Re: for_each_domain()/sched_domain_span() has offline CPUs (was Re: [PATCH 2/2] timers: Fix removed self-IPI on global timer's enqueue in nohz_full)

[Valentin Schneider]

On 27/03/24 15:28, Valentin Schneider wrote:
> On 27/03/24 13:42, Frederic Weisbecker wrote:
>> Le Tue, Mar 26, 2024 at 05:46:07PM +0100, Valentin Schneider a écrit :
>>> > Then with that patch I ran TREE07, just some short iterations:
>>> >
>>> > tools/testing/selftests/rcutorture/bin/kvm.sh --configs "10*TREE07" --allcpus --bootargs "rcutorture.onoff_interval=200" --duration 2
>>> >
>>> > And the warning triggers very quickly. At least since v6.3 but maybe since
>>> > earlier. Is this expected behaviour or am I right to assume that
>>> > for_each_domain()/sched_domain_span() shouldn't return an offline CPU?
>>> >
>>> 
>>> I would very much assume an offline CPU shouldn't show up in a
>>> sched_domain_span().
>>> 
>>> Now, on top of the above, there's one more thing worth noting:
>>>   cpu_up_down_serialize_trainwrecks()
>>> 
>>> This just flushes the cpuset work, so after that the sched_domain topology
>>> should be sane. However I see it's invoked at the tail end of _cpu_down(),
>>> IOW /after/ takedown_cpu() has run, which sounds too late. The comments
>>> around this vs. lock ordering aren't very reassuring however, so I need to
>>> look into this more.
>>
>> Ouch...
>>
>>> 
>>> Maybe as a "quick" test to see if this is the right culprit, you could try
>>> that with CONFIG_CPUSET=n? Because in that case the sched_domain update is
>>> ran within sched_cpu_deactivate().
>>
>> I just tried and I fear that doesn't help. It still triggers even without
>> cpusets :-s
>>
>
> What, you mean I can't always blame cgroups? What has the world come to?
>
> That's interesting, it means the deferred work item isn't the (only)
> issue. I'll grab your test patch and try to reproduce on TREE07.
>

Unfortunately I haven't been able to trigger your warning with ~20 runs of
TREE07 & CONFIG_CPUSETS=n, however it does trigger reliably with
CONFIG_CPUSETS=y, so I'm back to thinking the cpuset work is a likely
culprit...

>> Thanks.