Re: [PATCH] sched/topology: Fix overlapping sched_group build

[Valentin Schneider]

On Wed, Mar 25 2020, Valentin Schneider wrote:
> On Wed, Mar 25 2020, Valentin Schneider wrote:
>> On Tue, Mar 24 2020, Valentin Schneider wrote:
>>>  kernel/sched/topology.c | 23 ++++++++++++++++++++---
>>>  1 file changed, 20 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
>>> index 8344757bba6e..7033b27e5162 100644
>>> --- a/kernel/sched/topology.c
>>> +++ b/kernel/sched/topology.c
>>> @@ -866,7 +866,7 @@ build_balance_mask(struct sched_domain *sd, struct sched_group *sg, struct cpuma
>>>                       continue;
>>>
>>>               /* If we would not end up here, we can't continue from here */
>>> -		if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
>>> +		if (!cpumask_subset(sg_span, sched_domain_span(sibling->child)))
>>
>> So this is one source of issues; what I've done here is a bit stupid
>> since we include CPUs that *cannot* end up there. What I should've done
>> is something like:
>>
>>   cpumask_and(tmp, sched_domain_span(sibling->child), sched_domain_span(sd));
>>   if (!cpumask_equal(sg_span, tmp))
>>       ...
>>
>> But even with that I just unfold even more horrors: this breaks the
>> overlapping sched_group_capacity (see 1676330ecfa8 ("sched/topology: Fix
>> overlapping sched_group_capacity")).
>>
>> For instance, here I would have
>>
>>   CPU0-domain2-group4: span=4-5
>>   CPU4-domain2-group4: span=4-7 mask=4-5
>>
>
> ^ That's using Dietmar's qemu setup; on the D06 that is
>
>   CPU0-domain2-group48: span=48-71
>   CPU48-domain2-group48: span=48-95 mask=48-71
>

I realize this isn't clear at all;

- build_balance_mask() for CPU0-domain2-group48 (span=48-71) builds a
  mask that contains CPUs 48-71.
- build_balance_mask() for CPU48-domain2-group48 (span=48-95) builds a
  mask that contains CPUs 48-95.

Both masks have the same first CPU, so they both use the same
sched_group_capacity structure. However, CPU0-domain2-sgc48 was first
and set the sgc cpumask to what it wanted, so when it's turn for
CPU48-domain2-group48 we hit the WARN_ON_ONCE() and leave the
mask as-is.

And now I'll stop spewing nonsense for today.

>> Both groups are at the same topology level and have the same first CPU,
>> so they point to the same sched_group_capacity structure - but they
>> don't have the same span. They would without my "fix", but then the
>> group spans are back to being wrong. I'm starting to think this is
>> doomed, at least in the current state of things :/