There is one case can cause data corruption. - write 4k to fileA - fsync fileA, 4k data is writebacked to lbaA - write 4k to fileA - kworker flushs 4k to lbaB; dnode contain lbaB didn't be persisted yet - write 4k to fileB - kworker flush 4k to lbaA due to SSR - SPOR -> dnode with lbaA will be recovered, however lbaA contains fileB's data
One solution is tracking all fsynced file's block history, and disallow SSR overwrite on newly invalidated block on that file. However, during recovery, no matter the dnode is flushed or fsynced, all previous dnodes until last fsynced one in node chain can be recovered, that means we need to record all block change in flushed dnode, which will cause heavy cost, so let's just use simple fix by forbidding SSR overwrite directly. Signed-off-by: Chao Yu <yuch...@huawei.com> --- fs/f2fs/segment.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c index 9d9d9a050d59..69b3b553ee6b 100644 --- a/fs/f2fs/segment.c +++ b/fs/f2fs/segment.c @@ -2205,9 +2205,11 @@ static void update_sit_entry(struct f2fs_sb_info *sbi, block_t blkaddr, int del) if (!f2fs_test_and_set_bit(offset, se->discard_map)) sbi->discard_blks--; - /* don't overwrite by SSR to keep node chain */ - if (IS_NODESEG(se->type) && - !is_sbi_flag_set(sbi, SBI_CP_DISABLED)) { + /* + * SSR should never reuse block which is checkpointed + * or newly invalidated. + */ + if (!is_sbi_flag_set(sbi, SBI_CP_DISABLED)) { if (!f2fs_test_and_set_bit(offset, se->ckpt_valid_map)) se->ckpt_valid_blocks++; } -- 2.18.0.rc1