Re: [RFC PATCH] bfq: fix waker_bfqq inconsistency crash

[Khazhy Kumykov]

On Thu, Nov 3, 2022 at 1:47 AM Jan Kara <jack@xxxxxxx> wrote:
>
> On Thu 03-11-22 11:51:15, Yu Kuai wrote:
> > Hi,
> >
> > 在 2022/11/03 11:05, Khazhy Kumykov 写道:
> > > On Wed, Nov 2, 2022 at 7:56 PM Yu Kuai <yukuai1@xxxxxxxxxxxxxxx> wrote:
> > > >
> > > > Hi,
> > > >
> > > > 在 2022/11/03 9:39, Khazhismel Kumykov 写道:
> > > > > This fixes crashes in bfq_add_bfqq_busy due to waker_bfqq being NULL,
> > > > > but woken_list_node still being hashed. This would happen when
> > > > > bfq_init_rq() expects a brand new allocated queue to be returned from
> > > >
> > > >   From what I see, bfqq->waker_bfqq is updated in bfq_init_rq() only if
> > > > 'new_queue' is false, but if 'new_queue' is false, the returned 'bfqq'
> > > > from bfq_get_bfqq_handle_split() will never be oom_bfqq, so I'm confused
> > > > here...
> > > There's two calls for bfq_get_bfqq_handle_split in this function - the
> > > second one is after the check you mentioned, and is the problematic
> > > one.
> > Yes, thanks for the explanation. Now I understand how the problem
> > triggers.
> >
> > > >
> > > > > bfq_get_bfqq_handle_split() and unconditionally updates waker_bfqq
> > > > > without resetting woken_list_node. Since we can always return oom_bfqq
> > > > > when attempting to allocate, we cannot assume waker_bfqq starts as NULL.
> > > > > We must either reset woken_list_node, or avoid setting woken_list at all
> > > > > for oom_bfqq - opt to do the former.
> > > >
> > > > Once oom_bfqq is used, I think the io is treated as issued from root
> > > > group. Hence I don't think it's necessary to set woken_list or
> > > > waker_bfqq for oom_bfqq.
> > > Ack, I was wondering what's right here since, evidently, *someone* had
> > > already set oom_bfqq->waker_bfqq to *something* (although... maybe it
> > > was an earlier init_rq). But maybe it's better to do nothing if we
> > > *know* it's oom_bfqq.
> >
> > I need to have a check how oom_bfqq get involved with waker_bfqq, and
> > then see if it's reasonable.
> >
> > Probably Jan and Paolo will have better view on this.
>
> Thanks for the CC Kuai and thanks to Khazy for spotting the bug. The
> oom_bfqq is just a fallback bfqq and as such it should be extempted from
> all special handling like waker detection etc. All this stuff is just for
> optimizing performance and when we are OOM, we have far larger troubles
> than to optimize performance.
>
> So how I think we should really fix this is that we extempt oom_bfqq from
> waker detection in bfq_check_waker() by adding:
>
>         bfqq == bfqd->oom_bfqq ||
>         bfqd->last_completed_rq_bfq == bfqd->oom_bfqq)
>
> to the initial check and then also if bfq_get_bfqq_handle_split() returns
> oom_bfqq we should just skip carrying over the waker information.
Thanks for the tip! I'll send a followup, including your suggestions.

I do have some other questions in this area, if you could help me
understand. Looking again at bfq_init_rq, inside of the !new_queue
section - we call bfq_split_bfqq() to "split" our bfqq, then in the
next line bfq_get_bfqq_handle_split inspects bic_to_bfqq(bic,
is_sync), and if it's NULL, allocates a new queue. However, if we have
an async rq, this call will return the pre-existing async bfqq, as the
call to bfq_split_bfqq() only clears the sync bfqq, ever. The best
understanding I have now is: "bic->bfqq[aync] is never NULL (and we
don't merge async queues) so we'll never reach this !new_queue section
anyways if it's async". Is that accurate?

Thanks,
Khazhy


>
>                                                                 Honza
> --
> Jan Kara <jack@xxxxxxxx>
> SUSE Labs, CR

Attachment: smime.p7s
Description: S/MIME Cryptographic Signature