Date:   Fri, 13 Jan 2023 11:13:28 -0500
From:   Waiman Long <longman@...hat.com>
To:     Florian Weimer <fweimer@...hat.com>,
        Mathieu Desnoyers <mathieu.desnoyers@...icios.com>
Cc:     linux-kernel@...r.kernel.org, Ingo Molnar <mingo@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        libc-alpha@...rceware.org
Subject: Re: rseq CPU ID not correct on 6.0 kernels for pinned threads

On 1/13/23 11:06, Florian Weimer wrote:
> * Mathieu Desnoyers:
>
>> On 2023-01-12 11:33, Florian Weimer wrote:
>>> * Mathieu Desnoyers:
>>>
>>>> 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 ?
>>> Attached.
>>> cpupower frequency-info says:
>>> analyzing CPU 0:
>>>     driver: intel_cpufreq
>>>     CPUs which run at the same hardware frequency: 0
>>>     CPUs which need to have their frequency coordinated by software: 0
>>>     maximum transition latency: 20.0 us
>>>     hardware limits: 800 MHz - 4.60 GHz
>>>     available cpufreq governors: conservative ondemand userspace powersave performance schedutil
>>>     current policy: frequency should be within 800 MHz and 4.60 GHz.
>>>                     The governor "schedutil" may decide which speed to use
>>>                     within this range.
>>>     current CPU frequency: Unable to call hardware
>>>     current CPU frequency: 3.20 GHz (asserted by call to kernel)
>>>     boost state support:
>>>       Supported: yes
>>>       Active: yes
>>> And I have: kernel.sched_energy_aware = 1
>>>
>>>> Is this on a physical machine or in a virtual machine ?
>>> I think it happened on both.
>>> I added additional error reporting to the test (running on kernel
>>> 6.0.18-300.fc37.x86_64), and it seems that there is something that is
>>> mucking with affinity masks:
>>> info: Detected CPU set size (in bits): 64
>>> info: Maximum test CPU: 19
>>> error: Pinned thread 17 ran on impossible cpu 7
>>> info: getcpu reported CPU 7, node 0
>>> info: CPU affinity mask: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
>>> error: Pinned thread 3 ran on impossible cpu 13
>>> info: getcpu reported CPU 13, node 0
>>> info: CPU affinity mask: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19
>>> info: Main thread ran on 2 CPU(s) of 20 available CPU(s)
>>> info: Other threads ran on 20 CPU(s)
>>> For each of these threads, the affinity mask should be a singleton
>>> set.
>>> Now I need to find out if there is a process that changes affinity
>>> settings.
>> If it's not cpu hotunplug, then perhaps something like systemd
>> modifies the AllowedCPUs of your cpuset concurrently ?
> It's probably just this kernel bug:
>
> commit da019032819a1f09943d3af676892ec8c627668e
> Author: Waiman Long <longman@...hat.com>
> Date:   Thu Sep 22 14:00:39 2022 -0400
>
>      sched: Enforce user requested affinity
>      
>      It was found that the user requested affinity via sched_setaffinity()
>      can be easily overwritten by other kernel subsystems without an easy way
>      to reset it back to what the user requested. For example, any change
>      to the current cpuset hierarchy may reset the cpumask of the tasks in
>      the affected cpusets to the default cpuset value even if those tasks
>      have pre-existing user requested affinity. That is especially easy to
>      trigger under a cgroup v2 environment where writing "+cpuset" to the
>      root cgroup's cgroup.subtree_control file will reset the cpus affinity
>      of all the processes in the system.
>      
>      That is problematic in a nohz_full environment where the tasks running
>      in the nohz_full CPUs usually have their cpus affinity explicitly set
>      and will behave incorrectly if cpus affinity changes.
>      
>      Fix this problem by looking at user_cpus_ptr in __set_cpus_allowed_ptr()
>      and use it to restrcit the given cpumask unless there is no overlap. In
>      that case, it will fallback to the given one. The SCA_USER flag is
>      reused to indicate intent to set user_cpus_ptr and so user_cpus_ptr
>      masking should be skipped. In addition, masking should also be skipped
>      if any of the SCA_MIGRATE_* flag is set.
>      
>      All callers of set_cpus_allowed_ptr() will be affected by this change.
>      A scratch cpumask is added to percpu runqueues structure for doing
>      additional masking when user_cpus_ptr is set.
>      
>      Signed-off-by: Waiman Long <longman@...hat.com>
>      Signed-off-by: Peter Zijlstra (Intel) <peterz@...radead.org>
>      Link: https://lkml.kernel.org/r/20220922180041.1768141-4-longman@redhat.com
>
> I don't think it's been merged into any stable kernels yet?

This patch will be in the v6.2 kernel. Since it is not marked as a fix, 
it won't go into a stable kernel by default.

Cheers,
Longman

Powered by blists - more mailing lists