Re: [PATCH] sched: fix rq lock recursion issue

[Satya Durga Srinivasu Prabhala <quic_satyap@quicinc.com>]

On 7/1/22 4:48 AM, Qais Yousef wrote:
> On 07/01/22 10:33, Peter Zijlstra wrote:
>> On Thu, Jun 30, 2022 at 10:53:10PM +0100, Qais Yousef wrote:
>>> Hi Satya
>>>
>>> On 06/24/22 00:42, Satya Durga Srinivasu Prabhala wrote:
>>>> Below recursion is observed in a rare scenario where __schedule()
>>>> takes rq lock, at around same time task's affinity is being changed,
>>>> bpf function for tracing sched_switch calls migrate_enabled(),
>>>> checks for affinity change (cpus_ptr != cpus_mask) lands into
>>>> __set_cpus_allowed_ptr which tries acquire rq lock and causing the
>>>> recursion bug.
>>>>
>>>> Fix the issue by switching to preempt_enable/disable() for non-RT
>>>> Kernels.
>>> Interesting bug. Thanks for the report. Unfortunately I can't see this being
>>> a fix as it just limits the bug visibility to PREEMPT_RT kernels, but won't fix
>>> anything, no? ie: Kernels compiled with PREEMPT_RT will still hit this failure.
>> Worse, there's !RT stuff that grew to rely on the preemptible migrate
>> disable stuff, so this actively breaks things.
>>
>>> I'm curious how the race with set affinity is happening. I would have thought
>>> user space would get blocked as __schedule() will hold the rq lock.
>>>
>>> Do you have more details on that?
>> Yeah, I'm not seeing how this works either, in order for
>> migrate_enable() to actually call __set_cpus_allowed_ptr(), it needs to
>> have done migrate_disable() *before* schedule, schedule() will then have
>> to call migrate_disable_swich(), and *then* migrate_enable() does this.
>>
>> However, if things are nicely balanced (as they should be), then
>> trace_call_bpf() using migrate_disable()/migrate_enable() should never
>> hit this path.
>>
>> If, OTOH, migrate_disable() was called prior to schedule() and we did do
>> migrate_disable_switch(), then it should be impossible for the
>> tracepoint/bpf stuff to reach p->migration_disabled == 0.
> I think it's worth to confirm which kernel Satya is on too. If it's GKI, then
> worth checking first this is actually reproducible on/applicable to mainline.
We are seeing the issue on 5.15 GKI Kernel. On older Kernels, like 5.10 
Kernel
migrate_disable/enable() end-up calling preempt_disable/enable(). I will 
cross
check further on comments and inputs I received so far.

>
> Cheers
>
> --
> Qais Yousef