lkml.org 
[lkml]   [2023]   [Feb]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH] cgroup/cpuset: Don't filter offline CPUs in cpuset_cpus_allowed() for top cpuset tasks
From
On 2/3/23 16:00, Tejun Heo wrote:
> On Fri, Feb 03, 2023 at 11:40:40AM -0500, Waiman Long wrote:
>> Since commit 8f9ea86fdf99 ("sched: Always preserve the user
>> requested cpumask"), relax_compatible_cpus_allowed_ptr() is calling
>> __sched_setaffinity() unconditionally. This helps to expose a bug in
>> the current cpuset hotplug code where the cpumasks of the tasks in
>> the top cpuset are not updated at all when some CPUs become online or
>> offline. It is likely caused by the fact that some of the tasks in the
>> top cpuset, like percpu kthreads, cannot have their cpu affinity changed.
>>
>> One way to reproduce this as suggested by Peter is:
>> - boot machine
>> - offline all CPUs except one
>> - taskset -p ffffffff $$
>> - online all CPUs
>>
>> Fix this by allowing cpuset_cpus_allowed() to return a wider mask that
>> includes offline CPUs for those tasks that are in the top cpuset. For
>> tasks not in the top cpuset, the old rule applies and only online CPUs
>> will be returned in the mask since hotplug events will update their
>> cpumasks accordingly.
>>
>> Fixes: 8f9ea86fdf99 ("sched: Always preserve the user requested cpumask")
>> Reported-by: Will Deacon <will@kernel.org>
>> Originally-from: Peter Zijlstra (Intel) <peterz@infradead.org>
>> Signed-off-by: Waiman Long <longman@redhat.com>
> So, this is the replacement for the first patch[1] Will posted, right?

Yes, if Will and Peter has no objection. I think it is less risky and
handle the partition case better.

With v1, Will's patch should get similar result as the existing
guarantee_online_cpus() function since we can infer offline cpus from
cpus_allowed. With v2, it does include offline cpus correctly, I
believe, as long as no partition is enabled. However, the hotplug code
is able to update the cpumasks when a CPU is onlined. So the presence of
offline CPUs is nice to have, but not essential.

>
>> void cpuset_cpus_allowed(struct task_struct *tsk, struct cpumask *pmask)
>> {
>> unsigned long flags;
>> + struct cpuset *cs;
>>
>> spin_lock_irqsave(&callback_lock, flags);
>> - guarantee_online_cpus(tsk, pmask);
>> + rcu_read_lock();
>> +
>> + cs = task_cs(tsk);
>> + if (cs != &top_cpuset)
>> + guarantee_online_cpus(tsk, pmask);
>> + /*
>> + * TODO: Tasks in the top cpuset won't get update to their cpumasks
>> + * when a hotplug online/offline event happens. So we include all
>> + * offline cpus in the allowed cpu list.
>> + */
>> + if ((cs == &top_cpuset) || cpumask_empty(pmask)) {
>> + const struct cpumask *possible_mask = task_cpu_possible_mask(tsk);
>> +
>> + /*
>> + * We first exclude cpus allocated to partitions. If there is no
>> + * allowable online cpu left, we fall back to all possible cpus.
>> + */
>> + cpumask_andnot(pmask, possible_mask, top_cpuset.subparts_cpus);
> and the differences are that
>
> * It's only applied to the root cgroup.
>
> * Cpus taken up by partitions are excluded.
>
> Is my understanding correct?
Yes, that is correct.
>
>> + if (!cpumask_intersects(pmask, cpu_online_mask))
>> + cpumask_copy(pmask, possible_mask);
>> + }
>> +
>> + rcu_read_unlock();
>> spin_unlock_irqrestore(&callback_lock, flags);
> So, I suppose you're suggesting applying this patch instead of the one Will
> Deacon posted[1] and we need Will's second patch[2] on top, right?
Right. Let hear if Will and Peter agree with this plan. I have tested
this patch and it passed Peter's reproducer test correctly. During
testing, I uncovered another bug in the cpu affinity code which results
in a separate scheduler patch to fix it.
>
> [1] http://lkml.kernel.org/r/20230131221719.3176-3-will@kernel.org
> [2] http://lkml.kernel.org/r/20230131221719.3176-3-will@kernel.org
>
> Thanks.
Cheers,
Longman

\
 
 \ /
  Last update: 2023-03-27 00:09    [W:0.072 / U:0.240 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site