Hello, I've been seeing failures with generic/388 test (with 4.12-based distro kernel). Now after debugging it for some time it seems it is a problem in EXT4_GOING_FLAGS_NOLOGFLUSH implementation. What seems to be happening is that e.g. directory is being created, we set EXT4_FLAGS_SHUTDOWN in the middle and so the creation fails. We do iput() which should delete the unused inode but since the filesystem is marked as EXT4_FLAGS_SHUTDOWN, those changes don't hit the disk. *But* the transaction which has allocated the inode still manages to commit before we abort the journal (as there is a window provided by msleep() in EXT4_GOING_FLAGS_NOLOGFLUSH implementation). So after journal recovery, we have unattached inode and e2fsck complains. Now this would not happen if the system just crashed - in such case, either the allocating transaction won't be able to commit, or iput() would have finished successfully and logged inode deletion in the same transaction. So why is journal abort happening after EXT4_FLAGS_SHUTDOWN being set and why is that window even prolonged by msleep? That would deserve a comment if nothing else BTW... Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
