Re: [PATCH v3 4/8] sched/deadline: Rebuild root domain accounting after every update

[Dietmar Eggemann]

On 11/03/2025 13:34, Waiman Long wrote:
> On 3/11/25 7:59 AM, Juri Lelli wrote:
>> On 10/03/25 20:16, Waiman Long wrote:
>>> On 3/10/25 3:18 PM, Waiman Long wrote:
>>>> On 3/10/25 2:54 PM, Dietmar Eggemann wrote:
>>>>> On 10/03/2025 10:37, Juri Lelli wrote:
>>>>>> Rebuilding of root domains accounting information (total_bw) is
>>>>>> currently broken on some cases, e.g. suspend/resume on aarch64.
>>>>>> Problem
>>>>> Nit: Couldn't spot any arch dependency here. I guess it was just
>>>>> tested
>>>>> on Arm64 platforms so far.
>>>>>
>>>>> [...]
>>>>>
>>>>>> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
>>>>>> index 44093339761c..363ad268a25b 100644
>>>>>> --- a/kernel/sched/topology.c
>>>>>> +++ b/kernel/sched/topology.c
>>>>>> @@ -2791,6 +2791,7 @@ void partition_sched_domains_locked(int
>>>>>> ndoms_new, cpumask_var_t doms_new[],
>>>>>>        ndoms_cur = ndoms_new;
>>>>>>          update_sched_domain_debugfs();
>>>>>> +    dl_rebuild_rd_accounting();
>>>>> Won't dl_rebuild_rd_accounting()'s lockdep_assert_held(&cpuset_mutex)
>>>>> barf when called via cpuhp's:
>>>>>
>>>>> sched_cpu_deactivate()
>>>>>
>>>>>     cpuset_cpu_inactive()
>>>>>
>>>>>       partition_sched_domains()
>>>>>
>>>>>         partition_sched_domains_locked()
>>>>>
>>>>>           dl_rebuild_rd_accounting()
>>>>>
>>>>> ?
>> Good catch. Guess I didn't notice while testing with LOCKDEP as I was
>> never able to hit this call path on my systems.
>>
>>>> Right. If cpuhp_tasks_frozen is true, partition_sched_domains() will be
>>>> called without holding cpuset mutex.
>>>>
>>>> Well, I think we will need an additional wrapper in cpuset.c that
>>>> acquires the cpuset_mutex first before calling
>>>> partition_sched_domains()
>>>> and use the new wrapper in these cases.
>>> Actually, partition_sched_domains() is called with the special
>>> arguments (1,
>>> NULL, NULL) to reset the domain to a single one. So perhaps something
>>> like
>>> the following will be enough to avoid this problem.
>> I think this would work, as we will still rebuild the accounting after
>> last CPU comes back from suspend. The thing I am still not sure about is
>> what we want to do in case we have DEADLINE tasks around, since with
>> this I belive we would be ignoring them and let suspend proceed.
> 
> That is the current behavior. You can certainly create a test case to
> trigger such condition and see what to do about it. Alternatively, you
> can document that and come up with a follow-up patch later on.

But don't we rely on that partition_sched_domains_locked() calls
dl_rebuild_rd_accounting() even in the reset_domain=1 case?

Testcase: suspend/resume

on Arm64 big.LITTLE cpumask=[LITTLE][big]=[0,3-5][1-2]
plus cmd line option 'isolcpus=3,4'.

with Waiman's snippet:
https://lkml.kernel.org/r/fd4d6143-9bd2-4a7c-80dc-1e19e4d1b2d1@xxxxxxxxxx

...
[  234.831675] --- > partition_sched_domains_locked() reset_domain=1
[  234.835966] psci: CPU4 killed (polled 0 ms)
[  234.838912] Error taking CPU3 down: -16
[  234.838952] Non-boot CPUs are not disabled
[  234.838986] Enabling non-boot CPUs ...
...

IIRC, that's the old DL accounting issue.