xfs: preallocate blocks for worst-case btree expansion

To gracefully handle the situation where a CoW operation turns a
single refcount extent into a lot of tiny ones and then run out of
space when a tree split has to happen, use the per-AG reserved block
pool to pre-allocate all the space we'll ever need for a maximal
btree.  For a 4K block size, this only costs an overhead of 0.3% of
available disk space.

When reflink is enabled, we have an unfortunate problem with rmap --
since we can share a block billions of times, this means that the
reverse mapping btree can expand basically infinitely.  When an AG is
so full that there are no free blocks with which to expand the rmapbt,
the filesystem will shut down hard.

This is rather annoying to the user, so use the AG reservation code to
reserve a "reasonable" amount of space for rmap.  We'll prevent
reflinks and CoW operations if we think we're getting close to
exhausting an AG's free space rather than shutting down, but this
permanent reservation should be enough for "most" users.  Hopefully.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
[hch@lst.de: ensure that we invalidate the freed btree buffer]
Signed-off-by: Christoph Hellwig <hch@lst.de>

1 files changed, 65 insertions, 3 deletions

diff --git a/fs/xfs/libxfs/xfs_refcount_btree.c b/fs/xfs/libxfs/xfs_refcount_btree.c
index 81d58b09af71..453bb2757ec2 100644
--- a/fs/xfs/libxfs/xfs_refcount_btree.c
+++ b/fs/xfs/libxfs/xfs_refcount_btree.c
@@ -79,6 +79,8 @@ xfs_refcountbt_alloc_block(
 	struct xfs_alloc_arg	args;		/* block allocation args */
 	int			error;		/* error return value */
 
+	XFS_BTREE_TRACE_CURSOR(cur, XBT_ENTRY);
+
 	memset(&args, 0, sizeof(args));
 	args.tp = cur->bc_tp;
 	args.mp = cur->bc_mp;
@@ -88,6 +90,7 @@ xfs_refcountbt_alloc_block(
 	args.firstblock = args.fsbno;
 	xfs_rmap_ag_owner(&args.oinfo, XFS_RMAP_OWN_REFC);
 	args.minlen = args.maxlen = args.prod = 1;
+	args.resv = XFS_AG_RESV_METADATA;
 
 	error = xfs_alloc_vextent(&args);
 	if (error)
@@ -125,16 +128,19 @@ xfs_refcountbt_free_block(
 	struct xfs_agf		*agf = XFS_BUF_TO_AGF(agbp);
 	xfs_fsblock_t		fsbno = XFS_DADDR_TO_FSB(mp, XFS_BUF_ADDR(bp));
 	struct xfs_owner_info	oinfo;
+	int			error;
 
 	trace_xfs_refcountbt_free_block(cur->bc_mp, cur->bc_private.a.agno,
 			XFS_FSB_TO_AGBNO(cur->bc_mp, fsbno), 1);
 	xfs_rmap_ag_owner(&oinfo, XFS_RMAP_OWN_REFC);
 	be32_add_cpu(&agf->agf_refcount_blocks, -1);
 	xfs_alloc_log_agf(cur->bc_tp, agbp, XFS_AGF_REFCOUNT_BLOCKS);
-	xfs_bmap_add_free(mp, cur->bc_private.a.dfops, fsbno, 1,
-			&oinfo);
+	error = xfs_free_extent(cur->bc_tp, fsbno, 1, &oinfo,
+			XFS_AG_RESV_METADATA);
+	if (error)
+		return error;
 
-	return 0;
+	return error;
 }
 
 STATIC int
@@ -387,3 +393,59 @@ xfs_refcountbt_compute_maxlevels(
 	mp->m_refc_maxlevels = xfs_btree_compute_maxlevels(mp,
 			mp->m_refc_mnr, mp->m_sb.sb_agblocks);
 }
+
+/* Calculate the refcount btree size for some records. */
+xfs_extlen_t
+xfs_refcountbt_calc_size(
+	struct xfs_mount	*mp,
+	unsigned long long	len)
+{
+	return xfs_btree_calc_size(mp, mp->m_refc_mnr, len);
+}
+
+/*
+ * Calculate the maximum refcount btree size.
+ */
+xfs_extlen_t
+xfs_refcountbt_max_size(
+	struct xfs_mount	*mp)
+{
+	/* Bail out if we're uninitialized, which can happen in mkfs. */
+	if (mp->m_refc_mxr[0] == 0)
+		return 0;
+
+	return xfs_refcountbt_calc_size(mp, mp->m_sb.sb_agblocks);
+}
+
+/*
+ * Figure out how many blocks to reserve and how many are used by this btree.
+ */
+int
+xfs_refcountbt_calc_reserves(
+	struct xfs_mount	*mp,
+	xfs_agnumber_t		agno,
+	xfs_extlen_t		*ask,
+	xfs_extlen_t		*used)
+{
+	struct xfs_buf		*agbp;
+	struct xfs_agf		*agf;
+	xfs_extlen_t		tree_len;
+	int			error;
+
+	if (!xfs_sb_version_hasreflink(&mp->m_sb))
+		return 0;
+
+	*ask += xfs_refcountbt_max_size(mp);
+
+	error = xfs_alloc_read_agf(mp, NULL, agno, 0, &agbp);
+	if (error)
+		return error;
+
+	agf = XFS_BUF_TO_AGF(agbp);
+	tree_len = be32_to_cpu(agf->agf_refcount_blocks);
+	xfs_buf_relse(agbp);
+
+	*used += tree_len;
+
+	return error;
+}