Merge tag 'realtime-groups-6.13_2024-11-05' of https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux into staging-merge

xfs: shard the realtime section [v5.5 06/10]

Right now, the realtime section uses a single pair of metadata inodes to
store the free space information.  This presents a scalability problem
since every thread trying to allocate or free rt extents have to lock
these files.  Solve this problem by sharding the realtime section into
separate realtime allocation groups.

While we're at it, define a superblock to be stamped into the start of
the rt section.  This enables utilities such as blkid to identify block
devices containing realtime sections, and avoids the situation where
anything written into block 0 of the realtime extent can be
misinterpreted as file data.

The best advantage for rtgroups will become evident later when we get to
adding rmap and reflink to the realtime volume, since the geometry
constraints are the same for rt groups and AGs.  Hence we can reuse all
that code directly.

This is a very large patchset, but it catches us up with 20 years of
technical debt that have accumulated.

With a bit of luck, this should all go splendidly.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>

1 files changed, 519 insertions, 54 deletions

diff --git a/fs/xfs/xfs_rtalloc.c b/fs/xfs/xfs_rtalloc.c
index d9fb5e9c0aaf..7ecea7623a15 100644
--- a/fs/xfs/xfs_rtalloc.c
+++ b/fs/xfs/xfs_rtalloc.c
@@ -29,6 +29,7 @@
 #include "xfs_metafile.h"
 #include "xfs_rtgroup.h"
 #include "xfs_error.h"
+#include "xfs_trace.h"
 
 /*
  * Return whether there are any free extents in the size range given
@@ -745,12 +746,16 @@ xfs_growfs_rt_alloc_fake_mount(
 	nmp = kmemdup(mp, sizeof(*mp), GFP_KERNEL);
 	if (!nmp)
 		return NULL;
-	nmp->m_sb.sb_rextsize = rextsize;
-	xfs_mount_sb_set_rextsize(nmp, &nmp->m_sb);
+	xfs_mount_sb_set_rextsize(nmp, &nmp->m_sb, rextsize);
 	nmp->m_sb.sb_rblocks = rblocks;
 	nmp->m_sb.sb_rextents = xfs_blen_to_rtbxlen(nmp, nmp->m_sb.sb_rblocks);
 	nmp->m_sb.sb_rbmblocks = xfs_rtbitmap_blockcount(nmp);
 	nmp->m_sb.sb_rextslog = xfs_compute_rextslog(nmp->m_sb.sb_rextents);
+	if (xfs_has_rtgroups(nmp))
+		nmp->m_sb.sb_rgcount = howmany_64(nmp->m_sb.sb_rextents,
+						  nmp->m_sb.sb_rgextents);
+	else
+		nmp->m_sb.sb_rgcount = 1;
 	nmp->m_rsumblocks = xfs_rtsummary_blockcount(nmp, &nmp->m_rsumlevels);
 
 	if (rblocks > 0)
@@ -761,6 +766,88 @@ xfs_growfs_rt_alloc_fake_mount(
 	return nmp;
 }
 
+/* Free all the new space and return the number of extents actually freed. */
+static int
+xfs_growfs_rt_free_new(
+	struct xfs_rtgroup	*rtg,
+	struct xfs_rtalloc_args	*nargs,
+	xfs_rtbxlen_t		*freed_rtx)
+{
+	struct xfs_mount	*mp = rtg_mount(rtg);
+	xfs_rgnumber_t		rgno = rtg_rgno(rtg);
+	xfs_rtxnum_t		start_rtx = 0, end_rtx;
+
+	if (rgno < mp->m_sb.sb_rgcount)
+		start_rtx = xfs_rtgroup_extents(mp, rgno);
+	end_rtx = xfs_rtgroup_extents(nargs->mp, rgno);
+
+	/*
+	 * Compute the first new extent that we want to free, being careful to
+	 * skip past a realtime superblock at the start of the realtime volume.
+	 */
+	if (xfs_has_rtsb(nargs->mp) && rgno == 0 && start_rtx == 0)
+		start_rtx++;
+	*freed_rtx = end_rtx - start_rtx;
+	return xfs_rtfree_range(nargs, start_rtx, *freed_rtx);
+}
+
+static xfs_rfsblock_t
+xfs_growfs_rt_nrblocks(
+	struct xfs_rtgroup	*rtg,
+	xfs_rfsblock_t		nrblocks,
+	xfs_agblock_t		rextsize,
+	xfs_fileoff_t		bmbno)
+{
+	struct xfs_mount	*mp = rtg_mount(rtg);
+	xfs_rfsblock_t		step;
+
+	step = (bmbno + 1) * mp->m_rtx_per_rbmblock * rextsize;
+	if (xfs_has_rtgroups(mp)) {
+		xfs_rfsblock_t	rgblocks = mp->m_sb.sb_rgextents * rextsize;
+
+		step = min(rgblocks, step) + rgblocks * rtg_rgno(rtg);
+	}
+
+	return min(nrblocks, step);
+}
+
+/*
+ * If the post-grow filesystem will have an rtsb; we're initializing the first
+ * rtgroup; and the filesystem didn't have a realtime section, write the rtsb
+ * now, and attach the rtsb buffer to the real mount.
+ */
+static int
+xfs_growfs_rt_init_rtsb(
+	const struct xfs_rtalloc_args	*nargs,
+	const struct xfs_rtgroup	*rtg,
+	const struct xfs_rtalloc_args	*args)
+{
+	struct xfs_mount		*mp = args->mp;
+	struct xfs_buf			*rtsb_bp;
+	int				error;
+
+	if (!xfs_has_rtsb(nargs->mp))
+		return 0;
+	if (rtg_rgno(rtg) > 0)
+		return 0;
+	if (mp->m_sb.sb_rblocks)
+		return 0;
+
+	error = xfs_buf_get_uncached(mp->m_rtdev_targp, XFS_FSB_TO_BB(mp, 1),
+			0, &rtsb_bp);
+	if (error)
+		return error;
+
+	rtsb_bp->b_maps[0].bm_bn = XFS_RTSB_DADDR;
+	rtsb_bp->b_ops = &xfs_rtsb_buf_ops;
+
+	xfs_update_rtsb(rtsb_bp, mp->m_sb_bp);
+	mp->m_rtsb_bp = rtsb_bp;
+	error = xfs_bwrite(rtsb_bp);
+	xfs_buf_unlock(rtsb_bp);
+	return error;
+}
+
 static int
 xfs_growfs_rt_bmblock(
 	struct xfs_rtgroup	*rtg,
@@ -779,20 +866,21 @@ xfs_growfs_rt_bmblock(
 		.rtg		= rtg,
 	};
 	struct xfs_mount	*nmp;
-	xfs_rfsblock_t		nrblocks_step;
 	xfs_rtbxlen_t		freed_rtx;
 	int			error;
 
 	/*
-	 * Calculate new sb and mount fields for this round.
+	 * Calculate new sb and mount fields for this round.  Also ensure the
+	 * rtg_extents value is uptodate as the rtbitmap code relies on it.
 	 */
-	nrblocks_step = (bmbno + 1) * NBBY * mp->m_sb.sb_blocksize * rextsize;
 	nmp = nargs.mp = xfs_growfs_rt_alloc_fake_mount(mp,
-			min(nrblocks, nrblocks_step), rextsize);
+			xfs_growfs_rt_nrblocks(rtg, nrblocks, rextsize, bmbno),
+			rextsize);
 	if (!nmp)
 		return -ENOMEM;
 
-	rtg->rtg_extents = xfs_rtgroup_extents(nmp, rtg_rgno(rtg));
+	xfs_rtgroup_calc_geometry(nmp, rtg, rtg_rgno(rtg),
+			nmp->m_sb.sb_rgcount, nmp->m_sb.sb_rextents);
 
 	/*
 	 * Recompute the growfsrt reservation from the new rsumsize, so that the
@@ -837,6 +925,10 @@ xfs_growfs_rt_bmblock(
 			goto out_cancel;
 	}
 
+	error = xfs_growfs_rt_init_rtsb(&nargs, rtg, &args);
+	if (error)
+		goto out_cancel;
+
 	/*
 	 * Update superblock fields.
 	 */
@@ -855,12 +947,14 @@ xfs_growfs_rt_bmblock(
 	if (nmp->m_sb.sb_rextslog != mp->m_sb.sb_rextslog)
 		xfs_trans_mod_sb(args.tp, XFS_TRANS_SB_REXTSLOG,
 			nmp->m_sb.sb_rextslog - mp->m_sb.sb_rextslog);
+	if (nmp->m_sb.sb_rgcount != mp->m_sb.sb_rgcount)
+		xfs_trans_mod_sb(args.tp, XFS_TRANS_SB_RGCOUNT,
+			nmp->m_sb.sb_rgcount - mp->m_sb.sb_rgcount);
 
 	/*
 	 * Free the new extent.
 	 */
-	freed_rtx = nmp->m_sb.sb_rextents - mp->m_sb.sb_rextents;
-	error = xfs_rtfree_range(&nargs, mp->m_sb.sb_rextents, freed_rtx);
+	error = xfs_growfs_rt_free_new(rtg, &nargs, &freed_rtx);
 	xfs_rtbuf_cache_relse(&nargs);
 	if (error)
 		goto out_cancel;
@@ -875,7 +969,6 @@ xfs_growfs_rt_bmblock(
 	 */
 	mp->m_rsumlevels = nmp->m_rsumlevels;
 	mp->m_rsumblocks = nmp->m_rsumblocks;
-	xfs_mount_sb_set_rextsize(mp, &mp->m_sb);
 
 	/*
 	 * Recompute the growfsrt reservation from the new rsumsize.
@@ -901,6 +994,15 @@ out_free:
 	return error;
 }
 
+static xfs_rtxnum_t
+xfs_last_rtgroup_extents(
+	struct xfs_mount	*mp)
+{
+	return mp->m_sb.sb_rextents -
+		((xfs_rtxnum_t)(mp->m_sb.sb_rgcount - 1) *
+		 mp->m_sb.sb_rgextents);
+}
+
 /*
  * Calculate the last rbmblock currently used.
  *
@@ -911,11 +1013,20 @@ xfs_last_rt_bmblock(
 	struct xfs_rtgroup	*rtg)
 {
 	struct xfs_mount	*mp = rtg_mount(rtg);
-	xfs_fileoff_t		bmbno = mp->m_sb.sb_rbmblocks;
+	xfs_rgnumber_t		rgno = rtg_rgno(rtg);
+	xfs_fileoff_t		bmbno = 0;
+
+	ASSERT(!mp->m_sb.sb_rgcount || rgno >= mp->m_sb.sb_rgcount - 1);
+
+	if (mp->m_sb.sb_rgcount && rgno == mp->m_sb.sb_rgcount - 1) {
+		xfs_rtxnum_t	nrext = xfs_last_rtgroup_extents(mp);
+
+		/* Also fill up the previous block if not entirely full. */
+		bmbno = xfs_rtbitmap_blockcount_len(mp, nrext);
+		if (xfs_rtx_to_rbmword(mp, nrext) != 0)
+			bmbno--;
+	}
 
-	/* Skip the current block if it is exactly full. */
-	if (xfs_rtx_to_rbmword(mp, mp->m_sb.sb_rextents) != 0)
-		bmbno--;
 	return bmbno;
 }
 
@@ -932,38 +1043,56 @@ xfs_growfs_rt_alloc_blocks(
 	struct xfs_mount	*mp = rtg_mount(rtg);
 	struct xfs_inode	*rbmip = rtg->rtg_inodes[XFS_RTGI_BITMAP];
 	struct xfs_inode	*rsumip = rtg->rtg_inodes[XFS_RTGI_SUMMARY];
-	xfs_extlen_t		orbmblocks;
-	xfs_extlen_t		orsumblocks;
-	xfs_extlen_t		nrsumblocks;
+	xfs_extlen_t		orbmblocks = 0;
+	xfs_extlen_t		orsumblocks = 0;
 	struct xfs_mount	*nmp;
-	int			error;
-
-	/*
-	 * Get the old block counts for bitmap and summary inodes.
-	 * These can't change since other growfs callers are locked out.
-	 */
-	orbmblocks = XFS_B_TO_FSB(mp, rbmip->i_disk_size);
-	orsumblocks = XFS_B_TO_FSB(mp, rsumip->i_disk_size);
+	int			error = 0;
 
 	nmp = xfs_growfs_rt_alloc_fake_mount(mp, nrblocks, rextsize);
 	if (!nmp)
 		return -ENOMEM;
-
 	*nrbmblocks = nmp->m_sb.sb_rbmblocks;
-	nrsumblocks = nmp->m_rsumblocks;
-	kfree(nmp);
+
+	if (xfs_has_rtgroups(mp)) {
+		/*
+		 * For file systems with the rtgroups feature, the RT bitmap and
+		 * summary are always fully allocated, which means that we never
+		 * need to grow the existing files.
+		 *
+		 * But we have to be careful to only fill the bitmap until the
+		 * end of the actually used range.
+		 */
+		if (rtg_rgno(rtg) == nmp->m_sb.sb_rgcount - 1)
+			*nrbmblocks = xfs_rtbitmap_blockcount_len(nmp,
+					xfs_last_rtgroup_extents(nmp));
+
+		if (mp->m_sb.sb_rgcount &&
+		    rtg_rgno(rtg) == mp->m_sb.sb_rgcount - 1)
+			goto out_free;
+	} else {
+		/*
+		 * Get the old block counts for bitmap and summary inodes.
+		 * These can't change since other growfs callers are locked out.
+		 */
+		orbmblocks = XFS_B_TO_FSB(mp, rbmip->i_disk_size);
+		orsumblocks = XFS_B_TO_FSB(mp, rsumip->i_disk_size);
+	}
 
 	error = xfs_rtfile_initialize_blocks(rtg, XFS_RTGI_BITMAP, orbmblocks,
-			*nrbmblocks, NULL);
+			nmp->m_sb.sb_rbmblocks, NULL);
 	if (error)
-		return error;
-	return xfs_rtfile_initialize_blocks(rtg, XFS_RTGI_SUMMARY, orsumblocks,
-			nrsumblocks, NULL);
+		goto out_free;
+	error = xfs_rtfile_initialize_blocks(rtg, XFS_RTGI_SUMMARY, orsumblocks,
+			nmp->m_rsumblocks, NULL);
+out_free:
+	kfree(nmp);
+	return error;
 }
 
 static int
 xfs_growfs_rtg(
 	struct xfs_mount	*mp,
+	xfs_rgnumber_t		rgno,
 	xfs_rfsblock_t		nrblocks,
 	xfs_agblock_t		rextsize)
 {
@@ -974,7 +1103,7 @@ xfs_growfs_rtg(
 	unsigned int		i;
 	int			error;
 
-	rtg = xfs_rtgroup_grab(mp, 0);
+	rtg = xfs_rtgroup_grab(mp, rgno);
 	if (!rtg)
 		return -EINVAL;
 
@@ -1010,7 +1139,8 @@ out_error:
 	/*
 	 * Reset rtg_extents to the old value if adding more blocks failed.
 	 */
-	rtg->rtg_extents = xfs_rtgroup_extents(rtg_mount(rtg), rtg_rgno(rtg));
+	xfs_rtgroup_calc_geometry(mp, rtg, rtg_rgno(rtg), mp->m_sb.sb_rgcount,
+			mp->m_sb.sb_rextents);
 	if (old_rsum_cache) {
 		kvfree(rtg->rtg_rsum_cache);
 		rtg->rtg_rsum_cache = old_rsum_cache;
@@ -1046,13 +1176,66 @@ xfs_growfs_check_rtgeom(
 }
 
 /*
+ * Compute the new number of rt groups and ensure that /rtgroups exists.
+ *
+ * Changing the rtgroup size is not allowed (even if the rt volume hasn't yet
+ * been initialized) because the userspace ABI doesn't support it.
+ */
+static int
+xfs_growfs_rt_prep_groups(
+	struct xfs_mount	*mp,
+	xfs_rfsblock_t		rblocks,
+	xfs_extlen_t		rextsize,
+	xfs_rgnumber_t		*new_rgcount)
+{
+	int			error;
+
+	*new_rgcount = howmany_64(rblocks, mp->m_sb.sb_rgextents * rextsize);
+	if (*new_rgcount > XFS_MAX_RGNUMBER)
+		return -EINVAL;
+
+	/* Make sure the /rtgroups dir has been created */
+	if (!mp->m_rtdirip) {
+		struct xfs_trans	*tp;
+
+		error = xfs_trans_alloc_empty(mp, &tp);
+		if (error)
+			return error;
+		error = xfs_rtginode_load_parent(tp);
+		xfs_trans_cancel(tp);
+
+		if (error == -ENOENT)
+			error = xfs_rtginode_mkdir_parent(mp);
+		if (error)
+			return error;
+	}
+
+	return 0;
+}
+
+static bool
+xfs_grow_last_rtg(
+	struct xfs_mount	*mp)
+{
+	if (!xfs_has_rtgroups(mp))
+		return true;
+	if (mp->m_sb.sb_rgcount == 0)
+		return false;
+	return xfs_rtgroup_extents(mp, mp->m_sb.sb_rgcount - 1) <=
+			mp->m_sb.sb_rgextents;
+}
+
+/*
  * Grow the realtime area of the filesystem.
  */
 int
 xfs_growfs_rt(
-	xfs_mount_t	*mp,		/* mount point for filesystem */
-	xfs_growfs_rt_t	*in)		/* growfs rt input struct */
+	struct xfs_mount	*mp,
+	struct xfs_growfs_rt	*in)
 {
+	xfs_rgnumber_t		old_rgcount = mp->m_sb.sb_rgcount;
+	xfs_rgnumber_t		new_rgcount = 1;
+	xfs_rgnumber_t		rgno;
 	struct xfs_buf		*bp;
 	xfs_agblock_t		old_rextsize = mp->m_sb.sb_rextsize;
 	int			error;
@@ -1110,24 +1293,112 @@ xfs_growfs_rt(
 	if (error)
 		goto out_unlock;
 
-	error = xfs_growfs_rtg(mp, in->newblocks, in->extsize);
-	if (error)
-		goto out_unlock;
+	if (xfs_has_rtgroups(mp)) {
+		error = xfs_growfs_rt_prep_groups(mp, in->newblocks,
+				in->extsize, &new_rgcount);
+		if (error)
+			goto out_unlock;
+	}
 
-	if (old_rextsize != in->extsize) {
-		error = xfs_growfs_rt_fixup_extsize(mp);
+	if (xfs_grow_last_rtg(mp)) {
+		error = xfs_growfs_rtg(mp, old_rgcount - 1, in->newblocks,
+				in->extsize);
 		if (error)
 			goto out_unlock;
 	}
 
-	/* Update secondary superblocks now the physical grow has completed */
-	error = xfs_update_secondary_sbs(mp);
+	for (rgno = old_rgcount; rgno < new_rgcount; rgno++) {
+		xfs_rtbxlen_t	rextents = div_u64(in->newblocks, in->extsize);
+
+		error = xfs_rtgroup_alloc(mp, rgno, new_rgcount, rextents);
+		if (error)
+			goto out_unlock;
+
+		error = xfs_growfs_rtg(mp, rgno, in->newblocks, in->extsize);
+		if (error) {
+			struct xfs_rtgroup	*rtg;
+
+			rtg = xfs_rtgroup_grab(mp, rgno);
+			if (!WARN_ON_ONCE(!rtg)) {
+				xfs_rtunmount_rtg(rtg);
+				xfs_rtgroup_rele(rtg);
+				xfs_rtgroup_free(mp, rgno);
+			}
+			break;
+		}
+	}
+
+	if (!error && old_rextsize != in->extsize)
+		error = xfs_growfs_rt_fixup_extsize(mp);
+
+	/*
+	 * Update secondary superblocks now the physical grow has completed.
+	 *
+	 * Also do this in case of an error as we might have already
+	 * successfully updated one or more RTGs and incremented sb_rgcount.
+	 */
+	if (!xfs_is_shutdown(mp)) {
+		int error2 = xfs_update_secondary_sbs(mp);
+
+		if (!error)
+			error = error2;
+	}
 
 out_unlock:
 	mutex_unlock(&mp->m_growlock);
 	return error;
 }
 
+/* Read the realtime superblock and attach it to the mount. */
+int
+xfs_rtmount_readsb(
+	struct xfs_mount	*mp)
+{
+	struct xfs_buf		*bp;
+	int			error;
+
+	if (!xfs_has_rtsb(mp))
+		return 0;
+	if (mp->m_sb.sb_rblocks == 0)
+		return 0;
+	if (mp->m_rtdev_targp == NULL) {
+		xfs_warn(mp,
+	"Filesystem has a realtime volume, use rtdev=device option");
+		return -ENODEV;
+	}
+
+	/* m_blkbb_log is not set up yet */
+	error = xfs_buf_read_uncached(mp->m_rtdev_targp, XFS_RTSB_DADDR,
+			mp->m_sb.sb_blocksize >> BBSHIFT, XBF_NO_IOACCT, &bp,
+			&xfs_rtsb_buf_ops);
+	if (error) {
+		xfs_warn(mp, "rt sb validate failed with error %d.", error);
+		/* bad CRC means corrupted metadata */
+		if (error == -EFSBADCRC)
+			error = -EFSCORRUPTED;
+		return error;
+	}
+
+	mp->m_rtsb_bp = bp;
+	xfs_buf_unlock(bp);
+	return 0;
+}
+
+/* Detach the realtime superblock from the mount and free it. */
+void
+xfs_rtmount_freesb(
+	struct xfs_mount	*mp)
+{
+	struct xfs_buf		*bp = mp->m_rtsb_bp;
+
+	if (!bp)
+		return;
+
+	xfs_buf_lock(bp);
+	mp->m_rtsb_bp = NULL;
+	xfs_buf_relse(bp);
+}
+
 /*
  * Initialize realtime fields in the mount structure.
  */
@@ -1251,8 +1522,6 @@ xfs_rtmount_rtg(
 {
 	int			error, i;
 
-	rtg->rtg_extents = xfs_rtgroup_extents(mp, rtg_rgno(rtg));
-
 	for (i = 0; i < XFS_RTGI_MAX; i++) {
 		error = xfs_rtginode_load(rtg, i, tp);
 		if (error)
@@ -1391,9 +1660,118 @@ xfs_rtalloc_align_minmax(
 	*raminlen = newminlen;
 }
 
+/* Given a free extent, find any part of it that isn't busy, if possible. */
+STATIC bool
+xfs_rtalloc_check_busy(
+	struct xfs_rtalloc_args	*args,
+	xfs_rtxnum_t		start,
+	xfs_rtxlen_t		minlen_rtx,
+	xfs_rtxlen_t		maxlen_rtx,
+	xfs_rtxlen_t		len_rtx,
+	xfs_rtxlen_t		prod,
+	xfs_rtxnum_t		rtx,
+	xfs_rtxlen_t		*reslen,
+	xfs_rtxnum_t		*resrtx,
+	unsigned		*busy_gen)
+{
+	struct xfs_rtgroup	*rtg = args->rtg;
+	struct xfs_mount	*mp = rtg_mount(rtg);
+	xfs_agblock_t		rgbno = xfs_rtx_to_rgbno(rtg, rtx);
+	xfs_rgblock_t		min_rgbno = xfs_rtx_to_rgbno(rtg, start);
+	xfs_extlen_t		minlen = xfs_rtxlen_to_extlen(mp, minlen_rtx);
+	xfs_extlen_t		len = xfs_rtxlen_to_extlen(mp, len_rtx);
+	xfs_extlen_t		diff;
+	bool			busy;
+
+	busy = xfs_extent_busy_trim(rtg_group(rtg), minlen,
+			xfs_rtxlen_to_extlen(mp, maxlen_rtx), &rgbno, &len,
+			busy_gen);
+
+	/*
+	 * If we have a largish extent that happens to start before min_rgbno,
+	 * see if we can shift it into range...
+	 */
+	if (rgbno < min_rgbno && rgbno + len > min_rgbno) {
+		diff = min_rgbno - rgbno;
+		if (len > diff) {
+			rgbno += diff;
+			len -= diff;
+		}
+	}
+
+	if (prod > 1 && len >= minlen) {
+		xfs_rgblock_t	aligned_rgbno = roundup(rgbno, prod);
+
+		diff = aligned_rgbno - rgbno;
+
+		*resrtx = xfs_rgbno_to_rtx(mp, aligned_rgbno);
+		*reslen = xfs_extlen_to_rtxlen(mp,
+				diff >= len ? 0 : len - diff);
+	} else {
+		*resrtx = xfs_rgbno_to_rtx(mp, rgbno);
+		*reslen = xfs_extlen_to_rtxlen(mp, len);
+	}
+
+	return busy;
+}
+
+/*
+ * Adjust the given free extent so that it isn't busy, or flush the log and
+ * wait for the space to become unbusy.  Only needed for rtgroups.
+ */
+STATIC int
+xfs_rtallocate_adjust_for_busy(
+	struct xfs_rtalloc_args	*args,
+	xfs_rtxnum_t		start,
+	xfs_rtxlen_t		minlen,
+	xfs_rtxlen_t		maxlen,
+	xfs_rtxlen_t		*len,
+	xfs_rtxlen_t		prod,
+	xfs_rtxnum_t		*rtx)
+{
+	xfs_rtxnum_t		resrtx;
+	xfs_rtxlen_t		reslen;
+	unsigned		busy_gen;
+	bool			busy;
+	int			error;
+
+again:
+	busy = xfs_rtalloc_check_busy(args, start, minlen, maxlen, *len, prod,
+			*rtx, &reslen, &resrtx, &busy_gen);
+	if (!busy)
+		return 0;
+
+	if (reslen < minlen || (start != 0 && resrtx != *rtx)) {
+		/*
+		 * Enough of the extent was busy that we cannot satisfy the
+		 * allocation, or this is a near allocation and the start of
+		 * the extent is busy.  Flush the log and wait for the busy
+		 * situation to resolve.
+		 */
+		trace_xfs_rtalloc_extent_busy(args->rtg, start, minlen, maxlen,
+				*len, prod, *rtx, busy_gen);
+
+		error = xfs_extent_busy_flush(args->tp, rtg_group(args->rtg),
+				busy_gen, 0);
+		if (error)
+			return error;
+
+		goto again;
+	}
+
+	/* Some of the free space wasn't busy, hand that back to the caller. */
+	trace_xfs_rtalloc_extent_busy_trim(args->rtg, *rtx, *len, resrtx,
+			reslen);
+	*len = reslen;
+	*rtx = resrtx;
+
+	return 0;
+}
+
 static int
-xfs_rtallocate(
+xfs_rtallocate_rtg(
 	struct xfs_trans	*tp,
+	xfs_rgnumber_t		rgno,
 	xfs_rtblock_t		bno_hint,
 	xfs_rtxlen_t		minlen,
 	xfs_rtxlen_t		maxlen,
@@ -1413,16 +1791,33 @@ xfs_rtallocate(
 	xfs_rtxlen_t		len = 0;
 	int			error = 0;
 
-	args.rtg = xfs_rtgroup_grab(args.mp, 0);
+	args.rtg = xfs_rtgroup_grab(args.mp, rgno);
 	if (!args.rtg)
 		return -ENOSPC;
 
 	/*
-	 * Lock out modifications to both the RT bitmap and summary inodes.
+	 * We need to lock out modifications to both the RT bitmap and summary
+	 * inodes for finding free space in xfs_rtallocate_extent_{near,size}
+	 * and join the bitmap and summary inodes for the actual allocation
+	 * down in xfs_rtallocate_range.
+	 *
+	 * For RTG-enabled file system we don't want to join the inodes to the
+	 * transaction until we are committed to allocate to allocate from this
+	 * RTG so that only one inode of each type is locked at a time.
+	 *
+	 * But for pre-RTG file systems we need to already to join the bitmap
+	 * inode to the transaction for xfs_rtpick_extent, which bumps the
+	 * sequence number in it, so we'll have to join the inode to the
+	 * transaction early here.
+	 *
+	 * This is all a bit messy, but at least the mess is contained in
+	 * this function.
 	 */
 	if (!*rtlocked) {
 		xfs_rtgroup_lock(args.rtg, XFS_RTGLOCK_BITMAP);
-		xfs_rtgroup_trans_join(tp, args.rtg, XFS_RTGLOCK_BITMAP);
+		if (!xfs_has_rtgroups(args.mp))
+			xfs_rtgroup_trans_join(tp, args.rtg,
+					XFS_RTGLOCK_BITMAP);
 		*rtlocked = true;
 	}
 
@@ -1432,7 +1827,7 @@ xfs_rtallocate(
 	 */
 	if (bno_hint)
 		start = xfs_rtb_to_rtx(args.mp, bno_hint);
-	else if (initial_user_data)
+	else if (!xfs_has_rtgroups(args.mp) && initial_user_data)
 		start = xfs_rtpick_extent(args.rtg, tp, maxlen);
 
 	if (start) {
@@ -1453,8 +1848,20 @@ xfs_rtallocate(
 				prod, &rtx);
 	}
 
-	if (error)
+	if (error) {
+		if (xfs_has_rtgroups(args.mp))
+			goto out_unlock;
 		goto out_release;
+	}
+
+	if (xfs_has_rtgroups(args.mp)) {
+		error = xfs_rtallocate_adjust_for_busy(&args, start, minlen,
+				maxlen, &len, prod, &rtx);
+		if (error)
+			goto out_unlock;
+
+		xfs_rtgroup_trans_join(tp, args.rtg, XFS_RTGLOCK_BITMAP);
+	}
 
 	error = xfs_rtallocate_range(&args, rtx, len);
 	if (error)
@@ -1470,6 +1877,57 @@ out_release:
 	xfs_rtgroup_rele(args.rtg);
 	xfs_rtbuf_cache_relse(&args);
 	return error;
+out_unlock:
+	xfs_rtgroup_unlock(args.rtg, XFS_RTGLOCK_BITMAP);
+	*rtlocked = false;
+	goto out_release;
+}
+
+static int
+xfs_rtallocate_rtgs(
+	struct xfs_trans	*tp,
+	xfs_fsblock_t		bno_hint,
+	xfs_rtxlen_t		minlen,
+	xfs_rtxlen_t		maxlen,
+	xfs_rtxlen_t		prod,
+	bool			wasdel,
+	bool			initial_user_data,
+	xfs_rtblock_t		*bno,
+	xfs_extlen_t		*blen)
+{
+	struct xfs_mount	*mp = tp->t_mountp;
+	xfs_rgnumber_t		start_rgno, rgno;
+	int			error;
+
+	/*
+	 * For now this just blindly iterates over the RTGs for an initial
+	 * allocation.  We could try to keep an in-memory rtg_longest member
+	 * to avoid the locking when just looking for big enough free space,
+	 * but for now this keeps things simple.
+	 */
+	if (bno_hint != NULLFSBLOCK)
+		start_rgno = xfs_rtb_to_rgno(mp, bno_hint);
+	else
+		start_rgno = (atomic_inc_return(&mp->m_rtgrotor) - 1) %
+				mp->m_sb.sb_rgcount;
+
+	rgno = start_rgno;
+	do {
+		bool		rtlocked = false;
+
+		error = xfs_rtallocate_rtg(tp, rgno, bno_hint, minlen, maxlen,
+				prod, wasdel, initial_user_data, &rtlocked,
+				bno, blen);
+		if (error != -ENOSPC)
+			return error;
+		ASSERT(!rtlocked);
+
+		if (++rgno == mp->m_sb.sb_rgcount)
+			rgno = 0;
+		bno_hint = NULLFSBLOCK;
+	} while (rgno != start_rgno);
+
+	return -ENOSPC;
 }
 
 static int
@@ -1566,9 +2024,16 @@ retry:
 	if (xfs_bmap_adjacent(ap))
 		bno_hint = ap->blkno;
 
-	error = xfs_rtallocate(ap->tp, bno_hint, raminlen, ralen, prod,
-			ap->wasdel, initial_user_data, &rtlocked,
-			&ap->blkno, &ap->length);
+	if (xfs_has_rtgroups(ap->ip->i_mount)) {
+		error = xfs_rtallocate_rtgs(ap->tp, bno_hint, raminlen, ralen,
+				prod, ap->wasdel, initial_user_data,
+				&ap->blkno, &ap->length);
+	} else {
+		error = xfs_rtallocate_rtg(ap->tp, 0, bno_hint, raminlen, ralen,
+				prod, ap->wasdel, initial_user_data,
+				&rtlocked, &ap->blkno, &ap->length);
+	}
+
 	if (error == -ENOSPC) {
 		if (!noalign) {
 			/*