[PATCH] f2fs: write the directory inode in fsync before checkpoint

[Daeho Jeong]

From: Daeho Jeong <daehojeong@xxxxxxxxxx>

fsync() on a directory is done by a checkpoint, but the checkpoint
does not write timestamp-only inode updates. Since commit 7c45729a4d6d
("f2fs: keep dirty inodes selectively for checkpoint"), such updates
are marked with sync=false and left to inode writeback on purpose.

fsync() on a regular file calls f2fs_write_inode() first when the
inode is dirty, but fsync() on a directory goes to the checkpoint
directly and misses it. So if the last change to a directory only
updated its timestamps, e.g. unlinking a file in it, the new mtime is
lost after a sudden power off even though fsync() returned. generic/547
fails sometimes because of this.

Call f2fs_write_inode() for a directory before the checkpoint, as the
regular file path does. The inode block becomes dirty and is written by
the same checkpoint, so no extra checkpoint or flush is added.
Timestamp-only updates without fsync() are still not written by the
checkpoint.

Fixes: 7c45729a4d6d ("f2fs: keep dirty inodes selectively for checkpoint")
Signed-off-by: Daeho Jeong <daehojeong@xxxxxxxxxx>
---
 fs/f2fs/file.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
index 56c686b8580..b012cc22307 100644
--- a/fs/f2fs/file.c
+++ b/fs/f2fs/file.c
@@ -351,8 +351,15 @@ static int f2fs_do_sync_file(struct file *file, loff_t start, loff_t end,
 
 	trace_f2fs_sync_file_enter(inode);
 
-	if (S_ISDIR(inode->i_mode))
+	if (S_ISDIR(inode->i_mode)) {
+		/*
+		 * Checkpoint writes only the inodes on the dirty meta list.
+		 * Timestamp updates of a directory, e.g. by unlinking a file
+		 * in it, do not put it on the list, so write it here.
+		 */
+		f2fs_write_inode(inode, NULL);
 		goto go_write;
+	}
 
 	/* if fdatasync is triggered, let's do in-place-update */
 	if (datasync || get_dirty_pages(inode) <= SM_I(sbi)->min_fsync_blocks)
-- 
2.56.0.360.g66cac248cb-goog