1 files changed, 813 insertions, 0 deletions

diff --git a/fs/bcachefs/btree_key_cache.c b/fs/bcachefs/btree_key_cache.c
new file mode 100644
index 000000000000..244610b1d0b5
--- /dev/null
+++ b/fs/bcachefs/btree_key_cache.c
@@ -0,0 +1,813 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "btree_cache.h"
+#include "btree_iter.h"
+#include "btree_key_cache.h"
+#include "btree_locking.h"
+#include "btree_update.h"
+#include "errcode.h"
+#include "error.h"
+#include "journal.h"
+#include "journal_reclaim.h"
+#include "trace.h"
+
+#include <linux/sched/mm.h>
+
+static inline bool btree_uses_pcpu_readers(enum btree_id id)
+{
+	return id == BTREE_ID_subvolumes;
+}
+
+static struct kmem_cache *bch2_key_cache;
+
+static int bch2_btree_key_cache_cmp_fn(struct rhashtable_compare_arg *arg,
+				       const void *obj)
+{
+	const struct bkey_cached *ck = obj;
+	const struct bkey_cached_key *key = arg->key;
+
+	return ck->key.btree_id != key->btree_id ||
+		!bpos_eq(ck->key.pos, key->pos);
+}
+
+static const struct rhashtable_params bch2_btree_key_cache_params = {
+	.head_offset		= offsetof(struct bkey_cached, hash),
+	.key_offset		= offsetof(struct bkey_cached, key),
+	.key_len		= sizeof(struct bkey_cached_key),
+	.obj_cmpfn		= bch2_btree_key_cache_cmp_fn,
+	.automatic_shrinking	= true,
+};
+
+static inline void btree_path_cached_set(struct btree_trans *trans, struct btree_path *path,
+					 struct bkey_cached *ck,
+					 enum btree_node_locked_type lock_held)
+{
+	path->l[0].lock_seq	= six_lock_seq(&ck->c.lock);
+	path->l[0].b		= (void *) ck;
+	mark_btree_node_locked(trans, path, 0, lock_held);
+}
+
+__flatten
+inline struct bkey_cached *
+bch2_btree_key_cache_find(struct bch_fs *c, enum btree_id btree_id, struct bpos pos)
+{
+	struct bkey_cached_key key = {
+		.btree_id	= btree_id,
+		.pos		= pos,
+	};
+
+	return rhashtable_lookup_fast(&c->btree_key_cache.table, &key,
+				      bch2_btree_key_cache_params);
+}
+
+static bool bkey_cached_lock_for_evict(struct bkey_cached *ck)
+{
+	if (!six_trylock_intent(&ck->c.lock))
+		return false;
+
+	if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+		six_unlock_intent(&ck->c.lock);
+		return false;
+	}
+
+	if (!six_trylock_write(&ck->c.lock)) {
+		six_unlock_intent(&ck->c.lock);
+		return false;
+	}
+
+	return true;
+}
+
+static bool bkey_cached_evict(struct btree_key_cache *c,
+			      struct bkey_cached *ck)
+{
+	bool ret = !rhashtable_remove_fast(&c->table, &ck->hash,
+				      bch2_btree_key_cache_params);
+	if (ret) {
+		memset(&ck->key, ~0, sizeof(ck->key));
+		atomic_long_dec(&c->nr_keys);
+	}
+
+	return ret;
+}
+
+static void __bkey_cached_free(struct rcu_pending *pending, struct rcu_head *rcu)
+{
+	struct bch_fs *c = container_of(pending->srcu, struct bch_fs, btree_trans_barrier);
+	struct bkey_cached *ck = container_of(rcu, struct bkey_cached, rcu);
+
+	this_cpu_dec(*c->btree_key_cache.nr_pending);
+	kmem_cache_free(bch2_key_cache, ck);
+}
+
+static void bkey_cached_free(struct btree_key_cache *bc,
+			     struct bkey_cached *ck)
+{
+	kfree(ck->k);
+	ck->k		= NULL;
+	ck->u64s	= 0;
+
+	six_unlock_write(&ck->c.lock);
+	six_unlock_intent(&ck->c.lock);
+
+	bool pcpu_readers = ck->c.lock.readers != NULL;
+	rcu_pending_enqueue(&bc->pending[pcpu_readers], &ck->rcu);
+	this_cpu_inc(*bc->nr_pending);
+}
+
+static struct bkey_cached *__bkey_cached_alloc(unsigned key_u64s, gfp_t gfp)
+{
+	gfp |= __GFP_ACCOUNT|__GFP_RECLAIMABLE;
+
+	struct bkey_cached *ck = kmem_cache_zalloc(bch2_key_cache, gfp);
+	if (unlikely(!ck))
+		return NULL;
+	ck->k = kmalloc(key_u64s * sizeof(u64), gfp);
+	if (unlikely(!ck->k)) {
+		kmem_cache_free(bch2_key_cache, ck);
+		return NULL;
+	}
+	ck->u64s = key_u64s;
+	return ck;
+}
+
+static struct bkey_cached *
+bkey_cached_alloc(struct btree_trans *trans, struct btree_path *path, unsigned key_u64s)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_key_cache *bc = &c->btree_key_cache;
+	bool pcpu_readers = btree_uses_pcpu_readers(path->btree_id);
+	int ret;
+
+	struct bkey_cached *ck = container_of_or_null(
+				rcu_pending_dequeue(&bc->pending[pcpu_readers]),
+				struct bkey_cached, rcu);
+	if (ck)
+		goto lock;
+
+	ck = allocate_dropping_locks(trans, ret,
+				     __bkey_cached_alloc(key_u64s, _gfp));
+	if (ret) {
+		if (ck)
+			kfree(ck->k);
+		kmem_cache_free(bch2_key_cache, ck);
+		return ERR_PTR(ret);
+	}
+
+	if (ck) {
+		bch2_btree_lock_init(&ck->c, pcpu_readers ? SIX_LOCK_INIT_PCPU : 0);
+		ck->c.cached = true;
+		goto lock;
+	}
+
+	ck = container_of_or_null(rcu_pending_dequeue_from_all(&bc->pending[pcpu_readers]),
+				  struct bkey_cached, rcu);
+	if (ck)
+		goto lock;
+lock:
+	six_lock_intent(&ck->c.lock, NULL, NULL);
+	six_lock_write(&ck->c.lock, NULL, NULL);
+	return ck;
+}
+
+static struct bkey_cached *
+bkey_cached_reuse(struct btree_key_cache *c)
+{
+	struct bucket_table *tbl;
+	struct rhash_head *pos;
+	struct bkey_cached *ck;
+	unsigned i;
+
+	rcu_read_lock();
+	tbl = rht_dereference_rcu(c->table.tbl, &c->table);
+	for (i = 0; i < tbl->size; i++)
+		rht_for_each_entry_rcu(ck, pos, tbl, i, hash) {
+			if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
+			    bkey_cached_lock_for_evict(ck)) {
+				if (bkey_cached_evict(c, ck))
+					goto out;
+				six_unlock_write(&ck->c.lock);
+				six_unlock_intent(&ck->c.lock);
+			}
+		}
+	ck = NULL;
+out:
+	rcu_read_unlock();
+	return ck;
+}
+
+static int btree_key_cache_create(struct btree_trans *trans, struct btree_path *path,
+				  struct bkey_s_c k)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_key_cache *bc = &c->btree_key_cache;
+
+	/*
+	 * bch2_varint_decode can read past the end of the buffer by at
+	 * most 7 bytes (it won't be used):
+	 */
+	unsigned key_u64s = k.k->u64s + 1;
+
+	/*
+	 * Allocate some extra space so that the transaction commit path is less
+	 * likely to have to reallocate, since that requires a transaction
+	 * restart:
+	 */
+	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);
+	int ret = PTR_ERR_OR_ZERO(ck);
+	if (ret)
+		return ret;
+
+	if (unlikely(!ck)) {
+		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));
+			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->flags		= 1U << BKEY_CACHED_ACCESSED;
+
+	if (unlikely(key_u64s > ck->u64s)) {
+		mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
+
+		struct bkey_i *new_k = allocate_dropping_locks(trans, ret,
+				kmalloc(key_u64s * sizeof(u64), _gfp));
+		if (unlikely(!new_k)) {
+			bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
+				bch2_btree_id_str(ck->key.btree_id), key_u64s);
+			ret = -BCH_ERR_ENOMEM_btree_key_cache_fill;
+		} else if (ret) {
+			kfree(new_k);
+			goto err;
+		}
+
+		kfree(ck->k);
+		ck->k = new_k;
+		ck->u64s = key_u64s;
+	}
+
+	bkey_reassemble(ck->k, k);
+
+	ret = rhashtable_lookup_insert_fast(&bc->table, &ck->hash, bch2_btree_key_cache_params);
+	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);
+	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;
+	return 0;
+err:
+	bkey_cached_free(bc, ck);
+	mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
+
+	return ret;
+}
+
+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;
+		return 0;
+	}
+
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+
+	bch2_trans_iter_init(trans, &iter, ck_path->btree_id, ck_path->pos,
+			     BTREE_ITER_key_cache_fill|
+			     BTREE_ITER_cached_nofill);
+	iter.flags &= ~BTREE_ITER_with_journal;
+	k = bch2_btree_iter_peek_slot(&iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	/* Recheck after btree lookup, before allocating: */
+	ret = bch2_btree_key_cache_find(c, ck_path->btree_id, ck_path->pos) ? -EEXIST : 0;
+	if (unlikely(ret))
+		goto out;
+
+	ret = btree_key_cache_create(trans, ck_path, k);
+	if (ret)
+		goto err;
+out:
+	/* We're not likely to need this iterator again: */
+	bch2_set_btree_iter_dontneed(&iter);
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+static inline int btree_path_traverse_cached_fast(struct btree_trans *trans,
+						  struct btree_path *path)
+{
+	struct bch_fs *c = trans->c;
+	struct bkey_cached *ck;
+retry:
+	ck = bch2_btree_key_cache_find(c, path->btree_id, path->pos);
+	if (!ck)
+		return -ENOENT;
+
+	enum six_lock_type lock_want = __btree_lock_want(path, 0);
+
+	int ret = btree_node_lock(trans, path, (void *) ck, 0, lock_want, _THIS_IP_);
+	if (ret)
+		return ret;
+
+	if (ck->key.btree_id != path->btree_id ||
+	    !bpos_eq(ck->key.pos, path->pos)) {
+		six_unlock_type(&ck->c.lock, lock_want);
+		goto retry;
+	}
+
+	if (!test_bit(BKEY_CACHED_ACCESSED, &ck->flags))
+		set_bit(BKEY_CACHED_ACCESSED, &ck->flags);
+
+	btree_path_cached_set(trans, path, ck, (enum btree_node_locked_type) lock_want);
+	path->uptodate = BTREE_ITER_UPTODATE;
+	return 0;
+}
+
+int bch2_btree_path_traverse_cached(struct btree_trans *trans, struct btree_path *path,
+				    unsigned flags)
+{
+	EBUG_ON(path->level);
+
+	path->l[1].b = NULL;
+
+	int ret;
+	do {
+		ret = btree_path_traverse_cached_fast(trans, path);
+		if (unlikely(ret == -ENOENT))
+			ret = btree_key_cache_fill(trans, path, flags);
+	} while (ret == -EEXIST);
+
+	if (unlikely(ret)) {
+		path->uptodate = BTREE_ITER_NEED_TRAVERSE;
+		if (!bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
+			btree_node_unlock(trans, path, 0);
+			path->l[0].b = ERR_PTR(ret);
+		}
+	}
+	return ret;
+}
+
+static int btree_key_cache_flush_pos(struct btree_trans *trans,
+				     struct bkey_cached_key key,
+				     u64 journal_seq,
+				     unsigned commit_flags,
+				     bool evict)
+{
+	struct bch_fs *c = trans->c;
+	struct journal *j = &c->journal;
+	struct btree_iter c_iter, b_iter;
+	struct bkey_cached *ck = NULL;
+	int ret;
+
+	bch2_trans_iter_init(trans, &b_iter, key.btree_id, key.pos,
+			     BTREE_ITER_slots|
+			     BTREE_ITER_intent|
+			     BTREE_ITER_all_snapshots);
+	bch2_trans_iter_init(trans, &c_iter, key.btree_id, key.pos,
+			     BTREE_ITER_cached|
+			     BTREE_ITER_intent);
+	b_iter.flags &= ~BTREE_ITER_with_key_cache;
+
+	ret = bch2_btree_iter_traverse(&c_iter);
+	if (ret)
+		goto out;
+
+	ck = (void *) btree_iter_path(trans, &c_iter)->l[0].b;
+	if (!ck)
+		goto out;
+
+	if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+		if (evict)
+			goto evict;
+		goto out;
+	}
+
+	if (journal_seq && ck->journal.seq != journal_seq)
+		goto out;
+
+	trans->journal_res.seq = ck->journal.seq;
+
+	/*
+	 * If we're at the end of the journal, we really want to free up space
+	 * in the journal right away - we don't want to pin that old journal
+	 * sequence number with a new btree node write, we want to re-journal
+	 * the update
+	 */
+	if (ck->journal.seq == journal_last_seq(j))
+		commit_flags |= BCH_WATERMARK_reclaim;
+
+	if (ck->journal.seq != journal_last_seq(j) ||
+	    !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,
+				  BTREE_UPDATE_key_cache_reclaim|
+				  BTREE_UPDATE_internal_snapshot_node|
+				  BTREE_TRIGGER_norun) ?:
+		bch2_trans_commit(trans, NULL, NULL,
+				  BCH_TRANS_COMMIT_no_check_rw|
+				  BCH_TRANS_COMMIT_no_enospc|
+				  commit_flags);
+
+	bch2_fs_fatal_err_on(ret &&
+			     !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
+			     !bch2_err_matches(ret, BCH_ERR_journal_reclaim_would_deadlock) &&
+			     !bch2_journal_error(j), c,
+			     "flushing key cache: %s", bch2_err_str(ret));
+	if (ret)
+		goto out;
+
+	bch2_journal_pin_drop(j, &ck->journal);
+
+	struct btree_path *path = btree_iter_path(trans, &c_iter);
+	BUG_ON(!btree_node_locked(path, 0));
+
+	if (!evict) {
+		if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+			clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
+			atomic_long_dec(&c->btree_key_cache.nr_dirty);
+		}
+	} else {
+		struct btree_path *path2;
+		unsigned i;
+evict:
+		trans_for_each_path(trans, path2, i)
+			if (path2 != path)
+				__bch2_btree_path_unlock(trans, path2);
+
+		bch2_btree_node_lock_write_nofail(trans, path, &ck->c);
+
+		if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+			clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
+			atomic_long_dec(&c->btree_key_cache.nr_dirty);
+		}
+
+		mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
+		if (bkey_cached_evict(&c->btree_key_cache, ck)) {
+			bkey_cached_free(&c->btree_key_cache, ck);
+		} else {
+			six_unlock_write(&ck->c.lock);
+			six_unlock_intent(&ck->c.lock);
+		}
+	}
+out:
+	bch2_trans_iter_exit(trans, &b_iter);
+	bch2_trans_iter_exit(trans, &c_iter);
+	return ret;
+}
+
+int bch2_btree_key_cache_journal_flush(struct journal *j,
+				struct journal_entry_pin *pin, u64 seq)
+{
+	struct bch_fs *c = container_of(j, struct bch_fs, journal);
+	struct bkey_cached *ck =
+		container_of(pin, struct bkey_cached, journal);
+	struct bkey_cached_key key;
+	struct btree_trans *trans = bch2_trans_get(c);
+	int srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
+	int ret = 0;
+
+	btree_node_lock_nopath_nofail(trans, &ck->c, SIX_LOCK_read);
+	key = ck->key;
+
+	if (ck->journal.seq != seq ||
+	    !test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+		six_unlock_read(&ck->c.lock);
+		goto unlock;
+	}
+
+	if (ck->seq != seq) {
+		bch2_journal_pin_update(&c->journal, ck->seq, &ck->journal,
+					bch2_btree_key_cache_journal_flush);
+		six_unlock_read(&ck->c.lock);
+		goto unlock;
+	}
+	six_unlock_read(&ck->c.lock);
+
+	ret = lockrestart_do(trans,
+		btree_key_cache_flush_pos(trans, key, seq,
+				BCH_TRANS_COMMIT_journal_reclaim, false));
+unlock:
+	srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
+
+	bch2_trans_put(trans);
+	return ret;
+}
+
+bool bch2_btree_insert_key_cached(struct btree_trans *trans,
+				  unsigned flags,
+				  struct btree_insert_entry *insert_entry)
+{
+	struct bch_fs *c = trans->c;
+	struct bkey_cached *ck = (void *) (trans->paths + insert_entry->path)->l[0].b;
+	struct bkey_i *insert = insert_entry->k;
+	bool kick_reclaim = false;
+
+	BUG_ON(insert->k.u64s > ck->u64s);
+
+	bkey_copy(ck->k, insert);
+
+	if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+		EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
+		set_bit(BKEY_CACHED_DIRTY, &ck->flags);
+		atomic_long_inc(&c->btree_key_cache.nr_dirty);
+
+		if (bch2_nr_btree_keys_need_flush(c))
+			kick_reclaim = true;
+	}
+
+	/*
+	 * To minimize lock contention, we only add the journal pin here and
+	 * defer pin updates to the flush callback via ->seq. Be careful not to
+	 * update ->seq on nojournal commits because we don't want to update the
+	 * pin to a seq that doesn't include journal updates on disk. Otherwise
+	 * we risk losing the update after a crash.
+	 *
+	 * The only exception is if the pin is not active in the first place. We
+	 * have to add the pin because journal reclaim drives key cache
+	 * flushing. The flush callback will not proceed unless ->seq matches
+	 * the latest pin, so make sure it starts with a consistent value.
+	 */
+	if (!(insert_entry->flags & BTREE_UPDATE_nojournal) ||
+	    !journal_pin_active(&ck->journal)) {
+		ck->seq = trans->journal_res.seq;
+	}
+	bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
+			     &ck->journal, bch2_btree_key_cache_journal_flush);
+
+	if (kick_reclaim)
+		journal_reclaim_kick(&c->journal);
+	return true;
+}
+
+void bch2_btree_key_cache_drop(struct btree_trans *trans,
+			       struct btree_path *path)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_key_cache *bc = &c->btree_key_cache;
+	struct bkey_cached *ck = (void *) path->l[0].b;
+
+	/*
+	 * We just did an update to the btree, bypassing the key cache: the key
+	 * cache key is now stale and must be dropped, even if dirty:
+	 */
+	if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+		clear_bit(BKEY_CACHED_DIRTY, &ck->flags);
+		atomic_long_dec(&c->btree_key_cache.nr_dirty);
+		bch2_journal_pin_drop(&c->journal, &ck->journal);
+	}
+
+	bkey_cached_evict(bc, ck);
+	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;
+}
+
+static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
+					   struct shrink_control *sc)
+{
+	struct bch_fs *c = shrink->private_data;
+	struct btree_key_cache *bc = &c->btree_key_cache;
+	struct bucket_table *tbl;
+	struct bkey_cached *ck;
+	size_t scanned = 0, freed = 0, nr = sc->nr_to_scan;
+	unsigned iter, start;
+	int srcu_idx;
+
+	srcu_idx = srcu_read_lock(&c->btree_trans_barrier);
+	rcu_read_lock();
+
+	tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
+
+	/*
+	 * Scanning is expensive while a rehash is in progress - most elements
+	 * will be on the new hashtable, if it's in progress
+	 *
+	 * A rehash could still start while we're scanning - that's ok, we'll
+	 * still see most elements.
+	 */
+	if (unlikely(tbl->nest)) {
+		rcu_read_unlock();
+		srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
+		return SHRINK_STOP;
+	}
+
+	iter = bc->shrink_iter;
+	if (iter >= tbl->size)
+		iter = 0;
+	start = iter;
+
+	do {
+		struct rhash_head *pos, *next;
+
+		pos = rht_ptr_rcu(&tbl->buckets[iter]);
+
+		while (!rht_is_a_nulls(pos)) {
+			next = rht_dereference_bucket_rcu(pos->next, tbl, iter);
+			ck = container_of(pos, struct bkey_cached, hash);
+
+			if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
+				bc->skipped_dirty++;
+			} else if (test_bit(BKEY_CACHED_ACCESSED, &ck->flags)) {
+				clear_bit(BKEY_CACHED_ACCESSED, &ck->flags);
+				bc->skipped_accessed++;
+			} else if (!bkey_cached_lock_for_evict(ck)) {
+				bc->skipped_lock_fail++;
+			} else if (bkey_cached_evict(bc, ck)) {
+				bkey_cached_free(bc, ck);
+				bc->freed++;
+				freed++;
+			} else {
+				six_unlock_write(&ck->c.lock);
+				six_unlock_intent(&ck->c.lock);
+			}
+
+			scanned++;
+			if (scanned >= nr)
+				goto out;
+
+			pos = next;
+		}
+
+		iter++;
+		if (iter >= tbl->size)
+			iter = 0;
+	} while (scanned < nr && iter != start);
+out:
+	bc->shrink_iter = iter;
+
+	rcu_read_unlock();
+	srcu_read_unlock(&c->btree_trans_barrier, srcu_idx);
+
+	return freed;
+}
+
+static unsigned long bch2_btree_key_cache_count(struct shrinker *shrink,
+					    struct shrink_control *sc)
+{
+	struct bch_fs *c = shrink->private_data;
+	struct btree_key_cache *bc = &c->btree_key_cache;
+	long nr = atomic_long_read(&bc->nr_keys) -
+		atomic_long_read(&bc->nr_dirty);
+
+	/*
+	 * Avoid hammering our shrinker too much if it's nearly empty - the
+	 * shrinker code doesn't take into account how big our cache is, if it's
+	 * mostly empty but the system is under memory pressure it causes nasty
+	 * lock contention:
+	 */
+	nr -= 128;
+
+	return max(0L, nr);
+}
+
+void bch2_fs_btree_key_cache_exit(struct btree_key_cache *bc)
+{
+	struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
+	struct bucket_table *tbl;
+	struct bkey_cached *ck;
+	struct rhash_head *pos;
+	LIST_HEAD(items);
+	unsigned i;
+
+	shrinker_free(bc->shrink);
+
+	/*
+	 * The loop is needed to guard against racing with rehash:
+	 */
+	while (atomic_long_read(&bc->nr_keys)) {
+		rcu_read_lock();
+		tbl = rht_dereference_rcu(bc->table.tbl, &bc->table);
+		if (tbl) {
+			if (tbl->nest) {
+				/* wait for in progress rehash */
+				rcu_read_unlock();
+				mutex_lock(&bc->table.mutex);
+				mutex_unlock(&bc->table.mutex);
+				rcu_read_lock();
+				continue;
+			}
+			for (i = 0; i < tbl->size; i++)
+				while (pos = rht_ptr_rcu(&tbl->buckets[i]), !rht_is_a_nulls(pos)) {
+					ck = container_of(pos, struct bkey_cached, hash);
+					BUG_ON(!bkey_cached_evict(bc, ck));
+					kfree(ck->k);
+					kmem_cache_free(bch2_key_cache, ck);
+				}
+		}
+		rcu_read_unlock();
+	}
+
+	if (atomic_long_read(&bc->nr_dirty) &&
+	    !bch2_journal_error(&c->journal) &&
+	    test_bit(BCH_FS_was_rw, &c->flags))
+		panic("btree key cache shutdown error: nr_dirty nonzero (%li)\n",
+		      atomic_long_read(&bc->nr_dirty));
+
+	if (atomic_long_read(&bc->nr_keys))
+		panic("btree key cache shutdown error: nr_keys nonzero (%li)\n",
+		      atomic_long_read(&bc->nr_keys));
+
+	if (bc->table_init_done)
+		rhashtable_destroy(&bc->table);
+
+	rcu_pending_exit(&bc->pending[0]);
+	rcu_pending_exit(&bc->pending[1]);
+
+	free_percpu(bc->nr_pending);
+}
+
+void bch2_fs_btree_key_cache_init_early(struct btree_key_cache *c)
+{
+}
+
+int bch2_fs_btree_key_cache_init(struct btree_key_cache *bc)
+{
+	struct bch_fs *c = container_of(bc, struct bch_fs, btree_key_cache);
+	struct shrinker *shrink;
+
+	bc->nr_pending = alloc_percpu(size_t);
+	if (!bc->nr_pending)
+		return -BCH_ERR_ENOMEM_fs_btree_cache_init;
+
+	if (rcu_pending_init(&bc->pending[0], &c->btree_trans_barrier, __bkey_cached_free) ||
+	    rcu_pending_init(&bc->pending[1], &c->btree_trans_barrier, __bkey_cached_free))
+		return -BCH_ERR_ENOMEM_fs_btree_cache_init;
+
+	if (rhashtable_init(&bc->table, &bch2_btree_key_cache_params))
+		return -BCH_ERR_ENOMEM_fs_btree_cache_init;
+
+	bc->table_init_done = true;
+
+	shrink = shrinker_alloc(0, "%s-btree_key_cache", c->name);
+	if (!shrink)
+		return -BCH_ERR_ENOMEM_fs_btree_cache_init;
+	bc->shrink = shrink;
+	shrink->count_objects	= bch2_btree_key_cache_count;
+	shrink->scan_objects	= bch2_btree_key_cache_scan;
+	shrink->batch		= 1 << 14;
+	shrink->seeks		= 0;
+	shrink->private_data	= c;
+	shrinker_register(shrink);
+	return 0;
+}
+
+void bch2_btree_key_cache_to_text(struct printbuf *out, struct btree_key_cache *bc)
+{
+	printbuf_tabstop_push(out, 24);
+	printbuf_tabstop_push(out, 12);
+
+	prt_printf(out, "keys:\t%lu\r\n",		atomic_long_read(&bc->nr_keys));
+	prt_printf(out, "dirty:\t%lu\r\n",		atomic_long_read(&bc->nr_dirty));
+	prt_printf(out, "table size:\t%u\r\n",		bc->table.tbl->size);
+	prt_newline(out);
+	prt_printf(out, "shrinker:\n");
+	prt_printf(out, "requested_to_free:\t%lu\r\n",	bc->requested_to_free);
+	prt_printf(out, "freed:\t%lu\r\n",		bc->freed);
+	prt_printf(out, "skipped_dirty:\t%lu\r\n",	bc->skipped_dirty);
+	prt_printf(out, "skipped_accessed:\t%lu\r\n",	bc->skipped_accessed);
+	prt_printf(out, "skipped_lock_fail:\t%lu\r\n",	bc->skipped_lock_fail);
+	prt_newline(out);
+	prt_printf(out, "pending:\t%zu\r\n",		per_cpu_sum(bc->nr_pending));
+}
+
+void bch2_btree_key_cache_exit(void)
+{
+	kmem_cache_destroy(bch2_key_cache);
+}
+
+int __init bch2_btree_key_cache_init(void)
+{
+	bch2_key_cache = KMEM_CACHE(bkey_cached, SLAB_RECLAIM_ACCOUNT);
+	if (!bch2_key_cache)
+		return -ENOMEM;
+
+	return 0;
+}