Hello! On Wed 24-06-20 16:56:18, Costa Sapuntzakis wrote: > Our workload: taking snapshots repeatedly of an active ext4 filesystem > (vdbench fwiw). e2fsck discovered a snapshot that had a corrupted > superblock after journal replay. Diffing the corrupted superblock to > the superblock before journal replay revealed that only s_last_orphan > and the checksum had changed. > > The following race could explain it: > > Thread 1 (T1): ext4_orphan_del -> update s_last_orphan to value A -> > ext4_handle_dirty_super -> ext4_superblock_csum_set -- PAUSE right > before setting es->s_checksum > > T2: ext4_orphan_del -> update s_last_orphan to value B -> > ext4_handle_dirty_super -> ext4_superblock_csum_set runs to completion > > T1: Resume and assign es->s_checksum > > Is there higher level synchronization going on that makes this race benign? Thanks for report and the analysis. What you describe indeed seems possible. > If not, a spinlock around the calculation and assignment should fix it. Yes, probably ext4_superblock_csum_set() should use lock_buffer(EXT4_SB(sb)->s_sbh) to synchronize updating of superblock checksum. Will you send a patch? > The spinlock still has the race where s_last_orphan is being updated > while the checksum is calculated. But the last thread to set > s_last_orphan will also eventually try to recalculate the checksum and > set it right (though it's possible some other thread will do it for > it). And I'm guessing/hoping jbd2 won't flush the superblock to the > journal and close a transaction until the references from > journal_get_write_access drain. The checksum is recalculated before > the get_write_access reference is dropped. Yes, jbd2 layer will make sure that inconsistent block contents will not make it to disk in this case. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
