[PATCH] f2fs: fix to record both APPEND and UPDATE ino entries

[PATCH] f2fs: fix to record both APPEND and UPDATE ino entries

[Zhiguo Niu]

In f2fs_post_evict_inode(), record_bits is assigned with BIT(APPEND_INO)
and BIT(UPDATE_INO) respectively, so the second assignment overwrites
the first one: when both FI_APPEND_WRITE and FI_UPDATE_WRITE are set,
only the UPDATE_INO entry is recorded.

The two flags are not mutually exclusive.Once the APPEND_INO entry is
lost after the inode eviction, the shortcut in f2fs_do_sync_file() can
take the flush_out path in the next fsync(), skipping f2fs_fsync_node_pages().
As a result, the dnodes holding the newly allocated blocks are not
fsync-marked, and after a sudden power loss, the data appended before the
eviction is silently lost.

This restores the previous behavior where the two ino entries were
added independently.

Fixes: 314c9e476ffc ("f2fs: call __add_ino_entry out of the eviction path")
Signed-off-by: Zhiguo Niu <zhiguo.niu@unisoc.com>
---
 fs/f2fs/inode.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/fs/f2fs/inode.c b/fs/f2fs/inode.c
index 82ebcb338b0c..c20fcd8c2b97 100644
--- a/fs/f2fs/inode.c
+++ b/fs/f2fs/inode.c
@@ -1058,9 +1058,9 @@ static void f2fs_post_evict_inode(struct inode *inode)
 		goto skip_record;
 
 	if (is_inode_flag_set(inode, FI_APPEND_WRITE))
-		record_bits = BIT(APPEND_INO);
+		record_bits |= BIT(APPEND_INO);
 	if (is_inode_flag_set(inode, FI_UPDATE_WRITE))
-		record_bits = BIT(UPDATE_INO);
+		record_bits |= BIT(UPDATE_INO);
 
 	if (!record_bits)
 		goto skip_record;
-- 
2.43.0

Re: [PATCH] f2fs: fix to record both APPEND and UPDATE ino entries

[Chao Yu]

On 9/10/26 11:02, Zhiguo Niu wrote:
> In f2fs_post_evict_inode(), record_bits is assigned with BIT(APPEND_INO)
> and BIT(UPDATE_INO) respectively, so the second assignment overwrites
> the first one: when both FI_APPEND_WRITE and FI_UPDATE_WRITE are set,
> only the UPDATE_INO entry is recorded.
> 
> The two flags are not mutually exclusive.Once the APPEND_INO entry is
> lost after the inode eviction, the shortcut in f2fs_do_sync_file() can
> take the flush_out path in the next fsync(), skipping f2fs_fsync_node_pages().
> As a result, the dnodes holding the newly allocated blocks are not
> fsync-marked, and after a sudden power loss, the data appended before the
> eviction is silently lost.
> 
> This restores the previous behavior where the two ino entries were
> added independently.
> 
> Fixes: 314c9e476ffc ("f2fs: call __add_ino_entry out of the eviction path")
> Signed-off-by: Zhiguo Niu <zhiguo.niu@unisoc.com>
Reviewed-by: Chao Yu <chao@kernel.org>

Thanks,