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, 7 insertions, 0 deletions

diff --git a/fs/xfs/libxfs/xfs_refcount_btree.h b/fs/xfs/libxfs/xfs_refcount_btree.h
index 9e9ad7c6d267..3be7768bd51a 100644
--- a/fs/xfs/libxfs/xfs_refcount_btree.h
+++ b/fs/xfs/libxfs/xfs_refcount_btree.h
@@ -64,4 +64,11 @@ extern int xfs_refcountbt_maxrecs(struct xfs_mount *mp, int blocklen,
 		bool leaf);
 extern void xfs_refcountbt_compute_maxlevels(struct xfs_mount *mp);
 
+extern xfs_extlen_t xfs_refcountbt_calc_size(struct xfs_mount *mp,
+		unsigned long long len);
+extern xfs_extlen_t xfs_refcountbt_max_size(struct xfs_mount *mp);
+
+extern int xfs_refcountbt_calc_reserves(struct xfs_mount *mp,
+		xfs_agnumber_t agno, xfs_extlen_t *ask, xfs_extlen_t *used);
+
 #endif	/* __XFS_REFCOUNT_BTREE_H__ */