[PATCH v2 3/3] xfs: incorporate increased AGFL min requirement for minleft allocs

[Brian Foster <bfoster@redhat.com>]

Matt Fleming reports a filesystem shutdown due to inobt block
allocation failure during sparse chunk allocation. Inode creation
can involve multiple allocations in a transaction: the initial chunk
allocation and inode btree blocks via inobt record insertion. This
is expected to be safe by using the minleft parameter on the chunk
allocation to guarantee the selected AG has blocks available for
a followup inobt insertion.

The sequence that leads to this failure is that the alloc and inode
btrees are all full (require a full split on next insertion) and the
AG has just enough free space to satisfy a sparse chunk allocation
with minleft set (i.e. 7 blocks in this example). The chunk
allocation splits a free extent, triggers full allocbt splits, and
consumes 4 free blocks for the chunk and 4 AGFL blocks for the
btrees.

Next, the inobt record insertion triggers an inobt split. The AG has
enough free blocks, but the allocbt splits caused by the chunk
allocation have increased the min AGFL requirement for the AG due to
btree level increases. The AGFL requirement as calculated by
xfs_alloc_fix_freelist() is:

	free + AGFL - res - minfree - minleft = avail

This evaluates to the following on initial chunk allocation:

	2514 + 8 - 2505 - 8 - 2 = 7

... and then after the chunk allocation but before the inobt block
allocation:

	2510 + 4 - 2505 - 12 - 0 = -3

This causes the inobt alloc to fail despite minleft being set in the
first allocation. The error path cancels the dirty transaction and
shuts down the fs.

The problem here is that while minleft ensures free blocks are
available for the inobt insert, it is not sufficient to cover the
increase of the AGFL min free requirement. To address this, create a
variant of the AGFL min free calculation for minleft allocations
that incorporates an additional allocbt level increase.

We do not add the additional blocks directly to min_free because
this would lead to spurious AGFL block allocations and frees in the
common case (i.e. no btree splits). Instead, add the surplus block
requirement to the minleft value used to select the AG. This ensures
the AG has enough blocks for the caller's minleft value plus the
worst case increase in the AGFL. In the example above, the initial
calculation now evaluates to 3 blocks available instead of 7 and the
sparse inode allocation fails gracefully with -ENOSPC.

Reported-by: Matt Fleming <matt@readmodwrite.com>
Assisted-by: LLM
Signed-off-by: Brian Foster <bfoster@redhat.com>
---
 fs/xfs/libxfs/xfs_alloc.c | 29 ++++++++++++++++++++++++++++-
 fs/xfs/libxfs/xfs_alloc.h |  2 ++
 fs/xfs/libxfs/xfs_bmap.c  |  2 +-
 3 files changed, 31 insertions(+), 2 deletions(-)

diff --git a/fs/xfs/libxfs/xfs_alloc.c b/fs/xfs/libxfs/xfs_alloc.c
index dbb85fb6314b..74c5b587c87b 100644
--- a/fs/xfs/libxfs/xfs_alloc.c
+++ b/fs/xfs/libxfs/xfs_alloc.c
@@ -2500,6 +2500,20 @@ xfs_alloc_min_freelist(
 	return __xfs_alloc_min_freelist(mp, pag, 0);
 }
 
+/*
+ * Return the minimum freelist requirement considering a potential allocbt split
+ * from the current allocation. Use this when computing longest free extent for
+ * allocations with minleft set to ensure that the available extent length
+ * accounts for the subsequent allocation's increased AGFL requirement.
+ */
+unsigned int
+xfs_alloc_min_freelist_minleft(
+	struct xfs_mount	*mp,
+	struct xfs_perag	*pag)
+{
+	return __xfs_alloc_min_freelist(mp, pag, 1);
+}
+
 /*
  * Check if the operation we are fixing up the freelist for should go ahead or
  * not. If we are freeing blocks, we always allow it, otherwise the allocation
@@ -2517,6 +2531,7 @@ xfs_alloc_space_available(
 	xfs_extlen_t		reservation; /* blocks that are still reserved */
 	int			available;
 	xfs_extlen_t		agflcount;
+	xfs_extlen_t		minleft;
 
 	if (flags & XFS_ALLOC_FLAG_FREEING)
 		return true;
@@ -2533,10 +2548,22 @@ xfs_alloc_space_available(
 	 * Do we have enough free space remaining for the allocation? Don't
 	 * account extra agfl blocks because we are about to defer free them,
 	 * making them unavailable until the current transaction commits.
+	 *
+	 * If minleft is set, this allocation might cause an allocbt split that
+	 * increases the AGFL minimum for the next allocation in the
+	 * transaction. Reserve that space from the available block count
+	 * (without prematurely growing the AGFL) to prevent the subsequent
+	 * allocation from failing due to an increased min_free requirement.
 	 */
+	minleft = args->minleft;
+	if (minleft) {
+		minleft += xfs_alloc_min_freelist_minleft(args->mp, pag) -
+					min_free;
+	}
+
 	agflcount = min_t(xfs_extlen_t, pag->pagf_flcount, min_free);
 	available = (int)(pag->pagf_freeblks + agflcount -
-			  reservation - min_free - args->minleft);
+			  reservation - min_free - minleft);
 	if (available < (int)max(args->total, alloc_len))
 		return false;
 
diff --git a/fs/xfs/libxfs/xfs_alloc.h b/fs/xfs/libxfs/xfs_alloc.h
index 50ef79a1ed41..026b61a63994 100644
--- a/fs/xfs/libxfs/xfs_alloc.h
+++ b/fs/xfs/libxfs/xfs_alloc.h
@@ -73,6 +73,8 @@ xfs_extlen_t xfs_alloc_longest_free_extent(struct xfs_perag *pag,
 		xfs_extlen_t need, xfs_extlen_t reserved);
 unsigned int xfs_alloc_min_freelist(struct xfs_mount *mp,
 		struct xfs_perag *pag);
+unsigned int xfs_alloc_min_freelist_minleft(struct xfs_mount *mp,
+		struct xfs_perag *pag);
 int xfs_alloc_get_freelist(struct xfs_perag *pag, struct xfs_trans *tp,
 		struct xfs_buf *agfbp, xfs_agblock_t *bnop, int	 btreeblk);
 int xfs_alloc_put_freelist(struct xfs_perag *pag, struct xfs_trans *tp,
diff --git a/fs/xfs/libxfs/xfs_bmap.c b/fs/xfs/libxfs/xfs_bmap.c
index d64defeda645..f396df864cf4 100644
--- a/fs/xfs/libxfs/xfs_bmap.c
+++ b/fs/xfs/libxfs/xfs_bmap.c
@@ -3160,7 +3160,7 @@ xfs_bmap_longest_free_extent(
 	}
 
 	longest = xfs_alloc_longest_free_extent(pag,
-				xfs_alloc_min_freelist(pag_mount(pag), pag),
+				xfs_alloc_min_freelist_minleft(pag_mount(pag), pag),
 				xfs_ag_resv_needed(pag, XFS_AG_RESV_NONE));
 	if (*blen < longest)
 		*blen = longest;
-- 
2.55.0