On Thu 20-01-11 04:37:55, Nick Piggin wrote: > On Thu, Jan 20, 2011 at 3:50 AM, Jan Kara <jack@xxxxxxx> wrote: > > On Thu 20-01-11 03:03:23, Nick Piggin wrote: > >> On Thu, Jan 20, 2011 at 12:21 AM, Jan Kara <jack@xxxxxxx> wrote: > >> > Well, we are not required to cancel all the outstanding AIO because of the > >> > API requirement, that's granted. But we must do it because of the way how > >> > the code is written. Outstanding IO requests reference ioctx but they are > >> > not counted in ctx->users but in ctx->reqs_active. So the code relies on > >> > the fact that the reference held by the hash table protects ctx from being > >> > freed and io_destroy() waits for requests before dropping the last > >> > reference to ctx. But there's the second race I describe making it possible > >> > for new IO to be created after io_destroy() has waited for all IO to > >> > finish... > >> > >> Yes there is that race too I agree. I just didn't follow through the code far > >> enough to see it was a problem -- I thought it was by design. > >> > >> I'd like to solve it without synchronize_rcu() though. > > Ah, OK. I don't find io_destroy() performance critical but I can > > Probably not performance critical, but it could be a very > large slowdown so somebody might complain. > > > understand that you need not like synchronize_rcu() there. ;) Then it > > should be possible to make IO requests count in ctx->users which would > > solve the race as well. We'd just have to be prepared that request > > completion might put the last reference to ioctx and free it but that > > shouldn't be an issue. Do you like that solution better? > > I think so, if it can be done without slowing things down > and adding locks or atomics if possible. Actually, I found that freeing ioctx upon IO completion isn't straightforward because freeing ioctx may need to sleep (it is destroying work queue) and aio_complete() can be called from an interrupt context. We could offload the sleeping work to the RCU callback (basically we'd have to offload the whole __put_ioctx() to RCU callback) but I'm not convinced it's worth it so I rather chose a bit more subtle approach for fixing the race (see my patch). Honza -- Jan Kara <jack@xxxxxxx> SUSE Labs, CR -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
