On Wed 04-11-20 14:12:35, Jan Kara wrote: > On Tue 03-11-20 09:16:19, Costa Sapuntzakis wrote: > > Jan, does this fixup from Hillf look ok to you? You originally argued for > > lock_buffer/unlock_buffer. > > > > I think the problem here is that the ext4 code assumes that > > ext4_commit_super will not sleep if sync == 0 (or at least > > __ext4_grp_locked_error deos). Perhaps there should be a comment on > > ext4_commit_super documenting this constraint. > > Hum, right. I forgot about that. The spinlock Hillf suggests kind of works > but it still doesn't quite handle the case where superblock is modified in > parallel from another place (that can still lead to sb checksum mismatch on > next load). When we are going for a more complex solution I'd rather solve > this as well... I'm looking into possible solutions now. Just an update: I'm still working on this and it's like peeling an onion. The mixing of journalled superblock updates with unjournalled one is really evil and commit acaa532687cd fixes just a small part of the problems. So for now I suggest to just revert acaa532687cd (to avoid sleep in atomic context) and I'll submit larger set of fixes once they are ready. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
