abs_sectors = ptr_disk_sectors(level ? btree_sectors(c) : k.k->size, p);
-	*sectors = insert ? abs_sectors : -abs_sectors;
+	struct bkey_i_backpointer bp;
+	bch2_extent_ptr_to_bp(c, btree_id, level, k, p, entry, &bp);
+
+	*sectors = insert ? bp.v.bucket_len : -(s64) bp.v.bucket_len;
 
 	struct bch_dev *ca = bch2_dev_tryget(c, p.ptr.dev);
 	if (unlikely(!ca)) {
@@ -580,41 +581,36 @@ static int bch2_trigger_pointer(struct btree_trans *trans,
 		goto err;
 	}
 
-	struct bpos bucket;
-	struct bch_backpointer bp;
-	__bch2_extent_ptr_to_bp(trans->c, ca, btree_id, level, k, p, entry, &bucket, &bp, abs_sectors);
+	struct bpos bucket = PTR_BUCKET_POS(ca, &p.ptr);
 
 	if (flags & BTREE_TRIGGER_transactional) {
 		struct bkey_i_alloc_v4 *a = bch2_trans_start_alloc_update(trans, bucket, 0);
 		ret = PTR_ERR_OR_ZERO(a) ?:
-			__mark_pointer(trans, ca, k, &p, *sectors, bp.data_type, &a->v);
+			__mark_pointer(trans, ca, k, &p, *sectors, bp.v.data_type, &a->v, insert);
 		if (ret)
 			goto err;
 
 		if (!p.ptr.cached) {
-			ret = bch2_bucket_backpointer_mod(trans, ca, bucket, bp, k, insert);
+			ret = bch2_bucket_backpointer_mod(trans, k, &bp, insert);
 			if (ret)
 				goto err;
 		}
 	}
 
 	if (flags & BTREE_TRIGGER_gc) {
-		percpu_down_read(&c->mark_lock);
 		struct bucket *g = gc_bucket(ca, bucket.offset);
 		if (bch2_fs_inconsistent_on(!g, c, "reference to invalid bucket on device %u\n  %s",
 					    p.ptr.dev,
 					    (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 			ret = -BCH_ERR_trigger_pointer;
-			goto err_unlock;
+			goto err;
 		}
 
 		bucket_lock(g);
 		struct bch_alloc_v4 old = bucket_m_to_alloc(*g), new = old;
-		ret = __mark_pointer(trans, ca, k, &p, *sectors, bp.data_type, &new);
+		ret = __mark_pointer(trans, ca, k, &p, *sectors, bp.v.data_type, &new, insert);
 		alloc_to_bucket(g, new);
 		bucket_unlock(g);
-err_unlock:
-		percpu_up_read(&c->mark_lock);
 
 		if (!ret)
 			ret = bch2_alloc_key_to_dev_counters(trans, ca, &old, &new, flags);
@@ -951,6 +947,7 @@ static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
 				    enum bch_data_type type,
 				    unsigned sectors)
 {
+	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
 	int ret = 0;
 
@@ -960,8 +957,8 @@ static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
 		return PTR_ERR(a);
 
 	if (a->v.data_type && type && a->v.data_type != type) {
-		bch2_fsck_err(trans, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
-			      bucket_metadata_type_mismatch,
+		bch2_run_explicit_recovery_pass(c, BCH_RECOVERY_PASS_check_allocations);
+		log_fsck_err(trans, bucket_metadata_type_mismatch,
 			"bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
 			"while marking %s",
 			iter.pos.inode, iter.pos.offset, a->v.gen,
@@ -979,6 +976,7 @@ static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
 		ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
 	}
 err:
+fsck_err:
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
@@ -990,11 +988,10 @@ static int bch2_mark_metadata_bucket(struct btree_trans *trans, struct bch_dev *
 	struct bch_fs *c = trans->c;
 	int ret = 0;
 
-	percpu_down_read(&c->mark_lock);
 	struct bucket *g = gc_bucket(ca, b);
 	if (bch2_fs_inconsistent_on(!g, c, "reference to invalid bucket on device %u when marking metadata type %s",
 				    ca->dev_idx, bch2_data_type_str(data_type)))
-		goto err_unlock;
+		goto err;
 
 	bucket_lock(g);
 	struct bch_alloc_v4 old = bucket_m_to_alloc(*g);
@@ -1004,26 +1001,24 @@ static int bch2_mark_metadata_bucket(struct btree_trans *trans, struct bch_dev *
 			"different types of data in same bucket: %s, %s",
 			bch2_data_type_str(g->data_type),
 			bch2_data_type_str(data_type)))
-		goto err;
+		goto err_unlock;
 
 	if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 			"bucket %u:%llu gen %u data type %s sector count overflow: %u + %u > bucket size",
 			ca->dev_idx, b, g->gen,
 			bch2_data_type_str(g->data_type ?: data_type),
 			g->dirty_sectors, sectors))
-		goto err;
+		goto err_unlock;
 
 	g->data_type = data_type;
 	g->dirty_sectors += sectors;
 	struct bch_alloc_v4 new = bucket_m_to_alloc(*g);
 	bucket_unlock(g);
-	percpu_up_read(&c->mark_lock);
 	ret = bch2_alloc_key_to_dev_counters(trans, ca, &old, &new, flags);
 	return ret;
-err:
-	bucket_unlock(g);
 err_unlock:
-	percpu_up_read(&c->mark_lock);
+	bucket_unlock(g);
+err:
 	return -BCH_ERR_metadata_bucket_inconsistency;
 }
 
@@ -1155,6 +1150,31 @@ int bch2_trans_mark_dev_sbs(struct bch_fs *c)
 	return bch2_trans_mark_dev_sbs_flags(c, BTREE_TRIGGER_transactional);
 }
 
+bool bch2_is_superblock_bucket(struct bch_dev *ca, u64 b)
+{
+	struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
+	u64 b_offset	= bucket_to_sector(ca, b);
+	u64 b_end	= bucket_to_sector(ca, b + 1);
+	unsigned i;
+
+	if (!b)
+		return true;
+
+	for (i = 0; i < layout->nr_superblocks; i++) {
+		u64 offset = le64_to_cpu(layout->sb_offset[i]);
+		u64 end = offset + (1 << layout->sb_max_size_bits);
+
+		if (!(offset >= b_end || end <= b_offset))
+			return true;
+	}
+
+	for (i = 0; i < ca->journal.nr; i++)
+		if (b == ca->journal.buckets[i])
+			return true;
+
+	return false;
+}
+
 /* Disk reservations: */
 
 #define SECTORS_CACHE	1024
@@ -1238,7 +1258,7 @@ int bch2_buckets_nouse_alloc(struct bch_fs *c)
 	for_each_member_device(c, ca) {
 		BUG_ON(ca->buckets_nouse);
 
-		ca->buckets_nouse = kvmalloc(BITS_TO_LONGS(ca->mi.nbuckets) *
+		ca->buckets_nouse = bch2_kvmalloc(BITS_TO_LONGS(ca->mi.nbuckets) *
 					    sizeof(unsigned long),
 					    GFP_KERNEL|__GFP_ZERO);
 		if (!ca->buckets_nouse) {
@@ -1264,10 +1284,15 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 	bool resize = ca->bucket_gens != NULL;
 	int ret;
 
-	BUG_ON(resize && ca->buckets_nouse);
+	if (resize)
+		lockdep_assert_held(&c->state_lock);
+
+	if (resize && ca->buckets_nouse)
+		return -BCH_ERR_no_resize_with_buckets_nouse;
 
-	if (!(bucket_gens	= kvmalloc(sizeof(struct bucket_gens) + nbuckets,
-					   GFP_KERNEL|__GFP_ZERO))) {
+	bucket_gens = bch2_kvmalloc(struct_size(bucket_gens, b, nbuckets),
+				    GFP_KERNEL|__GFP_ZERO);
+	if (!bucket_gens) {
 		ret = -BCH_ERR_ENOMEM_bucket_gens;
 		goto err;
 	}
@@ -1277,19 +1302,16 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 	bucket_gens->nbuckets_minus_first =
 		bucket_gens->nbuckets - bucket_gens->first_bucket;
 
-	if (resize) {
-		down_write(&ca->bucket_lock);
-		percpu_down_write(&c->mark_lock);
-	}
-
 	old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
 
 	if (resize) {
-		size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
-
+		bucket_gens->nbuckets = min(bucket_gens->nbuckets,
+					    old_bucket_gens->nbuckets);
+		bucket_gens->nbuckets_minus_first =
+			bucket_gens->nbuckets - bucket_gens->first_bucket;
 		memcpy(bucket_gens->b,
 		       old_bucket_gens->b,
-		       n);
+		       bucket_gens->nbuckets);
 	}
 
 	rcu_assign_pointer(ca->bucket_gens, bucket_gens);
@@ -1297,11 +1319,6 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 
 	nbuckets = ca->mi.nbuckets;
 
-	if (resize) {
-		percpu_up_write(&c->mark_lock);
-		up_write(&ca->bucket_lock);
-	}
-
 	ret = 0;
 err:
 	if (bucket_gens)
diff --git a/fs/bcachefs/buckets.h b/fs/bcachefs/buckets.h
index ccc78bfe2fd4..a9acdd6c0c86 100644
--- a/fs/bcachefs/buckets.h
+++ b/fs/bcachefs/buckets.h
@@ -82,16 +82,15 @@ static inline void bucket_lock(struct bucket *b)
 
 static inline struct bucket *gc_bucket(struct bch_dev *ca, size_t b)
 {
-	return genradix_ptr(&ca->buckets_gc, b);
+	return bucket_valid(ca, b)
+		? genradix_ptr(&ca->buckets_gc, b)
+		: NULL;
 }
 
 static inline struct bucket_gens *bucket_gens(struct bch_dev *ca)
 {
 	return rcu_dereference_check(ca->bucket_gens,
-				     !ca->fs ||
-				     percpu_rwsem_is_held(&ca->fs->mark_lock) ||
-				     lockdep_is_held(&ca->fs->state_lock) ||
-				     lockdep_is_held(&ca->bucket_lock));
+				     lockdep_is_held(&ca->fs->state_lock));
 }
 
 static inline u8 *bucket_gen(struct bch_dev *ca, size_t b)
@@ -308,26 +307,7 @@ int bch2_trans_mark_dev_sbs_flags(struct bch_fs *,
 				    enum btree_iter_update_trigger_flags);
 int bch2_trans_mark_dev_sbs(struct bch_fs *);
 
-static inline bool is_superblock_bucket(struct bch_dev *ca, u64 b)
-{
-	struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
-	u64 b_offset	= bucket_to_sector(ca, b);
-	u64 b_end	= bucket_to_sector(ca, b + 1);
-	unsigned i;
-
-	if (!b)
-		return true;
-
-	for (i = 0; i < layout->nr_superblocks; i++) {
-		u64 offset = le64_to_cpu(layout->sb_offset[i]);
-		u64 end = offset + (1 << layout->sb_max_size_bits);
-
-		if (!(offset >= b_end || end <= b_offset))
-			return true;
-	}
-
-	return false;
-}
+bool bch2_is_superblock_bucket(struct bch_dev *, u64);
 
 static inline const char *bch2_data_type_str(enum bch_data_type type)
 {
diff --git a/fs/bcachefs/buckets_types.h b/fs/bcachefs/buckets_types.h
index 28bd09a253c8..7174047b8e92 100644
--- a/fs/bcachefs/buckets_types.h
+++ b/fs/bcachefs/buckets_types.h
@@ -24,7 +24,7 @@ struct bucket_gens {
 	u16			first_bucket;
 	size_t			nbuckets;
 	size_t			nbuckets_minus_first;
-	u8			b[];
+	u8			b[] __counted_by(nbuckets);
 };
 
 struct bch_dev_usage {
diff --git a/fs/bcachefs/chardev.c b/fs/bcachefs/chardev.c
index 2182b555c112..46e9e32105a9 100644
--- a/fs/bcachefs/chardev.c
+++ b/fs/bcachefs/chardev.c
@@ -6,11 +6,11 @@
 #include "buckets.h"
 #include "chardev.h"
 #include "disk_accounting.h"
+#include "fsck.h"
 #include "journal.h"
 #include "move.h"
 #include "recovery_passes.h"
 #include "replicas.h"
-#include "super.h"
 #include "super-io.h"
 #include "thread_with_file.h"
 
@@ -127,130 +127,6 @@ static long bch2_ioctl_incremental(struct bch_ioctl_incremental __user *user_arg
 }
 #endif
 
-struct fsck_thread {
-	struct thread_with_stdio thr;
-	struct bch_fs		*c;
-	struct bch_opts		opts;
-};
-
-static void bch2_fsck_thread_exit(struct thread_with_stdio *_thr)
-{
-	struct fsck_thread *thr = container_of(_thr, struct fsck_thread, thr);
-	kfree(thr);
-}
-
-static int bch2_fsck_offline_thread_fn(struct thread_with_stdio *stdio)
-{
-	struct fsck_thread *thr = container_of(stdio, struct fsck_thread, thr);
-	struct bch_fs *c = thr->c;
-
-	int ret = PTR_ERR_OR_ZERO(c);
-	if (ret)
-		return ret;
-
-	ret = bch2_fs_start(thr->c);
-	if (ret)
-		goto err;
-
-	if (test_bit(BCH_FS_errors_fixed, &c->flags)) {
-		bch2_stdio_redirect_printf(&stdio->stdio, false, "%s: errors fixed\n", c->name);
-		ret |= 1;
-	}
-	if (test_bit(BCH_FS_error, &c->flags)) {
-		bch2_stdio_redirect_printf(&stdio->stdio, false, "%s: still has errors\n", c->name);
-		ret |= 4;
-	}
-err:
-	bch2_fs_stop(c);
-	return ret;
-}
-
-static const struct thread_with_stdio_ops bch2_offline_fsck_ops = {
-	.exit		= bch2_fsck_thread_exit,
-	.fn		= bch2_fsck_offline_thread_fn,
-};
-
-static long bch2_ioctl_fsck_offline(struct bch_ioctl_fsck_offline __user *user_arg)
-{
-	struct bch_ioctl_fsck_offline arg;
-	struct fsck_thread *thr = NULL;
-	darray_str(devs) = {};
-	long ret = 0;
-
-	if (copy_from_user(&arg, user_arg, sizeof(arg)))
-		return -EFAULT;
-
-	if (arg.flags)
-		return -EINVAL;
-
-	if (!capable(CAP_SYS_ADMIN))
-		return -EPERM;
-
-	for (size_t i = 0; i < arg.nr_devs; i++) {
-		u64 dev_u64;
-		ret = copy_from_user_errcode(&dev_u64, &user_arg->devs[i], sizeof(u64));
-		if (ret)
-			goto err;
-
-		char *dev_str = strndup_user((char __user *)(unsigned long) dev_u64, PATH_MAX);
-		ret = PTR_ERR_OR_ZERO(dev_str);
-		if (ret)
-			goto err;
-
-		ret = darray_push(&devs, dev_str);
-		if (ret) {
-			kfree(dev_str);
-			goto err;
-		}
-	}
-
-	thr = kzalloc(sizeof(*thr), GFP_KERNEL);
-	if (!thr) {
-		ret = -ENOMEM;
-		goto err;
-	}
-
-	thr->opts = bch2_opts_empty();
-
-	if (arg.opts) {
-		char *optstr = strndup_user((char __user *)(unsigned long) arg.opts, 1 << 16);
-		ret =   PTR_ERR_OR_ZERO(optstr) ?:
-			bch2_parse_mount_opts(NULL, &thr->opts, NULL, optstr);
-		if (!IS_ERR(optstr))
-			kfree(optstr);
-
-		if (ret)
-			goto err;
-	}
-
-	opt_set(thr->opts, stdio, (u64)(unsigned long)&thr->thr.stdio);
-	opt_set(thr->opts, read_only, 1);
-	opt_set(thr->opts, ratelimit_errors, 0);
-
-	/* We need request_key() to be called before we punt to kthread: */
-	opt_set(thr->opts, nostart, true);
-
-	bch2_thread_with_stdio_init(&thr->thr, &bch2_offline_fsck_ops);
-
-	thr->c = bch2_fs_open(devs.data, arg.nr_devs, thr->opts);
-
-	if (!IS_ERR(thr->c) &&
-	    thr->c->opts.errors == BCH_ON_ERROR_panic)
-		thr->c->opts.errors = BCH_ON_ERROR_ro;
-
-	ret = __bch2_run_thread_with_stdio(&thr->thr);
-out:
-	darray_for_each(devs, i)
-		kfree(*i);
-	darray_exit(&devs);
-	return ret;
-err:
-	if (thr)
-		bch2_fsck_thread_exit(&thr->thr);
-	pr_err("ret %s", bch2_err_str(ret));
-	goto out;
-}
-
 static long bch2_global_ioctl(unsigned cmd, void __user *arg)
 {
 	long ret;
@@ -775,99 +651,6 @@ static long bch2_ioctl_disk_resize_journal(struct bch_fs *c,
 	return ret;
 }
 
-static int bch2_fsck_online_thread_fn(struct thread_with_stdio *stdio)
-{
-	struct fsck_thread *thr = container_of(stdio, struct fsck_thread, thr);
-	struct bch_fs *c = thr->c;
-
-	c->stdio_filter = current;
-	c->stdio = &thr->thr.stdio;
-
-	/*
-	 * XXX: can we figure out a way to do this without mucking with c->opts?
-	 */
-	unsigned old_fix_errors = c->opts.fix_errors;
-	if (opt_defined(thr->opts, fix_errors))
-		c->opts.fix_errors = thr->opts.fix_errors;
-	else
-		c->opts.fix_errors = FSCK_FIX_ask;
-
-	c->opts.fsck = true;
-	set_bit(BCH_FS_fsck_running, &c->flags);
-
-	c->curr_recovery_pass = BCH_RECOVERY_PASS_check_alloc_info;
-	int ret = bch2_run_online_recovery_passes(c);
-
-	clear_bit(BCH_FS_fsck_running, &c->flags);
-	bch_err_fn(c, ret);
-
-	c->stdio = NULL;
-	c->stdio_filter = NULL;
-	c->opts.fix_errors = old_fix_errors;
-
-	up(&c->online_fsck_mutex);
-	bch2_ro_ref_put(c);
-	return ret;
-}
-
-static const struct thread_with_stdio_ops bch2_online_fsck_ops = {
-	.exit		= bch2_fsck_thread_exit,
-	.fn		= bch2_fsck_online_thread_fn,
-};
-
-static long bch2_ioctl_fsck_online(struct bch_fs *c,
-				   struct bch_ioctl_fsck_online arg)
-{
-	struct fsck_thread *thr = NULL;
-	long ret = 0;
-
-	if (arg.flags)
-		return -EINVAL;
-
-	if (!capable(CAP_SYS_ADMIN))
-		return -EPERM;
-
-	if (!bch2_ro_ref_tryget(c))
-		return -EROFS;
-
-	if (down_trylock(&c->online_fsck_mutex)) {
-		bch2_ro_ref_put(c);
-		return -EAGAIN;
-	}
-
-	thr = kzalloc(sizeof(*thr), GFP_KERNEL);
-	if (!thr) {
-		ret = -ENOMEM;
-		goto err;
-	}
-
-	thr->c = c;
-	thr->opts = bch2_opts_empty();
-
-	if (arg.opts) {
-		char *optstr = strndup_user((char __user *)(unsigned long) arg.opts, 1 << 16);
-
-		ret =   PTR_ERR_OR_ZERO(optstr) ?:
-			bch2_parse_mount_opts(c, &thr->opts, NULL, optstr);
-		if (!IS_ERR(optstr))
-			kfree(optstr);
-
-		if (ret)
-			goto err;
-	}
-
-	ret = bch2_run_thread_with_stdio(&thr->thr, &bch2_online_fsck_ops);
-err:
-	if (ret < 0) {
-		bch_err_fn(c, ret);
-		if (thr)
-			bch2_fsck_thread_exit(&thr->thr);
-		up(&c->online_fsck_mutex);
-		bch2_ro_ref_put(c);
-	}
-	return ret;
-}
-
 #define BCH_IOCTL(_name, _argtype)					\
 do {									\
 	_argtype i;							\
diff --git a/fs/bcachefs/checksum.c b/fs/bcachefs/checksum.c
index ce8fc677bef9..23a383577d4c 100644
--- a/fs/bcachefs/checksum.c
+++ b/fs/bcachefs/checksum.c
@@ -2,6 +2,7 @@
 #include "bcachefs.h"
 #include "checksum.h"
 #include "errcode.h"
+#include "error.h"
 #include "super.h"
 #include "super-io.h"
 
@@ -252,6 +253,10 @@ int bch2_encrypt(struct bch_fs *c, unsigned type,
 	if (!bch2_csum_type_is_encryption(type))
 		return 0;
 
+	if (bch2_fs_inconsistent_on(!c->chacha20,
+				    c, "attempting to encrypt without encryption key"))
+		return -BCH_ERR_no_encryption_key;
+
 	return do_encrypt(c->chacha20, nonce, data, len);
 }
 
@@ -337,8 +342,9 @@ int __bch2_encrypt_bio(struct bch_fs *c, unsigned type,
 	size_t sgl_len = 0;
 	int ret = 0;
 
-	if (!bch2_csum_type_is_encryption(type))
-		return 0;
+	if (bch2_fs_inconsistent_on(!c->chacha20,
+				    c, "attempting to encrypt without encryption key"))
+		return -BCH_ERR_no_encryption_key;
 
 	darray_init(&sgl);
 
diff --git a/fs/bcachefs/checksum.h b/fs/bcachefs/checksum.h
index e40499fde9a4..43b9d71f2f2b 100644
--- a/fs/bcachefs/checksum.h
+++ b/fs/bcachefs/checksum.h
@@ -109,7 +109,7 @@ int bch2_enable_encryption(struct bch_fs *, bool);
 void bch2_fs_encryption_exit(struct bch_fs *);
 int bch2_fs_encryption_init(struct bch_fs *);
 
-static inline enum bch_csum_type bch2_csum_opt_to_type(enum bch_csum_opts type,
+static inline enum bch_csum_type bch2_csum_opt_to_type(enum bch_csum_opt type,
 						       bool data)
 {
 	switch (type) {
diff --git a/fs/bcachefs/compress.c b/fs/bcachefs/compress.c
index 1410365a8891..f99ff1819597 100644
--- a/fs/bcachefs/compress.c
+++ b/fs/bcachefs/compress.c
@@ -2,13 +2,33 @@
 #include "bcachefs.h"
 #include "checksum.h"
 #include "compress.h"
+#include "error.h"
 #include "extents.h"
+#include "opts.h"
 #include "super-io.h"
 
 #include <linux/lz4.h>
 #include <linux/zlib.h>
 #include <linux/zstd.h>
 
+static inline enum bch_compression_opts bch2_compression_type_to_opt(enum bch_compression_type type)
+{
+	switch (type) {
+	case BCH_COMPRESSION_TYPE_none:
+	case BCH_COMPRESSION_TYPE_incompressible:
+		return BCH_COMPRESSION_OPT_none;
+	case BCH_COMPRESSION_TYPE_lz4_old:
+	case BCH_COMPRESSION_TYPE_lz4:
+		return BCH_COMPRESSION_OPT_lz4;
+	case BCH_COMPRESSION_TYPE_gzip:
+		return BCH_COMPRESSION_OPT_gzip;
+	case BCH_COMPRESSION_TYPE_zstd:
+		return BCH_COMPRESSION_OPT_zstd;
+	default:
+		BUG();
+	}
+}
+
 /* Bounce buffer: */
 struct bbuf {
 	void		*b;
@@ -158,6 +178,19 @@ static int __bio_uncompress(struct bch_fs *c, struct bio *src,
 	void *workspace;
 	int ret;
 
+	enum bch_compression_opts opt = bch2_compression_type_to_opt(crc.compression_type);
+	mempool_t *workspace_pool = &c->compress_workspace[opt];
+	if (unlikely(!mempool_initialized(workspace_pool))) {
+		if (fsck_err(c, compression_type_not_marked_in_sb,
+			     "compression type %s set but not marked in superblock",
+			     __bch2_compression_types[crc.compression_type]))
+			ret = bch2_check_set_has_compressed_data(c, opt);
+		else
+			ret = -BCH_ERR_compression_workspace_not_initialized;
+		if (ret)
+			goto out;
+	}
+
 	src_data = bio_map_or_bounce(c, src, READ);
 
 	switch (crc.compression_type) {
@@ -176,13 +209,13 @@ static int __bio_uncompress(struct bch_fs *c, struct bio *src,
 			.avail_out	= dst_len,
 		};
 
-		workspace = mempool_alloc(&c->decompress_workspace, GFP_NOFS);
+		workspace = mempool_alloc(workspace_pool, GFP_NOFS);
 
 		zlib_set_workspace(&strm, workspace);
 		zlib_inflateInit2(&strm, -MAX_WBITS);
 		ret = zlib_inflate(&strm, Z_FINISH);
 
-		mempool_free(workspace, &c->decompress_workspace);
+		mempool_free(workspace, workspace_pool);
 
 		if (ret != Z_STREAM_END)
 			goto err;
@@ -195,14 +228,14 @@ static int __bio_uncompress(struct bch_fs *c, struct bio *src,
 		if (real_src_len > src_len - 4)
 			goto err;
 
-		workspace = mempool_alloc(&c->decompress_workspace, GFP_NOFS);
+		workspace = mempool_alloc(workspace_pool, GFP_NOFS);
 		ctx = zstd_init_dctx(workspace, zstd_dctx_workspace_bound());
 
 		ret = zstd_decompress_dctx(ctx,
 				dst_data,	dst_len,
 				src_data.b + 4, real_src_len);
 
-		mempool_free(workspace, &c->decompress_workspace);
+		mempool_free(workspace, workspace_pool);
 
 		if (ret != dst_len)
 			goto err;
@@ -212,6 +245,7 @@ static int __bio_uncompress(struct bch_fs *c, struct bio *src,
 		BUG();
 	}
 	ret = 0;
+fsck_err:
 out:
 	bio_unmap_or_unbounce(c, src_data);
 	return ret;
@@ -394,8 +428,21 @@ static unsigned __bio_compress(struct bch_fs *c,
 	unsigned pad;
 	int ret = 0;
 
-	BUG_ON(compression_type >= BCH_COMPRESSION_TYPE_NR);
-	BUG_ON(!mempool_initialized(&c->compress_workspace[compression_type]));
+	/* bch2_compression_decode catches unknown compression types: */
+	BUG_ON(compression.type >= BCH_COMPRESSION_OPT_NR);
+
+	mempool_t *workspace_pool = &c->compress_workspace[compression.type];
+	if (unlikely(!mempool_initialized(workspace_pool))) {
+		if (fsck_err(c, compression_opt_not_marked_in_sb,
+			     "compression opt %s set but not marked in superblock",
+			     bch2_compression_opts[compression.type])) {
+			ret = bch2_check_set_has_compressed_data(c, compression.type);
+			if (ret) /* memory allocation failure, don't compress */
+				return 0;
+		} else {
+			return 0;
+		}
+	}
 
 	/* If it's only one block, don't bother trying to compress: */
 	if (src->bi_iter.bi_size <= c->opts.block_size)
@@ -404,7 +451,7 @@ static unsigned __bio_compress(struct bch_fs *c,
 	dst_data = bio_map_or_bounce(c, dst, WRITE);
 	src_data = bio_map_or_bounce(c, src, READ);
 
-	workspace = mempool_alloc(&c->compress_workspace[compression_type], GFP_NOFS);
+	workspace = mempool_alloc(workspace_pool, GFP_NOFS);
 
 	*src_len = src->bi_iter.bi_size;
 	*dst_len = dst->bi_iter.bi_size;
@@ -447,7 +494,7 @@ static unsigned __bio_compress(struct bch_fs *c,
 		*src_len = round_down(*src_len, block_bytes(c));
 	}
 
-	mempool_free(workspace, &c->compress_workspace[compression_type]);
+	mempool_free(workspace, workspace_pool);
 
 	if (ret)
 		goto err;
@@ -477,6 +524,9 @@ out:
 err:
 	ret = BCH_COMPRESSION_TYPE_incompressible;
 	goto out;
+fsck_err:
+	ret = 0;
+	goto out;
 }
 
 unsigned bch2_bio_compress(struct bch_fs *c,
@@ -559,7 +609,6 @@ void bch2_fs_compress_exit(struct bch_fs *c)
 {
 	unsigned i;
 
-	mempool_exit(&c->decompress_workspace);
 	for (i = 0; i < ARRAY_SIZE(c->compress_workspace); i++)
 		mempool_exit(&c->compress_workspace[i]);
 	mempool_exit(&c->compression_bounce[WRITE]);
@@ -568,7 +617,6 @@ void bch2_fs_compress_exit(struct bch_fs *c)
 
 static int __bch2_fs_compress_init(struct bch_fs *c, u64 features)
 {
-	size_t decompress_workspace_size = 0;
 	ZSTD_parameters params = zstd_get_params(zstd_max_clevel(),
 						 c->opts.encoded_extent_max);
 
@@ -576,19 +624,17 @@ static int __bch2_fs_compress_init(struct bch_fs *c, u64 features)
 
 	struct {
 		unsigned			feature;
-		enum bch_compression_type	type;
+		enum bch_compression_opts	type;
 		size_t				compress_workspace;
-		size_t				decompress_workspace;
 	} compression_types[] = {
-		{ BCH_FEATURE_lz4, BCH_COMPRESSION_TYPE_lz4,
-			max_t(size_t, LZ4_MEM_COMPRESS, LZ4HC_MEM_COMPRESS),
-			0 },
-		{ BCH_FEATURE_gzip, BCH_COMPRESSION_TYPE_gzip,
-			zlib_deflate_workspacesize(MAX_WBITS, DEF_MEM_LEVEL),
-			zlib_inflate_workspacesize(), },
-		{ BCH_FEATURE_zstd, BCH_COMPRESSION_TYPE_zstd,
-			c->zstd_workspace_size,
-			zstd_dctx_workspace_bound() },
+		{ BCH_FEATURE_lz4, BCH_COMPRESSION_OPT_lz4,
+			max_t(size_t, LZ4_MEM_COMPRESS, LZ4HC_MEM_COMPRESS) },
+		{ BCH_FEATURE_gzip, BCH_COMPRESSION_OPT_gzip,
+			max(zlib_deflate_workspacesize(MAX_WBITS, DEF_MEM_LEVEL),
+			    zlib_inflate_workspacesize()) },
+		{ BCH_FEATURE_zstd, BCH_COMPRESSION_OPT_zstd,
+			max(c->zstd_workspace_size,
+			    zstd_dctx_workspace_bound()) },
 	}, *i;
 	bool have_compressed = false;
 
@@ -613,9 +659,6 @@ static int __bch2_fs_compress_init(struct bch_fs *c, u64 features)
 	for (i = compression_types;
 	     i < compression_types + ARRAY_SIZE(compression_types);
 	     i++) {
-		decompress_workspace_size =
-			max(decompress_workspace_size, i->decompress_workspace);
-
 		if (!(features & (1 << i->feature)))
 			continue;
 
@@ -628,11 +671,6 @@ static int __bch2_fs_compress_init(struct bch_fs *c, u64 features)
 			return -BCH_ERR_ENOMEM_compression_workspace_init;
 	}
 
-	if (!mempool_initialized(&c->decompress_workspace) &&
-	    mempool_init_kvmalloc_pool(&c->decompress_workspace,
-				       1, decompress_workspace_size))
-		return -BCH_ERR_ENOMEM_decompression_workspace_init;
-
 	return 0;
 }
 
diff --git a/fs/bcachefs/darray.h b/fs/bcachefs/darray.h
index 8f4c3f0665c4..c6151495985f 100644
--- a/fs/bcachefs/darray.h
+++ b/fs/bcachefs/darray.h
@@ -83,7 +83,7 @@ int __bch2_darray_resize_noprof(darray_char *, size_t, size_t, gfp_t);
 	for (typeof(&(_d).data[0]) _i = (_d).data; _i < (_d).data + (_d).nr; _i++)
 
 #define darray_for_each_reverse(_d, _i)					\
-	for (typeof(&(_d).data[0]) _i = (_d).data + (_d).nr - 1; _i >= (_d).data; --_i)
+	for (typeof(&(_d).data[0]) _i = (_d).data + (_d).nr - 1; _i >= (_d).data && (_d).nr; --_i)
 
 #define darray_init(_d)							\
 do {									\
diff --git a/fs/bcachefs/data_update.c b/fs/bcachefs/data_update.c
index 8e75a852b358..585214931e05 100644
--- a/fs/bcachefs/data_update.c
+++ b/fs/bcachefs/data_update.c
@@ -110,11 +110,8 @@ static void trace_move_extent_fail2(struct data_update *m,
 {
 	struct bch_fs *c = m->op.c;
 	struct bkey_s_c old = bkey_i_to_s_c(m->k.k);
-	const union bch_extent_entry *entry;
-	struct bch_extent_ptr *ptr;
-	struct extent_ptr_decoded p;
 	struct printbuf buf = PRINTBUF;
-	unsigned i, rewrites_found = 0;
+	unsigned rewrites_found = 0;
 
 	if (!trace_move_extent_fail_enabled())
 		return;
@@ -122,27 +119,25 @@ static void trace_move_extent_fail2(struct data_update *m,
 	prt_str(&buf, msg);
 
 	if (insert) {
-		i = 0;
+		const union bch_extent_entry *entry;
+		struct bch_extent_ptr *ptr;
+		struct extent_ptr_decoded p;
+
+		unsigned ptr_bit = 1;
 		bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry) {
-			if (((1U << i) & m->data_opts.rewrite_ptrs) &&
+			if ((ptr_bit & m->data_opts.rewrite_ptrs) &&
 			    (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
 			    !ptr->cached)
-				rewrites_found |= 1U << i;
-			i++;
+				rewrites_found |= ptr_bit;
+			ptr_bit <<= 1;
 		}
 	}
 
-	prt_printf(&buf, "\nrewrite ptrs:   %u%u%u%u",
-		   (m->data_opts.rewrite_ptrs & (1 << 0)) != 0,
-		   (m->data_opts.rewrite_ptrs & (1 << 1)) != 0,
-		   (m->data_opts.rewrite_ptrs & (1 << 2)) != 0,
-		   (m->data_opts.rewrite_ptrs & (1 << 3)) != 0);
+	prt_str(&buf, "rewrites found:\t");
+	bch2_prt_u64_base2(&buf, rewrites_found);
+	prt_newline(&buf);
 
-	prt_printf(&buf, "\nrewrites found: %u%u%u%u",
-		   (rewrites_found & (1 << 0)) != 0,
-		   (rewrites_found & (1 << 1)) != 0,
-		   (rewrites_found & (1 << 2)) != 0,
-		   (rewrites_found & (1 << 3)) != 0);
+	bch2_data_update_opts_to_text(&buf, c, &m->op.opts, &m->data_opts);
 
 	prt_str(&buf, "\nold:    ");
 	bch2_bkey_val_to_text(&buf, c, old);
@@ -194,7 +189,7 @@ static int __bch2_data_update_index_update(struct btree_trans *trans,
 		struct bpos next_pos;
 		bool should_check_enospc;
 		s64 i_sectors_delta = 0, disk_sectors_delta = 0;
-		unsigned rewrites_found = 0, durability, i;
+		unsigned rewrites_found = 0, durability, ptr_bit;
 
 		bch2_trans_begin(trans);
 
@@ -231,16 +226,16 @@ static int __bch2_data_update_index_update(struct btree_trans *trans,
 		 *
 		 * Fist, drop rewrite_ptrs from @new:
 		 */
-		i = 0;
+		ptr_bit = 1;
 		bkey_for_each_ptr_decode(old.k, bch2_bkey_ptrs_c(old), p, entry_c) {
-			if (((1U << i) & m->data_opts.rewrite_ptrs) &&
+			if ((ptr_bit & m->data_opts.rewrite_ptrs) &&
 			    (ptr = bch2_extent_has_ptr(old, p, bkey_i_to_s(insert))) &&
 			    !ptr->cached) {
 				bch2_extent_ptr_set_cached(c, &m->op.opts,
 							   bkey_i_to_s(insert), ptr);
-				rewrites_found |= 1U << i;
+				rewrites_found |= ptr_bit;
 			}
-			i++;
+			ptr_bit <<= 1;
 		}
 
 		if (m->data_opts.rewrite_ptrs &&
@@ -323,8 +318,11 @@ restart_drop_extra_replicas:
 		 * it's been hard to reproduce, so this should give us some more
 		 * information when it does occur:
 		 */
-		int invalid = bch2_bkey_validate(c, bkey_i_to_s_c(insert), __btree_node_type(0, m->btree_id),
-						 BCH_VALIDATE_commit);
+		int invalid = bch2_bkey_validate(c, bkey_i_to_s_c(insert),
+						 (struct bkey_validate_context) {
+							.btree	= m->btree_id,
+							.flags	= BCH_VALIDATE_commit,
+						 });
 		if (invalid) {
 			struct printbuf buf = PRINTBUF;
 
@@ -362,7 +360,7 @@ restart_drop_extra_replicas:
 						k.k->p, bkey_start_pos(&insert->k)) ?:
 			bch2_insert_snapshot_whiteouts(trans, m->btree_id,
 						k.k->p, insert->k.p) ?:
-			bch2_bkey_set_needs_rebalance(c, insert, &op->opts) ?:
+			bch2_bkey_set_needs_rebalance(c, &op->opts, insert) ?:
 			bch2_trans_update(trans, &iter, insert,
 				BTREE_UPDATE_internal_snapshot_node) ?:
 			bch2_trans_commit(trans, &op->res,
@@ -540,7 +538,7 @@ void bch2_data_update_opts_to_text(struct printbuf *out, struct bch_fs *c,
 	prt_newline(out);
 
 	prt_str(out, "compression:\t");
-	bch2_compression_opt_to_text(out, background_compression(*io_opts));
+	bch2_compression_opt_to_text(out, io_opts->background_compression);
 	prt_newline(out);
 
 	prt_str(out, "opts.replicas:\t");
@@ -614,7 +612,7 @@ int bch2_data_update_init(struct btree_trans *trans,
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 	const union bch_extent_entry *entry;
 	struct extent_ptr_decoded p;
-	unsigned i, reserve_sectors = k.k->size * data_opts.extra_replicas;
+	unsigned reserve_sectors = k.k->size * data_opts.extra_replicas;
 	int ret = 0;
 
 	/*
@@ -622,7 +620,7 @@ int bch2_data_update_init(struct btree_trans *trans,
 	 * and we have to check for this because we go rw before repairing the
 	 * snapshots table - just skip it, we can move it later.
 	 */
-	if (unlikely(k.k->p.snapshot && !bch2_snapshot_equiv(c, k.k->p.snapshot)))
+	if (unlikely(k.k->p.snapshot && !bch2_snapshot_exists(c, k.k->p.snapshot)))
 		return -BCH_ERR_data_update_done;
 
 	if (!bkey_get_dev_refs(c, k))
@@ -652,22 +650,22 @@ int bch2_data_update_init(struct btree_trans *trans,
 		BCH_WRITE_DATA_ENCODED|
 		BCH_WRITE_MOVE|
 		m->data_opts.write_flags;
-	m->op.compression_opt	= background_compression(io_opts);
+	m->op.compression_opt	= io_opts.background_compression;
 	m->op.watermark		= m->data_opts.btree_insert_flags & BCH_WATERMARK_MASK;
 
 	unsigned durability_have = 0, durability_removing = 0;
 
-	i = 0;
+	unsigned ptr_bit = 1;
 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 		if (!p.ptr.cached) {
 			rcu_read_lock();
-			if (BIT(i) & m->data_opts.rewrite_ptrs) {
+			if (ptr_bit & m->data_opts.rewrite_ptrs) {
 				if (crc_is_compressed(p.crc))
 					reserve_sectors += k.k->size;
 
 				m->op.nr_replicas += bch2_extent_ptr_desired_durability(c, &p);
 				durability_removing += bch2_extent_ptr_desired_durability(c, &p);
-			} else if (!(BIT(i) & m->data_opts.kill_ptrs)) {
+			} else if (!(ptr_bit & m->data_opts.kill_ptrs)) {
 				bch2_dev_list_add_dev(&m->op.devs_have, p.ptr.dev);
 				durability_have += bch2_extent_ptr_durability(c, &p);
 			}
@@ -687,7 +685,7 @@ int bch2_data_update_init(struct btree_trans *trans,
 		if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
 			m->op.incompressible = true;
 
-		i++;
+		ptr_bit <<= 1;
 	}
 
 	unsigned durability_required = max(0, (int) (io_opts.data_replicas - durability_have));
@@ -750,14 +748,14 @@ out:
 void bch2_data_update_opts_normalize(struct bkey_s_c k, struct data_update_opts *opts)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	unsigned i = 0;
+	unsigned ptr_bit = 1;
 
 	bkey_for_each_ptr(ptrs, ptr) {
-		if ((opts->rewrite_ptrs & (1U << i)) && ptr->cached) {
-			opts->kill_ptrs |= 1U << i;
-			opts->rewrite_ptrs ^= 1U << i;
+		if ((opts->rewrite_ptrs & ptr_bit) && ptr->cached) {
+			opts->kill_ptrs |= ptr_bit;
+			opts->rewrite_ptrs ^= ptr_bit;
 		}
 
-		i++;
+		ptr_bit <<= 1;
 	}
 }
diff --git a/fs/bcachefs/debug.c b/fs/bcachefs/debug.c
index 45aec1afdb0e..b5de52a50d10 100644
--- a/fs/bcachefs/debug.c
+++ b/fs/bcachefs/debug.c
@@ -472,7 +472,9 @@ static void bch2_cached_btree_node_to_text(struct printbuf *out, struct bch_fs *
 	if (!out->nr_tabstops)
 		printbuf_tabstop_push(out, 32);
 
-	prt_printf(out, "%px btree=%s l=%u\n", b, bch2_btree_id_str(b->c.btree_id), b->c.level);
+	prt_printf(out, "%px ", b);
+	bch2_btree_id_level_to_text(out, b->c.btree_id, b->c.level);
+	prt_printf(out, "\n");
 
 	printbuf_indent_add(out, 2);
 
diff --git a/fs/bcachefs/dirent.c b/fs/bcachefs/dirent.c
index faffc98d5605..600eee936f13 100644
--- a/fs/bcachefs/dirent.c
+++ b/fs/bcachefs/dirent.c
@@ -101,7 +101,7 @@ const struct bch_hash_desc bch2_dirent_hash_desc = {
 };
 
 int bch2_dirent_validate(struct bch_fs *c, struct bkey_s_c k,
-			 enum bch_validate_flags flags)
+			 struct bkey_validate_context from)
 {
 	struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
 	struct qstr d_name = bch2_dirent_get_name(d);
@@ -120,7 +120,7 @@ int bch2_dirent_validate(struct bch_fs *c, struct bkey_s_c k,
 	 * Check new keys don't exceed the max length
 	 * (older keys may be larger.)
 	 */
-	bkey_fsck_err_on((flags & BCH_VALIDATE_commit) && d_name.len > BCH_NAME_MAX,
+	bkey_fsck_err_on((from.flags & BCH_VALIDATE_commit) && d_name.len > BCH_NAME_MAX,
 			 c, dirent_name_too_long,
 			 "dirent name too big (%u > %u)",
 			 d_name.len, BCH_NAME_MAX);
@@ -266,7 +266,7 @@ int bch2_dirent_read_target(struct btree_trans *trans, subvol_inum dir,
 	} else {
 		target->subvol	= le32_to_cpu(d.v->d_child_subvol);
 
-		ret = bch2_subvolume_get(trans, target->subvol, true, BTREE_ITER_cached, &s);
+		ret = bch2_subvolume_get(trans, target->subvol, true, &s);
 
 		target->inum	= le64_to_cpu(s.inode);
 	}
@@ -500,7 +500,7 @@ int bch2_empty_dir_snapshot(struct btree_trans *trans, u64 dir, u32 subvol, u32
 	struct bkey_s_c k;
 	int ret;
 
-	for_each_btree_key_upto_norestart(trans, iter, BTREE_ID_dirents,
+	for_each_btree_key_max_norestart(trans, iter, BTREE_ID_dirents,
 			   SPOS(dir, 0, snapshot),
 			   POS(dir, U64_MAX), 0, k, ret)
 		if (k.k->type == KEY_TYPE_dirent) {
@@ -549,7 +549,7 @@ int bch2_readdir(struct bch_fs *c, subvol_inum inum, struct dir_context *ctx)
 	bch2_bkey_buf_init(&sk);
 
 	int ret = bch2_trans_run(c,
-		for_each_btree_key_in_subvolume_upto(trans, iter, BTREE_ID_dirents,
+		for_each_btree_key_in_subvolume_max(trans, iter, BTREE_ID_dirents,
 				   POS(inum.inum, ctx->pos),
 				   POS(inum.inum, U64_MAX),
 				   inum.subvol, 0, k, ({
diff --git a/fs/bcachefs/dirent.h b/fs/bcachefs/dirent.h
index 53ad99666022..a633f83c1ac7 100644
--- a/fs/bcachefs/dirent.h
+++ b/fs/bcachefs/dirent.h
@@ -4,10 +4,10 @@
 
 #include "str_hash.h"
 
-enum bch_validate_flags;
 extern const struct bch_hash_desc bch2_dirent_hash_desc;
 
-int bch2_dirent_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_dirent_validate(struct bch_fs *, struct bkey_s_c,
+			 struct bkey_validate_context);
 void bch2_dirent_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
 #define bch2_bkey_ops_dirent ((struct bkey_ops) {	\
@@ -31,6 +31,11 @@ static inline unsigned dirent_val_u64s(unsigned len)
 			    sizeof(u64));
 }
 
+static inline unsigned int dirent_occupied_size(const struct qstr *name)
+{
+	return (BKEY_U64s + dirent_val_u64s(name->len)) * sizeof(u64);
+}
+
 int bch2_dirent_read_target(struct btree_trans *, subvol_inum,
 			    struct bkey_s_c_dirent, subvol_inum *);
 
diff --git a/fs/bcachefs/disk_accounting.c b/fs/bcachefs/disk_accounting.c
index 07eb8fa1b026..b32e91ba8be8 100644
--- a/fs/bcachefs/disk_accounting.c
+++ b/fs/bcachefs/disk_accounting.c
@@ -79,6 +79,8 @@ static inline void accounting_key_init(struct bkey_i *k, struct disk_accounting_
 	memcpy_u64s_small(acc->v.d, d, nr);
 }
 
+static int bch2_accounting_update_sb_one(struct bch_fs *, struct bpos);
+
 int bch2_disk_accounting_mod(struct btree_trans *trans,
 			     struct disk_accounting_pos *k,
 			     s64 *d, unsigned nr, bool gc)
@@ -96,9 +98,16 @@ int bch2_disk_accounting_mod(struct btree_trans *trans,
 
 	accounting_key_init(&k_i.k, k, d, nr);
 
-	return likely(!gc)
-		? bch2_trans_update_buffered(trans, BTREE_ID_accounting, &k_i.k)
-		: bch2_accounting_mem_add(trans, bkey_i_to_s_c_accounting(&k_i.k), true);
+	if (unlikely(gc)) {
+		int ret = bch2_accounting_mem_add(trans, bkey_i_to_s_c_accounting(&k_i.k), true);
+		if (ret == -BCH_ERR_btree_insert_need_mark_replicas)
+			ret = drop_locks_do(trans,
+				bch2_accounting_update_sb_one(trans->c, disk_accounting_pos_to_bpos(k))) ?:
+				bch2_accounting_mem_add(trans, bkey_i_to_s_c_accounting(&k_i.k), true);
+		return ret;
+	} else {
+		return bch2_trans_update_buffered(trans, BTREE_ID_accounting, &k_i.k);
+	}
 }
 
 int bch2_mod_dev_cached_sectors(struct btree_trans *trans,
@@ -127,14 +136,15 @@ static inline bool is_zero(char *start, char *end)
 #define field_end(p, member)	(((void *) (&p.member)) + sizeof(p.member))
 
 int bch2_accounting_validate(struct bch_fs *c, struct bkey_s_c k,
-			     enum bch_validate_flags flags)
+			     struct bkey_validate_context from)
 {
 	struct disk_accounting_pos acc_k;
 	bpos_to_disk_accounting_pos(&acc_k, k.k->p);
 	void *end = &acc_k + 1;
 	int ret = 0;
 
-	bkey_fsck_err_on(bversion_zero(k.k->bversion),
+	bkey_fsck_err_on((from.flags & BCH_VALIDATE_commit) &&
+			 bversion_zero(k.k->bversion),
 			 c, accounting_key_version_0,
 			 "accounting key with version=0");
 
@@ -217,7 +227,8 @@ void bch2_accounting_key_to_text(struct printbuf *out, struct disk_accounting_po
 		prt_printf(out, "id=%u", k->snapshot.id);
 		break;
 	case BCH_DISK_ACCOUNTING_btree:
-		prt_printf(out, "btree=%s", bch2_btree_id_str(k->btree.id));
+		prt_str(out, "btree=");
+		bch2_btree_id_to_text(out, k->btree.id);
 		break;
 	}
 }
@@ -243,10 +254,10 @@ void bch2_accounting_swab(struct bkey_s k)
 }
 
 static inline void __accounting_to_replicas(struct bch_replicas_entry_v1 *r,
-					    struct disk_accounting_pos acc)
+					    struct disk_accounting_pos *acc)
 {
-	unsafe_memcpy(r, &acc.replicas,
-		      replicas_entry_bytes(&acc.replicas),
+	unsafe_memcpy(r, &acc->replicas,
+		      replicas_entry_bytes(&acc->replicas),
 		      "variable length struct");
 }
 
@@ -257,7 +268,7 @@ static inline bool accounting_to_replicas(struct bch_replicas_entry_v1 *r, struc
 
 	switch (acc_k.type) {
 	case BCH_DISK_ACCOUNTING_replicas:
-		__accounting_to_replicas(r, acc_k);
+		__accounting_to_replicas(r, &acc_k);
 		return true;
 	default:
 		return false;
@@ -322,6 +333,14 @@ static int __bch2_accounting_mem_insert(struct bch_fs *c, struct bkey_s_c_accoun
 
 	eytzinger0_sort(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
 			accounting_pos_cmp, NULL);
+
+	if (trace_accounting_mem_insert_enabled()) {
+		struct printbuf buf = PRINTBUF;
+
+		bch2_accounting_to_text(&buf, c, a.s_c);
+		trace_accounting_mem_insert(c, buf.buf);
+		printbuf_exit(&buf);
+	}
 	return 0;
 err:
 	free_percpu(n.v[1]);
@@ -461,32 +480,6 @@ int bch2_fs_accounting_read(struct bch_fs *c, darray_char *out_buf, unsigned acc
 	return ret;
 }
 
-void bch2_fs_accounting_to_text(struct printbuf *out, struct bch_fs *c)
-{
-	struct bch_accounting_mem *acc = &c->accounting;
-
-	percpu_down_read(&c->mark_lock);
-	out->atomic++;
-
-	eytzinger0_for_each(i, acc->k.nr) {
-		struct disk_accounting_pos acc_k;
-		bpos_to_disk_accounting_pos(&acc_k, acc->k.data[i].pos);
-
-		bch2_accounting_key_to_text(out, &acc_k);
-
-		u64 v[BCH_ACCOUNTING_MAX_COUNTERS];
-		bch2_accounting_mem_read_counters(acc, i, v, ARRAY_SIZE(v), false);
-
-		prt_str(out, ":");
-		for (unsigned j = 0; j < acc->k.data[i].nr_counters; j++)
-			prt_printf(out, " %llu", v[j]);
-		prt_newline(out);
-	}
-
-	--out->atomic;
-	percpu_up_read(&c->mark_lock);
-}
-
 static void bch2_accounting_free_counters(struct bch_accounting_mem *acc, bool gc)
 {
 	darray_for_each(acc->k, e) {
@@ -625,7 +618,7 @@ static int bch2_disk_accounting_validate_late(struct btree_trans *trans,
 	switch (acc.type) {
 	case BCH_DISK_ACCOUNTING_replicas: {
 		struct bch_replicas_padded r;
-		__accounting_to_replicas(&r.e, acc);
+		__accounting_to_replicas(&r.e, &acc);
 
 		for (unsigned i = 0; i < r.e.nr_devs; i++)
 			if (r.e.devs[i] != BCH_SB_MEMBER_INVALID &&
@@ -699,11 +692,45 @@ int bch2_accounting_read(struct bch_fs *c)
 	struct btree_trans *trans = bch2_trans_get(c);
 	struct printbuf buf = PRINTBUF;
 
-	int ret = for_each_btree_key(trans, iter,
-				BTREE_ID_accounting, POS_MIN,
+	/*
+	 * We might run more than once if we rewind to start topology repair or
+	 * btree node scan - and those might cause us to get different results,
+	 * so we can't just skip if we've already run.
+	 *
+	 * Instead, zero out any accounting we have:
+	 */
+	percpu_down_write(&c->mark_lock);
+	darray_for_each(acc->k, e)
+		percpu_memset(e->v[0], 0, sizeof(u64) * e->nr_counters);
+	for_each_member_device(c, ca)
+		percpu_memset(ca->usage, 0, sizeof(*ca->usage));
+	percpu_memset(c->usage, 0, sizeof(*c->usage));
+	percpu_up_write(&c->mark_lock);
+
+	struct btree_iter iter;
+	bch2_trans_iter_init(trans, &iter, BTREE_ID_accounting, POS_MIN,
+			     BTREE_ITER_prefetch|BTREE_ITER_all_snapshots);
+	iter.flags &= ~BTREE_ITER_with_journal;
+	int ret = for_each_btree_key_continue(trans, iter,
 				BTREE_ITER_prefetch|BTREE_ITER_all_snapshots, k, ({
 			struct bkey u;
 			struct bkey_s_c k = bch2_btree_path_peek_slot_exact(btree_iter_path(trans, &iter), &u);
+
+			if (k.k->type != KEY_TYPE_accounting)
+				continue;
+
+			struct disk_accounting_pos acc_k;
+			bpos_to_disk_accounting_pos(&acc_k, k.k->p);
+
+			if (acc_k.type >= BCH_DISK_ACCOUNTING_TYPE_NR)
+				break;
+
+			if (!bch2_accounting_is_mem(acc_k)) {
+				struct disk_accounting_pos next = { .type = acc_k.type + 1 };
+				bch2_btree_iter_set_pos(&iter, disk_accounting_pos_to_bpos(&next));
+				continue;
+			}
+
 			accounting_read_key(trans, k);
 		}));
 	if (ret)
@@ -715,6 +742,12 @@ int bch2_accounting_read(struct bch_fs *c)
 
 	darray_for_each(*keys, i) {
 		if (i->k->k.type == KEY_TYPE_accounting) {
+			struct disk_accounting_pos acc_k;
+			bpos_to_disk_accounting_pos(&acc_k, i->k->k.p);
+
+			if (!bch2_accounting_is_mem(acc_k))
+				continue;
+
 			struct bkey_s_c k = bkey_i_to_s_c(i->k);
 			unsigned idx = eytzinger0_find(acc->k.data, acc->k.nr,
 						sizeof(acc->k.data[0]),
@@ -748,15 +781,16 @@ int bch2_accounting_read(struct bch_fs *c)
 	keys->gap = keys->nr = dst - keys->data;
 
 	percpu_down_write(&c->mark_lock);
-	unsigned i = 0;
-	while (i < acc->k.nr) {
-		unsigned idx = inorder_to_eytzinger0(i, acc->k.nr);
 
+	darray_for_each_reverse(acc->k, i) {
 		struct disk_accounting_pos acc_k;
-		bpos_to_disk_accounting_pos(&acc_k, acc->k.data[idx].pos);
+		bpos_to_disk_accounting_pos(&acc_k, i->pos);
 
 		u64 v[BCH_ACCOUNTING_MAX_COUNTERS];
-		bch2_accounting_mem_read_counters(acc, idx, v, ARRAY_SIZE(v), false);
+		memset(v, 0, sizeof(v));
+
+		for (unsigned j = 0; j < i->nr_counters; j++)
+			v[j] = percpu_u64_get(i->v[0] + j);
 
 		/*
 		 * If the entry counters are zeroed, it should be treated as
@@ -765,26 +799,25 @@ int bch2_accounting_read(struct bch_fs *c)
 		 * Remove it, so that if it's re-added it gets re-marked in the
 		 * superblock:
 		 */
-		ret = bch2_is_zero(v, sizeof(v[0]) * acc->k.data[idx].nr_counters)
+		ret = bch2_is_zero(v, sizeof(v[0]) * i->nr_counters)
 			? -BCH_ERR_remove_disk_accounting_entry
-			: bch2_disk_accounting_validate_late(trans, acc_k,
-							v, acc->k.data[idx].nr_counters);
+			: bch2_disk_accounting_validate_late(trans, acc_k, v, i->nr_counters);
 
 		if (ret == -BCH_ERR_remove_disk_accounting_entry) {
-			free_percpu(acc->k.data[idx].v[0]);
-			free_percpu(acc->k.data[idx].v[1]);
-			darray_remove_item(&acc->k, &acc->k.data[idx]);
-			eytzinger0_sort(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
-					accounting_pos_cmp, NULL);
+			free_percpu(i->v[0]);
+			free_percpu(i->v[1]);
+			darray_remove_item(&acc->k, i);
 			ret = 0;
 			continue;
 		}
 
 		if (ret)
 			goto fsck_err;
-		i++;
 	}
 
+	eytzinger0_sort(acc->k.data, acc->k.nr, sizeof(acc->k.data[0]),
+			accounting_pos_cmp, NULL);
+
 	preempt_disable();
 	struct bch_fs_usage_base *usage = this_cpu_ptr(c->usage);
 
@@ -804,7 +837,7 @@ int bch2_accounting_read(struct bch_fs *c)
 			break;
 		case BCH_DISK_ACCOUNTING_dev_data_type:
 			rcu_read_lock();
-			struct bch_dev *ca = bch2_dev_rcu(c, k.dev_data_type.dev);
+			struct bch_dev *ca = bch2_dev_rcu_noerror(c, k.dev_data_type.dev);
 			if (ca) {
 				struct bch_dev_usage_type __percpu *d = &ca->usage->d[k.dev_data_type.data_type];
 				percpu_u64_set(&d->buckets,	v[0]);
@@ -881,10 +914,13 @@ void bch2_verify_accounting_clean(struct bch_fs *c)
 			bpos_to_disk_accounting_pos(&acc_k, k.k->p);
 
 			if (acc_k.type >= BCH_DISK_ACCOUNTING_TYPE_NR)
-				continue;
+				break;
 
-			if (acc_k.type == BCH_DISK_ACCOUNTING_inum)
+			if (!bch2_accounting_is_mem(acc_k)) {
+				struct disk_accounting_pos next = { .type = acc_k.type + 1 };
+				bch2_btree_iter_set_pos(&iter, disk_accounting_pos_to_bpos(&next));
 				continue;
+			}
 
 			bch2_accounting_mem_read(c, k.k->p, v, nr);
 
@@ -910,7 +946,7 @@ void bch2_verify_accounting_clean(struct bch_fs *c)
 				break;
 			case BCH_DISK_ACCOUNTING_dev_data_type: {
 				rcu_read_lock();
-				struct bch_dev *ca = bch2_dev_rcu(c, acc_k.dev_data_type.dev);
+				struct bch_dev *ca = bch2_dev_rcu_noerror(c, acc_k.dev_data_type.dev);
 				if (!ca) {
 					rcu_read_unlock();
 					continue;
diff --git a/fs/bcachefs/disk_accounting.h b/fs/bcachefs/disk_accounting.h
index 4ea6c8a092bc..5360cbb3ec29 100644
--- a/fs/bcachefs/disk_accounting.h
+++ b/fs/bcachefs/disk_accounting.h
@@ -2,6 +2,7 @@
 #ifndef _BCACHEFS_DISK_ACCOUNTING_H
 #define _BCACHEFS_DISK_ACCOUNTING_H
 
+#include "btree_update.h"
 #include "eytzinger.h"
 #include "sb-members.h"
 
@@ -62,27 +63,32 @@ static inline void fs_usage_data_type_to_base(struct bch_fs_usage_base *fs_usage
 
 static inline void bpos_to_disk_accounting_pos(struct disk_accounting_pos *acc, struct bpos p)
 {
-	acc->_pad = p;
+	BUILD_BUG_ON(sizeof(*acc) != sizeof(p));
+
 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
-	bch2_bpos_swab(&acc->_pad);
+	acc->_pad = p;
+#else
+	memcpy_swab(acc, &p, sizeof(p));
 #endif
 }
 
-static inline struct bpos disk_accounting_pos_to_bpos(struct disk_accounting_pos *k)
+static inline struct bpos disk_accounting_pos_to_bpos(struct disk_accounting_pos *acc)
 {
-	struct bpos ret = k->_pad;
-
+	struct bpos p;
 #if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
-	bch2_bpos_swab(&ret);
+	p = acc->_pad;
+#else
+	memcpy_swab(&p, acc, sizeof(p));
 #endif
-	return ret;
+	return p;
 }
 
 int bch2_disk_accounting_mod(struct btree_trans *, struct disk_accounting_pos *,
 			     s64 *, unsigned, bool);
 int bch2_mod_dev_cached_sectors(struct btree_trans *, unsigned, s64, bool);
 
-int bch2_accounting_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_accounting_validate(struct bch_fs *, struct bkey_s_c,
+			     struct bkey_validate_context);
 void bch2_accounting_key_to_text(struct printbuf *, struct disk_accounting_pos *);
 void bch2_accounting_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 void bch2_accounting_swab(struct bkey_s);
@@ -112,6 +118,12 @@ enum bch_accounting_mode {
 int bch2_accounting_mem_insert(struct bch_fs *, struct bkey_s_c_accounting, enum bch_accounting_mode);
 void bch2_accounting_mem_gc(struct bch_fs *);
 
+static inline bool bch2_accounting_is_mem(struct disk_accounting_pos acc)
+{
+	return acc.type < BCH_DISK_ACCOUNTING_TYPE_NR &&
+		acc.type != BCH_DISK_ACCOUNTING_inum;
+}
+
 /*
  * Update in memory counters so they match the btree update we're doing; called
  * from transaction commit path
@@ -126,9 +138,10 @@ static inline int bch2_accounting_mem_mod_locked(struct btree_trans *trans,
 	bpos_to_disk_accounting_pos(&acc_k, a.k->p);
 	bool gc = mode == BCH_ACCOUNTING_gc;
 
-	EBUG_ON(gc && !acc->gc_running);
+	if (gc && !acc->gc_running)
+		return 0;
 
-	if (acc_k.type == BCH_DISK_ACCOUNTING_inum)
+	if (!bch2_accounting_is_mem(acc_k))
 		return 0;
 
 	if (mode == BCH_ACCOUNTING_normal) {
@@ -141,7 +154,7 @@ static inline int bch2_accounting_mem_mod_locked(struct btree_trans *trans,
 			break;
 		case BCH_DISK_ACCOUNTING_dev_data_type:
 			rcu_read_lock();
-			struct bch_dev *ca = bch2_dev_rcu(c, acc_k.dev_data_type.dev);
+			struct bch_dev *ca = bch2_dev_rcu_noerror(c, acc_k.dev_data_type.dev);
 			if (ca) {
 				this_cpu_add(ca->usage->d[acc_k.dev_data_type.data_type].buckets, a.v->d[0]);
 				this_cpu_add(ca->usage->d[acc_k.dev_data_type.data_type].sectors, a.v->d[1]);
@@ -204,9 +217,45 @@ static inline void bch2_accounting_mem_read(struct bch_fs *c, struct bpos p,
 	bch2_accounting_mem_read_counters(acc, idx, v, nr, false);
 }
 
+static inline struct bversion journal_pos_to_bversion(struct journal_res *res, unsigned offset)
+{
+	EBUG_ON(!res->ref);
+
+	return (struct bversion) {
+		.hi = res->seq >> 32,
+		.lo = (res->seq << 32) | (res->offset + offset),
+	};
+}
+
+static inline int bch2_accounting_trans_commit_hook(struct btree_trans *trans,
+						    struct bkey_i_accounting *a,
+						    unsigned commit_flags)
+{
+	a->k.bversion = journal_pos_to_bversion(&trans->journal_res,
+						(u64 *) a - (u64 *) trans->journal_entries);
+
+	EBUG_ON(bversion_zero(a->k.bversion));
+
+	return likely(!(commit_flags & BCH_TRANS_COMMIT_skip_accounting_apply))
+		? bch2_accounting_mem_mod_locked(trans, accounting_i_to_s_c(a), BCH_ACCOUNTING_normal)
+		: 0;
+}
+
+static inline void bch2_accounting_trans_commit_revert(struct btree_trans *trans,
+						       struct bkey_i_accounting *a_i,
+						       unsigned commit_flags)
+{
+	if (likely(!(commit_flags & BCH_TRANS_COMMIT_skip_accounting_apply))) {
+		struct bkey_s_accounting a = accounting_i_to_s(a_i);
+
+		bch2_accounting_neg(a);
+		bch2_accounting_mem_mod_locked(trans, a.c, BCH_ACCOUNTING_normal);
+		bch2_accounting_neg(a);
+	}
+}
+
 int bch2_fs_replicas_usage_read(struct bch_fs *, darray_char *);
 int bch2_fs_accounting_read(struct bch_fs *, darray_char *, unsigned);
-void bch2_fs_accounting_to_text(struct printbuf *, struct bch_fs *);
 
 int bch2_gc_accounting_start(struct bch_fs *);
 int bch2_gc_accounting_done(struct bch_fs *);
diff --git a/fs/bcachefs/ec.c b/fs/bcachefs/ec.c
index 6094afb0c6be..d2a5e76e6479 100644
--- a/fs/bcachefs/ec.c
+++ b/fs/bcachefs/ec.c
@@ -26,6 +26,7 @@
 #include "util.h"
 
 #include <linux/sort.h>
+#include <linux/string_choices.h>
 
 #ifdef __KERNEL__
 
@@ -109,7 +110,7 @@ struct ec_bio {
 /* Stripes btree keys: */
 
 int bch2_stripe_validate(struct bch_fs *c, struct bkey_s_c k,
-			 enum bch_validate_flags flags)
+			 struct bkey_validate_context from)
 {
 	const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 	int ret = 0;
@@ -129,7 +130,7 @@ int bch2_stripe_validate(struct bch_fs *c, struct bkey_s_c k,
 			 "invalid csum granularity (%u >= 64)",
 			 s->csum_granularity_bits);
 
-	ret = bch2_bkey_ptrs_validate(c, k, flags);
+	ret = bch2_bkey_ptrs_validate(c, k, from);
 fsck_err:
 	return ret;
 }
@@ -304,13 +305,12 @@ static int mark_stripe_bucket(struct btree_trans *trans,
 	}
 
 	if (flags & BTREE_TRIGGER_gc) {
-		percpu_down_read(&c->mark_lock);
 		struct bucket *g = gc_bucket(ca, bucket.offset);
 		if (bch2_fs_inconsistent_on(!g, c, "reference to invalid bucket on device %u\n  %s",
 					    ptr->dev,
 					    (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
 			ret = -BCH_ERR_mark_stripe;
-			goto err_unlock;
+			goto err;
 		}
 
 		bucket_lock(g);
@@ -318,8 +318,7 @@ static int mark_stripe_bucket(struct btree_trans *trans,
 		ret = __mark_stripe_bucket(trans, ca, s, ptr_idx, deleting, bucket, &new, flags);
 		alloc_to_bucket(g, new);
 		bucket_unlock(g);
-err_unlock:
-		percpu_up_read(&c->mark_lock);
+
 		if (!ret)
 			ret = bch2_alloc_key_to_dev_counters(trans, ca, &old, &new, flags);
 	}
@@ -732,7 +731,7 @@ static void ec_block_endio(struct bio *bio)
 			       ? BCH_MEMBER_ERROR_write
 			       : BCH_MEMBER_ERROR_read,
 			       "erasure coding %s error: %s",
-			       bio_data_dir(bio) ? "write" : "read",
+			       str_write_read(bio_data_dir(bio)),
 			       bch2_blk_status_to_str(bio->bi_status)))
 		clear_bit(ec_bio->idx, ec_bio->buf->valid);
 
@@ -909,7 +908,7 @@ err:
 	bch2_bkey_val_to_text(&msgbuf, c, orig_k);
 	bch_err_ratelimited(c,
 			    "error doing reconstruct read: %s\n  %s", msg, msgbuf.buf);
-	printbuf_exit(&msgbuf);;
+	printbuf_exit(&msgbuf);
 	ret = -BCH_ERR_stripe_reconstruct;
 	goto out;
 }
@@ -1266,11 +1265,11 @@ static int ec_stripe_update_extent(struct btree_trans *trans,
 				   struct bch_dev *ca,
 				   struct bpos bucket, u8 gen,
 				   struct ec_stripe_buf *s,
-				   struct bpos *bp_pos)
+				   struct bkey_s_c_backpointer bp,
+				   struct bkey_buf *last_flushed)
 {
 	struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
 	struct bch_fs *c = trans->c;
-	struct bch_backpointer bp;
 	struct btree_iter iter;
 	struct bkey_s_c k;
 	const struct bch_extent_ptr *ptr_c;
@@ -1279,33 +1278,26 @@ static int ec_stripe_update_extent(struct btree_trans *trans,
 	struct bkey_i *n;
 	int ret, dev, block;
 
-	ret = bch2_get_next_backpointer(trans, ca, bucket, gen,
-				bp_pos, &bp, BTREE_ITER_cached);
-	if (ret)
-		return ret;
-	if (bpos_eq(*bp_pos, SPOS_MAX))
-		return 0;
-
-	if (bp.level) {
+	if (bp.v->level) {
 		struct printbuf buf = PRINTBUF;
 		struct btree_iter node_iter;
 		struct btree *b;
 
-		b = bch2_backpointer_get_node(trans, &node_iter, *bp_pos, bp);
+		b = bch2_backpointer_get_node(trans, bp, &node_iter, last_flushed);
 		bch2_trans_iter_exit(trans, &node_iter);
 
 		if (!b)
 			return 0;
 
 		prt_printf(&buf, "found btree node in erasure coded bucket: b=%px\n", b);
-		bch2_backpointer_to_text(&buf, &bp);
+		bch2_bkey_val_to_text(&buf, c, bp.s_c);
 
 		bch2_fs_inconsistent(c, "%s", buf.buf);
 		printbuf_exit(&buf);
 		return -EIO;
 	}
 
-	k = bch2_backpointer_get_key(trans, &iter, *bp_pos, bp, BTREE_ITER_intent);
+	k = bch2_backpointer_get_key(trans, bp, &iter, BTREE_ITER_intent, last_flushed);
 	ret = bkey_err(k);
 	if (ret)
 		return ret;
@@ -1364,7 +1356,6 @@ static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_b
 	struct bch_fs *c = trans->c;
 	struct bch_stripe *v = &bkey_i_to_stripe(&s->key)->v;
 	struct bch_extent_ptr ptr = v->ptrs[block];
-	struct bpos bp_pos = POS_MIN;
 	int ret = 0;
 
 	struct bch_dev *ca = bch2_dev_tryget(c, ptr.dev);
@@ -1373,19 +1364,27 @@ static int ec_stripe_update_bucket(struct btree_trans *trans, struct ec_stripe_b
 
 	struct bpos bucket_pos = PTR_BUCKET_POS(ca, &ptr);
 
-	while (1) {
-		ret = commit_do(trans, NULL, NULL,
-				BCH_TRANS_COMMIT_no_check_rw|
-				BCH_TRANS_COMMIT_no_enospc,
-			ec_stripe_update_extent(trans, ca, bucket_pos, ptr.gen, s, &bp_pos));
-		if (ret)
-			break;
-		if (bkey_eq(bp_pos, POS_MAX))
+	struct bkey_buf last_flushed;
+	bch2_bkey_buf_init(&last_flushed);
+	bkey_init(&last_flushed.k->k);
+
+	ret = for_each_btree_key_max_commit(trans, bp_iter, BTREE_ID_backpointers,
+			bucket_pos_to_bp_start(ca, bucket_pos),
+			bucket_pos_to_bp_end(ca, bucket_pos), 0, bp_k,
+			NULL, NULL,
+			BCH_TRANS_COMMIT_no_check_rw|
+			BCH_TRANS_COMMIT_no_enospc, ({
+		if (bkey_ge(bp_k.k->p, bucket_pos_to_bp(ca, bpos_nosnap_successor(bucket_pos), 0)))
 			break;
 
-		bp_pos = bpos_nosnap_successor(bp_pos);
-	}
+		if (bp_k.k->type != KEY_TYPE_backpointer)
+			continue;
 
+		ec_stripe_update_extent(trans, ca, bucket_pos, ptr.gen, s,
+					bkey_s_c_to_backpointer(bp_k), &last_flushed);
+	}));
+
+	bch2_bkey_buf_exit(&last_flushed, c);
 	bch2_dev_put(ca);
 	return ret;
 }
@@ -1707,7 +1706,7 @@ static void ec_stripe_key_init(struct bch_fs *c,
 	set_bkey_val_u64s(&s->k, u64s);
 }
 
-static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
+static struct ec_stripe_new *ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
 {
 	struct ec_stripe_new *s;
 
@@ -1715,7 +1714,7 @@ static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
 
 	s = kzalloc(sizeof(*s), GFP_KERNEL);
 	if (!s)
-		return -BCH_ERR_ENOMEM_ec_new_stripe_alloc;
+		return NULL;
 
 	mutex_init(&s->lock);
 	closure_init(&s->iodone, NULL);
@@ -1730,10 +1729,7 @@ static int ec_new_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
 	ec_stripe_key_init(c, &s->new_stripe.key,
 			   s->nr_data, s->nr_parity,
 			   h->blocksize, h->disk_label);
-
-	h->s = s;
-	h->nr_created++;
-	return 0;
+	return s;
 }
 
 static void ec_stripe_head_devs_update(struct bch_fs *c, struct ec_stripe_head *h)
@@ -1878,25 +1874,26 @@ err:
 	return h;
 }
 
-static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_head *h,
+static int new_stripe_alloc_buckets(struct btree_trans *trans,
+				    struct ec_stripe_head *h, struct ec_stripe_new *s,
 				    enum bch_watermark watermark, struct closure *cl)
 {
 	struct bch_fs *c = trans->c;
 	struct bch_devs_mask devs = h->devs;
 	struct open_bucket *ob;
 	struct open_buckets buckets;
-	struct bch_stripe *v = &bkey_i_to_stripe(&h->s->new_stripe.key)->v;
+	struct bch_stripe *v = &bkey_i_to_stripe(&s->new_stripe.key)->v;
 	unsigned i, j, nr_have_parity = 0, nr_have_data = 0;
 	bool have_cache = true;
 	int ret = 0;
 
-	BUG_ON(v->nr_blocks	!= h->s->nr_data + h->s->nr_parity);
-	BUG_ON(v->nr_redundant	!= h->s->nr_parity);
+	BUG_ON(v->nr_blocks	!= s->nr_data + s->nr_parity);
+	BUG_ON(v->nr_redundant	!= s->nr_parity);
 
 	/* * We bypass the sector allocator which normally does this: */
 	bitmap_and(devs.d, devs.d, c->rw_devs[BCH_DATA_user].d, BCH_SB_MEMBERS_MAX);
 
-	for_each_set_bit(i, h->s->blocks_gotten, v->nr_blocks) {
+	for_each_set_bit(i, s->blocks_gotten, v->nr_blocks) {
 		/*
 		 * Note: we don't yet repair invalid blocks (failed/removed
 		 * devices) when reusing stripes - we still need a codepath to
@@ -1906,21 +1903,21 @@ static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_
 		if (v->ptrs[i].dev != BCH_SB_MEMBER_INVALID)
 			__clear_bit(v->ptrs[i].dev, devs.d);
 
-		if (i < h->s->nr_data)
+		if (i < s->nr_data)
 			nr_have_data++;
 		else
 			nr_have_parity++;
 	}
 
-	BUG_ON(nr_have_data	> h->s->nr_data);
-	BUG_ON(nr_have_parity	> h->s->nr_parity);
+	BUG_ON(nr_have_data	> s->nr_data);
+	BUG_ON(nr_have_parity	> s->nr_parity);
 
 	buckets.nr = 0;
-	if (nr_have_parity < h->s->nr_parity) {
+	if (nr_have_parity < s->nr_parity) {
 		ret = bch2_bucket_alloc_set_trans(trans, &buckets,
 					    &h->parity_stripe,
 					    &devs,
-					    h->s->nr_parity,
+					    s->nr_parity,
 					    &nr_have_parity,
 					    &have_cache, 0,
 					    BCH_DATA_parity,
@@ -1928,14 +1925,14 @@ static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_
 					    cl);
 
 		open_bucket_for_each(c, &buckets, ob, i) {
-			j = find_next_zero_bit(h->s->blocks_gotten,
-					       h->s->nr_data + h->s->nr_parity,
-					       h->s->nr_data);
-			BUG_ON(j >= h->s->nr_data + h->s->nr_parity);
+			j = find_next_zero_bit(s->blocks_gotten,
+					       s->nr_data + s->nr_parity,
+					       s->nr_data);
+			BUG_ON(j >= s->nr_data + s->nr_parity);
 
-			h->s->blocks[j] = buckets.v[i];
+			s->blocks[j] = buckets.v[i];
 			v->ptrs[j] = bch2_ob_ptr(c, ob);
-			__set_bit(j, h->s->blocks_gotten);
+			__set_bit(j, s->blocks_gotten);
 		}
 
 		if (ret)
@@ -1943,11 +1940,11 @@ static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_
 	}
 
 	buckets.nr = 0;
-	if (nr_have_data < h->s->nr_data) {
+	if (nr_have_data < s->nr_data) {
 		ret = bch2_bucket_alloc_set_trans(trans, &buckets,
 					    &h->block_stripe,
 					    &devs,
-					    h->s->nr_data,
+					    s->nr_data,
 					    &nr_have_data,
 					    &have_cache, 0,
 					    BCH_DATA_user,
@@ -1955,13 +1952,13 @@ static int new_stripe_alloc_buckets(struct btree_trans *trans, struct ec_stripe_
 					    cl);
 
 		open_bucket_for_each(c, &buckets, ob, i) {
-			j = find_next_zero_bit(h->s->blocks_gotten,
-					       h->s->nr_data, 0);
-			BUG_ON(j >= h->s->nr_data);
+			j = find_next_zero_bit(s->blocks_gotten,
+					       s->nr_data, 0);
+			BUG_ON(j >= s->nr_data);
 
-			h->s->blocks[j] = buckets.v[i];
+			s->blocks[j] = buckets.v[i];
 			v->ptrs[j] = bch2_ob_ptr(c, ob);
-			__set_bit(j, h->s->blocks_gotten);
+			__set_bit(j, s->blocks_gotten);
 		}
 
 		if (ret)
@@ -2007,73 +2004,78 @@ static s64 get_existing_stripe(struct bch_fs *c,
 	return ret;
 }
 
-static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h)
+static int init_new_stripe_from_existing(struct bch_fs *c, struct ec_stripe_new *s)
 {
-	struct bch_fs *c = trans->c;
-	struct bch_stripe *new_v = &bkey_i_to_stripe(&h->s->new_stripe.key)->v;
-	struct bch_stripe *existing_v;
+	struct bch_stripe *new_v = &bkey_i_to_stripe(&s->new_stripe.key)->v;
+	struct bch_stripe *existing_v = &bkey_i_to_stripe(&s->existing_stripe.key)->v;
 	unsigned i;
-	s64 idx;
-	int ret;
-
-	/*
-	 * If we can't allocate a new stripe, and there's no stripes with empty
-	 * blocks for us to reuse, that means we have to wait on copygc:
-	 */
-	idx = get_existing_stripe(c, h);
-	if (idx < 0)
-		return -BCH_ERR_stripe_alloc_blocked;
-
-	ret = get_stripe_key_trans(trans, idx, &h->s->existing_stripe);
-	bch2_fs_fatal_err_on(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart), c,
-			     "reading stripe key: %s", bch2_err_str(ret));
-	if (ret) {
-		bch2_stripe_close(c, h->s);
-		return ret;
-	}
 
-	existing_v = &bkey_i_to_stripe(&h->s->existing_stripe.key)->v;
-
-	BUG_ON(existing_v->nr_redundant != h->s->nr_parity);
-	h->s->nr_data = existing_v->nr_blocks -
+	BUG_ON(existing_v->nr_redundant != s->nr_parity);
+	s->nr_data = existing_v->nr_blocks -
 		existing_v->nr_redundant;
 
-	ret = ec_stripe_buf_init(&h->s->existing_stripe, 0, h->blocksize);
+	int ret = ec_stripe_buf_init(&s->existing_stripe, 0, le16_to_cpu(existing_v->sectors));
 	if (ret) {
-		bch2_stripe_close(c, h->s);
+		bch2_stripe_close(c, s);
 		return ret;
 	}
 
-	BUG_ON(h->s->existing_stripe.size != h->blocksize);
-	BUG_ON(h->s->existing_stripe.size != le16_to_cpu(existing_v->sectors));
+	BUG_ON(s->existing_stripe.size != le16_to_cpu(existing_v->sectors));
 
 	/*
 	 * Free buckets we initially allocated - they might conflict with
 	 * blocks from the stripe we're reusing:
 	 */
-	for_each_set_bit(i, h->s->blocks_gotten, new_v->nr_blocks) {
-		bch2_open_bucket_put(c, c->open_buckets + h->s->blocks[i]);
-		h->s->blocks[i] = 0;
+	for_each_set_bit(i, s->blocks_gotten, new_v->nr_blocks) {
+		bch2_open_bucket_put(c, c->open_buckets + s->blocks[i]);
+		s->blocks[i] = 0;
 	}
-	memset(h->s->blocks_gotten, 0, sizeof(h->s->blocks_gotten));
-	memset(h->s->blocks_allocated, 0, sizeof(h->s->blocks_allocated));
+	memset(s->blocks_gotten, 0, sizeof(s->blocks_gotten));
+	memset(s->blocks_allocated, 0, sizeof(s->blocks_allocated));
 
-	for (i = 0; i < existing_v->nr_blocks; i++) {
+	for (unsigned i = 0; i < existing_v->nr_blocks; i++) {
 		if (stripe_blockcount_get(existing_v, i)) {
-			__set_bit(i, h->s->blocks_gotten);
-			__set_bit(i, h->s->blocks_allocated);
+			__set_bit(i, s->blocks_gotten);
+			__set_bit(i, s->blocks_allocated);
 		}
 
-		ec_block_io(c, &h->s->existing_stripe, READ, i, &h->s->iodone);
+		ec_block_io(c, &s->existing_stripe, READ, i, &s->iodone);
 	}
 
-	bkey_copy(&h->s->new_stripe.key, &h->s->existing_stripe.key);
-	h->s->have_existing_stripe = true;
+	bkey_copy(&s->new_stripe.key, &s->existing_stripe.key);
+	s->have_existing_stripe = true;
 
 	return 0;
 }
 
-static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h)
+static int __bch2_ec_stripe_head_reuse(struct btree_trans *trans, struct ec_stripe_head *h,
+				       struct ec_stripe_new *s)
+{
+	struct bch_fs *c = trans->c;
+	s64 idx;
+	int ret;
+
+	/*
+	 * If we can't allocate a new stripe, and there's no stripes with empty
+	 * blocks for us to reuse, that means we have to wait on copygc:
+	 */
+	idx = get_existing_stripe(c, h);
+	if (idx < 0)
+		return -BCH_ERR_stripe_alloc_blocked;
+
+	ret = get_stripe_key_trans(trans, idx, &s->existing_stripe);
+	bch2_fs_fatal_err_on(ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart), c,
+			     "reading stripe key: %s", bch2_err_str(ret));
+	if (ret) {
+		bch2_stripe_close(c, s);
+		return ret;
+	}
+
+	return init_new_stripe_from_existing(c, s);
+}
+
+static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_stripe_head *h,
+					 struct ec_stripe_new *s)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
@@ -2082,15 +2084,19 @@ static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_st
 	struct bpos start_pos = bpos_max(min_pos, POS(0, c->ec_stripe_hint));
 	int ret;
 
-	if (!h->s->res.sectors) {
-		ret = bch2_disk_reservation_get(c, &h->s->res,
+	if (!s->res.sectors) {
+		ret = bch2_disk_reservation_get(c, &s->res,
 					h->blocksize,
-					h->s->nr_parity,
+					s->nr_parity,
 					BCH_DISK_RESERVATION_NOFAIL);
 		if (ret)
 			return ret;
 	}
 
+	/*
+	 * Allocate stripe slot
+	 * XXX: we're going to need a bitrange btree of free stripes
+	 */
 	for_each_btree_key_norestart(trans, iter, BTREE_ID_stripes, start_pos,
 			   BTREE_ITER_slots|BTREE_ITER_intent, k, ret) {
 		if (bkey_gt(k.k->p, POS(0, U32_MAX))) {
@@ -2105,7 +2111,7 @@ static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_st
 		}
 
 		if (bkey_deleted(k.k) &&
-		    bch2_try_open_stripe(c, h->s, k.k->p.offset))
+		    bch2_try_open_stripe(c, s, k.k->p.offset))
 			break;
 	}
 
@@ -2116,16 +2122,16 @@ static int __bch2_ec_stripe_head_reserve(struct btree_trans *trans, struct ec_st
 
 	ret = ec_stripe_mem_alloc(trans, &iter);
 	if (ret) {
-		bch2_stripe_close(c, h->s);
+		bch2_stripe_close(c, s);
 		goto err;
 	}
 
-	h->s->new_stripe.key.k.p = iter.pos;
+	s->new_stripe.key.k.p = iter.pos;
 out:
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 err:
-	bch2_disk_reservation_put(c, &h->s->res);
+	bch2_disk_reservation_put(c, &s->res);
 	goto out;
 }
 
@@ -2156,22 +2162,27 @@ struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
 		return h;
 
 	if (!h->s) {
-		ret = ec_new_stripe_alloc(c, h);
-		if (ret) {
+		h->s = ec_new_stripe_alloc(c, h);
+		if (!h->s) {
+			ret = -BCH_ERR_ENOMEM_ec_new_stripe_alloc;
 			bch_err(c, "failed to allocate new stripe");
 			goto err;
 		}
+
+		h->nr_created++;
 	}
 
-	if (h->s->allocated)
+	struct ec_stripe_new *s = h->s;
+
+	if (s->allocated)
 		goto allocated;
 
-	if (h->s->have_existing_stripe)
+	if (s->have_existing_stripe)
 		goto alloc_existing;
 
 	/* First, try to allocate a full stripe: */
-	ret =   new_stripe_alloc_buckets(trans, h, BCH_WATERMARK_stripe, NULL) ?:
-		__bch2_ec_stripe_head_reserve(trans, h);
+	ret =   new_stripe_alloc_buckets(trans, h, s, BCH_WATERMARK_stripe, NULL) ?:
+		__bch2_ec_stripe_head_reserve(trans, h, s);
 	if (!ret)
 		goto allocate_buf;
 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
@@ -2183,15 +2194,15 @@ struct ec_stripe_head *bch2_ec_stripe_head_get(struct btree_trans *trans,
 	 * existing stripe:
 	 */
 	while (1) {
-		ret = __bch2_ec_stripe_head_reuse(trans, h);
+		ret = __bch2_ec_stripe_head_reuse(trans, h, s);
 		if (!ret)
 			break;
 		if (waiting || !cl || ret != -BCH_ERR_stripe_alloc_blocked)
 			goto err;
 
 		if (watermark == BCH_WATERMARK_copygc) {
-			ret =   new_stripe_alloc_buckets(trans, h, watermark, NULL) ?:
-				__bch2_ec_stripe_head_reserve(trans, h);
+			ret =   new_stripe_alloc_buckets(trans, h, s, watermark, NULL) ?:
+				__bch2_ec_stripe_head_reserve(trans, h, s);
 			if (ret)
 				goto err;
 			goto allocate_buf;
@@ -2209,19 +2220,19 @@ alloc_existing:
 	 * Retry allocating buckets, with the watermark for this
 	 * particular write:
 	 */
-	ret = new_stripe_alloc_buckets(trans, h, watermark, cl);
+	ret = new_stripe_alloc_buckets(trans, h, s, watermark, cl);
 	if (ret)
 		goto err;
 
 allocate_buf:
-	ret = ec_stripe_buf_init(&h->s->new_stripe, 0, h->blocksize);
+	ret = ec_stripe_buf_init(&s->new_stripe, 0, h->blocksize);
 	if (ret)
 		goto err;
 
-	h->s->allocated = true;
+	s->allocated = true;
 allocated:
-	BUG_ON(!h->s->idx);
-	BUG_ON(!h->s->new_stripe.data[0]);
+	BUG_ON(!s->idx);
+	BUG_ON(!s->new_stripe.data[0]);
 	BUG_ON(trans->restarted);
 	return h;
 err:
@@ -2286,7 +2297,7 @@ err:
 int bch2_dev_remove_stripes(struct bch_fs *c, unsigned dev_idx)
 {
 	return bch2_trans_run(c,
-		for_each_btree_key_upto_commit(trans, iter,
+		for_each_btree_key_max_commit(trans, iter,
 				  BTREE_ID_alloc, POS(dev_idx, 0), POS(dev_idx, U64_MAX),
 				  BTREE_ITER_intent, k,
 				  NULL, NULL, 0, ({
@@ -2449,11 +2460,9 @@ void bch2_fs_ec_exit(struct bch_fs *c)
 
 	while (1) {
 		mutex_lock(&c->ec_stripe_head_lock);
-		h = list_first_entry_or_null(&c->ec_stripe_head_list,
-					     struct ec_stripe_head, list);
-		if (h)
-			list_del(&h->list);
+		h = list_pop_entry(&c->ec_stripe_head_list, struct ec_stripe_head, list);
 		mutex_unlock(&c->ec_stripe_head_lock);
+
 		if (!h)
 			break;
 
diff --git a/fs/bcachefs/ec.h b/fs/bcachefs/ec.h
index 43326370b410..583ca6a226da 100644
--- a/fs/bcachefs/ec.h
+++ b/fs/bcachefs/ec.h
@@ -6,9 +6,8 @@
 #include "buckets_types.h"
 #include "extents_types.h"
 
-enum bch_validate_flags;
-
-int bch2_stripe_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_stripe_validate(struct bch_fs *, struct bkey_s_c,
+			 struct bkey_validate_context);
 void bch2_stripe_to_text(struct printbuf *, struct bch_fs *,
 			 struct bkey_s_c);
 int bch2_trigger_stripe(struct btree_trans *, enum btree_id, unsigned,
diff --git a/fs/bcachefs/ec_format.h b/fs/bcachefs/ec_format.h
index 64ef52e00078..b9770f24f213 100644
--- a/fs/bcachefs/ec_format.h
+++ b/fs/bcachefs/ec_format.h
@@ -20,6 +20,23 @@ struct bch_stripe {
 	 */
 	__u8			disk_label;
 
+	/*
+	 * Variable length sections:
+	 * - Pointers
+	 * - Checksums
+	 *   2D array of [stripe block/device][csum block], with checksum block
+	 *   size given by csum_granularity_bits
+	 * - Block sector counts: per-block array of u16s
+	 *
+	 * XXX:
+	 * Either checksums should have come last, or we should have included a
+	 * checksum_size field (the size in bytes of the checksum itself, not
+	 * the blocksize the checksum covers).
+	 *
+	 * Currently we aren't able to access the block sector counts if the
+	 * checksum type is unknown.
+	 */
+
 	struct bch_extent_ptr	ptrs[];
 } __packed __aligned(8);
 
diff --git a/fs/bcachefs/errcode.h b/fs/bcachefs/errcode.h
index 9c4fe5cdbfb7..4590cd0c7c90 100644
--- a/fs/bcachefs/errcode.h
+++ b/fs/bcachefs/errcode.h
@@ -54,7 +54,8 @@
 	x(ENOMEM,			ENOMEM_compression_bounce_read_init)	\
 	x(ENOMEM,			ENOMEM_compression_bounce_write_init)	\
 	x(ENOMEM,			ENOMEM_compression_workspace_init)	\
-	x(ENOMEM,			ENOMEM_decompression_workspace_init)	\
+	x(ENOMEM,			ENOMEM_backpointer_mismatches_bitmap)	\
+	x(EIO,				compression_workspace_not_initialized)	\
 	x(ENOMEM,			ENOMEM_bucket_gens)			\
 	x(ENOMEM,			ENOMEM_buckets_nouse)			\
 	x(ENOMEM,			ENOMEM_usage_init)			\
@@ -116,6 +117,8 @@
 	x(ENOENT,			ENOENT_dirent_doesnt_match_inode)	\
 	x(ENOENT,			ENOENT_dev_not_found)			\
 	x(ENOENT,			ENOENT_dev_idx_not_found)		\
+	x(ENOENT,			ENOENT_inode_no_backpointer)		\
+	x(ENOENT,			ENOENT_no_snapshot_tree_subvol)		\
 	x(ENOTEMPTY,			ENOTEMPTY_dir_not_empty)		\
 	x(ENOTEMPTY,			ENOTEMPTY_subvol_not_empty)		\
 	x(EEXIST,			EEXIST_str_hash_set)			\
@@ -148,6 +151,7 @@
 	x(BCH_ERR_transaction_restart,	transaction_restart_split_race)		\
 	x(BCH_ERR_transaction_restart,	transaction_restart_write_buffer_flush)	\
 	x(BCH_ERR_transaction_restart,	transaction_restart_nested)		\
+	x(BCH_ERR_transaction_restart,	transaction_restart_commit)		\
 	x(0,				no_btree_node)				\
 	x(BCH_ERR_no_btree_node,	no_btree_node_relock)			\
 	x(BCH_ERR_no_btree_node,	no_btree_node_upgrade)			\
@@ -164,7 +168,6 @@
 	x(BCH_ERR_btree_insert_fail,	btree_insert_need_journal_res)		\
 	x(BCH_ERR_btree_insert_fail,	btree_insert_need_journal_reclaim)	\
 	x(0,				backpointer_to_overwritten_btree_node)	\
-	x(0,				lock_fail_root_changed)			\
 	x(0,				journal_reclaim_would_deadlock)		\
 	x(EINVAL,			fsck)					\
 	x(BCH_ERR_fsck,			fsck_fix)				\
@@ -173,7 +176,9 @@
 	x(BCH_ERR_fsck,			fsck_errors_not_fixed)			\
 	x(BCH_ERR_fsck,			fsck_repair_unimplemented)		\
 	x(BCH_ERR_fsck,			fsck_repair_impossible)			\
-	x(0,				restart_recovery)			\
+	x(EINVAL,			restart_recovery)			\
+	x(EINVAL,			not_in_recovery)			\
+	x(EINVAL,			cannot_rewind_recovery)			\
 	x(0,				data_update_done)			\
 	x(EINVAL,			device_state_not_allowed)		\
 	x(EINVAL,			member_info_missing)			\
@@ -192,7 +197,9 @@
 	x(EINVAL,			opt_parse_error)			\
 	x(EINVAL,			remove_with_metadata_missing_unimplemented)\
 	x(EINVAL,			remove_would_lose_data)			\
-	x(EINVAL,			btree_iter_with_journal_not_supported)	\
+	x(EINVAL,			no_resize_with_buckets_nouse)		\
+	x(EINVAL,			inode_unpack_error)			\
+	x(EINVAL,			varint_decode_error)			\
 	x(EROFS,			erofs_trans_commit)			\
 	x(EROFS,			erofs_no_writes)			\
 	x(EROFS,			erofs_journal_err)			\
@@ -241,7 +248,10 @@
 	x(BCH_ERR_invalid_sb,		invalid_sb_downgrade)			\
 	x(BCH_ERR_invalid,		invalid_bkey)				\
 	x(BCH_ERR_operation_blocked,    nocow_lock_blocked)			\
+	x(EIO,				journal_shutdown)			\
+	x(EIO,				journal_flush_err)			\
 	x(EIO,				btree_node_read_err)			\
+	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_cached)		\
 	x(EIO,				sb_not_downgraded)			\
 	x(EIO,				btree_node_write_all_failed)		\
 	x(EIO,				btree_node_read_error)			\
@@ -257,6 +267,8 @@
 	x(EIO,				no_device_to_read_from)			\
 	x(EIO,				missing_indirect_extent)		\
 	x(EIO,				invalidate_stripe_to_dev)		\
+	x(EIO,				no_encryption_key)			\
+	x(EIO,				insufficient_journal_devices)		\
 	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_fixable)		\
 	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_want_retry)		\
 	x(BCH_ERR_btree_node_read_err,	btree_node_read_err_must_retry)		\
@@ -305,6 +317,7 @@ static inline long bch2_err_class(long err)
 
 #define BLK_STS_REMOVED		((__force blk_status_t)128)
 
+#include <linux/blk_types.h>
 const char *bch2_blk_status_to_str(blk_status_t);
 
 #endif /* _BCACHFES_ERRCODE_H */
diff --git a/fs/bcachefs/error.c b/fs/bcachefs/error.c
index b679def8fb98..038da6a61f6b 100644
--- a/fs/bcachefs/error.c
+++ b/fs/bcachefs/error.c
@@ -1,7 +1,9 @@
 // SPDX-License-Identifier: GPL-2.0
 #include "bcachefs.h"
+#include "btree_cache.h"
 #include "btree_iter.h"
 #include "error.h"
+#include "fs-common.h"
 #include "journal.h"
 #include "recovery_passes.h"
 #include "super.h"
@@ -33,7 +35,7 @@ bool bch2_inconsistent_error(struct bch_fs *c)
 int bch2_topology_error(struct bch_fs *c)
 {
 	set_bit(BCH_FS_topology_error, &c->flags);
-	if (!test_bit(BCH_FS_fsck_running, &c->flags)) {
+	if (!test_bit(BCH_FS_recovery_running, &c->flags)) {
 		bch2_inconsistent_error(c);
 		return -BCH_ERR_btree_need_topology_repair;
 	} else {
@@ -218,6 +220,30 @@ static const u8 fsck_flags_extra[] = {
 #undef x
 };
 
+static int do_fsck_ask_yn(struct bch_fs *c,
+			  struct btree_trans *trans,
+			  struct printbuf *question,
+			  const char *action)
+{
+	prt_str(question, ", ");
+	prt_str(question, action);
+
+	if (bch2_fs_stdio_redirect(c))
+		bch2_print(c, "%s", question->buf);
+	else
+		bch2_print_string_as_lines(KERN_ERR, question->buf);
+
+	int ask = bch2_fsck_ask_yn(c, trans);
+
+	if (trans) {
+		int ret = bch2_trans_relock(trans);
+		if (ret)
+			return ret;
+	}
+
+	return ask;
+}
+
 int __bch2_fsck_err(struct bch_fs *c,
 		  struct btree_trans *trans,
 		  enum bch_fsck_flags flags,
@@ -226,7 +252,7 @@ int __bch2_fsck_err(struct bch_fs *c,
 {
 	struct fsck_err_state *s = NULL;
 	va_list args;
-	bool print = true, suppressing = false, inconsistent = false;
+	bool print = true, suppressing = false, inconsistent = false, exiting = false;
 	struct printbuf buf = PRINTBUF, *out = &buf;
 	int ret = -BCH_ERR_fsck_ignore;
 	const char *action_orig = "fix?", *action = action_orig;
@@ -256,9 +282,10 @@ int __bch2_fsck_err(struct bch_fs *c,
 		!trans &&
 		bch2_current_has_btree_trans(c));
 
-	if ((flags & FSCK_CAN_FIX) &&
-	    test_bit(err, c->sb.errors_silent))
-		return -BCH_ERR_fsck_fix;
+	if (test_bit(err, c->sb.errors_silent))
+		return flags & FSCK_CAN_FIX
+			? -BCH_ERR_fsck_fix
+			: -BCH_ERR_fsck_ignore;
 
 	bch2_sb_error_count(c, err);
 
@@ -289,16 +316,14 @@ int __bch2_fsck_err(struct bch_fs *c,
 		 */
 		if (s->last_msg && !strcmp(buf.buf, s->last_msg)) {
 			ret = s->ret;
-			mutex_unlock(&c->fsck_error_msgs_lock);
-			goto err;
+			goto err_unlock;
 		}
 
 		kfree(s->last_msg);
 		s->last_msg = kstrdup(buf.buf, GFP_KERNEL);
 		if (!s->last_msg) {
-			mutex_unlock(&c->fsck_error_msgs_lock);
 			ret = -ENOMEM;
-			goto err;
+			goto err_unlock;
 		}
 
 		if (c->opts.ratelimit_errors &&
@@ -318,13 +343,19 @@ int __bch2_fsck_err(struct bch_fs *c,
 		prt_printf(out, bch2_log_msg(c, ""));
 #endif
 
-	if ((flags & FSCK_CAN_FIX) &&
-	    (flags & FSCK_AUTOFIX) &&
+	if ((flags & FSCK_AUTOFIX) &&
 	    (c->opts.errors == BCH_ON_ERROR_continue ||
 	     c->opts.errors == BCH_ON_ERROR_fix_safe)) {
 		prt_str(out, ", ");
-		prt_actioning(out, action);
-		ret = -BCH_ERR_fsck_fix;
+		if (flags & FSCK_CAN_FIX) {
+			prt_actioning(out, action);
+			ret = -BCH_ERR_fsck_fix;
+		} else {
+			prt_str(out, ", continuing");
+			ret = -BCH_ERR_fsck_ignore;
+		}
+
+		goto print;
 	} else if (!test_bit(BCH_FS_fsck_running, &c->flags)) {
 		if (c->opts.errors != BCH_ON_ERROR_continue ||
 		    !(flags & (FSCK_CAN_FIX|FSCK_CAN_IGNORE))) {
@@ -348,31 +379,18 @@ int __bch2_fsck_err(struct bch_fs *c,
 			: c->opts.fix_errors;
 
 		if (fix == FSCK_FIX_ask) {
-			prt_str(out, ", ");
-			prt_str(out, action);
-
-			if (bch2_fs_stdio_redirect(c))
-				bch2_print(c, "%s", out->buf);
-			else
-				bch2_print_string_as_lines(KERN_ERR, out->buf);
 			print = false;
 
-			int ask = bch2_fsck_ask_yn(c, trans);
-
-			if (trans) {
-				ret = bch2_trans_relock(trans);
-				if (ret) {
-					mutex_unlock(&c->fsck_error_msgs_lock);
-					goto err;
-				}
-			}
+			ret = do_fsck_ask_yn(c, trans, out, action);
+			if (ret < 0)
+				goto err_unlock;
 
-			if (ask >= YN_ALLNO && s)
-				s->fix = ask == YN_ALLNO
+			if (ret >= YN_ALLNO && s)
+				s->fix = ret == YN_ALLNO
 					? FSCK_FIX_no
 					: FSCK_FIX_yes;
 
-			ret = ask & 1
+			ret = ret & 1
 				? -BCH_ERR_fsck_fix
 				: -BCH_ERR_fsck_ignore;
 		} else if (fix == FSCK_FIX_yes ||
@@ -385,9 +403,7 @@ int __bch2_fsck_err(struct bch_fs *c,
 			prt_str(out, ", not ");
 			prt_actioning(out, action);
 		}
-	} else if (flags & FSCK_NEED_FSCK) {
-		prt_str(out, " (run fsck to correct)");
-	} else {
+	} else if (!(flags & FSCK_CAN_IGNORE)) {
 		prt_str(out, " (repair unimplemented)");
 	}
 
@@ -396,14 +412,13 @@ int __bch2_fsck_err(struct bch_fs *c,
 	     !(flags & FSCK_CAN_IGNORE)))
 		ret = -BCH_ERR_fsck_errors_not_fixed;
 
-	bool exiting =
-		test_bit(BCH_FS_fsck_running, &c->flags) &&
-		(ret != -BCH_ERR_fsck_fix &&
-		 ret != -BCH_ERR_fsck_ignore);
-
-	if (exiting)
+	if (test_bit(BCH_FS_fsck_running, &c->flags) &&
+	    (ret != -BCH_ERR_fsck_fix &&
+	     ret != -BCH_ERR_fsck_ignore)) {
+		exiting = true;
 		print = true;
-
+	}
+print:
 	if (print) {
 		if (bch2_fs_stdio_redirect(c))
 			bch2_print(c, "%s\n", out->buf);
@@ -419,17 +434,24 @@ int __bch2_fsck_err(struct bch_fs *c,
 	if (s)
 		s->ret = ret;
 
-	mutex_unlock(&c->fsck_error_msgs_lock);
-
 	if (inconsistent)
 		bch2_inconsistent_error(c);
 
-	if (ret == -BCH_ERR_fsck_fix) {
-		set_bit(BCH_FS_errors_fixed, &c->flags);
-	} else {
-		set_bit(BCH_FS_errors_not_fixed, &c->flags);
-		set_bit(BCH_FS_error, &c->flags);
+	/*
+	 * We don't yet track whether the filesystem currently has errors, for
+	 * log_fsck_err()s: that would require us to track for every error type
+	 * which recovery pass corrects it, to get the fsck exit status correct:
+	 */
+	if (flags & FSCK_CAN_FIX) {
+		if (ret == -BCH_ERR_fsck_fix) {
+			set_bit(BCH_FS_errors_fixed, &c->flags);
+		} else {
+			set_bit(BCH_FS_errors_not_fixed, &c->flags);
+			set_bit(BCH_FS_error, &c->flags);
+		}
 	}
+err_unlock:
+	mutex_unlock(&c->fsck_error_msgs_lock);
 err:
 	if (action != action_orig)
 		kfree(action);
@@ -437,28 +459,52 @@ err:
 	return ret;
 }
 
+static const char * const bch2_bkey_validate_contexts[] = {
+#define x(n) #n,
+	BKEY_VALIDATE_CONTEXTS()
+#undef x
+	NULL
+};
+
 int __bch2_bkey_fsck_err(struct bch_fs *c,
 			 struct bkey_s_c k,
-			 enum bch_validate_flags validate_flags,
+			 struct bkey_validate_context from,
 			 enum bch_sb_error_id err,
 			 const char *fmt, ...)
 {
-	if (validate_flags & BCH_VALIDATE_silent)
+	if (from.flags & BCH_VALIDATE_silent)
 		return -BCH_ERR_fsck_delete_bkey;
 
 	unsigned fsck_flags = 0;
-	if (!(validate_flags & (BCH_VALIDATE_write|BCH_VALIDATE_commit)))
+	if (!(from.flags & (BCH_VALIDATE_write|BCH_VALIDATE_commit))) {
+		if (test_bit(err, c->sb.errors_silent))
+			return -BCH_ERR_fsck_delete_bkey;
+
 		fsck_flags |= FSCK_AUTOFIX|FSCK_CAN_FIX;
+	}
+	if (!WARN_ON(err >= ARRAY_SIZE(fsck_flags_extra)))
+		fsck_flags |= fsck_flags_extra[err];
 
 	struct printbuf buf = PRINTBUF;
-	va_list args;
+	prt_printf(&buf, "invalid bkey in %s",
+		   bch2_bkey_validate_contexts[from.from]);
+
+	if (from.from == BKEY_VALIDATE_journal)
+		prt_printf(&buf, " journal seq=%llu offset=%u",
+			   from.journal_seq, from.journal_offset);
+
+	prt_str(&buf, " btree=");
+	bch2_btree_id_to_text(&buf, from.btree);
+	prt_printf(&buf, " level=%u: ", from.level);
 
-	prt_str(&buf, "invalid bkey ");
 	bch2_bkey_val_to_text(&buf, c, k);
 	prt_str(&buf, "\n  ");
+
+	va_list args;
 	va_start(args, fmt);
 	prt_vprintf(&buf, fmt, args);
 	va_end(args);
+
 	prt_str(&buf, ": delete?");
 
 	int ret = __bch2_fsck_err(c, NULL, fsck_flags, err, "%s", buf.buf);
@@ -483,3 +529,36 @@ void bch2_flush_fsck_errs(struct bch_fs *c)
 
 	mutex_unlock(&c->fsck_error_msgs_lock);
 }
+
+int bch2_inum_err_msg_trans(struct btree_trans *trans, struct printbuf *out, subvol_inum inum)
+{
+	u32 restart_count = trans->restart_count;
+	int ret = 0;
+
+	/* XXX: we don't yet attempt to print paths when we don't know the subvol */
+	if (inum.subvol)
+		ret = lockrestart_do(trans, bch2_inum_to_path(trans, inum, out));
+	if (!inum.subvol || ret)
+		prt_printf(out, "inum %llu:%llu", inum.subvol, inum.inum);
+
+	return trans_was_restarted(trans, restart_count);
+}
+
+int bch2_inum_offset_err_msg_trans(struct btree_trans *trans, struct printbuf *out,
+				    subvol_inum inum, u64 offset)
+{
+	int ret = bch2_inum_err_msg_trans(trans, out, inum);
+	prt_printf(out, " offset %llu: ", offset);
+	return ret;
+}
+
+void bch2_inum_err_msg(struct bch_fs *c, struct printbuf *out, subvol_inum inum)
+{
+	bch2_trans_run(c, bch2_inum_err_msg_trans(trans, out, inum));
+}
+
+void bch2_inum_offset_err_msg(struct bch_fs *c, struct printbuf *out,
+			      subvol_inum inum, u64 offset)
+{
+	bch2_trans_run(c, bch2_inum_offset_err_msg_trans(trans, out, inum, offset));
+}
diff --git a/fs/bcachefs/error.h b/fs/bcachefs/error.h
index 6551ada926b6..7acf2a27ca28 100644
--- a/fs/bcachefs/error.h
+++ b/fs/bcachefs/error.h
@@ -45,32 +45,11 @@ int bch2_topology_error(struct bch_fs *);
 	bch2_inconsistent_error(c);					\
 })
 
-#define bch2_fs_inconsistent_on(cond, c, ...)				\
+#define bch2_fs_inconsistent_on(cond, ...)				\
 ({									\
 	bool _ret = unlikely(!!(cond));					\
-									\
-	if (_ret)							\
-		bch2_fs_inconsistent(c, __VA_ARGS__);			\
-	_ret;								\
-})
-
-/*
- * Later we might want to mark only the particular device inconsistent, not the
- * entire filesystem:
- */
-
-#define bch2_dev_inconsistent(ca, ...)					\
-do {									\
-	bch_err(ca, __VA_ARGS__);					\
-	bch2_inconsistent_error((ca)->fs);				\
-} while (0)
-
-#define bch2_dev_inconsistent_on(cond, ca, ...)				\
-({									\
-	bool _ret = unlikely(!!(cond));					\
-									\
 	if (_ret)							\
-		bch2_dev_inconsistent(ca, __VA_ARGS__);			\
+		bch2_fs_inconsistent(__VA_ARGS__);			\
 	_ret;								\
 })
 
@@ -123,9 +102,9 @@ int __bch2_fsck_err(struct bch_fs *, struct btree_trans *,
 
 void bch2_flush_fsck_errs(struct bch_fs *);
 
-#define __fsck_err(c, _flags, _err_type, ...)				\
+#define fsck_err_wrap(_do)						\
 ({									\
-	int _ret = bch2_fsck_err(c, _flags, _err_type, __VA_ARGS__);	\
+	int _ret = _do;							\
 	if (_ret != -BCH_ERR_fsck_fix &&				\
 	    _ret != -BCH_ERR_fsck_ignore) {				\
 		ret = _ret;						\
@@ -135,6 +114,8 @@ void bch2_flush_fsck_errs(struct bch_fs *);
 	_ret == -BCH_ERR_fsck_fix;					\
 })
 
+#define __fsck_err(...)		fsck_err_wrap(bch2_fsck_err(__VA_ARGS__))
+
 /* These macros return true if error should be fixed: */
 
 /* XXX: mark in superblock that filesystem contains errors, if we ignore: */
@@ -149,12 +130,6 @@ void bch2_flush_fsck_errs(struct bch_fs *);
 	(unlikely(cond) ? __fsck_err(c, _flags, _err_type, __VA_ARGS__) : false);\
 })
 
-#define need_fsck_err_on(cond, c, _err_type, ...)				\
-	__fsck_err_on(cond, c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK, _err_type, __VA_ARGS__)
-
-#define need_fsck_err(c, _err_type, ...)				\
-	__fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK, _err_type, __VA_ARGS__)
-
 #define mustfix_fsck_err(c, _err_type, ...)				\
 	__fsck_err(c, FSCK_CAN_FIX, _err_type, __VA_ARGS__)
 
@@ -167,11 +142,22 @@ void bch2_flush_fsck_errs(struct bch_fs *);
 #define fsck_err_on(cond, c, _err_type, ...)				\
 	__fsck_err_on(cond, c, FSCK_CAN_FIX|FSCK_CAN_IGNORE, _err_type, __VA_ARGS__)
 
+#define log_fsck_err(c, _err_type, ...)					\
+	__fsck_err(c, FSCK_CAN_IGNORE, _err_type, __VA_ARGS__)
+
+#define log_fsck_err_on(cond, ...)					\
+({									\
+	bool _ret = unlikely(!!(cond));					\
+	if (_ret)							\
+		log_fsck_err(__VA_ARGS__);				\
+	_ret;								\
+})
+
 enum bch_validate_flags;
 __printf(5, 6)
 int __bch2_bkey_fsck_err(struct bch_fs *,
 			 struct bkey_s_c,
-			 enum bch_validate_flags,
+			 struct bkey_validate_context from,
 			 enum bch_sb_error_id,
 			 const char *, ...);
 
@@ -181,7 +167,7 @@ int __bch2_bkey_fsck_err(struct bch_fs *,
  */
 #define bkey_fsck_err(c, _err_type, _err_msg, ...)			\
 do {									\
-	int _ret = __bch2_bkey_fsck_err(c, k, flags,			\
+	int _ret = __bch2_bkey_fsck_err(c, k, from,			\
 				BCH_FSCK_ERR_##_err_type,		\
 				_err_msg, ##__VA_ARGS__);		\
 	if (_ret != -BCH_ERR_fsck_fix &&				\
@@ -252,4 +238,10 @@ void bch2_io_error(struct bch_dev *, enum bch_member_error_type);
 	_ret;								\
 })
 
+int bch2_inum_err_msg_trans(struct btree_trans *, struct printbuf *, subvol_inum);
+int bch2_inum_offset_err_msg_trans(struct btree_trans *, struct printbuf *, subvol_inum, u64);
+
+void bch2_inum_err_msg(struct bch_fs *, struct printbuf *, subvol_inum);
+void bch2_inum_offset_err_msg(struct bch_fs *, struct printbuf *, subvol_inum, u64);
+
 #endif /* _BCACHEFS_ERROR_H */
diff --git a/fs/bcachefs/extent_update.c b/fs/bcachefs/extent_update.c
index 5f4fecb358da..6aac579a692a 100644
--- a/fs/bcachefs/extent_update.c
+++ b/fs/bcachefs/extent_update.c
@@ -64,7 +64,7 @@ static int count_iters_for_insert(struct btree_trans *trans,
 		break;
 	case KEY_TYPE_reflink_p: {
 		struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
-		u64 idx = le64_to_cpu(p.v->idx);
+		u64 idx = REFLINK_P_IDX(p.v);
 		unsigned sectors = bpos_min(*end, p.k->p).offset -
 			bkey_start_offset(p.k);
 		struct btree_iter iter;
@@ -128,7 +128,7 @@ int bch2_extent_atomic_end(struct btree_trans *trans,
 
 	bch2_trans_copy_iter(&copy, iter);
 
-	for_each_btree_key_upto_continue_norestart(copy, insert->k.p, 0, k, ret) {
+	for_each_btree_key_max_continue_norestart(copy, insert->k.p, 0, k, ret) {
 		unsigned offset = 0;
 
 		if (bkey_gt(bkey_start_pos(&insert->k), bkey_start_pos(k.k)))
diff --git a/fs/bcachefs/extents.c b/fs/bcachefs/extents.c
index 37e3d69bec06..05d5f71a7ca9 100644
--- a/fs/bcachefs/extents.c
+++ b/fs/bcachefs/extents.c
@@ -21,6 +21,7 @@
 #include "extents.h"
 #include "inode.h"
 #include "journal.h"
+#include "rebalance.h"
 #include "replicas.h"
 #include "super.h"
 #include "super-io.h"
@@ -88,6 +89,14 @@ static inline bool ptr_better(struct bch_fs *c,
 		u64 l1 = dev_latency(c, p1.ptr.dev);
 		u64 l2 = dev_latency(c, p2.ptr.dev);
 
+		/*
+		 * Square the latencies, to bias more in favor of the faster
+		 * device - we never want to stop issuing reads to the slower
+		 * device altogether, so that we can update our latency numbers:
+		 */
+		l1 *= l1;
+		l2 *= l2;
+
 		/* Pick at random, biased in favor of the faster device: */
 
 		return bch2_rand_range(l1 + l2) > l1;
@@ -169,7 +178,7 @@ int bch2_bkey_pick_read_device(struct bch_fs *c, struct bkey_s_c k,
 /* KEY_TYPE_btree_ptr: */
 
 int bch2_btree_ptr_validate(struct bch_fs *c, struct bkey_s_c k,
-			    enum bch_validate_flags flags)
+			    struct bkey_validate_context from)
 {
 	int ret = 0;
 
@@ -177,7 +186,7 @@ int bch2_btree_ptr_validate(struct bch_fs *c, struct bkey_s_c k,
 			 c, btree_ptr_val_too_big,
 			 "value too big (%zu > %u)", bkey_val_u64s(k.k), BCH_REPLICAS_MAX);
 
-	ret = bch2_bkey_ptrs_validate(c, k, flags);
+	ret = bch2_bkey_ptrs_validate(c, k, from);
 fsck_err:
 	return ret;
 }
@@ -189,7 +198,7 @@ void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 int bch2_btree_ptr_v2_validate(struct bch_fs *c, struct bkey_s_c k,
-			       enum bch_validate_flags flags)
+			       struct bkey_validate_context from)
 {
 	struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
 	int ret = 0;
@@ -203,12 +212,13 @@ int bch2_btree_ptr_v2_validate(struct bch_fs *c, struct bkey_s_c k,
 			 c, btree_ptr_v2_min_key_bad,
 			 "min_key > key");
 
-	if (flags & BCH_VALIDATE_write)
+	if ((from.flags & BCH_VALIDATE_write) &&
+	    c->sb.version_min >= bcachefs_metadata_version_btree_ptr_sectors_written)
 		bkey_fsck_err_on(!bp.v->sectors_written,
 				 c, btree_ptr_v2_written_0,
 				 "sectors_written == 0");
 
-	ret = bch2_bkey_ptrs_validate(c, k, flags);
+	ret = bch2_bkey_ptrs_validate(c, k, from);
 fsck_err:
 	return ret;
 }
@@ -395,7 +405,7 @@ bool bch2_extent_merge(struct bch_fs *c, struct bkey_s l, struct bkey_s_c r)
 /* KEY_TYPE_reservation: */
 
 int bch2_reservation_validate(struct bch_fs *c, struct bkey_s_c k,
-			      enum bch_validate_flags flags)
+			      struct bkey_validate_context from)
 {
 	struct bkey_s_c_reservation r = bkey_s_c_to_reservation(k);
 	int ret = 0;
@@ -1120,6 +1130,57 @@ void bch2_extent_crc_unpacked_to_text(struct printbuf *out, struct bch_extent_cr
 	bch2_prt_compression_type(out, crc->compression_type);
 }
 
+static void bch2_extent_rebalance_to_text(struct printbuf *out, struct bch_fs *c,
+					  const struct bch_extent_rebalance *r)
+{
+	prt_str(out, "rebalance:");
+
+	prt_printf(out, " replicas=%u", r->data_replicas);
+	if (r->data_replicas_from_inode)
+		prt_str(out, " (inode)");
+
+	prt_str(out, " checksum=");
+	bch2_prt_csum_opt(out, r->data_checksum);
+	if (r->data_checksum_from_inode)
+		prt_str(out, " (inode)");
+
+	if (r->background_compression || r->background_compression_from_inode) {
+		prt_str(out, " background_compression=");
+		bch2_compression_opt_to_text(out, r->background_compression);
+
+		if (r->background_compression_from_inode)
+			prt_str(out, " (inode)");
+	}
+
+	if (r->background_target || r->background_target_from_inode) {
+		prt_str(out, " background_target=");
+		if (c)
+			bch2_target_to_text(out, c, r->background_target);
+		else
+			prt_printf(out, "%u", r->background_target);
+
+		if (r->background_target_from_inode)
+			prt_str(out, " (inode)");
+	}
+
+	if (r->promote_target || r->promote_target_from_inode) {
+		prt_str(out, " promote_target=");
+		if (c)
+			bch2_target_to_text(out, c, r->promote_target);
+		else
+			prt_printf(out, "%u", r->promote_target);
+
+		if (r->promote_target_from_inode)
+			prt_str(out, " (inode)");
+	}
+
+	if (r->erasure_code || r->erasure_code_from_inode) {
+		prt_printf(out, " ec=%u", r->erasure_code);
+		if (r->erasure_code_from_inode)
+			prt_str(out, " (inode)");
+	}
+}
+
 void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 			    struct bkey_s_c k)
 {
@@ -1155,18 +1216,10 @@ void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 			       (u64) ec->idx, ec->block);
 			break;
 		}
-		case BCH_EXTENT_ENTRY_rebalance: {
-			const struct bch_extent_rebalance *r = &entry->rebalance;
-
-			prt_str(out, "rebalance: target ");
-			if (c)
-				bch2_target_to_text(out, c, r->target);
-			else
-				prt_printf(out, "%u", r->target);
-			prt_str(out, " compression ");
-			bch2_compression_opt_to_text(out, r->compression);
+		case BCH_EXTENT_ENTRY_rebalance:
+			bch2_extent_rebalance_to_text(out, c, &entry->rebalance);
 			break;
-		}
+
 		default:
 			prt_printf(out, "(invalid extent entry %.16llx)", *((u64 *) entry));
 			return;
@@ -1178,13 +1231,19 @@ void bch2_bkey_ptrs_to_text(struct printbuf *out, struct bch_fs *c,
 
 static int extent_ptr_validate(struct bch_fs *c,
 			       struct bkey_s_c k,
-			       enum bch_validate_flags flags,
+			       struct bkey_validate_context from,
 			       const struct bch_extent_ptr *ptr,
 			       unsigned size_ondisk,
 			       bool metadata)
 {
 	int ret = 0;
 
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	bkey_for_each_ptr(ptrs, ptr2)
+		bkey_fsck_err_on(ptr != ptr2 && ptr->dev == ptr2->dev,
+				 c, ptr_to_duplicate_device,
+				 "multiple pointers to same device (%u)", ptr->dev);
+
 	/* bad pointers are repaired by check_fix_ptrs(): */
 	rcu_read_lock();
 	struct bch_dev *ca = bch2_dev_rcu_noerror(c, ptr->dev);
@@ -1199,13 +1258,6 @@ static int extent_ptr_validate(struct bch_fs *c,
 	unsigned bucket_size	= ca->mi.bucket_size;
 	rcu_read_unlock();
 
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	bkey_for_each_ptr(ptrs, ptr2)
-		bkey_fsck_err_on(ptr != ptr2 && ptr->dev == ptr2->dev,
-				 c, ptr_to_duplicate_device,
-				 "multiple pointers to same device (%u)", ptr->dev);
-
-
 	bkey_fsck_err_on(bucket >= nbuckets,
 			 c, ptr_after_last_bucket,
 			 "pointer past last bucket (%llu > %llu)", bucket, nbuckets);
@@ -1221,7 +1273,7 @@ fsck_err:
 }
 
 int bch2_bkey_ptrs_validate(struct bch_fs *c, struct bkey_s_c k,
-			    enum bch_validate_flags flags)
+			    struct bkey_validate_context from)
 {
 	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 	const union bch_extent_entry *entry;
@@ -1248,7 +1300,7 @@ int bch2_bkey_ptrs_validate(struct bch_fs *c, struct bkey_s_c k,
 
 		switch (extent_entry_type(entry)) {
 		case BCH_EXTENT_ENTRY_ptr:
-			ret = extent_ptr_validate(c, k, flags, &entry->ptr, size_ondisk, false);
+			ret = extent_ptr_validate(c, k, from, &entry->ptr, size_ondisk, false);
 			if (ret)
 				return ret;
 
@@ -1270,9 +1322,6 @@ int bch2_bkey_ptrs_validate(struct bch_fs *c, struct bkey_s_c k,
 		case BCH_EXTENT_ENTRY_crc128:
 			crc = bch2_extent_crc_unpack(k.k, entry_to_crc(entry));
 
-			bkey_fsck_err_on(crc.offset + crc.live_size > crc.uncompressed_size,
-					 c, ptr_crc_uncompressed_size_too_small,
-					 "checksum offset + key size > uncompressed size");
 			bkey_fsck_err_on(!bch2_checksum_type_valid(c, crc.csum_type),
 					 c, ptr_crc_csum_type_unknown,
 					 "invalid checksum type");
@@ -1280,6 +1329,19 @@ int bch2_bkey_ptrs_validate(struct bch_fs *c, struct bkey_s_c k,
 					 c, ptr_crc_compression_type_unknown,
 					 "invalid compression type");
 
+			bkey_fsck_err_on(crc.offset + crc.live_size > crc.uncompressed_size,
+					 c, ptr_crc_uncompressed_size_too_small,
+					 "checksum offset + key size > uncompressed size");
+			bkey_fsck_err_on(crc_is_encoded(crc) &&
+					 (crc.uncompressed_size > c->opts.encoded_extent_max >> 9) &&
+					 (from.flags & (BCH_VALIDATE_write|BCH_VALIDATE_commit)),
+					 c, ptr_crc_uncompressed_size_too_big,
+					 "too large encoded extent");
+			bkey_fsck_err_on(!crc_is_compressed(crc) &&
+					 crc.compressed_size != crc.uncompressed_size,
+					 c, ptr_crc_uncompressed_size_mismatch,
+					 "not compressed but compressed != uncompressed size");
+
 			if (bch2_csum_type_is_encryption(crc.csum_type)) {
 				if (nonce == UINT_MAX)
 					nonce = crc.offset + crc.nonce;
@@ -1293,12 +1355,6 @@ int bch2_bkey_ptrs_validate(struct bch_fs *c, struct bkey_s_c k,
 					 "redundant crc entry");
 			crc_since_last_ptr = true;
 
-			bkey_fsck_err_on(crc_is_encoded(crc) &&
-					 (crc.uncompressed_size > c->opts.encoded_extent_max >> 9) &&
-					 (flags & (BCH_VALIDATE_write|BCH_VALIDATE_commit)),
-					 c, ptr_crc_uncompressed_size_too_big,
-					 "too large encoded extent");
-
 			size_ondisk = crc.compressed_size;
 			break;
 		case BCH_EXTENT_ENTRY_stripe_ptr:
@@ -1391,166 +1447,6 @@ void bch2_ptr_swab(struct bkey_s k)
 	}
 }
 
-const struct bch_extent_rebalance *bch2_bkey_rebalance_opts(struct bkey_s_c k)
-{
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const union bch_extent_entry *entry;
-
-	bkey_extent_entry_for_each(ptrs, entry)
-		if (__extent_entry_type(entry) == BCH_EXTENT_ENTRY_rebalance)
-			return &entry->rebalance;
-
-	return NULL;
-}
-
-unsigned bch2_bkey_ptrs_need_rebalance(struct bch_fs *c, struct bkey_s_c k,
-				       unsigned target, unsigned compression)
-{
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	unsigned rewrite_ptrs = 0;
-
-	if (compression) {
-		unsigned compression_type = bch2_compression_opt_to_type(compression);
-		const union bch_extent_entry *entry;
-		struct extent_ptr_decoded p;
-		unsigned i = 0;
-
-		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-			if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible ||
-			    p.ptr.unwritten) {
-				rewrite_ptrs = 0;
-				goto incompressible;
-			}
-
-			if (!p.ptr.cached && p.crc.compression_type != compression_type)
-				rewrite_ptrs |= 1U << i;
-			i++;
-		}
-	}
-incompressible:
-	if (target && bch2_target_accepts_data(c, BCH_DATA_user, target)) {
-		unsigned i = 0;
-
-		bkey_for_each_ptr(ptrs, ptr) {
-			if (!ptr->cached && !bch2_dev_in_target(c, ptr->dev, target))
-				rewrite_ptrs |= 1U << i;
-			i++;
-		}
-	}
-
-	return rewrite_ptrs;
-}
-
-bool bch2_bkey_needs_rebalance(struct bch_fs *c, struct bkey_s_c k)
-{
-	const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k);
-
-	/*
-	 * If it's an indirect extent, we don't delete the rebalance entry when
-	 * done so that we know what options were applied - check if it still
-	 * needs work done:
-	 */
-	if (r &&
-	    k.k->type == KEY_TYPE_reflink_v &&
-	    !bch2_bkey_ptrs_need_rebalance(c, k, r->target, r->compression))
-		r = NULL;
-
-	return r != NULL;
-}
-
-static u64 __bch2_bkey_sectors_need_rebalance(struct bch_fs *c, struct bkey_s_c k,
-				       unsigned target, unsigned compression)
-{
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const union bch_extent_entry *entry;
-	struct extent_ptr_decoded p;
-	u64 sectors = 0;
-
-	if (compression) {
-		unsigned compression_type = bch2_compression_opt_to_type(compression);
-
-		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-			if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible ||
-			    p.ptr.unwritten) {
-				sectors = 0;
-				goto incompressible;
-			}
-
-			if (!p.ptr.cached && p.crc.compression_type != compression_type)
-				sectors += p.crc.compressed_size;
-		}
-	}
-incompressible:
-	if (target && bch2_target_accepts_data(c, BCH_DATA_user, target)) {
-		bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
-			if (!p.ptr.cached && !bch2_dev_in_target(c, p.ptr.dev, target))
-				sectors += p.crc.compressed_size;
-	}
-
-	return sectors;
-}
-
-u64 bch2_bkey_sectors_need_rebalance(struct bch_fs *c, struct bkey_s_c k)
-{
-	const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k);
-
-	return r ? __bch2_bkey_sectors_need_rebalance(c, k, r->target, r->compression) : 0;
-}
-
-int bch2_bkey_set_needs_rebalance(struct bch_fs *c, struct bkey_i *_k,
-				  struct bch_io_opts *opts)
-{
-	struct bkey_s k = bkey_i_to_s(_k);
-	struct bch_extent_rebalance *r;
-	unsigned target = opts->background_target;
-	unsigned compression = background_compression(*opts);
-	bool needs_rebalance;
-
-	if (!bkey_extent_is_direct_data(k.k))
-		return 0;
-
-	/* get existing rebalance entry: */
-	r = (struct bch_extent_rebalance *) bch2_bkey_rebalance_opts(k.s_c);
-	if (r) {
-		if (k.k->type == KEY_TYPE_reflink_v) {
-			/*
-			 * indirect extents: existing options take precedence,
-			 * so that we don't move extents back and forth if
-			 * they're referenced by different inodes with different
-			 * options:
-			 */
-			if (r->target)
-				target = r->target;
-			if (r->compression)
-				compression = r->compression;
-		}
-
-		r->target	= target;
-		r->compression	= compression;
-	}
-
-	needs_rebalance = bch2_bkey_ptrs_need_rebalance(c, k.s_c, target, compression);
-
-	if (needs_rebalance && !r) {
-		union bch_extent_entry *new = bkey_val_end(k);
-
-		new->rebalance.type		= 1U << BCH_EXTENT_ENTRY_rebalance;
-		new->rebalance.compression	= compression;
-		new->rebalance.target		= target;
-		new->rebalance.unused		= 0;
-		k.k->u64s += extent_entry_u64s(new);
-	} else if (!needs_rebalance && r && k.k->type != KEY_TYPE_reflink_v) {
-		/*
-		 * For indirect extents, don't delete the rebalance entry when
-		 * we're finished so that we know we specifically moved it or
-		 * compressed it to its current location/compression type
-		 */
-		extent_entry_drop(k, (union bch_extent_entry *) r);
-	}
-
-	return 0;
-}
-
 /* Generic extent code: */
 
 int bch2_cut_front_s(struct bpos where, struct bkey_s k)
@@ -1610,7 +1506,7 @@ int bch2_cut_front_s(struct bpos where, struct bkey_s k)
 	case KEY_TYPE_reflink_p: {
 		struct bkey_s_reflink_p p = bkey_s_to_reflink_p(k);
 
-		le64_add_cpu(&p.v->idx, sub);
+		SET_REFLINK_P_IDX(p.v, REFLINK_P_IDX(p.v) + sub);
 		break;
 	}
 	case KEY_TYPE_inline_data:
diff --git a/fs/bcachefs/extents.h b/fs/bcachefs/extents.h
index bcffcf60aaaf..620b284aa34f 100644
--- a/fs/bcachefs/extents.h
+++ b/fs/bcachefs/extents.h
@@ -8,7 +8,6 @@
 
 struct bch_fs;
 struct btree_trans;
-enum bch_validate_flags;
 
 /* extent entries: */
 
@@ -410,12 +409,12 @@ int bch2_bkey_pick_read_device(struct bch_fs *, struct bkey_s_c,
 /* KEY_TYPE_btree_ptr: */
 
 int bch2_btree_ptr_validate(struct bch_fs *, struct bkey_s_c,
-			    enum bch_validate_flags);
+			    struct bkey_validate_context);
 void bch2_btree_ptr_to_text(struct printbuf *, struct bch_fs *,
 			    struct bkey_s_c);
 
 int bch2_btree_ptr_v2_validate(struct bch_fs *, struct bkey_s_c,
-			       enum bch_validate_flags);
+			       struct bkey_validate_context);
 void bch2_btree_ptr_v2_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 void bch2_btree_ptr_v2_compat(enum btree_id, unsigned, unsigned,
 			      int, struct bkey_s);
@@ -452,7 +451,7 @@ bool bch2_extent_merge(struct bch_fs *, struct bkey_s, struct bkey_s_c);
 /* KEY_TYPE_reservation: */
 
 int bch2_reservation_validate(struct bch_fs *, struct bkey_s_c,
-			      enum bch_validate_flags);
+			      struct bkey_validate_context);
 void bch2_reservation_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 bool bch2_reservation_merge(struct bch_fs *, struct bkey_s, struct bkey_s_c);
 
@@ -696,7 +695,7 @@ void bch2_extent_ptr_to_text(struct printbuf *out, struct bch_fs *, const struct
 void bch2_bkey_ptrs_to_text(struct printbuf *, struct bch_fs *,
 			    struct bkey_s_c);
 int bch2_bkey_ptrs_validate(struct bch_fs *, struct bkey_s_c,
-			    enum bch_validate_flags);
+			    struct bkey_validate_context);
 
 static inline bool bch2_extent_ptr_eq(struct bch_extent_ptr ptr1,
 				      struct bch_extent_ptr ptr2)
@@ -710,15 +709,6 @@ static inline bool bch2_extent_ptr_eq(struct bch_extent_ptr ptr1,
 
 void bch2_ptr_swab(struct bkey_s);
 
-const struct bch_extent_rebalance *bch2_bkey_rebalance_opts(struct bkey_s_c);
-unsigned bch2_bkey_ptrs_need_rebalance(struct bch_fs *, struct bkey_s_c,
-				       unsigned, unsigned);
-bool bch2_bkey_needs_rebalance(struct bch_fs *, struct bkey_s_c);
-u64 bch2_bkey_sectors_need_rebalance(struct bch_fs *, struct bkey_s_c);
-
-int bch2_bkey_set_needs_rebalance(struct bch_fs *, struct bkey_i *,
-				  struct bch_io_opts *);
-
 /* Generic extent code: */
 
 enum bch_extent_overlap {
diff --git a/fs/bcachefs/extents_format.h b/fs/bcachefs/extents_format.h
index 3bd2fdbb0817..c198dfc376d6 100644
--- a/fs/bcachefs/extents_format.h
+++ b/fs/bcachefs/extents_format.h
@@ -201,19 +201,8 @@ struct bch_extent_stripe_ptr {
 #endif
 };
 
-struct bch_extent_rebalance {
-#if defined(__LITTLE_ENDIAN_BITFIELD)
-	__u64			type:6,
-				unused:34,
-				compression:8, /* enum bch_compression_opt */
-				target:16;
-#elif defined (__BIG_ENDIAN_BITFIELD)
-	__u64			target:16,
-				compression:8,
-				unused:34,
-				type:6;
-#endif
-};
+/* bch_extent_rebalance: */
+#include "rebalance_format.h"
 
 union bch_extent_entry {
 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ ||  __BITS_PER_LONG == 64
diff --git a/fs/bcachefs/fs-common.c b/fs/bcachefs/fs-common.c
index 7e10a9ddcfd9..d70d9f634cea 100644
--- a/fs/bcachefs/fs-common.c
+++ b/fs/bcachefs/fs-common.c
@@ -69,9 +69,7 @@ int bch2_create_trans(struct btree_trans *trans,
 		if (!snapshot_src.inum) {
 			/* Inode wasn't specified, just snapshot: */
 			struct bch_subvolume s;
-
-			ret = bch2_subvolume_get(trans, snapshot_src.subvol, true,
-						 BTREE_ITER_cached, &s);
+			ret = bch2_subvolume_get(trans, snapshot_src.subvol, true, &s);
 			if (ret)
 				goto err;
 
@@ -154,6 +152,7 @@ int bch2_create_trans(struct btree_trans *trans,
 		if (is_subdir_for_nlink(new_inode))
 			dir_u->bi_nlink++;
 		dir_u->bi_mtime = dir_u->bi_ctime = now;
+		dir_u->bi_size += dirent_occupied_size(name);
 
 		ret = bch2_inode_write(trans, &dir_iter, dir_u);
 		if (ret)
@@ -172,6 +171,10 @@ int bch2_create_trans(struct btree_trans *trans,
 		new_inode->bi_dir_offset	= dir_offset;
 	}
 
+	if (S_ISDIR(mode) &&
+	    !new_inode->bi_subvol)
+		new_inode->bi_depth = dir_u->bi_depth + 1;
+
 	inode_iter.flags &= ~BTREE_ITER_all_snapshots;
 	bch2_btree_iter_set_snapshot(&inode_iter, snapshot);
 
@@ -218,6 +221,7 @@ int bch2_link_trans(struct btree_trans *trans,
 	}
 
 	dir_u->bi_mtime = dir_u->bi_ctime = now;
+	dir_u->bi_size += dirent_occupied_size(name);
 
 	dir_hash = bch2_hash_info_init(c, dir_u);
 
@@ -320,6 +324,7 @@ int bch2_unlink_trans(struct btree_trans *trans,
 
 	dir_u->bi_mtime = dir_u->bi_ctime = inode_u->bi_ctime = now;
 	dir_u->bi_nlink -= is_subdir_for_nlink(inode_u);
+	dir_u->bi_size	-= dirent_occupied_size(name);
 
 	ret =   bch2_hash_delete_at(trans, bch2_dirent_hash_desc,
 				    &dir_hash, &dirent_iter,
@@ -458,6 +463,14 @@ int bch2_rename_trans(struct btree_trans *trans,
 		goto err;
 	}
 
+	if (mode == BCH_RENAME) {
+		src_dir_u->bi_size -= dirent_occupied_size(src_name);
+		dst_dir_u->bi_size += dirent_occupied_size(dst_name);
+	}
+
+	if (mode == BCH_RENAME_OVERWRITE)
+		src_dir_u->bi_size -= dirent_occupied_size(src_name);
+
 	if (src_inode_u->bi_parent_subvol)
 		src_inode_u->bi_parent_subvol = dst_dir.subvol;
 
@@ -512,6 +525,15 @@ int bch2_rename_trans(struct btree_trans *trans,
 		dst_dir_u->bi_nlink++;
 	}
 
+	if (S_ISDIR(src_inode_u->bi_mode) &&
+	    !src_inode_u->bi_subvol)
+		src_inode_u->bi_depth = dst_dir_u->bi_depth + 1;
+
+	if (mode == BCH_RENAME_EXCHANGE &&
+	    S_ISDIR(dst_inode_u->bi_mode) &&
+	    !dst_inode_u->bi_subvol)
+		dst_inode_u->bi_depth = src_dir_u->bi_depth + 1;
+
 	if (dst_inum.inum && is_subdir_for_nlink(dst_inode_u)) {
 		dst_dir_u->bi_nlink--;
 		src_dir_u->bi_nlink += mode == BCH_RENAME_EXCHANGE;
@@ -548,3 +570,94 @@ err:
 	bch2_trans_iter_exit(trans, &src_dir_iter);
 	return ret;
 }
+
+static inline void prt_bytes_reversed(struct printbuf *out, const void *b, unsigned n)
+{
+	bch2_printbuf_make_room(out, n);
+
+	unsigned can_print = min(n, printbuf_remaining(out));
+
+	b += n;
+
+	for (unsigned i = 0; i < can_print; i++)
+		out->buf[out->pos++] = *((char *) --b);
+
+	printbuf_nul_terminate(out);
+}
+
+static inline void prt_str_reversed(struct printbuf *out, const char *s)
+{
+	prt_bytes_reversed(out, s, strlen(s));
+}
+
+static inline void reverse_bytes(void *b, size_t n)
+{
+	char *e = b + n, *s = b;
+
+	while (s < e) {
+		--e;
+		swap(*s, *e);
+		s++;
+	}
+}
+
+/* XXX: we don't yet attempt to print paths when we don't know the subvol */
+int bch2_inum_to_path(struct btree_trans *trans, subvol_inum inum, struct printbuf *path)
+{
+	unsigned orig_pos = path->pos;
+	int ret = 0;
+
+	while (!(inum.subvol == BCACHEFS_ROOT_SUBVOL &&
+		 inum.inum   == BCACHEFS_ROOT_INO)) {
+		struct bch_inode_unpacked inode;
+		ret = bch2_inode_find_by_inum_trans(trans, inum, &inode);
+		if (ret)
+			goto disconnected;
+
+		if (!inode.bi_dir && !inode.bi_dir_offset) {
+			ret = -BCH_ERR_ENOENT_inode_no_backpointer;
+			goto disconnected;
+		}
+
+		inum.subvol	= inode.bi_parent_subvol ?: inum.subvol;
+		inum.inum	= inode.bi_dir;
+
+		u32 snapshot;
+		ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
+		if (ret)
+			goto disconnected;
+
+		struct btree_iter d_iter;
+		struct bkey_s_c_dirent d = bch2_bkey_get_iter_typed(trans, &d_iter,
+				BTREE_ID_dirents, SPOS(inode.bi_dir, inode.bi_dir_offset, snapshot),
+				0, dirent);
+		ret = bkey_err(d.s_c);
+		if (ret)
+			goto disconnected;
+
+		struct qstr dirent_name = bch2_dirent_get_name(d);
+		prt_bytes_reversed(path, dirent_name.name, dirent_name.len);
+
+		prt_char(path, '/');
+
+		bch2_trans_iter_exit(trans, &d_iter);
+	}
+
+	if (orig_pos == path->pos)
+		prt_char(path, '/');
+out:
+	ret = path->allocation_failure ? -ENOMEM : 0;
+	if (ret)
+		goto err;
+
+	reverse_bytes(path->buf + orig_pos, path->pos - orig_pos);
+	return 0;
+err:
+	return ret;
+disconnected:
+	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+		goto err;
+
+	prt_str_reversed(path, "(disconnected)");
+	goto out;
+}
diff --git a/fs/bcachefs/fs-common.h b/fs/bcachefs/fs-common.h
index c934e807b380..2b59210bb5e8 100644
--- a/fs/bcachefs/fs-common.h
+++ b/fs/bcachefs/fs-common.h
@@ -42,4 +42,6 @@ int bch2_rename_trans(struct btree_trans *,
 bool bch2_reinherit_attrs(struct bch_inode_unpacked *,
 			  struct bch_inode_unpacked *);
 
+int bch2_inum_to_path(struct btree_trans *, subvol_inum, struct printbuf *);
+
 #endif /* _BCACHEFS_FS_COMMON_H */
diff --git a/fs/bcachefs/fs-io-buffered.c b/fs/bcachefs/fs-io-buffered.c
index 95972809e76d..ab1d5db2fa56 100644
--- a/fs/bcachefs/fs-io-buffered.c
+++ b/fs/bcachefs/fs-io-buffered.c
@@ -164,7 +164,8 @@ static void bchfs_read(struct btree_trans *trans,
 			     BTREE_ITER_slots);
 	while (1) {
 		struct bkey_s_c k;
-		unsigned bytes, sectors, offset_into_extent;
+		unsigned bytes, sectors;
+		s64 offset_into_extent;
 		enum btree_id data_btree = BTREE_ID_extents;
 
 		bch2_trans_begin(trans);
@@ -197,7 +198,7 @@ static void bchfs_read(struct btree_trans *trans,
 
 		k = bkey_i_to_s_c(sk.k);
 
-		sectors = min(sectors, k.k->size - offset_into_extent);
+		sectors = min_t(unsigned, sectors, k.k->size - offset_into_extent);
 
 		if (readpages_iter) {
 			ret = readpage_bio_extend(trans, readpages_iter, &rbio->bio, sectors,
@@ -230,10 +231,12 @@ err:
 	bch2_trans_iter_exit(trans, &iter);
 
 	if (ret) {
-		bch_err_inum_offset_ratelimited(c,
-				iter.pos.inode,
-				iter.pos.offset << 9,
-				"read error %i from btree lookup", ret);
+		struct printbuf buf = PRINTBUF;
+		bch2_inum_offset_err_msg_trans(trans, &buf, inum, iter.pos.offset << 9);
+		prt_printf(&buf, "read error %i from btree lookup", ret);
+		bch_err_ratelimited(c, "%s", buf.buf);
+		printbuf_exit(&buf);
+
 		rbio->bio.bi_status = BLK_STS_IOERR;
 		bio_endio(&rbio->bio);
 	}
@@ -248,6 +251,7 @@ void bch2_readahead(struct readahead_control *ractl)
 	struct bch_io_opts opts;
 	struct folio *folio;
 	struct readpages_iter readpages_iter;
+	struct blk_plug plug;
 
 	bch2_inode_opts_get(&opts, c, &inode->ei_inode);
 
@@ -255,6 +259,16 @@ void bch2_readahead(struct readahead_control *ractl)
 	if (ret)
 		return;
 
+	/*
+	 * Besides being a general performance optimization, plugging helps with
+	 * avoiding btree transaction srcu warnings - submitting a bio can
+	 * block, and we don't want todo that with the transaction locked.
+	 *
+	 * However, plugged bios are submitted when we schedule; we ideally
+	 * would have our own scheduler hook to call unlock_long() before
+	 * scheduling.
+	 */
+	blk_start_plug(&plug);
 	bch2_pagecache_add_get(inode);
 
 	struct btree_trans *trans = bch2_trans_get(c);
@@ -281,7 +295,7 @@ void bch2_readahead(struct readahead_control *ractl)
 	bch2_trans_put(trans);
 
 	bch2_pagecache_add_put(inode);
-
+	blk_finish_plug(&plug);
 	darray_exit(&readpages_iter.folios);
 }
 
@@ -296,9 +310,13 @@ int bch2_read_single_folio(struct folio *folio, struct address_space *mapping)
 	struct bch_fs *c = inode->v.i_sb->s_fs_info;
 	struct bch_read_bio *rbio;
 	struct bch_io_opts opts;
+	struct blk_plug plug;
 	int ret;
 	DECLARE_COMPLETION_ONSTACK(done);
 
+	BUG_ON(folio_test_uptodate(folio));
+	BUG_ON(folio_test_dirty(folio));
+
 	if (!bch2_folio_create(folio, GFP_KERNEL))
 		return -ENOMEM;
 
@@ -313,7 +331,9 @@ int bch2_read_single_folio(struct folio *folio, struct address_space *mapping)
 	rbio->bio.bi_iter.bi_sector = folio_sector(folio);
 	BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
 
+	blk_start_plug(&plug);
 	bch2_trans_run(c, (bchfs_read(trans, rbio, inode_inum(inode), NULL), 0));
+	blk_finish_plug(&plug);
 	wait_for_completion(&done);
 
 	ret = blk_status_to_errno(rbio->bio.bi_status);
@@ -605,15 +625,6 @@ do_io:
 		BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio,
 				     sectors << 9, offset << 9));
 
-		/* Check for writing past i_size: */
-		WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
-			  round_up(i_size, block_bytes(c)) &&
-			  !test_bit(BCH_FS_emergency_ro, &c->flags),
-			  "writing past i_size: %llu > %llu (unrounded %llu)\n",
-			  bio_end_sector(&w->io->op.wbio.bio) << 9,
-			  round_up(i_size, block_bytes(c)),
-			  i_size);
-
 		w->io->op.res.sectors += reserved_sectors;
 		w->io->op.i_sectors_delta -= dirty_sectors;
 		w->io->op.new_i_size = i_size;
@@ -669,7 +680,7 @@ int bch2_write_begin(struct file *file, struct address_space *mapping,
 	folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
 				    FGP_WRITEBEGIN | fgf_set_order(len),
 				    mapping_gfp_mask(mapping));
-	if (IS_ERR_OR_NULL(folio))
+	if (IS_ERR(folio))
 		goto err_unlock;
 
 	offset = pos - folio_pos(folio);
diff --git a/fs/bcachefs/fs-io-direct.c b/fs/bcachefs/fs-io-direct.c
index 6d3a05ae5da8..2089c36b5866 100644
--- a/fs/bcachefs/fs-io-direct.c
+++ b/fs/bcachefs/fs-io-direct.c
@@ -70,6 +70,7 @@ static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
 	struct bch_io_opts opts;
 	struct dio_read *dio;
 	struct bio *bio;
+	struct blk_plug plug;
 	loff_t offset = req->ki_pos;
 	bool sync = is_sync_kiocb(req);
 	size_t shorten;
@@ -128,6 +129,8 @@ static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
 	 */
 	dio->should_dirty = iter_is_iovec(iter);
 
+	blk_start_plug(&plug);
+
 	goto start;
 	while (iter->count) {
 		bio = bio_alloc_bioset(NULL,
@@ -160,6 +163,8 @@ start:
 		bch2_read(c, rbio_init(bio, opts), inode_inum(inode));
 	}
 
+	blk_finish_plug(&plug);
+
 	iter->count += shorten;
 
 	if (sync) {
diff --git a/fs/bcachefs/fs-io-pagecache.c b/fs/bcachefs/fs-io-pagecache.c
index 1d4910ea0f1d..e072900e6a5b 100644
--- a/fs/bcachefs/fs-io-pagecache.c
+++ b/fs/bcachefs/fs-io-pagecache.c
@@ -29,7 +29,7 @@ int bch2_filemap_get_contig_folios_d(struct address_space *mapping,
 			break;
 
 		f = __filemap_get_folio(mapping, pos >> PAGE_SHIFT, fgp_flags, gfp);
-		if (IS_ERR_OR_NULL(f))
+		if (IS_ERR(f))
 			break;
 
 		BUG_ON(fs->nr && folio_pos(f) != pos);
@@ -199,7 +199,7 @@ int bch2_folio_set(struct bch_fs *c, subvol_inum inum,
 	unsigned folio_idx = 0;
 
 	return bch2_trans_run(c,
-		for_each_btree_key_in_subvolume_upto(trans, iter, BTREE_ID_extents,
+		for_each_btree_key_in_subvolume_max(trans, iter, BTREE_ID_extents,
 				   POS(inum.inum, offset),
 				   POS(inum.inum, U64_MAX),
 				   inum.subvol, BTREE_ITER_slots, k, ({
diff --git a/fs/bcachefs/fs-io.c b/fs/bcachefs/fs-io.c
index 2456c41b215e..94bf34b9b65f 100644
--- a/fs/bcachefs/fs-io.c
+++ b/fs/bcachefs/fs-io.c
@@ -167,6 +167,34 @@ void __bch2_i_sectors_acct(struct bch_fs *c, struct bch_inode_info *inode,
 
 /* fsync: */
 
+static int bch2_get_inode_journal_seq_trans(struct btree_trans *trans, subvol_inum inum,
+					    u64 *seq)
+{
+	struct printbuf buf = PRINTBUF;
+	struct bch_inode_unpacked u;
+	struct btree_iter iter;
+	int ret = bch2_inode_peek(trans, &iter, &u, inum, 0);
+	if (ret)
+		return ret;
+
+	u64 cur_seq = journal_cur_seq(&trans->c->journal);
+	*seq = min(cur_seq, u.bi_journal_seq);
+
+	if (fsck_err_on(u.bi_journal_seq > cur_seq,
+			trans, inode_journal_seq_in_future,
+			"inode journal seq in future (currently at %llu)\n%s",
+			cur_seq,
+			(bch2_inode_unpacked_to_text(&buf, &u),
+			buf.buf))) {
+		u.bi_journal_seq = cur_seq;
+		ret = bch2_inode_write(trans, &iter, &u);
+	}
+fsck_err:
+	bch2_trans_iter_exit(trans, &iter);
+	printbuf_exit(&buf);
+	return ret;
+}
+
 /*
  * inode->ei_inode.bi_journal_seq won't be up to date since it's set in an
  * insert trigger: look up the btree inode instead
@@ -180,9 +208,10 @@ static int bch2_flush_inode(struct bch_fs *c,
 	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_fsync))
 		return -EROFS;
 
-	struct bch_inode_unpacked u;
-	int ret = bch2_inode_find_by_inum(c, inode_inum(inode), &u) ?:
-		  bch2_journal_flush_seq(&c->journal, u.bi_journal_seq, TASK_INTERRUPTIBLE) ?:
+	u64 seq;
+	int ret = bch2_trans_commit_do(c, NULL, NULL, 0,
+			bch2_get_inode_journal_seq_trans(trans, inode_inum(inode), &seq)) ?:
+		  bch2_journal_flush_seq(&c->journal, seq, TASK_INTERRUPTIBLE) ?:
 		  bch2_inode_flush_nocow_writes(c, inode);
 	bch2_write_ref_put(c, BCH_WRITE_REF_fsync);
 	return ret;
@@ -222,7 +251,7 @@ static inline int range_has_data(struct bch_fs *c, u32 subvol,
 				 struct bpos end)
 {
 	return bch2_trans_run(c,
-		for_each_btree_key_in_subvolume_upto(trans, iter, BTREE_ID_extents, start, end,
+		for_each_btree_key_in_subvolume_max(trans, iter, BTREE_ID_extents, start, end,
 						    subvol, 0, k, ({
 			bkey_extent_is_data(k.k) && !bkey_extent_is_unwritten(k);
 		})));
@@ -256,7 +285,7 @@ static int __bch2_truncate_folio(struct bch_inode_info *inode,
 
 		folio = __filemap_get_folio(mapping, index,
 					    FGP_LOCK|FGP_CREAT, GFP_KERNEL);
-		if (IS_ERR_OR_NULL(folio)) {
+		if (IS_ERR(folio)) {
 			ret = -ENOMEM;
 			goto out;
 		}
@@ -806,7 +835,7 @@ static int quota_reserve_range(struct bch_inode_info *inode,
 	u64 sectors = end - start;
 
 	int ret = bch2_trans_run(c,
-		for_each_btree_key_in_subvolume_upto(trans, iter,
+		for_each_btree_key_in_subvolume_max(trans, iter,
 				BTREE_ID_extents,
 				POS(inode->v.i_ino, start),
 				POS(inode->v.i_ino, end - 1),
@@ -877,11 +906,18 @@ loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
 	bch2_mark_pagecache_unallocated(src, pos_src >> 9,
 				   (pos_src + aligned_len) >> 9);
 
+	/*
+	 * XXX: we'd like to be telling bch2_remap_range() if we have
+	 * permission to write to the source file, and thus if io path option
+	 * changes should be propagated through the copy, but we need mnt_idmap
+	 * from the pathwalk, awkward
+	 */
 	ret = bch2_remap_range(c,
 			       inode_inum(dst), pos_dst >> 9,
 			       inode_inum(src), pos_src >> 9,
 			       aligned_len >> 9,
-			       pos_dst + len, &i_sectors_delta);
+			       pos_dst + len, &i_sectors_delta,
+			       false);
 	if (ret < 0)
 		goto err;
 
@@ -922,7 +958,7 @@ static loff_t bch2_seek_data(struct file *file, u64 offset)
 		return -ENXIO;
 
 	int ret = bch2_trans_run(c,
-		for_each_btree_key_in_subvolume_upto(trans, iter, BTREE_ID_extents,
+		for_each_btree_key_in_subvolume_max(trans, iter, BTREE_ID_extents,
 				   POS(inode->v.i_ino, offset >> 9),
 				   POS(inode->v.i_ino, U64_MAX),
 				   inum.subvol, 0, k, ({
@@ -958,7 +994,7 @@ static loff_t bch2_seek_hole(struct file *file, u64 offset)
 		return -ENXIO;
 
 	int ret = bch2_trans_run(c,
-		for_each_btree_key_in_subvolume_upto(trans, iter, BTREE_ID_extents,
+		for_each_btree_key_in_subvolume_max(trans, iter, BTREE_ID_extents,
 				   POS(inode->v.i_ino, offset >> 9),
 				   POS(inode->v.i_ino, U64_MAX),
 				   inum.subvol, BTREE_ITER_slots, k, ({
diff --git a/fs/bcachefs/fs-ioctl.c b/fs/bcachefs/fs-ioctl.c
index 405cf08bda34..15725b4ce393 100644
--- a/fs/bcachefs/fs-ioctl.c
+++ b/fs/bcachefs/fs-ioctl.c
@@ -406,7 +406,7 @@ static long bch2_ioctl_subvolume_create(struct bch_fs *c, struct file *filp,
 		sync_inodes_sb(c->vfs_sb);
 		up_read(&c->vfs_sb->s_umount);
 	}
-retry:
+
 	if (arg.src_ptr) {
 		error = user_path_at(arg.dirfd,
 				(const char __user *)(unsigned long)arg.src_ptr,
@@ -486,11 +486,6 @@ err3:
 err2:
 	if (arg.src_ptr)
 		path_put(&src_path);
-
-	if (retry_estale(error, lookup_flags)) {
-		lookup_flags |= LOOKUP_REVAL;
-		goto retry;
-	}
 err1:
 	return error;
 }
diff --git a/fs/bcachefs/fs.c b/fs/bcachefs/fs.c
index a41d0d8a2f7b..90ade8f648d9 100644
--- a/fs/bcachefs/fs.c
+++ b/fs/bcachefs/fs.c
@@ -23,6 +23,7 @@
 #include "journal.h"
 #include "keylist.h"
 #include "quota.h"
+#include "rebalance.h"
 #include "snapshot.h"
 #include "super.h"
 #include "xattr.h"
@@ -38,6 +39,7 @@
 #include <linux/posix_acl.h>
 #include <linux/random.h>
 #include <linux/seq_file.h>
+#include <linux/siphash.h>
 #include <linux/statfs.h>
 #include <linux/string.h>
 #include <linux/xattr.h>
@@ -65,6 +67,9 @@ void bch2_inode_update_after_write(struct btree_trans *trans,
 	i_gid_write(&inode->v, bi->bi_gid);
 	inode->v.i_mode	= bi->bi_mode;
 
+	if (fields & ATTR_SIZE)
+		i_size_write(&inode->v, bi->bi_size);
+
 	if (fields & ATTR_ATIME)
 		inode_set_atime_to_ts(&inode->v, bch2_time_to_timespec(c, bi->bi_atime));
 	if (fields & ATTR_MTIME)
@@ -89,10 +94,25 @@ int __must_check bch2_write_inode(struct bch_fs *c,
 retry:
 	bch2_trans_begin(trans);
 
-	ret   = bch2_inode_peek(trans, &iter, &inode_u, inode_inum(inode),
-				BTREE_ITER_intent) ?:
-		(set ? set(trans, inode, &inode_u, p) : 0) ?:
-		bch2_inode_write(trans, &iter, &inode_u) ?:
+	ret = bch2_inode_peek(trans, &iter, &inode_u, inode_inum(inode), BTREE_ITER_intent);
+	if (ret)
+		goto err;
+
+	struct bch_extent_rebalance old_r = bch2_inode_rebalance_opts_get(c, &inode_u);
+
+	ret = (set ? set(trans, inode, &inode_u, p) : 0);
+	if (ret)
+		goto err;
+
+	struct bch_extent_rebalance new_r = bch2_inode_rebalance_opts_get(c, &inode_u);
+
+	if (memcmp(&old_r, &new_r, sizeof(new_r))) {
+		ret = bch2_set_rebalance_needs_scan_trans(trans, inode_u.bi_inum);
+		if (ret)
+			goto err;
+	}
+
+	ret   = bch2_inode_write(trans, &iter, &inode_u) ?:
 		bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
 
 	/*
@@ -101,7 +121,7 @@ retry:
 	 */
 	if (!ret)
 		bch2_inode_update_after_write(trans, inode, &inode_u, fields);
-
+err:
 	bch2_trans_iter_exit(trans, &iter);
 
 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
@@ -160,8 +180,9 @@ static bool subvol_inum_eq(subvol_inum a, subvol_inum b)
 static u32 bch2_vfs_inode_hash_fn(const void *data, u32 len, u32 seed)
 {
 	const subvol_inum *inum = data;
+	siphash_key_t k = { .key[0] = seed };
 
-	return jhash(&inum->inum, sizeof(inum->inum), seed);
+	return siphash_2u64(inum->subvol, inum->inum, &k);
 }
 
 static u32 bch2_vfs_inode_obj_hash_fn(const void *data, u32 len, u32 seed)
@@ -190,11 +211,18 @@ static const struct rhashtable_params bch2_vfs_inodes_params = {
 	.automatic_shrinking	= true,
 };
 
+static const struct rhashtable_params bch2_vfs_inodes_by_inum_params = {
+	.head_offset		= offsetof(struct bch_inode_info, by_inum_hash),
+	.key_offset		= offsetof(struct bch_inode_info, ei_inum.inum),
+	.key_len		= sizeof(u64),
+	.automatic_shrinking	= true,
+};
+
 int bch2_inode_or_descendents_is_open(struct btree_trans *trans, struct bpos p)
 {
 	struct bch_fs *c = trans->c;
-	struct rhashtable *ht = &c->vfs_inodes_table;
-	subvol_inum inum = (subvol_inum) { .inum = p.offset };
+	struct rhltable *ht = &c->vfs_inodes_by_inum_table;
+	u64 inum = p.offset;
 	DARRAY(u32) subvols;
 	int ret = 0;
 
@@ -219,15 +247,15 @@ restart_from_top:
 	struct rhash_lock_head __rcu *const *bkt;
 	struct rhash_head *he;
 	unsigned int hash;
-	struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
+	struct bucket_table *tbl = rht_dereference_rcu(ht->ht.tbl, &ht->ht);
 restart:
-	hash = rht_key_hashfn(ht, tbl, &inum, bch2_vfs_inodes_params);
+	hash = rht_key_hashfn(&ht->ht, tbl, &inum, bch2_vfs_inodes_by_inum_params);
 	bkt = rht_bucket(tbl, hash);
 	do {
 		struct bch_inode_info *inode;
 
 		rht_for_each_entry_rcu_from(inode, he, rht_ptr_rcu(bkt), tbl, hash, hash) {
-			if (inode->ei_inum.inum == inum.inum) {
+			if (inode->ei_inum.inum == inum) {
 				ret = darray_push_gfp(&subvols, inode->ei_inum.subvol,
 						      GFP_NOWAIT|__GFP_NOWARN);
 				if (ret) {
@@ -248,7 +276,7 @@ restart:
 	/* Ensure we see any new tables. */
 	smp_rmb();
 
-	tbl = rht_dereference_rcu(tbl->future_tbl, ht);
+	tbl = rht_dereference_rcu(tbl->future_tbl, &ht->ht);
 	if (unlikely(tbl))
 		goto restart;
 	rcu_read_unlock();
@@ -327,7 +355,11 @@ static void bch2_inode_hash_remove(struct bch_fs *c, struct bch_inode_info *inod
 	spin_unlock(&inode->v.i_lock);
 
 	if (remove) {
-		int ret = rhashtable_remove_fast(&c->vfs_inodes_table,
+		int ret = rhltable_remove(&c->vfs_inodes_by_inum_table,
+					&inode->by_inum_hash, bch2_vfs_inodes_by_inum_params);
+		BUG_ON(ret);
+
+		ret = rhashtable_remove_fast(&c->vfs_inodes_table,
 					&inode->hash, bch2_vfs_inodes_params);
 		BUG_ON(ret);
 		inode->v.i_hash.pprev = NULL;
@@ -372,6 +404,11 @@ retry:
 		discard_new_inode(&inode->v);
 		return old;
 	} else {
+		int ret = rhltable_insert(&c->vfs_inodes_by_inum_table,
+					  &inode->by_inum_hash,
+					  bch2_vfs_inodes_by_inum_params);
+		BUG_ON(ret);
+
 		inode_fake_hash(&inode->v);
 
 		inode_sb_list_add(&inode->v);
@@ -465,7 +502,7 @@ struct inode *bch2_vfs_inode_get(struct bch_fs *c, subvol_inum inum)
 	struct bch_inode_unpacked inode_u;
 	struct bch_subvolume subvol;
 	int ret = lockrestart_do(trans,
-		bch2_subvolume_get(trans, inum.subvol, true, 0, &subvol) ?:
+		bch2_subvolume_get(trans, inum.subvol, true, &subvol) ?:
 		bch2_inode_find_by_inum_trans(trans, inum, &inode_u)) ?:
 		PTR_ERR_OR_ZERO(inode = bch2_inode_hash_init_insert(trans, inum, &inode_u, &subvol));
 	bch2_trans_put(trans);
@@ -535,8 +572,7 @@ retry:
 	inum.subvol = inode_u.bi_subvol ?: dir->ei_inum.subvol;
 	inum.inum = inode_u.bi_inum;
 
-	ret   = bch2_subvolume_get(trans, inum.subvol, true,
-				   BTREE_ITER_with_updates, &subvol) ?:
+	ret   = bch2_subvolume_get(trans, inum.subvol, true, &subvol) ?:
 		bch2_trans_commit(trans, NULL, &journal_seq, 0);
 	if (unlikely(ret)) {
 		bch2_quota_acct(c, bch_qid(&inode_u), Q_INO, -1,
@@ -549,7 +585,7 @@ err_before_quota:
 
 	if (!(flags & BCH_CREATE_TMPFILE)) {
 		bch2_inode_update_after_write(trans, dir, &dir_u,
-					      ATTR_MTIME|ATTR_CTIME);
+					      ATTR_MTIME|ATTR_CTIME|ATTR_SIZE);
 		mutex_unlock(&dir->ei_update_lock);
 	}
 
@@ -617,7 +653,7 @@ static struct bch_inode_info *bch2_lookup_trans(struct btree_trans *trans,
 
 	struct bch_subvolume subvol;
 	struct bch_inode_unpacked inode_u;
-	ret =   bch2_subvolume_get(trans, inum.subvol, true, 0, &subvol) ?:
+	ret =   bch2_subvolume_get(trans, inum.subvol, true, &subvol) ?:
 		bch2_inode_find_by_inum_nowarn_trans(trans, inum, &inode_u) ?:
 		PTR_ERR_OR_ZERO(inode = bch2_inode_hash_init_insert(trans, inum, &inode_u, &subvol));
 
@@ -628,7 +664,7 @@ static struct bch_inode_info *bch2_lookup_trans(struct btree_trans *trans,
 		goto err;
 
 	/* regular files may have hardlinks: */
-	if (bch2_fs_inconsistent_on(bch2_inode_should_have_bp(&inode_u) &&
+	if (bch2_fs_inconsistent_on(bch2_inode_should_have_single_bp(&inode_u) &&
 				    !bkey_eq(k.k->p, POS(inode_u.bi_dir, inode_u.bi_dir_offset)),
 				    c,
 				    "dirent points to inode that does not point back:\n  %s",
@@ -706,7 +742,7 @@ static int __bch2_link(struct bch_fs *c,
 
 	if (likely(!ret)) {
 		bch2_inode_update_after_write(trans, dir, &dir_u,
-					      ATTR_MTIME|ATTR_CTIME);
+					      ATTR_MTIME|ATTR_CTIME|ATTR_SIZE);
 		bch2_inode_update_after_write(trans, inode, &inode_u, ATTR_CTIME);
 	}
 
@@ -759,7 +795,7 @@ int __bch2_unlink(struct inode *vdir, struct dentry *dentry,
 		goto err;
 
 	bch2_inode_update_after_write(trans, dir, &dir_u,
-				      ATTR_MTIME|ATTR_CTIME);
+				      ATTR_MTIME|ATTR_CTIME|ATTR_SIZE);
 	bch2_inode_update_after_write(trans, inode, &inode_u,
 				      ATTR_MTIME);
 
@@ -937,11 +973,11 @@ err_tx_restart:
 	       dst_inode->v.i_ino != dst_inode_u.bi_inum);
 
 	bch2_inode_update_after_write(trans, src_dir, &src_dir_u,
-				      ATTR_MTIME|ATTR_CTIME);
+				      ATTR_MTIME|ATTR_CTIME|ATTR_SIZE);
 
 	if (src_dir != dst_dir)
 		bch2_inode_update_after_write(trans, dst_dir, &dst_dir_u,
-					      ATTR_MTIME|ATTR_CTIME);
+					      ATTR_MTIME|ATTR_CTIME|ATTR_SIZE);
 
 	bch2_inode_update_after_write(trans, src_inode, &src_inode_u,
 				      ATTR_CTIME);
@@ -1245,7 +1281,6 @@ static int bch2_fiemap(struct inode *vinode, struct fiemap_extent_info *info,
 	struct btree_iter iter;
 	struct bkey_s_c k;
 	struct bkey_buf cur, prev;
-	unsigned offset_into_extent, sectors;
 	bool have_extent = false;
 	int ret = 0;
 
@@ -1278,7 +1313,7 @@ static int bch2_fiemap(struct inode *vinode, struct fiemap_extent_info *info,
 
 		bch2_btree_iter_set_snapshot(&iter, snapshot);
 
-		k = bch2_btree_iter_peek_upto(&iter, end);
+		k = bch2_btree_iter_peek_max(&iter, end);
 		ret = bkey_err(k);
 		if (ret)
 			continue;
@@ -1292,9 +1327,8 @@ static int bch2_fiemap(struct inode *vinode, struct fiemap_extent_info *info,
 			continue;
 		}
 
-		offset_into_extent	= iter.pos.offset -
-			bkey_start_offset(k.k);
-		sectors			= k.k->size - offset_into_extent;
+		s64 offset_into_extent	= iter.pos.offset - bkey_start_offset(k.k);
+		unsigned sectors	= k.k->size - offset_into_extent;
 
 		bch2_bkey_buf_reassemble(&cur, c, k);
 
@@ -1306,7 +1340,7 @@ static int bch2_fiemap(struct inode *vinode, struct fiemap_extent_info *info,
 		k = bkey_i_to_s_c(cur.k);
 		bch2_bkey_buf_realloc(&prev, c, k.k->u64s);
 
-		sectors = min(sectors, k.k->size - offset_into_extent);
+		sectors = min_t(unsigned, sectors, k.k->size - offset_into_extent);
 
 		bch2_cut_front(POS(k.k->p.inode,
 				   bkey_start_offset(k.k) +
@@ -1736,7 +1770,6 @@ static void bch2_vfs_inode_init(struct btree_trans *trans,
 	bch2_inode_update_after_write(trans, inode, bi, ~0);
 
 	inode->v.i_blocks	= bi->bi_sectors;
-	inode->v.i_ino		= bi->bi_inum;
 	inode->v.i_rdev		= bi->bi_dev;
 	inode->v.i_generation	= bi->bi_generation;
 	inode->v.i_size		= bi->bi_size;
@@ -2200,7 +2233,8 @@ got_sb:
 	sb->s_time_gran		= c->sb.nsec_per_time_unit;
 	sb->s_time_min		= div_s64(S64_MIN, c->sb.time_units_per_sec) + 1;
 	sb->s_time_max		= div_s64(S64_MAX, c->sb.time_units_per_sec);
-	sb->s_uuid		= c->sb.user_uuid;
+	super_set_uuid(sb, c->sb.user_uuid.b, sizeof(c->sb.user_uuid));
+	super_set_sysfs_name_uuid(sb);
 	sb->s_shrink->seeks	= 0;
 	c->vfs_sb		= sb;
 	strscpy(sb->s_id, c->name, sizeof(sb->s_id));
@@ -2345,13 +2379,16 @@ static int bch2_init_fs_context(struct fs_context *fc)
 
 void bch2_fs_vfs_exit(struct bch_fs *c)
 {
+	if (c->vfs_inodes_by_inum_table.ht.tbl)
+		rhltable_destroy(&c->vfs_inodes_by_inum_table);
 	if (c->vfs_inodes_table.tbl)
 		rhashtable_destroy(&c->vfs_inodes_table);
 }
 
 int bch2_fs_vfs_init(struct bch_fs *c)
 {
-	return rhashtable_init(&c->vfs_inodes_table, &bch2_vfs_inodes_params);
+	return rhashtable_init(&c->vfs_inodes_table, &bch2_vfs_inodes_params) ?:
+		rhltable_init(&c->vfs_inodes_by_inum_table, &bch2_vfs_inodes_by_inum_params);
 }
 
 static struct file_system_type bcache_fs_type = {
diff --git a/fs/bcachefs/fs.h b/fs/bcachefs/fs.h
index 59f9f7ae728d..dd2198541455 100644
--- a/fs/bcachefs/fs.h
+++ b/fs/bcachefs/fs.h
@@ -14,6 +14,7 @@
 struct bch_inode_info {
 	struct inode		v;
 	struct rhash_head	hash;
+	struct rhlist_head	by_inum_hash;
 	subvol_inum		ei_inum;
 
 	struct list_head	ei_vfs_inode_list;
diff --git a/fs/bcachefs/fsck.c b/fs/bcachefs/fsck.c
index 75c8a97a6954..8fcf7c8e5ede 100644
--- a/fs/bcachefs/fsck.c
+++ b/fs/bcachefs/fsck.c
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 
 #include "bcachefs.h"
+#include "bcachefs_ioctl.h"
 #include "bkey_buf.h"
 #include "btree_cache.h"
 #include "btree_update.h"
@@ -16,6 +17,7 @@
 #include "recovery_passes.h"
 #include "snapshot.h"
 #include "super.h"
+#include "thread_with_file.h"
 #include "xattr.h"
 
 #include <linux/bsearch.h>
@@ -73,7 +75,7 @@ static s64 bch2_count_inode_sectors(struct btree_trans *trans, u64 inum,
 {
 	u64 sectors = 0;
 
-	int ret = for_each_btree_key_upto(trans, iter, BTREE_ID_extents,
+	int ret = for_each_btree_key_max(trans, iter, BTREE_ID_extents,
 				SPOS(inum, 0, snapshot),
 				POS(inum, U64_MAX),
 				0, k, ({
@@ -90,7 +92,7 @@ static s64 bch2_count_subdirs(struct btree_trans *trans, u64 inum,
 {
 	u64 subdirs = 0;
 
-	int ret = for_each_btree_key_upto(trans, iter, BTREE_ID_dirents,
+	int ret = for_each_btree_key_max(trans, iter, BTREE_ID_dirents,
 				    SPOS(inum, 0, snapshot),
 				    POS(inum, U64_MAX),
 				    0, k, ({
@@ -107,7 +109,7 @@ static int subvol_lookup(struct btree_trans *trans, u32 subvol,
 			 u32 *snapshot, u64 *inum)
 {
 	struct bch_subvolume s;
-	int ret = bch2_subvolume_get(trans, subvol, false, 0, &s);
+	int ret = bch2_subvolume_get(trans, subvol, false, &s);
 
 	*snapshot = le32_to_cpu(s.snapshot);
 	*inum = le64_to_cpu(s.inode);
@@ -170,7 +172,7 @@ static int lookup_dirent_in_snapshot(struct btree_trans *trans,
 	if (ret)
 		return ret;
 
-	struct bkey_s_c_dirent d = bkey_s_c_to_dirent(bch2_btree_iter_peek_slot(&iter));
+	struct bkey_s_c_dirent d = bkey_s_c_to_dirent(k);
 	*target = le64_to_cpu(d.v->d_inum);
 	*type = d.v->d_type;
 	bch2_trans_iter_exit(trans, &iter);
@@ -203,6 +205,36 @@ err:
 	return ret;
 }
 
+/*
+ * Find any subvolume associated with a tree of snapshots
+ * We can't rely on master_subvol - it might have been deleted.
+ */
+static int find_snapshot_tree_subvol(struct btree_trans *trans,
+				     u32 tree_id, u32 *subvol)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+
+	for_each_btree_key_norestart(trans, iter, BTREE_ID_snapshots, POS_MIN, 0, k, ret) {
+		if (k.k->type != KEY_TYPE_snapshot)
+			continue;
+
+		struct bkey_s_c_snapshot s = bkey_s_c_to_snapshot(k);
+		if (le32_to_cpu(s.v->tree) != tree_id)
+			continue;
+
+		if (s.v->subvol) {
+			*subvol = le32_to_cpu(s.v->subvol);
+			goto found;
+		}
+	}
+	ret = -BCH_ERR_ENOENT_no_snapshot_tree_subvol;
+found:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
 /* Get lost+found, create if it doesn't exist: */
 static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
 			    struct bch_inode_unpacked *lostfound,
@@ -210,6 +242,7 @@ static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
 {
 	struct bch_fs *c = trans->c;
 	struct qstr lostfound_str = QSTR("lost+found");
+	struct btree_iter lostfound_iter = { NULL };
 	u64 inum = 0;
 	unsigned d_type = 0;
 	int ret;
@@ -220,20 +253,24 @@ static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
 	if (ret)
 		return ret;
 
-	subvol_inum root_inum = { .subvol = le32_to_cpu(st.master_subvol) };
+	u32 subvolid;
+	ret = find_snapshot_tree_subvol(trans,
+				bch2_snapshot_tree(c, snapshot), &subvolid);
+	bch_err_msg(c, ret, "finding subvol associated with snapshot tree %u",
+		    bch2_snapshot_tree(c, snapshot));
+	if (ret)
+		return ret;
 
 	struct bch_subvolume subvol;
-	ret = bch2_subvolume_get(trans, le32_to_cpu(st.master_subvol),
-				 false, 0, &subvol);
-	bch_err_msg(c, ret, "looking up root subvol %u for snapshot %u",
-		    le32_to_cpu(st.master_subvol), snapshot);
+	ret = bch2_subvolume_get(trans, subvolid, false, &subvol);
+	bch_err_msg(c, ret, "looking up subvol %u for snapshot %u", subvolid, snapshot);
 	if (ret)
 		return ret;
 
 	if (!subvol.inode) {
 		struct btree_iter iter;
 		struct bkey_i_subvolume *subvol = bch2_bkey_get_mut_typed(trans, &iter,
-				BTREE_ID_subvolumes, POS(0, le32_to_cpu(st.master_subvol)),
+				BTREE_ID_subvolumes, POS(0, subvolid),
 				0, subvolume);
 		ret = PTR_ERR_OR_ZERO(subvol);
 		if (ret)
@@ -243,13 +280,16 @@ static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
 		bch2_trans_iter_exit(trans, &iter);
 	}
 
-	root_inum.inum = le64_to_cpu(subvol.inode);
+	subvol_inum root_inum = {
+		.subvol = subvolid,
+		.inum = le64_to_cpu(subvol.inode)
+	};
 
 	struct bch_inode_unpacked root_inode;
 	struct bch_hash_info root_hash_info;
 	ret = lookup_inode(trans, root_inum.inum, snapshot, &root_inode);
 	bch_err_msg(c, ret, "looking up root inode %llu for subvol %u",
-		    root_inum.inum, le32_to_cpu(st.master_subvol));
+		    root_inum.inum, subvolid);
 	if (ret)
 		return ret;
 
@@ -288,11 +328,16 @@ create_lostfound:
 	 * XXX: we could have a nicer log message here  if we had a nice way to
 	 * walk backpointers to print a path
 	 */
-	bch_notice(c, "creating lost+found in subvol %llu snapshot %u",
-		   root_inum.subvol, le32_to_cpu(st.root_snapshot));
+	struct printbuf path = PRINTBUF;
+	ret = bch2_inum_to_path(trans, root_inum, &path);
+	if (ret)
+		goto err;
+
+	bch_notice(c, "creating %s/lost+found in subvol %llu snapshot %u",
+		   path.buf, root_inum.subvol, snapshot);
+	printbuf_exit(&path);
 
 	u64 now = bch2_current_time(c);
-	struct btree_iter lostfound_iter = { NULL };
 	u64 cpu = raw_smp_processor_id();
 
 	bch2_inode_init_early(c, lostfound);
@@ -451,7 +496,9 @@ static int reattach_inode(struct btree_trans *trans, struct bch_inode_unpacked *
 				continue;
 
 			struct bch_inode_unpacked child_inode;
-			bch2_inode_unpack(k, &child_inode);
+			ret = bch2_inode_unpack(k, &child_inode);
+			if (ret)
+				break;
 
 			if (!inode_should_reattach(&child_inode)) {
 				ret = maybe_delete_dirent(trans,
@@ -482,6 +529,13 @@ static int reattach_inode(struct btree_trans *trans, struct bch_inode_unpacked *
 	return ret;
 }
 
+static struct bkey_s_c_dirent dirent_get_by_pos(struct btree_trans *trans,
+						struct btree_iter *iter,
+						struct bpos pos)
+{
+	return bch2_bkey_get_iter_typed(trans, iter, BTREE_ID_dirents, pos, 0, dirent);
+}
+
 static int remove_backpointer(struct btree_trans *trans,
 			      struct bch_inode_unpacked *inode)
 {
@@ -490,13 +544,11 @@ static int remove_backpointer(struct btree_trans *trans,
 
 	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
-	struct bkey_s_c_dirent d =
-		bch2_bkey_get_iter_typed(trans, &iter, BTREE_ID_dirents,
-				     SPOS(inode->bi_dir, inode->bi_dir_offset, inode->bi_snapshot), 0,
-				     dirent);
-	int ret =   bkey_err(d) ?:
-		dirent_points_to_inode(c, d, inode) ?:
-		__remove_dirent(trans, d.k->p);
+	struct bkey_s_c_dirent d = dirent_get_by_pos(trans, &iter,
+				     SPOS(inode->bi_dir, inode->bi_dir_offset, inode->bi_snapshot));
+	int ret = bkey_err(d) ?:
+		  dirent_points_to_inode(c, d, inode) ?:
+		  __remove_dirent(trans, d.k->p);
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
@@ -613,7 +665,7 @@ static int reconstruct_inode(struct btree_trans *trans, enum btree_id btree, u32
 		struct btree_iter iter = {};
 
 		bch2_trans_iter_init(trans, &iter, BTREE_ID_extents, SPOS(inum, U64_MAX, snapshot), 0);
-		struct bkey_s_c k = bch2_btree_iter_peek_prev(&iter);
+		struct bkey_s_c k = bch2_btree_iter_peek_prev_min(&iter, POS(inum, 0));
 		bch2_trans_iter_exit(trans, &iter);
 		int ret = bkey_err(k);
 		if (ret)
@@ -780,11 +832,13 @@ struct inode_walker {
 	struct bpos			last_pos;
 
 	DARRAY(struct inode_walker_entry) inodes;
+	snapshot_id_list		deletes;
 };
 
 static void inode_walker_exit(struct inode_walker *w)
 {
 	darray_exit(&w->inodes);
+	darray_exit(&w->deletes);
 }
 
 static struct inode_walker inode_walker_init(void)
@@ -797,9 +851,8 @@ static int add_inode(struct bch_fs *c, struct inode_walker *w,
 {
 	struct bch_inode_unpacked u;
 
-	BUG_ON(bch2_inode_unpack(inode, &u));
-
-	return darray_push(&w->inodes, ((struct inode_walker_entry) {
+	return bch2_inode_unpack(inode, &u) ?:
+		darray_push(&w->inodes, ((struct inode_walker_entry) {
 		.inode		= u,
 		.snapshot	= inode.k->p.snapshot,
 	}));
@@ -909,8 +962,9 @@ static int get_visible_inodes(struct btree_trans *trans,
 	int ret;
 
 	w->inodes.nr = 0;
+	w->deletes.nr = 0;
 
-	for_each_btree_key_norestart(trans, iter, BTREE_ID_inodes, POS(0, inum),
+	for_each_btree_key_reverse_norestart(trans, iter, BTREE_ID_inodes, SPOS(0, inum, s->pos.snapshot),
 			   BTREE_ITER_all_snapshots, k, ret) {
 		if (k.k->p.offset != inum)
 			break;
@@ -918,10 +972,13 @@ static int get_visible_inodes(struct btree_trans *trans,
 		if (!ref_visible(c, s, s->pos.snapshot, k.k->p.snapshot))
 			continue;
 
-		if (bkey_is_inode(k.k))
-			add_inode(c, w, k);
+		if (snapshot_list_has_ancestor(c, &w->deletes, k.k->p.snapshot))
+			continue;
 
-		if (k.k->p.snapshot >= s->pos.snapshot)
+		ret = bkey_is_inode(k.k)
+			? add_inode(c, w, k)
+			: snapshot_list_add(c, &w->deletes, k.k->p.snapshot);
+		if (ret)
 			break;
 	}
 	bch2_trans_iter_exit(trans, &iter);
@@ -929,69 +986,16 @@ static int get_visible_inodes(struct btree_trans *trans,
 	return ret;
 }
 
-static int dirent_has_target(struct btree_trans *trans, struct bkey_s_c_dirent d)
-{
-	if (d.v->d_type == DT_SUBVOL) {
-		u32 snap;
-		u64 inum;
-		int ret = subvol_lookup(trans, le32_to_cpu(d.v->d_child_subvol), &snap, &inum);
-		if (ret && !bch2_err_matches(ret, ENOENT))
-			return ret;
-		return !ret;
-	} else {
-		struct btree_iter iter;
-		struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
-				SPOS(0, le64_to_cpu(d.v->d_inum), d.k->p.snapshot), 0);
-		int ret = bkey_err(k);
-		if (ret)
-			return ret;
-
-		ret = bkey_is_inode(k.k);
-		bch2_trans_iter_exit(trans, &iter);
-		return ret;
-	}
-}
-
 /*
  * Prefer to delete the first one, since that will be the one at the wrong
  * offset:
  * return value: 0 -> delete k1, 1 -> delete k2
  */
-static int hash_pick_winner(struct btree_trans *trans,
-			    const struct bch_hash_desc desc,
-			    struct bch_hash_info *hash_info,
-			    struct bkey_s_c k1,
-			    struct bkey_s_c k2)
-{
-	if (bkey_val_bytes(k1.k) == bkey_val_bytes(k2.k) &&
-	    !memcmp(k1.v, k2.v, bkey_val_bytes(k1.k)))
-		return 0;
-
-	switch (desc.btree_id) {
-	case BTREE_ID_dirents: {
-		int ret = dirent_has_target(trans, bkey_s_c_to_dirent(k1));
-		if (ret < 0)
-			return ret;
-		if (!ret)
-			return 0;
-
-		ret = dirent_has_target(trans, bkey_s_c_to_dirent(k2));
-		if (ret < 0)
-			return ret;
-		if (!ret)
-			return 1;
-		return 2;
-	}
-	default:
-		return 0;
-	}
-}
-
-static int fsck_update_backpointers(struct btree_trans *trans,
-				    struct snapshots_seen *s,
-				    const struct bch_hash_desc desc,
-				    struct bch_hash_info *hash_info,
-				    struct bkey_i *new)
+int bch2_fsck_update_backpointers(struct btree_trans *trans,
+				  struct snapshots_seen *s,
+				  const struct bch_hash_desc desc,
+				  struct bch_hash_info *hash_info,
+				  struct bkey_i *new)
 {
 	if (new->k.type != KEY_TYPE_dirent)
 		return 0;
@@ -1019,160 +1023,6 @@ err:
 	return ret;
 }
 
-static int fsck_rename_dirent(struct btree_trans *trans,
-			      struct snapshots_seen *s,
-			      const struct bch_hash_desc desc,
-			      struct bch_hash_info *hash_info,
-			      struct bkey_s_c_dirent old)
-{
-	struct qstr old_name = bch2_dirent_get_name(old);
-	struct bkey_i_dirent *new = bch2_trans_kmalloc(trans, bkey_bytes(old.k) + 32);
-	int ret = PTR_ERR_OR_ZERO(new);
-	if (ret)
-		return ret;
-
-	bkey_dirent_init(&new->k_i);
-	dirent_copy_target(new, old);
-	new->k.p = old.k->p;
-
-	for (unsigned i = 0; i < 1000; i++) {
-		unsigned len = sprintf(new->v.d_name, "%.*s.fsck_renamed-%u",
-				       old_name.len, old_name.name, i);
-		unsigned u64s = BKEY_U64s + dirent_val_u64s(len);
-
-		if (u64s > U8_MAX)
-			return -EINVAL;
-
-		new->k.u64s = u64s;
-
-		ret = bch2_hash_set_in_snapshot(trans, bch2_dirent_hash_desc, hash_info,
-						(subvol_inum) { 0, old.k->p.inode },
-						old.k->p.snapshot, &new->k_i,
-						BTREE_UPDATE_internal_snapshot_node);
-		if (!bch2_err_matches(ret, EEXIST))
-			break;
-	}
-
-	if (ret)
-		return ret;
-
-	return fsck_update_backpointers(trans, s, desc, hash_info, &new->k_i);
-}
-
-static int hash_check_key(struct btree_trans *trans,
-			  struct snapshots_seen *s,
-			  const struct bch_hash_desc desc,
-			  struct bch_hash_info *hash_info,
-			  struct btree_iter *k_iter, struct bkey_s_c hash_k)
-{
-	struct bch_fs *c = trans->c;
-	struct btree_iter iter = { NULL };
-	struct printbuf buf = PRINTBUF;
-	struct bkey_s_c k;
-	u64 hash;
-	int ret = 0;
-
-	if (hash_k.k->type != desc.key_type)
-		return 0;
-
-	hash = desc.hash_bkey(hash_info, hash_k);
-
-	if (likely(hash == hash_k.k->p.offset))
-		return 0;
-
-	if (hash_k.k->p.offset < hash)
-		goto bad_hash;
-
-	for_each_btree_key_norestart(trans, iter, desc.btree_id,
-				     SPOS(hash_k.k->p.inode, hash, hash_k.k->p.snapshot),
-				     BTREE_ITER_slots, k, ret) {
-		if (bkey_eq(k.k->p, hash_k.k->p))
-			break;
-
-		if (k.k->type == desc.key_type &&
-		    !desc.cmp_bkey(k, hash_k))
-			goto duplicate_entries;
-
-		if (bkey_deleted(k.k)) {
-			bch2_trans_iter_exit(trans, &iter);
-			goto bad_hash;
-		}
-	}
-out:
-	bch2_trans_iter_exit(trans, &iter);
-	printbuf_exit(&buf);
-	return ret;
-bad_hash:
-	if (fsck_err(trans, hash_table_key_wrong_offset,
-		     "hash table key at wrong offset: btree %s inode %llu offset %llu, hashed to %llu\n  %s",
-		     bch2_btree_id_str(desc.btree_id), hash_k.k->p.inode, hash_k.k->p.offset, hash,
-		     (printbuf_reset(&buf),
-		      bch2_bkey_val_to_text(&buf, c, hash_k), buf.buf))) {
-		struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, hash_k);
-		if (IS_ERR(new))
-			return PTR_ERR(new);
-
-		k = bch2_hash_set_or_get_in_snapshot(trans, &iter, desc, hash_info,
-				       (subvol_inum) { 0, hash_k.k->p.inode },
-				       hash_k.k->p.snapshot, new,
-				       STR_HASH_must_create|
-				       BTREE_ITER_with_updates|
-				       BTREE_UPDATE_internal_snapshot_node);
-		ret = bkey_err(k);
-		if (ret)
-			goto out;
-		if (k.k)
-			goto duplicate_entries;
-
-		ret =   bch2_hash_delete_at(trans, desc, hash_info, k_iter,
-					    BTREE_UPDATE_internal_snapshot_node) ?:
-			fsck_update_backpointers(trans, s, desc, hash_info, new) ?:
-			bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc) ?:
-			-BCH_ERR_transaction_restart_nested;
-		goto out;
-	}
-fsck_err:
-	goto out;
-duplicate_entries:
-	ret = hash_pick_winner(trans, desc, hash_info, hash_k, k);
-	if (ret < 0)
-		goto out;
-
-	if (!fsck_err(trans, hash_table_key_duplicate,
-		      "duplicate hash table keys%s:\n%s",
-		      ret != 2 ? "" : ", both point to valid inodes",
-		      (printbuf_reset(&buf),
-		       bch2_bkey_val_to_text(&buf, c, hash_k),
-		       prt_newline(&buf),
-		       bch2_bkey_val_to_text(&buf, c, k),
-		       buf.buf)))
-		goto out;
-
-	switch (ret) {
-	case 0:
-		ret = bch2_hash_delete_at(trans, desc, hash_info, k_iter, 0);
-		break;
-	case 1:
-		ret = bch2_hash_delete_at(trans, desc, hash_info, &iter, 0);
-		break;
-	case 2:
-		ret = fsck_rename_dirent(trans, s, desc, hash_info, bkey_s_c_to_dirent(hash_k)) ?:
-			bch2_hash_delete_at(trans, desc, hash_info, k_iter, 0);
-		goto out;
-	}
-
-	ret = bch2_trans_commit(trans, NULL, NULL, 0) ?:
-		-BCH_ERR_transaction_restart_nested;
-	goto out;
-}
-
-static struct bkey_s_c_dirent dirent_get_by_pos(struct btree_trans *trans,
-						struct btree_iter *iter,
-						struct bpos pos)
-{
-	return bch2_bkey_get_iter_typed(trans, iter, BTREE_ID_dirents, pos, 0, dirent);
-}
-
 static struct bkey_s_c_dirent inode_get_dirent(struct btree_trans *trans,
 					       struct btree_iter *iter,
 					       struct bch_inode_unpacked *inode,
@@ -1260,12 +1110,43 @@ static int get_snapshot_root_inode(struct btree_trans *trans,
 		goto err;
 	BUG();
 found_root:
-	BUG_ON(bch2_inode_unpack(k, root));
+	ret = bch2_inode_unpack(k, root);
 err:
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
 }
 
+static int check_directory_size(struct btree_trans *trans,
+				struct bch_inode_unpacked *inode_u,
+				struct bkey_s_c inode_k, bool *write_inode)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	u64 new_size = 0;
+	int ret;
+
+	for_each_btree_key_max_norestart(trans, iter, BTREE_ID_dirents,
+			SPOS(inode_k.k->p.offset, 0, inode_k.k->p.snapshot),
+			POS(inode_k.k->p.offset, U64_MAX),
+			0, k, ret) {
+		if (k.k->type != KEY_TYPE_dirent)
+			continue;
+
+		struct bkey_s_c_dirent dirent = bkey_s_c_to_dirent(k);
+		struct qstr name = bch2_dirent_get_name(dirent);
+
+		new_size += dirent_occupied_size(&name);
+	}
+	bch2_trans_iter_exit(trans, &iter);
+
+	if (!ret && inode_u->bi_size != new_size) {
+		inode_u->bi_size = new_size;
+		*write_inode = true;
+	}
+
+	return ret;
+}
+
 static int check_inode(struct btree_trans *trans,
 		       struct btree_iter *iter,
 		       struct bkey_s_c k,
@@ -1291,7 +1172,9 @@ static int check_inode(struct btree_trans *trans,
 	if (!bkey_is_inode(k.k))
 		return 0;
 
-	BUG_ON(bch2_inode_unpack(k, &u));
+	ret = bch2_inode_unpack(k, &u);
+	if (ret)
+		goto err;
 
 	if (snapshot_root->bi_inum != u.bi_inum) {
 		ret = get_snapshot_root_inode(trans, snapshot_root, u.bi_inum);
@@ -1302,7 +1185,7 @@ static int check_inode(struct btree_trans *trans,
 	if (fsck_err_on(u.bi_hash_seed		!= snapshot_root->bi_hash_seed ||
 			INODE_STR_HASH(&u)	!= INODE_STR_HASH(snapshot_root),
 			trans, inode_snapshot_mismatch,
-			"inodes in different snapshots don't match")) {
+			"inode hash info in different snapshots don't match")) {
 		u.bi_hash_seed = snapshot_root->bi_hash_seed;
 		SET_INODE_STR_HASH(&u, INODE_STR_HASH(snapshot_root));
 		do_update = true;
@@ -1392,7 +1275,7 @@ static int check_inode(struct btree_trans *trans,
 
 			if (fsck_err_on(!ret,
 					trans, inode_unlinked_and_not_open,
-				      "inode %llu%u unlinked and not open",
+				      "inode %llu:%u unlinked and not open",
 				      u.bi_inum, u.bi_snapshot)) {
 				ret = bch2_inode_rm_snapshot(trans, u.bi_inum, iter->pos.snapshot);
 				bch_err_msg(c, ret, "in fsck deleting inode");
@@ -1415,7 +1298,7 @@ static int check_inode(struct btree_trans *trans,
 	if (u.bi_subvol) {
 		struct bch_subvolume s;
 
-		ret = bch2_subvolume_get(trans, u.bi_subvol, false, 0, &s);
+		ret = bch2_subvolume_get(trans, u.bi_subvol, false, &s);
 		if (ret && !bch2_err_matches(ret, ENOENT))
 			goto err;
 
@@ -1441,6 +1324,27 @@ static int check_inode(struct btree_trans *trans,
 			do_update = true;
 		}
 	}
+
+	if (fsck_err_on(u.bi_journal_seq > journal_cur_seq(&c->journal),
+			trans, inode_journal_seq_in_future,
+			"inode journal seq in future (currently at %llu)\n%s",
+			journal_cur_seq(&c->journal),
+			(printbuf_reset(&buf),
+			 bch2_inode_unpacked_to_text(&buf, &u),
+			buf.buf))) {
+		u.bi_journal_seq = journal_cur_seq(&c->journal);
+		do_update = true;
+	}
+
+	if (S_ISDIR(u.bi_mode)) {
+		ret = check_directory_size(trans, &u, k, &do_update);
+
+		fsck_err_on(ret,
+			    trans, directory_size_mismatch,
+			    "directory inode %llu:%u with the mismatch directory size",
+			    u.bi_inum, k.k->p.snapshot);
+		ret = 0;
+	}
 do_update:
 	if (do_update) {
 		ret = __bch2_fsck_write_inode(trans, &u);
@@ -1502,7 +1406,9 @@ static int find_oldest_inode_needs_reattach(struct btree_trans *trans,
 			break;
 
 		struct bch_inode_unpacked parent_inode;
-		bch2_inode_unpack(k, &parent_inode);
+		ret = bch2_inode_unpack(k, &parent_inode);
+		if (ret)
+			break;
 
 		if (!inode_should_reattach(&parent_inode))
 			break;
@@ -1525,7 +1431,9 @@ static int check_unreachable_inode(struct btree_trans *trans,
 		return 0;
 
 	struct bch_inode_unpacked inode;
-	BUG_ON(bch2_inode_unpack(k, &inode));
+	ret = bch2_inode_unpack(k, &inode);
+	if (ret)
+		return ret;
 
 	if (!inode_should_reattach(&inode))
 		return 0;
@@ -1649,7 +1557,7 @@ static int check_i_sectors_notnested(struct btree_trans *trans, struct inode_wal
 		if (i->count != count2) {
 			bch_err_ratelimited(c, "fsck counted i_sectors wrong for inode %llu:%u: got %llu should be %llu",
 					    w->last_pos.inode, i->snapshot, i->count, count2);
-			return -BCH_ERR_internal_fsck_err;
+			i->count = count2;
 		}
 
 		if (fsck_err_on(!(i->inode.bi_flags & BCH_INODE_i_sectors_dirty),
@@ -1753,7 +1661,7 @@ static int overlapping_extents_found(struct btree_trans *trans,
 	bch2_trans_iter_init(trans, &iter1, btree, pos1,
 			     BTREE_ITER_all_snapshots|
 			     BTREE_ITER_not_extents);
-	k1 = bch2_btree_iter_peek_upto(&iter1, POS(pos1.inode, U64_MAX));
+	k1 = bch2_btree_iter_peek_max(&iter1, POS(pos1.inode, U64_MAX));
 	ret = bkey_err(k1);
 	if (ret)
 		goto err;
@@ -1778,7 +1686,7 @@ static int overlapping_extents_found(struct btree_trans *trans,
 	while (1) {
 		bch2_btree_iter_advance(&iter2);
 
-		k2 = bch2_btree_iter_peek_upto(&iter2, POS(pos1.inode, U64_MAX));
+		k2 = bch2_btree_iter_peek_max(&iter2, POS(pos1.inode, U64_MAX));
 		ret = bkey_err(k2);
 		if (ret)
 			goto err;
@@ -2156,7 +2064,7 @@ static int check_dirent_inode_dirent(struct btree_trans *trans,
 		return __bch2_fsck_write_inode(trans, target);
 	}
 
-	if (bch2_inode_should_have_bp(target) &&
+	if (bch2_inode_should_have_single_bp(target) &&
 	    !fsck_err(trans, inode_wrong_backpointer,
 		      "dirent points to inode that does not point back:\n  %s",
 		      (bch2_bkey_val_to_text(&buf, c, d.s_c),
@@ -2480,7 +2388,7 @@ static int check_dirent(struct btree_trans *trans, struct btree_iter *iter,
 		*hash_info = bch2_hash_info_init(c, &i->inode);
 	dir->first_this_inode = false;
 
-	ret = hash_check_key(trans, s, bch2_dirent_hash_desc, hash_info, iter, k);
+	ret = bch2_str_hash_check_key(trans, s, &bch2_dirent_hash_desc, hash_info, iter, k);
 	if (ret < 0)
 		goto err;
 	if (ret) {
@@ -2519,6 +2427,30 @@ static int check_dirent(struct btree_trans *trans, struct btree_iter *iter,
 			if (ret)
 				goto err;
 		}
+
+		darray_for_each(target->deletes, i)
+			if (fsck_err_on(!snapshot_list_has_id(&s->ids, *i),
+					trans, dirent_to_overwritten_inode,
+					"dirent points to inode overwritten in snapshot %u:\n%s",
+					*i,
+					(printbuf_reset(&buf),
+					 bch2_bkey_val_to_text(&buf, c, k),
+					 buf.buf))) {
+				struct btree_iter delete_iter;
+				bch2_trans_iter_init(trans, &delete_iter,
+						     BTREE_ID_dirents,
+						     SPOS(k.k->p.inode, k.k->p.offset, *i),
+						     BTREE_ITER_intent);
+				ret =   bch2_btree_iter_traverse(&delete_iter) ?:
+					bch2_hash_delete_at(trans, bch2_dirent_hash_desc,
+							  hash_info,
+							  &delete_iter,
+							  BTREE_UPDATE_internal_snapshot_node);
+				bch2_trans_iter_exit(trans, &delete_iter);
+				if (ret)
+					goto err;
+
+			}
 	}
 
 	ret = bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
@@ -2594,7 +2526,7 @@ static int check_xattr(struct btree_trans *trans, struct btree_iter *iter,
 		*hash_info = bch2_hash_info_init(c, &i->inode);
 	inode->first_this_inode = false;
 
-	ret = hash_check_key(trans, NULL, bch2_xattr_hash_desc, hash_info, iter, k);
+	ret = bch2_str_hash_check_key(trans, NULL, &bch2_xattr_hash_desc, hash_info, iter, k);
 	bch_err_fn(c, ret);
 	return ret;
 }
@@ -2774,6 +2706,48 @@ struct pathbuf_entry {
 
 typedef DARRAY(struct pathbuf_entry) pathbuf;
 
+static int bch2_bi_depth_renumber_one(struct btree_trans *trans, struct pathbuf_entry *p,
+				      u32 new_depth)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
+					       SPOS(0, p->inum, p->snapshot), 0);
+
+	struct bch_inode_unpacked inode;
+	int ret = bkey_err(k) ?:
+		!bkey_is_inode(k.k) ? -BCH_ERR_ENOENT_inode
+		: bch2_inode_unpack(k, &inode);
+	if (ret)
+		goto err;
+
+	if (inode.bi_depth != new_depth) {
+		inode.bi_depth = new_depth;
+		ret = __bch2_fsck_write_inode(trans, &inode) ?:
+			bch2_trans_commit(trans, NULL, NULL, 0);
+	}
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+static int bch2_bi_depth_renumber(struct btree_trans *trans, pathbuf *path, u32 new_bi_depth)
+{
+	u32 restart_count = trans->restart_count;
+	int ret = 0;
+
+	darray_for_each_reverse(*path, i) {
+		ret = nested_lockrestart_do(trans,
+				bch2_bi_depth_renumber_one(trans, i, new_bi_depth));
+		bch_err_fn(trans->c, ret);
+		if (ret)
+			break;
+
+		new_bi_depth++;
+	}
+
+	return ret ?: trans_was_restarted(trans, restart_count);
+}
+
 static bool path_is_dup(pathbuf *p, u64 inum, u32 snapshot)
 {
 	darray_for_each(*p, i)
@@ -2783,21 +2757,21 @@ static bool path_is_dup(pathbuf *p, u64 inum, u32 snapshot)
 	return false;
 }
 
-static int check_path(struct btree_trans *trans, pathbuf *p, struct bkey_s_c inode_k)
+static int check_path_loop(struct btree_trans *trans, struct bkey_s_c inode_k)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter inode_iter = {};
-	struct bch_inode_unpacked inode;
+	pathbuf path = {};
 	struct printbuf buf = PRINTBUF;
 	u32 snapshot = inode_k.k->p.snapshot;
+	bool redo_bi_depth = false;
+	u32 min_bi_depth = U32_MAX;
 	int ret = 0;
 
-	p->nr = 0;
-
-	BUG_ON(bch2_inode_unpack(inode_k, &inode));
-
-	if (!S_ISDIR(inode.bi_mode))
-		return 0;
+	struct bch_inode_unpacked inode;
+	ret = bch2_inode_unpack(inode_k, &inode);
+	if (ret)
+		return ret;
 
 	while (!inode.bi_subvol) {
 		struct btree_iter dirent_iter;
@@ -2807,7 +2781,7 @@ static int check_path(struct btree_trans *trans, pathbuf *p, struct bkey_s_c ino
 		d = inode_get_dirent(trans, &dirent_iter, &inode, &parent_snapshot);
 		ret = bkey_err(d.s_c);
 		if (ret && !bch2_err_matches(ret, ENOENT))
-			break;
+			goto out;
 
 		if (!ret && (ret = dirent_points_to_inode(c, d, &inode)))
 			bch2_trans_iter_exit(trans, &dirent_iter);
@@ -2822,7 +2796,7 @@ static int check_path(struct btree_trans *trans, pathbuf *p, struct bkey_s_c ino
 
 		bch2_trans_iter_exit(trans, &dirent_iter);
 
-		ret = darray_push(p, ((struct pathbuf_entry) {
+		ret = darray_push(&path, ((struct pathbuf_entry) {
 			.inum		= inode.bi_inum,
 			.snapshot	= snapshot,
 		}));
@@ -2834,22 +2808,32 @@ static int check_path(struct btree_trans *trans, pathbuf *p, struct bkey_s_c ino
 		bch2_trans_iter_exit(trans, &inode_iter);
 		inode_k = bch2_bkey_get_iter(trans, &inode_iter, BTREE_ID_inodes,
 					     SPOS(0, inode.bi_dir, snapshot), 0);
+
+		struct bch_inode_unpacked parent_inode;
 		ret = bkey_err(inode_k) ?:
 			!bkey_is_inode(inode_k.k) ? -BCH_ERR_ENOENT_inode
-			: bch2_inode_unpack(inode_k, &inode);
+			: bch2_inode_unpack(inode_k, &parent_inode);
 		if (ret) {
 			/* Should have been caught in dirents pass */
 			bch_err_msg(c, ret, "error looking up parent directory");
-			break;
+			goto out;
 		}
 
+		min_bi_depth = parent_inode.bi_depth;
+
+		if (parent_inode.bi_depth < inode.bi_depth &&
+		    min_bi_depth < U16_MAX)
+			break;
+
+		inode = parent_inode;
 		snapshot = inode_k.k->p.snapshot;
+		redo_bi_depth = true;
 
-		if (path_is_dup(p, inode.bi_inum, snapshot)) {
+		if (path_is_dup(&path, inode.bi_inum, snapshot)) {
 			/* XXX print path */
 			bch_err(c, "directory structure loop");
 
-			darray_for_each(*p, i)
+			darray_for_each(path, i)
 				pr_err("%llu:%u", i->inum, i->snapshot);
 			pr_err("%llu:%u", inode.bi_inum, snapshot);
 
@@ -2862,12 +2846,20 @@ static int check_path(struct btree_trans *trans, pathbuf *p, struct bkey_s_c ino
 				ret = reattach_inode(trans, &inode);
 				bch_err_msg(c, ret, "reattaching inode %llu", inode.bi_inum);
 			}
-			break;
+
+			goto out;
 		}
 	}
+
+	if (inode.bi_subvol)
+		min_bi_depth = 0;
+
+	if (redo_bi_depth)
+		ret = bch2_bi_depth_renumber(trans, &path, min_bi_depth);
 out:
 fsck_err:
 	bch2_trans_iter_exit(trans, &inode_iter);
+	darray_exit(&path);
 	printbuf_exit(&buf);
 	bch_err_fn(c, ret);
 	return ret;
@@ -2879,24 +2871,20 @@ fsck_err:
  */
 int bch2_check_directory_structure(struct bch_fs *c)
 {
-	pathbuf path = { 0, };
-	int ret;
-
-	ret = bch2_trans_run(c,
+	int ret = bch2_trans_run(c,
 		for_each_btree_key_commit(trans, iter, BTREE_ID_inodes, POS_MIN,
 					  BTREE_ITER_intent|
 					  BTREE_ITER_prefetch|
 					  BTREE_ITER_all_snapshots, k,
 					  NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
-			if (!bkey_is_inode(k.k))
+			if (!S_ISDIR(bkey_inode_mode(k)))
 				continue;
 
 			if (bch2_inode_flags(k) & BCH_INODE_unlinked)
 				continue;
 
-			check_path(trans, &path, k);
+			check_path_loop(trans, k);
 		})));
-	darray_exit(&path);
 
 	bch_err_fn(c, ret);
 	return ret;
@@ -2994,7 +2982,9 @@ static int check_nlinks_find_hardlinks(struct bch_fs *c,
 
 			/* Should never fail, checked by bch2_inode_invalid: */
 			struct bch_inode_unpacked u;
-			BUG_ON(bch2_inode_unpack(k, &u));
+			_ret3 = bch2_inode_unpack(k, &u);
+			if (_ret3)
+				break;
 
 			/*
 			 * Backpointer and directory structure checks are sufficient for
@@ -3072,7 +3062,9 @@ static int check_nlinks_update_inode(struct btree_trans *trans, struct btree_ite
 	if (!bkey_is_inode(k.k))
 		return 0;
 
-	BUG_ON(bch2_inode_unpack(k, &u));
+	ret = bch2_inode_unpack(k, &u);
+	if (ret)
+		return ret;
 
 	if (S_ISDIR(u.bi_mode))
 		return 0;
@@ -3194,3 +3186,223 @@ int bch2_fix_reflink_p(struct bch_fs *c)
 	bch_err_fn(c, ret);
 	return ret;
 }
+
+#ifndef NO_BCACHEFS_CHARDEV
+
+struct fsck_thread {
+	struct thread_with_stdio thr;
+	struct bch_fs		*c;
+	struct bch_opts		opts;
+};
+
+static void bch2_fsck_thread_exit(struct thread_with_stdio *_thr)
+{
+	struct fsck_thread *thr = container_of(_thr, struct fsck_thread, thr);
+	kfree(thr);
+}
+
+static int bch2_fsck_offline_thread_fn(struct thread_with_stdio *stdio)
+{
+	struct fsck_thread *thr = container_of(stdio, struct fsck_thread, thr);
+	struct bch_fs *c = thr->c;
+
+	int ret = PTR_ERR_OR_ZERO(c);
+	if (ret)
+		return ret;
+
+	ret = bch2_fs_start(thr->c);
+	if (ret)
+		goto err;
+
+	if (test_bit(BCH_FS_errors_fixed, &c->flags)) {
+		bch2_stdio_redirect_printf(&stdio->stdio, false, "%s: errors fixed\n", c->name);
+		ret |= 1;
+	}
+	if (test_bit(BCH_FS_error, &c->flags)) {
+		bch2_stdio_redirect_printf(&stdio->stdio, false, "%s: still has errors\n", c->name);
+		ret |= 4;
+	}
+err:
+	bch2_fs_stop(c);
+	return ret;
+}
+
+static const struct thread_with_stdio_ops bch2_offline_fsck_ops = {
+	.exit		= bch2_fsck_thread_exit,
+	.fn		= bch2_fsck_offline_thread_fn,
+};
+
+long bch2_ioctl_fsck_offline(struct bch_ioctl_fsck_offline __user *user_arg)
+{
+	struct bch_ioctl_fsck_offline arg;
+	struct fsck_thread *thr = NULL;
+	darray_str(devs) = {};
+	long ret = 0;
+
+	if (copy_from_user(&arg, user_arg, sizeof(arg)))
+		return -EFAULT;
+
+	if (arg.flags)
+		return -EINVAL;
+
+	if (!capable(CAP_SYS_ADMIN))
+		return -EPERM;
+
+	for (size_t i = 0; i < arg.nr_devs; i++) {
+		u64 dev_u64;
+		ret = copy_from_user_errcode(&dev_u64, &user_arg->devs[i], sizeof(u64));
+		if (ret)
+			goto err;
+
+		char *dev_str = strndup_user((char __user *)(unsigned long) dev_u64, PATH_MAX);
+		ret = PTR_ERR_OR_ZERO(dev_str);
+		if (ret)
+			goto err;
+
+		ret = darray_push(&devs, dev_str);
+		if (ret) {
+			kfree(dev_str);
+			goto err;
+		}
+	}
+
+	thr = kzalloc(sizeof(*thr), GFP_KERNEL);
+	if (!thr) {
+		ret = -ENOMEM;
+		goto err;
+	}
+
+	thr->opts = bch2_opts_empty();
+
+	if (arg.opts) {
+		char *optstr = strndup_user((char __user *)(unsigned long) arg.opts, 1 << 16);
+		ret =   PTR_ERR_OR_ZERO(optstr) ?:
+			bch2_parse_mount_opts(NULL, &thr->opts, NULL, optstr);
+		if (!IS_ERR(optstr))
+			kfree(optstr);
+
+		if (ret)
+			goto err;
+	}
+
+	opt_set(thr->opts, stdio, (u64)(unsigned long)&thr->thr.stdio);
+	opt_set(thr->opts, read_only, 1);
+	opt_set(thr->opts, ratelimit_errors, 0);
+
+	/* We need request_key() to be called before we punt to kthread: */
+	opt_set(thr->opts, nostart, true);
+
+	bch2_thread_with_stdio_init(&thr->thr, &bch2_offline_fsck_ops);
+
+	thr->c = bch2_fs_open(devs.data, arg.nr_devs, thr->opts);
+
+	if (!IS_ERR(thr->c) &&
+	    thr->c->opts.errors == BCH_ON_ERROR_panic)
+		thr->c->opts.errors = BCH_ON_ERROR_ro;
+
+	ret = __bch2_run_thread_with_stdio(&thr->thr);
+out:
+	darray_for_each(devs, i)
+		kfree(*i);
+	darray_exit(&devs);
+	return ret;
+err:
+	if (thr)
+		bch2_fsck_thread_exit(&thr->thr);
+	pr_err("ret %s", bch2_err_str(ret));
+	goto out;
+}
+
+static int bch2_fsck_online_thread_fn(struct thread_with_stdio *stdio)
+{
+	struct fsck_thread *thr = container_of(stdio, struct fsck_thread, thr);
+	struct bch_fs *c = thr->c;
+
+	c->stdio_filter = current;
+	c->stdio = &thr->thr.stdio;
+
+	/*
+	 * XXX: can we figure out a way to do this without mucking with c->opts?
+	 */
+	unsigned old_fix_errors = c->opts.fix_errors;
+	if (opt_defined(thr->opts, fix_errors))
+		c->opts.fix_errors = thr->opts.fix_errors;
+	else
+		c->opts.fix_errors = FSCK_FIX_ask;
+
+	c->opts.fsck = true;
+	set_bit(BCH_FS_fsck_running, &c->flags);
+
+	c->curr_recovery_pass = BCH_RECOVERY_PASS_check_alloc_info;
+	int ret = bch2_run_online_recovery_passes(c);
+
+	clear_bit(BCH_FS_fsck_running, &c->flags);
+	bch_err_fn(c, ret);
+
+	c->stdio = NULL;
+	c->stdio_filter = NULL;
+	c->opts.fix_errors = old_fix_errors;
+
+	up(&c->online_fsck_mutex);
+	bch2_ro_ref_put(c);
+	return ret;
+}
+
+static const struct thread_with_stdio_ops bch2_online_fsck_ops = {
+	.exit		= bch2_fsck_thread_exit,
+	.fn		= bch2_fsck_online_thread_fn,
+};
+
+long bch2_ioctl_fsck_online(struct bch_fs *c, struct bch_ioctl_fsck_online arg)
+{
+	struct fsck_thread *thr = NULL;
+	long ret = 0;
+
+	if (arg.flags)
+		return -EINVAL;
+
+	if (!capable(CAP_SYS_ADMIN))
+		return -EPERM;
+
+	if (!bch2_ro_ref_tryget(c))
+		return -EROFS;
+
+	if (down_trylock(&c->online_fsck_mutex)) {
+		bch2_ro_ref_put(c);
+		return -EAGAIN;
+	}
+
+	thr = kzalloc(sizeof(*thr), GFP_KERNEL);
+	if (!thr) {
+		ret = -ENOMEM;
+		goto err;
+	}
+
+	thr->c = c;
+	thr->opts = bch2_opts_empty();
+
+	if (arg.opts) {
+		char *optstr = strndup_user((char __user *)(unsigned long) arg.opts, 1 << 16);
+
+		ret =   PTR_ERR_OR_ZERO(optstr) ?:
+			bch2_parse_mount_opts(c, &thr->opts, NULL, optstr);
+		if (!IS_ERR(optstr))
+			kfree(optstr);
+
+		if (ret)
+			goto err;
+	}
+
+	ret = bch2_run_thread_with_stdio(&thr->thr, &bch2_online_fsck_ops);
+err:
+	if (ret < 0) {
+		bch_err_fn(c, ret);
+		if (thr)
+			bch2_fsck_thread_exit(&thr->thr);
+		up(&c->online_fsck_mutex);
+		bch2_ro_ref_put(c);
+	}
+	return ret;
+}
+
+#endif /* NO_BCACHEFS_CHARDEV */
diff --git a/fs/bcachefs/fsck.h b/fs/bcachefs/fsck.h
index 1cca31011530..574948278cd4 100644
--- a/fs/bcachefs/fsck.h
+++ b/fs/bcachefs/fsck.h
@@ -2,6 +2,14 @@
 #ifndef _BCACHEFS_FSCK_H
 #define _BCACHEFS_FSCK_H
 
+#include "str_hash.h"
+
+int bch2_fsck_update_backpointers(struct btree_trans *,
+				  struct snapshots_seen *,
+				  const struct bch_hash_desc,
+				  struct bch_hash_info *,
+				  struct bkey_i *);
+
 int bch2_check_inodes(struct bch_fs *);
 int bch2_check_extents(struct bch_fs *);
 int bch2_check_indirect_extents(struct bch_fs *);
@@ -14,4 +22,7 @@ int bch2_check_directory_structure(struct bch_fs *);
 int bch2_check_nlinks(struct bch_fs *);
 int bch2_fix_reflink_p(struct bch_fs *);
 
+long bch2_ioctl_fsck_offline(struct bch_ioctl_fsck_offline __user *);
+long bch2_ioctl_fsck_online(struct bch_fs *, struct bch_ioctl_fsck_online);
+
 #endif /* _BCACHEFS_FSCK_H */
diff --git a/fs/bcachefs/inode.c b/fs/bcachefs/inode.c
index 039cb7a22244..04ec05206f8c 100644
--- a/fs/bcachefs/inode.c
+++ b/fs/bcachefs/inode.c
@@ -14,6 +14,7 @@
 #include "extent_update.h"
 #include "fs.h"
 #include "inode.h"
+#include "opts.h"
 #include "str_hash.h"
 #include "snapshot.h"
 #include "subvolume.h"
@@ -47,10 +48,10 @@ static int inode_decode_field(const u8 *in, const u8 *end,
 	u8 *p;
 
 	if (in >= end)
-		return -1;
+		return -BCH_ERR_inode_unpack_error;
 
 	if (!*in)
-		return -1;
+		return -BCH_ERR_inode_unpack_error;
 
 	/*
 	 * position of highest set bit indicates number of bytes:
@@ -60,7 +61,7 @@ static int inode_decode_field(const u8 *in, const u8 *end,
 	bytes	= byte_table[shift - 1];
 
 	if (in + bytes > end)
-		return -1;
+		return -BCH_ERR_inode_unpack_error;
 
 	p = (u8 *) be + 16 - bytes;
 	memcpy(p, in, bytes);
@@ -176,7 +177,7 @@ static noinline int bch2_inode_unpack_v1(struct bkey_s_c_inode inode,
 		return ret;						\
 									\
 	if (field_bits > sizeof(unpacked->_name) * 8)			\
-		return -1;						\
+		return -BCH_ERR_inode_unpack_error;			\
 									\
 	unpacked->_name = field[1];					\
 	in += ret;
@@ -217,7 +218,7 @@ static int bch2_inode_unpack_v2(struct bch_inode_unpacked *unpacked,
 									\
 	unpacked->_name = v[0];						\
 	if (v[1] || v[0] != unpacked->_name)				\
-		return -1;						\
+		return -BCH_ERR_inode_unpack_error;			\
 	fieldnr++;
 
 	BCH_INODE_FIELDS_v2()
@@ -268,7 +269,7 @@ static int bch2_inode_unpack_v3(struct bkey_s_c k,
 									\
 	unpacked->_name = v[0];						\
 	if (v[1] || v[0] != unpacked->_name)				\
-		return -1;						\
+		return -BCH_ERR_inode_unpack_error;			\
 	fieldnr++;
 
 	BCH_INODE_FIELDS_v3()
@@ -428,7 +429,7 @@ struct bkey_i *bch2_inode_to_v3(struct btree_trans *trans, struct bkey_i *k)
 }
 
 static int __bch2_inode_validate(struct bch_fs *c, struct bkey_s_c k,
-				 enum bch_validate_flags flags)
+				 struct bkey_validate_context from)
 {
 	struct bch_inode_unpacked unpacked;
 	int ret = 0;
@@ -468,7 +469,7 @@ fsck_err:
 }
 
 int bch2_inode_validate(struct bch_fs *c, struct bkey_s_c k,
-			enum bch_validate_flags flags)
+			struct bkey_validate_context from)
 {
 	struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 	int ret = 0;
@@ -478,13 +479,13 @@ int bch2_inode_validate(struct bch_fs *c, struct bkey_s_c k,
 			 "invalid str hash type (%llu >= %u)",
 			 INODEv1_STR_HASH(inode.v), BCH_STR_HASH_NR);
 
-	ret = __bch2_inode_validate(c, k, flags);
+	ret = __bch2_inode_validate(c, k, from);
 fsck_err:
 	return ret;
 }
 
 int bch2_inode_v2_validate(struct bch_fs *c, struct bkey_s_c k,
-			   enum bch_validate_flags flags)
+			   struct bkey_validate_context from)
 {
 	struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 	int ret = 0;
@@ -494,13 +495,13 @@ int bch2_inode_v2_validate(struct bch_fs *c, struct bkey_s_c k,
 			 "invalid str hash type (%llu >= %u)",
 			 INODEv2_STR_HASH(inode.v), BCH_STR_HASH_NR);
 
-	ret = __bch2_inode_validate(c, k, flags);
+	ret = __bch2_inode_validate(c, k, from);
 fsck_err:
 	return ret;
 }
 
 int bch2_inode_v3_validate(struct bch_fs *c, struct bkey_s_c k,
-			   enum bch_validate_flags flags)
+			   struct bkey_validate_context from)
 {
 	struct bkey_s_c_inode_v3 inode = bkey_s_c_to_inode_v3(k);
 	int ret = 0;
@@ -518,7 +519,7 @@ int bch2_inode_v3_validate(struct bch_fs *c, struct bkey_s_c k,
 			 "invalid str hash type (%llu >= %u)",
 			 INODEv3_STR_HASH(inode.v), BCH_STR_HASH_NR);
 
-	ret = __bch2_inode_validate(c, k, flags);
+	ret = __bch2_inode_validate(c, k, from);
 fsck_err:
 	return ret;
 }
@@ -617,7 +618,7 @@ bch2_bkey_get_iter_snapshot_parent(struct btree_trans *trans, struct btree_iter
 	struct bkey_s_c k;
 	int ret = 0;
 
-	for_each_btree_key_upto_norestart(trans, *iter, btree,
+	for_each_btree_key_max_norestart(trans, *iter, btree,
 					  bpos_successor(pos),
 					  SPOS(pos.inode, pos.offset, U32_MAX),
 					  flags|BTREE_ITER_all_snapshots, k, ret)
@@ -652,7 +653,7 @@ int __bch2_inode_has_child_snapshots(struct btree_trans *trans, struct bpos pos)
 	struct bkey_s_c k;
 	int ret = 0;
 
-	for_each_btree_key_upto_norestart(trans, iter,
+	for_each_btree_key_max_norestart(trans, iter,
 			BTREE_ID_inodes, POS(0, pos.offset), bpos_predecessor(pos),
 			BTREE_ITER_all_snapshots|
 			BTREE_ITER_with_updates, k, ret)
@@ -779,7 +780,7 @@ int bch2_trigger_inode(struct btree_trans *trans,
 }
 
 int bch2_inode_generation_validate(struct bch_fs *c, struct bkey_s_c k,
-				   enum bch_validate_flags flags)
+				   struct bkey_validate_context from)
 {
 	int ret = 0;
 
@@ -798,6 +799,28 @@ void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 	prt_printf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 }
 
+int bch2_inode_alloc_cursor_validate(struct bch_fs *c, struct bkey_s_c k,
+				   struct bkey_validate_context from)
+{
+	int ret = 0;
+
+	bkey_fsck_err_on(k.k->p.inode != LOGGED_OPS_INUM_inode_cursors,
+			 c, inode_alloc_cursor_inode_bad,
+			 "k.p.inode bad");
+fsck_err:
+	return ret;
+}
+
+void bch2_inode_alloc_cursor_to_text(struct printbuf *out, struct bch_fs *c,
+				     struct bkey_s_c k)
+{
+	struct bkey_s_c_inode_alloc_cursor i = bkey_s_c_to_inode_alloc_cursor(k);
+
+	prt_printf(out, "idx %llu generation %llu",
+		   le64_to_cpu(i.v->idx),
+		   le64_to_cpu(i.v->gen));
+}
+
 void bch2_inode_init_early(struct bch_fs *c,
 			   struct bch_inode_unpacked *inode_u)
 {
@@ -858,43 +881,78 @@ static inline u32 bkey_generation(struct bkey_s_c k)
 	}
 }
 
-/*
- * This just finds an empty slot:
- */
-int bch2_inode_create(struct btree_trans *trans,
-		      struct btree_iter *iter,
-		      struct bch_inode_unpacked *inode_u,
-		      u32 snapshot, u64 cpu)
+static struct bkey_i_inode_alloc_cursor *
+bch2_inode_alloc_cursor_get(struct btree_trans *trans, u64 cpu, u64 *min, u64 *max)
 {
 	struct bch_fs *c = trans->c;
-	struct bkey_s_c k;
-	u64 min, max, start, pos, *hint;
-	int ret = 0;
-	unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 
-	if (c->opts.shard_inode_numbers) {
-		bits -= c->inode_shard_bits;
+	u64 cursor_idx = c->opts.inodes_32bit ? 0 : cpu + 1;
 
-		min = (cpu << bits);
-		max = (cpu << bits) | ~(ULLONG_MAX << bits);
+	cursor_idx &= ~(~0ULL << c->opts.shard_inode_numbers_bits);
 
-		min = max_t(u64, min, BLOCKDEV_INODE_MAX);
-		hint = c->unused_inode_hints + cpu;
+	struct btree_iter iter;
+	struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter,
+					BTREE_ID_logged_ops,
+					POS(LOGGED_OPS_INUM_inode_cursors, cursor_idx),
+					BTREE_ITER_cached);
+	int ret = bkey_err(k);
+	if (ret)
+		return ERR_PTR(ret);
+
+	struct bkey_i_inode_alloc_cursor *cursor =
+		k.k->type == KEY_TYPE_inode_alloc_cursor
+		? bch2_bkey_make_mut_typed(trans, &iter, &k, 0, inode_alloc_cursor)
+		: bch2_bkey_alloc(trans, &iter, 0, inode_alloc_cursor);
+	ret = PTR_ERR_OR_ZERO(cursor);
+	if (ret)
+		goto err;
+
+	if (c->opts.inodes_32bit) {
+		*min = BLOCKDEV_INODE_MAX;
+		*max = INT_MAX;
 	} else {
-		min = BLOCKDEV_INODE_MAX;
-		max = ~(ULLONG_MAX << bits);
-		hint = c->unused_inode_hints;
+		cursor->v.bits = c->opts.shard_inode_numbers_bits;
+
+		unsigned bits = 63 - c->opts.shard_inode_numbers_bits;
+
+		*min = max(cpu << bits, (u64) INT_MAX + 1);
+		*max = (cpu << bits) | ~(ULLONG_MAX << bits);
 	}
 
-	start = READ_ONCE(*hint);
+	if (le64_to_cpu(cursor->v.idx)  < *min)
+		cursor->v.idx = cpu_to_le64(*min);
 
-	if (start >= max || start < min)
-		start = min;
+	if (le64_to_cpu(cursor->v.idx) >= *max) {
+		cursor->v.idx = cpu_to_le64(*min);
+		le32_add_cpu(&cursor->v.gen, 1);
+	}
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret ? ERR_PTR(ret) : cursor;
+}
+
+/*
+ * This just finds an empty slot:
+ */
+int bch2_inode_create(struct btree_trans *trans,
+		      struct btree_iter *iter,
+		      struct bch_inode_unpacked *inode_u,
+		      u32 snapshot, u64 cpu)
+{
+	u64 min, max;
+	struct bkey_i_inode_alloc_cursor *cursor =
+		bch2_inode_alloc_cursor_get(trans, cpu, &min, &max);
+	int ret = PTR_ERR_OR_ZERO(cursor);
+	if (ret)
+		return ret;
+
+	u64 start = le64_to_cpu(cursor->v.idx);
+	u64 pos = start;
 
-	pos = start;
 	bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, pos),
 			     BTREE_ITER_all_snapshots|
 			     BTREE_ITER_intent);
+	struct bkey_s_c k;
 again:
 	while ((k = bch2_btree_iter_peek(iter)).k &&
 	       !(ret = bkey_err(k)) &&
@@ -924,6 +982,7 @@ again:
 	/* Retry from start */
 	pos = start = min;
 	bch2_btree_iter_set_pos(iter, POS(0, pos));
+	le32_add_cpu(&cursor->v.gen, 1);
 	goto again;
 found_slot:
 	bch2_btree_iter_set_pos(iter, SPOS(0, pos, snapshot));
@@ -934,9 +993,9 @@ found_slot:
 		return ret;
 	}
 
-	*hint			= k.k->p.offset;
 	inode_u->bi_inum	= k.k->p.offset;
-	inode_u->bi_generation	= bkey_generation(k);
+	inode_u->bi_generation	= le64_to_cpu(cursor->v.gen);
+	cursor->v.idx		= cpu_to_le64(k.k->p.offset + 1);
 	return 0;
 }
 
@@ -966,7 +1025,7 @@ static int bch2_inode_delete_keys(struct btree_trans *trans,
 
 		bch2_btree_iter_set_snapshot(&iter, snapshot);
 
-		k = bch2_btree_iter_peek_upto(&iter, end);
+		k = bch2_btree_iter_peek_max(&iter, end);
 		ret = bkey_err(k);
 		if (ret)
 			goto err;
@@ -998,8 +1057,6 @@ int bch2_inode_rm(struct bch_fs *c, subvol_inum inum)
 {
 	struct btree_trans *trans = bch2_trans_get(c);
 	struct btree_iter iter = { NULL };
-	struct bkey_i_inode_generation delete;
-	struct bch_inode_unpacked inode_u;
 	struct bkey_s_c k;
 	u32 snapshot;
 	int ret;
@@ -1039,13 +1096,7 @@ retry:
 		goto err;
 	}
 
-	bch2_inode_unpack(k, &inode_u);
-
-	bkey_inode_generation_init(&delete.k_i);
-	delete.k.p = iter.pos;
-	delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
-
-	ret   = bch2_trans_update(trans, &iter, &delete.k_i, 0) ?:
+	ret   = bch2_btree_delete_at(trans, &iter, 0) ?:
 		bch2_trans_commit(trans, NULL, NULL,
 				BCH_TRANS_COMMIT_no_enospc);
 err:
@@ -1141,12 +1192,17 @@ struct bch_opts bch2_inode_opts_to_opts(struct bch_inode_unpacked *inode)
 void bch2_inode_opts_get(struct bch_io_opts *opts, struct bch_fs *c,
 			 struct bch_inode_unpacked *inode)
 {
-#define x(_name, _bits)		opts->_name = inode_opt_get(c, inode, _name);
+#define x(_name, _bits)							\
+	if ((inode)->bi_##_name) {					\
+		opts->_name = inode->bi_##_name - 1;			\
+		opts->_name##_from_inode = true;			\
+	} else {							\
+		opts->_name = c->opts._name;				\
+	}
 	BCH_INODE_OPTS()
 #undef x
 
-	if (opts->nocow)
-		opts->compression = opts->background_compression = opts->data_checksum = opts->erasure_code = 0;
+	bch2_io_opts_fixups(opts);
 }
 
 int bch2_inum_opts_get(struct btree_trans *trans, subvol_inum inum, struct bch_io_opts *opts)
@@ -1380,7 +1436,8 @@ again:
 					NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
 		ret = may_delete_deleted_inode(trans, &iter, k.k->p, &need_another_pass);
 		if (ret > 0) {
-			bch_verbose(c, "deleting unlinked inode %llu:%u", k.k->p.offset, k.k->p.snapshot);
+			bch_verbose_ratelimited(c, "deleting unlinked inode %llu:%u",
+						k.k->p.offset, k.k->p.snapshot);
 
 			ret = bch2_inode_rm_snapshot(trans, k.k->p.offset, k.k->p.snapshot);
 			/*
diff --git a/fs/bcachefs/inode.h b/fs/bcachefs/inode.h
index eab82b5eb897..d2e134528f0e 100644
--- a/fs/bcachefs/inode.h
+++ b/fs/bcachefs/inode.h
@@ -7,15 +7,14 @@
 #include "opts.h"
 #include "snapshot.h"
 
-enum bch_validate_flags;
 extern const char * const bch2_inode_opts[];
 
 int bch2_inode_validate(struct bch_fs *, struct bkey_s_c,
-		       enum bch_validate_flags);
+			struct bkey_validate_context);
 int bch2_inode_v2_validate(struct bch_fs *, struct bkey_s_c,
-			  enum bch_validate_flags);
+			   struct bkey_validate_context);
 int bch2_inode_v3_validate(struct bch_fs *, struct bkey_s_c,
-			  enum bch_validate_flags);
+			   struct bkey_validate_context);
 void bch2_inode_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
 int __bch2_inode_has_child_snapshots(struct btree_trans *, struct bpos);
@@ -60,7 +59,7 @@ static inline bool bkey_is_inode(const struct bkey *k)
 }
 
 int bch2_inode_generation_validate(struct bch_fs *, struct bkey_s_c,
-				  enum bch_validate_flags);
+				   struct bkey_validate_context);
 void bch2_inode_generation_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
 #define bch2_bkey_ops_inode_generation ((struct bkey_ops) {	\
@@ -69,6 +68,16 @@ void bch2_inode_generation_to_text(struct printbuf *, struct bch_fs *, struct bk
 	.min_val_size	= 8,					\
 })
 
+int bch2_inode_alloc_cursor_validate(struct bch_fs *, struct bkey_s_c,
+				     struct bkey_validate_context);
+void bch2_inode_alloc_cursor_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
+
+#define bch2_bkey_ops_inode_alloc_cursor ((struct bkey_ops) {	\
+	.key_validate	= bch2_inode_alloc_cursor_validate,	\
+	.val_to_text	= bch2_inode_alloc_cursor_to_text,	\
+	.min_val_size	= 16,					\
+})
+
 #if 0
 typedef struct {
 	u64			lo;
@@ -220,6 +229,20 @@ static inline u32 bch2_inode_flags(struct bkey_s_c k)
 	}
 }
 
+static inline unsigned bkey_inode_mode(struct bkey_s_c k)
+{
+	switch (k.k->type) {
+	case KEY_TYPE_inode:
+		return le16_to_cpu(bkey_s_c_to_inode(k).v->bi_mode);
+	case KEY_TYPE_inode_v2:
+		return le16_to_cpu(bkey_s_c_to_inode_v2(k).v->bi_mode);
+	case KEY_TYPE_inode_v3:
+		return INODEv3_MODE(bkey_s_c_to_inode_v3(k).v);
+	default:
+		return 0;
+	}
+}
+
 /* i_nlink: */
 
 static inline unsigned nlink_bias(umode_t mode)
@@ -249,7 +272,7 @@ static inline void bch2_inode_nlink_set(struct bch_inode_unpacked *bi,
 int bch2_inode_nlink_inc(struct bch_inode_unpacked *);
 void bch2_inode_nlink_dec(struct btree_trans *, struct bch_inode_unpacked *);
 
-static inline bool bch2_inode_should_have_bp(struct bch_inode_unpacked *inode)
+static inline bool bch2_inode_should_have_single_bp(struct bch_inode_unpacked *inode)
 {
 	bool inode_has_bp = inode->bi_dir || inode->bi_dir_offset;
 
@@ -262,6 +285,14 @@ void bch2_inode_opts_get(struct bch_io_opts *, struct bch_fs *,
 			 struct bch_inode_unpacked *);
 int bch2_inum_opts_get(struct btree_trans*, subvol_inum, struct bch_io_opts *);
 
+static inline struct bch_extent_rebalance
+bch2_inode_rebalance_opts_get(struct bch_fs *c, struct bch_inode_unpacked *inode)
+{
+	struct bch_io_opts io_opts;
+	bch2_inode_opts_get(&io_opts, c, inode);
+	return io_opts_to_rebalance_opts(&io_opts);
+}
+
 int bch2_inode_rm_snapshot(struct btree_trans *, u64, u32);
 int bch2_delete_dead_inodes(struct bch_fs *);
 
diff --git a/fs/bcachefs/inode_format.h b/fs/bcachefs/inode_format.h
index 7928d0c6954f..b99a5bf1a75e 100644
--- a/fs/bcachefs/inode_format.h
+++ b/fs/bcachefs/inode_format.h
@@ -101,7 +101,9 @@ struct bch_inode_generation {
 	x(bi_dir_offset,		64)	\
 	x(bi_subvol,			32)	\
 	x(bi_parent_subvol,		32)	\
-	x(bi_nocow,			8)
+	x(bi_nocow,			8)	\
+	x(bi_depth,			32)	\
+	x(bi_inodes_32bit,		8)
 
 /* subset of BCH_INODE_FIELDS */
 #define BCH_INODE_OPTS()			\
@@ -114,7 +116,8 @@ struct bch_inode_generation {
 	x(foreground_target,		16)	\
 	x(background_target,		16)	\
 	x(erasure_code,			16)	\
-	x(nocow,			8)
+	x(nocow,			8)	\
+	x(inodes_32bit,			8)
 
 enum inode_opt_id {
 #define x(name, ...)				\
@@ -164,4 +167,12 @@ LE64_BITMASK(INODEv3_FIELDS_START,
 				struct bch_inode_v3, bi_flags, 31, 36);
 LE64_BITMASK(INODEv3_MODE,	struct bch_inode_v3, bi_flags, 36, 52);
 
+struct bch_inode_alloc_cursor {
+	struct bch_val		v;
+	__u8			bits;
+	__u8			pad;
+	__le32			gen;
+	__le64			idx;
+};
+
 #endif /* _BCACHEFS_INODE_FORMAT_H */
diff --git a/fs/bcachefs/io_misc.c b/fs/bcachefs/io_misc.c
index f283051758d6..5353979117b0 100644
--- a/fs/bcachefs/io_misc.c
+++ b/fs/bcachefs/io_misc.c
@@ -113,11 +113,13 @@ int bch2_extent_fallocate(struct btree_trans *trans,
 err:
 	if (!ret && sectors_allocated)
 		bch2_increment_clock(c, sectors_allocated, WRITE);
-	if (should_print_err(ret))
-		bch_err_inum_offset_ratelimited(c,
-			inum.inum,
-			iter->pos.offset << 9,
-			"%s(): error: %s", __func__, bch2_err_str(ret));
+	if (should_print_err(ret)) {
+		struct printbuf buf = PRINTBUF;
+		bch2_inum_offset_err_msg_trans(trans, &buf, inum, iter->pos.offset << 9);
+		prt_printf(&buf, "fallocate error: %s", bch2_err_str(ret));
+		bch_err_ratelimited(c, "%s", buf.buf);
+		printbuf_exit(&buf);
+	}
 err_noprint:
 	bch2_open_buckets_put(c, &open_buckets);
 	bch2_disk_reservation_put(c, &disk_res);
@@ -164,9 +166,9 @@ int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
 		bch2_btree_iter_set_snapshot(iter, snapshot);
 
 		/*
-		 * peek_upto() doesn't have ideal semantics for extents:
+		 * peek_max() doesn't have ideal semantics for extents:
 		 */
-		k = bch2_btree_iter_peek_upto(iter, end_pos);
+		k = bch2_btree_iter_peek_max(iter, end_pos);
 		if (!k.k)
 			break;
 
@@ -426,8 +428,8 @@ case LOGGED_OP_FINSERT_shift_extents:
 		bch2_btree_iter_set_pos(&iter, SPOS(inum.inum, pos, snapshot));
 
 		k = insert
-			? bch2_btree_iter_peek_prev(&iter)
-			: bch2_btree_iter_peek_upto(&iter, POS(inum.inum, U64_MAX));
+			? bch2_btree_iter_peek_prev_min(&iter, POS(inum.inum, 0))
+			: bch2_btree_iter_peek_max(&iter, POS(inum.inum, U64_MAX));
 		if ((ret = bkey_err(k)))
 			goto btree_err;
 
@@ -461,7 +463,7 @@ case LOGGED_OP_FINSERT_shift_extents:
 
 		op->v.pos = cpu_to_le64(insert ? bkey_start_offset(&delete.k) : delete.k.p.offset);
 
-		ret =   bch2_bkey_set_needs_rebalance(c, copy, &opts) ?:
+		ret =   bch2_bkey_set_needs_rebalance(c, &opts, copy) ?:
 			bch2_btree_insert_trans(trans, BTREE_ID_extents, &delete, 0) ?:
 			bch2_btree_insert_trans(trans, BTREE_ID_extents, copy, 0) ?:
 			bch2_logged_op_update(trans, &op->k_i) ?:
diff --git a/fs/bcachefs/io_read.c b/fs/bcachefs/io_read.c
index b3b934a87c6d..8c7b2d3d779d 100644
--- a/fs/bcachefs/io_read.c
+++ b/fs/bcachefs/io_read.c
@@ -21,6 +21,7 @@
 #include "io_read.h"
 #include "io_misc.h"
 #include "io_write.h"
+#include "reflink.h"
 #include "subvolume.h"
 #include "trace.h"
 
@@ -90,13 +91,18 @@ static const struct rhashtable_params bch_promote_params = {
 	.automatic_shrinking	= true,
 };
 
+static inline bool have_io_error(struct bch_io_failures *failed)
+{
+	return failed && failed->nr;
+}
+
 static inline int should_promote(struct bch_fs *c, struct bkey_s_c k,
 				  struct bpos pos,
 				  struct bch_io_opts opts,
 				  unsigned flags,
 				  struct bch_io_failures *failed)
 {
-	if (!failed) {
+	if (!have_io_error(failed)) {
 		BUG_ON(!opts.promote_target);
 
 		if (!(flags & BCH_READ_MAY_PROMOTE))
@@ -223,7 +229,7 @@ static struct promote_op *__promote_alloc(struct btree_trans *trans,
 
 	struct data_update_opts update_opts = {};
 
-	if (!failed) {
+	if (!have_io_error(failed)) {
 		update_opts.target = opts.promote_target;
 		update_opts.extra_replicas = 1;
 		update_opts.write_flags = BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED;
@@ -231,11 +237,11 @@ static struct promote_op *__promote_alloc(struct btree_trans *trans,
 		update_opts.target = opts.foreground_target;
 
 		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-		unsigned i = 0;
+		unsigned ptr_bit = 1;
 		bkey_for_each_ptr(ptrs, ptr) {
 			if (bch2_dev_io_failures(failed, ptr->dev))
-				update_opts.rewrite_ptrs |= BIT(i);
-			i++;
+				update_opts.rewrite_ptrs |= ptr_bit;
+			ptr_bit <<= 1;
 		}
 	}
 
@@ -285,7 +291,7 @@ static struct promote_op *promote_alloc(struct btree_trans *trans,
 	 * if failed != NULL we're not actually doing a promote, we're
 	 * recovering from an io/checksum error
 	 */
-	bool promote_full = (failed ||
+	bool promote_full = (have_io_error(failed) ||
 			     *read_full ||
 			     READ_ONCE(c->opts.promote_whole_extents));
 	/* data might have to be decompressed in the write path: */
@@ -321,6 +327,20 @@ nopromote:
 
 /* Read */
 
+static int bch2_read_err_msg_trans(struct btree_trans *trans, struct printbuf *out,
+				   struct bch_read_bio *rbio, struct bpos read_pos)
+{
+	return bch2_inum_offset_err_msg_trans(trans, out,
+		(subvol_inum) { rbio->subvol, read_pos.inode },
+		read_pos.offset << 9);
+}
+
+static void bch2_read_err_msg(struct bch_fs *c, struct printbuf *out,
+			      struct bch_read_bio *rbio, struct bpos read_pos)
+{
+	bch2_trans_run(c, bch2_read_err_msg_trans(trans, out, rbio, read_pos));
+}
+
 #define READ_RETRY_AVOID	1
 #define READ_RETRY		2
 #define READ_ERR		3
@@ -499,6 +519,29 @@ static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
 	}
 }
 
+static void bch2_read_io_err(struct work_struct *work)
+{
+	struct bch_read_bio *rbio =
+		container_of(work, struct bch_read_bio, work);
+	struct bio *bio = &rbio->bio;
+	struct bch_fs *c	= rbio->c;
+	struct bch_dev *ca = rbio->have_ioref ? bch2_dev_have_ref(c, rbio->pick.ptr.dev) : NULL;
+	struct printbuf buf = PRINTBUF;
+
+	bch2_read_err_msg(c, &buf, rbio, rbio->read_pos);
+	prt_printf(&buf, "data read error: %s", bch2_blk_status_to_str(bio->bi_status));
+
+	if (ca) {
+		bch2_io_error(ca, BCH_MEMBER_ERROR_read);
+		bch_err_ratelimited(ca, "%s", buf.buf);
+	} else {
+		bch_err_ratelimited(c, "%s", buf.buf);
+	}
+
+	printbuf_exit(&buf);
+	bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
+}
+
 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
 				   struct bch_read_bio *rbio)
 {
@@ -562,6 +605,73 @@ static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
 			     __bch2_rbio_narrow_crcs(trans, rbio));
 }
 
+static void bch2_read_csum_err(struct work_struct *work)
+{
+	struct bch_read_bio *rbio =
+		container_of(work, struct bch_read_bio, work);
+	struct bch_fs *c	= rbio->c;
+	struct bio *src		= &rbio->bio;
+	struct bch_extent_crc_unpacked crc = rbio->pick.crc;
+	struct nonce nonce = extent_nonce(rbio->version, crc);
+	struct bch_csum csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
+	struct printbuf buf = PRINTBUF;
+
+	bch2_read_err_msg(c, &buf, rbio, rbio->read_pos);
+	prt_str(&buf, "data ");
+	bch2_csum_err_msg(&buf, crc.csum_type, rbio->pick.crc.csum, csum);
+
+	struct bch_dev *ca = rbio->have_ioref ? bch2_dev_have_ref(c, rbio->pick.ptr.dev) : NULL;
+	if (ca) {
+		bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
+		bch_err_ratelimited(ca, "%s", buf.buf);
+	} else {
+		bch_err_ratelimited(c, "%s", buf.buf);
+	}
+
+	bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+	printbuf_exit(&buf);
+}
+
+static void bch2_read_decompress_err(struct work_struct *work)
+{
+	struct bch_read_bio *rbio =
+		container_of(work, struct bch_read_bio, work);
+	struct bch_fs *c	= rbio->c;
+	struct printbuf buf = PRINTBUF;
+
+	bch2_read_err_msg(c, &buf, rbio, rbio->read_pos);
+	prt_str(&buf, "decompression error");
+
+	struct bch_dev *ca = rbio->have_ioref ? bch2_dev_have_ref(c, rbio->pick.ptr.dev) : NULL;
+	if (ca)
+		bch_err_ratelimited(ca, "%s", buf.buf);
+	else
+		bch_err_ratelimited(c, "%s", buf.buf);
+
+	bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
+	printbuf_exit(&buf);
+}
+
+static void bch2_read_decrypt_err(struct work_struct *work)
+{
+	struct bch_read_bio *rbio =
+		container_of(work, struct bch_read_bio, work);
+	struct bch_fs *c	= rbio->c;
+	struct printbuf buf = PRINTBUF;
+
+	bch2_read_err_msg(c, &buf, rbio, rbio->read_pos);
+	prt_str(&buf, "decrypt error");
+
+	struct bch_dev *ca = rbio->have_ioref ? bch2_dev_have_ref(c, rbio->pick.ptr.dev) : NULL;
+	if (ca)
+		bch_err_ratelimited(ca, "%s", buf.buf);
+	else
+		bch_err_ratelimited(c, "%s", buf.buf);
+
+	bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
+	printbuf_exit(&buf);
+}
+
 /* Inner part that may run in process context */
 static void __bch2_read_endio(struct work_struct *work)
 {
@@ -668,33 +778,13 @@ csum_err:
 		goto out;
 	}
 
-	struct printbuf buf = PRINTBUF;
-	buf.atomic++;
-	prt_str(&buf, "data ");
-	bch2_csum_err_msg(&buf, crc.csum_type, rbio->pick.crc.csum, csum);
-
-	struct bch_dev *ca = rbio->have_ioref ? bch2_dev_have_ref(c, rbio->pick.ptr.dev) : NULL;
-	if (ca) {
-		bch_err_inum_offset_ratelimited(ca,
-			rbio->read_pos.inode,
-			rbio->read_pos.offset << 9,
-			"data %s", buf.buf);
-		bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
-	}
-	printbuf_exit(&buf);
-	bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
+	bch2_rbio_punt(rbio, bch2_read_csum_err, RBIO_CONTEXT_UNBOUND, system_unbound_wq);
 	goto out;
 decompression_err:
-	bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
-					rbio->read_pos.offset << 9,
-					"decompression error");
-	bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
+	bch2_rbio_punt(rbio, bch2_read_decompress_err, RBIO_CONTEXT_UNBOUND, system_unbound_wq);
 	goto out;
 decrypt_err:
-	bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
-					rbio->read_pos.offset << 9,
-					"decrypt error");
-	bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
+	bch2_rbio_punt(rbio, bch2_read_decrypt_err, RBIO_CONTEXT_UNBOUND, system_unbound_wq);
 	goto out;
 }
 
@@ -715,16 +805,8 @@ static void bch2_read_endio(struct bio *bio)
 	if (!rbio->split)
 		rbio->bio.bi_end_io = rbio->end_io;
 
-	if (bio->bi_status) {
-		if (ca) {
-			bch_err_inum_offset_ratelimited(ca,
-				rbio->read_pos.inode,
-				rbio->read_pos.offset,
-				"data read error: %s",
-				bch2_blk_status_to_str(bio->bi_status));
-			bch2_io_error(ca, BCH_MEMBER_ERROR_read);
-		}
-		bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
+	if (unlikely(bio->bi_status)) {
+		bch2_rbio_punt(rbio, bch2_read_io_err, RBIO_CONTEXT_UNBOUND, system_unbound_wq);
 		return;
 	}
 
@@ -750,45 +832,6 @@ static void bch2_read_endio(struct bio *bio)
 	bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
 }
 
-int __bch2_read_indirect_extent(struct btree_trans *trans,
-				unsigned *offset_into_extent,
-				struct bkey_buf *orig_k)
-{
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	u64 reflink_offset;
-	int ret;
-
-	reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
-		*offset_into_extent;
-
-	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_reflink,
-			       POS(0, reflink_offset), 0);
-	ret = bkey_err(k);
-	if (ret)
-		goto err;
-
-	if (k.k->type != KEY_TYPE_reflink_v &&
-	    k.k->type != KEY_TYPE_indirect_inline_data) {
-		bch_err_inum_offset_ratelimited(trans->c,
-			orig_k->k->k.p.inode,
-			orig_k->k->k.p.offset << 9,
-			"%llu len %u points to nonexistent indirect extent %llu",
-			orig_k->k->k.p.offset,
-			orig_k->k->k.size,
-			reflink_offset);
-		bch2_inconsistent_error(trans->c);
-		ret = -BCH_ERR_missing_indirect_extent;
-		goto err;
-	}
-
-	*offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
-	bch2_bkey_buf_reassemble(orig_k, trans->c, k);
-err:
-	bch2_trans_iter_exit(trans, &iter);
-	return ret;
-}
-
 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
 						   struct bch_dev *ca,
 						   struct bkey_s_c k,
@@ -868,15 +911,24 @@ retry_pick:
 	if (!pick_ret)
 		goto hole;
 
-	if (pick_ret < 0) {
+	if (unlikely(pick_ret < 0)) {
+		struct printbuf buf = PRINTBUF;
+		bch2_read_err_msg_trans(trans, &buf, orig, read_pos);
+		prt_printf(&buf, "no device to read from: %s\n  ", bch2_err_str(pick_ret));
+		bch2_bkey_val_to_text(&buf, c, k);
+
+		bch_err_ratelimited(c, "%s", buf.buf);
+		printbuf_exit(&buf);
+		goto err;
+	}
+
+	if (unlikely(bch2_csum_type_is_encryption(pick.crc.csum_type)) && !c->chacha20) {
 		struct printbuf buf = PRINTBUF;
+		bch2_read_err_msg_trans(trans, &buf, orig, read_pos);
+		prt_printf(&buf, "attempting to read encrypted data without encryption key\n  ");
 		bch2_bkey_val_to_text(&buf, c, k);
 
-		bch_err_inum_offset_ratelimited(c,
-				read_pos.inode, read_pos.offset << 9,
-				"no device to read from: %s\n  %s",
-				bch2_err_str(pick_ret),
-				buf.buf);
+		bch_err_ratelimited(c, "%s", buf.buf);
 		printbuf_exit(&buf);
 		goto err;
 	}
@@ -942,7 +994,7 @@ retry_pick:
 		bounce = true;
 	}
 
-	if (orig->opts.promote_target)// || failed)
+	if (orig->opts.promote_target || have_io_error(failed))
 		promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,
 					&rbio, &bounce, &read_full, failed);
 
@@ -1062,11 +1114,15 @@ get_bio:
 	}
 
 	if (!rbio->pick.idx) {
-		if (!rbio->have_ioref) {
-			bch_err_inum_offset_ratelimited(c,
-					read_pos.inode,
-					read_pos.offset << 9,
-					"no device to read from");
+		if (unlikely(!rbio->have_ioref)) {
+			struct printbuf buf = PRINTBUF;
+			bch2_read_err_msg_trans(trans, &buf, rbio, read_pos);
+			prt_printf(&buf, "no device to read from:\n  ");
+			bch2_bkey_val_to_text(&buf, c, k);
+
+			bch_err_ratelimited(c, "%s", buf.buf);
+			printbuf_exit(&buf);
+
 			bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
 			goto out;
 		}
@@ -1164,7 +1220,6 @@ void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
 			     BTREE_ITER_slots);
 
 	while (1) {
-		unsigned bytes, sectors, offset_into_extent;
 		enum btree_id data_btree = BTREE_ID_extents;
 
 		bch2_trans_begin(trans);
@@ -1184,9 +1239,9 @@ void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
 		if (ret)
 			goto err;
 
-		offset_into_extent = iter.pos.offset -
+		s64 offset_into_extent = iter.pos.offset -
 			bkey_start_offset(k.k);
-		sectors = k.k->size - offset_into_extent;
+		unsigned sectors = k.k->size - offset_into_extent;
 
 		bch2_bkey_buf_reassemble(&sk, c, k);
 
@@ -1201,9 +1256,9 @@ void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
 		 * With indirect extents, the amount of data to read is the min
 		 * of the original extent and the indirect extent:
 		 */
-		sectors = min(sectors, k.k->size - offset_into_extent);
+		sectors = min_t(unsigned, sectors, k.k->size - offset_into_extent);
 
-		bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
+		unsigned bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
 		swap(bvec_iter.bi_size, bytes);
 
 		if (bvec_iter.bi_size == bytes)
@@ -1229,16 +1284,20 @@ err:
 	}
 
 	bch2_trans_iter_exit(trans, &iter);
-	bch2_trans_put(trans);
-	bch2_bkey_buf_exit(&sk, c);
 
 	if (ret) {
-		bch_err_inum_offset_ratelimited(c, inum.inum,
-						bvec_iter.bi_sector << 9,
-						"read error %i from btree lookup", ret);
+		struct printbuf buf = PRINTBUF;
+		bch2_inum_offset_err_msg_trans(trans, &buf, inum, bvec_iter.bi_sector << 9);
+		prt_printf(&buf, "read error %i from btree lookup", ret);
+		bch_err_ratelimited(c, "%s", buf.buf);
+		printbuf_exit(&buf);
+
 		rbio->bio.bi_status = BLK_STS_IOERR;
 		bch2_rbio_done(rbio);
 	}
+
+	bch2_trans_put(trans);
+	bch2_bkey_buf_exit(&sk, c);
 }
 
 void bch2_fs_io_read_exit(struct bch_fs *c)
diff --git a/fs/bcachefs/io_read.h b/fs/bcachefs/io_read.h
index d9c18bb7d403..a82e8a94ccb6 100644
--- a/fs/bcachefs/io_read.h
+++ b/fs/bcachefs/io_read.h
@@ -3,6 +3,7 @@
 #define _BCACHEFS_IO_READ_H
 
 #include "bkey_buf.h"
+#include "reflink.h"
 
 struct bch_read_bio {
 	struct bch_fs		*c;
@@ -79,19 +80,32 @@ struct bch_devs_mask;
 struct cache_promote_op;
 struct extent_ptr_decoded;
 
-int __bch2_read_indirect_extent(struct btree_trans *, unsigned *,
-				struct bkey_buf *);
-
 static inline int bch2_read_indirect_extent(struct btree_trans *trans,
 					    enum btree_id *data_btree,
-					    unsigned *offset_into_extent,
-					    struct bkey_buf *k)
+					    s64 *offset_into_extent,
+					    struct bkey_buf *extent)
 {
-	if (k->k->k.type != KEY_TYPE_reflink_p)
+	if (extent->k->k.type != KEY_TYPE_reflink_p)
 		return 0;
 
 	*data_btree = BTREE_ID_reflink;
-	return __bch2_read_indirect_extent(trans, offset_into_extent, k);
+	struct btree_iter iter;
+	struct bkey_s_c k = bch2_lookup_indirect_extent(trans, &iter,
+						offset_into_extent,
+						bkey_i_to_s_c_reflink_p(extent->k),
+						true, 0);
+	int ret = bkey_err(k);
+	if (ret)
+		return ret;
+
+	if (bkey_deleted(k.k)) {
+		bch2_trans_iter_exit(trans, &iter);
+		return -BCH_ERR_missing_indirect_extent;
+	}
+
+	bch2_bkey_buf_reassemble(extent, trans->c, k);
+	bch2_trans_iter_exit(trans, &iter);
+	return 0;
 }
 
 enum bch_read_flags {
diff --git a/fs/bcachefs/io_write.c b/fs/bcachefs/io_write.c
index 96720adcfee0..3e71860f66b9 100644
--- a/fs/bcachefs/io_write.c
+++ b/fs/bcachefs/io_write.c
@@ -164,7 +164,7 @@ int bch2_sum_sector_overwrites(struct btree_trans *trans,
 
 	bch2_trans_copy_iter(&iter, extent_iter);
 
-	for_each_btree_key_upto_continue_norestart(iter,
+	for_each_btree_key_max_continue_norestart(iter,
 				new->k.p, BTREE_ITER_slots, old, ret) {
 		s64 sectors = min(new->k.p.offset, old.k->p.offset) -
 			max(bkey_start_offset(&new->k),
@@ -216,6 +216,7 @@ static inline int bch2_extent_update_i_size_sectors(struct btree_trans *trans,
 			      SPOS(0,
 				   extent_iter->pos.inode,
 				   extent_iter->snapshot),
+			      BTREE_ITER_intent|
 			      BTREE_ITER_cached);
 	int ret = bkey_err(k);
 	if (unlikely(ret))
@@ -369,7 +370,7 @@ static int bch2_write_index_default(struct bch_write_op *op)
 				     bkey_start_pos(&sk.k->k),
 				     BTREE_ITER_slots|BTREE_ITER_intent);
 
-		ret =   bch2_bkey_set_needs_rebalance(c, sk.k, &op->opts) ?:
+		ret =   bch2_bkey_set_needs_rebalance(c, &op->opts, sk.k) ?:
 			bch2_extent_update(trans, inum, &iter, sk.k,
 					&op->res,
 					op->new_i_size, &op->i_sectors_delta,
@@ -395,6 +396,21 @@ static int bch2_write_index_default(struct bch_write_op *op)
 
 /* Writes */
 
+static void __bch2_write_op_error(struct printbuf *out, struct bch_write_op *op,
+				  u64 offset)
+{
+	bch2_inum_offset_err_msg(op->c, out,
+				 (subvol_inum) { op->subvol, op->pos.inode, },
+				 offset << 9);
+	prt_printf(out, "write error%s: ",
+		   op->flags & BCH_WRITE_MOVE ? "(internal move)" : "");
+}
+
+static void bch2_write_op_error(struct printbuf *out, struct bch_write_op *op)
+{
+	__bch2_write_op_error(out, op, op->pos.offset);
+}
+
 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
 			       enum bch_data_type type,
 			       const struct bkey_i *k,
@@ -531,14 +547,14 @@ static void __bch2_write_index(struct bch_write_op *op)
 
 		op->written += sectors_start - keylist_sectors(keys);
 
-		if (ret && !bch2_err_matches(ret, EROFS)) {
+		if (unlikely(ret && !bch2_err_matches(ret, EROFS))) {
 			struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
 
-			bch_err_inum_offset_ratelimited(c,
-				insert->k.p.inode, insert->k.p.offset << 9,
-				"%s write error while doing btree update: %s",
-				op->flags & BCH_WRITE_MOVE ? "move" : "user",
-				bch2_err_str(ret));
+			struct printbuf buf = PRINTBUF;
+			__bch2_write_op_error(&buf, op, bkey_start_offset(&insert->k));
+			prt_printf(&buf, "btree update error: %s", bch2_err_str(ret));
+			bch_err_ratelimited(c, "%s", buf.buf);
+			printbuf_exit(&buf);
 		}
 
 		if (ret)
@@ -621,9 +637,7 @@ void bch2_write_point_do_index_updates(struct work_struct *work)
 
 	while (1) {
 		spin_lock_irq(&wp->writes_lock);
-		op = list_first_entry_or_null(&wp->writes, struct bch_write_op, wp_list);
-		if (op)
-			list_del(&op->wp_list);
+		op = list_pop_entry(&wp->writes, struct bch_write_op, wp_list);
 		wp_update_state(wp, op != NULL);
 		spin_unlock_irq(&wp->writes_lock);
 
@@ -1080,11 +1094,14 @@ do_write:
 	*_dst = dst;
 	return more;
 csum_err:
-	bch_err_inum_offset_ratelimited(c,
-		op->pos.inode,
-		op->pos.offset << 9,
-		"%s write error: error verifying existing checksum while rewriting existing data (memory corruption?)",
-		op->flags & BCH_WRITE_MOVE ? "move" : "user");
+	{
+		struct printbuf buf = PRINTBUF;
+		bch2_write_op_error(&buf, op);
+		prt_printf(&buf, "error verifying existing checksum while rewriting existing data (memory corruption?)");
+		bch_err_ratelimited(c, "%s", buf.buf);
+		printbuf_exit(&buf);
+	}
+
 	ret = -EIO;
 err:
 	if (to_wbio(dst)->bounce)
@@ -1165,7 +1182,7 @@ static void bch2_nocow_write_convert_unwritten(struct bch_write_op *op)
 	struct btree_trans *trans = bch2_trans_get(c);
 
 	for_each_keylist_key(&op->insert_keys, orig) {
-		int ret = for_each_btree_key_upto_commit(trans, iter, BTREE_ID_extents,
+		int ret = for_each_btree_key_max_commit(trans, iter, BTREE_ID_extents,
 				     bkey_start_pos(&orig->k), orig->k.p,
 				     BTREE_ITER_intent, k,
 				     NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
@@ -1175,11 +1192,11 @@ static void bch2_nocow_write_convert_unwritten(struct bch_write_op *op)
 		if (ret && !bch2_err_matches(ret, EROFS)) {
 			struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
 
-			bch_err_inum_offset_ratelimited(c,
-				insert->k.p.inode, insert->k.p.offset << 9,
-				"%s write error while doing btree update: %s",
-				op->flags & BCH_WRITE_MOVE ? "move" : "user",
-				bch2_err_str(ret));
+			struct printbuf buf = PRINTBUF;
+			__bch2_write_op_error(&buf, op, bkey_start_offset(&insert->k));
+			prt_printf(&buf, "btree update error: %s", bch2_err_str(ret));
+			bch_err_ratelimited(c, "%s", buf.buf);
+			printbuf_exit(&buf);
 		}
 
 		if (ret) {
@@ -1339,17 +1356,19 @@ err:
 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 		goto retry;
 
+	bch2_trans_put(trans);
+	darray_exit(&buckets);
+
 	if (ret) {
-		bch_err_inum_offset_ratelimited(c,
-			op->pos.inode, op->pos.offset << 9,
-			"%s: btree lookup error %s", __func__, bch2_err_str(ret));
+		struct printbuf buf = PRINTBUF;
+		bch2_write_op_error(&buf, op);
+		prt_printf(&buf, "%s(): btree lookup error: %s", __func__, bch2_err_str(ret));
+		bch_err_ratelimited(c, "%s", buf.buf);
+		printbuf_exit(&buf);
 		op->error = ret;
 		op->flags |= BCH_WRITE_SUBMITTED;
 	}
 
-	bch2_trans_put(trans);
-	darray_exit(&buckets);
-
 	/* fallback to cow write path? */
 	if (!(op->flags & BCH_WRITE_SUBMITTED)) {
 		closure_sync(&op->cl);
@@ -1462,14 +1481,14 @@ err:
 		if (ret <= 0) {
 			op->flags |= BCH_WRITE_SUBMITTED;
 
-			if (ret < 0) {
-				if (!(op->flags & BCH_WRITE_ALLOC_NOWAIT))
-					bch_err_inum_offset_ratelimited(c,
-						op->pos.inode,
-						op->pos.offset << 9,
-						"%s(): %s error: %s", __func__,
-						op->flags & BCH_WRITE_MOVE ? "move" : "user",
-						bch2_err_str(ret));
+			if (unlikely(ret < 0)) {
+				if (!(op->flags & BCH_WRITE_ALLOC_NOWAIT)) {
+					struct printbuf buf = PRINTBUF;
+					bch2_write_op_error(&buf, op);
+					prt_printf(&buf, "%s(): %s", __func__, bch2_err_str(ret));
+					bch_err_ratelimited(c, "%s", buf.buf);
+					printbuf_exit(&buf);
+				}
 				op->error = ret;
 				break;
 			}
@@ -1595,12 +1614,11 @@ CLOSURE_CALLBACK(bch2_write)
 	bch2_keylist_init(&op->insert_keys, op->inline_keys);
 	wbio_init(bio)->put_bio = false;
 
-	if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
-		bch_err_inum_offset_ratelimited(c,
-			op->pos.inode,
-			op->pos.offset << 9,
-			"%s write error: misaligned write",
-			op->flags & BCH_WRITE_MOVE ? "move" : "user");
+	if (unlikely(bio->bi_iter.bi_size & (c->opts.block_size - 1))) {
+		struct printbuf buf = PRINTBUF;
+		bch2_write_op_error(&buf, op);
+		prt_printf(&buf, "misaligned write");
+		printbuf_exit(&buf);
 		op->error = -EIO;
 		goto err;
 	}
diff --git a/fs/bcachefs/journal.c b/fs/bcachefs/journal.c
index 2dc0d60c1745..2cd20114b74b 100644
--- a/fs/bcachefs/journal.c
+++ b/fs/bcachefs/journal.c
@@ -217,6 +217,12 @@ void bch2_journal_buf_put_final(struct journal *j, u64 seq)
 	if (__bch2_journal_pin_put(j, seq))
 		bch2_journal_reclaim_fast(j);
 	bch2_journal_do_writes(j);
+
+	/*
+	 * for __bch2_next_write_buffer_flush_journal_buf(), when quiescing an
+	 * open journal entry
+	 */
+	wake_up(&j->wait);
 }
 
 /*
@@ -251,6 +257,9 @@ static void __journal_entry_close(struct journal *j, unsigned closed_val, bool t
 	if (!__journal_entry_is_open(old))
 		return;
 
+	if (old.cur_entry_offset == JOURNAL_ENTRY_BLOCKED_VAL)
+		old.cur_entry_offset = j->cur_entry_offset_if_blocked;
+
 	/* Close out old buffer: */
 	buf->data->u64s		= cpu_to_le32(old.cur_entry_offset);
 
@@ -373,6 +382,10 @@ static int journal_entry_open(struct journal *j)
 	if (nr_unwritten_journal_entries(j) == ARRAY_SIZE(j->buf))
 		return JOURNAL_ERR_max_in_flight;
 
+	if (bch2_fs_fatal_err_on(journal_cur_seq(j) >= JOURNAL_SEQ_MAX,
+				 c, "cannot start: journal seq overflow"))
+		return JOURNAL_ERR_insufficient_devices; /* -EROFS */
+
 	BUG_ON(!j->cur_entry_sectors);
 
 	buf->expires		=
@@ -664,7 +677,7 @@ out:
  * @seq:	seq to flush
  * @parent:	closure object to wait with
  * Returns:	1 if @seq has already been flushed, 0 if @seq is being flushed,
- *		-EIO if @seq will never be flushed
+ *		-BCH_ERR_journal_flush_err if @seq will never be flushed
  *
  * Like bch2_journal_wait_on_seq, except that it triggers a write immediately if
  * necessary
@@ -687,7 +700,7 @@ int bch2_journal_flush_seq_async(struct journal *j, u64 seq,
 
 	/* Recheck under lock: */
 	if (j->err_seq && seq >= j->err_seq) {
-		ret = -EIO;
+		ret = -BCH_ERR_journal_flush_err;
 		goto out;
 	}
 
@@ -794,10 +807,11 @@ int bch2_journal_flush(struct journal *j)
 }
 
 /*
- * bch2_journal_noflush_seq - tell the journal not to issue any flushes before
+ * bch2_journal_noflush_seq - ask the journal not to issue any flushes in the
+ * range [start, end)
  * @seq
  */
-bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
+bool bch2_journal_noflush_seq(struct journal *j, u64 start, u64 end)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	u64 unwritten_seq;
@@ -806,15 +820,15 @@ bool bch2_journal_noflush_seq(struct journal *j, u64 seq)
 	if (!(c->sb.features & (1ULL << BCH_FEATURE_journal_no_flush)))
 		return false;
 
-	if (seq <= c->journal.flushed_seq_ondisk)
+	if (c->journal.flushed_seq_ondisk >= start)
 		return false;
 
 	spin_lock(&j->lock);
-	if (seq <= c->journal.flushed_seq_ondisk)
+	if (c->journal.flushed_seq_ondisk >= start)
 		goto out;
 
 	for (unwritten_seq = journal_last_unwritten_seq(j);
-	     unwritten_seq < seq;
+	     unwritten_seq < end;
 	     unwritten_seq++) {
 		struct journal_buf *buf = journal_seq_to_buf(j, unwritten_seq);
 
@@ -831,19 +845,14 @@ out:
 	return ret;
 }
 
-int bch2_journal_meta(struct journal *j)
+static int __bch2_journal_meta(struct journal *j)
 {
-	struct journal_buf *buf;
-	struct journal_res res;
-	int ret;
-
-	memset(&res, 0, sizeof(res));
-
-	ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
+	struct journal_res res = {};
+	int ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
 	if (ret)
 		return ret;
 
-	buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
+	struct journal_buf *buf = j->buf + (res.seq & JOURNAL_BUF_MASK);
 	buf->must_flush = true;
 
 	if (!buf->flush_time) {
@@ -856,27 +865,70 @@ int bch2_journal_meta(struct journal *j)
 	return bch2_journal_flush_seq(j, res.seq, TASK_UNINTERRUPTIBLE);
 }
 
+int bch2_journal_meta(struct journal *j)
+{
+	struct bch_fs *c = container_of(j, struct bch_fs, journal);
+
+	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_journal))
+		return -EROFS;
+
+	int ret = __bch2_journal_meta(j);
+	bch2_write_ref_put(c, BCH_WRITE_REF_journal);
+	return ret;
+}
+
 /* block/unlock the journal: */
 
 void bch2_journal_unblock(struct journal *j)
 {
 	spin_lock(&j->lock);
-	j->blocked--;
+	if (!--j->blocked &&
+	    j->cur_entry_offset_if_blocked < JOURNAL_ENTRY_CLOSED_VAL &&
+	    j->reservations.cur_entry_offset == JOURNAL_ENTRY_BLOCKED_VAL) {
+		union journal_res_state old, new;
+
+		old.v = atomic64_read(&j->reservations.counter);
+		do {
+			new.v = old.v;
+			new.cur_entry_offset = j->cur_entry_offset_if_blocked;
+		} while (!atomic64_try_cmpxchg(&j->reservations.counter, &old.v, new.v));
+	}
 	spin_unlock(&j->lock);
 
 	journal_wake(j);
 }
 
+static void __bch2_journal_block(struct journal *j)
+{
+	if (!j->blocked++) {
+		union journal_res_state old, new;
+
+		old.v = atomic64_read(&j->reservations.counter);
+		do {
+			j->cur_entry_offset_if_blocked = old.cur_entry_offset;
+
+			if (j->cur_entry_offset_if_blocked >= JOURNAL_ENTRY_CLOSED_VAL)
+				break;
+
+			new.v = old.v;
+			new.cur_entry_offset = JOURNAL_ENTRY_BLOCKED_VAL;
+		} while (!atomic64_try_cmpxchg(&j->reservations.counter, &old.v, new.v));
+
+		journal_cur_buf(j)->data->u64s = cpu_to_le32(old.cur_entry_offset);
+	}
+}
+
 void bch2_journal_block(struct journal *j)
 {
 	spin_lock(&j->lock);
-	j->blocked++;
+	__bch2_journal_block(j);
 	spin_unlock(&j->lock);
 
 	journal_quiesce(j);
 }
 
-static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
+static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct journal *j,
+						u64 max_seq, bool *blocked)
 {
 	struct journal_buf *ret = NULL;
 
@@ -893,13 +945,17 @@ static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct jou
 		struct journal_buf *buf = j->buf + idx;
 
 		if (buf->need_flush_to_write_buffer) {
-			if (seq == journal_cur_seq(j))
-				__journal_entry_close(j, JOURNAL_ENTRY_CLOSED_VAL, true);
-
 			union journal_res_state s;
 			s.v = atomic64_read_acquire(&j->reservations.counter);
 
-			ret = journal_state_count(s, idx)
+			unsigned open = seq == journal_cur_seq(j) && __journal_entry_is_open(s);
+
+			if (open && !*blocked) {
+				__bch2_journal_block(j);
+				*blocked = true;
+			}
+
+			ret = journal_state_count(s, idx) > open
 				? ERR_PTR(-EAGAIN)
 				: buf;
 			break;
@@ -912,11 +968,17 @@ static struct journal_buf *__bch2_next_write_buffer_flush_journal_buf(struct jou
 	return ret;
 }
 
-struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq)
+struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j,
+							     u64 max_seq, bool *blocked)
 {
 	struct journal_buf *ret;
+	*blocked = false;
+
+	wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j,
+						max_seq, blocked)) != ERR_PTR(-EAGAIN));
+	if (IS_ERR_OR_NULL(ret) && *blocked)
+		bch2_journal_unblock(j);
 
-	wait_event(j->wait, (ret = __bch2_next_write_buffer_flush_journal_buf(j, max_seq)) != ERR_PTR(-EAGAIN));
 	return ret;
 }
 
@@ -945,19 +1007,17 @@ static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 	}
 
 	for (nr_got = 0; nr_got < nr_want; nr_got++) {
-		if (new_fs) {
-			bu[nr_got] = bch2_bucket_alloc_new_fs(ca);
-			if (bu[nr_got] < 0) {
-				ret = -BCH_ERR_ENOSPC_bucket_alloc;
-				break;
-			}
-		} else {
-			ob[nr_got] = bch2_bucket_alloc(c, ca, BCH_WATERMARK_normal,
-						       BCH_DATA_journal, cl);
-			ret = PTR_ERR_OR_ZERO(ob[nr_got]);
-			if (ret)
-				break;
+		enum bch_watermark watermark = new_fs
+			? BCH_WATERMARK_btree
+			: BCH_WATERMARK_normal;
 
+		ob[nr_got] = bch2_bucket_alloc(c, ca, watermark,
+					       BCH_DATA_journal, cl);
+		ret = PTR_ERR_OR_ZERO(ob[nr_got]);
+		if (ret)
+			break;
+
+		if (!new_fs) {
 			ret = bch2_trans_run(c,
 				bch2_trans_mark_metadata_bucket(trans, ca,
 						ob[nr_got]->bucket, BCH_DATA_journal,
@@ -967,9 +1027,9 @@ static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 				bch_err_msg(c, ret, "marking new journal buckets");
 				break;
 			}
-
-			bu[nr_got] = ob[nr_got]->bucket;
 		}
+
+		bu[nr_got] = ob[nr_got]->bucket;
 	}
 
 	if (!nr_got)
@@ -1009,8 +1069,7 @@ static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
 	if (ret)
 		goto err_unblock;
 
-	if (!new_fs)
-		bch2_write_super(c);
+	bch2_write_super(c);
 
 	/* Commit: */
 	if (c)
@@ -1044,9 +1103,8 @@ err_unblock:
 						bu[i], BCH_DATA_free, 0,
 						BTREE_TRIGGER_transactional));
 err_free:
-	if (!new_fs)
-		for (i = 0; i < nr_got; i++)
-			bch2_open_bucket_put(c, ob[i]);
+	for (i = 0; i < nr_got; i++)
+		bch2_open_bucket_put(c, ob[i]);
 
 	kfree(new_bucket_seq);
 	kfree(new_buckets);
@@ -1193,7 +1251,7 @@ void bch2_fs_journal_stop(struct journal *j)
 	 * Always write a new journal entry, to make sure the clock hands are up
 	 * to date (and match the superblock)
 	 */
-	bch2_journal_meta(j);
+	__bch2_journal_meta(j);
 
 	journal_quiesce(j);
 	cancel_delayed_work_sync(&j->write_work);
@@ -1217,6 +1275,11 @@ int bch2_fs_journal_start(struct journal *j, u64 cur_seq)
 	bool had_entries = false;
 	u64 last_seq = cur_seq, nr, seq;
 
+	if (cur_seq >= JOURNAL_SEQ_MAX) {
+		bch_err(c, "cannot start: journal seq overflow");
+		return -EINVAL;
+	}
+
 	genradix_for_each_reverse(&c->journal_entries, iter, _i) {
 		i = *_i;
 
@@ -1474,6 +1537,9 @@ void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
 	case JOURNAL_ENTRY_CLOSED_VAL:
 		prt_printf(out, "closed\n");
 		break;
+	case JOURNAL_ENTRY_BLOCKED_VAL:
+		prt_printf(out, "blocked\n");
+		break;
 	default:
 		prt_printf(out, "%u/%u\n", s.cur_entry_offset, j->cur_entry_u64s);
 		break;
@@ -1499,6 +1565,9 @@ void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
 	printbuf_indent_sub(out, 2);
 
 	for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
+		if (!ca->mi.durability)
+			continue;
+
 		struct journal_device *ja = &ca->journal;
 
 		if (!test_bit(ca->dev_idx, c->rw_devs[BCH_DATA_journal].d))
@@ -1508,6 +1577,7 @@ void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
 			continue;
 
 		prt_printf(out, "dev %u:\n",			ca->dev_idx);
+		prt_printf(out, "durability %u:\n",		ca->mi.durability);
 		printbuf_indent_add(out, 2);
 		prt_printf(out, "nr\t%u\n",			ja->nr);
 		prt_printf(out, "bucket size\t%u\n",		ca->mi.bucket_size);
@@ -1519,6 +1589,8 @@ void __bch2_journal_debug_to_text(struct printbuf *out, struct journal *j)
 		printbuf_indent_sub(out, 2);
 	}
 
+	prt_printf(out, "replicas want %u need %u\n", c->opts.metadata_replicas, c->opts.metadata_replicas_required);
+
 	rcu_read_unlock();
 
 	--out->atomic;
diff --git a/fs/bcachefs/journal.h b/fs/bcachefs/journal.h
index 2762be6f9814..cb0df0663946 100644
--- a/fs/bcachefs/journal.h
+++ b/fs/bcachefs/journal.h
@@ -285,7 +285,8 @@ static inline void bch2_journal_buf_put(struct journal *j, unsigned idx, u64 seq
 		spin_lock(&j->lock);
 		bch2_journal_buf_put_final(j, seq);
 		spin_unlock(&j->lock);
-	}
+	} else if (unlikely(s.cur_entry_offset == JOURNAL_ENTRY_BLOCKED_VAL))
+		wake_up(&j->wait);
 }
 
 /*
@@ -403,7 +404,7 @@ void bch2_journal_flush_async(struct journal *, struct closure *);
 
 int bch2_journal_flush_seq(struct journal *, u64, unsigned);
 int bch2_journal_flush(struct journal *);
-bool bch2_journal_noflush_seq(struct journal *, u64);
+bool bch2_journal_noflush_seq(struct journal *, u64, u64);
 int bch2_journal_meta(struct journal *);
 
 void bch2_journal_halt(struct journal *);
@@ -411,7 +412,7 @@ void bch2_journal_halt(struct journal *);
 static inline int bch2_journal_error(struct journal *j)
 {
 	return j->reservations.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL
-		? -EIO : 0;
+		? -BCH_ERR_journal_shutdown : 0;
 }
 
 struct bch_dev;
@@ -424,7 +425,7 @@ static inline void bch2_journal_set_replay_done(struct journal *j)
 
 void bch2_journal_unblock(struct journal *);
 void bch2_journal_block(struct journal *);
-struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *j, u64 max_seq);
+struct journal_buf *bch2_next_write_buffer_flush_journal_buf(struct journal *, u64, bool *);
 
 void __bch2_journal_debug_to_text(struct printbuf *, struct journal *);
 void bch2_journal_debug_to_text(struct printbuf *, struct journal *);
diff --git a/fs/bcachefs/journal_io.c b/fs/bcachefs/journal_io.c
index fb35dd336331..7f2efe85a805 100644
--- a/fs/bcachefs/journal_io.c
+++ b/fs/bcachefs/journal_io.c
@@ -17,6 +17,8 @@
 #include "sb-clean.h"
 #include "trace.h"
 
+#include <linux/string_choices.h>
+
 void bch2_journal_pos_from_member_info_set(struct bch_fs *c)
 {
 	lockdep_assert_held(&c->sb_lock);
@@ -299,7 +301,7 @@ static void journal_entry_err_msg(struct printbuf *out,
 	journal_entry_err_msg(&_buf, version, jset, entry);		\
 	prt_printf(&_buf, msg, ##__VA_ARGS__);				\
 									\
-	switch (flags & BCH_VALIDATE_write) {				\
+	switch (from.flags & BCH_VALIDATE_write) {			\
 	case READ:							\
 		mustfix_fsck_err(c, _err, "%s", _buf.buf);		\
 		break;							\
@@ -325,11 +327,11 @@ static void journal_entry_err_msg(struct printbuf *out,
 static int journal_validate_key(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
-				unsigned level, enum btree_id btree_id,
 				struct bkey_i *k,
-				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from,
+				unsigned version, int big_endian)
 {
+	enum bch_validate_flags flags = from.flags;
 	int write = flags & BCH_VALIDATE_write;
 	void *next = vstruct_next(entry);
 	int ret = 0;
@@ -364,11 +366,10 @@ static int journal_validate_key(struct bch_fs *c,
 	}
 
 	if (!write)
-		bch2_bkey_compat(level, btree_id, version, big_endian,
+		bch2_bkey_compat(from.level, from.btree, version, big_endian,
 				 write, NULL, bkey_to_packed(k));
 
-	ret = bch2_bkey_validate(c, bkey_i_to_s_c(k),
-				 __btree_node_type(level, btree_id), write);
+	ret = bch2_bkey_validate(c, bkey_i_to_s_c(k), from);
 	if (ret == -BCH_ERR_fsck_delete_bkey) {
 		le16_add_cpu(&entry->u64s, -((u16) k->k.u64s));
 		memmove(k, bkey_next(k), next - (void *) bkey_next(k));
@@ -379,7 +380,7 @@ static int journal_validate_key(struct bch_fs *c,
 		goto fsck_err;
 
 	if (write)
-		bch2_bkey_compat(level, btree_id, version, big_endian,
+		bch2_bkey_compat(from.level, from.btree, version, big_endian,
 				 write, NULL, bkey_to_packed(k));
 fsck_err:
 	return ret;
@@ -389,16 +390,15 @@ static int journal_entry_btree_keys_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct bkey_i *k = entry->start;
 
+	from.level	= entry->level;
+	from.btree	= entry->btree_id;
+
 	while (k != vstruct_last(entry)) {
-		int ret = journal_validate_key(c, jset, entry,
-					       entry->level,
-					       entry->btree_id,
-					       k, version, big_endian,
-					       flags|BCH_VALIDATE_journal);
+		int ret = journal_validate_key(c, jset, entry, k, from, version, big_endian);
 		if (ret == FSCK_DELETED_KEY)
 			continue;
 		else if (ret)
@@ -421,7 +421,8 @@ static void journal_entry_btree_keys_to_text(struct printbuf *out, struct bch_fs
 			bch2_prt_jset_entry_type(out, entry->type);
 			prt_str(out, ": ");
 		}
-		prt_printf(out, "btree=%s l=%u ", bch2_btree_id_str(entry->btree_id), entry->level);
+		bch2_btree_id_level_to_text(out, entry->btree_id, entry->level);
+		prt_char(out, ' ');
 		bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(k));
 		first = false;
 	}
@@ -431,11 +432,15 @@ static int journal_entry_btree_root_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct bkey_i *k = entry->start;
 	int ret = 0;
 
+	from.root	= true;
+	from.level	= entry->level + 1;
+	from.btree	= entry->btree_id;
+
 	if (journal_entry_err_on(!entry->u64s ||
 				 le16_to_cpu(entry->u64s) != k->k.u64s,
 				 c, version, jset, entry,
@@ -452,8 +457,7 @@ static int journal_entry_btree_root_validate(struct bch_fs *c,
 		return 0;
 	}
 
-	ret = journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
-				   version, big_endian, flags);
+	ret = journal_validate_key(c, jset, entry, k, from, version, big_endian);
 	if (ret == FSCK_DELETED_KEY)
 		ret = 0;
 fsck_err:
@@ -470,7 +474,7 @@ static int journal_entry_prio_ptrs_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	/* obsolete, don't care: */
 	return 0;
@@ -485,7 +489,7 @@ static int journal_entry_blacklist_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	int ret = 0;
 
@@ -512,7 +516,7 @@ static int journal_entry_blacklist_v2_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct jset_entry_blacklist_v2 *bl_entry;
 	int ret = 0;
@@ -554,7 +558,7 @@ static int journal_entry_usage_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct jset_entry_usage *u =
 		container_of(entry, struct jset_entry_usage, entry);
@@ -588,7 +592,7 @@ static int journal_entry_data_usage_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct jset_entry_data_usage *u =
 		container_of(entry, struct jset_entry_data_usage, entry);
@@ -632,7 +636,7 @@ static int journal_entry_clock_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct jset_entry_clock *clock =
 		container_of(entry, struct jset_entry_clock, entry);
@@ -665,14 +669,14 @@ static void journal_entry_clock_to_text(struct printbuf *out, struct bch_fs *c,
 	struct jset_entry_clock *clock =
 		container_of(entry, struct jset_entry_clock, entry);
 
-	prt_printf(out, "%s=%llu", clock->rw ? "write" : "read", le64_to_cpu(clock->time));
+	prt_printf(out, "%s=%llu", str_write_read(clock->rw), le64_to_cpu(clock->time));
 }
 
 static int journal_entry_dev_usage_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	struct jset_entry_dev_usage *u =
 		container_of(entry, struct jset_entry_dev_usage, entry);
@@ -729,7 +733,7 @@ static int journal_entry_log_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	return 0;
 }
@@ -738,19 +742,19 @@ static void journal_entry_log_to_text(struct printbuf *out, struct bch_fs *c,
 				      struct jset_entry *entry)
 {
 	struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
-	unsigned bytes = vstruct_bytes(entry) - offsetof(struct jset_entry_log, d);
 
-	prt_printf(out, "%.*s", bytes, l->d);
+	prt_printf(out, "%.*s", jset_entry_log_msg_bytes(l), l->d);
 }
 
 static int journal_entry_overwrite_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
+	from.flags = 0;
 	return journal_entry_btree_keys_validate(c, jset, entry,
-				version, big_endian, READ);
+				version, big_endian, from);
 }
 
 static void journal_entry_overwrite_to_text(struct printbuf *out, struct bch_fs *c,
@@ -763,10 +767,10 @@ static int journal_entry_write_buffer_keys_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	return journal_entry_btree_keys_validate(c, jset, entry,
-				version, big_endian, READ);
+				version, big_endian, from);
 }
 
 static void journal_entry_write_buffer_keys_to_text(struct printbuf *out, struct bch_fs *c,
@@ -779,7 +783,7 @@ static int journal_entry_datetime_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	unsigned bytes = vstruct_bytes(entry);
 	unsigned expected = 16;
@@ -809,7 +813,7 @@ static void journal_entry_datetime_to_text(struct printbuf *out, struct bch_fs *
 struct jset_entry_ops {
 	int (*validate)(struct bch_fs *, struct jset *,
 			struct jset_entry *, unsigned, int,
-			enum bch_validate_flags);
+			struct bkey_validate_context);
 	void (*to_text)(struct printbuf *, struct bch_fs *, struct jset_entry *);
 };
 
@@ -827,11 +831,11 @@ int bch2_journal_entry_validate(struct bch_fs *c,
 				struct jset *jset,
 				struct jset_entry *entry,
 				unsigned version, int big_endian,
-				enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	return entry->type < BCH_JSET_ENTRY_NR
 		? bch2_jset_entry_ops[entry->type].validate(c, jset, entry,
-				version, big_endian, flags)
+				version, big_endian, from)
 		: 0;
 }
 
@@ -849,10 +853,18 @@ void bch2_journal_entry_to_text(struct printbuf *out, struct bch_fs *c,
 static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 				 enum bch_validate_flags flags)
 {
+	struct bkey_validate_context from = {
+		.flags		= flags,
+		.from		= BKEY_VALIDATE_journal,
+		.journal_seq	= le64_to_cpu(jset->seq),
+	};
+
 	unsigned version = le32_to_cpu(jset->version);
 	int ret = 0;
 
 	vstruct_for_each(jset, entry) {
+		from.journal_offset = (u64 *) entry - jset->_data;
+
 		if (journal_entry_err_on(vstruct_next(entry) > vstruct_last(jset),
 				c, version, jset, entry,
 				journal_entry_past_jset_end,
@@ -861,8 +873,8 @@ static int jset_validate_entries(struct bch_fs *c, struct jset *jset,
 			break;
 		}
 
-		ret = bch2_journal_entry_validate(c, jset, entry,
-					version, JSET_BIG_ENDIAN(jset), flags);
+		ret = bch2_journal_entry_validate(c, jset, entry, version,
+						  JSET_BIG_ENDIAN(jset), from);
 		if (ret)
 			break;
 	}
@@ -875,13 +887,17 @@ static int jset_validate(struct bch_fs *c,
 			 struct jset *jset, u64 sector,
 			 enum bch_validate_flags flags)
 {
-	unsigned version;
+	struct bkey_validate_context from = {
+		.flags		= flags,
+		.from		= BKEY_VALIDATE_journal,
+		.journal_seq	= le64_to_cpu(jset->seq),
+	};
 	int ret = 0;
 
 	if (le64_to_cpu(jset->magic) != jset_magic(c))
 		return JOURNAL_ENTRY_NONE;
 
-	version = le32_to_cpu(jset->version);
+	unsigned version = le32_to_cpu(jset->version);
 	if (journal_entry_err_on(!bch2_version_compatible(version),
 			c, version, jset, NULL,
 			jset_unsupported_version,
@@ -926,15 +942,16 @@ static int jset_validate_early(struct bch_fs *c,
 			 unsigned bucket_sectors_left,
 			 unsigned sectors_read)
 {
-	size_t bytes = vstruct_bytes(jset);
-	unsigned version;
-	enum bch_validate_flags flags = BCH_VALIDATE_journal;
+	struct bkey_validate_context from = {
+		.from		= BKEY_VALIDATE_journal,
+		.journal_seq	= le64_to_cpu(jset->seq),
+	};
 	int ret = 0;
 
 	if (le64_to_cpu(jset->magic) != jset_magic(c))
 		return JOURNAL_ENTRY_NONE;
 
-	version = le32_to_cpu(jset->version);
+	unsigned version = le32_to_cpu(jset->version);
 	if (journal_entry_err_on(!bch2_version_compatible(version),
 			c, version, jset, NULL,
 			jset_unsupported_version,
@@ -947,6 +964,7 @@ static int jset_validate_early(struct bch_fs *c,
 		return -EINVAL;
 	}
 
+	size_t bytes = vstruct_bytes(jset);
 	if (bytes > (sectors_read << 9) &&
 	    sectors_read < bucket_sectors_left)
 		return JOURNAL_ENTRY_REREAD;
@@ -1231,8 +1249,6 @@ int bch2_journal_read(struct bch_fs *c,
 	 * those entries will be blacklisted:
 	 */
 	genradix_for_each_reverse(&c->journal_entries, radix_iter, _i) {
-		enum bch_validate_flags flags = BCH_VALIDATE_journal;
-
 		i = *_i;
 
 		if (journal_replay_ignore(i))
@@ -1252,6 +1268,10 @@ int bch2_journal_read(struct bch_fs *c,
 			continue;
 		}
 
+		struct bkey_validate_context from = {
+			.from		= BKEY_VALIDATE_journal,
+			.journal_seq	= le64_to_cpu(i->j.seq),
+		};
 		if (journal_entry_err_on(le64_to_cpu(i->j.last_seq) > le64_to_cpu(i->j.seq),
 					 c, le32_to_cpu(i->j.version), &i->j, NULL,
 					 jset_last_seq_newer_than_seq,
@@ -1411,27 +1431,50 @@ fsck_err:
 
 /* journal write: */
 
+static void journal_advance_devs_to_next_bucket(struct journal *j,
+						struct dev_alloc_list *devs,
+						unsigned sectors, u64 seq)
+{
+	struct bch_fs *c = container_of(j, struct bch_fs, journal);
+
+	darray_for_each(*devs, i) {
+		struct bch_dev *ca = rcu_dereference(c->devs[*i]);
+		if (!ca)
+			continue;
+
+		struct journal_device *ja = &ca->journal;
+
+		if (sectors > ja->sectors_free &&
+		    sectors <= ca->mi.bucket_size &&
+		    bch2_journal_dev_buckets_available(j, ja,
+					journal_space_discarded)) {
+			ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
+			ja->sectors_free = ca->mi.bucket_size;
+
+			/*
+			 * ja->bucket_seq[ja->cur_idx] must always have
+			 * something sensible:
+			 */
+			ja->bucket_seq[ja->cur_idx] = le64_to_cpu(seq);
+		}
+	}
+}
+
 static void __journal_write_alloc(struct journal *j,
 				  struct journal_buf *w,
-				  struct dev_alloc_list *devs_sorted,
+				  struct dev_alloc_list *devs,
 				  unsigned sectors,
 				  unsigned *replicas,
 				  unsigned replicas_want)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
-	struct journal_device *ja;
-	struct bch_dev *ca;
-	unsigned i;
 
-	if (*replicas >= replicas_want)
-		return;
-
-	for (i = 0; i < devs_sorted->nr; i++) {
-		ca = rcu_dereference(c->devs[devs_sorted->devs[i]]);
+	darray_for_each(*devs, i) {
+		struct bch_dev *ca = rcu_dereference(c->devs[*i]);
 		if (!ca)
 			continue;
 
-		ja = &ca->journal;
+		struct journal_device *ja = &ca->journal;
 
 		/*
 		 * Check that we can use this device, and aren't already using
@@ -1477,65 +1520,53 @@ static int journal_write_alloc(struct journal *j, struct journal_buf *w)
 {
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	struct bch_devs_mask devs;
-	struct journal_device *ja;
-	struct bch_dev *ca;
 	struct dev_alloc_list devs_sorted;
 	unsigned sectors = vstruct_sectors(w->data, c->block_bits);
 	unsigned target = c->opts.metadata_target ?:
 		c->opts.foreground_target;
-	unsigned i, replicas = 0, replicas_want =
+	unsigned replicas = 0, replicas_want =
 		READ_ONCE(c->opts.metadata_replicas);
 	unsigned replicas_need = min_t(unsigned, replicas_want,
 				       READ_ONCE(c->opts.metadata_replicas_required));
+	bool advance_done = false;
 
 	rcu_read_lock();
-retry:
-	devs = target_rw_devs(c, BCH_DATA_journal, target);
 
-	devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
+	/* We might run more than once if we have to stop and do discards: */
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(&w->key));
+	bkey_for_each_ptr(ptrs, p) {
+		struct bch_dev *ca = bch2_dev_rcu_noerror(c, p->dev);
+		if (ca)
+			replicas += ca->mi.durability;
+	}
 
-	__journal_write_alloc(j, w, &devs_sorted,
-			      sectors, &replicas, replicas_want);
+retry_target:
+	devs = target_rw_devs(c, BCH_DATA_journal, target);
+	devs_sorted = bch2_dev_alloc_list(c, &j->wp.stripe, &devs);
+retry_alloc:
+	__journal_write_alloc(j, w, &devs_sorted, sectors, &replicas, replicas_want);
 
-	if (replicas >= replicas_want)
+	if (likely(replicas >= replicas_want))
 		goto done;
 
-	for (i = 0; i < devs_sorted.nr; i++) {
-		ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
-		if (!ca)
-			continue;
-
-		ja = &ca->journal;
-
-		if (sectors > ja->sectors_free &&
-		    sectors <= ca->mi.bucket_size &&
-		    bch2_journal_dev_buckets_available(j, ja,
-					journal_space_discarded)) {
-			ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
-			ja->sectors_free = ca->mi.bucket_size;
-
-			/*
-			 * ja->bucket_seq[ja->cur_idx] must always have
-			 * something sensible:
-			 */
-			ja->bucket_seq[ja->cur_idx] = le64_to_cpu(w->data->seq);
-		}
+	if (!advance_done) {
+		journal_advance_devs_to_next_bucket(j, &devs_sorted, sectors, w->data->seq);
+		advance_done = true;
+		goto retry_alloc;
 	}
 
-	__journal_write_alloc(j, w, &devs_sorted,
-			      sectors, &replicas, replicas_want);
-
 	if (replicas < replicas_want && target) {
 		/* Retry from all devices: */
 		target = 0;
-		goto retry;
+		advance_done = false;
+		goto retry_target;
 	}
 done:
 	rcu_read_unlock();
 
 	BUG_ON(bkey_val_u64s(&w->key.k) > BCH_REPLICAS_MAX);
 
-	return replicas >= replicas_need ? 0 : -EROFS;
+	return replicas >= replicas_need ? 0 : -BCH_ERR_insufficient_journal_devices;
 }
 
 static void journal_buf_realloc(struct journal *j, struct journal_buf *buf)
@@ -2023,19 +2054,21 @@ CLOSURE_CALLBACK(bch2_journal_write)
 		bch2_journal_do_discards(j);
 	}
 
-	if (ret) {
+	if (ret && !bch2_journal_error(j)) {
 		struct printbuf buf = PRINTBUF;
 		buf.atomic++;
 
-		prt_printf(&buf, bch2_fmt(c, "Unable to allocate journal write at seq %llu: %s"),
+		prt_printf(&buf, bch2_fmt(c, "Unable to allocate journal write at seq %llu for %zu sectors: %s"),
 					  le64_to_cpu(w->data->seq),
+					  vstruct_sectors(w->data, c->block_bits),
 					  bch2_err_str(ret));
 		__bch2_journal_debug_to_text(&buf, j);
 		spin_unlock(&j->lock);
 		bch2_print_string_as_lines(KERN_ERR, buf.buf);
 		printbuf_exit(&buf);
-		goto err;
 	}
+	if (ret)
+		goto err;
 
 	/*
 	 * write is allocated, no longer need to account for it in
diff --git a/fs/bcachefs/journal_io.h b/fs/bcachefs/journal_io.h
index 2ca9cde30ea8..12b39fcb4424 100644
--- a/fs/bcachefs/journal_io.h
+++ b/fs/bcachefs/journal_io.h
@@ -63,7 +63,7 @@ static inline struct jset_entry *__jset_entry_type_next(struct jset *jset,
 
 int bch2_journal_entry_validate(struct bch_fs *, struct jset *,
 				struct jset_entry *, unsigned, int,
-				enum bch_validate_flags);
+				struct bkey_validate_context);
 void bch2_journal_entry_to_text(struct printbuf *, struct bch_fs *,
 				struct jset_entry *);
 
diff --git a/fs/bcachefs/journal_reclaim.c b/fs/bcachefs/journal_reclaim.c
index ace291f175dd..3c8242606da7 100644
--- a/fs/bcachefs/journal_reclaim.c
+++ b/fs/bcachefs/journal_reclaim.c
@@ -38,6 +38,9 @@ unsigned bch2_journal_dev_buckets_available(struct journal *j,
 					    struct journal_device *ja,
 					    enum journal_space_from from)
 {
+	if (!ja->nr)
+		return 0;
+
 	unsigned available = (journal_space_from(ja, from) -
 			      ja->cur_idx - 1 + ja->nr) % ja->nr;
 
@@ -137,14 +140,18 @@ static struct journal_space __journal_space_available(struct journal *j, unsigne
 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
 	unsigned pos, nr_devs = 0;
 	struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
+	unsigned min_bucket_size = U32_MAX;
 
 	BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
 
 	rcu_read_lock();
 	for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
-		if (!ca->journal.nr)
+		if (!ca->journal.nr ||
+		    !ca->mi.durability)
 			continue;
 
+		min_bucket_size = min(min_bucket_size, ca->mi.bucket_size);
+
 		space = journal_dev_space_available(j, ca, from);
 		if (!space.next_entry)
 			continue;
@@ -164,7 +171,9 @@ static struct journal_space __journal_space_available(struct journal *j, unsigne
 	 * We sorted largest to smallest, and we want the smallest out of the
 	 * @nr_devs_want largest devices:
 	 */
-	return dev_space[nr_devs_want - 1];
+	space = dev_space[nr_devs_want - 1];
+	space.next_entry = min(space.next_entry, min_bucket_size);
+	return space;
 }
 
 void bch2_journal_space_available(struct journal *j)
@@ -758,10 +767,12 @@ static int bch2_journal_reclaim_thread(void *arg)
 			journal_empty = fifo_empty(&j->pin);
 			spin_unlock(&j->lock);
 
+			long timeout = j->next_reclaim - jiffies;
+
 			if (journal_empty)
 				schedule();
-			else if (time_after(j->next_reclaim, jiffies))
-				schedule_timeout(j->next_reclaim - jiffies);
+			else if (timeout > 0)
+				schedule_timeout(timeout);
 			else
 				break;
 		}
diff --git a/fs/bcachefs/journal_types.h b/fs/bcachefs/journal_types.h
index 19183fcf7ad7..e9bd716fbb71 100644
--- a/fs/bcachefs/journal_types.h
+++ b/fs/bcachefs/journal_types.h
@@ -9,6 +9,9 @@
 #include "super_types.h"
 #include "fifo.h"
 
+/* btree write buffer steals 8 bits for its own purposes: */
+#define JOURNAL_SEQ_MAX		((1ULL << 56) - 1)
+
 #define JOURNAL_BUF_BITS	2
 #define JOURNAL_BUF_NR		(1U << JOURNAL_BUF_BITS)
 #define JOURNAL_BUF_MASK	(JOURNAL_BUF_NR - 1)
@@ -112,6 +115,7 @@ union journal_res_state {
  */
 #define JOURNAL_ENTRY_OFFSET_MAX	((1U << 20) - 1)
 
+#define JOURNAL_ENTRY_BLOCKED_VAL	(JOURNAL_ENTRY_OFFSET_MAX - 2)
 #define JOURNAL_ENTRY_CLOSED_VAL	(JOURNAL_ENTRY_OFFSET_MAX - 1)
 #define JOURNAL_ENTRY_ERROR_VAL		(JOURNAL_ENTRY_OFFSET_MAX)
 
@@ -193,6 +197,7 @@ struct journal {
 	 * insufficient devices:
 	 */
 	enum journal_errors	cur_entry_error;
+	unsigned		cur_entry_offset_if_blocked;
 
 	unsigned		buf_size_want;
 	/*
diff --git a/fs/bcachefs/logged_ops.c b/fs/bcachefs/logged_ops.c
index 60e00702d1a4..75f27ec26f85 100644
--- a/fs/bcachefs/logged_ops.c
+++ b/fs/bcachefs/logged_ops.c
@@ -63,8 +63,10 @@ fsck_err:
 int bch2_resume_logged_ops(struct bch_fs *c)
 {
 	int ret = bch2_trans_run(c,
-		for_each_btree_key(trans, iter,
-				   BTREE_ID_logged_ops, POS_MIN,
+		for_each_btree_key_max(trans, iter,
+				   BTREE_ID_logged_ops,
+				   POS(LOGGED_OPS_INUM_logged_ops, 0),
+				   POS(LOGGED_OPS_INUM_logged_ops, U64_MAX),
 				   BTREE_ITER_prefetch, k,
 			resume_logged_op(trans, &iter, k)));
 	bch_err_fn(c, ret);
@@ -74,9 +76,8 @@ int bch2_resume_logged_ops(struct bch_fs *c)
 static int __bch2_logged_op_start(struct btree_trans *trans, struct bkey_i *k)
 {
 	struct btree_iter iter;
-	int ret;
-
-	ret = bch2_bkey_get_empty_slot(trans, &iter, BTREE_ID_logged_ops, POS_MAX);
+	int ret = bch2_bkey_get_empty_slot(trans, &iter,
+				 BTREE_ID_logged_ops, POS(LOGGED_OPS_INUM_logged_ops, U64_MAX));
 	if (ret)
 		return ret;
 
diff --git a/fs/bcachefs/logged_ops_format.h b/fs/bcachefs/logged_ops_format.h
index 6a4bf7129dba..cfb67c95d4c8 100644
--- a/fs/bcachefs/logged_ops_format.h
+++ b/fs/bcachefs/logged_ops_format.h
@@ -2,6 +2,11 @@
 #ifndef _BCACHEFS_LOGGED_OPS_FORMAT_H
 #define _BCACHEFS_LOGGED_OPS_FORMAT_H
 
+enum logged_ops_inums {
+	LOGGED_OPS_INUM_logged_ops,
+	LOGGED_OPS_INUM_inode_cursors,
+};
+
 struct bch_logged_op_truncate {
 	struct bch_val		v;
 	__le32			subvol;
diff --git a/fs/bcachefs/lru.c b/fs/bcachefs/lru.c
index 10857eccdeaf..ce794d55818f 100644
--- a/fs/bcachefs/lru.c
+++ b/fs/bcachefs/lru.c
@@ -12,7 +12,7 @@
 
 /* KEY_TYPE_lru is obsolete: */
 int bch2_lru_validate(struct bch_fs *c, struct bkey_s_c k,
-		     enum bch_validate_flags flags)
+		      struct bkey_validate_context from)
 {
 	int ret = 0;
 
@@ -192,7 +192,7 @@ int bch2_check_lrus(struct bch_fs *c)
 	int ret = bch2_trans_run(c,
 		for_each_btree_key_commit(trans, iter,
 				BTREE_ID_lru, POS_MIN, BTREE_ITER_prefetch, k,
-				NULL, NULL, BCH_TRANS_COMMIT_no_enospc|BCH_TRANS_COMMIT_lazy_rw,
+				NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 			bch2_check_lru_key(trans, &iter, k, &last_flushed)));
 
 	bch2_bkey_buf_exit(&last_flushed, c);
diff --git a/fs/bcachefs/lru.h b/fs/bcachefs/lru.h
index e6a7d8241bb8..f31a6cf1514c 100644
--- a/fs/bcachefs/lru.h
+++ b/fs/bcachefs/lru.h
@@ -33,7 +33,7 @@ static inline enum bch_lru_type lru_type(struct bkey_s_c l)
 	return BCH_LRU_read;
 }
 
-int bch2_lru_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_lru_validate(struct bch_fs *, struct bkey_s_c, struct bkey_validate_context);
 void bch2_lru_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
 void bch2_lru_pos_to_text(struct printbuf *, struct bpos);
diff --git a/fs/bcachefs/move.c b/fs/bcachefs/move.c
index 0ef4a86850bb..c493ea625553 100644
--- a/fs/bcachefs/move.c
+++ b/fs/bcachefs/move.c
@@ -21,6 +21,8 @@
 #include "journal_reclaim.h"
 #include "keylist.h"
 #include "move.h"
+#include "rebalance.h"
+#include "reflink.h"
 #include "replicas.h"
 #include "snapshot.h"
 #include "super-io.h"
@@ -196,6 +198,13 @@ void bch2_moving_ctxt_exit(struct moving_context *ctxt)
 	list_del(&ctxt->list);
 	mutex_unlock(&c->moving_context_lock);
 
+	/*
+	 * Generally, releasing a transaction within a transaction restart means
+	 * an unhandled transaction restart: but this can happen legitimately
+	 * within the move code, e.g. when bch2_move_ratelimit() tells us to
+	 * exit before we've retried
+	 */
+	bch2_trans_begin(ctxt->trans);
 	bch2_trans_put(ctxt->trans);
 	memset(ctxt, 0, sizeof(*ctxt));
 }
@@ -379,34 +388,42 @@ err:
 	return ret;
 }
 
-struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
+static struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
 			  struct per_snapshot_io_opts *io_opts,
+			  struct bpos extent_pos, /* extent_iter, extent_k may be in reflink btree */
+			  struct btree_iter *extent_iter,
 			  struct bkey_s_c extent_k)
 {
 	struct bch_fs *c = trans->c;
 	u32 restart_count = trans->restart_count;
+	struct bch_io_opts *opts_ret = &io_opts->fs_io_opts;
 	int ret = 0;
 
-	if (io_opts->cur_inum != extent_k.k->p.inode) {
+	if (extent_k.k->type == KEY_TYPE_reflink_v)
+		goto out;
+
+	if (io_opts->cur_inum != extent_pos.inode) {
 		io_opts->d.nr = 0;
 
-		ret = for_each_btree_key(trans, iter, BTREE_ID_inodes, POS(0, extent_k.k->p.inode),
+		ret = for_each_btree_key(trans, iter, BTREE_ID_inodes, POS(0, extent_pos.inode),
 					 BTREE_ITER_all_snapshots, k, ({
-			if (k.k->p.offset != extent_k.k->p.inode)
+			if (k.k->p.offset != extent_pos.inode)
 				break;
 
 			if (!bkey_is_inode(k.k))
 				continue;
 
 			struct bch_inode_unpacked inode;
-			BUG_ON(bch2_inode_unpack(k, &inode));
+			_ret3 = bch2_inode_unpack(k, &inode);
+			if (_ret3)
+				break;
 
 			struct snapshot_io_opts_entry e = { .snapshot = k.k->p.snapshot };
 			bch2_inode_opts_get(&e.io_opts, trans->c, &inode);
 
 			darray_push(&io_opts->d, e);
 		}));
-		io_opts->cur_inum = extent_k.k->p.inode;
+		io_opts->cur_inum = extent_pos.inode;
 	}
 
 	ret = ret ?: trans_was_restarted(trans, restart_count);
@@ -415,43 +432,46 @@ struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
 
 	if (extent_k.k->p.snapshot)
 		darray_for_each(io_opts->d, i)
-			if (bch2_snapshot_is_ancestor(c, extent_k.k->p.snapshot, i->snapshot))
-				return &i->io_opts;
-
-	return &io_opts->fs_io_opts;
+			if (bch2_snapshot_is_ancestor(c, extent_k.k->p.snapshot, i->snapshot)) {
+				opts_ret = &i->io_opts;
+				break;
+			}
+out:
+	ret = bch2_get_update_rebalance_opts(trans, opts_ret, extent_iter, extent_k);
+	if (ret)
+		return ERR_PTR(ret);
+	return opts_ret;
 }
 
 int bch2_move_get_io_opts_one(struct btree_trans *trans,
 			      struct bch_io_opts *io_opts,
+			      struct btree_iter *extent_iter,
 			      struct bkey_s_c extent_k)
 {
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	int ret;
+	struct bch_fs *c = trans->c;
+
+	*io_opts = bch2_opts_to_inode_opts(c->opts);
 
 	/* reflink btree? */
-	if (!extent_k.k->p.inode) {
-		*io_opts = bch2_opts_to_inode_opts(trans->c->opts);
-		return 0;
-	}
+	if (!extent_k.k->p.inode)
+		goto out;
 
-	k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
+	struct btree_iter inode_iter;
+	struct bkey_s_c inode_k = bch2_bkey_get_iter(trans, &inode_iter, BTREE_ID_inodes,
 			       SPOS(0, extent_k.k->p.inode, extent_k.k->p.snapshot),
 			       BTREE_ITER_cached);
-	ret = bkey_err(k);
+	int ret = bkey_err(inode_k);
 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 		return ret;
 
-	if (!ret && bkey_is_inode(k.k)) {
+	if (!ret && bkey_is_inode(inode_k.k)) {
 		struct bch_inode_unpacked inode;
-		bch2_inode_unpack(k, &inode);
-		bch2_inode_opts_get(io_opts, trans->c, &inode);
-	} else {
-		*io_opts = bch2_opts_to_inode_opts(trans->c->opts);
+		bch2_inode_unpack(inode_k, &inode);
+		bch2_inode_opts_get(io_opts, c, &inode);
 	}
-
-	bch2_trans_iter_exit(trans, &iter);
-	return 0;
+	bch2_trans_iter_exit(trans, &inode_iter);
+out:
+	return bch2_get_update_rebalance_opts(trans, io_opts, extent_iter, extent_k);
 }
 
 int bch2_move_ratelimit(struct moving_context *ctxt)
@@ -509,9 +529,15 @@ static int bch2_move_data_btree(struct moving_context *ctxt,
 	struct per_snapshot_io_opts snapshot_io_opts;
 	struct bch_io_opts *io_opts;
 	struct bkey_buf sk;
-	struct btree_iter iter;
+	struct btree_iter iter, reflink_iter = {};
 	struct bkey_s_c k;
 	struct data_update_opts data_opts;
+	/*
+	 * If we're moving a single file, also process reflinked data it points
+	 * to (this includes propagating changed io_opts from the inode to the
+	 * extent):
+	 */
+	bool walk_indirect = start.inode == end.inode;
 	int ret = 0, ret2;
 
 	per_snapshot_io_opts_init(&snapshot_io_opts, c);
@@ -531,6 +557,8 @@ static int bch2_move_data_btree(struct moving_context *ctxt,
 		bch2_ratelimit_reset(ctxt->rate);
 
 	while (!bch2_move_ratelimit(ctxt)) {
+		struct btree_iter *extent_iter = &iter;
+
 		bch2_trans_begin(trans);
 
 		k = bch2_btree_iter_peek(&iter);
@@ -549,10 +577,36 @@ static int bch2_move_data_btree(struct moving_context *ctxt,
 		if (ctxt->stats)
 			ctxt->stats->pos = BBPOS(iter.btree_id, iter.pos);
 
+		if (walk_indirect &&
+		    k.k->type == KEY_TYPE_reflink_p &&
+		    REFLINK_P_MAY_UPDATE_OPTIONS(bkey_s_c_to_reflink_p(k).v)) {
+			struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
+			s64 offset_into_extent	= iter.pos.offset - bkey_start_offset(k.k);
+
+			bch2_trans_iter_exit(trans, &reflink_iter);
+			k = bch2_lookup_indirect_extent(trans, &reflink_iter, &offset_into_extent, p, true, 0);
+			ret = bkey_err(k);
+			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+				continue;
+			if (ret)
+				break;
+
+			if (bkey_deleted(k.k))
+				goto next_nondata;
+
+			/*
+			 * XXX: reflink pointers may point to multiple indirect
+			 * extents, so don't advance past the entire reflink
+			 * pointer - need to fixup iter->k
+			 */
+			extent_iter = &reflink_iter;
+		}
+
 		if (!bkey_extent_is_direct_data(k.k))
 			goto next_nondata;
 
-		io_opts = bch2_move_get_io_opts(trans, &snapshot_io_opts, k);
+		io_opts = bch2_move_get_io_opts(trans, &snapshot_io_opts,
+						iter.pos, extent_iter, k);
 		ret = PTR_ERR_OR_ZERO(io_opts);
 		if (ret)
 			continue;
@@ -568,7 +622,7 @@ static int bch2_move_data_btree(struct moving_context *ctxt,
 		bch2_bkey_buf_reassemble(&sk, c, k);
 		k = bkey_i_to_s_c(sk.k);
 
-		ret2 = bch2_move_extent(ctxt, NULL, &iter, k, *io_opts, data_opts);
+		ret2 = bch2_move_extent(ctxt, NULL, extent_iter, k, *io_opts, data_opts);
 		if (ret2) {
 			if (bch2_err_matches(ret2, BCH_ERR_transaction_restart))
 				continue;
@@ -589,6 +643,7 @@ next_nondata:
 		bch2_btree_iter_advance(&iter);
 	}
 
+	bch2_trans_iter_exit(trans, &reflink_iter);
 	bch2_trans_iter_exit(trans, &iter);
 	bch2_bkey_buf_exit(&sk, c);
 	per_snapshot_io_opts_exit(&snapshot_io_opts);
@@ -654,16 +709,12 @@ int bch2_evacuate_bucket(struct moving_context *ctxt,
 	struct bch_fs *c = trans->c;
 	bool is_kthread = current->flags & PF_KTHREAD;
 	struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
-	struct btree_iter iter;
+	struct btree_iter iter = {}, bp_iter = {};
 	struct bkey_buf sk;
-	struct bch_backpointer bp;
-	struct bch_alloc_v4 a_convert;
-	const struct bch_alloc_v4 *a;
 	struct bkey_s_c k;
 	struct data_update_opts data_opts;
-	unsigned dirty_sectors, bucket_size;
-	u64 fragmentation;
-	struct bpos bp_pos = POS_MIN;
+	unsigned sectors_moved = 0;
+	struct bkey_buf last_flushed;
 	int ret = 0;
 
 	struct bch_dev *ca = bch2_dev_tryget(c, bucket.inode);
@@ -672,6 +723,8 @@ int bch2_evacuate_bucket(struct moving_context *ctxt,
 
 	trace_bucket_evacuate(c, &bucket);
 
+	bch2_bkey_buf_init(&last_flushed);
+	bkey_init(&last_flushed.k->k);
 	bch2_bkey_buf_init(&sk);
 
 	/*
@@ -679,21 +732,13 @@ int bch2_evacuate_bucket(struct moving_context *ctxt,
 	 */
 	bch2_trans_begin(trans);
 
-	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
-			     bucket, BTREE_ITER_cached);
-	ret = lockrestart_do(trans,
-			bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
-	bch2_trans_iter_exit(trans, &iter);
+	bch2_trans_iter_init(trans, &bp_iter, BTREE_ID_backpointers,
+			     bucket_pos_to_bp_start(ca, bucket), 0);
 
 	bch_err_msg(c, ret, "looking up alloc key");
 	if (ret)
 		goto err;
 
-	a = bch2_alloc_to_v4(k, &a_convert);
-	dirty_sectors = bch2_bucket_sectors_dirty(*a);
-	bucket_size = ca->mi.bucket_size;
-	fragmentation = alloc_lru_idx_fragmentation(*a, ca);
-
 	ret = bch2_btree_write_buffer_tryflush(trans);
 	bch_err_msg(c, ret, "flushing btree write buffer");
 	if (ret)
@@ -705,18 +750,23 @@ int bch2_evacuate_bucket(struct moving_context *ctxt,
 
 		bch2_trans_begin(trans);
 
-		ret = bch2_get_next_backpointer(trans, ca, bucket, gen,
-						&bp_pos, &bp,
-						BTREE_ITER_cached);
+		k = bch2_btree_iter_peek(&bp_iter);
+		ret = bkey_err(k);
 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 			continue;
 		if (ret)
 			goto err;
-		if (bkey_eq(bp_pos, POS_MAX))
+
+		if (!k.k || bkey_gt(k.k->p, bucket_pos_to_bp_end(ca, bucket)))
 			break;
 
-		if (!bp.level) {
-			k = bch2_backpointer_get_key(trans, &iter, bp_pos, bp, 0);
+		if (k.k->type != KEY_TYPE_backpointer)
+			goto next;
+
+		struct bkey_s_c_backpointer bp = bkey_s_c_to_backpointer(k);
+
+		if (!bp.v->level) {
+			k = bch2_backpointer_get_key(trans, bp, &iter, 0, &last_flushed);
 			ret = bkey_err(k);
 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 				continue;
@@ -728,7 +778,7 @@ int bch2_evacuate_bucket(struct moving_context *ctxt,
 			bch2_bkey_buf_reassemble(&sk, c, k);
 			k = bkey_i_to_s_c(sk.k);
 
-			ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
+			ret = bch2_move_get_io_opts_one(trans, &io_opts, &iter, k);
 			if (ret) {
 				bch2_trans_iter_exit(trans, &iter);
 				continue;
@@ -738,14 +788,18 @@ int bch2_evacuate_bucket(struct moving_context *ctxt,
 			data_opts.target	= io_opts.background_target;
 			data_opts.rewrite_ptrs = 0;
 
+			unsigned sectors = bp.v->bucket_len; /* move_extent will drop locks */
 			unsigned i = 0;
-			bkey_for_each_ptr(bch2_bkey_ptrs_c(k), ptr) {
-				if (ptr->dev == bucket.inode) {
-					data_opts.rewrite_ptrs |= 1U << i;
-					if (ptr->cached) {
+			const union bch_extent_entry *entry;
+			struct extent_ptr_decoded p;
+			bkey_for_each_ptr_decode(k.k, bch2_bkey_ptrs_c(k), p, entry) {
+				if (p.ptr.dev == bucket.inode) {
+					if (p.ptr.cached) {
 						bch2_trans_iter_exit(trans, &iter);
 						goto next;
 					}
+					data_opts.rewrite_ptrs |= 1U << i;
+					break;
 				}
 				i++;
 			}
@@ -765,14 +819,15 @@ int bch2_evacuate_bucket(struct moving_context *ctxt,
 				goto err;
 
 			if (ctxt->stats)
-				atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
+				atomic64_add(sectors, &ctxt->stats->sectors_seen);
+			sectors_moved += sectors;
 		} else {
 			struct btree *b;
 
-			b = bch2_backpointer_get_node(trans, &iter, bp_pos, bp);
+			b = bch2_backpointer_get_node(trans, bp, &iter, &last_flushed);
 			ret = PTR_ERR_OR_ZERO(b);
 			if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
-				continue;
+				goto next;
 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 				continue;
 			if (ret)
@@ -796,15 +851,18 @@ int bch2_evacuate_bucket(struct moving_context *ctxt,
 				atomic64_add(sectors, &ctxt->stats->sectors_seen);
 				atomic64_add(sectors, &ctxt->stats->sectors_moved);
 			}
+			sectors_moved += btree_sectors(c);
 		}
 next:
-		bp_pos = bpos_nosnap_successor(bp_pos);
+		bch2_btree_iter_advance(&bp_iter);
 	}
 
-	trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
+	trace_evacuate_bucket(c, &bucket, sectors_moved, ca->mi.bucket_size, ret);
 err:
+	bch2_trans_iter_exit(trans, &bp_iter);
 	bch2_dev_put(ca);
 	bch2_bkey_buf_exit(&sk, c);
+	bch2_bkey_buf_exit(&last_flushed, c);
 	return ret;
 }
 
diff --git a/fs/bcachefs/move.h b/fs/bcachefs/move.h
index 9baf3093a678..51e0505a8156 100644
--- a/fs/bcachefs/move.h
+++ b/fs/bcachefs/move.h
@@ -110,9 +110,8 @@ static inline void per_snapshot_io_opts_exit(struct per_snapshot_io_opts *io_opt
 	darray_exit(&io_opts->d);
 }
 
-struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *,
-				struct per_snapshot_io_opts *, struct bkey_s_c);
-int bch2_move_get_io_opts_one(struct btree_trans *, struct bch_io_opts *, struct bkey_s_c);
+int bch2_move_get_io_opts_one(struct btree_trans *, struct bch_io_opts *,
+			      struct btree_iter *, struct bkey_s_c);
 
 int bch2_scan_old_btree_nodes(struct bch_fs *, struct bch_move_stats *);
 
diff --git a/fs/bcachefs/movinggc.c b/fs/bcachefs/movinggc.c
index d658be90f737..85c361e78ba5 100644
--- a/fs/bcachefs/movinggc.c
+++ b/fs/bcachefs/movinggc.c
@@ -167,7 +167,7 @@ static int bch2_copygc_get_buckets(struct moving_context *ctxt,
 
 	bch2_trans_begin(trans);
 
-	ret = for_each_btree_key_upto(trans, iter, BTREE_ID_lru,
+	ret = for_each_btree_key_max(trans, iter, BTREE_ID_lru,
 				  lru_pos(BCH_LRU_FRAGMENTATION_START, 0, 0),
 				  lru_pos(BCH_LRU_FRAGMENTATION_START, U64_MAX, LRU_TIME_MAX),
 				  0, k, ({
@@ -350,9 +350,9 @@ static int bch2_copygc_thread(void *arg)
 		bch2_trans_unlock_long(ctxt.trans);
 		cond_resched();
 
-		if (!c->copy_gc_enabled) {
+		if (!c->opts.copygc_enabled) {
 			move_buckets_wait(&ctxt, buckets, true);
-			kthread_wait_freezable(c->copy_gc_enabled ||
+			kthread_wait_freezable(c->opts.copygc_enabled ||
 					       kthread_should_stop());
 		}
 
diff --git a/fs/bcachefs/opts.c b/fs/bcachefs/opts.c
index 0e2ee262fbd4..6772faf385a5 100644
--- a/fs/bcachefs/opts.c
+++ b/fs/bcachefs/opts.c
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 
 #include <linux/kernel.h>
+#include <linux/fs_parser.h>
 
 #include "bcachefs.h"
 #include "compress.h"
@@ -48,12 +49,12 @@ static const char * const __bch2_csum_types[] = {
 	NULL
 };
 
-const char * const bch2_csum_opts[] = {
+const char * const __bch2_csum_opts[] = {
 	BCH_CSUM_OPTS()
 	NULL
 };
 
-static const char * const __bch2_compression_types[] = {
+const char * const __bch2_compression_types[] = {
 	BCH_COMPRESSION_TYPES()
 	NULL
 };
@@ -113,6 +114,7 @@ void bch2_prt_##name(struct printbuf *out, type t)				\
 PRT_STR_OPT_BOUNDSCHECKED(jset_entry_type,	enum bch_jset_entry_type);
 PRT_STR_OPT_BOUNDSCHECKED(fs_usage_type,	enum bch_fs_usage_type);
 PRT_STR_OPT_BOUNDSCHECKED(data_type,		enum bch_data_type);
+PRT_STR_OPT_BOUNDSCHECKED(csum_opt,		enum bch_csum_opt);
 PRT_STR_OPT_BOUNDSCHECKED(csum_type,		enum bch_csum_type);
 PRT_STR_OPT_BOUNDSCHECKED(compression_type,	enum bch_compression_type);
 PRT_STR_OPT_BOUNDSCHECKED(str_hash_type,	enum bch_str_hash_type);
@@ -333,17 +335,18 @@ int bch2_opt_parse(struct bch_fs *c,
 	switch (opt->type) {
 	case BCH_OPT_BOOL:
 		if (val) {
-			ret = kstrtou64(val, 10, res);
+			ret = lookup_constant(bool_names, val, -BCH_ERR_option_not_bool);
+			if (ret != -BCH_ERR_option_not_bool) {
+				*res = ret;
+			} else {
+				if (err)
+					prt_printf(err, "%s: must be bool", opt->attr.name);
+				return ret;
+			}
 		} else {
-			ret = 0;
 			*res = 1;
 		}
 
-		if (ret < 0 || (*res != 0 && *res != 1)) {
-			if (err)
-				prt_printf(err, "%s: must be bool", opt->attr.name);
-			return ret < 0 ? ret : -BCH_ERR_option_not_bool;
-		}
 		break;
 	case BCH_OPT_UINT:
 		if (!val) {
@@ -710,11 +713,14 @@ void bch2_opt_set_sb(struct bch_fs *c, struct bch_dev *ca,
 
 struct bch_io_opts bch2_opts_to_inode_opts(struct bch_opts src)
 {
-	return (struct bch_io_opts) {
+	struct bch_io_opts opts = {
 #define x(_name, _bits)	._name = src._name,
 	BCH_INODE_OPTS()
 #undef x
 	};
+
+	bch2_io_opts_fixups(&opts);
+	return opts;
 }
 
 bool bch2_opt_is_inode_opt(enum bch_opt_id id)
diff --git a/fs/bcachefs/opts.h b/fs/bcachefs/opts.h
index 23dda014e331..e763d52e0f38 100644
--- a/fs/bcachefs/opts.h
+++ b/fs/bcachefs/opts.h
@@ -16,7 +16,8 @@ extern const char * const bch2_version_upgrade_opts[];
 extern const char * const bch2_sb_features[];
 extern const char * const bch2_sb_compat[];
 extern const char * const __bch2_btree_ids[];
-extern const char * const bch2_csum_opts[];
+extern const char * const __bch2_csum_opts[];
+extern const char * const __bch2_compression_types[];
 extern const char * const bch2_compression_opts[];
 extern const char * const __bch2_str_hash_types[];
 extern const char * const bch2_str_hash_opts[];
@@ -27,6 +28,7 @@ extern const char * const bch2_d_types[];
 void bch2_prt_jset_entry_type(struct printbuf *,	enum bch_jset_entry_type);
 void bch2_prt_fs_usage_type(struct printbuf *,		enum bch_fs_usage_type);
 void bch2_prt_data_type(struct printbuf *,		enum bch_data_type);
+void bch2_prt_csum_opt(struct printbuf *,		enum bch_csum_opt);
 void bch2_prt_csum_type(struct printbuf *,		enum bch_csum_type);
 void bch2_prt_compression_type(struct printbuf *,	enum bch_compression_type);
 void bch2_prt_str_hash_type(struct printbuf *,		enum bch_str_hash_type);
@@ -171,12 +173,12 @@ enum fsck_err_opts {
 	  "size",	"Maximum size of checksummed/compressed extents")\
 	x(metadata_checksum,		u8,				\
 	  OPT_FS|OPT_FORMAT|OPT_MOUNT|OPT_RUNTIME,			\
-	  OPT_STR(bch2_csum_opts),					\
+	  OPT_STR(__bch2_csum_opts),					\
 	  BCH_SB_META_CSUM_TYPE,	BCH_CSUM_OPT_crc32c,		\
 	  NULL,		NULL)						\
 	x(data_checksum,		u8,				\
 	  OPT_FS|OPT_INODE|OPT_FORMAT|OPT_MOUNT|OPT_RUNTIME,		\
-	  OPT_STR(bch2_csum_opts),					\
+	  OPT_STR(__bch2_csum_opts),					\
 	  BCH_SB_DATA_CSUM_TYPE,	BCH_CSUM_OPT_crc32c,		\
 	  NULL,		NULL)						\
 	x(compression,			u8,				\
@@ -220,14 +222,14 @@ enum fsck_err_opts {
 	  BCH_SB_ERASURE_CODE,		false,				\
 	  NULL,		"Enable erasure coding (DO NOT USE YET)")	\
 	x(inodes_32bit,			u8,				\
-	  OPT_FS|OPT_FORMAT|OPT_MOUNT|OPT_RUNTIME,			\
+	  OPT_FS|OPT_INODE|OPT_FORMAT|OPT_MOUNT|OPT_RUNTIME,		\
 	  OPT_BOOL(),							\
 	  BCH_SB_INODE_32BIT,		true,				\
 	  NULL,		"Constrain inode numbers to 32 bits")		\
-	x(shard_inode_numbers,		u8,				\
-	  OPT_FS|OPT_FORMAT|OPT_MOUNT|OPT_RUNTIME,			\
-	  OPT_BOOL(),							\
-	  BCH_SB_SHARD_INUMS,		true,				\
+	x(shard_inode_numbers_bits,	u8,				\
+	  OPT_FS|OPT_FORMAT,						\
+	  OPT_UINT(0, 8),						\
+	  BCH_SB_SHARD_INUMS_NBITS,	0,				\
 	  NULL,		"Shard new inode numbers by CPU id")		\
 	x(inodes_use_key_cache,	u8,					\
 	  OPT_FS|OPT_FORMAT|OPT_MOUNT,					\
@@ -473,6 +475,18 @@ enum fsck_err_opts {
 	  BCH2_NO_SB_OPT,			true,			\
 	  NULL,		"Enable nocow mode: enables runtime locking in\n"\
 			"data move path needed if nocow will ever be in use\n")\
+	x(copygc_enabled,		u8,				\
+	  OPT_FS|OPT_MOUNT,						\
+	  OPT_BOOL(),							\
+	  BCH2_NO_SB_OPT,			true,			\
+	  NULL,		"Enable copygc: disable for debugging, or to\n"\
+			"quiet the system when doing performance testing\n")\
+	x(rebalance_enabled,		u8,				\
+	  OPT_FS|OPT_MOUNT,						\
+	  OPT_BOOL(),							\
+	  BCH2_NO_SB_OPT,			true,			\
+	  NULL,		"Enable rebalance: disable for debugging, or to\n"\
+			"quiet the system when doing performance testing\n")\
 	x(no_data_io,			u8,				\
 	  OPT_MOUNT,							\
 	  OPT_BOOL(),							\
@@ -488,7 +502,7 @@ enum fsck_err_opts {
 	  OPT_DEVICE,							\
 	  OPT_UINT(0, S64_MAX),						\
 	  BCH2_NO_SB_OPT,		0,				\
-	  "size",	"Size of filesystem on device")			\
+	  "size",	"Specifies the bucket size; must be greater than the btree node size")\
 	x(durability,			u8,				\
 	  OPT_DEVICE|OPT_SB_FIELD_ONE_BIAS,				\
 	  OPT_UINT(0, BCH_REPLICAS_MAX),				\
@@ -624,14 +638,39 @@ struct bch_io_opts {
 #define x(_name, _bits)	u##_bits _name;
 	BCH_INODE_OPTS()
 #undef x
+#define x(_name, _bits)	u64 _name##_from_inode:1;
+	BCH_INODE_OPTS()
+#undef x
 };
 
-static inline unsigned background_compression(struct bch_io_opts opts)
+static inline void bch2_io_opts_fixups(struct bch_io_opts *opts)
 {
-	return opts.background_compression ?: opts.compression;
+	if (!opts->background_target)
+		opts->background_target = opts->foreground_target;
+	if (!opts->background_compression)
+		opts->background_compression = opts->compression;
+	if (opts->nocow) {
+		opts->compression = opts->background_compression = 0;
+		opts->data_checksum = 0;
+		opts->erasure_code = 0;
+	}
 }
 
 struct bch_io_opts bch2_opts_to_inode_opts(struct bch_opts);
 bool bch2_opt_is_inode_opt(enum bch_opt_id);
 
+/* rebalance opts: */
+
+static inline struct bch_extent_rebalance io_opts_to_rebalance_opts(struct bch_io_opts *opts)
+{
+	return (struct bch_extent_rebalance) {
+		.type = BIT(BCH_EXTENT_ENTRY_rebalance),
+#define x(_name)							\
+		._name = opts->_name,					\
+		._name##_from_inode = opts->_name##_from_inode,
+		BCH_REBALANCE_OPTS()
+#undef x
+	};
+};
+
 #endif /* _BCACHEFS_OPTS_H */
diff --git a/fs/bcachefs/printbuf.h b/fs/bcachefs/printbuf.h
index 1d570387b77f..d0dd398baa2b 100644
--- a/fs/bcachefs/printbuf.h
+++ b/fs/bcachefs/printbuf.h
@@ -251,16 +251,23 @@ static inline void prt_hex_byte_upper(struct printbuf *out, u8 byte)
 	printbuf_nul_terminate_reserved(out);
 }
 
+static inline void printbuf_reset_keep_tabstops(struct printbuf *buf)
+{
+	buf->pos		= 0;
+	buf->allocation_failure	= 0;
+	buf->last_newline	= 0;
+	buf->last_field		= 0;
+	buf->indent		= 0;
+	buf->cur_tabstop	= 0;
+}
+
 /**
  * printbuf_reset - re-use a printbuf without freeing and re-initializing it:
  */
 static inline void printbuf_reset(struct printbuf *buf)
 {
-	buf->pos		= 0;
-	buf->allocation_failure	= 0;
-	buf->indent		= 0;
+	printbuf_reset_keep_tabstops(buf);
 	buf->nr_tabstops	= 0;
-	buf->cur_tabstop	= 0;
 }
 
 /**
diff --git a/fs/bcachefs/quota.c b/fs/bcachefs/quota.c
index 74f45a8162ad..8b857fc33244 100644
--- a/fs/bcachefs/quota.c
+++ b/fs/bcachefs/quota.c
@@ -60,7 +60,7 @@ const struct bch_sb_field_ops bch_sb_field_ops_quota = {
 };
 
 int bch2_quota_validate(struct bch_fs *c, struct bkey_s_c k,
-			enum bch_validate_flags flags)
+			struct bkey_validate_context from)
 {
 	int ret = 0;
 
diff --git a/fs/bcachefs/quota.h b/fs/bcachefs/quota.h
index a62abcc5332a..1551800ff44c 100644
--- a/fs/bcachefs/quota.h
+++ b/fs/bcachefs/quota.h
@@ -5,10 +5,10 @@
 #include "inode.h"
 #include "quota_types.h"
 
-enum bch_validate_flags;
 extern const struct bch_sb_field_ops bch_sb_field_ops_quota;
 
-int bch2_quota_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_quota_validate(struct bch_fs *, struct bkey_s_c,
+			struct bkey_validate_context);
 void bch2_quota_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
 #define bch2_bkey_ops_quota ((struct bkey_ops) {	\
diff --git a/fs/bcachefs/rcu_pending.c b/fs/bcachefs/rcu_pending.c
index 40a20192eee8..bef2aa1b8bcd 100644
--- a/fs/bcachefs/rcu_pending.c
+++ b/fs/bcachefs/rcu_pending.c
@@ -25,21 +25,37 @@ enum rcu_pending_special {
 #define RCU_PENDING_KVFREE_FN		((rcu_pending_process_fn) (ulong) RCU_PENDING_KVFREE)
 #define RCU_PENDING_CALL_RCU_FN		((rcu_pending_process_fn) (ulong) RCU_PENDING_CALL_RCU)
 
-static inline unsigned long __get_state_synchronize_rcu(struct srcu_struct *ssp)
+#ifdef __KERNEL__
+typedef unsigned long			rcu_gp_poll_state_t;
+
+static inline bool rcu_gp_poll_cookie_eq(rcu_gp_poll_state_t l, rcu_gp_poll_state_t r)
+{
+	return l == r;
+}
+#else
+typedef struct urcu_gp_poll_state	rcu_gp_poll_state_t;
+
+static inline bool rcu_gp_poll_cookie_eq(rcu_gp_poll_state_t l, rcu_gp_poll_state_t r)
+{
+	return l.grace_period_id == r.grace_period_id;
+}
+#endif
+
+static inline rcu_gp_poll_state_t __get_state_synchronize_rcu(struct srcu_struct *ssp)
 {
 	return ssp
 		? get_state_synchronize_srcu(ssp)
 		: get_state_synchronize_rcu();
 }
 
-static inline unsigned long __start_poll_synchronize_rcu(struct srcu_struct *ssp)
+static inline rcu_gp_poll_state_t __start_poll_synchronize_rcu(struct srcu_struct *ssp)
 {
 	return ssp
 		? start_poll_synchronize_srcu(ssp)
 		: start_poll_synchronize_rcu();
 }
 
-static inline bool __poll_state_synchronize_rcu(struct srcu_struct *ssp, unsigned long cookie)
+static inline bool __poll_state_synchronize_rcu(struct srcu_struct *ssp, rcu_gp_poll_state_t cookie)
 {
 	return ssp
 		? poll_state_synchronize_srcu(ssp, cookie)
@@ -71,13 +87,13 @@ struct rcu_pending_seq {
 	GENRADIX(struct rcu_head *)	objs;
 	size_t				nr;
 	struct rcu_head			**cursor;
-	unsigned long			seq;
+	rcu_gp_poll_state_t		seq;
 };
 
 struct rcu_pending_list {
 	struct rcu_head			*head;
 	struct rcu_head			*tail;
-	unsigned long			seq;
+	rcu_gp_poll_state_t		seq;
 };
 
 struct rcu_pending_pcpu {
@@ -316,10 +332,10 @@ static void rcu_pending_rcu_cb(struct rcu_head *rcu)
 }
 
 static __always_inline struct rcu_pending_seq *
-get_object_radix(struct rcu_pending_pcpu *p, unsigned long seq)
+get_object_radix(struct rcu_pending_pcpu *p, rcu_gp_poll_state_t seq)
 {
 	darray_for_each_reverse(p->objs, objs)
-		if (objs->seq == seq)
+		if (rcu_gp_poll_cookie_eq(objs->seq, seq))
 			return objs;
 
 	if (darray_push_gfp(&p->objs, ((struct rcu_pending_seq) { .seq = seq }), GFP_ATOMIC))
@@ -329,7 +345,7 @@ get_object_radix(struct rcu_pending_pcpu *p, unsigned long seq)
 }
 
 static noinline bool
-rcu_pending_enqueue_list(struct rcu_pending_pcpu *p, unsigned long seq,
+rcu_pending_enqueue_list(struct rcu_pending_pcpu *p, rcu_gp_poll_state_t seq,
 			 struct rcu_head *head, void *ptr,
 			 unsigned long *flags)
 {
@@ -364,7 +380,7 @@ rcu_pending_enqueue_list(struct rcu_pending_pcpu *p, unsigned long seq,
 again:
 	for (struct rcu_pending_list *i = p->lists;
 	     i < p->lists + NUM_ACTIVE_RCU_POLL_OLDSTATE; i++) {
-		if (i->seq == seq) {
+		if (rcu_gp_poll_cookie_eq(i->seq, seq)) {
 			rcu_pending_list_add(i, head);
 			return false;
 		}
@@ -408,7 +424,7 @@ __rcu_pending_enqueue(struct rcu_pending *pending, struct rcu_head *head,
 	struct rcu_pending_pcpu *p;
 	struct rcu_pending_seq *objs;
 	struct genradix_node *new_node = NULL;
-	unsigned long seq, flags;
+	unsigned long flags;
 	bool start_gp = false;
 
 	BUG_ON((ptr != NULL) != (pending->process == RCU_PENDING_KVFREE_FN));
@@ -416,7 +432,7 @@ __rcu_pending_enqueue(struct rcu_pending *pending, struct rcu_head *head,
 	local_irq_save(flags);
 	p = this_cpu_ptr(pending->p);
 	spin_lock(&p->lock);
-	seq = __get_state_synchronize_rcu(pending->srcu);
+	rcu_gp_poll_state_t seq = __get_state_synchronize_rcu(pending->srcu);
 restart:
 	if (may_sleep &&
 	    unlikely(process_finished_items(pending, p, flags)))
diff --git a/fs/bcachefs/rebalance.c b/fs/bcachefs/rebalance.c
index cd6647374353..4adc74cd3f70 100644
--- a/fs/bcachefs/rebalance.c
+++ b/fs/bcachefs/rebalance.c
@@ -24,6 +24,192 @@
 #include <linux/kthread.h>
 #include <linux/sched/cputime.h>
 
+/* bch_extent_rebalance: */
+
+static const struct bch_extent_rebalance *bch2_bkey_rebalance_opts(struct bkey_s_c k)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+
+	bkey_extent_entry_for_each(ptrs, entry)
+		if (__extent_entry_type(entry) == BCH_EXTENT_ENTRY_rebalance)
+			return &entry->rebalance;
+
+	return NULL;
+}
+
+static inline unsigned bch2_bkey_ptrs_need_compress(struct bch_fs *c,
+					   struct bch_io_opts *opts,
+					   struct bkey_s_c k,
+					   struct bkey_ptrs_c ptrs)
+{
+	if (!opts->background_compression)
+		return 0;
+
+	unsigned compression_type = bch2_compression_opt_to_type(opts->background_compression);
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
+	unsigned ptr_bit = 1;
+	unsigned rewrite_ptrs = 0;
+
+	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+		if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible ||
+		    p.ptr.unwritten)
+			return 0;
+
+		if (!p.ptr.cached && p.crc.compression_type != compression_type)
+			rewrite_ptrs |= ptr_bit;
+		ptr_bit <<= 1;
+	}
+
+	return rewrite_ptrs;
+}
+
+static inline unsigned bch2_bkey_ptrs_need_move(struct bch_fs *c,
+				       struct bch_io_opts *opts,
+				       struct bkey_ptrs_c ptrs)
+{
+	if (!opts->background_target ||
+	    !bch2_target_accepts_data(c, BCH_DATA_user, opts->background_target))
+		return 0;
+
+	unsigned ptr_bit = 1;
+	unsigned rewrite_ptrs = 0;
+
+	bkey_for_each_ptr(ptrs, ptr) {
+		if (!ptr->cached && !bch2_dev_in_target(c, ptr->dev, opts->background_target))
+			rewrite_ptrs |= ptr_bit;
+		ptr_bit <<= 1;
+	}
+
+	return rewrite_ptrs;
+}
+
+static unsigned bch2_bkey_ptrs_need_rebalance(struct bch_fs *c,
+					      struct bch_io_opts *opts,
+					      struct bkey_s_c k)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+
+	return bch2_bkey_ptrs_need_compress(c, opts, k, ptrs) |
+		bch2_bkey_ptrs_need_move(c, opts, ptrs);
+}
+
+u64 bch2_bkey_sectors_need_rebalance(struct bch_fs *c, struct bkey_s_c k)
+{
+	const struct bch_extent_rebalance *opts = bch2_bkey_rebalance_opts(k);
+	if (!opts)
+		return 0;
+
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
+	u64 sectors = 0;
+
+	if (opts->background_compression) {
+		unsigned compression_type = bch2_compression_opt_to_type(opts->background_compression);
+
+		bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
+			if (p.crc.compression_type == BCH_COMPRESSION_TYPE_incompressible ||
+			    p.ptr.unwritten) {
+				sectors = 0;
+				goto incompressible;
+			}
+
+			if (!p.ptr.cached && p.crc.compression_type != compression_type)
+				sectors += p.crc.compressed_size;
+		}
+	}
+incompressible:
+	if (opts->background_target &&
+	    bch2_target_accepts_data(c, BCH_DATA_user, opts->background_target)) {
+		bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
+			if (!p.ptr.cached && !bch2_dev_in_target(c, p.ptr.dev, opts->background_target))
+				sectors += p.crc.compressed_size;
+	}
+
+	return sectors;
+}
+
+static bool bch2_bkey_rebalance_needs_update(struct bch_fs *c, struct bch_io_opts *opts,
+					     struct bkey_s_c k)
+{
+	if (!bkey_extent_is_direct_data(k.k))
+		return 0;
+
+	const struct bch_extent_rebalance *old = bch2_bkey_rebalance_opts(k);
+
+	if (k.k->type == KEY_TYPE_reflink_v || bch2_bkey_ptrs_need_rebalance(c, opts, k)) {
+		struct bch_extent_rebalance new = io_opts_to_rebalance_opts(opts);
+		return old == NULL || memcmp(old, &new, sizeof(new));
+	} else {
+		return old != NULL;
+	}
+}
+
+int bch2_bkey_set_needs_rebalance(struct bch_fs *c, struct bch_io_opts *opts,
+				  struct bkey_i *_k)
+{
+	if (!bkey_extent_is_direct_data(&_k->k))
+		return 0;
+
+	struct bkey_s k = bkey_i_to_s(_k);
+	struct bch_extent_rebalance *old =
+		(struct bch_extent_rebalance *) bch2_bkey_rebalance_opts(k.s_c);
+
+	if (k.k->type == KEY_TYPE_reflink_v || bch2_bkey_ptrs_need_rebalance(c, opts, k.s_c)) {
+		if (!old) {
+			old = bkey_val_end(k);
+			k.k->u64s += sizeof(*old) / sizeof(u64);
+		}
+
+		*old = io_opts_to_rebalance_opts(opts);
+	} else {
+		if (old)
+			extent_entry_drop(k, (union bch_extent_entry *) old);
+	}
+
+	return 0;
+}
+
+int bch2_get_update_rebalance_opts(struct btree_trans *trans,
+				   struct bch_io_opts *io_opts,
+				   struct btree_iter *iter,
+				   struct bkey_s_c k)
+{
+	BUG_ON(iter->flags & BTREE_ITER_is_extents);
+	BUG_ON(iter->flags & BTREE_ITER_filter_snapshots);
+
+	const struct bch_extent_rebalance *r = k.k->type == KEY_TYPE_reflink_v
+		? bch2_bkey_rebalance_opts(k) : NULL;
+	if (r) {
+#define x(_name)							\
+		if (r->_name##_from_inode) {				\
+			io_opts->_name = r->_name;			\
+			io_opts->_name##_from_inode = true;		\
+		}
+		BCH_REBALANCE_OPTS()
+#undef x
+	}
+
+	if (!bch2_bkey_rebalance_needs_update(trans->c, io_opts, k))
+		return 0;
+
+	struct bkey_i *n = bch2_trans_kmalloc(trans, bkey_bytes(k.k) + 8);
+	int ret = PTR_ERR_OR_ZERO(n);
+	if (ret)
+		return ret;
+
+	bkey_reassemble(n, k);
+
+	/* On successfull transaction commit, @k was invalidated: */
+
+	return bch2_bkey_set_needs_rebalance(trans->c, io_opts, n) ?:
+		bch2_trans_update(trans, iter, n, BTREE_UPDATE_internal_snapshot_node) ?:
+		bch2_trans_commit(trans, NULL, NULL, 0) ?:
+		-BCH_ERR_transaction_restart_nested;
+}
+
 #define REBALANCE_WORK_SCAN_OFFSET	(U64_MAX - 1)
 
 static const char * const bch2_rebalance_state_strs[] = {
@@ -33,7 +219,7 @@ static const char * const bch2_rebalance_state_strs[] = {
 #undef x
 };
 
-static int __bch2_set_rebalance_needs_scan(struct btree_trans *trans, u64 inum)
+int bch2_set_rebalance_needs_scan_trans(struct btree_trans *trans, u64 inum)
 {
 	struct btree_iter iter;
 	struct bkey_s_c k;
@@ -71,9 +257,8 @@ err:
 int bch2_set_rebalance_needs_scan(struct bch_fs *c, u64 inum)
 {
 	int ret = bch2_trans_commit_do(c, NULL, NULL,
-				       BCH_TRANS_COMMIT_no_enospc|
-				       BCH_TRANS_COMMIT_lazy_rw,
-			    __bch2_set_rebalance_needs_scan(trans, inum));
+				       BCH_TRANS_COMMIT_no_enospc,
+			    bch2_set_rebalance_needs_scan_trans(trans, inum));
 	rebalance_wakeup(c);
 	return ret;
 }
@@ -121,6 +306,9 @@ static int bch2_bkey_clear_needs_rebalance(struct btree_trans *trans,
 					   struct btree_iter *iter,
 					   struct bkey_s_c k)
 {
+	if (!bch2_bkey_rebalance_opts(k))
+		return 0;
+
 	struct bkey_i *n = bch2_bkey_make_mut(trans, iter, &k, 0);
 	int ret = PTR_ERR_OR_ZERO(n);
 	if (ret)
@@ -134,31 +322,27 @@ static int bch2_bkey_clear_needs_rebalance(struct btree_trans *trans,
 static struct bkey_s_c next_rebalance_extent(struct btree_trans *trans,
 			struct bpos work_pos,
 			struct btree_iter *extent_iter,
+			struct bch_io_opts *io_opts,
 			struct data_update_opts *data_opts)
 {
 	struct bch_fs *c = trans->c;
-	struct bkey_s_c k;
 
 	bch2_trans_iter_exit(trans, extent_iter);
 	bch2_trans_iter_init(trans, extent_iter,
 			     work_pos.inode ? BTREE_ID_extents : BTREE_ID_reflink,
 			     work_pos,
 			     BTREE_ITER_all_snapshots);
-	k = bch2_btree_iter_peek_slot(extent_iter);
+	struct bkey_s_c k = bch2_btree_iter_peek_slot(extent_iter);
 	if (bkey_err(k))
 		return k;
 
-	const struct bch_extent_rebalance *r = k.k ? bch2_bkey_rebalance_opts(k) : NULL;
-	if (!r) {
-		/* raced due to btree write buffer, nothing to do */
-		return bkey_s_c_null;
-	}
+	int ret = bch2_move_get_io_opts_one(trans, io_opts, extent_iter, k);
+	if (ret)
+		return bkey_s_c_err(ret);
 
 	memset(data_opts, 0, sizeof(*data_opts));
-
-	data_opts->rewrite_ptrs		=
-		bch2_bkey_ptrs_need_rebalance(c, k, r->target, r->compression);
-	data_opts->target		= r->target;
+	data_opts->rewrite_ptrs		= bch2_bkey_ptrs_need_rebalance(c, io_opts, k);
+	data_opts->target		= io_opts->background_target;
 	data_opts->write_flags		|= BCH_WRITE_ONLY_SPECIFIED_DEVS;
 
 	if (!data_opts->rewrite_ptrs) {
@@ -178,12 +362,28 @@ static struct bkey_s_c next_rebalance_extent(struct btree_trans *trans,
 	if (trace_rebalance_extent_enabled()) {
 		struct printbuf buf = PRINTBUF;
 
-		prt_str(&buf, "target=");
-		bch2_target_to_text(&buf, c, r->target);
-		prt_str(&buf, " compression=");
-		bch2_compression_opt_to_text(&buf, r->compression);
-		prt_str(&buf, " ");
 		bch2_bkey_val_to_text(&buf, c, k);
+		prt_newline(&buf);
+
+		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+
+		unsigned p = bch2_bkey_ptrs_need_compress(c, io_opts, k, ptrs);
+		if (p) {
+			prt_str(&buf, "compression=");
+			bch2_compression_opt_to_text(&buf, io_opts->background_compression);
+			prt_str(&buf, " ");
+			bch2_prt_u64_base2(&buf, p);
+			prt_newline(&buf);
+		}
+
+		p = bch2_bkey_ptrs_need_move(c, io_opts, ptrs);
+		if (p) {
+			prt_str(&buf, "move=");
+			bch2_target_to_text(&buf, c, io_opts->background_target);
+			prt_str(&buf, " ");
+			bch2_prt_u64_base2(&buf, p);
+			prt_newline(&buf);
+		}
 
 		trace_rebalance_extent(c, buf.buf);
 		printbuf_exit(&buf);
@@ -212,14 +412,10 @@ static int do_rebalance_extent(struct moving_context *ctxt,
 	bch2_bkey_buf_init(&sk);
 
 	ret = bkey_err(k = next_rebalance_extent(trans, work_pos,
-						 extent_iter, &data_opts));
+				extent_iter, &io_opts, &data_opts));
 	if (ret || !k.k)
 		goto out;
 
-	ret = bch2_move_get_io_opts_one(trans, &io_opts, k);
-	if (ret)
-		goto out;
-
 	atomic64_add(k.k->size, &ctxt->stats->sectors_seen);
 
 	/*
@@ -253,20 +449,8 @@ static bool rebalance_pred(struct bch_fs *c, void *arg,
 			   struct bch_io_opts *io_opts,
 			   struct data_update_opts *data_opts)
 {
-	unsigned target, compression;
-
-	if (k.k->p.inode) {
-		target		= io_opts->background_target;
-		compression	= background_compression(*io_opts);
-	} else {
-		const struct bch_extent_rebalance *r = bch2_bkey_rebalance_opts(k);
-
-		target		= r ? r->target : io_opts->background_target;
-		compression	= r ? r->compression : background_compression(*io_opts);
-	}
-
-	data_opts->rewrite_ptrs		= bch2_bkey_ptrs_need_rebalance(c, k, target, compression);
-	data_opts->target		= target;
+	data_opts->rewrite_ptrs		= bch2_bkey_ptrs_need_rebalance(c, io_opts, k);
+	data_opts->target		= io_opts->background_target;
 	data_opts->write_flags		|= BCH_WRITE_ONLY_SPECIFIED_DEVS;
 	return data_opts->rewrite_ptrs != 0;
 }
@@ -338,9 +522,9 @@ static int do_rebalance(struct moving_context *ctxt)
 			     BTREE_ITER_all_snapshots);
 
 	while (!bch2_move_ratelimit(ctxt)) {
-		if (!r->enabled) {
+		if (!c->opts.rebalance_enabled) {
 			bch2_moving_ctxt_flush_all(ctxt);
-			kthread_wait_freezable(r->enabled ||
+			kthread_wait_freezable(c->opts.rebalance_enabled ||
 					       kthread_should_stop());
 		}
 
diff --git a/fs/bcachefs/rebalance.h b/fs/bcachefs/rebalance.h
index 28a52638f16c..0a0821ab895d 100644
--- a/fs/bcachefs/rebalance.h
+++ b/fs/bcachefs/rebalance.h
@@ -2,8 +2,18 @@
 #ifndef _BCACHEFS_REBALANCE_H
 #define _BCACHEFS_REBALANCE_H
 
+#include "compress.h"
+#include "disk_groups.h"
 #include "rebalance_types.h"
 
+u64 bch2_bkey_sectors_need_rebalance(struct bch_fs *, struct bkey_s_c);
+int bch2_bkey_set_needs_rebalance(struct bch_fs *, struct bch_io_opts *, struct bkey_i *);
+int bch2_get_update_rebalance_opts(struct btree_trans *,
+				   struct bch_io_opts *,
+				   struct btree_iter *,
+				   struct bkey_s_c);
+
+int bch2_set_rebalance_needs_scan_trans(struct btree_trans *, u64);
 int bch2_set_rebalance_needs_scan(struct bch_fs *, u64 inum);
 int bch2_set_fs_needs_rebalance(struct bch_fs *);
 
diff --git a/fs/bcachefs/rebalance_format.h b/fs/bcachefs/rebalance_format.h
new file mode 100644
index 000000000000..ff9a1342a22b
--- /dev/null
+++ b/fs/bcachefs/rebalance_format.h
@@ -0,0 +1,53 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BCACHEFS_REBALANCE_FORMAT_H
+#define _BCACHEFS_REBALANCE_FORMAT_H
+
+struct bch_extent_rebalance {
+#if defined(__LITTLE_ENDIAN_BITFIELD)
+	__u64			type:6,
+				unused:3,
+
+				promote_target_from_inode:1,
+				erasure_code_from_inode:1,
+				data_checksum_from_inode:1,
+				background_compression_from_inode:1,
+				data_replicas_from_inode:1,
+				background_target_from_inode:1,
+
+				promote_target:16,
+				erasure_code:1,
+				data_checksum:4,
+				data_replicas:4,
+				background_compression:8, /* enum bch_compression_opt */
+				background_target:16;
+#elif defined (__BIG_ENDIAN_BITFIELD)
+	__u64			background_target:16,
+				background_compression:8,
+				data_replicas:4,
+				data_checksum:4,
+				erasure_code:1,
+				promote_target:16,
+
+				background_target_from_inode:1,
+				data_replicas_from_inode:1,
+				background_compression_from_inode:1,
+				data_checksum_from_inode:1,
+				erasure_code_from_inode:1,
+				promote_target_from_inode:1,
+
+				unused:3,
+				type:6;
+#endif
+};
+
+/* subset of BCH_INODE_OPTS */
+#define BCH_REBALANCE_OPTS()			\
+	x(data_checksum)			\
+	x(background_compression)		\
+	x(data_replicas)			\
+	x(promote_target)			\
+	x(background_target)			\
+	x(erasure_code)
+
+#endif /* _BCACHEFS_REBALANCE_FORMAT_H */
+
diff --git a/fs/bcachefs/rebalance_types.h b/fs/bcachefs/rebalance_types.h
index 0fffb536c1d0..fe5098c17dfc 100644
--- a/fs/bcachefs/rebalance_types.h
+++ b/fs/bcachefs/rebalance_types.h
@@ -30,8 +30,6 @@ struct bch_fs_rebalance {
 	struct bbpos			scan_start;
 	struct bbpos			scan_end;
 	struct bch_move_stats		scan_stats;
-
-	unsigned			enabled:1;
 };
 
 #endif /* _BCACHEFS_REBALANCE_TYPES_H */
diff --git a/fs/bcachefs/recovery.c b/fs/bcachefs/recovery.c
index 3c7f941dde39..98825437381c 100644
--- a/fs/bcachefs/recovery.c
+++ b/fs/bcachefs/recovery.c
@@ -34,21 +34,83 @@
 
 #define QSTR(n) { { { .len = strlen(n) } }, .name = n }
 
-void bch2_btree_lost_data(struct bch_fs *c, enum btree_id btree)
+int bch2_btree_lost_data(struct bch_fs *c, enum btree_id btree)
 {
-	if (btree >= BTREE_ID_NR_MAX)
-		return;
-
 	u64 b = BIT_ULL(btree);
+	int ret = 0;
+
+	mutex_lock(&c->sb_lock);
+	struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
 
 	if (!(c->sb.btrees_lost_data & b)) {
-		bch_err(c, "flagging btree %s lost data", bch2_btree_id_str(btree));
+		struct printbuf buf = PRINTBUF;
+		bch2_btree_id_to_text(&buf, btree);
+		bch_err(c, "flagging btree %s lost data", buf.buf);
+		printbuf_exit(&buf);
+		ext->btrees_lost_data |= cpu_to_le64(b);
+	}
 
-		mutex_lock(&c->sb_lock);
-		bch2_sb_field_get(c->disk_sb.sb, ext)->btrees_lost_data |= cpu_to_le64(b);
-		bch2_write_super(c);
-		mutex_unlock(&c->sb_lock);
+	/* Once we have runtime self healing for topology errors we won't need this: */
+	ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_topology) ?: ret;
+
+	/* Btree node accounting will be off: */
+	__set_bit_le64(BCH_FSCK_ERR_accounting_mismatch, ext->errors_silent);
+	ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_allocations) ?: ret;
+
+#ifdef CONFIG_BCACHEFS_DEBUG
+	/*
+	 * These are much more minor, and don't need to be corrected right away,
+	 * but in debug mode we want the next fsck run to be clean:
+	 */
+	ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_lrus) ?: ret;
+	ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_backpointers_to_extents) ?: ret;
+#endif
+
+	switch (btree) {
+	case BTREE_ID_alloc:
+		ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_alloc_info) ?: ret;
+
+		__set_bit_le64(BCH_FSCK_ERR_alloc_key_data_type_wrong, ext->errors_silent);
+		__set_bit_le64(BCH_FSCK_ERR_alloc_key_gen_wrong, ext->errors_silent);
+		__set_bit_le64(BCH_FSCK_ERR_alloc_key_dirty_sectors_wrong, ext->errors_silent);
+		__set_bit_le64(BCH_FSCK_ERR_alloc_key_cached_sectors_wrong, ext->errors_silent);
+		__set_bit_le64(BCH_FSCK_ERR_alloc_key_stripe_wrong, ext->errors_silent);
+		__set_bit_le64(BCH_FSCK_ERR_alloc_key_stripe_redundancy_wrong, ext->errors_silent);
+		goto out;
+	case BTREE_ID_backpointers:
+		ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_btree_backpointers) ?: ret;
+		ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_extents_to_backpointers) ?: ret;
+		goto out;
+	case BTREE_ID_need_discard:
+		ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_alloc_info) ?: ret;
+		goto out;
+	case BTREE_ID_freespace:
+		ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_alloc_info) ?: ret;
+		goto out;
+	case BTREE_ID_bucket_gens:
+		ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_alloc_info) ?: ret;
+		goto out;
+	case BTREE_ID_lru:
+		ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_alloc_info) ?: ret;
+		goto out;
+	case BTREE_ID_accounting:
+		ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_check_allocations) ?: ret;
+		goto out;
+	default:
+		ret = bch2_run_explicit_recovery_pass_persistent_locked(c, BCH_RECOVERY_PASS_scan_for_btree_nodes) ?: ret;
+		goto out;
 	}
+out:
+	bch2_write_super(c);
+	mutex_unlock(&c->sb_lock);
+
+	return ret;
+}
+
+static void kill_btree(struct bch_fs *c, enum btree_id btree)
+{
+	bch2_btree_id_root(c, btree)->alive = false;
+	bch2_shoot_down_journal_keys(c, btree, 0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
 }
 
 /* for -o reconstruct_alloc: */
@@ -79,6 +141,8 @@ static void bch2_reconstruct_alloc(struct bch_fs *c)
 	__set_bit_le64(BCH_FSCK_ERR_fs_usage_persistent_reserved_wrong, ext->errors_silent);
 	__set_bit_le64(BCH_FSCK_ERR_fs_usage_replicas_wrong, ext->errors_silent);
 
+	__set_bit_le64(BCH_FSCK_ERR_alloc_key_to_missing_lru_entry, ext->errors_silent);
+
 	__set_bit_le64(BCH_FSCK_ERR_alloc_key_data_type_wrong, ext->errors_silent);
 	__set_bit_le64(BCH_FSCK_ERR_alloc_key_gen_wrong, ext->errors_silent);
 	__set_bit_le64(BCH_FSCK_ERR_alloc_key_dirty_sectors_wrong, ext->errors_silent);
@@ -99,16 +163,9 @@ static void bch2_reconstruct_alloc(struct bch_fs *c)
 	bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
-	bch2_shoot_down_journal_keys(c, BTREE_ID_alloc,
-				     0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
-	bch2_shoot_down_journal_keys(c, BTREE_ID_backpointers,
-				     0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
-	bch2_shoot_down_journal_keys(c, BTREE_ID_need_discard,
-				     0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
-	bch2_shoot_down_journal_keys(c, BTREE_ID_freespace,
-				     0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
-	bch2_shoot_down_journal_keys(c, BTREE_ID_bucket_gens,
-				     0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
+	for (unsigned i = 0; i < btree_id_nr_alive(c); i++)
+		if (btree_id_is_alloc(i))
+			kill_btree(c, i);
 }
 
 /*
@@ -354,10 +411,13 @@ int bch2_journal_replay(struct bch_fs *c)
 				 ? BCH_TRANS_COMMIT_no_journal_res|BCH_WATERMARK_reclaim
 				 : 0),
 			     bch2_journal_replay_key(trans, k));
-		bch_err_msg(c, ret, "while replaying key at btree %s level %u:",
-			    bch2_btree_id_str(k->btree_id), k->level);
-		if (ret)
+		if (ret) {
+			struct printbuf buf = PRINTBUF;
+			bch2_btree_id_level_to_text(&buf, k->btree_id, k->level);
+			bch_err_msg(c, ret, "while replaying key at %s:", buf.buf);
+			printbuf_exit(&buf);
 			goto err;
+		}
 
 		BUG_ON(k->btree_id != BTREE_ID_accounting && !k->overwritten);
 	}
@@ -403,7 +463,9 @@ static int journal_replay_entry_early(struct bch_fs *c,
 
 	switch (entry->type) {
 	case BCH_JSET_ENTRY_btree_root: {
-		struct btree_root *r;
+
+		if (unlikely(!entry->u64s))
+			return 0;
 
 		if (fsck_err_on(entry->btree_id >= BTREE_ID_NR_MAX,
 				c, invalid_btree_id,
@@ -417,15 +479,11 @@ static int journal_replay_entry_early(struct bch_fs *c,
 				return ret;
 		}
 
-		r = bch2_btree_id_root(c, entry->btree_id);
+		struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
 
-		if (entry->u64s) {
-			r->level = entry->level;
-			bkey_copy(&r->key, (struct bkey_i *) entry->start);
-			r->error = 0;
-		} else {
-			r->error = -BCH_ERR_btree_node_read_error;
-		}
+		r->level = entry->level;
+		bkey_copy(&r->key, (struct bkey_i *) entry->start);
+		r->error = 0;
 		r->alive = true;
 		break;
 	}
@@ -505,6 +563,7 @@ static int journal_replay_early(struct bch_fs *c,
 
 static int read_btree_roots(struct bch_fs *c)
 {
+	struct printbuf buf = PRINTBUF;
 	int ret = 0;
 
 	for (unsigned i = 0; i < btree_id_nr_alive(c); i++) {
@@ -513,33 +572,22 @@ static int read_btree_roots(struct bch_fs *c)
 		if (!r->alive)
 			continue;
 
-		if (btree_id_is_alloc(i) && c->opts.reconstruct_alloc)
-			continue;
+		printbuf_reset(&buf);
+		bch2_btree_id_level_to_text(&buf, i, r->level);
 
 		if (mustfix_fsck_err_on((ret = r->error),
 					c, btree_root_bkey_invalid,
 					"invalid btree root %s",
-					bch2_btree_id_str(i)) ||
+					buf.buf) ||
 		    mustfix_fsck_err_on((ret = r->error = bch2_btree_root_read(c, i, &r->key, r->level)),
 					c, btree_root_read_error,
-					"error reading btree root %s l=%u: %s",
-					bch2_btree_id_str(i), r->level, bch2_err_str(ret))) {
-			if (btree_id_is_alloc(i)) {
-				c->opts.recovery_passes |= BIT_ULL(BCH_RECOVERY_PASS_check_allocations);
-				c->opts.recovery_passes |= BIT_ULL(BCH_RECOVERY_PASS_check_alloc_info);
-				c->opts.recovery_passes |= BIT_ULL(BCH_RECOVERY_PASS_check_lrus);
-				c->opts.recovery_passes |= BIT_ULL(BCH_RECOVERY_PASS_check_extents_to_backpointers);
-				c->opts.recovery_passes |= BIT_ULL(BCH_RECOVERY_PASS_check_alloc_to_lru_refs);
-				c->sb.compat &= ~(1ULL << BCH_COMPAT_alloc_info);
+					"error reading btree root %s: %s",
+					buf.buf, bch2_err_str(ret))) {
+			if (btree_id_is_alloc(i))
 				r->error = 0;
-			} else if (!(c->opts.recovery_passes & BIT_ULL(BCH_RECOVERY_PASS_scan_for_btree_nodes))) {
-				bch_info(c, "will run btree node scan");
-				c->opts.recovery_passes |= BIT_ULL(BCH_RECOVERY_PASS_scan_for_btree_nodes);
-				c->opts.recovery_passes |= BIT_ULL(BCH_RECOVERY_PASS_check_topology);
-			}
 
-			ret = 0;
-			bch2_btree_lost_data(c, i);
+			ret = bch2_btree_lost_data(c, i);
+			BUG_ON(ret);
 		}
 	}
 
@@ -553,6 +601,7 @@ static int read_btree_roots(struct bch_fs *c)
 		}
 	}
 fsck_err:
+	printbuf_exit(&buf);
 	return ret;
 }
 
@@ -563,6 +612,7 @@ static bool check_version_upgrade(struct bch_fs *c)
 					 bch2_latest_compatible_version(c->sb.version));
 	unsigned old_version = c->sb.version_upgrade_complete ?: c->sb.version;
 	unsigned new_version = 0;
+	bool ret = false;
 
 	if (old_version < bcachefs_metadata_required_upgrade_below) {
 		if (c->opts.version_upgrade == BCH_VERSION_UPGRADE_incompatible ||
@@ -618,14 +668,32 @@ static bool check_version_upgrade(struct bch_fs *c)
 		}
 
 		bch_info(c, "%s", buf.buf);
+		printbuf_exit(&buf);
 
-		bch2_sb_upgrade(c, new_version);
+		ret = true;
+	}
 
+	if (new_version > c->sb.version_incompat &&
+	    c->opts.version_upgrade == BCH_VERSION_UPGRADE_incompatible) {
+		struct printbuf buf = PRINTBUF;
+
+		prt_str(&buf, "Now allowing incompatible features up to ");
+		bch2_version_to_text(&buf, new_version);
+		prt_str(&buf, ", previously allowed up to ");
+		bch2_version_to_text(&buf, c->sb.version_incompat_allowed);
+		prt_newline(&buf);
+
+		bch_info(c, "%s", buf.buf);
 		printbuf_exit(&buf);
-		return true;
+
+		ret = true;
 	}
 
-	return false;
+	if (ret)
+		bch2_sb_upgrade(c, new_version,
+				c->opts.version_upgrade == BCH_VERSION_UPGRADE_incompatible);
+
+	return ret;
 }
 
 int bch2_fs_recovery(struct bch_fs *c)
@@ -660,8 +728,13 @@ int bch2_fs_recovery(struct bch_fs *c)
 		goto err;
 	}
 
-	if (c->opts.norecovery)
-		c->opts.recovery_pass_last = BCH_RECOVERY_PASS_journal_replay - 1;
+	if (c->opts.norecovery) {
+		c->opts.recovery_pass_last = c->opts.recovery_pass_last
+			? min(c->opts.recovery_pass_last, BCH_RECOVERY_PASS_snapshots_read)
+			: BCH_RECOVERY_PASS_snapshots_read;
+		c->opts.nochanges = true;
+		c->opts.read_only = true;
+	}
 
 	mutex_lock(&c->sb_lock);
 	struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
@@ -708,17 +781,20 @@ int bch2_fs_recovery(struct bch_fs *c)
 
 	c->opts.recovery_passes |= bch2_recovery_passes_from_stable(le64_to_cpu(ext->recovery_passes_required[0]));
 
+	if (c->sb.version_upgrade_complete < bcachefs_metadata_version_autofix_errors) {
+		SET_BCH_SB_ERROR_ACTION(c->disk_sb.sb, BCH_ON_ERROR_fix_safe);
+		write_sb = true;
+	}
+
 	if (write_sb)
 		bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
-	if (c->opts.fsck && IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
-		c->opts.recovery_passes |= BIT_ULL(BCH_RECOVERY_PASS_check_topology);
-
 	if (c->opts.fsck)
 		set_bit(BCH_FS_fsck_running, &c->flags);
 	if (c->sb.clean)
 		set_bit(BCH_FS_clean_recovery, &c->flags);
+	set_bit(BCH_FS_recovery_running, &c->flags);
 
 	ret = bch2_blacklist_table_initialize(c);
 	if (ret) {
@@ -807,15 +883,15 @@ use_clean:
 	c->journal_replay_seq_start	= last_seq;
 	c->journal_replay_seq_end	= blacklist_seq - 1;
 
-	if (c->opts.reconstruct_alloc)
-		bch2_reconstruct_alloc(c);
-
 	zero_out_btree_mem_ptr(&c->journal_keys);
 
 	ret = journal_replay_early(c, clean);
 	if (ret)
 		goto err;
 
+	if (c->opts.reconstruct_alloc)
+		bch2_reconstruct_alloc(c);
+
 	/*
 	 * After an unclean shutdown, skip then next few journal sequence
 	 * numbers as they may have been referenced by btree writes that
@@ -870,16 +946,17 @@ use_clean:
 	 */
 	set_bit(BCH_FS_may_go_rw, &c->flags);
 	clear_bit(BCH_FS_fsck_running, &c->flags);
+	clear_bit(BCH_FS_recovery_running, &c->flags);
 
 	/* in case we don't run journal replay, i.e. norecovery mode */
 	set_bit(BCH_FS_accounting_replay_done, &c->flags);
 
+	bch2_async_btree_node_rewrites_flush(c);
+
 	/* fsync if we fixed errors */
-	if (test_bit(BCH_FS_errors_fixed, &c->flags) &&
-	    bch2_write_ref_tryget(c, BCH_WRITE_REF_fsync)) {
+	if (test_bit(BCH_FS_errors_fixed, &c->flags)) {
 		bch2_journal_flush_all_pins(&c->journal);
 		bch2_journal_meta(&c->journal);
-		bch2_write_ref_put(c, BCH_WRITE_REF_fsync);
 	}
 
 	/* If we fixed errors, verify that fs is actually clean now: */
@@ -1021,7 +1098,7 @@ int bch2_fs_initialize(struct bch_fs *c)
 	bch2_check_version_downgrade(c);
 
 	if (c->opts.version_upgrade != BCH_VERSION_UPGRADE_none) {
-		bch2_sb_upgrade(c, bcachefs_metadata_version_current);
+		bch2_sb_upgrade(c, bcachefs_metadata_version_current, false);
 		SET_BCH_SB_VERSION_UPGRADE_COMPLETE(c->disk_sb.sb, bcachefs_metadata_version_current);
 		bch2_write_super(c);
 	}
@@ -1035,7 +1112,6 @@ int bch2_fs_initialize(struct bch_fs *c)
 	bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
-	c->curr_recovery_pass = BCH_RECOVERY_PASS_NR;
 	set_bit(BCH_FS_btree_running, &c->flags);
 	set_bit(BCH_FS_may_go_rw, &c->flags);
 
@@ -1076,9 +1152,6 @@ int bch2_fs_initialize(struct bch_fs *c)
 	if (ret)
 		goto err;
 
-	for_each_online_member(c, ca)
-		ca->new_fs_bucket_idx = 0;
-
 	ret = bch2_fs_freespace_init(c);
 	if (ret)
 		goto err;
@@ -1137,6 +1210,7 @@ int bch2_fs_initialize(struct bch_fs *c)
 	bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
+	c->curr_recovery_pass = BCH_RECOVERY_PASS_NR;
 	return 0;
 err:
 	bch_err_fn(c, ret);
diff --git a/fs/bcachefs/recovery.h b/fs/bcachefs/recovery.h
index 4bf818de1f2f..b0d55754b21b 100644
--- a/fs/bcachefs/recovery.h
+++ b/fs/bcachefs/recovery.h
@@ -2,7 +2,7 @@
 #ifndef _BCACHEFS_RECOVERY_H
 #define _BCACHEFS_RECOVERY_H
 
-void bch2_btree_lost_data(struct bch_fs *, enum btree_id);
+int bch2_btree_lost_data(struct bch_fs *, enum btree_id);
 
 int bch2_journal_replay(struct bch_fs *);
 
diff --git a/fs/bcachefs/recovery_passes.c b/fs/bcachefs/recovery_passes.c
index dff589ddc984..0b3c951c32da 100644
--- a/fs/bcachefs/recovery_passes.c
+++ b/fs/bcachefs/recovery_passes.c
@@ -46,7 +46,7 @@ static int bch2_set_may_go_rw(struct bch_fs *c)
 
 	set_bit(BCH_FS_may_go_rw, &c->flags);
 
-	if (keys->nr || c->opts.fsck || !c->sb.clean || c->opts.recovery_passes)
+	if (keys->nr || !c->opts.read_only || c->opts.fsck || !c->sb.clean || c->opts.recovery_passes)
 		return bch2_fs_read_write_early(c);
 	return 0;
 }
@@ -100,20 +100,34 @@ u64 bch2_recovery_passes_from_stable(u64 v)
 /*
  * For when we need to rewind recovery passes and run a pass we skipped:
  */
-int bch2_run_explicit_recovery_pass(struct bch_fs *c,
-				    enum bch_recovery_pass pass)
+static int __bch2_run_explicit_recovery_pass(struct bch_fs *c,
+					     enum bch_recovery_pass pass)
 {
-	if (c->opts.recovery_passes & BIT_ULL(pass))
+	if (c->curr_recovery_pass == ARRAY_SIZE(recovery_pass_fns))
+		return -BCH_ERR_not_in_recovery;
+
+	if (c->recovery_passes_complete & BIT_ULL(pass))
 		return 0;
 
-	bch_info(c, "running explicit recovery pass %s (%u), currently at %s (%u)",
-		 bch2_recovery_passes[pass], pass,
-		 bch2_recovery_passes[c->curr_recovery_pass], c->curr_recovery_pass);
+	bool print = !(c->opts.recovery_passes & BIT_ULL(pass));
+
+	if (pass < BCH_RECOVERY_PASS_set_may_go_rw &&
+	    c->curr_recovery_pass >= BCH_RECOVERY_PASS_set_may_go_rw) {
+		if (print)
+			bch_info(c, "need recovery pass %s (%u), but already rw",
+				 bch2_recovery_passes[pass], pass);
+		return -BCH_ERR_cannot_rewind_recovery;
+	}
+
+	if (print)
+		bch_info(c, "running explicit recovery pass %s (%u), currently at %s (%u)",
+			 bch2_recovery_passes[pass], pass,
+			 bch2_recovery_passes[c->curr_recovery_pass], c->curr_recovery_pass);
 
 	c->opts.recovery_passes |= BIT_ULL(pass);
 
-	if (c->curr_recovery_pass >= pass) {
-		c->curr_recovery_pass = pass;
+	if (c->curr_recovery_pass > pass) {
+		c->next_recovery_pass = pass;
 		c->recovery_passes_complete &= (1ULL << pass) >> 1;
 		return -BCH_ERR_restart_recovery;
 	} else {
@@ -121,6 +135,27 @@ int bch2_run_explicit_recovery_pass(struct bch_fs *c,
 	}
 }
 
+int bch2_run_explicit_recovery_pass(struct bch_fs *c,
+				    enum bch_recovery_pass pass)
+{
+	unsigned long flags;
+	spin_lock_irqsave(&c->recovery_pass_lock, flags);
+	int ret = __bch2_run_explicit_recovery_pass(c, pass);
+	spin_unlock_irqrestore(&c->recovery_pass_lock, flags);
+	return ret;
+}
+
+int bch2_run_explicit_recovery_pass_persistent_locked(struct bch_fs *c,
+					       enum bch_recovery_pass pass)
+{
+	lockdep_assert_held(&c->sb_lock);
+
+	struct bch_sb_field_ext *ext = bch2_sb_field_get(c->disk_sb.sb, ext);
+	__set_bit_le64(bch2_recovery_pass_to_stable(pass), ext->recovery_passes_required);
+
+	return bch2_run_explicit_recovery_pass(c, pass);
+}
+
 int bch2_run_explicit_recovery_pass_persistent(struct bch_fs *c,
 					       enum bch_recovery_pass pass)
 {
@@ -233,31 +268,48 @@ int bch2_run_recovery_passes(struct bch_fs *c)
 	 */
 	c->opts.recovery_passes_exclude &= ~BCH_RECOVERY_PASS_set_may_go_rw;
 
-	while (c->curr_recovery_pass < ARRAY_SIZE(recovery_pass_fns)) {
-		if (c->opts.recovery_pass_last &&
-		    c->curr_recovery_pass > c->opts.recovery_pass_last)
-			break;
-
-		if (should_run_recovery_pass(c, c->curr_recovery_pass)) {
-			unsigned pass = c->curr_recovery_pass;
+	while (c->curr_recovery_pass < ARRAY_SIZE(recovery_pass_fns) && !ret) {
+		c->next_recovery_pass = c->curr_recovery_pass + 1;
 
-			ret =   bch2_run_recovery_pass(c, c->curr_recovery_pass) ?:
-				bch2_journal_flush(&c->journal);
-			if (bch2_err_matches(ret, BCH_ERR_restart_recovery) ||
-			    (ret && c->curr_recovery_pass < pass))
-				continue;
-			if (ret)
-				break;
+		spin_lock_irq(&c->recovery_pass_lock);
+		unsigned pass = c->curr_recovery_pass;
 
-			c->recovery_passes_complete |= BIT_ULL(c->curr_recovery_pass);
+		if (c->opts.recovery_pass_last &&
+		    c->curr_recovery_pass > c->opts.recovery_pass_last) {
+			spin_unlock_irq(&c->recovery_pass_lock);
+			break;
 		}
 
-		c->recovery_pass_done = max(c->recovery_pass_done, c->curr_recovery_pass);
-
-		if (!test_bit(BCH_FS_error, &c->flags))
-			bch2_clear_recovery_pass_required(c, c->curr_recovery_pass);
-
-		c->curr_recovery_pass++;
+		if (!should_run_recovery_pass(c, pass)) {
+			c->curr_recovery_pass++;
+			c->recovery_pass_done = max(c->recovery_pass_done, pass);
+			spin_unlock_irq(&c->recovery_pass_lock);
+			continue;
+		}
+		spin_unlock_irq(&c->recovery_pass_lock);
+
+		ret =   bch2_run_recovery_pass(c, pass) ?:
+			bch2_journal_flush(&c->journal);
+
+		if (!ret && !test_bit(BCH_FS_error, &c->flags))
+			bch2_clear_recovery_pass_required(c, pass);
+
+		spin_lock_irq(&c->recovery_pass_lock);
+		if (c->next_recovery_pass < c->curr_recovery_pass) {
+			/*
+			 * bch2_run_explicit_recovery_pass() was called: we
+			 * can't always catch -BCH_ERR_restart_recovery because
+			 * it may have been called from another thread (btree
+			 * node read completion)
+			 */
+			ret = 0;
+			c->recovery_passes_complete &= ~(~0ULL << c->curr_recovery_pass);
+		} else {
+			c->recovery_passes_complete |= BIT_ULL(pass);
+			c->recovery_pass_done = max(c->recovery_pass_done, pass);
+		}
+		c->curr_recovery_pass = c->next_recovery_pass;
+		spin_unlock_irq(&c->recovery_pass_lock);
 	}
 
 	return ret;
diff --git a/fs/bcachefs/recovery_passes.h b/fs/bcachefs/recovery_passes.h
index 99b464e127b8..7d7339c8fa29 100644
--- a/fs/bcachefs/recovery_passes.h
+++ b/fs/bcachefs/recovery_passes.h
@@ -9,6 +9,7 @@ u64 bch2_recovery_passes_from_stable(u64 v);
 u64 bch2_fsck_recovery_passes(void);
 
 int bch2_run_explicit_recovery_pass(struct bch_fs *, enum bch_recovery_pass);
+int bch2_run_explicit_recovery_pass_persistent_locked(struct bch_fs *, enum bch_recovery_pass);
 int bch2_run_explicit_recovery_pass_persistent(struct bch_fs *, enum bch_recovery_pass);
 
 int bch2_run_online_recovery_passes(struct bch_fs *);
diff --git a/fs/bcachefs/recovery_passes_types.h b/fs/bcachefs/recovery_passes_types.h
index 94dc20ca2065..418557960ed6 100644
--- a/fs/bcachefs/recovery_passes_types.h
+++ b/fs/bcachefs/recovery_passes_types.h
@@ -8,53 +8,59 @@
 #define PASS_ALWAYS		BIT(3)
 #define PASS_ONLINE		BIT(4)
 
+#ifdef CONFIG_BCACHEFS_DEBUG
+#define PASS_FSCK_DEBUG		BIT(1)
+#else
+#define PASS_FSCK_DEBUG		0
+#endif
+
 /*
  * Passes may be reordered, but the second field is a persistent identifier and
  * must never change:
  */
-#define BCH_RECOVERY_PASSES()							\
-	x(recovery_pass_empty,			41, PASS_SILENT)		\
-	x(scan_for_btree_nodes,			37, 0)				\
-	x(check_topology,			 4, 0)				\
-	x(accounting_read,			39, PASS_ALWAYS)		\
-	x(alloc_read,				 0, PASS_ALWAYS)		\
-	x(stripes_read,				 1, PASS_ALWAYS)		\
-	x(initialize_subvolumes,		 2, 0)				\
-	x(snapshots_read,			 3, PASS_ALWAYS)		\
-	x(check_allocations,			 5, PASS_FSCK)			\
-	x(trans_mark_dev_sbs,			 6, PASS_ALWAYS|PASS_SILENT)	\
-	x(fs_journal_alloc,			 7, PASS_ALWAYS|PASS_SILENT)	\
-	x(set_may_go_rw,			 8, PASS_ALWAYS|PASS_SILENT)	\
-	x(journal_replay,			 9, PASS_ALWAYS)		\
-	x(check_alloc_info,			10, PASS_ONLINE|PASS_FSCK)	\
-	x(check_lrus,				11, PASS_ONLINE|PASS_FSCK)	\
-	x(check_btree_backpointers,		12, PASS_ONLINE|PASS_FSCK)	\
-	x(check_backpointers_to_extents,	13, PASS_ONLINE|PASS_FSCK)	\
-	x(check_extents_to_backpointers,	14, PASS_ONLINE|PASS_FSCK)	\
-	x(check_alloc_to_lru_refs,		15, PASS_ONLINE|PASS_FSCK)	\
-	x(fs_freespace_init,			16, PASS_ALWAYS|PASS_SILENT)	\
-	x(bucket_gens_init,			17, 0)				\
-	x(reconstruct_snapshots,		38, 0)				\
-	x(check_snapshot_trees,			18, PASS_ONLINE|PASS_FSCK)	\
-	x(check_snapshots,			19, PASS_ONLINE|PASS_FSCK)	\
-	x(check_subvols,			20, PASS_ONLINE|PASS_FSCK)	\
-	x(check_subvol_children,		35, PASS_ONLINE|PASS_FSCK)	\
-	x(delete_dead_snapshots,		21, PASS_ONLINE|PASS_FSCK)	\
-	x(fs_upgrade_for_subvolumes,		22, 0)				\
-	x(check_inodes,				24, PASS_FSCK)			\
-	x(check_extents,			25, PASS_FSCK)			\
-	x(check_indirect_extents,		26, PASS_FSCK)			\
-	x(check_dirents,			27, PASS_FSCK)			\
-	x(check_xattrs,				28, PASS_FSCK)			\
-	x(check_root,				29, PASS_ONLINE|PASS_FSCK)	\
-	x(check_unreachable_inodes,		40, PASS_ONLINE|PASS_FSCK)	\
-	x(check_subvolume_structure,		36, PASS_ONLINE|PASS_FSCK)	\
-	x(check_directory_structure,		30, PASS_ONLINE|PASS_FSCK)	\
-	x(check_nlinks,				31, PASS_FSCK)			\
-	x(resume_logged_ops,			23, PASS_ALWAYS)		\
-	x(delete_dead_inodes,			32, PASS_ALWAYS)		\
-	x(fix_reflink_p,			33, 0)				\
-	x(set_fs_needs_rebalance,		34, 0)				\
+#define BCH_RECOVERY_PASSES()								\
+	x(recovery_pass_empty,			41, PASS_SILENT)			\
+	x(scan_for_btree_nodes,			37, 0)					\
+	x(check_topology,			 4, 0)					\
+	x(accounting_read,			39, PASS_ALWAYS)			\
+	x(alloc_read,				 0, PASS_ALWAYS)			\
+	x(stripes_read,				 1, PASS_ALWAYS)			\
+	x(initialize_subvolumes,		 2, 0)					\
+	x(snapshots_read,			 3, PASS_ALWAYS)			\
+	x(check_allocations,			 5, PASS_FSCK)				\
+	x(trans_mark_dev_sbs,			 6, PASS_ALWAYS|PASS_SILENT)		\
+	x(fs_journal_alloc,			 7, PASS_ALWAYS|PASS_SILENT)		\
+	x(set_may_go_rw,			 8, PASS_ALWAYS|PASS_SILENT)		\
+	x(journal_replay,			 9, PASS_ALWAYS)			\
+	x(check_alloc_info,			10, PASS_ONLINE|PASS_FSCK)		\
+	x(check_lrus,				11, PASS_ONLINE|PASS_FSCK)		\
+	x(check_btree_backpointers,		12, PASS_ONLINE|PASS_FSCK)		\
+	x(check_backpointers_to_extents,	13, PASS_ONLINE|PASS_FSCK_DEBUG)	\
+	x(check_extents_to_backpointers,	14, PASS_ONLINE|PASS_FSCK)		\
+	x(check_alloc_to_lru_refs,		15, PASS_ONLINE|PASS_FSCK)		\
+	x(fs_freespace_init,			16, PASS_ALWAYS|PASS_SILENT)		\
+	x(bucket_gens_init,			17, 0)					\
+	x(reconstruct_snapshots,		38, 0)					\
+	x(check_snapshot_trees,			18, PASS_ONLINE|PASS_FSCK)		\
+	x(check_snapshots,			19, PASS_ONLINE|PASS_FSCK)		\
+	x(check_subvols,			20, PASS_ONLINE|PASS_FSCK)		\
+	x(check_subvol_children,		35, PASS_ONLINE|PASS_FSCK)		\
+	x(delete_dead_snapshots,		21, PASS_ONLINE|PASS_FSCK)		\
+	x(fs_upgrade_for_subvolumes,		22, 0)					\
+	x(check_inodes,				24, PASS_FSCK)				\
+	x(check_extents,			25, PASS_FSCK)				\
+	x(check_indirect_extents,		26, PASS_ONLINE|PASS_FSCK)		\
+	x(check_dirents,			27, PASS_FSCK)				\
+	x(check_xattrs,				28, PASS_FSCK)				\
+	x(check_root,				29, PASS_ONLINE|PASS_FSCK)		\
+	x(check_unreachable_inodes,		40, PASS_FSCK)				\
+	x(check_subvolume_structure,		36, PASS_ONLINE|PASS_FSCK)		\
+	x(check_directory_structure,		30, PASS_ONLINE|PASS_FSCK)		\
+	x(check_nlinks,				31, PASS_FSCK)				\
+	x(resume_logged_ops,			23, PASS_ALWAYS)			\
+	x(delete_dead_inodes,			32, PASS_ALWAYS)			\
+	x(fix_reflink_p,			33, 0)					\
+	x(set_fs_needs_rebalance,		34, 0)
 
 /* We normally enumerate recovery passes in the order we run them: */
 enum bch_recovery_pass {
diff --git a/fs/bcachefs/reflink.c b/fs/bcachefs/reflink.c
index f457925fa362..93ba4f4e47ca 100644
--- a/fs/bcachefs/reflink.c
+++ b/fs/bcachefs/reflink.c
@@ -15,6 +15,17 @@
 
 #include <linux/sched/signal.h>
 
+static inline bool bkey_extent_is_reflink_data(const struct bkey *k)
+{
+	switch (k->type) {
+	case KEY_TYPE_reflink_v:
+	case KEY_TYPE_indirect_inline_data:
+		return true;
+	default:
+		return false;
+	}
+}
+
 static inline unsigned bkey_type_to_indirect(const struct bkey *k)
 {
 	switch (k->type) {
@@ -30,15 +41,15 @@ static inline unsigned bkey_type_to_indirect(const struct bkey *k)
 /* reflink pointers */
 
 int bch2_reflink_p_validate(struct bch_fs *c, struct bkey_s_c k,
-			    enum bch_validate_flags flags)
+			    struct bkey_validate_context from)
 {
 	struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
 	int ret = 0;
 
-	bkey_fsck_err_on(le64_to_cpu(p.v->idx) < le32_to_cpu(p.v->front_pad),
+	bkey_fsck_err_on(REFLINK_P_IDX(p.v) < le32_to_cpu(p.v->front_pad),
 			 c, reflink_p_front_pad_bad,
 			 "idx < front_pad (%llu < %u)",
-			 le64_to_cpu(p.v->idx), le32_to_cpu(p.v->front_pad));
+			 REFLINK_P_IDX(p.v), le32_to_cpu(p.v->front_pad));
 fsck_err:
 	return ret;
 }
@@ -49,7 +60,7 @@ void bch2_reflink_p_to_text(struct printbuf *out, struct bch_fs *c,
 	struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
 
 	prt_printf(out, "idx %llu front_pad %u back_pad %u",
-	       le64_to_cpu(p.v->idx),
+	       REFLINK_P_IDX(p.v),
 	       le32_to_cpu(p.v->front_pad),
 	       le32_to_cpu(p.v->back_pad));
 }
@@ -65,49 +76,250 @@ bool bch2_reflink_p_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r
 	 */
 	return false;
 
-	if (le64_to_cpu(l.v->idx) + l.k->size != le64_to_cpu(r.v->idx))
+	if (REFLINK_P_IDX(l.v) + l.k->size != REFLINK_P_IDX(r.v))
+		return false;
+
+	if (REFLINK_P_ERROR(l.v) != REFLINK_P_ERROR(r.v))
 		return false;
 
 	bch2_key_resize(l.k, l.k->size + r.k->size);
 	return true;
 }
 
+/* indirect extents */
+
+int bch2_reflink_v_validate(struct bch_fs *c, struct bkey_s_c k,
+			    struct bkey_validate_context from)
+{
+	int ret = 0;
+
+	bkey_fsck_err_on(bkey_gt(k.k->p, POS(0, REFLINK_P_IDX_MAX)),
+			 c, reflink_v_pos_bad,
+			 "indirect extent above maximum position 0:%llu",
+			 REFLINK_P_IDX_MAX);
+
+	ret = bch2_bkey_ptrs_validate(c, k, from);
+fsck_err:
+	return ret;
+}
+
+void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
+			    struct bkey_s_c k)
+{
+	struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
+
+	prt_printf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
+
+	bch2_bkey_ptrs_to_text(out, c, k);
+}
+
+#if 0
+Currently disabled, needs to be debugged:
+
+bool bch2_reflink_v_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
+{
+	struct bkey_s_reflink_v   l = bkey_s_to_reflink_v(_l);
+	struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(_r);
+
+	return l.v->refcount == r.v->refcount && bch2_extent_merge(c, _l, _r);
+}
+#endif
+
+/* indirect inline data */
+
+int bch2_indirect_inline_data_validate(struct bch_fs *c, struct bkey_s_c k,
+				       struct bkey_validate_context from)
+{
+	return 0;
+}
+
+void bch2_indirect_inline_data_to_text(struct printbuf *out,
+				       struct bch_fs *c, struct bkey_s_c k)
+{
+	struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
+	unsigned datalen = bkey_inline_data_bytes(k.k);
+
+	prt_printf(out, "refcount %llu datalen %u: %*phN",
+	       le64_to_cpu(d.v->refcount), datalen,
+	       min(datalen, 32U), d.v->data);
+}
+
+/* lookup */
+
+static int bch2_indirect_extent_not_missing(struct btree_trans *trans, struct bkey_s_c_reflink_p p,
+					    bool should_commit)
+{
+	struct bkey_i_reflink_p *new = bch2_bkey_make_mut_noupdate_typed(trans, p.s_c, reflink_p);
+	int ret = PTR_ERR_OR_ZERO(new);
+	if (ret)
+		return ret;
+
+	SET_REFLINK_P_ERROR(&new->v, false);
+	ret = bch2_btree_insert_trans(trans, BTREE_ID_extents, &new->k_i, BTREE_TRIGGER_norun);
+	if (ret)
+		return ret;
+
+	if (!should_commit)
+		return 0;
+
+	return bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc) ?:
+		-BCH_ERR_transaction_restart_nested;
+}
+
+static int bch2_indirect_extent_missing_error(struct btree_trans *trans,
+					      struct bkey_s_c_reflink_p p,
+					      u64 missing_start, u64 missing_end,
+					      bool should_commit)
+{
+	if (REFLINK_P_ERROR(p.v))
+		return -BCH_ERR_missing_indirect_extent;
+
+	struct bch_fs *c = trans->c;
+	u64 live_start	= REFLINK_P_IDX(p.v);
+	u64 live_end	= REFLINK_P_IDX(p.v) + p.k->size;
+	u64 refd_start	= live_start	- le32_to_cpu(p.v->front_pad);
+	u64 refd_end	= live_end	+ le32_to_cpu(p.v->back_pad);
+	struct printbuf buf = PRINTBUF;
+	int ret = 0;
+
+	BUG_ON(missing_start	< refd_start);
+	BUG_ON(missing_end	> refd_end);
+
+	if (fsck_err(trans, reflink_p_to_missing_reflink_v,
+		     "pointer to missing indirect extent\n"
+		     "  %s\n"
+		     "  missing range %llu-%llu",
+		     (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
+		     missing_start, missing_end)) {
+		struct bkey_i_reflink_p *new = bch2_bkey_make_mut_noupdate_typed(trans, p.s_c, reflink_p);
+		ret = PTR_ERR_OR_ZERO(new);
+		if (ret)
+			goto err;
+
+		/*
+		 * Is the missing range not actually needed?
+		 *
+		 * p.v->idx refers to the data that we actually want, but if the
+		 * indirect extent we point to was bigger, front_pad and back_pad
+		 * indicate the range we took a reference on.
+		 */
+
+		if (missing_end <= live_start) {
+			new->v.front_pad = cpu_to_le32(live_start - missing_end);
+		} else if (missing_start >= live_end) {
+			new->v.back_pad = cpu_to_le32(missing_start - live_end);
+		} else {
+			struct bpos new_start	= bkey_start_pos(&new->k);
+			struct bpos new_end	= new->k.p;
+
+			if (missing_start > live_start)
+				new_start.offset += missing_start - live_start;
+			if (missing_end < live_end)
+				new_end.offset -= live_end - missing_end;
+
+			bch2_cut_front(new_start, &new->k_i);
+			bch2_cut_back(new_end, &new->k_i);
+
+			SET_REFLINK_P_ERROR(&new->v, true);
+		}
+
+		ret = bch2_btree_insert_trans(trans, BTREE_ID_extents, &new->k_i, BTREE_TRIGGER_norun);
+		if (ret)
+			goto err;
+
+		if (should_commit)
+			ret =   bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc) ?:
+				-BCH_ERR_transaction_restart_nested;
+	}
+err:
+fsck_err:
+	printbuf_exit(&buf);
+	return ret;
+}
+
+/*
+ * This is used from the read path, which doesn't expect to have to do a
+ * transaction commit, and from triggers, which should not be doing a commit:
+ */
+struct bkey_s_c bch2_lookup_indirect_extent(struct btree_trans *trans,
+					    struct btree_iter *iter,
+					    s64 *offset_into_extent,
+					    struct bkey_s_c_reflink_p p,
+					    bool should_commit,
+					    unsigned iter_flags)
+{
+	BUG_ON(*offset_into_extent < -((s64) le32_to_cpu(p.v->front_pad)));
+	BUG_ON(*offset_into_extent >= p.k->size + le32_to_cpu(p.v->back_pad));
+
+	u64 reflink_offset = REFLINK_P_IDX(p.v) + *offset_into_extent;
+
+	struct bkey_s_c k = bch2_bkey_get_iter(trans, iter, BTREE_ID_reflink,
+				       POS(0, reflink_offset), iter_flags);
+	if (bkey_err(k))
+		return k;
+
+	if (unlikely(!bkey_extent_is_reflink_data(k.k))) {
+		bch2_trans_iter_exit(trans, iter);
+
+		unsigned size = min((u64) k.k->size,
+				    REFLINK_P_IDX(p.v) + p.k->size + le32_to_cpu(p.v->back_pad) -
+				    reflink_offset);
+		bch2_key_resize(&iter->k, size);
+
+		int ret = bch2_indirect_extent_missing_error(trans, p, reflink_offset,
+							     k.k->p.offset, should_commit);
+		if (ret)
+			return bkey_s_c_err(ret);
+	} else if (unlikely(REFLINK_P_ERROR(p.v))) {
+		bch2_trans_iter_exit(trans, iter);
+
+		int ret = bch2_indirect_extent_not_missing(trans, p, should_commit);
+		if (ret)
+			return bkey_s_c_err(ret);
+	}
+
+	*offset_into_extent = reflink_offset - bkey_start_offset(k.k);
+	return k;
+}
+
+/* reflink pointer trigger */
+
 static int trans_trigger_reflink_p_segment(struct btree_trans *trans,
 			struct bkey_s_c_reflink_p p, u64 *idx,
 			enum btree_iter_update_trigger_flags flags)
 {
 	struct bch_fs *c = trans->c;
-	struct btree_iter iter;
-	struct bkey_i *k;
-	__le64 *refcount;
-	int add = !(flags & BTREE_TRIGGER_overwrite) ? 1 : -1;
 	struct printbuf buf = PRINTBUF;
-	int ret;
 
-	k = bch2_bkey_get_mut_noupdate(trans, &iter,
-			BTREE_ID_reflink, POS(0, *idx),
-			BTREE_ITER_with_updates);
-	ret = PTR_ERR_OR_ZERO(k);
+	s64 offset_into_extent = *idx - REFLINK_P_IDX(p.v);
+	struct btree_iter iter;
+	struct bkey_s_c k = bch2_lookup_indirect_extent(trans, &iter, &offset_into_extent, p, false,
+							BTREE_ITER_intent|
+							BTREE_ITER_with_updates);
+	int ret = bkey_err(k);
 	if (ret)
-		goto err;
+		return ret;
 
-	refcount = bkey_refcount(bkey_i_to_s(k));
-	if (!refcount) {
-		bch2_bkey_val_to_text(&buf, c, p.s_c);
-		bch2_trans_inconsistent(trans,
-			"nonexistent indirect extent at %llu while marking\n  %s",
-			*idx, buf.buf);
-		ret = -EIO;
-		goto err;
+	if (bkey_deleted(k.k)) {
+		if (!(flags & BTREE_TRIGGER_overwrite))
+			ret = -BCH_ERR_missing_indirect_extent;
+		goto next;
 	}
 
+	struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
+	ret = PTR_ERR_OR_ZERO(new);
+	if (ret)
+		goto err;
+
+	__le64 *refcount = bkey_refcount(bkey_i_to_s(new));
 	if (!*refcount && (flags & BTREE_TRIGGER_overwrite)) {
 		bch2_bkey_val_to_text(&buf, c, p.s_c);
-		bch2_trans_inconsistent(trans,
-			"indirect extent refcount underflow at %llu while marking\n  %s",
-			*idx, buf.buf);
-		ret = -EIO;
-		goto err;
+		prt_printf(&buf, "\n  ");
+		bch2_bkey_val_to_text(&buf, c, k);
+		log_fsck_err(trans, reflink_refcount_underflow,
+			     "indirect extent refcount underflow while marking\n  %s",
+			   buf.buf);
+		goto next;
 	}
 
 	if (flags & BTREE_TRIGGER_insert) {
@@ -115,25 +327,26 @@ static int trans_trigger_reflink_p_segment(struct btree_trans *trans,
 		u64 pad;
 
 		pad = max_t(s64, le32_to_cpu(v->front_pad),
-			    le64_to_cpu(v->idx) - bkey_start_offset(&k->k));
+			    REFLINK_P_IDX(v) - bkey_start_offset(&new->k));
 		BUG_ON(pad > U32_MAX);
 		v->front_pad = cpu_to_le32(pad);
 
 		pad = max_t(s64, le32_to_cpu(v->back_pad),
-			    k->k.p.offset - p.k->size - le64_to_cpu(v->idx));
+			    new->k.p.offset - p.k->size - REFLINK_P_IDX(v));
 		BUG_ON(pad > U32_MAX);
 		v->back_pad = cpu_to_le32(pad);
 	}
 
-	le64_add_cpu(refcount, add);
+	le64_add_cpu(refcount, !(flags & BTREE_TRIGGER_overwrite) ? 1 : -1);
 
 	bch2_btree_iter_set_pos_to_extent_start(&iter);
-	ret = bch2_trans_update(trans, &iter, k, 0);
+	ret = bch2_trans_update(trans, &iter, new, 0);
 	if (ret)
 		goto err;
-
-	*idx = k->k.p.offset;
+next:
+	*idx = k.k->p.offset;
 err:
+fsck_err:
 	bch2_trans_iter_exit(trans, &iter);
 	printbuf_exit(&buf);
 	return ret;
@@ -147,9 +360,7 @@ static s64 gc_trigger_reflink_p_segment(struct btree_trans *trans,
 	struct bch_fs *c = trans->c;
 	struct reflink_gc *r;
 	int add = !(flags & BTREE_TRIGGER_overwrite) ? 1 : -1;
-	u64 start = le64_to_cpu(p.v->idx);
-	u64 end = le64_to_cpu(p.v->idx) + p.k->size;
-	u64 next_idx = end + le32_to_cpu(p.v->back_pad);
+	u64 next_idx = REFLINK_P_IDX(p.v) + p.k->size + le32_to_cpu(p.v->back_pad);
 	s64 ret = 0;
 	struct printbuf buf = PRINTBUF;
 
@@ -168,36 +379,14 @@ static s64 gc_trigger_reflink_p_segment(struct btree_trans *trans,
 	*idx = r->offset;
 	return 0;
 not_found:
-	BUG_ON(!(flags & BTREE_TRIGGER_check_repair));
-
-	if (fsck_err(trans, reflink_p_to_missing_reflink_v,
-		     "pointer to missing indirect extent\n"
-		     "  %s\n"
-		     "  missing range %llu-%llu",
-		     (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
-		     *idx, next_idx)) {
-		struct bkey_i *update = bch2_bkey_make_mut_noupdate(trans, p.s_c);
-		ret = PTR_ERR_OR_ZERO(update);
+	if (flags & BTREE_TRIGGER_check_repair) {
+		ret = bch2_indirect_extent_missing_error(trans, p, *idx, next_idx, false);
 		if (ret)
 			goto err;
-
-		if (next_idx <= start) {
-			bkey_i_to_reflink_p(update)->v.front_pad = cpu_to_le32(start - next_idx);
-		} else if (*idx >= end) {
-			bkey_i_to_reflink_p(update)->v.back_pad = cpu_to_le32(*idx - end);
-		} else {
-			bkey_error_init(update);
-			update->k.p		= p.k->p;
-			update->k.size		= p.k->size;
-			set_bkey_val_u64s(&update->k, 0);
-		}
-
-		ret = bch2_btree_insert_trans(trans, BTREE_ID_extents, update, BTREE_TRIGGER_norun);
 	}
 
 	*idx = next_idx;
 err:
-fsck_err:
 	printbuf_exit(&buf);
 	return ret;
 }
@@ -210,8 +399,8 @@ static int __trigger_reflink_p(struct btree_trans *trans,
 	struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
 	int ret = 0;
 
-	u64 idx = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
-	u64 end = le64_to_cpu(p.v->idx) + p.k->size + le32_to_cpu(p.v->back_pad);
+	u64 idx = REFLINK_P_IDX(p.v) - le32_to_cpu(p.v->front_pad);
+	u64 end = REFLINK_P_IDX(p.v) + p.k->size + le32_to_cpu(p.v->back_pad);
 
 	if (flags & BTREE_TRIGGER_transactional) {
 		while (idx < end && !ret)
@@ -253,35 +442,7 @@ int bch2_trigger_reflink_p(struct btree_trans *trans,
 	return trigger_run_overwrite_then_insert(__trigger_reflink_p, trans, btree_id, level, old, new, flags);
 }
 
-/* indirect extents */
-
-int bch2_reflink_v_validate(struct bch_fs *c, struct bkey_s_c k,
-			    enum bch_validate_flags flags)
-{
-	return bch2_bkey_ptrs_validate(c, k, flags);
-}
-
-void bch2_reflink_v_to_text(struct printbuf *out, struct bch_fs *c,
-			    struct bkey_s_c k)
-{
-	struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(k);
-
-	prt_printf(out, "refcount: %llu ", le64_to_cpu(r.v->refcount));
-
-	bch2_bkey_ptrs_to_text(out, c, k);
-}
-
-#if 0
-Currently disabled, needs to be debugged:
-
-bool bch2_reflink_v_merge(struct bch_fs *c, struct bkey_s _l, struct bkey_s_c _r)
-{
-	struct bkey_s_reflink_v   l = bkey_s_to_reflink_v(_l);
-	struct bkey_s_c_reflink_v r = bkey_s_c_to_reflink_v(_r);
-
-	return l.v->refcount == r.v->refcount && bch2_extent_merge(c, _l, _r);
-}
-#endif
+/* indirect extent trigger */
 
 static inline void
 check_indirect_extent_deleting(struct bkey_s new,
@@ -307,25 +468,6 @@ int bch2_trigger_reflink_v(struct btree_trans *trans,
 	return bch2_trigger_extent(trans, btree_id, level, old, new, flags);
 }
 
-/* indirect inline data */
-
-int bch2_indirect_inline_data_validate(struct bch_fs *c, struct bkey_s_c k,
-				      enum bch_validate_flags flags)
-{
-	return 0;
-}
-
-void bch2_indirect_inline_data_to_text(struct printbuf *out,
-				       struct bch_fs *c, struct bkey_s_c k)
-{
-	struct bkey_s_c_indirect_inline_data d = bkey_s_c_to_indirect_inline_data(k);
-	unsigned datalen = bkey_inline_data_bytes(k.k);
-
-	prt_printf(out, "refcount %llu datalen %u: %*phN",
-	       le64_to_cpu(d.v->refcount), datalen,
-	       min(datalen, 32U), d.v->data);
-}
-
 int bch2_trigger_indirect_inline_data(struct btree_trans *trans,
 			      enum btree_id btree_id, unsigned level,
 			      struct bkey_s_c old, struct bkey_s new,
@@ -336,9 +478,12 @@ int bch2_trigger_indirect_inline_data(struct btree_trans *trans,
 	return 0;
 }
 
+/* create */
+
 static int bch2_make_extent_indirect(struct btree_trans *trans,
 				     struct btree_iter *extent_iter,
-				     struct bkey_i *orig)
+				     struct bkey_i *orig,
+				     bool reflink_p_may_update_opts_field)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter reflink_iter = { NULL };
@@ -358,6 +503,14 @@ static int bch2_make_extent_indirect(struct btree_trans *trans,
 	if (ret)
 		goto err;
 
+	/*
+	 * XXX: we're assuming that 56 bits will be enough for the life of the
+	 * filesystem: we need to implement wraparound, with a cursor in the
+	 * logged ops btree:
+	 */
+	if (bkey_ge(reflink_iter.pos, POS(0, REFLINK_P_IDX_MAX - orig->k.size)))
+		return -ENOSPC;
+
 	r_v = bch2_trans_kmalloc(trans, sizeof(__le64) + bkey_bytes(&orig->k));
 	ret = PTR_ERR_OR_ZERO(r_v);
 	if (ret)
@@ -394,7 +547,10 @@ static int bch2_make_extent_indirect(struct btree_trans *trans,
 	memset(&r_p->v, 0, sizeof(r_p->v));
 #endif
 
-	r_p->v.idx = cpu_to_le64(bkey_start_offset(&r_v->k));
+	SET_REFLINK_P_IDX(&r_p->v, bkey_start_offset(&r_v->k));
+
+	if (reflink_p_may_update_opts_field)
+		SET_REFLINK_P_MAY_UPDATE_OPTIONS(&r_p->v, true);
 
 	ret = bch2_trans_update(trans, extent_iter, &r_p->k_i,
 				BTREE_UPDATE_internal_snapshot_node);
@@ -409,7 +565,7 @@ static struct bkey_s_c get_next_src(struct btree_iter *iter, struct bpos end)
 	struct bkey_s_c k;
 	int ret;
 
-	for_each_btree_key_upto_continue_norestart(*iter, end, 0, k, ret) {
+	for_each_btree_key_max_continue_norestart(*iter, end, 0, k, ret) {
 		if (bkey_extent_is_unwritten(k))
 			continue;
 
@@ -426,7 +582,8 @@ s64 bch2_remap_range(struct bch_fs *c,
 		     subvol_inum dst_inum, u64 dst_offset,
 		     subvol_inum src_inum, u64 src_offset,
 		     u64 remap_sectors,
-		     u64 new_i_size, s64 *i_sectors_delta)
+		     u64 new_i_size, s64 *i_sectors_delta,
+		     bool may_change_src_io_path_opts)
 {
 	struct btree_trans *trans;
 	struct btree_iter dst_iter, src_iter;
@@ -439,6 +596,8 @@ s64 bch2_remap_range(struct bch_fs *c,
 	struct bpos src_want;
 	u64 dst_done = 0;
 	u32 dst_snapshot, src_snapshot;
+	bool reflink_p_may_update_opts_field =
+		bch2_request_incompat_feature(c, bcachefs_metadata_version_reflink_p_may_update_opts);
 	int ret = 0, ret2 = 0;
 
 	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_reflink))
@@ -520,7 +679,8 @@ s64 bch2_remap_range(struct bch_fs *c,
 			src_k = bkey_i_to_s_c(new_src.k);
 
 			ret = bch2_make_extent_indirect(trans, &src_iter,
-						new_src.k);
+						new_src.k,
+						reflink_p_may_update_opts_field);
 			if (ret)
 				continue;
 
@@ -533,11 +693,15 @@ s64 bch2_remap_range(struct bch_fs *c,
 			struct bkey_i_reflink_p *dst_p =
 				bkey_reflink_p_init(new_dst.k);
 
-			u64 offset = le64_to_cpu(src_p.v->idx) +
+			u64 offset = REFLINK_P_IDX(src_p.v) +
 				(src_want.offset -
 				 bkey_start_offset(src_k.k));
 
-			dst_p->v.idx = cpu_to_le64(offset);
+			SET_REFLINK_P_IDX(&dst_p->v, offset);
+
+			if (reflink_p_may_update_opts_field &&
+			    may_change_src_io_path_opts)
+				SET_REFLINK_P_MAY_UPDATE_OPTIONS(&dst_p->v, true);
 		} else {
 			BUG();
 		}
@@ -547,7 +711,7 @@ s64 bch2_remap_range(struct bch_fs *c,
 				min(src_k.k->p.offset - src_want.offset,
 				    dst_end.offset - dst_iter.pos.offset));
 
-		ret =   bch2_bkey_set_needs_rebalance(c, new_dst.k, &opts) ?:
+		ret =   bch2_bkey_set_needs_rebalance(c, &opts, new_dst.k) ?:
 			bch2_extent_update(trans, dst_inum, &dst_iter,
 					new_dst.k, &disk_res,
 					new_i_size, i_sectors_delta,
@@ -591,3 +755,97 @@ err:
 
 	return dst_done ?: ret ?: ret2;
 }
+
+/* fsck */
+
+static int bch2_gc_write_reflink_key(struct btree_trans *trans,
+				     struct btree_iter *iter,
+				     struct bkey_s_c k,
+				     size_t *idx)
+{
+	struct bch_fs *c = trans->c;
+	const __le64 *refcount = bkey_refcount_c(k);
+	struct printbuf buf = PRINTBUF;
+	struct reflink_gc *r;
+	int ret = 0;
+
+	if (!refcount)
+		return 0;
+
+	while ((r = genradix_ptr(&c->reflink_gc_table, *idx)) &&
+	       r->offset < k.k->p.offset)
+		++*idx;
+
+	if (!r ||
+	    r->offset != k.k->p.offset ||
+	    r->size != k.k->size) {
+		bch_err(c, "unexpected inconsistency walking reflink table at gc finish");
+		return -EINVAL;
+	}
+
+	if (fsck_err_on(r->refcount != le64_to_cpu(*refcount),
+			trans, reflink_v_refcount_wrong,
+			"reflink key has wrong refcount:\n"
+			"  %s\n"
+			"  should be %u",
+			(bch2_bkey_val_to_text(&buf, c, k), buf.buf),
+			r->refcount)) {
+		struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
+		ret = PTR_ERR_OR_ZERO(new);
+		if (ret)
+			goto out;
+
+		if (!r->refcount)
+			new->k.type = KEY_TYPE_deleted;
+		else
+			*bkey_refcount(bkey_i_to_s(new)) = cpu_to_le64(r->refcount);
+		ret = bch2_trans_update(trans, iter, new, 0);
+	}
+out:
+fsck_err:
+	printbuf_exit(&buf);
+	return ret;
+}
+
+int bch2_gc_reflink_done(struct bch_fs *c)
+{
+	size_t idx = 0;
+
+	int ret = bch2_trans_run(c,
+		for_each_btree_key_commit(trans, iter,
+				BTREE_ID_reflink, POS_MIN,
+				BTREE_ITER_prefetch, k,
+				NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
+			bch2_gc_write_reflink_key(trans, &iter, k, &idx)));
+	c->reflink_gc_nr = 0;
+	return ret;
+}
+
+int bch2_gc_reflink_start(struct bch_fs *c)
+{
+	c->reflink_gc_nr = 0;
+
+	int ret = bch2_trans_run(c,
+		for_each_btree_key(trans, iter, BTREE_ID_reflink, POS_MIN,
+				   BTREE_ITER_prefetch, k, ({
+			const __le64 *refcount = bkey_refcount_c(k);
+
+			if (!refcount)
+				continue;
+
+			struct reflink_gc *r = genradix_ptr_alloc(&c->reflink_gc_table,
+							c->reflink_gc_nr++, GFP_KERNEL);
+			if (!r) {
+				ret = -BCH_ERR_ENOMEM_gc_reflink_start;
+				break;
+			}
+
+			r->offset	= k.k->p.offset;
+			r->size		= k.k->size;
+			r->refcount	= 0;
+			0;
+		})));
+
+	bch_err_fn(c, ret);
+	return ret;
+}
diff --git a/fs/bcachefs/reflink.h b/fs/bcachefs/reflink.h
index 51afe11d8ed6..1632780bdf18 100644
--- a/fs/bcachefs/reflink.h
+++ b/fs/bcachefs/reflink.h
@@ -2,9 +2,8 @@
 #ifndef _BCACHEFS_REFLINK_H
 #define _BCACHEFS_REFLINK_H
 
-enum bch_validate_flags;
-
-int bch2_reflink_p_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_reflink_p_validate(struct bch_fs *, struct bkey_s_c,
+			    struct bkey_validate_context);
 void bch2_reflink_p_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 bool bch2_reflink_p_merge(struct bch_fs *, struct bkey_s, struct bkey_s_c);
 int bch2_trigger_reflink_p(struct btree_trans *, enum btree_id, unsigned,
@@ -19,7 +18,8 @@ int bch2_trigger_reflink_p(struct btree_trans *, enum btree_id, unsigned,
 	.min_val_size	= 16,					\
 })
 
-int bch2_reflink_v_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_reflink_v_validate(struct bch_fs *, struct bkey_s_c,
+			    struct bkey_validate_context);
 void bch2_reflink_v_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 int bch2_trigger_reflink_v(struct btree_trans *, enum btree_id, unsigned,
 			   struct bkey_s_c, struct bkey_s,
@@ -34,7 +34,7 @@ int bch2_trigger_reflink_v(struct btree_trans *, enum btree_id, unsigned,
 })
 
 int bch2_indirect_inline_data_validate(struct bch_fs *, struct bkey_s_c,
-				      enum bch_validate_flags);
+				       struct bkey_validate_context);
 void bch2_indirect_inline_data_to_text(struct printbuf *,
 				struct bch_fs *, struct bkey_s_c);
 int bch2_trigger_indirect_inline_data(struct btree_trans *,
@@ -73,7 +73,15 @@ static inline __le64 *bkey_refcount(struct bkey_s k)
 	}
 }
 
+struct bkey_s_c bch2_lookup_indirect_extent(struct btree_trans *, struct btree_iter *,
+					    s64 *, struct bkey_s_c_reflink_p,
+					    bool, unsigned);
+
 s64 bch2_remap_range(struct bch_fs *, subvol_inum, u64,
-		     subvol_inum, u64, u64, u64, s64 *);
+		     subvol_inum, u64, u64, u64, s64 *,
+		     bool);
+
+int bch2_gc_reflink_done(struct bch_fs *);
+int bch2_gc_reflink_start(struct bch_fs *);
 
 #endif /* _BCACHEFS_REFLINK_H */
diff --git a/fs/bcachefs/reflink_format.h b/fs/bcachefs/reflink_format.h
index 6772eebb1fc6..92995e4f898e 100644
--- a/fs/bcachefs/reflink_format.h
+++ b/fs/bcachefs/reflink_format.h
@@ -4,7 +4,7 @@
 
 struct bch_reflink_p {
 	struct bch_val		v;
-	__le64			idx;
+	__le64			idx_flags;
 	/*
 	 * A reflink pointer might point to an indirect extent which is then
 	 * later split (by copygc or rebalance). If we only pointed to part of
@@ -17,6 +17,11 @@ struct bch_reflink_p {
 	__le32			back_pad;
 } __packed __aligned(8);
 
+LE64_BITMASK(REFLINK_P_IDX,	struct bch_reflink_p, idx_flags,  0, 56);
+LE64_BITMASK(REFLINK_P_ERROR,	struct bch_reflink_p, idx_flags, 56, 57);
+LE64_BITMASK(REFLINK_P_MAY_UPDATE_OPTIONS,
+				struct bch_reflink_p, idx_flags, 57, 58);
+
 struct bch_reflink_v {
 	struct bch_val		v;
 	__le64			refcount;
diff --git a/fs/bcachefs/sb-clean.c b/fs/bcachefs/sb-clean.c
index 005275281804..59c8770e4a0e 100644
--- a/fs/bcachefs/sb-clean.c
+++ b/fs/bcachefs/sb-clean.c
@@ -23,6 +23,10 @@
 int bch2_sb_clean_validate_late(struct bch_fs *c, struct bch_sb_field_clean *clean,
 				int write)
 {
+	struct bkey_validate_context from = {
+		.flags		= write,
+		.from		= BKEY_VALIDATE_superblock,
+	};
 	struct jset_entry *entry;
 	int ret;
 
@@ -40,7 +44,7 @@ int bch2_sb_clean_validate_late(struct bch_fs *c, struct bch_sb_field_clean *cle
 		ret = bch2_journal_entry_validate(c, NULL, entry,
 						  le16_to_cpu(c->disk_sb.sb->version),
 						  BCH_SB_BIG_ENDIAN(c->disk_sb.sb),
-						  write);
+						  from);
 		if (ret)
 			return ret;
 	}
diff --git a/fs/bcachefs/sb-counters_format.h b/fs/bcachefs/sb-counters_format.h
index 62ea478215d0..fdcf598f08b1 100644
--- a/fs/bcachefs/sb-counters_format.h
+++ b/fs/bcachefs/sb-counters_format.h
@@ -2,86 +2,91 @@
 #ifndef _BCACHEFS_SB_COUNTERS_FORMAT_H
 #define _BCACHEFS_SB_COUNTERS_FORMAT_H
 
-#define BCH_PERSISTENT_COUNTERS()				\
-	x(io_read,					0)	\
-	x(io_write,					1)	\
-	x(io_move,					2)	\
-	x(bucket_invalidate,				3)	\
-	x(bucket_discard,				4)	\
-	x(bucket_alloc,					5)	\
-	x(bucket_alloc_fail,				6)	\
-	x(btree_cache_scan,				7)	\
-	x(btree_cache_reap,				8)	\
-	x(btree_cache_cannibalize,			9)	\
-	x(btree_cache_cannibalize_lock,			10)	\
-	x(btree_cache_cannibalize_lock_fail,		11)	\
-	x(btree_cache_cannibalize_unlock,		12)	\
-	x(btree_node_write,				13)	\
-	x(btree_node_read,				14)	\
-	x(btree_node_compact,				15)	\
-	x(btree_node_merge,				16)	\
-	x(btree_node_split,				17)	\
-	x(btree_node_rewrite,				18)	\
-	x(btree_node_alloc,				19)	\
-	x(btree_node_free,				20)	\
-	x(btree_node_set_root,				21)	\
-	x(btree_path_relock_fail,			22)	\
-	x(btree_path_upgrade_fail,			23)	\
-	x(btree_reserve_get_fail,			24)	\
-	x(journal_entry_full,				25)	\
-	x(journal_full,					26)	\
-	x(journal_reclaim_finish,			27)	\
-	x(journal_reclaim_start,			28)	\
-	x(journal_write,				29)	\
-	x(read_promote,					30)	\
-	x(read_bounce,					31)	\
-	x(read_split,					33)	\
-	x(read_retry,					32)	\
-	x(read_reuse_race,				34)	\
-	x(move_extent_read,				35)	\
-	x(move_extent_write,				36)	\
-	x(move_extent_finish,				37)	\
-	x(move_extent_fail,				38)	\
-	x(move_extent_start_fail,			39)	\
-	x(copygc,					40)	\
-	x(copygc_wait,					41)	\
-	x(gc_gens_end,					42)	\
-	x(gc_gens_start,				43)	\
-	x(trans_blocked_journal_reclaim,		44)	\
-	x(trans_restart_btree_node_reused,		45)	\
-	x(trans_restart_btree_node_split,		46)	\
-	x(trans_restart_fault_inject,			47)	\
-	x(trans_restart_iter_upgrade,			48)	\
-	x(trans_restart_journal_preres_get,		49)	\
-	x(trans_restart_journal_reclaim,		50)	\
-	x(trans_restart_journal_res_get,		51)	\
-	x(trans_restart_key_cache_key_realloced,	52)	\
-	x(trans_restart_key_cache_raced,		53)	\
-	x(trans_restart_mark_replicas,			54)	\
-	x(trans_restart_mem_realloced,			55)	\
-	x(trans_restart_memory_allocation_failure,	56)	\
-	x(trans_restart_relock,				57)	\
-	x(trans_restart_relock_after_fill,		58)	\
-	x(trans_restart_relock_key_cache_fill,		59)	\
-	x(trans_restart_relock_next_node,		60)	\
-	x(trans_restart_relock_parent_for_fill,		61)	\
-	x(trans_restart_relock_path,			62)	\
-	x(trans_restart_relock_path_intent,		63)	\
-	x(trans_restart_too_many_iters,			64)	\
-	x(trans_restart_traverse,			65)	\
-	x(trans_restart_upgrade,			66)	\
-	x(trans_restart_would_deadlock,			67)	\
-	x(trans_restart_would_deadlock_write,		68)	\
-	x(trans_restart_injected,			69)	\
-	x(trans_restart_key_cache_upgrade,		70)	\
-	x(trans_traverse_all,				71)	\
-	x(transaction_commit,				72)	\
-	x(write_super,					73)	\
-	x(trans_restart_would_deadlock_recursion_limit,	74)	\
-	x(trans_restart_write_buffer_flush,		75)	\
-	x(trans_restart_split_race,			76)	\
-	x(write_buffer_flush_slowpath,			77)	\
-	x(write_buffer_flush_sync,			78)
+enum counters_flags {
+	TYPE_COUNTER	= BIT(0),	/* event counters */
+	TYPE_SECTORS	= BIT(1),	/* amount counters, the unit is sectors */
+};
+
+#define BCH_PERSISTENT_COUNTERS()					\
+	x(io_read,					0,	TYPE_SECTORS)	\
+	x(io_write,					1,	TYPE_SECTORS)	\
+	x(io_move,					2,	TYPE_SECTORS)	\
+	x(bucket_invalidate,				3,	TYPE_COUNTER)	\
+	x(bucket_discard,				4,	TYPE_COUNTER)	\
+	x(bucket_alloc,					5,	TYPE_COUNTER)	\
+	x(bucket_alloc_fail,				6,	TYPE_COUNTER)	\
+	x(btree_cache_scan,				7,	TYPE_COUNTER)	\
+	x(btree_cache_reap,				8,	TYPE_COUNTER)	\
+	x(btree_cache_cannibalize,			9,	TYPE_COUNTER)	\
+	x(btree_cache_cannibalize_lock,			10,	TYPE_COUNTER)	\
+	x(btree_cache_cannibalize_lock_fail,		11,	TYPE_COUNTER)	\
+	x(btree_cache_cannibalize_unlock,		12,	TYPE_COUNTER)	\
+	x(btree_node_write,				13,	TYPE_COUNTER)	\
+	x(btree_node_read,				14,	TYPE_COUNTER)	\
+	x(btree_node_compact,				15,	TYPE_COUNTER)	\
+	x(btree_node_merge,				16,	TYPE_COUNTER)	\
+	x(btree_node_split,				17,	TYPE_COUNTER)	\
+	x(btree_node_rewrite,				18,	TYPE_COUNTER)	\
+	x(btree_node_alloc,				19,	TYPE_COUNTER)	\
+	x(btree_node_free,				20,	TYPE_COUNTER)	\
+	x(btree_node_set_root,				21,	TYPE_COUNTER)	\
+	x(btree_path_relock_fail,			22,	TYPE_COUNTER)	\
+	x(btree_path_upgrade_fail,			23,	TYPE_COUNTER)	\
+	x(btree_reserve_get_fail,			24,	TYPE_COUNTER)	\
+	x(journal_entry_full,				25,	TYPE_COUNTER)	\
+	x(journal_full,					26,	TYPE_COUNTER)	\
+	x(journal_reclaim_finish,			27,	TYPE_COUNTER)	\
+	x(journal_reclaim_start,			28,	TYPE_COUNTER)	\
+	x(journal_write,				29,	TYPE_COUNTER)	\
+	x(read_promote,					30,	TYPE_COUNTER)	\
+	x(read_bounce,					31,	TYPE_COUNTER)	\
+	x(read_split,					33,	TYPE_COUNTER)	\
+	x(read_retry,					32,	TYPE_COUNTER)	\
+	x(read_reuse_race,				34,	TYPE_COUNTER)	\
+	x(move_extent_read,				35,	TYPE_SECTORS)	\
+	x(move_extent_write,				36,	TYPE_SECTORS)	\
+	x(move_extent_finish,				37,	TYPE_SECTORS)	\
+	x(move_extent_fail,				38,	TYPE_COUNTER)	\
+	x(move_extent_start_fail,			39,	TYPE_COUNTER)	\
+	x(copygc,					40,	TYPE_COUNTER)	\
+	x(copygc_wait,					41,	TYPE_COUNTER)	\
+	x(gc_gens_end,					42,	TYPE_COUNTER)	\
+	x(gc_gens_start,				43,	TYPE_COUNTER)	\
+	x(trans_blocked_journal_reclaim,		44,	TYPE_COUNTER)	\
+	x(trans_restart_btree_node_reused,		45,	TYPE_COUNTER)	\
+	x(trans_restart_btree_node_split,		46,	TYPE_COUNTER)	\
+	x(trans_restart_fault_inject,			47,	TYPE_COUNTER)	\
+	x(trans_restart_iter_upgrade,			48,	TYPE_COUNTER)	\
+	x(trans_restart_journal_preres_get,		49,	TYPE_COUNTER)	\
+	x(trans_restart_journal_reclaim,		50,	TYPE_COUNTER)	\
+	x(trans_restart_journal_res_get,		51,	TYPE_COUNTER)	\
+	x(trans_restart_key_cache_key_realloced,	52,	TYPE_COUNTER)	\
+	x(trans_restart_key_cache_raced,		53,	TYPE_COUNTER)	\
+	x(trans_restart_mark_replicas,			54,	TYPE_COUNTER)	\
+	x(trans_restart_mem_realloced,			55,	TYPE_COUNTER)	\
+	x(trans_restart_memory_allocation_failure,	56,	TYPE_COUNTER)	\
+	x(trans_restart_relock,				57,	TYPE_COUNTER)	\
+	x(trans_restart_relock_after_fill,		58,	TYPE_COUNTER)	\
+	x(trans_restart_relock_key_cache_fill,		59,	TYPE_COUNTER)	\
+	x(trans_restart_relock_next_node,		60,	TYPE_COUNTER)	\
+	x(trans_restart_relock_parent_for_fill,		61,	TYPE_COUNTER)	\
+	x(trans_restart_relock_path,			62,	TYPE_COUNTER)	\
+	x(trans_restart_relock_path_intent,		63,	TYPE_COUNTER)	\
+	x(trans_restart_too_many_iters,			64,	TYPE_COUNTER)	\
+	x(trans_restart_traverse,			65,	TYPE_COUNTER)	\
+	x(trans_restart_upgrade,			66,	TYPE_COUNTER)	\
+	x(trans_restart_would_deadlock,			67,	TYPE_COUNTER)	\
+	x(trans_restart_would_deadlock_write,		68,	TYPE_COUNTER)	\
+	x(trans_restart_injected,			69,	TYPE_COUNTER)	\
+	x(trans_restart_key_cache_upgrade,		70,	TYPE_COUNTER)	\
+	x(trans_traverse_all,				71,	TYPE_COUNTER)	\
+	x(transaction_commit,				72,	TYPE_COUNTER)	\
+	x(write_super,					73,	TYPE_COUNTER)	\
+	x(trans_restart_would_deadlock_recursion_limit,	74,	TYPE_COUNTER)	\
+	x(trans_restart_write_buffer_flush,		75,	TYPE_COUNTER)	\
+	x(trans_restart_split_race,			76,	TYPE_COUNTER)	\
+	x(write_buffer_flush_slowpath,			77,	TYPE_COUNTER)	\
+	x(write_buffer_flush_sync,			78,	TYPE_COUNTER)
 
 enum bch_persistent_counters {
 #define x(t, n, ...) BCH_COUNTER_##t,
diff --git a/fs/bcachefs/sb-downgrade.c b/fs/bcachefs/sb-downgrade.c
index 8767c33c2b51..14f6b6a5fb38 100644
--- a/fs/bcachefs/sb-downgrade.c
+++ b/fs/bcachefs/sb-downgrade.c
@@ -81,7 +81,19 @@
 	  BCH_FSCK_ERR_accounting_mismatch)			\
 	x(inode_has_child_snapshots,				\
 	  BIT_ULL(BCH_RECOVERY_PASS_check_inodes),		\
-	  BCH_FSCK_ERR_inode_has_child_snapshots_wrong)
+	  BCH_FSCK_ERR_inode_has_child_snapshots_wrong)		\
+	x(backpointer_bucket_gen,				\
+	  BIT_ULL(BCH_RECOVERY_PASS_check_extents_to_backpointers),\
+	  BCH_FSCK_ERR_backpointer_to_missing_ptr,		\
+	  BCH_FSCK_ERR_ptr_to_missing_backpointer)		\
+	x(disk_accounting_big_endian,				\
+	  BIT_ULL(BCH_RECOVERY_PASS_check_allocations),		\
+	  BCH_FSCK_ERR_accounting_mismatch,			\
+	  BCH_FSCK_ERR_accounting_key_replicas_nr_devs_0,	\
+	  BCH_FSCK_ERR_accounting_key_junk_at_end)		\
+	x(directory_size,					\
+	  BIT_ULL(BCH_RECOVERY_PASS_check_inodes),		\
+	  BCH_FSCK_ERR_directory_size_mismatch)			\
 
 #define DOWNGRADE_TABLE()					\
 	x(bucket_stripe_sectors,				\
@@ -117,7 +129,19 @@
 	  BCH_FSCK_ERR_bkey_version_in_future)			\
 	x(rebalance_work_acct_fix,				\
 	  BIT_ULL(BCH_RECOVERY_PASS_check_allocations),		\
-	  BCH_FSCK_ERR_accounting_mismatch)
+	  BCH_FSCK_ERR_accounting_mismatch,			\
+	  BCH_FSCK_ERR_accounting_key_replicas_nr_devs_0,	\
+	  BCH_FSCK_ERR_accounting_key_junk_at_end)		\
+	x(backpointer_bucket_gen,				\
+	  BIT_ULL(BCH_RECOVERY_PASS_check_extents_to_backpointers),\
+	  BCH_FSCK_ERR_backpointer_bucket_offset_wrong,		\
+	  BCH_FSCK_ERR_backpointer_to_missing_ptr,		\
+	  BCH_FSCK_ERR_ptr_to_missing_backpointer)		\
+	x(disk_accounting_big_endian,				\
+	  BIT_ULL(BCH_RECOVERY_PASS_check_allocations),		\
+	  BCH_FSCK_ERR_accounting_mismatch,			\
+	  BCH_FSCK_ERR_accounting_key_replicas_nr_devs_0,	\
+	  BCH_FSCK_ERR_accounting_key_junk_at_end)
 
 struct upgrade_downgrade_entry {
 	u64		recovery_passes;
diff --git a/fs/bcachefs/sb-errors_format.h b/fs/bcachefs/sb-errors_format.h
index 9feb6739f77a..0b4fe899209b 100644
--- a/fs/bcachefs/sb-errors_format.h
+++ b/fs/bcachefs/sb-errors_format.h
@@ -5,9 +5,8 @@
 enum bch_fsck_flags {
 	FSCK_CAN_FIX		= 1 << 0,
 	FSCK_CAN_IGNORE		= 1 << 1,
-	FSCK_NEED_FSCK		= 1 << 2,
-	FSCK_NO_RATELIMIT	= 1 << 3,
-	FSCK_AUTOFIX		= 1 << 4,
+	FSCK_NO_RATELIMIT	= 1 << 2,
+	FSCK_AUTOFIX		= 1 << 3,
 };
 
 #define BCH_SB_ERRS()									\
@@ -59,7 +58,7 @@ enum bch_fsck_flags {
 	x(bset_empty,						 45,	0)		\
 	x(bset_bad_seq,						 46,	0)		\
 	x(bset_blacklisted_journal_seq,				 47,	0)		\
-	x(first_bset_blacklisted_journal_seq,			 48,	0)		\
+	x(first_bset_blacklisted_journal_seq,			 48,	FSCK_AUTOFIX)	\
 	x(btree_node_bad_btree,					 49,	0)		\
 	x(btree_node_bad_level,					 50,	0)		\
 	x(btree_node_bad_min_key,				 51,	0)		\
@@ -68,17 +67,17 @@ enum bch_fsck_flags {
 	x(btree_node_bkey_past_bset_end,			 54,	0)		\
 	x(btree_node_bkey_bad_format,				 55,	0)		\
 	x(btree_node_bad_bkey,					 56,	0)		\
-	x(btree_node_bkey_out_of_order,				 57,	0)		\
-	x(btree_root_bkey_invalid,				 58,	0)		\
-	x(btree_root_read_error,				 59,	0)		\
+	x(btree_node_bkey_out_of_order,				 57,	FSCK_AUTOFIX)	\
+	x(btree_root_bkey_invalid,				 58,	FSCK_AUTOFIX)	\
+	x(btree_root_read_error,				 59,	FSCK_AUTOFIX)	\
 	x(btree_root_bad_min_key,				 60,	0)		\
 	x(btree_root_bad_max_key,				 61,	0)		\
-	x(btree_node_read_error,				 62,	0)		\
-	x(btree_node_topology_bad_min_key,			 63,	0)		\
-	x(btree_node_topology_bad_max_key,			 64,	0)		\
-	x(btree_node_topology_overwritten_by_prev_node,		 65,	0)		\
-	x(btree_node_topology_overwritten_by_next_node,		 66,	0)		\
-	x(btree_node_topology_interior_node_empty,		 67,	0)		\
+	x(btree_node_read_error,				 62,	FSCK_AUTOFIX)	\
+	x(btree_node_topology_bad_min_key,			 63,	FSCK_AUTOFIX)	\
+	x(btree_node_topology_bad_max_key,			 64,	FSCK_AUTOFIX)	\
+	x(btree_node_topology_overwritten_by_prev_node,		 65,	FSCK_AUTOFIX)	\
+	x(btree_node_topology_overwritten_by_next_node,		 66,	FSCK_AUTOFIX)	\
+	x(btree_node_topology_interior_node_empty,		 67,	FSCK_AUTOFIX)	\
 	x(fs_usage_hidden_wrong,				 68,	FSCK_AUTOFIX)	\
 	x(fs_usage_btree_wrong,					 69,	FSCK_AUTOFIX)	\
 	x(fs_usage_data_wrong,					 70,	FSCK_AUTOFIX)	\
@@ -123,11 +122,12 @@ enum bch_fsck_flags {
 	x(alloc_key_cached_sectors_wrong,			109,	FSCK_AUTOFIX)	\
 	x(alloc_key_stripe_wrong,				110,	FSCK_AUTOFIX)	\
 	x(alloc_key_stripe_redundancy_wrong,			111,	FSCK_AUTOFIX)	\
+	x(alloc_key_journal_seq_in_future,			298,	FSCK_AUTOFIX)	\
 	x(bucket_sector_count_overflow,				112,	0)		\
 	x(bucket_metadata_type_mismatch,			113,	0)		\
-	x(need_discard_key_wrong,				114,	0)		\
-	x(freespace_key_wrong,					115,	0)		\
-	x(freespace_hole_missing,				116,	0)		\
+	x(need_discard_key_wrong,				114,	FSCK_AUTOFIX)	\
+	x(freespace_key_wrong,					115,	FSCK_AUTOFIX)	\
+	x(freespace_hole_missing,				116,	FSCK_AUTOFIX)	\
 	x(bucket_gens_val_size_bad,				117,	0)		\
 	x(bucket_gens_key_wrong,				118,	FSCK_AUTOFIX)	\
 	x(bucket_gens_hole_wrong,				119,	FSCK_AUTOFIX)	\
@@ -139,9 +139,10 @@ enum bch_fsck_flags {
 	x(discarding_bucket_not_in_need_discard_btree,		291,	0)		\
 	x(backpointer_bucket_offset_wrong,			125,	0)		\
 	x(backpointer_level_bad,				294,	0)		\
-	x(backpointer_to_missing_device,			126,	0)		\
-	x(backpointer_to_missing_alloc,				127,	0)		\
-	x(backpointer_to_missing_ptr,				128,	0)		\
+	x(backpointer_dev_bad,					297,	0)		\
+	x(backpointer_to_missing_device,			126,	FSCK_AUTOFIX)	\
+	x(backpointer_to_missing_alloc,				127,	FSCK_AUTOFIX)	\
+	x(backpointer_to_missing_ptr,				128,	FSCK_AUTOFIX)	\
 	x(lru_entry_at_time_0,					129,	FSCK_AUTOFIX)	\
 	x(lru_entry_to_invalid_bucket,				130,	FSCK_AUTOFIX)	\
 	x(lru_entry_bad,					131,	FSCK_AUTOFIX)	\
@@ -167,14 +168,15 @@ enum bch_fsck_flags {
 	x(ptr_to_incorrect_stripe,				151,	0)		\
 	x(ptr_gen_newer_than_bucket_gen,			152,	0)		\
 	x(ptr_too_stale,					153,	0)		\
-	x(stale_dirty_ptr,					154,	0)		\
+	x(stale_dirty_ptr,					154,	FSCK_AUTOFIX)	\
 	x(ptr_bucket_data_type_mismatch,			155,	0)		\
 	x(ptr_cached_and_erasure_coded,				156,	0)		\
 	x(ptr_crc_uncompressed_size_too_small,			157,	0)		\
+	x(ptr_crc_uncompressed_size_too_big,			161,	0)		\
+	x(ptr_crc_uncompressed_size_mismatch,			300,	0)		\
 	x(ptr_crc_csum_type_unknown,				158,	0)		\
 	x(ptr_crc_compression_type_unknown,			159,	0)		\
 	x(ptr_crc_redundant,					160,	0)		\
-	x(ptr_crc_uncompressed_size_too_big,			161,	0)		\
 	x(ptr_crc_nonce_mismatch,				162,	0)		\
 	x(ptr_stripe_redundant,					163,	0)		\
 	x(reservation_key_nr_replicas_invalid,			164,	0)		\
@@ -209,6 +211,7 @@ enum bch_fsck_flags {
 	x(bkey_in_missing_snapshot,				190,	0)		\
 	x(inode_pos_inode_nonzero,				191,	0)		\
 	x(inode_pos_blockdev_range,				192,	0)		\
+	x(inode_alloc_cursor_inode_bad,				301,	0)		\
 	x(inode_unpack_error,					193,	0)		\
 	x(inode_str_hash_invalid,				194,	0)		\
 	x(inode_v3_fields_start_bad,				195,	0)		\
@@ -232,6 +235,7 @@ enum bch_fsck_flags {
 	x(inode_wrong_nlink,					209,	FSCK_AUTOFIX)	\
 	x(inode_has_child_snapshots_wrong,			287,	0)		\
 	x(inode_unreachable,					210,	FSCK_AUTOFIX)	\
+	x(inode_journal_seq_in_future,				299,	FSCK_AUTOFIX)	\
 	x(deleted_inode_but_clean,				211,	FSCK_AUTOFIX)	\
 	x(deleted_inode_missing,				212,	FSCK_AUTOFIX)	\
 	x(deleted_inode_is_dir,					213,	FSCK_AUTOFIX)	\
@@ -252,6 +256,7 @@ enum bch_fsck_flags {
 	x(dirent_in_missing_dir_inode,				227,	0)		\
 	x(dirent_in_non_dir_inode,				228,	0)		\
 	x(dirent_to_missing_inode,				229,	0)		\
+	x(dirent_to_overwritten_inode,				302,	0)		\
 	x(dirent_to_missing_subvol,				230,	0)		\
 	x(dirent_to_itself,					231,	0)		\
 	x(quota_type_invalid,					232,	0)		\
@@ -288,7 +293,7 @@ enum bch_fsck_flags {
 	x(btree_root_unreadable_and_scan_found_nothing,		263,	0)		\
 	x(snapshot_node_missing,				264,	0)		\
 	x(dup_backpointer_to_bad_csum_extent,			265,	0)		\
-	x(btree_bitmap_not_marked,				266,	0)		\
+	x(btree_bitmap_not_marked,				266,	FSCK_AUTOFIX)	\
 	x(sb_clean_entry_overrun,				267,	0)		\
 	x(btree_ptr_v2_written_0,				268,	0)		\
 	x(subvol_snapshot_bad,					269,	0)		\
@@ -306,7 +311,10 @@ enum bch_fsck_flags {
 	x(accounting_key_replicas_devs_unsorted,		280,	FSCK_AUTOFIX)	\
 	x(accounting_key_version_0,				282,	FSCK_AUTOFIX)	\
 	x(logged_op_but_clean,					283,	FSCK_AUTOFIX)	\
-	x(MAX,							295,	0)
+	x(compression_opt_not_marked_in_sb,			295,	FSCK_AUTOFIX)	\
+	x(compression_type_not_marked_in_sb,			296,	FSCK_AUTOFIX)	\
+	x(directory_size_mismatch,				303,	FSCK_AUTOFIX)	\
+	x(MAX,							304,	0)
 
 enum bch_sb_error_id {
 #define x(t, n, ...) BCH_FSCK_ERR_##t = n,
diff --git a/fs/bcachefs/six.c b/fs/bcachefs/six.c
index 617d07e53b20..7e7c66a1e1a6 100644
--- a/fs/bcachefs/six.c
+++ b/fs/bcachefs/six.c
@@ -491,8 +491,12 @@ static int six_lock_slowpath(struct six_lock *lock, enum six_lock_type type,
 				list_del(&wait->list);
 			raw_spin_unlock(&lock->wait_lock);
 
-			if (unlikely(acquired))
+			if (unlikely(acquired)) {
 				do_six_unlock_type(lock, type);
+			} else if (type == SIX_LOCK_write) {
+				six_clear_bitmask(lock, SIX_LOCK_HELD_write);
+				six_lock_wakeup(lock, atomic_read(&lock->state), SIX_LOCK_read);
+			}
 			break;
 		}
 
@@ -501,10 +505,6 @@ static int six_lock_slowpath(struct six_lock *lock, enum six_lock_type type,
 
 	__set_current_state(TASK_RUNNING);
 out:
-	if (ret && type == SIX_LOCK_write) {
-		six_clear_bitmask(lock, SIX_LOCK_HELD_write);
-		six_lock_wakeup(lock, atomic_read(&lock->state), SIX_LOCK_read);
-	}
 	trace_contention_end(lock, 0);
 
 	return ret;
@@ -616,8 +616,6 @@ void six_unlock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long
 
 	if (type != SIX_LOCK_write)
 		six_release(&lock->dep_map, ip);
-	else
-		lock->seq++;
 
 	if (type == SIX_LOCK_intent &&
 	    lock->intent_lock_recurse) {
@@ -625,6 +623,15 @@ void six_unlock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long
 		return;
 	}
 
+	if (type == SIX_LOCK_write &&
+	    lock->write_lock_recurse) {
+		--lock->write_lock_recurse;
+		return;
+	}
+
+	if (type == SIX_LOCK_write)
+		lock->seq++;
+
 	do_six_unlock_type(lock, type);
 }
 EXPORT_SYMBOL_GPL(six_unlock_ip);
@@ -735,13 +742,13 @@ void six_lock_increment(struct six_lock *lock, enum six_lock_type type)
 			atomic_add(l[type].lock_val, &lock->state);
 		}
 		break;
+	case SIX_LOCK_write:
+		lock->write_lock_recurse++;
+		fallthrough;
 	case SIX_LOCK_intent:
 		EBUG_ON(!(atomic_read(&lock->state) & SIX_LOCK_HELD_intent));
 		lock->intent_lock_recurse++;
 		break;
-	case SIX_LOCK_write:
-		BUG();
-		break;
 	}
 }
 EXPORT_SYMBOL_GPL(six_lock_increment);
diff --git a/fs/bcachefs/six.h b/fs/bcachefs/six.h
index 68d46fd7f391..c142e06b7a3a 100644
--- a/fs/bcachefs/six.h
+++ b/fs/bcachefs/six.h
@@ -137,6 +137,7 @@ struct six_lock {
 	atomic_t		state;
 	u32			seq;
 	unsigned		intent_lock_recurse;
+	unsigned		write_lock_recurse;
 	struct task_struct	*owner;
 	unsigned __percpu	*readers;
 	raw_spinlock_t		wait_lock;
diff --git a/fs/bcachefs/snapshot.c b/fs/bcachefs/snapshot.c
index ae57638506c3..c54091a28909 100644
--- a/fs/bcachefs/snapshot.c
+++ b/fs/bcachefs/snapshot.c
@@ -2,6 +2,7 @@
 
 #include "bcachefs.h"
 #include "bkey_buf.h"
+#include "btree_cache.h"
 #include "btree_key_cache.h"
 #include "btree_update.h"
 #include "buckets.h"
@@ -32,7 +33,7 @@ void bch2_snapshot_tree_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 int bch2_snapshot_tree_validate(struct bch_fs *c, struct bkey_s_c k,
-			       enum bch_validate_flags flags)
+				struct bkey_validate_context from)
 {
 	int ret = 0;
 
@@ -225,7 +226,7 @@ void bch2_snapshot_to_text(struct printbuf *out, struct bch_fs *c,
 }
 
 int bch2_snapshot_validate(struct bch_fs *c, struct bkey_s_c k,
-			  enum bch_validate_flags flags)
+			   struct bkey_validate_context from)
 {
 	struct bkey_s_c_snapshot s;
 	u32 i, id;
@@ -279,23 +280,6 @@ fsck_err:
 	return ret;
 }
 
-static void __set_is_ancestor_bitmap(struct bch_fs *c, u32 id)
-{
-	struct snapshot_t *t = snapshot_t_mut(c, id);
-	u32 parent = id;
-
-	while ((parent = bch2_snapshot_parent_early(c, parent)) &&
-	       parent - id - 1 < IS_ANCESTOR_BITMAP)
-		__set_bit(parent - id - 1, t->is_ancestor);
-}
-
-static void set_is_ancestor_bitmap(struct bch_fs *c, u32 id)
-{
-	mutex_lock(&c->snapshot_table_lock);
-	__set_is_ancestor_bitmap(c, id);
-	mutex_unlock(&c->snapshot_table_lock);
-}
-
 static int __bch2_mark_snapshot(struct btree_trans *trans,
 		       enum btree_id btree, unsigned level,
 		       struct bkey_s_c old, struct bkey_s_c new,
@@ -317,6 +301,7 @@ static int __bch2_mark_snapshot(struct btree_trans *trans,
 	if (new.k->type == KEY_TYPE_snapshot) {
 		struct bkey_s_c_snapshot s = bkey_s_c_to_snapshot(new);
 
+		t->live		= true;
 		t->parent	= le32_to_cpu(s.v->parent);
 		t->children[0]	= le32_to_cpu(s.v->children[0]);
 		t->children[1]	= le32_to_cpu(s.v->children[1]);
@@ -335,7 +320,11 @@ static int __bch2_mark_snapshot(struct btree_trans *trans,
 			t->skip[2]	= 0;
 		}
 
-		__set_is_ancestor_bitmap(c, id);
+		u32 parent = id;
+
+		while ((parent = bch2_snapshot_parent_early(c, parent)) &&
+		       parent - id - 1 < IS_ANCESTOR_BITMAP)
+			__set_bit(parent - id - 1, t->is_ancestor);
 
 		if (BCH_SNAPSHOT_DELETED(s.v)) {
 			set_bit(BCH_FS_need_delete_dead_snapshots, &c->flags);
@@ -365,70 +354,6 @@ int bch2_snapshot_lookup(struct btree_trans *trans, u32 id,
 				       BTREE_ITER_with_updates, snapshot, s);
 }
 
-static int bch2_snapshot_live(struct btree_trans *trans, u32 id)
-{
-	struct bch_snapshot v;
-	int ret;
-
-	if (!id)
-		return 0;
-
-	ret = bch2_snapshot_lookup(trans, id, &v);
-	if (bch2_err_matches(ret, ENOENT))
-		bch_err(trans->c, "snapshot node %u not found", id);
-	if (ret)
-		return ret;
-
-	return !BCH_SNAPSHOT_DELETED(&v);
-}
-
-/*
- * If @k is a snapshot with just one live child, it's part of a linear chain,
- * which we consider to be an equivalence class: and then after snapshot
- * deletion cleanup, there should only be a single key at a given position in
- * this equivalence class.
- *
- * This sets the equivalence class of @k to be the child's equivalence class, if
- * it's part of such a linear chain: this correctly sets equivalence classes on
- * startup if we run leaf to root (i.e. in natural key order).
- */
-static int bch2_snapshot_set_equiv(struct btree_trans *trans, struct bkey_s_c k)
-{
-	struct bch_fs *c = trans->c;
-	unsigned i, nr_live = 0, live_idx = 0;
-	struct bkey_s_c_snapshot snap;
-	u32 id = k.k->p.offset, child[2];
-
-	if (k.k->type != KEY_TYPE_snapshot)
-		return 0;
-
-	snap = bkey_s_c_to_snapshot(k);
-
-	child[0] = le32_to_cpu(snap.v->children[0]);
-	child[1] = le32_to_cpu(snap.v->children[1]);
-
-	for (i = 0; i < 2; i++) {
-		int ret = bch2_snapshot_live(trans, child[i]);
-
-		if (ret < 0)
-			return ret;
-
-		if (ret)
-			live_idx = i;
-		nr_live += ret;
-	}
-
-	mutex_lock(&c->snapshot_table_lock);
-
-	snapshot_t_mut(c, id)->equiv = nr_live == 1
-		? snapshot_t_mut(c, child[live_idx])->equiv
-		: id;
-
-	mutex_unlock(&c->snapshot_table_lock);
-
-	return 0;
-}
-
 /* fsck: */
 
 static u32 bch2_snapshot_child(struct bch_fs *c, u32 id, unsigned child)
@@ -506,7 +431,6 @@ static int bch2_snapshot_tree_master_subvol(struct btree_trans *trans,
 			break;
 		}
 	}
-
 	bch2_trans_iter_exit(trans, &iter);
 
 	if (!ret && !found) {
@@ -536,6 +460,7 @@ static int check_snapshot_tree(struct btree_trans *trans,
 	struct bch_snapshot s;
 	struct bch_subvolume subvol;
 	struct printbuf buf = PRINTBUF;
+	struct btree_iter snapshot_iter = {};
 	u32 root_id;
 	int ret;
 
@@ -545,22 +470,35 @@ static int check_snapshot_tree(struct btree_trans *trans,
 	st = bkey_s_c_to_snapshot_tree(k);
 	root_id = le32_to_cpu(st.v->root_snapshot);
 
-	ret = bch2_snapshot_lookup(trans, root_id, &s);
+	struct bkey_s_c_snapshot snapshot_k =
+		bch2_bkey_get_iter_typed(trans, &snapshot_iter, BTREE_ID_snapshots,
+					 POS(0, root_id), 0, snapshot);
+	ret = bkey_err(snapshot_k);
 	if (ret && !bch2_err_matches(ret, ENOENT))
 		goto err;
 
+	if (!ret)
+		bkey_val_copy(&s, snapshot_k);
+
 	if (fsck_err_on(ret ||
 			root_id != bch2_snapshot_root(c, root_id) ||
 			st.k->p.offset != le32_to_cpu(s.tree),
 			trans, snapshot_tree_to_missing_snapshot,
 			"snapshot tree points to missing/incorrect snapshot:\n  %s",
-			(bch2_bkey_val_to_text(&buf, c, st.s_c), buf.buf))) {
+			(bch2_bkey_val_to_text(&buf, c, st.s_c),
+			 prt_newline(&buf),
+			 ret
+			 ? prt_printf(&buf, "(%s)", bch2_err_str(ret))
+			 : bch2_bkey_val_to_text(&buf, c, snapshot_k.s_c),
+			 buf.buf))) {
 		ret = bch2_btree_delete_at(trans, iter, 0);
 		goto err;
 	}
 
-	ret = bch2_subvolume_get(trans, le32_to_cpu(st.v->master_subvol),
-				 false, 0, &subvol);
+	if (!st.v->master_subvol)
+		goto out;
+
+	ret = bch2_subvolume_get(trans, le32_to_cpu(st.v->master_subvol), false, &subvol);
 	if (ret && !bch2_err_matches(ret, ENOENT))
 		goto err;
 
@@ -603,8 +541,10 @@ static int check_snapshot_tree(struct btree_trans *trans,
 		u->v.master_subvol = cpu_to_le32(subvol_id);
 		st = snapshot_tree_i_to_s_c(u);
 	}
+out:
 err:
 fsck_err:
+	bch2_trans_iter_exit(trans, &snapshot_iter);
 	printbuf_exit(&buf);
 	return ret;
 }
@@ -799,7 +739,7 @@ static int check_snapshot(struct btree_trans *trans,
 
 	if (should_have_subvol) {
 		id = le32_to_cpu(s.subvol);
-		ret = bch2_subvolume_get(trans, id, 0, false, &subvol);
+		ret = bch2_subvolume_get(trans, id, false, &subvol);
 		if (bch2_err_matches(ret, ENOENT))
 			bch_err(c, "snapshot points to nonexistent subvolume:\n  %s",
 				(bch2_bkey_val_to_text(&buf, c, k), buf.buf));
@@ -902,7 +842,7 @@ static int check_snapshot_exists(struct btree_trans *trans, u32 id)
 {
 	struct bch_fs *c = trans->c;
 
-	if (bch2_snapshot_equiv(c, id))
+	if (bch2_snapshot_exists(c, id))
 		return 0;
 
 	/* Do we need to reconstruct the snapshot_tree entry as well? */
@@ -951,8 +891,7 @@ static int check_snapshot_exists(struct btree_trans *trans, u32 id)
 
 	return  bch2_btree_insert_trans(trans, BTREE_ID_snapshots, &snapshot->k_i, 0) ?:
 		bch2_mark_snapshot(trans, BTREE_ID_snapshots, 0,
-				   bkey_s_c_null, bkey_i_to_s(&snapshot->k_i), 0) ?:
-		bch2_snapshot_set_equiv(trans, bkey_i_to_s_c(&snapshot->k_i));
+				   bkey_s_c_null, bkey_i_to_s(&snapshot->k_i), 0);
 }
 
 /* Figure out which snapshot nodes belong in the same tree: */
@@ -1050,7 +989,7 @@ int bch2_reconstruct_snapshots(struct bch_fs *c)
 		snapshot_id_list_to_text(&buf, t);
 
 		darray_for_each(*t, id) {
-			if (fsck_err_on(!bch2_snapshot_equiv(c, *id),
+			if (fsck_err_on(!bch2_snapshot_exists(c, *id),
 					trans, snapshot_node_missing,
 					"snapshot node %u from tree %s missing, recreate?", *id, buf.buf)) {
 				if (t->nr > 1) {
@@ -1083,10 +1022,12 @@ int bch2_check_key_has_snapshot(struct btree_trans *trans,
 	struct printbuf buf = PRINTBUF;
 	int ret = 0;
 
-	if (fsck_err_on(!bch2_snapshot_equiv(c, k.k->p.snapshot),
+	if (fsck_err_on(!bch2_snapshot_exists(c, k.k->p.snapshot),
 			trans, bkey_in_missing_snapshot,
 			"key in missing snapshot %s, delete?",
-			(bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
+			(bch2_btree_id_to_text(&buf, iter->btree_id),
+			 prt_char(&buf, ' '),
+			 bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
 		ret = bch2_btree_delete_at(trans, iter,
 					    BTREE_UPDATE_internal_snapshot_node) ?: 1;
 fsck_err:
@@ -1100,13 +1041,11 @@ fsck_err:
 int bch2_snapshot_node_set_deleted(struct btree_trans *trans, u32 id)
 {
 	struct btree_iter iter;
-	struct bkey_i_snapshot *s;
-	int ret = 0;
-
-	s = bch2_bkey_get_mut_typed(trans, &iter,
+	struct bkey_i_snapshot *s =
+		bch2_bkey_get_mut_typed(trans, &iter,
 				    BTREE_ID_snapshots, POS(0, id),
 				    0, snapshot);
-	ret = PTR_ERR_OR_ZERO(s);
+	int ret = PTR_ERR_OR_ZERO(s);
 	if (unlikely(ret)) {
 		bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT),
 					trans->c, "missing snapshot %u", id);
@@ -1294,10 +1233,6 @@ static int create_snapids(struct btree_trans *trans, u32 parent, u32 tree,
 			goto err;
 
 		new_snapids[i]	= iter.pos.offset;
-
-		mutex_lock(&c->snapshot_table_lock);
-		snapshot_t_mut(c, new_snapids[i])->equiv = new_snapids[i];
-		mutex_unlock(&c->snapshot_table_lock);
 	}
 err:
 	bch2_trans_iter_exit(trans, &iter);
@@ -1403,129 +1338,153 @@ int bch2_snapshot_node_create(struct btree_trans *trans, u32 parent,
  * that key to snapshot leaf nodes, where we can mutate it
  */
 
-static int delete_dead_snapshots_process_key(struct btree_trans *trans,
-			       struct btree_iter *iter,
-			       struct bkey_s_c k,
-			       snapshot_id_list *deleted,
-			       snapshot_id_list *equiv_seen,
-			       struct bpos *last_pos)
+struct snapshot_interior_delete {
+	u32	id;
+	u32	live_child;
+};
+typedef DARRAY(struct snapshot_interior_delete) interior_delete_list;
+
+static inline u32 interior_delete_has_id(interior_delete_list *l, u32 id)
 {
-	int ret = bch2_check_key_has_snapshot(trans, iter, k);
-	if (ret)
-		return ret < 0 ? ret : 0;
+	darray_for_each(*l, i)
+		if (i->id == id)
+			return i->live_child;
+	return 0;
+}
 
-	struct bch_fs *c = trans->c;
-	u32 equiv = bch2_snapshot_equiv(c, k.k->p.snapshot);
-	if (!equiv) /* key for invalid snapshot node, but we chose not to delete */
+static unsigned __live_child(struct snapshot_table *t, u32 id,
+			     snapshot_id_list *delete_leaves,
+			     interior_delete_list *delete_interior)
+{
+	struct snapshot_t *s = __snapshot_t(t, id);
+	if (!s)
 		return 0;
 
-	if (!bkey_eq(k.k->p, *last_pos))
-		equiv_seen->nr = 0;
+	for (unsigned i = 0; i < ARRAY_SIZE(s->children); i++)
+		if (s->children[i] &&
+		    !snapshot_list_has_id(delete_leaves, s->children[i]) &&
+		    !interior_delete_has_id(delete_interior, s->children[i]))
+			return s->children[i];
 
-	if (snapshot_list_has_id(deleted, k.k->p.snapshot))
-		return bch2_btree_delete_at(trans, iter,
-					    BTREE_UPDATE_internal_snapshot_node);
+	for (unsigned i = 0; i < ARRAY_SIZE(s->children); i++) {
+		u32 live_child = s->children[i]
+			? __live_child(t, s->children[i], delete_leaves, delete_interior)
+			: 0;
+		if (live_child)
+			return live_child;
+	}
 
-	if (!bpos_eq(*last_pos, k.k->p) &&
-	    snapshot_list_has_id(equiv_seen, equiv))
-		return bch2_btree_delete_at(trans, iter,
-					    BTREE_UPDATE_internal_snapshot_node);
+	return 0;
+}
 
-	*last_pos = k.k->p;
+static unsigned live_child(struct bch_fs *c, u32 id,
+			   snapshot_id_list *delete_leaves,
+			   interior_delete_list *delete_interior)
+{
+	rcu_read_lock();
+	u32 ret = __live_child(rcu_dereference(c->snapshots), id,
+			       delete_leaves, delete_interior);
+	rcu_read_unlock();
+	return ret;
+}
 
-	ret = snapshot_list_add_nodup(c, equiv_seen, equiv);
-	if (ret)
-		return ret;
+static int delete_dead_snapshots_process_key(struct btree_trans *trans,
+					     struct btree_iter *iter,
+					     struct bkey_s_c k,
+					     snapshot_id_list *delete_leaves,
+					     interior_delete_list *delete_interior)
+{
+	if (snapshot_list_has_id(delete_leaves, k.k->p.snapshot))
+		return bch2_btree_delete_at(trans, iter,
+					    BTREE_UPDATE_internal_snapshot_node);
 
-	/*
-	 * When we have a linear chain of snapshot nodes, we consider
-	 * those to form an equivalence class: we're going to collapse
-	 * them all down to a single node, and keep the leaf-most node -
-	 * which has the same id as the equivalence class id.
-	 *
-	 * If there are multiple keys in different snapshots at the same
-	 * position, we're only going to keep the one in the newest
-	 * snapshot (we delete the others above) - the rest have been
-	 * overwritten and are redundant, and for the key we're going to keep we
-	 * need to move it to the equivalance class ID if it's not there
-	 * already.
-	 */
-	if (equiv != k.k->p.snapshot) {
+	u32 live_child = interior_delete_has_id(delete_interior, k.k->p.snapshot);
+	if (live_child) {
 		struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
 		int ret = PTR_ERR_OR_ZERO(new);
 		if (ret)
 			return ret;
 
-		new->k.p.snapshot = equiv;
-
-		struct btree_iter new_iter;
-		bch2_trans_iter_init(trans, &new_iter, iter->btree_id, new->k.p,
-				     BTREE_ITER_all_snapshots|
-				     BTREE_ITER_cached|
-				     BTREE_ITER_intent);
+		new->k.p.snapshot = live_child;
 
-		ret =   bch2_btree_iter_traverse(&new_iter) ?:
-			bch2_trans_update(trans, &new_iter, new,
-					BTREE_UPDATE_internal_snapshot_node) ?:
-			bch2_btree_delete_at(trans, iter,
-					BTREE_UPDATE_internal_snapshot_node);
-		bch2_trans_iter_exit(trans, &new_iter);
+		struct btree_iter dst_iter;
+		struct bkey_s_c dst_k = bch2_bkey_get_iter(trans, &dst_iter,
+							   iter->btree_id, new->k.p,
+							   BTREE_ITER_all_snapshots|
+							   BTREE_ITER_intent);
+		ret = bkey_err(dst_k);
 		if (ret)
 			return ret;
+
+		ret =   (bkey_deleted(dst_k.k)
+			 ? bch2_trans_update(trans, &dst_iter, new,
+					     BTREE_UPDATE_internal_snapshot_node)
+			 : 0) ?:
+			bch2_btree_delete_at(trans, iter,
+					     BTREE_UPDATE_internal_snapshot_node);
+		bch2_trans_iter_exit(trans, &dst_iter);
+		return ret;
 	}
 
 	return 0;
 }
 
-static int bch2_snapshot_needs_delete(struct btree_trans *trans, struct bkey_s_c k)
+/*
+ * For a given snapshot, if it doesn't have a subvolume that points to it, and
+ * it doesn't have child snapshot nodes - it's now redundant and we can mark it
+ * as deleted.
+ */
+static int check_should_delete_snapshot(struct btree_trans *trans, struct bkey_s_c k,
+					snapshot_id_list *delete_leaves,
+					interior_delete_list *delete_interior)
 {
-	struct bkey_s_c_snapshot snap;
-	u32 children[2];
-	int ret;
-
 	if (k.k->type != KEY_TYPE_snapshot)
 		return 0;
 
-	snap = bkey_s_c_to_snapshot(k);
-	if (BCH_SNAPSHOT_DELETED(snap.v) ||
-	    BCH_SNAPSHOT_SUBVOL(snap.v))
+	struct bch_fs *c = trans->c;
+	struct bkey_s_c_snapshot s = bkey_s_c_to_snapshot(k);
+	unsigned live_children = 0;
+
+	if (BCH_SNAPSHOT_SUBVOL(s.v))
 		return 0;
 
-	children[0] = le32_to_cpu(snap.v->children[0]);
-	children[1] = le32_to_cpu(snap.v->children[1]);
+	for (unsigned i = 0; i < 2; i++) {
+		u32 child = le32_to_cpu(s.v->children[i]);
 
-	ret   = bch2_snapshot_live(trans, children[0]) ?:
-		bch2_snapshot_live(trans, children[1]);
-	if (ret < 0)
-		return ret;
-	return !ret;
-}
+		live_children += child &&
+			!snapshot_list_has_id(delete_leaves, child);
+	}
 
-/*
- * For a given snapshot, if it doesn't have a subvolume that points to it, and
- * it doesn't have child snapshot nodes - it's now redundant and we can mark it
- * as deleted.
- */
-static int bch2_delete_redundant_snapshot(struct btree_trans *trans, struct bkey_s_c k)
-{
-	int ret = bch2_snapshot_needs_delete(trans, k);
+	if (live_children == 0) {
+		return snapshot_list_add(c, delete_leaves, s.k->p.offset);
+	} else if (live_children == 1) {
+		struct snapshot_interior_delete d = {
+			.id		= s.k->p.offset,
+			.live_child	= live_child(c, s.k->p.offset, delete_leaves, delete_interior),
+		};
+
+		if (!d.live_child) {
+			bch_err(c, "error finding live child of snapshot %u", d.id);
+			return -EINVAL;
+		}
 
-	return ret <= 0
-		? ret
-		: bch2_snapshot_node_set_deleted(trans, k.k->p.offset);
+		return darray_push(delete_interior, d);
+	} else {
+		return 0;
+	}
 }
 
 static inline u32 bch2_snapshot_nth_parent_skip(struct bch_fs *c, u32 id, u32 n,
-						snapshot_id_list *skip)
+						interior_delete_list *skip)
 {
 	rcu_read_lock();
-	while (snapshot_list_has_id(skip, id))
+	while (interior_delete_has_id(skip, id))
 		id = __bch2_snapshot_parent(c, id);
 
 	while (n--) {
 		do {
 			id = __bch2_snapshot_parent(c, id);
-		} while (snapshot_list_has_id(skip, id));
+		} while (interior_delete_has_id(skip, id));
 	}
 	rcu_read_unlock();
 
@@ -1534,7 +1493,7 @@ static inline u32 bch2_snapshot_nth_parent_skip(struct bch_fs *c, u32 id, u32 n,
 
 static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
 					      struct btree_iter *iter, struct bkey_s_c k,
-					      snapshot_id_list *deleted)
+					      interior_delete_list *deleted)
 {
 	struct bch_fs *c = trans->c;
 	u32 nr_deleted_ancestors = 0;
@@ -1544,7 +1503,7 @@ static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
 	if (k.k->type != KEY_TYPE_snapshot)
 		return 0;
 
-	if (snapshot_list_has_id(deleted, k.k->p.offset))
+	if (interior_delete_has_id(deleted, k.k->p.offset))
 		return 0;
 
 	s = bch2_bkey_make_mut_noupdate_typed(trans, k, snapshot);
@@ -1553,7 +1512,7 @@ static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
 		return ret;
 
 	darray_for_each(*deleted, i)
-		nr_deleted_ancestors += bch2_snapshot_is_ancestor(c, s->k.p.offset, *i);
+		nr_deleted_ancestors += bch2_snapshot_is_ancestor(c, s->k.p.offset, i->id);
 
 	if (!nr_deleted_ancestors)
 		return 0;
@@ -1571,7 +1530,7 @@ static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
 		for (unsigned j = 0; j < ARRAY_SIZE(s->v.skip); j++) {
 			u32 id = le32_to_cpu(s->v.skip[j]);
 
-			if (snapshot_list_has_id(deleted, id)) {
+			if (interior_delete_has_id(deleted, id)) {
 				id = bch2_snapshot_nth_parent_skip(c,
 							parent,
 							depth > 1
@@ -1590,51 +1549,45 @@ static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
 
 int bch2_delete_dead_snapshots(struct bch_fs *c)
 {
-	struct btree_trans *trans;
-	snapshot_id_list deleted = { 0 };
-	snapshot_id_list deleted_interior = { 0 };
-	int ret = 0;
-
 	if (!test_and_clear_bit(BCH_FS_need_delete_dead_snapshots, &c->flags))
 		return 0;
 
-	trans = bch2_trans_get(c);
+	struct btree_trans *trans = bch2_trans_get(c);
+	snapshot_id_list delete_leaves = {};
+	interior_delete_list delete_interior = {};
+	int ret = 0;
 
 	/*
 	 * For every snapshot node: If we have no live children and it's not
 	 * pointed to by a subvolume, delete it:
 	 */
-	ret = for_each_btree_key_commit(trans, iter, BTREE_ID_snapshots,
-			POS_MIN, 0, k,
-			NULL, NULL, 0,
-		bch2_delete_redundant_snapshot(trans, k));
-	bch_err_msg(c, ret, "deleting redundant snapshots");
+	ret = for_each_btree_key(trans, iter, BTREE_ID_snapshots, POS_MIN, 0, k,
+		check_should_delete_snapshot(trans, k, &delete_leaves, &delete_interior));
+	if (!bch2_err_matches(ret, EROFS))
+		bch_err_msg(c, ret, "walking snapshots");
 	if (ret)
 		goto err;
 
-	ret = for_each_btree_key(trans, iter, BTREE_ID_snapshots,
-				 POS_MIN, 0, k,
-		bch2_snapshot_set_equiv(trans, k));
-	bch_err_msg(c, ret, "in bch2_snapshots_set_equiv");
-	if (ret)
+	if (!delete_leaves.nr && !delete_interior.nr)
 		goto err;
 
-	ret = for_each_btree_key(trans, iter, BTREE_ID_snapshots,
-				 POS_MIN, 0, k, ({
-		if (k.k->type != KEY_TYPE_snapshot)
-			continue;
+	{
+		struct printbuf buf = PRINTBUF;
+		prt_printf(&buf, "deleting leaves");
+		darray_for_each(delete_leaves, i)
+			prt_printf(&buf, " %u", *i);
 
-		BCH_SNAPSHOT_DELETED(bkey_s_c_to_snapshot(k).v)
-			? snapshot_list_add(c, &deleted, k.k->p.offset)
-			: 0;
-	}));
-	bch_err_msg(c, ret, "walking snapshots");
-	if (ret)
-		goto err;
+		prt_printf(&buf, " interior");
+		darray_for_each(delete_interior, i)
+			prt_printf(&buf, " %u->%u", i->id, i->live_child);
+
+		ret = commit_do(trans, NULL, NULL, 0, bch2_trans_log_msg(trans, &buf));
+		printbuf_exit(&buf);
+		if (ret)
+			goto err;
+	}
 
 	for (unsigned btree = 0; btree < BTREE_ID_NR; btree++) {
-		struct bpos last_pos = POS_MIN;
-		snapshot_id_list equiv_seen = { 0 };
 		struct disk_reservation res = { 0 };
 
 		if (!btree_type_has_snapshots(btree))
@@ -1644,33 +1597,26 @@ int bch2_delete_dead_snapshots(struct bch_fs *c)
 				btree, POS_MIN,
 				BTREE_ITER_prefetch|BTREE_ITER_all_snapshots, k,
 				&res, NULL, BCH_TRANS_COMMIT_no_enospc,
-			delete_dead_snapshots_process_key(trans, &iter, k, &deleted,
-							  &equiv_seen, &last_pos));
+			delete_dead_snapshots_process_key(trans, &iter, k,
+							  &delete_leaves,
+							  &delete_interior));
 
 		bch2_disk_reservation_put(c, &res);
-		darray_exit(&equiv_seen);
 
-		bch_err_msg(c, ret, "deleting keys from dying snapshots");
+		if (!bch2_err_matches(ret, EROFS))
+			bch_err_msg(c, ret, "deleting keys from dying snapshots");
 		if (ret)
 			goto err;
 	}
 
-	bch2_trans_unlock(trans);
-	down_write(&c->snapshot_create_lock);
-
-	ret = for_each_btree_key(trans, iter, BTREE_ID_snapshots,
-				 POS_MIN, 0, k, ({
-		u32 snapshot = k.k->p.offset;
-		u32 equiv = bch2_snapshot_equiv(c, snapshot);
-
-		equiv != snapshot
-			? snapshot_list_add(c, &deleted_interior, snapshot)
-			: 0;
-	}));
-
-	bch_err_msg(c, ret, "walking snapshots");
-	if (ret)
-		goto err_create_lock;
+	darray_for_each(delete_leaves, i) {
+		ret = commit_do(trans, NULL, NULL, 0,
+			bch2_snapshot_node_delete(trans, *i));
+		if (!bch2_err_matches(ret, EROFS))
+			bch_err_msg(c, ret, "deleting snapshot %u", *i);
+		if (ret)
+			goto err;
+	}
 
 	/*
 	 * Fixing children of deleted snapshots can't be done completely
@@ -1680,32 +1626,24 @@ int bch2_delete_dead_snapshots(struct bch_fs *c)
 	ret = for_each_btree_key_commit(trans, iter, BTREE_ID_snapshots, POS_MIN,
 				  BTREE_ITER_intent, k,
 				  NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
-		bch2_fix_child_of_deleted_snapshot(trans, &iter, k, &deleted_interior));
+		bch2_fix_child_of_deleted_snapshot(trans, &iter, k, &delete_interior));
 	if (ret)
-		goto err_create_lock;
-
-	darray_for_each(deleted, i) {
-		ret = commit_do(trans, NULL, NULL, 0,
-			bch2_snapshot_node_delete(trans, *i));
-		bch_err_msg(c, ret, "deleting snapshot %u", *i);
-		if (ret)
-			goto err_create_lock;
-	}
+		goto err;
 
-	darray_for_each(deleted_interior, i) {
+	darray_for_each(delete_interior, i) {
 		ret = commit_do(trans, NULL, NULL, 0,
-			bch2_snapshot_node_delete(trans, *i));
-		bch_err_msg(c, ret, "deleting snapshot %u", *i);
+			bch2_snapshot_node_delete(trans, i->id));
+		if (!bch2_err_matches(ret, EROFS))
+			bch_err_msg(c, ret, "deleting snapshot %u", i->id);
 		if (ret)
-			goto err_create_lock;
+			goto err;
 	}
-err_create_lock:
-	up_write(&c->snapshot_create_lock);
 err:
-	darray_exit(&deleted_interior);
-	darray_exit(&deleted);
+	darray_exit(&delete_interior);
+	darray_exit(&delete_leaves);
 	bch2_trans_put(trans);
-	bch_err_fn(c, ret);
+	if (!bch2_err_matches(ret, EROFS))
+		bch_err_fn(c, ret);
 	return ret;
 }
 
@@ -1721,8 +1659,12 @@ void bch2_delete_dead_snapshots_work(struct work_struct *work)
 
 void bch2_delete_dead_snapshots_async(struct bch_fs *c)
 {
-	if (bch2_write_ref_tryget(c, BCH_WRITE_REF_delete_dead_snapshots) &&
-	    !queue_work(c->write_ref_wq, &c->snapshot_delete_work))
+	if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_delete_dead_snapshots))
+		return;
+
+	BUG_ON(!test_bit(BCH_FS_may_go_rw, &c->flags));
+
+	if (!queue_work(c->write_ref_wq, &c->snapshot_delete_work))
 		bch2_write_ref_put(c, BCH_WRITE_REF_delete_dead_snapshots);
 }
 
@@ -1735,18 +1677,10 @@ int __bch2_key_has_snapshot_overwrites(struct btree_trans *trans,
 	struct bkey_s_c k;
 	int ret;
 
-	bch2_trans_iter_init(trans, &iter, id, pos,
-			     BTREE_ITER_not_extents|
-			     BTREE_ITER_all_snapshots);
-	while (1) {
-		k = bch2_btree_iter_prev(&iter);
-		ret = bkey_err(k);
-		if (ret)
-			break;
-
-		if (!k.k)
-			break;
-
+	for_each_btree_key_reverse_norestart(trans, iter, id, bpos_predecessor(pos),
+					     BTREE_ITER_not_extents|
+					     BTREE_ITER_all_snapshots,
+					     k, ret) {
 		if (!bkey_eq(pos, k.k->p))
 			break;
 
@@ -1760,37 +1694,36 @@ int __bch2_key_has_snapshot_overwrites(struct btree_trans *trans,
 	return ret;
 }
 
-static int bch2_check_snapshot_needs_deletion(struct btree_trans *trans, struct bkey_s_c k)
+static bool interior_snapshot_needs_delete(struct bkey_s_c_snapshot snap)
 {
-	struct bch_fs *c = trans->c;
-	struct bkey_s_c_snapshot snap;
-	int ret = 0;
+	/* If there's one child, it's redundant and keys will be moved to the child */
+	return !!snap.v->children[0] + !!snap.v->children[1] == 1;
+}
 
+static int bch2_check_snapshot_needs_deletion(struct btree_trans *trans, struct bkey_s_c k)
+{
 	if (k.k->type != KEY_TYPE_snapshot)
 		return 0;
 
-	snap = bkey_s_c_to_snapshot(k);
+	struct bkey_s_c_snapshot snap = bkey_s_c_to_snapshot(k);
 	if (BCH_SNAPSHOT_DELETED(snap.v) ||
-	    bch2_snapshot_equiv(c, k.k->p.offset) != k.k->p.offset ||
-	    (ret = bch2_snapshot_needs_delete(trans, k)) > 0) {
-		set_bit(BCH_FS_need_delete_dead_snapshots, &c->flags);
-		return 0;
-	}
+	    interior_snapshot_needs_delete(snap))
+		set_bit(BCH_FS_need_delete_dead_snapshots, &trans->c->flags);
 
-	return ret;
+	return 0;
 }
 
 int bch2_snapshots_read(struct bch_fs *c)
 {
+	/*
+	 * Initializing the is_ancestor bitmaps requires ancestors to already be
+	 * initialized - so mark in reverse:
+	 */
 	int ret = bch2_trans_run(c,
-		for_each_btree_key(trans, iter, BTREE_ID_snapshots,
-				   POS_MIN, 0, k,
+		for_each_btree_key_reverse(trans, iter, BTREE_ID_snapshots,
+				   POS_MAX, 0, k,
 			__bch2_mark_snapshot(trans, BTREE_ID_snapshots, 0, bkey_s_c_null, k, 0) ?:
-			bch2_snapshot_set_equiv(trans, k) ?:
-			bch2_check_snapshot_needs_deletion(trans, k)) ?:
-		for_each_btree_key(trans, iter, BTREE_ID_snapshots,
-				   POS_MIN, 0, k,
-			   (set_is_ancestor_bitmap(c, k.k->p.offset), 0)));
+			bch2_check_snapshot_needs_deletion(trans, k)));
 	bch_err_fn(c, ret);
 
 	/*
diff --git a/fs/bcachefs/snapshot.h b/fs/bcachefs/snapshot.h
index 29c94716293e..00373cf32e7b 100644
--- a/fs/bcachefs/snapshot.h
+++ b/fs/bcachefs/snapshot.h
@@ -2,11 +2,9 @@
 #ifndef _BCACHEFS_SNAPSHOT_H
 #define _BCACHEFS_SNAPSHOT_H
 
-enum bch_validate_flags;
-
 void bch2_snapshot_tree_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 int bch2_snapshot_tree_validate(struct bch_fs *, struct bkey_s_c,
-			       enum bch_validate_flags);
+				struct bkey_validate_context);
 
 #define bch2_bkey_ops_snapshot_tree ((struct bkey_ops) {	\
 	.key_validate	= bch2_snapshot_tree_validate,		\
@@ -19,7 +17,8 @@ struct bkey_i_snapshot_tree *__bch2_snapshot_tree_create(struct btree_trans *);
 int bch2_snapshot_tree_lookup(struct btree_trans *, u32, struct bch_snapshot_tree *);
 
 void bch2_snapshot_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
-int bch2_snapshot_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_snapshot_validate(struct bch_fs *, struct bkey_s_c,
+			   struct bkey_validate_context);
 int bch2_mark_snapshot(struct btree_trans *, enum btree_id, unsigned,
 		       struct bkey_s_c, struct bkey_s,
 		       enum btree_iter_update_trigger_flags);
@@ -120,19 +119,19 @@ static inline u32 bch2_snapshot_root(struct bch_fs *c, u32 id)
 	return id;
 }
 
-static inline u32 __bch2_snapshot_equiv(struct bch_fs *c, u32 id)
+static inline bool __bch2_snapshot_exists(struct bch_fs *c, u32 id)
 {
 	const struct snapshot_t *s = snapshot_t(c, id);
-	return s ? s->equiv : 0;
+	return s ? s->live : 0;
 }
 
-static inline u32 bch2_snapshot_equiv(struct bch_fs *c, u32 id)
+static inline bool bch2_snapshot_exists(struct bch_fs *c, u32 id)
 {
 	rcu_read_lock();
-	id = __bch2_snapshot_equiv(c, id);
+	bool ret = __bch2_snapshot_exists(c, id);
 	rcu_read_unlock();
 
-	return id;
+	return ret;
 }
 
 static inline int bch2_snapshot_is_internal_node(struct bch_fs *c, u32 id)
diff --git a/fs/bcachefs/str_hash.c b/fs/bcachefs/str_hash.c
new file mode 100644
index 000000000000..8c2c5539de2e
--- /dev/null
+++ b/fs/bcachefs/str_hash.c
@@ -0,0 +1,295 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "btree_cache.h"
+#include "btree_update.h"
+#include "dirent.h"
+#include "fsck.h"
+#include "str_hash.h"
+#include "subvolume.h"
+
+static int bch2_dirent_has_target(struct btree_trans *trans, struct bkey_s_c_dirent d)
+{
+	if (d.v->d_type == DT_SUBVOL) {
+		struct bch_subvolume subvol;
+		int ret = bch2_subvolume_get(trans, le32_to_cpu(d.v->d_child_subvol),
+					     false, &subvol);
+		if (ret && !bch2_err_matches(ret, ENOENT))
+			return ret;
+		return !ret;
+	} else {
+		struct btree_iter iter;
+		struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
+				SPOS(0, le64_to_cpu(d.v->d_inum), d.k->p.snapshot), 0);
+		int ret = bkey_err(k);
+		if (ret)
+			return ret;
+
+		ret = bkey_is_inode(k.k);
+		bch2_trans_iter_exit(trans, &iter);
+		return ret;
+	}
+}
+
+static int fsck_rename_dirent(struct btree_trans *trans,
+			      struct snapshots_seen *s,
+			      const struct bch_hash_desc desc,
+			      struct bch_hash_info *hash_info,
+			      struct bkey_s_c_dirent old)
+{
+	struct qstr old_name = bch2_dirent_get_name(old);
+	struct bkey_i_dirent *new = bch2_trans_kmalloc(trans, bkey_bytes(old.k) + 32);
+	int ret = PTR_ERR_OR_ZERO(new);
+	if (ret)
+		return ret;
+
+	bkey_dirent_init(&new->k_i);
+	dirent_copy_target(new, old);
+	new->k.p = old.k->p;
+
+	for (unsigned i = 0; i < 1000; i++) {
+		unsigned len = sprintf(new->v.d_name, "%.*s.fsck_renamed-%u",
+				       old_name.len, old_name.name, i);
+		unsigned u64s = BKEY_U64s + dirent_val_u64s(len);
+
+		if (u64s > U8_MAX)
+			return -EINVAL;
+
+		new->k.u64s = u64s;
+
+		ret = bch2_hash_set_in_snapshot(trans, bch2_dirent_hash_desc, hash_info,
+						(subvol_inum) { 0, old.k->p.inode },
+						old.k->p.snapshot, &new->k_i,
+						BTREE_UPDATE_internal_snapshot_node);
+		if (!bch2_err_matches(ret, EEXIST))
+			break;
+	}
+
+	if (ret)
+		return ret;
+
+	return bch2_fsck_update_backpointers(trans, s, desc, hash_info, &new->k_i);
+}
+
+static int hash_pick_winner(struct btree_trans *trans,
+			    const struct bch_hash_desc desc,
+			    struct bch_hash_info *hash_info,
+			    struct bkey_s_c k1,
+			    struct bkey_s_c k2)
+{
+	if (bkey_val_bytes(k1.k) == bkey_val_bytes(k2.k) &&
+	    !memcmp(k1.v, k2.v, bkey_val_bytes(k1.k)))
+		return 0;
+
+	switch (desc.btree_id) {
+	case BTREE_ID_dirents: {
+		int ret = bch2_dirent_has_target(trans, bkey_s_c_to_dirent(k1));
+		if (ret < 0)
+			return ret;
+		if (!ret)
+			return 0;
+
+		ret = bch2_dirent_has_target(trans, bkey_s_c_to_dirent(k2));
+		if (ret < 0)
+			return ret;
+		if (!ret)
+			return 1;
+		return 2;
+	}
+	default:
+		return 0;
+	}
+}
+
+static int repair_inode_hash_info(struct btree_trans *trans,
+				  struct bch_inode_unpacked *snapshot_root)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	for_each_btree_key_reverse_norestart(trans, iter, BTREE_ID_inodes,
+					     SPOS(0, snapshot_root->bi_inum, snapshot_root->bi_snapshot - 1),
+					     BTREE_ITER_all_snapshots, k, ret) {
+		if (k.k->p.offset != snapshot_root->bi_inum)
+			break;
+		if (!bkey_is_inode(k.k))
+			continue;
+
+		struct bch_inode_unpacked inode;
+		ret = bch2_inode_unpack(k, &inode);
+		if (ret)
+			break;
+
+		if (fsck_err_on(inode.bi_hash_seed	!= snapshot_root->bi_hash_seed ||
+				INODE_STR_HASH(&inode)	!= INODE_STR_HASH(snapshot_root),
+				trans, inode_snapshot_mismatch,
+				"inode hash info in different snapshots don't match")) {
+			inode.bi_hash_seed = snapshot_root->bi_hash_seed;
+			SET_INODE_STR_HASH(&inode, INODE_STR_HASH(snapshot_root));
+			ret = __bch2_fsck_write_inode(trans, &inode) ?:
+				bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc) ?:
+				-BCH_ERR_transaction_restart_nested;
+			break;
+		}
+	}
+fsck_err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+/*
+ * All versions of the same inode in different snapshots must have the same hash
+ * seed/type: verify that the hash info we're using matches the root
+ */
+static int check_inode_hash_info_matches_root(struct btree_trans *trans, u64 inum,
+					      struct bch_hash_info *hash_info)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	for_each_btree_key_reverse_norestart(trans, iter, BTREE_ID_inodes, SPOS(0, inum, U32_MAX),
+					     BTREE_ITER_all_snapshots, k, ret) {
+		if (k.k->p.offset != inum)
+			break;
+		if (bkey_is_inode(k.k))
+			goto found;
+	}
+	bch_err(c, "%s(): inum %llu not found", __func__, inum);
+	ret = -BCH_ERR_fsck_repair_unimplemented;
+	goto err;
+found:;
+	struct bch_inode_unpacked inode;
+	ret = bch2_inode_unpack(k, &inode);
+	if (ret)
+		goto err;
+
+	struct bch_hash_info hash2 = bch2_hash_info_init(c, &inode);
+	if (hash_info->type != hash2.type ||
+	    memcmp(&hash_info->siphash_key, &hash2.siphash_key, sizeof(hash2.siphash_key))) {
+		ret = repair_inode_hash_info(trans, &inode);
+		if (!ret) {
+			bch_err(c, "inode hash info mismatch with root, but mismatch not found\n"
+				"%u %llx %llx\n"
+				"%u %llx %llx",
+				hash_info->type,
+				hash_info->siphash_key.k0,
+				hash_info->siphash_key.k1,
+				hash2.type,
+				hash2.siphash_key.k0,
+				hash2.siphash_key.k1);
+			ret = -BCH_ERR_fsck_repair_unimplemented;
+		}
+	}
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+int __bch2_str_hash_check_key(struct btree_trans *trans,
+			      struct snapshots_seen *s,
+			      const struct bch_hash_desc *desc,
+			      struct bch_hash_info *hash_info,
+			      struct btree_iter *k_iter, struct bkey_s_c hash_k)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter = { NULL };
+	struct printbuf buf = PRINTBUF;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	u64 hash = desc->hash_bkey(hash_info, hash_k);
+	if (hash_k.k->p.offset < hash)
+		goto bad_hash;
+
+	for_each_btree_key_norestart(trans, iter, desc->btree_id,
+				     SPOS(hash_k.k->p.inode, hash, hash_k.k->p.snapshot),
+				     BTREE_ITER_slots, k, ret) {
+		if (bkey_eq(k.k->p, hash_k.k->p))
+			break;
+
+		if (k.k->type == desc->key_type &&
+		    !desc->cmp_bkey(k, hash_k))
+			goto duplicate_entries;
+
+		if (bkey_deleted(k.k)) {
+			bch2_trans_iter_exit(trans, &iter);
+			goto bad_hash;
+		}
+	}
+out:
+	bch2_trans_iter_exit(trans, &iter);
+	printbuf_exit(&buf);
+	return ret;
+bad_hash:
+	/*
+	 * Before doing any repair, check hash_info itself:
+	 */
+	ret = check_inode_hash_info_matches_root(trans, hash_k.k->p.inode, hash_info);
+	if (ret)
+		goto out;
+
+	if (fsck_err(trans, hash_table_key_wrong_offset,
+		     "hash table key at wrong offset: btree %s inode %llu offset %llu, hashed to %llu\n  %s",
+		     bch2_btree_id_str(desc->btree_id), hash_k.k->p.inode, hash_k.k->p.offset, hash,
+		     (printbuf_reset(&buf),
+		      bch2_bkey_val_to_text(&buf, c, hash_k), buf.buf))) {
+		struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, hash_k);
+		if (IS_ERR(new))
+			return PTR_ERR(new);
+
+		k = bch2_hash_set_or_get_in_snapshot(trans, &iter, *desc, hash_info,
+				       (subvol_inum) { 0, hash_k.k->p.inode },
+				       hash_k.k->p.snapshot, new,
+				       STR_HASH_must_create|
+				       BTREE_ITER_with_updates|
+				       BTREE_UPDATE_internal_snapshot_node);
+		ret = bkey_err(k);
+		if (ret)
+			goto out;
+		if (k.k)
+			goto duplicate_entries;
+
+		ret =   bch2_hash_delete_at(trans, *desc, hash_info, k_iter,
+					    BTREE_UPDATE_internal_snapshot_node) ?:
+			bch2_fsck_update_backpointers(trans, s, *desc, hash_info, new) ?:
+			bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc) ?:
+			-BCH_ERR_transaction_restart_nested;
+		goto out;
+	}
+fsck_err:
+	goto out;
+duplicate_entries:
+	ret = hash_pick_winner(trans, *desc, hash_info, hash_k, k);
+	if (ret < 0)
+		goto out;
+
+	if (!fsck_err(trans, hash_table_key_duplicate,
+		      "duplicate hash table keys%s:\n%s",
+		      ret != 2 ? "" : ", both point to valid inodes",
+		      (printbuf_reset(&buf),
+		       bch2_bkey_val_to_text(&buf, c, hash_k),
+		       prt_newline(&buf),
+		       bch2_bkey_val_to_text(&buf, c, k),
+		       buf.buf)))
+		goto out;
+
+	switch (ret) {
+	case 0:
+		ret = bch2_hash_delete_at(trans, *desc, hash_info, k_iter, 0);
+		break;
+	case 1:
+		ret = bch2_hash_delete_at(trans, *desc, hash_info, &iter, 0);
+		break;
+	case 2:
+		ret = fsck_rename_dirent(trans, s, *desc, hash_info, bkey_s_c_to_dirent(hash_k)) ?:
+			bch2_hash_delete_at(trans, *desc, hash_info, k_iter, 0);
+		goto out;
+	}
+
+	ret = bch2_trans_commit(trans, NULL, NULL, 0) ?:
+		-BCH_ERR_transaction_restart_nested;
+	goto out;
+}
diff --git a/fs/bcachefs/str_hash.h b/fs/bcachefs/str_hash.h
index ec2b1feea520..55a4ac7bf220 100644
--- a/fs/bcachefs/str_hash.h
+++ b/fs/bcachefs/str_hash.h
@@ -160,7 +160,7 @@ bch2_hash_lookup_in_snapshot(struct btree_trans *trans,
 	struct bkey_s_c k;
 	int ret;
 
-	for_each_btree_key_upto_norestart(trans, *iter, desc.btree_id,
+	for_each_btree_key_max_norestart(trans, *iter, desc.btree_id,
 			   SPOS(inum.inum, desc.hash_key(info, key), snapshot),
 			   POS(inum.inum, U64_MAX),
 			   BTREE_ITER_slots|flags, k, ret) {
@@ -210,7 +210,7 @@ bch2_hash_hole(struct btree_trans *trans,
 	if (ret)
 		return ret;
 
-	for_each_btree_key_upto_norestart(trans, *iter, desc.btree_id,
+	for_each_btree_key_max_norestart(trans, *iter, desc.btree_id,
 			   SPOS(inum.inum, desc.hash_key(info, key), snapshot),
 			   POS(inum.inum, U64_MAX),
 			   BTREE_ITER_slots|BTREE_ITER_intent, k, ret)
@@ -265,7 +265,7 @@ struct bkey_s_c bch2_hash_set_or_get_in_snapshot(struct btree_trans *trans,
 	bool found = false;
 	int ret;
 
-	for_each_btree_key_upto_norestart(trans, *iter, desc.btree_id,
+	for_each_btree_key_max_norestart(trans, *iter, desc.btree_id,
 			   SPOS(insert->k.p.inode,
 				desc.hash_bkey(info, bkey_i_to_s_c(insert)),
 				snapshot),
@@ -393,4 +393,26 @@ int bch2_hash_delete(struct btree_trans *trans,
 	return ret;
 }
 
+struct snapshots_seen;
+int __bch2_str_hash_check_key(struct btree_trans *,
+			      struct snapshots_seen *,
+			      const struct bch_hash_desc *,
+			      struct bch_hash_info *,
+			      struct btree_iter *, struct bkey_s_c);
+
+static inline int bch2_str_hash_check_key(struct btree_trans *trans,
+			    struct snapshots_seen *s,
+			    const struct bch_hash_desc *desc,
+			    struct bch_hash_info *hash_info,
+			    struct btree_iter *k_iter, struct bkey_s_c hash_k)
+{
+	if (hash_k.k->type != desc->key_type)
+		return 0;
+
+	if (likely(desc->hash_bkey(hash_info, hash_k) == hash_k.k->p.offset))
+		return 0;
+
+	return __bch2_str_hash_check_key(trans, s, desc, hash_info, k_iter, hash_k);
+}
+
 #endif /* _BCACHEFS_STR_HASH_H */
diff --git a/fs/bcachefs/subvolume.c b/fs/bcachefs/subvolume.c
index 80e5efaff524..e3d0475232e5 100644
--- a/fs/bcachefs/subvolume.c
+++ b/fs/bcachefs/subvolume.c
@@ -207,7 +207,7 @@ int bch2_check_subvol_children(struct bch_fs *c)
 /* Subvolumes: */
 
 int bch2_subvolume_validate(struct bch_fs *c, struct bkey_s_c k,
-			   enum bch_validate_flags flags)
+			    struct bkey_validate_context from)
 {
 	struct bkey_s_c_subvolume subvol = bkey_s_c_to_subvolume(k);
 	int ret = 0;
@@ -286,11 +286,11 @@ int bch2_subvol_has_children(struct btree_trans *trans, u32 subvol)
 static __always_inline int
 bch2_subvolume_get_inlined(struct btree_trans *trans, unsigned subvol,
 			   bool inconsistent_if_not_found,
-			   int iter_flags,
 			   struct bch_subvolume *s)
 {
 	int ret = bch2_bkey_get_val_typed(trans, BTREE_ID_subvolumes, POS(0, subvol),
-					  iter_flags, subvolume, s);
+					  BTREE_ITER_cached|
+					  BTREE_ITER_with_updates, subvolume, s);
 	bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT) &&
 				inconsistent_if_not_found,
 				trans->c, "missing subvolume %u", subvol);
@@ -299,16 +299,15 @@ bch2_subvolume_get_inlined(struct btree_trans *trans, unsigned subvol,
 
 int bch2_subvolume_get(struct btree_trans *trans, unsigned subvol,
 		       bool inconsistent_if_not_found,
-		       int iter_flags,
 		       struct bch_subvolume *s)
 {
-	return bch2_subvolume_get_inlined(trans, subvol, inconsistent_if_not_found, iter_flags, s);
+	return bch2_subvolume_get_inlined(trans, subvol, inconsistent_if_not_found, s);
 }
 
 int bch2_subvol_is_ro_trans(struct btree_trans *trans, u32 subvol)
 {
 	struct bch_subvolume s;
-	int ret = bch2_subvolume_get_inlined(trans, subvol, true, 0, &s);
+	int ret = bch2_subvolume_get_inlined(trans, subvol, true, &s);
 	if (ret)
 		return ret;
 
@@ -328,7 +327,7 @@ int bch2_snapshot_get_subvol(struct btree_trans *trans, u32 snapshot,
 	struct bch_snapshot snap;
 
 	return  bch2_snapshot_lookup(trans, snapshot, &snap) ?:
-		bch2_subvolume_get(trans, le32_to_cpu(snap.subvol), true, 0, subvol);
+		bch2_subvolume_get(trans, le32_to_cpu(snap.subvol), true, subvol);
 }
 
 int __bch2_subvolume_get_snapshot(struct btree_trans *trans, u32 subvolid,
@@ -396,8 +395,7 @@ static int bch2_subvolumes_reparent(struct btree_trans *trans, u32 subvolid_to_d
 	struct bch_subvolume s;
 
 	return lockrestart_do(trans,
-			bch2_subvolume_get(trans, subvolid_to_delete, true,
-				   BTREE_ITER_cached, &s)) ?:
+			bch2_subvolume_get(trans, subvolid_to_delete, true, &s)) ?:
 		for_each_btree_key_commit(trans, iter,
 				BTREE_ID_subvolumes, POS_MIN, BTREE_ITER_prefetch, k,
 				NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
@@ -411,26 +409,56 @@ static int bch2_subvolumes_reparent(struct btree_trans *trans, u32 subvolid_to_d
  */
 static int __bch2_subvolume_delete(struct btree_trans *trans, u32 subvolid)
 {
-	struct btree_iter iter;
-	struct bkey_s_c_subvolume subvol;
-	u32 snapid;
-	int ret = 0;
+	struct btree_iter subvol_iter = {}, snapshot_iter = {}, snapshot_tree_iter = {};
 
-	subvol = bch2_bkey_get_iter_typed(trans, &iter,
+	struct bkey_s_c_subvolume subvol =
+		bch2_bkey_get_iter_typed(trans, &subvol_iter,
 				BTREE_ID_subvolumes, POS(0, subvolid),
 				BTREE_ITER_cached|BTREE_ITER_intent,
 				subvolume);
-	ret = bkey_err(subvol);
+	int ret = bkey_err(subvol);
 	bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
 				"missing subvolume %u", subvolid);
 	if (ret)
-		return ret;
+		goto err;
 
-	snapid = le32_to_cpu(subvol.v->snapshot);
+	u32 snapid = le32_to_cpu(subvol.v->snapshot);
+
+	struct bkey_s_c_snapshot snapshot =
+		bch2_bkey_get_iter_typed(trans, &snapshot_iter,
+				BTREE_ID_snapshots, POS(0, snapid),
+				0, snapshot);
+	ret = bkey_err(subvol);
+	bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
+				"missing snapshot %u", snapid);
+	if (ret)
+		goto err;
+
+	u32 treeid = le32_to_cpu(snapshot.v->tree);
 
-	ret =   bch2_btree_delete_at(trans, &iter, 0) ?:
+	struct bkey_s_c_snapshot_tree snapshot_tree =
+		bch2_bkey_get_iter_typed(trans, &snapshot_tree_iter,
+				BTREE_ID_snapshot_trees, POS(0, treeid),
+				0, snapshot_tree);
+
+	if (le32_to_cpu(snapshot_tree.v->master_subvol) == subvolid) {
+		struct bkey_i_snapshot_tree *snapshot_tree_mut =
+			bch2_bkey_make_mut_typed(trans, &snapshot_tree_iter,
+						 &snapshot_tree.s_c,
+						 0, snapshot_tree);
+		ret = PTR_ERR_OR_ZERO(snapshot_tree_mut);
+		if (ret)
+			goto err;
+
+		snapshot_tree_mut->v.master_subvol = 0;
+	}
+
+	ret =   bch2_btree_delete_at(trans, &subvol_iter, 0) ?:
 		bch2_snapshot_node_set_deleted(trans, snapid);
-	bch2_trans_iter_exit(trans, &iter);
+err:
+	bch2_trans_iter_exit(trans, &snapshot_tree_iter);
+	bch2_trans_iter_exit(trans, &snapshot_iter);
+	bch2_trans_iter_exit(trans, &subvol_iter);
 	return ret;
 }
 
@@ -675,7 +703,7 @@ err:
 /* set bi_subvol on root inode */
 int bch2_fs_upgrade_for_subvolumes(struct bch_fs *c)
 {
-	int ret = bch2_trans_commit_do(c, NULL, NULL, BCH_TRANS_COMMIT_lazy_rw,
+	int ret = bch2_trans_commit_do(c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 				       __bch2_fs_upgrade_for_subvolumes(trans));
 	bch_err_fn(c, ret);
 	return ret;
diff --git a/fs/bcachefs/subvolume.h b/fs/bcachefs/subvolume.h
index f897d106e142..910f6196700e 100644
--- a/fs/bcachefs/subvolume.h
+++ b/fs/bcachefs/subvolume.h
@@ -5,12 +5,11 @@
 #include "darray.h"
 #include "subvolume_types.h"
 
-enum bch_validate_flags;
-
 int bch2_check_subvols(struct bch_fs *);
 int bch2_check_subvol_children(struct bch_fs *);
 
-int bch2_subvolume_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_subvolume_validate(struct bch_fs *, struct bkey_s_c,
+			    struct bkey_validate_context);
 void bch2_subvolume_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 int bch2_subvolume_trigger(struct btree_trans *, enum btree_id, unsigned,
 			   struct bkey_s_c, struct bkey_s,
@@ -25,7 +24,7 @@ int bch2_subvolume_trigger(struct btree_trans *, enum btree_id, unsigned,
 
 int bch2_subvol_has_children(struct btree_trans *, u32);
 int bch2_subvolume_get(struct btree_trans *, unsigned,
-		       bool, int, struct bch_subvolume *);
+		       bool, struct bch_subvolume *);
 int __bch2_subvolume_get_snapshot(struct btree_trans *, u32,
 				  u32 *, bool);
 int bch2_subvolume_get_snapshot(struct btree_trans *, u32, u32 *);
@@ -34,7 +33,7 @@ int bch2_subvol_is_ro_trans(struct btree_trans *, u32);
 int bch2_subvol_is_ro(struct bch_fs *, u32);
 
 static inline struct bkey_s_c
-bch2_btree_iter_peek_in_subvolume_upto_type(struct btree_iter *iter, struct bpos end,
+bch2_btree_iter_peek_in_subvolume_max_type(struct btree_iter *iter, struct bpos end,
 					    u32 subvolid, unsigned flags)
 {
 	u32 snapshot;
@@ -43,10 +42,10 @@ bch2_btree_iter_peek_in_subvolume_upto_type(struct btree_iter *iter, struct bpos
 		return bkey_s_c_err(ret);
 
 	bch2_btree_iter_set_snapshot(iter, snapshot);
-	return bch2_btree_iter_peek_upto_type(iter, end, flags);
+	return bch2_btree_iter_peek_max_type(iter, end, flags);
 }
 
-#define for_each_btree_key_in_subvolume_upto_continue(_trans, _iter,		\
+#define for_each_btree_key_in_subvolume_max_continue(_trans, _iter,		\
 					 _end, _subvolid, _flags, _k, _do)	\
 ({										\
 	struct bkey_s_c _k;							\
@@ -54,7 +53,7 @@ bch2_btree_iter_peek_in_subvolume_upto_type(struct btree_iter *iter, struct bpos
 										\
 	do {									\
 		_ret3 = lockrestart_do(_trans, ({				\
-			(_k) = bch2_btree_iter_peek_in_subvolume_upto_type(&(_iter),	\
+			(_k) = bch2_btree_iter_peek_in_subvolume_max_type(&(_iter),	\
 						_end, _subvolid, (_flags));	\
 			if (!(_k).k)						\
 				break;						\
@@ -67,14 +66,14 @@ bch2_btree_iter_peek_in_subvolume_upto_type(struct btree_iter *iter, struct bpos
 	_ret3;									\
 })
 
-#define for_each_btree_key_in_subvolume_upto(_trans, _iter, _btree_id,		\
+#define for_each_btree_key_in_subvolume_max(_trans, _iter, _btree_id,		\
 				_start, _end, _subvolid, _flags, _k, _do)	\
 ({										\
 	struct btree_iter _iter;						\
 	bch2_trans_iter_init((_trans), &(_iter), (_btree_id),			\
 			     (_start), (_flags));				\
 										\
-	for_each_btree_key_in_subvolume_upto_continue(_trans, _iter,		\
+	for_each_btree_key_in_subvolume_max_continue(_trans, _iter,		\
 					_end, _subvolid, _flags, _k, _do);	\
 })
 
diff --git a/fs/bcachefs/subvolume_types.h b/fs/bcachefs/subvolume_types.h
index f2ec4277c2a5..1549d6daf7af 100644
--- a/fs/bcachefs/subvolume_types.h
+++ b/fs/bcachefs/subvolume_types.h
@@ -9,13 +9,13 @@ typedef DARRAY(u32) snapshot_id_list;
 #define IS_ANCESTOR_BITMAP	128
 
 struct snapshot_t {
+	bool			live;
 	u32			parent;
 	u32			skip[3];
 	u32			depth;
 	u32			children[2];
 	u32			subvol; /* Nonzero only if a subvolume points to this node: */
 	u32			tree;
-	u32			equiv;
 	unsigned long		is_ancestor[BITS_TO_LONGS(IS_ANCESTOR_BITMAP)];
 };
 
diff --git a/fs/bcachefs/super-io.c b/fs/bcachefs/super-io.c
index 7c71594f6a8b..8037ccbacf6a 100644
--- a/fs/bcachefs/super-io.c
+++ b/fs/bcachefs/super-io.c
@@ -23,6 +23,7 @@
 
 #include <linux/backing-dev.h>
 #include <linux/sort.h>
+#include <linux/string_choices.h>
 
 static const struct blk_holder_ops bch2_sb_handle_bdev_ops = {
 };
@@ -41,7 +42,7 @@ static const struct bch2_metadata_version bch2_metadata_versions[] = {
 #undef x
 };
 
-void bch2_version_to_text(struct printbuf *out, unsigned v)
+void bch2_version_to_text(struct printbuf *out, enum bcachefs_metadata_version v)
 {
 	const char *str = "(unknown version)";
 
@@ -54,7 +55,7 @@ void bch2_version_to_text(struct printbuf *out, unsigned v)
 	prt_printf(out, "%u.%u: %s", BCH_VERSION_MAJOR(v), BCH_VERSION_MINOR(v), str);
 }
 
-unsigned bch2_latest_compatible_version(unsigned v)
+enum bcachefs_metadata_version bch2_latest_compatible_version(enum bcachefs_metadata_version v)
 {
 	if (!BCH_VERSION_MAJOR(v))
 		return v;
@@ -68,6 +69,16 @@ unsigned bch2_latest_compatible_version(unsigned v)
 	return v;
 }
 
+void bch2_set_version_incompat(struct bch_fs *c, enum bcachefs_metadata_version version)
+{
+	mutex_lock(&c->sb_lock);
+	SET_BCH_SB_VERSION_INCOMPAT(c->disk_sb.sb,
+		max(BCH_SB_VERSION_INCOMPAT(c->disk_sb.sb), version));
+	c->disk_sb.sb->features[0] |= cpu_to_le64(BCH_FEATURE_incompat_version_field);
+	bch2_write_super(c);
+	mutex_unlock(&c->sb_lock);
+}
+
 const char * const bch2_sb_fields[] = {
 #define x(name, nr)	#name,
 	BCH_SB_FIELDS()
@@ -368,6 +379,12 @@ static int bch2_sb_validate(struct bch_sb_handle *disk_sb,
 		return -BCH_ERR_invalid_sb_features;
 	}
 
+	if (BCH_VERSION_MAJOR(le16_to_cpu(sb->version)) > BCH_VERSION_MAJOR(bcachefs_metadata_version_current) ||
+	    BCH_SB_VERSION_INCOMPAT(sb) > bcachefs_metadata_version_current) {
+		prt_printf(out, "Filesystem has incompatible version");
+		return -BCH_ERR_invalid_sb_features;
+	}
+
 	block_size = le16_to_cpu(sb->block_size);
 
 	if (block_size > PAGE_SECTORS) {
@@ -406,6 +423,21 @@ static int bch2_sb_validate(struct bch_sb_handle *disk_sb,
 		return -BCH_ERR_invalid_sb_time_precision;
 	}
 
+	/* old versions didn't know to downgrade this field */
+	if (BCH_SB_VERSION_INCOMPAT_ALLOWED(sb) > le16_to_cpu(sb->version))
+		SET_BCH_SB_VERSION_INCOMPAT_ALLOWED(sb, le16_to_cpu(sb->version));
+
+	if (BCH_SB_VERSION_INCOMPAT(sb) > BCH_SB_VERSION_INCOMPAT_ALLOWED(sb)) {
+		prt_printf(out, "Invalid version_incompat ");
+		bch2_version_to_text(out, BCH_SB_VERSION_INCOMPAT(sb));
+		prt_str(out, " > incompat_allowed ");
+		bch2_version_to_text(out, BCH_SB_VERSION_INCOMPAT_ALLOWED(sb));
+		if (flags & BCH_VALIDATE_write)
+			return -BCH_ERR_invalid_sb_version;
+		else
+			SET_BCH_SB_VERSION_INCOMPAT_ALLOWED(sb, BCH_SB_VERSION_INCOMPAT(sb));
+	}
+
 	if (!flags) {
 		/*
 		 * Been seeing a bug where these are getting inexplicably
@@ -428,6 +460,11 @@ static int bch2_sb_validate(struct bch_sb_handle *disk_sb,
 			SET_BCH_SB_PROMOTE_WHOLE_EXTENTS(sb, true);
 	}
 
+#ifdef __KERNEL__
+	if (!BCH_SB_SHARD_INUMS_NBITS(sb))
+		SET_BCH_SB_SHARD_INUMS_NBITS(sb, ilog2(roundup_pow_of_two(num_online_cpus())));
+#endif
+
 	for (opt_id = 0; opt_id < bch2_opts_nr; opt_id++) {
 		const struct bch_option *opt = bch2_opt_table + opt_id;
 
@@ -519,6 +556,9 @@ static void bch2_sb_update(struct bch_fs *c)
 	c->sb.uuid		= src->uuid;
 	c->sb.user_uuid		= src->user_uuid;
 	c->sb.version		= le16_to_cpu(src->version);
+	c->sb.version_incompat	= BCH_SB_VERSION_INCOMPAT(src);
+	c->sb.version_incompat_allowed
+				= BCH_SB_VERSION_INCOMPAT_ALLOWED(src);
 	c->sb.version_min	= le16_to_cpu(src->version_min);
 	c->sb.version_upgrade_complete = BCH_SB_VERSION_UPGRADE_COMPLETE(src);
 	c->sb.nr_devices	= src->nr_devices;
@@ -676,7 +716,8 @@ reread:
 	}
 
 	enum bch_csum_type csum_type = BCH_SB_CSUM_TYPE(sb->sb);
-	if (csum_type >= BCH_CSUM_NR) {
+	if (csum_type >= BCH_CSUM_NR ||
+	    bch2_csum_type_is_encryption(csum_type)) {
 		prt_printf(err, "unknown checksum type %llu", BCH_SB_CSUM_TYPE(sb->sb));
 		return -BCH_ERR_invalid_sb_csum_type;
 	}
@@ -878,7 +919,7 @@ static void write_super_endio(struct bio *bio)
 			       ? BCH_MEMBER_ERROR_write
 			       : BCH_MEMBER_ERROR_read,
 			       "superblock %s error: %s",
-			       bio_data_dir(bio) ? "write" : "read",
+			       str_write_read(bio_data_dir(bio)),
 			       bch2_blk_status_to_str(bio->bi_status)))
 		ca->sb_write_error = 1;
 
@@ -891,14 +932,15 @@ static void read_back_super(struct bch_fs *c, struct bch_dev *ca)
 	struct bch_sb *sb = ca->disk_sb.sb;
 	struct bio *bio = ca->disk_sb.bio;
 
+	memset(ca->sb_read_scratch, 0, BCH_SB_READ_SCRATCH_BUF_SIZE);
+
 	bio_reset(bio, ca->disk_sb.bdev, REQ_OP_READ|REQ_SYNC|REQ_META);
 	bio->bi_iter.bi_sector	= le64_to_cpu(sb->layout.sb_offset[0]);
 	bio->bi_end_io		= write_super_endio;
 	bio->bi_private		= ca;
-	bch2_bio_map(bio, ca->sb_read_scratch, PAGE_SIZE);
+	bch2_bio_map(bio, ca->sb_read_scratch, BCH_SB_READ_SCRATCH_BUF_SIZE);
 
-	this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_sb],
-		     bio_sectors(bio));
+	this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_sb], bio_sectors(bio));
 
 	percpu_ref_get(&ca->io_ref);
 	closure_bio_submit(bio, &c->sb_write);
@@ -1042,9 +1084,16 @@ int bch2_write_super(struct bch_fs *c)
 				": Superblock write was silently dropped! (seq %llu expected %llu)",
 				le64_to_cpu(ca->sb_read_scratch->seq),
 				ca->disk_sb.seq);
-			bch2_fs_fatal_error(c, "%s", buf.buf);
+
+			if (c->opts.errors != BCH_ON_ERROR_continue &&
+			    c->opts.errors != BCH_ON_ERROR_fix_safe) {
+				ret = -BCH_ERR_erofs_sb_err;
+				bch2_fs_fatal_error(c, "%s", buf.buf);
+			} else {
+				bch_err(c, "%s", buf.buf);
+			}
+
 			printbuf_exit(&buf);
-			ret = -BCH_ERR_erofs_sb_err;
 		}
 
 		if (le64_to_cpu(ca->sb_read_scratch->seq) > ca->disk_sb.seq) {
@@ -1149,6 +1198,8 @@ bool bch2_check_version_downgrade(struct bch_fs *c)
 	 */
 	if (BCH_SB_VERSION_UPGRADE_COMPLETE(c->disk_sb.sb) > bcachefs_metadata_version_current)
 		SET_BCH_SB_VERSION_UPGRADE_COMPLETE(c->disk_sb.sb, bcachefs_metadata_version_current);
+	if (BCH_SB_VERSION_INCOMPAT_ALLOWED(c->disk_sb.sb) > bcachefs_metadata_version_current)
+		SET_BCH_SB_VERSION_INCOMPAT_ALLOWED(c->disk_sb.sb, bcachefs_metadata_version_current);
 	if (c->sb.version > bcachefs_metadata_version_current)
 		c->disk_sb.sb->version = cpu_to_le16(bcachefs_metadata_version_current);
 	if (c->sb.version_min > bcachefs_metadata_version_current)
@@ -1157,7 +1208,7 @@ bool bch2_check_version_downgrade(struct bch_fs *c)
 	return ret;
 }
 
-void bch2_sb_upgrade(struct bch_fs *c, unsigned new_version)
+void bch2_sb_upgrade(struct bch_fs *c, unsigned new_version, bool incompat)
 {
 	lockdep_assert_held(&c->sb_lock);
 
@@ -1167,6 +1218,10 @@ void bch2_sb_upgrade(struct bch_fs *c, unsigned new_version)
 
 	c->disk_sb.sb->version = cpu_to_le16(new_version);
 	c->disk_sb.sb->features[0] |= cpu_to_le64(BCH_SB_FEATURES_ALL);
+
+	if (incompat)
+		SET_BCH_SB_VERSION_INCOMPAT_ALLOWED(c->disk_sb.sb,
+			max(BCH_SB_VERSION_INCOMPAT_ALLOWED(c->disk_sb.sb), new_version));
 }
 
 static int bch2_sb_ext_validate(struct bch_sb *sb, struct bch_sb_field *f,
@@ -1331,6 +1386,14 @@ void bch2_sb_to_text(struct printbuf *out, struct bch_sb *sb,
 	bch2_version_to_text(out, le16_to_cpu(sb->version));
 	prt_newline(out);
 
+	prt_printf(out, "Incompatible features allowed:\t");
+	bch2_version_to_text(out, BCH_SB_VERSION_INCOMPAT_ALLOWED(sb));
+	prt_newline(out);
+
+	prt_printf(out, "Incompatible features in use:\t");
+	bch2_version_to_text(out, BCH_SB_VERSION_INCOMPAT(sb));
+	prt_newline(out);
+
 	prt_printf(out, "Version upgrade complete:\t");
 	bch2_version_to_text(out, BCH_SB_VERSION_UPGRADE_COMPLETE(sb));
 	prt_newline(out);
diff --git a/fs/bcachefs/super-io.h b/fs/bcachefs/super-io.h
index fadd364e2802..f1ab4f943720 100644
--- a/fs/bcachefs/super-io.h
+++ b/fs/bcachefs/super-io.h
@@ -10,14 +10,29 @@
 
 #include <asm/byteorder.h>
 
+#define BCH_SB_READ_SCRATCH_BUF_SIZE		4096
+
 static inline bool bch2_version_compatible(u16 version)
 {
 	return BCH_VERSION_MAJOR(version) <= BCH_VERSION_MAJOR(bcachefs_metadata_version_current) &&
 		version >= bcachefs_metadata_version_min;
 }
 
-void bch2_version_to_text(struct printbuf *, unsigned);
-unsigned bch2_latest_compatible_version(unsigned);
+void bch2_version_to_text(struct printbuf *, enum bcachefs_metadata_version);
+enum bcachefs_metadata_version bch2_latest_compatible_version(enum bcachefs_metadata_version);
+
+void bch2_set_version_incompat(struct bch_fs *, enum bcachefs_metadata_version);
+
+static inline bool bch2_request_incompat_feature(struct bch_fs *c,
+						 enum bcachefs_metadata_version version)
+{
+	if (unlikely(version > c->sb.version_incompat)) {
+		if (version > c->sb.version_incompat_allowed)
+			return false;
+		bch2_set_version_incompat(c, version);
+	}
+	return true;
+}
 
 static inline size_t bch2_sb_field_bytes(struct bch_sb_field *f)
 {
@@ -92,7 +107,7 @@ static inline void bch2_check_set_feature(struct bch_fs *c, unsigned feat)
 }
 
 bool bch2_check_version_downgrade(struct bch_fs *);
-void bch2_sb_upgrade(struct bch_fs *, unsigned);
+void bch2_sb_upgrade(struct bch_fs *, unsigned, bool);
 
 void __bch2_sb_field_to_text(struct printbuf *, struct bch_sb *,
 			     struct bch_sb_field *);
diff --git a/fs/bcachefs/super.c b/fs/bcachefs/super.c
index a6ed9a0bf1c7..d97ea7bd1171 100644
--- a/fs/bcachefs/super.c
+++ b/fs/bcachefs/super.c
@@ -290,7 +290,7 @@ static void __bch2_fs_read_only(struct bch_fs *c)
 
 	bch2_fs_journal_stop(&c->journal);
 
-	bch_info(c, "%sshutdown complete, journal seq %llu",
+	bch_info(c, "%sclean shutdown complete, journal seq %llu",
 		 test_bit(BCH_FS_clean_shutdown, &c->flags) ? "" : "un",
 		 c->journal.seq_ondisk);
 
@@ -441,6 +441,8 @@ static int __bch2_fs_read_write(struct bch_fs *c, bool early)
 {
 	int ret;
 
+	BUG_ON(!test_bit(BCH_FS_may_go_rw, &c->flags));
+
 	if (test_bit(BCH_FS_initial_gc_unfixed, &c->flags)) {
 		bch_err(c, "cannot go rw, unfixed btree errors");
 		return -BCH_ERR_erofs_unfixed_errors;
@@ -561,6 +563,7 @@ static void __bch2_fs_free(struct bch_fs *c)
 	bch2_io_clock_exit(&c->io_clock[WRITE]);
 	bch2_io_clock_exit(&c->io_clock[READ]);
 	bch2_fs_compress_exit(c);
+	bch2_fs_btree_gc_exit(c);
 	bch2_journal_keys_put_initial(c);
 	bch2_find_btree_nodes_exit(&c->found_btree_nodes);
 	BUG_ON(atomic_read(&c->journal_keys.ref));
@@ -584,7 +587,6 @@ static void __bch2_fs_free(struct bch_fs *c)
 #endif
 	kfree(rcu_dereference_protected(c->disk_groups, 1));
 	kfree(c->journal_seq_blacklist_table);
-	kfree(c->unused_inode_hints);
 
 	if (c->write_ref_wq)
 		destroy_workqueue(c->write_ref_wq);
@@ -766,21 +768,17 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 
 	refcount_set(&c->ro_ref, 1);
 	init_waitqueue_head(&c->ro_ref_wait);
+	spin_lock_init(&c->recovery_pass_lock);
 	sema_init(&c->online_fsck_mutex, 1);
 
-	init_rwsem(&c->gc_lock);
-	mutex_init(&c->gc_gens_lock);
-	atomic_set(&c->journal_keys.ref, 1);
-	c->journal_keys.initial_ref_held = true;
-
 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
 		bch2_time_stats_init(&c->times[i]);
 
-	bch2_fs_gc_init(c);
 	bch2_fs_copygc_init(c);
 	bch2_fs_btree_key_cache_init_early(&c->btree_key_cache);
 	bch2_fs_btree_iter_init_early(c);
 	bch2_fs_btree_interior_update_init_early(c);
+	bch2_fs_journal_keys_init(c);
 	bch2_fs_allocator_background_init(c);
 	bch2_fs_allocator_foreground_init(c);
 	bch2_fs_rebalance_init(c);
@@ -809,9 +807,6 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	INIT_LIST_HEAD(&c->vfs_inodes_list);
 	mutex_init(&c->vfs_inodes_lock);
 
-	c->copy_gc_enabled		= 1;
-	c->rebalance.enabled		= 1;
-
 	c->journal.flush_write_time	= &c->times[BCH_TIME_journal_flush_write];
 	c->journal.noflush_write_time	= &c->times[BCH_TIME_journal_noflush_write];
 	c->journal.flush_seq_time	= &c->times[BCH_TIME_journal_flush_seq];
@@ -873,8 +868,6 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 		(btree_blocks(c) + 1) * 2 *
 		sizeof(struct sort_iter_set);
 
-	c->inode_shard_bits = ilog2(roundup_pow_of_two(num_possible_cpus()));
-
 	if (!(c->btree_update_wq = alloc_workqueue("bcachefs",
 				WQ_HIGHPRI|WQ_FREEZABLE|WQ_MEM_RECLAIM|WQ_UNBOUND, 512)) ||
 	    !(c->btree_io_complete_wq = alloc_workqueue("bcachefs_btree_io",
@@ -901,9 +894,7 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	    !(c->online_reserved = alloc_percpu(u64)) ||
 	    mempool_init_kvmalloc_pool(&c->btree_bounce_pool, 1,
 				       c->opts.btree_node_size) ||
-	    mempool_init_kmalloc_pool(&c->large_bkey_pool, 1, 2048) ||
-	    !(c->unused_inode_hints = kcalloc(1U << c->inode_shard_bits,
-					      sizeof(u64), GFP_KERNEL))) {
+	    mempool_init_kmalloc_pool(&c->large_bkey_pool, 1, 2048)) {
 		ret = -BCH_ERR_ENOMEM_fs_other_alloc;
 		goto err;
 	}
@@ -917,6 +908,7 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	    bch2_fs_btree_cache_init(c) ?:
 	    bch2_fs_btree_key_cache_init(&c->btree_key_cache) ?:
 	    bch2_fs_btree_interior_update_init(c) ?:
+	    bch2_fs_btree_gc_init(c) ?:
 	    bch2_fs_buckets_waiting_for_journal_init(c) ?:
 	    bch2_fs_btree_write_buffer_init(c) ?:
 	    bch2_fs_subvolumes_init(c) ?:
@@ -1033,9 +1025,12 @@ int bch2_fs_start(struct bch_fs *c)
 		bch2_dev_allocator_add(c, ca);
 	bch2_recalc_capacity(c);
 
+	c->recovery_task = current;
 	ret = BCH_SB_INITIALIZED(c->disk_sb.sb)
 		? bch2_fs_recovery(c)
 		: bch2_fs_initialize(c);
+	c->recovery_task = NULL;
+
 	if (ret)
 		goto err;
 
@@ -1120,12 +1115,12 @@ static int bch2_dev_in_fs(struct bch_sb_handle *fs,
 
 		prt_bdevname(&buf, fs->bdev);
 		prt_char(&buf, ' ');
-		bch2_prt_datetime(&buf, le64_to_cpu(fs->sb->write_time));;
+		bch2_prt_datetime(&buf, le64_to_cpu(fs->sb->write_time));
 		prt_newline(&buf);
 
 		prt_bdevname(&buf, sb->bdev);
 		prt_char(&buf, ' ');
-		bch2_prt_datetime(&buf, le64_to_cpu(sb->sb->write_time));;
+		bch2_prt_datetime(&buf, le64_to_cpu(sb->sb->write_time));
 		prt_newline(&buf);
 
 		if (!opts->no_splitbrain_check)
@@ -1198,7 +1193,7 @@ static void bch2_dev_free(struct bch_dev *ca)
 
 	free_percpu(ca->io_done);
 	bch2_dev_buckets_free(ca);
-	free_page((unsigned long) ca->sb_read_scratch);
+	kfree(ca->sb_read_scratch);
 
 	bch2_time_stats_quantiles_exit(&ca->io_latency[WRITE]);
 	bch2_time_stats_quantiles_exit(&ca->io_latency[READ]);
@@ -1309,8 +1304,6 @@ static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
 	init_completion(&ca->ref_completion);
 	init_completion(&ca->io_ref_completion);
 
-	init_rwsem(&ca->bucket_lock);
-
 	INIT_WORK(&ca->io_error_work, bch2_io_error_work);
 
 	bch2_time_stats_quantiles_init(&ca->io_latency[READ]);
@@ -1337,7 +1330,7 @@ static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
 
 	if (percpu_ref_init(&ca->io_ref, bch2_dev_io_ref_complete,
 			    PERCPU_REF_INIT_DEAD, GFP_KERNEL) ||
-	    !(ca->sb_read_scratch = (void *) __get_free_page(GFP_KERNEL)) ||
+	    !(ca->sb_read_scratch = kmalloc(BCH_SB_READ_SCRATCH_BUF_SIZE, GFP_KERNEL)) ||
 	    bch2_dev_buckets_alloc(c, ca) ||
 	    !(ca->io_done	= alloc_percpu(*ca->io_done)))
 		goto err;
@@ -1366,7 +1359,6 @@ static int bch2_dev_alloc(struct bch_fs *c, unsigned dev_idx)
 {
 	struct bch_member member = bch2_sb_member_get(c->disk_sb.sb, dev_idx);
 	struct bch_dev *ca = NULL;
-	int ret = 0;
 
 	if (bch2_fs_init_fault("dev_alloc"))
 		goto err;
@@ -1378,10 +1370,8 @@ static int bch2_dev_alloc(struct bch_fs *c, unsigned dev_idx)
 	ca->fs = c;
 
 	bch2_dev_attach(c, ca, dev_idx);
-	return ret;
+	return 0;
 err:
-	if (ca)
-		bch2_dev_free(ca);
 	return -BCH_ERR_ENOMEM_dev_alloc;
 }
 
@@ -1751,11 +1741,6 @@ int bch2_dev_add(struct bch_fs *c, const char *path)
 	if (ret)
 		goto err;
 
-	ret = bch2_dev_journal_alloc(ca, true);
-	bch_err_msg(c, ret, "allocating journal");
-	if (ret)
-		goto err;
-
 	down_write(&c->state_lock);
 	mutex_lock(&c->sb_lock);
 
@@ -1806,11 +1791,14 @@ int bch2_dev_add(struct bch_fs *c, const char *path)
 	if (ret)
 		goto err_late;
 
-	ca->new_fs_bucket_idx = 0;
-
 	if (ca->mi.state == BCH_MEMBER_STATE_rw)
 		__bch2_dev_read_write(c, ca);
 
+	ret = bch2_dev_journal_alloc(ca, false);
+	bch_err_msg(c, ret, "allocating journal");
+	if (ret)
+		goto err_late;
+
 	up_write(&c->state_lock);
 	return 0;
 
diff --git a/fs/bcachefs/super.h b/fs/bcachefs/super.h
index dada09331d2e..fa6d52216510 100644
--- a/fs/bcachefs/super.h
+++ b/fs/bcachefs/super.h
@@ -34,16 +34,6 @@ void bch2_fs_read_only(struct bch_fs *);
 int bch2_fs_read_write(struct bch_fs *);
 int bch2_fs_read_write_early(struct bch_fs *);
 
-/*
- * Only for use in the recovery/fsck path:
- */
-static inline void bch2_fs_lazy_rw(struct bch_fs *c)
-{
-	if (!test_bit(BCH_FS_rw, &c->flags) &&
-	    !test_bit(BCH_FS_was_rw, &c->flags))
-		bch2_fs_read_write_early(c);
-}
-
 void __bch2_fs_stop(struct bch_fs *);
 void bch2_fs_free(struct bch_fs *);
 void bch2_fs_stop(struct bch_fs *);
diff --git a/fs/bcachefs/sysfs.c b/fs/bcachefs/sysfs.c
index 03e59f86f360..a7eb1f511484 100644
--- a/fs/bcachefs/sysfs.c
+++ b/fs/bcachefs/sysfs.c
@@ -146,7 +146,7 @@ write_attribute(trigger_journal_writes);
 write_attribute(trigger_btree_cache_shrink);
 write_attribute(trigger_btree_key_cache_shrink);
 write_attribute(trigger_freelist_wakeup);
-rw_attribute(gc_gens_pos);
+read_attribute(gc_gens_pos);
 
 read_attribute(uuid);
 read_attribute(minor);
@@ -203,7 +203,6 @@ read_attribute(disk_groups);
 
 read_attribute(has_data);
 read_attribute(alloc_debug);
-read_attribute(accounting);
 read_attribute(usage_base);
 
 #define x(t, n, ...) read_attribute(t);
@@ -211,12 +210,11 @@ BCH_PERSISTENT_COUNTERS()
 #undef x
 
 rw_attribute(discard);
+read_attribute(state);
 rw_attribute(label);
 
-rw_attribute(copy_gc_enabled);
 read_attribute(copy_gc_wait);
 
-rw_attribute(rebalance_enabled);
 sysfs_pd_controller_attribute(rebalance);
 read_attribute(rebalance_status);
 
@@ -237,11 +235,6 @@ write_attribute(perf_test);
 	BCH_TIME_STATS()
 #undef x
 
-static struct attribute sysfs_state_rw = {
-	.name = "state",
-	.mode =  0444,
-};
-
 static size_t bch2_btree_cache_size(struct bch_fs *c)
 {
 	struct btree_cache *bc = &c->btree_cache;
@@ -302,7 +295,8 @@ static int bch2_compression_stats_to_text(struct printbuf *out, struct bch_fs *c
 
 static void bch2_gc_gens_pos_to_text(struct printbuf *out, struct bch_fs *c)
 {
-	prt_printf(out, "%s: ", bch2_btree_id_str(c->gc_gens_btree));
+	bch2_btree_id_to_text(out, c->gc_gens_btree);
+	prt_printf(out, ": ");
 	bch2_bpos_to_text(out, c->gc_gens_pos);
 	prt_printf(out, "\n");
 }
@@ -339,9 +333,6 @@ SHOW(bch2_fs)
 	if (attr == &sysfs_gc_gens_pos)
 		bch2_gc_gens_pos_to_text(out, c);
 
-	sysfs_printf(copy_gc_enabled, "%i", c->copy_gc_enabled);
-
-	sysfs_printf(rebalance_enabled,		"%i", c->rebalance.enabled);
 	sysfs_pd_controller_show(rebalance,	&c->rebalance.pd); /* XXX */
 
 	if (attr == &sysfs_copy_gc_wait)
@@ -405,9 +396,6 @@ SHOW(bch2_fs)
 	if (attr == &sysfs_alloc_debug)
 		bch2_fs_alloc_debug_to_text(out, c);
 
-	if (attr == &sysfs_accounting)
-		bch2_fs_accounting_to_text(out, c);
-
 	if (attr == &sysfs_usage_base)
 		bch2_fs_usage_base_to_text(out, c);
 
@@ -418,23 +406,6 @@ STORE(bch2_fs)
 {
 	struct bch_fs *c = container_of(kobj, struct bch_fs, kobj);
 
-	if (attr == &sysfs_copy_gc_enabled) {
-		ssize_t ret = strtoul_safe(buf, c->copy_gc_enabled)
-			?: (ssize_t) size;
-
-		if (c->copygc_thread)
-			wake_up_process(c->copygc_thread);
-		return ret;
-	}
-
-	if (attr == &sysfs_rebalance_enabled) {
-		ssize_t ret = strtoul_safe(buf, c->rebalance.enabled)
-			?: (ssize_t) size;
-
-		rebalance_wakeup(c);
-		return ret;
-	}
-
 	sysfs_pd_controller_store(rebalance,	&c->rebalance.pd);
 
 	/* Debugging: */
@@ -534,15 +505,22 @@ SHOW(bch2_fs_counters)
 
 	printbuf_tabstop_push(out, 32);
 
-	#define x(t, ...) \
+	#define x(t, n, f, ...) \
 		if (attr == &sysfs_##t) {					\
 			counter             = percpu_u64_get(&c->counters[BCH_COUNTER_##t]);\
 			counter_since_mount = counter - c->counters_on_mount[BCH_COUNTER_##t];\
+			if (f & TYPE_SECTORS) {					\
+				counter <<= 9;					\
+				counter_since_mount <<= 9;			\
+			}							\
+										\
 			prt_printf(out, "since mount:\t");			\
+			(f & TYPE_COUNTER) ? prt_u64(out, counter_since_mount) :\
 			prt_human_readable_u64(out, counter_since_mount);	\
 			prt_newline(out);					\
 										\
 			prt_printf(out, "since filesystem creation:\t");	\
+			(f & TYPE_COUNTER) ? prt_u64(out, counter) :		\
 			prt_human_readable_u64(out, counter);			\
 			prt_newline(out);					\
 		}
@@ -610,10 +588,8 @@ struct attribute *bch2_fs_internal_files[] = {
 
 	&sysfs_gc_gens_pos,
 
-	&sysfs_copy_gc_enabled,
 	&sysfs_copy_gc_wait,
 
-	&sysfs_rebalance_enabled,
 	sysfs_pd_controller_files(rebalance),
 
 	&sysfs_moving_ctxts,
@@ -622,7 +598,6 @@ struct attribute *bch2_fs_internal_files[] = {
 
 	&sysfs_disk_groups,
 	&sysfs_alloc_debug,
-	&sysfs_accounting,
 	&sysfs_usage_base,
 	NULL
 };
@@ -682,6 +657,13 @@ STORE(bch2_fs_opts_dir)
 	     (id == Opt_compression && !c->opts.background_compression)))
 		bch2_set_rebalance_needs_scan(c, 0);
 
+	if (v && id == Opt_rebalance_enabled)
+		rebalance_wakeup(c);
+
+	if (v && id == Opt_copygc_enabled &&
+	    c->copygc_thread)
+		wake_up_process(c->copygc_thread);
+
 	ret = size;
 err:
 	bch2_write_ref_put(c, BCH_WRITE_REF_sysfs);
@@ -790,7 +772,7 @@ SHOW(bch2_dev)
 		prt_char(out, '\n');
 	}
 
-	if (attr == &sysfs_state_rw) {
+	if (attr == &sysfs_state) {
 		prt_string_option(out, bch2_member_states, ca->mi.state);
 		prt_char(out, '\n');
 	}
@@ -870,7 +852,7 @@ struct attribute *bch2_dev_files[] = {
 
 	/* settings: */
 	&sysfs_discard,
-	&sysfs_state_rw,
+	&sysfs_state,
 	&sysfs_label,
 
 	&sysfs_has_data,
diff --git a/fs/bcachefs/tests.c b/fs/bcachefs/tests.c
index fb5c1543e52f..6c6469814637 100644
--- a/fs/bcachefs/tests.c
+++ b/fs/bcachefs/tests.c
@@ -131,7 +131,7 @@ static int test_iterate(struct bch_fs *c, u64 nr)
 	i = 0;
 
 	ret = bch2_trans_run(c,
-		for_each_btree_key_upto(trans, iter, BTREE_ID_xattrs,
+		for_each_btree_key_max(trans, iter, BTREE_ID_xattrs,
 					SPOS(0, 0, U32_MAX), POS(0, U64_MAX),
 					0, k, ({
 			BUG_ON(k.k->p.offset != i++);
@@ -186,7 +186,7 @@ static int test_iterate_extents(struct bch_fs *c, u64 nr)
 	i = 0;
 
 	ret = bch2_trans_run(c,
-		for_each_btree_key_upto(trans, iter, BTREE_ID_extents,
+		for_each_btree_key_max(trans, iter, BTREE_ID_extents,
 					SPOS(0, 0, U32_MAX), POS(0, U64_MAX),
 					0, k, ({
 			BUG_ON(bkey_start_offset(k.k) != i);
@@ -242,7 +242,7 @@ static int test_iterate_slots(struct bch_fs *c, u64 nr)
 	i = 0;
 
 	ret = bch2_trans_run(c,
-		for_each_btree_key_upto(trans, iter, BTREE_ID_xattrs,
+		for_each_btree_key_max(trans, iter, BTREE_ID_xattrs,
 					  SPOS(0, 0, U32_MAX), POS(0, U64_MAX),
 					  0, k, ({
 			BUG_ON(k.k->p.offset != i);
@@ -259,7 +259,7 @@ static int test_iterate_slots(struct bch_fs *c, u64 nr)
 	i = 0;
 
 	ret = bch2_trans_run(c,
-		for_each_btree_key_upto(trans, iter, BTREE_ID_xattrs,
+		for_each_btree_key_max(trans, iter, BTREE_ID_xattrs,
 					SPOS(0, 0, U32_MAX), POS(0, U64_MAX),
 					BTREE_ITER_slots, k, ({
 			if (i >= nr * 2)
@@ -302,7 +302,7 @@ static int test_iterate_slots_extents(struct bch_fs *c, u64 nr)
 	i = 0;
 
 	ret = bch2_trans_run(c,
-		for_each_btree_key_upto(trans, iter, BTREE_ID_extents,
+		for_each_btree_key_max(trans, iter, BTREE_ID_extents,
 					SPOS(0, 0, U32_MAX), POS(0, U64_MAX),
 					0, k, ({
 			BUG_ON(bkey_start_offset(k.k) != i + 8);
@@ -320,7 +320,7 @@ static int test_iterate_slots_extents(struct bch_fs *c, u64 nr)
 	i = 0;
 
 	ret = bch2_trans_run(c,
-		for_each_btree_key_upto(trans, iter, BTREE_ID_extents,
+		for_each_btree_key_max(trans, iter, BTREE_ID_extents,
 					SPOS(0, 0, U32_MAX), POS(0, U64_MAX),
 					BTREE_ITER_slots, k, ({
 			if (i == nr)
@@ -349,10 +349,10 @@ static int test_peek_end(struct bch_fs *c, u64 nr)
 	bch2_trans_iter_init(trans, &iter, BTREE_ID_xattrs,
 			     SPOS(0, 0, U32_MAX), 0);
 
-	lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_upto(&iter, POS(0, U64_MAX))));
+	lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_max(&iter, POS(0, U64_MAX))));
 	BUG_ON(k.k);
 
-	lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_upto(&iter, POS(0, U64_MAX))));
+	lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_max(&iter, POS(0, U64_MAX))));
 	BUG_ON(k.k);
 
 	bch2_trans_iter_exit(trans, &iter);
@@ -369,10 +369,10 @@ static int test_peek_end_extents(struct bch_fs *c, u64 nr)
 	bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
 			     SPOS(0, 0, U32_MAX), 0);
 
-	lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_upto(&iter, POS(0, U64_MAX))));
+	lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_max(&iter, POS(0, U64_MAX))));
 	BUG_ON(k.k);
 
-	lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_upto(&iter, POS(0, U64_MAX))));
+	lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_max(&iter, POS(0, U64_MAX))));
 	BUG_ON(k.k);
 
 	bch2_trans_iter_exit(trans, &iter);
@@ -488,7 +488,7 @@ static int test_snapshot_filter(struct bch_fs *c, u32 snapid_lo, u32 snapid_hi)
 	trans = bch2_trans_get(c);
 	bch2_trans_iter_init(trans, &iter, BTREE_ID_xattrs,
 			     SPOS(0, 0, snapid_lo), 0);
-	lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_upto(&iter, POS(0, U64_MAX))));
+	lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_max(&iter, POS(0, U64_MAX))));
 
 	BUG_ON(k.k->p.snapshot != U32_MAX);
 
@@ -672,7 +672,7 @@ static int __do_delete(struct btree_trans *trans, struct bpos pos)
 
 	bch2_trans_iter_init(trans, &iter, BTREE_ID_xattrs, pos,
 			     BTREE_ITER_intent);
-	k = bch2_btree_iter_peek_upto(&iter, POS(0, U64_MAX));
+	k = bch2_btree_iter_peek_max(&iter, POS(0, U64_MAX));
 	ret = bkey_err(k);
 	if (ret)
 		goto err;
@@ -726,7 +726,7 @@ static int seq_insert(struct bch_fs *c, u64 nr)
 static int seq_lookup(struct bch_fs *c, u64 nr)
 {
 	return bch2_trans_run(c,
-		for_each_btree_key_upto(trans, iter, BTREE_ID_xattrs,
+		for_each_btree_key_max(trans, iter, BTREE_ID_xattrs,
 				  SPOS(0, 0, U32_MAX), POS(0, U64_MAX),
 				  0, k,
 		0));
diff --git a/fs/bcachefs/trace.h b/fs/bcachefs/trace.h
index 5597b9d6297f..9d40b7d4ea29 100644
--- a/fs/bcachefs/trace.h
+++ b/fs/bcachefs/trace.h
@@ -199,6 +199,30 @@ DECLARE_EVENT_CLASS(bio,
 		  (unsigned long long)__entry->sector, __entry->nr_sector)
 );
 
+/* disk_accounting.c */
+
+TRACE_EVENT(accounting_mem_insert,
+	TP_PROTO(struct bch_fs *c, const char *acc),
+	TP_ARGS(c, acc),
+
+	TP_STRUCT__entry(
+		__field(dev_t,		dev			)
+		__field(unsigned,	new_nr			)
+		__string(acc,		acc			)
+	),
+
+	TP_fast_assign(
+		__entry->dev		= c->dev;
+		__entry->new_nr		= c->accounting.k.nr;
+		__assign_str(acc);
+	),
+
+	TP_printk("%d,%d entries %u added %s",
+		  MAJOR(__entry->dev), MINOR(__entry->dev),
+		  __entry->new_nr,
+		  __get_str(acc))
+);
+
 /* fs.c: */
 TRACE_EVENT(bch2_sync_fs,
 	TP_PROTO(struct super_block *sb, int wait),
@@ -848,8 +872,8 @@ TRACE_EVENT(move_data,
 TRACE_EVENT(evacuate_bucket,
 	TP_PROTO(struct bch_fs *c, struct bpos *bucket,
 		 unsigned sectors, unsigned bucket_size,
-		 u64 fragmentation, int ret),
-	TP_ARGS(c, bucket, sectors, bucket_size, fragmentation, ret),
+		 int ret),
+	TP_ARGS(c, bucket, sectors, bucket_size, ret),
 
 	TP_STRUCT__entry(
 		__field(dev_t,		dev		)
@@ -857,7 +881,6 @@ TRACE_EVENT(evacuate_bucket,
 		__field(u64,		bucket		)
 		__field(u32,		sectors		)
 		__field(u32,		bucket_size	)
-		__field(u64,		fragmentation	)
 		__field(int,		ret		)
 	),
 
@@ -867,15 +890,14 @@ TRACE_EVENT(evacuate_bucket,
 		__entry->bucket			= bucket->offset;
 		__entry->sectors		= sectors;
 		__entry->bucket_size		= bucket_size;
-		__entry->fragmentation		= fragmentation;
 		__entry->ret			= ret;
 	),
 
-	TP_printk("%d,%d %llu:%llu sectors %u/%u fragmentation %llu ret %i",
+	TP_printk("%d,%d %llu:%llu sectors %u/%u ret %i",
 		  MAJOR(__entry->dev), MINOR(__entry->dev),
 		  __entry->member, __entry->bucket,
 		  __entry->sectors, __entry->bucket_size,
-		  __entry->fragmentation, __entry->ret)
+		  __entry->ret)
 );
 
 TRACE_EVENT(copygc,
@@ -1316,6 +1338,12 @@ TRACE_EVENT(trans_restart_key_cache_key_realloced,
 		  __entry->new_u64s)
 );
 
+DEFINE_EVENT(transaction_event,	trans_restart_write_buffer_flush,
+	TP_PROTO(struct btree_trans *trans,
+		 unsigned long caller_ip),
+	TP_ARGS(trans, caller_ip)
+);
+
 TRACE_EVENT(path_downgrade,
 	TP_PROTO(struct btree_trans *trans,
 		 unsigned long caller_ip,
@@ -1352,10 +1380,21 @@ TRACE_EVENT(path_downgrade,
 		  __entry->pos_snapshot)
 );
 
-DEFINE_EVENT(transaction_event,	trans_restart_write_buffer_flush,
-	TP_PROTO(struct btree_trans *trans,
-		 unsigned long caller_ip),
-	TP_ARGS(trans, caller_ip)
+TRACE_EVENT(key_cache_fill,
+	TP_PROTO(struct btree_trans *trans, const char *key),
+	TP_ARGS(trans, key),
+
+	TP_STRUCT__entry(
+		__array(char,		trans_fn, 32	)
+		__string(key,		key			)
+	),
+
+	TP_fast_assign(
+		strscpy(__entry->trans_fn, trans->fn, sizeof(__entry->trans_fn));
+		__assign_str(key);
+	),
+
+	TP_printk("%s %s", __entry->trans_fn, __get_str(key))
 );
 
 TRACE_EVENT(write_buffer_flush,
@@ -1414,6 +1453,24 @@ TRACE_EVENT(write_buffer_flush_slowpath,
 	TP_printk("%zu/%zu", __entry->slowpath, __entry->total)
 );
 
+TRACE_EVENT(write_buffer_maybe_flush,
+	TP_PROTO(struct btree_trans *trans, unsigned long caller_ip, const char *key),
+	TP_ARGS(trans, caller_ip, key),
+
+	TP_STRUCT__entry(
+		__array(char,			trans_fn, 32	)
+		__field(unsigned long,		caller_ip	)
+		__string(key,			key		)
+	),
+
+	TP_fast_assign(
+		strscpy(__entry->trans_fn, trans->fn, sizeof(__entry->trans_fn));
+		__assign_str(key);
+	),
+
+	TP_printk("%s %pS %s", __entry->trans_fn, (void *) __entry->caller_ip, __get_str(key))
+);
+
 DEFINE_EVENT(fs_str, rebalance_extent,
 	TP_PROTO(struct bch_fs *c, const char *str),
 	TP_ARGS(c, str)
diff --git a/fs/bcachefs/util.h b/fs/bcachefs/util.h
index fb02c1c36004..1a1720116071 100644
--- a/fs/bcachefs/util.h
+++ b/fs/bcachefs/util.h
@@ -55,6 +55,16 @@ static inline size_t buf_pages(void *p, size_t len)
 			    PAGE_SIZE);
 }
 
+static inline void *bch2_kvmalloc(size_t n, gfp_t flags)
+{
+	void *p = unlikely(n >= INT_MAX)
+		? vmalloc(n)
+		: kvmalloc(n, flags & ~__GFP_ZERO);
+	if (p && (flags & __GFP_ZERO))
+		memset(p, 0, n);
+	return p;
+}
+
 #define init_heap(heap, _size, gfp)					\
 ({									\
 	(heap)->nr = 0;						\
@@ -317,6 +327,19 @@ do {									\
 	_ptr ? container_of(_ptr, type, member) : NULL;			\
 })
 
+static inline struct list_head *list_pop(struct list_head *head)
+{
+	if (list_empty(head))
+		return NULL;
+
+	struct list_head *ret = head->next;
+	list_del_init(ret);
+	return ret;
+}
+
+#define list_pop_entry(head, type, member)		\
+	container_of_or_null(list_pop(head), type, member)
+
 /* Does linear interpolation between powers of two */
 static inline unsigned fract_exp_two(unsigned x, unsigned fract_bits)
 {
@@ -696,4 +719,13 @@ static inline bool test_bit_le64(size_t bit, __le64 *addr)
 	return (addr[bit / 64] & cpu_to_le64(BIT_ULL(bit % 64))) != 0;
 }
 
+static inline void memcpy_swab(void *_dst, void *_src, size_t len)
+{
+	u8 *dst = _dst + len;
+	u8 *src = _src;
+
+	while (len--)
+		*--dst = *src++;
+}
+
 #endif /* _BCACHEFS_UTIL_H */
diff --git a/fs/bcachefs/varint.c b/fs/bcachefs/varint.c
index 6a78553d9b0c..6620ecae26af 100644
--- a/fs/bcachefs/varint.c
+++ b/fs/bcachefs/varint.c
@@ -9,6 +9,7 @@
 #include <valgrind/memcheck.h>
 #endif
 
+#include "errcode.h"
 #include "varint.h"
 
 /**
@@ -53,7 +54,7 @@ int bch2_varint_decode(const u8 *in, const u8 *end, u64 *out)
 	u64 v;
 
 	if (unlikely(in + bytes > end))
-		return -1;
+		return -BCH_ERR_varint_decode_error;
 
 	if (likely(bytes < 9)) {
 		__le64 v_le = 0;
@@ -115,7 +116,7 @@ int bch2_varint_decode_fast(const u8 *in, const u8 *end, u64 *out)
 	unsigned bytes = ffz(*in) + 1;
 
 	if (unlikely(in + bytes > end))
-		return -1;
+		return -BCH_ERR_varint_decode_error;
 
 	if (likely(bytes < 9)) {
 		v >>= bytes;
diff --git a/fs/bcachefs/xattr.c b/fs/bcachefs/xattr.c
index 952aca400faf..aed7c6984173 100644
--- a/fs/bcachefs/xattr.c
+++ b/fs/bcachefs/xattr.c
@@ -71,7 +71,7 @@ const struct bch_hash_desc bch2_xattr_hash_desc = {
 };
 
 int bch2_xattr_validate(struct bch_fs *c, struct bkey_s_c k,
-		       enum bch_validate_flags flags)
+			struct bkey_validate_context from)
 {
 	struct bkey_s_c_xattr xattr = bkey_s_c_to_xattr(k);
 	unsigned val_u64s = xattr_val_u64s(xattr.v->x_name_len,
@@ -309,7 +309,7 @@ ssize_t bch2_xattr_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 	u64 offset = 0, inum = inode->ei_inode.bi_inum;
 
 	int ret = bch2_trans_run(c,
-		for_each_btree_key_in_subvolume_upto(trans, iter, BTREE_ID_xattrs,
+		for_each_btree_key_in_subvolume_max(trans, iter, BTREE_ID_xattrs,
 				   POS(inum, offset),
 				   POS(inum, U64_MAX),
 				   inode->ei_inum.subvol, 0, k, ({
@@ -565,13 +565,6 @@ static int bch2_xattr_bcachefs_set(const struct xattr_handler *handler,
 	ret = bch2_write_inode(c, inode, inode_opt_set_fn, &s, 0);
 err:
 	mutex_unlock(&inode->ei_update_lock);
-
-	if (value &&
-	    (opt_id == Opt_background_target ||
-	     opt_id == Opt_background_compression ||
-	     (opt_id == Opt_compression && !inode_opt_get(c, &inode->ei_inode, background_compression))))
-		bch2_set_rebalance_needs_scan(c, inode->ei_inode.bi_inum);
-
 err_class_exit:
 	return bch2_err_class(ret);
 }
@@ -609,7 +602,7 @@ static const struct xattr_handler bch_xattr_bcachefs_effective_handler = {
 
 #endif /* NO_BCACHEFS_FS */
 
-const struct xattr_handler *bch2_xattr_handlers[] = {
+const struct xattr_handler * const bch2_xattr_handlers[] = {
 	&bch_xattr_user_handler,
 	&bch_xattr_trusted_handler,
 	&bch_xattr_security_handler,
diff --git a/fs/bcachefs/xattr.h b/fs/bcachefs/xattr.h
index c188a5ad64ce..132fbbd15a66 100644
--- a/fs/bcachefs/xattr.h
+++ b/fs/bcachefs/xattr.h
@@ -6,7 +6,8 @@
 
 extern const struct bch_hash_desc bch2_xattr_hash_desc;
 
-int bch2_xattr_validate(struct bch_fs *, struct bkey_s_c, enum bch_validate_flags);
+int bch2_xattr_validate(struct bch_fs *, struct bkey_s_c,
+			struct bkey_validate_context);
 void bch2_xattr_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
 #define bch2_bkey_ops_xattr ((struct bkey_ops) {	\
@@ -44,6 +45,6 @@ int bch2_xattr_set(struct btree_trans *, subvol_inum,
 
 ssize_t bch2_xattr_list(struct dentry *, char *, size_t);
 
-extern const struct xattr_handler *bch2_xattr_handlers[];
+extern const struct xattr_handler * const bch2_xattr_handlers[];
 
 #endif /* _BCACHEFS_XATTR_H */
diff --git a/fs/fs_parser.c b/fs/fs_parser.c
index 16fa61ef56bf..e635a81e17d9 100644
--- a/fs/fs_parser.c
+++ b/fs/fs_parser.c
@@ -13,7 +13,7 @@
 #include <linux/namei.h>
 #include "internal.h"
 
-static const struct constant_table bool_names[] = {
+const struct constant_table bool_names[] = {
 	{ "0",		false },
 	{ "1",		true },
 	{ "false",	false },
@@ -22,6 +22,7 @@ static const struct constant_table bool_names[] = {
 	{ "yes",	true },
 	{ },
 };
+EXPORT_SYMBOL(bool_names);
 
 static const struct constant_table *
 __lookup_constant(const struct constant_table *tbl, const char *name)
diff --git a/include/linux/fs_parser.h b/include/linux/fs_parser.h
index 3cef566088fc..53e566efd5fd 100644
--- a/include/linux/fs_parser.h
+++ b/include/linux/fs_parser.h
@@ -84,6 +84,8 @@ extern int fs_lookup_param(struct fs_context *fc,
 
 extern int lookup_constant(const struct constant_table tbl[], const char *name, int not_found);
 
+extern const struct constant_table bool_names[];
+
 #ifdef CONFIG_VALIDATE_FS_PARSER
 extern bool validate_constant_table(const struct constant_table *tbl, size_t tbl_size,
 				    int low, int high, int special);
diff --git a/include/linux/min_heap.h b/include/linux/min_heap.h
index e781727c8916..6325f6ffb895 100644
--- a/include/linux/min_heap.h
+++ b/include/linux/min_heap.h
@@ -15,8 +15,8 @@
  */
 #define MIN_HEAP_PREALLOCATED(_type, _name, _nr)	\
 struct _name {	\
-	int nr;	\
-	int size;	\
+	size_t nr;	\
+	size_t size;	\
 	_type *data;	\
 	_type preallocated[_nr];	\
 }