Re: rseq CPU ID not correct on 6.0 kernels for pinned threads

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

On 2023-01-11 14:31, Mathieu Desnoyers wrote:
> On 2023-01-11 09:52, Mathieu Desnoyers wrote:
>> On 2023-01-11 06:26, Florian Weimer wrote:
>>> The glibc test suite contains a test that verifies that sched_getcpu
>>> returns the expected CPU number for a thread that is pinned (via
>>> sched_setaffinity) to a specific CPU.  There are other threads running
>>> which attempt to de-schedule the pinned thread from its CPU.  I believe
>>> the test is correctly doing what it is expected to do; it is invalid
>>> only if one believes that it is okay for the kernel to disregard the
>>> affinity mask for scheduling decisions.
>>>
>>> These days, we use the cpu_id rseq field as the return value of
>>> sched_getcpu if the kernel has rseq support (which it has in these
>>> cases).
>>>
>>> This test has started failing sporadically for us, some time around
>>> kernel 6.0.  I see failure occasionally on a Fedora builder, it runs:
>>>
>>> Linux buildvm-x86-26.iad2.fedoraproject.org 6.0.15-300.fc37.x86_64 #1 
>>> SMP PREEMPT_DYNAMIC Wed Dec 21 18:33:23 UTC 2022 x86_64 x86_64 x86_64 
>>> GNU/Linux
>>>
>>> I think I've seen it on the x86-64 builder only, but that might just be
>>> an accident.
>>>
>>> The failing tests log this output:
>>>
>>> =====FAIL: nptl/tst-thread-affinity-pthread.out=====
>>> info: Detected CPU set size (in bits): 64
>>> info: Maximum test CPU: 5
>>> error: Pinned thread 1 ran on impossible cpu 0
>>> error: Pinned thread 0 ran on impossible cpu 0
>>> info: Main thread ran on 4 CPU(s) of 6 available CPU(s)
>>> info: Other threads ran on 6 CPU(s)
>>> =====FAIL: nptl/tst-thread-affinity-pthread2.out=====
>>> info: Detected CPU set size (in bits): 64
>>> info: Maximum test CPU: 5
>>> error: Pinned thread 1 ran on impossible cpu 1
>>> error: Pinned thread 2 ran on impossible cpu 0
>>> error: Pinned thread 3 ran on impossible cpu 3
>>> info: Main thread ran on 5 CPU(s) of 6 available CPU(s)
>>> info: Other threads ran on 6 CPU(s)
>>>
>>> But I also encountered one local failure, but it is rare.  Maybe it's
>>> load-related.  There shouldn't be any CPU unplug or anything like that
>>> involved here.
>>>
>>> I am not entirely sure if something is changing CPU affinities from
>>> outside the process (which would be quite wrong, but not a kernel bug).
>>> But in the past, our glibc test has detected real rseq cpu_id
>>> brokenness, so I'm leaning towards that as the cause this time, too.
>>
>> It can be caused by rseq failing to update the cpu number field on 
>> return to userspace. Tthis could be validated by printing the regular 
>> getcpu vdso value and/or the value returned by the getcpu system call 
>> when the error is triggered, and see whether it matches the rseq cpu 
>> id value.
>>
>> It can also be caused by scheduler failure to take the affinity into
>> account.
>>
>> As you also point out, it can also be caused by some other task
>> modifying the affinity of your task concurrently. You could print
>> the result of sched_getaffinity on error to get a better idea of
>> the expected vs actual mask.
>>
>> Lastly, it could be caused by CPU hotplug which would set all bits
>> in the affinity mask as a fallback. As you mention it should not be
>> the cause there.
>>
>> Can you share your kernel configuration ?
> 
> Also, can you provide more information about the cpufreq driver and 
> governor used in your system ? e.g. output of
> 
> cpupower frequency-info
> 
> and also output of
> 
> sysctl kernel.sched_energy_aware
> 
> Is this on a physical machine or in a virtual machine ?

And one more thing: can you reproduce with a CONFIG_RSEQ=n kernel, or 
when disabling rseq with the glibc.pthread.rseq glibc tunable ?

Thanks,

Mathieu

> 
> Thanks,
> 
> Mathieu
> 
>>
>> Thanks,
>>
>> Mathieu
>>
>>>
>>> Thanks,
>>> Florian
>>>
>>
> 

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com