xfs: don't free EOF blocks on read close

When we have a workload that does open/read/close in parallel with other
allocation, the file becomes rapidly fragmented. This is due to close()
calling xfs_release() and removing the speculative preallocation beyond
EOF.

The existing open/*/close heuristic in xfs_release() does not catch this
as a sync writer does not leave delayed allocation blocks allocated on
the inode for later writeback that can be detected in xfs_release() and
hence XFS_IDIRTY_RELEASE never gets set.

In xfs_file_release(), we know more about the released file context, and
so we need to communicate some of the details to xfs_release() so it can
do the right thing here and skip EOF block truncation. This defers the
EOF block cleanup for synchronous write contexts to the background EOF
block cleaner which will clean up within a few minutes.

Before:

Test 1: sync write fragmentation counts

/mnt/scratch/file.0: 919
/mnt/scratch/file.1: 916
/mnt/scratch/file.2: 919
/mnt/scratch/file.3: 920
/mnt/scratch/file.4: 920
/mnt/scratch/file.5: 921
/mnt/scratch/file.6: 916
/mnt/scratch/file.7: 918

After:

Test 1: sync write fragmentation counts

/mnt/scratch/file.0: 24
/mnt/scratch/file.1: 24
/mnt/scratch/file.2: 11
/mnt/scratch/file.3: 24
/mnt/scratch/file.4: 3
/mnt/scratch/file.5: 24
/mnt/scratch/file.6: 24
/mnt/scratch/file.7: 23

Signed-off-by: Dave Chinner <dchinner@redhat.com>
[darrick: wordsmithing, fix commit message]
Signed-off-by: Darrick J. Wong <djwong@kernel.org>

3 files changed, 18 insertions, 7 deletions

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index cea3c14e3ba0..11f57c73f748 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -1413,12 +1413,22 @@ xfs_dir_open(
 	return error;
 }
 
+/*
+ * When we release the file, we don't want it to trim EOF blocks if it is a
+ * readonly context.  This avoids open/read/close workloads from removing
+ * EOF blocks that other writers depend upon to reduce fragmentation.
+ */
 STATIC int
 xfs_file_release(
 	struct inode	*inode,
-	struct file	*filp)
+	struct file	*file)
 {
-	return xfs_release(XFS_I(inode));
+	bool		free_eof_blocks = true;
+
+	if ((file->f_mode & (FMODE_WRITE | FMODE_READ)) == FMODE_READ)
+		free_eof_blocks = false;
+
+	return xfs_release(XFS_I(inode), free_eof_blocks);
 }
 
 STATIC int
diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
index f0d63dab19e1..80f9fb2a7799 100644
--- a/fs/xfs/xfs_inode.c
+++ b/fs/xfs/xfs_inode.c
@@ -1198,10 +1198,11 @@ out:
 
 int
 xfs_release(
-	xfs_inode_t	*ip)
+	struct xfs_inode	*ip,
+	bool			want_free_eofblocks)
 {
-	xfs_mount_t	*mp = ip->i_mount;
-	int		error = 0;
+	struct xfs_mount	*mp = ip->i_mount;
+	int			error = 0;
 
 	if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
 		return 0;
@@ -1243,7 +1244,7 @@ xfs_release(
 	 * another chance to drop them once the last reference to the inode is
 	 * dropped, so we'll never leak blocks permanently.
 	 */
-	if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL))
+	if (!want_free_eofblocks || !xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL))
 		return 0;
 
 	if (xfs_can_free_eofblocks(ip, false)) {
diff --git a/fs/xfs/xfs_inode.h b/fs/xfs/xfs_inode.h
index 979699248635..3a1321b8f1c8 100644
--- a/fs/xfs/xfs_inode.h
+++ b/fs/xfs/xfs_inode.h
@@ -416,7 +416,7 @@ enum layout_break_reason {
 #define XFS_INHERIT_GID(pip)	\
 	(xfs_has_grpid((pip)->i_mount) || (VFS_I(pip)->i_mode & S_ISGID))
 
-int		xfs_release(struct xfs_inode *ip);
+int		xfs_release(struct xfs_inode *ip, bool can_free_eofblocks);
 void		xfs_inactive(struct xfs_inode *ip);
 int		xfs_lookup(struct xfs_inode *dp, struct xfs_name *name,
 			   struct xfs_inode **ipp, struct xfs_name *ci_name);