We have a node chain to serialize node block writes, so if any IOs for
node block writes are reordered, we'll get broken node chain. IOWs,
roll-forward recovery will see all or none node blocks given fsync
mark.
E.g.,
Node chain consists of:
N1 -> N2 -> N3 -> NFSYNC -> N1' -> N2' -> N'FSYNC
Reordered to:
1) N1 -> N2 -> N3 -> N2' -> NFSYNC -> N'FSYNC -> power-cut
2) N1 -> N2 -> N3 -> N1' -> NFSYNC -> power-cut
3) N1 -> N2 -> NFSYNC -> N1' -> N'FSYNC -> N3 -> power-cut
4) N1 -> NFSYNC -> N1' -> N2' -> N'FSYNC -> N3 -> power-cut
Roll-forward recovery can proceed to:
1) N1 -> N2 -> N3 -> NFSYNC -> X
2) N1 -> N2 -> N3 -> NFSYNC -> N1' -> X
3) N1 -> N2 -> N3 -> FSYNC -> N1' -> X
4) N1 -> X
Signed-off-by: Jaegeuk Kim <jaegeuk@xxxxxxxxxx>
---
fs/f2fs/file.c | 16 +++++++++++++---
1 file changed, 13 insertions(+), 3 deletions(-)
diff --git a/fs/f2fs/file.c b/fs/f2fs/file.c
index a0413c951458..0246d19d96c3 100644
--- a/fs/f2fs/file.c
+++ b/fs/f2fs/file.c
@@ -274,9 +274,19 @@ static int f2fs_do_sync_file(struct file *file, loff_t start, loff_t end,
goto sync_nodes;
}
- ret = wait_on_node_pages_writeback(sbi, ino);
- if (ret)
- goto out;
+ /*
+ * If it's atomic_write, it's just fine to keep write ordering. So
+ * here we don't need to wait for node write completion, since we use
+ * node chain which serializes node blocks. If one of node writes are
+ * reordered, we can see simply broken chain, resulting in stopping
+ * roll-forward recovery. It means we'll recover all or none node blocks
+ * given fsync mark.
+ */
+ if (!atomic) {
+ ret = wait_on_node_pages_writeback(sbi, ino);
+ if (ret)
+ goto out;
+ }
/* once recovery info is written, don't need to tack this */
remove_ino_entry(sbi, ino, APPEND_INO);
--
2.13.0.rc1.294.g07d810a77f-goog
