Re: [PATCH bpf-next 2/2] bpf/stackmap: fix A-A deadlock in bpf_get_stack()

From: Song Liu
Date: Mon Oct 14 2019 - 16:13:55 EST


Thanks Peter!

> On Oct 14, 2019, at 2:09 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Thu, Oct 10, 2019 at 06:06:14PM +0000, Alexei Starovoitov wrote:
>> On 10/10/19 10:46 AM, Peter Zijlstra wrote:
>
>>> All of stack_map_get_build_id_offset() is just disguisting games; I did
>>> tell you guys how to do lockless vma lookups a few years ago -- and yes,
>>> that is invasive core mm surgery. But this is just disguisting hacks for
>>> not wanting to do it right.
>>
>> you mean speculative page fault stuff?
>> That was my hope as well and I offered Laurent all the help to land it.
>> Yet after a year since we've talked the patches are not any closer
>> to landing.
>> Any other 'invasive mm surgery' you have in mind?
>
> Indeed that series. It had RCU managed VMAs and lockless VMA lookups,
> which is exactly what you need here.

Lockless VMA lookups will be really useful. It would resolve all the
pains we are having here.

I remember Google folks also mentioned in LPC that they would like
better mechanism to confirm build-id in perf.

>
>>> Basically the only semi-sane thing to do with that trainwreck is
>>> s/in_nmi()/true/ and pray.
>>>
>>> On top of that I just hate buildids in general.
>>
>> Emotions aside... build_id is useful and used in production.
>> It's used widely because it solves real problems.
>
> AFAIU it solves the problem of you not knowing what version of the
> binary runs where; which I was hoping your cloud infrastructure thing
> would actually know already.
>
> Anyway, I know what it does, I just don't nessecarily agree it is the
> right way around that particular problem (also, the way I'm personally
> affected is that perf-record is dead slow by default due to built-id
> post processing).
>
> And it obviously leads to horrible hacks like the code currently under
> discussion :/
>
>> This dead lock is from real servers and not from some sanitizer wannabe.
>
> If you enable CFS bandwidth control and run this function on the
> trace_hrtimer_start() tracepoint, you should be able to trigger a real
> AB-BA lockup.
>
>> Hence we need to fix it as cleanly as possible and quickly.
>> s/in_nmi/true/ is certainly an option.
>
> That is the best option; because tracepoints / perf-overflow handlers
> really should not be taking any locks.
>
>> I'm worried about overhead of doing irq_work_queue() all the time.
>> But I'm not familiar with mechanism enough to justify the concerns.
>> Would it make sense to do s/in_nmi/irgs_disabled/ instead?
>
> irqs_disabled() should work in this particular case because rq->lock
> (and therefore all it's nested locks) are IRQ-safe.

We worry about the overhead of irq_work for every single stackmap
lookup. So we would like to go with the irqs_disabled() check. I just
sent v2 of the patch.

Thanks again,
Song