Re: [PATCH 2/4] sched/deadline: Improve admission control for asymmetric CPU capacities

[Valentin Schneider]

On 08/04/20 14:30, luca abeni wrote:
>>
>> I don't think this is strictly equivalent to what we have now for the
>> SMP case. 'cpus' used to come from dl_bw_cpus(), which is an ugly way
>> of writing
>>
>>      cpumask_weight(rd->span AND cpu_active_mask);
>>
>> The rd->cpu_capacity_orig field you added gets set once per domain
>> rebuild, so it also happens in sched_cpu_(de)activate() but is
>> separate from touching cpu_active_mask. AFAICT this mean we can
>> observe a CPU as !active but still see its capacity_orig accounted in
>> a root_domain.
>
> Sorry, I suspect this is my fault, because the bug comes from my
> original patch.
> When I wrote the original code, I believed that when a CPU is
> deactivated it is also removed from its root domain.
>
> I now see that I was wrong.
>

Well it is indeed the case, but sadly it's not an atomic step - AFAICT with
cpusets we do hold some cpuset lock when calling __dl_overflow() and when
rebuilding the domains, but not when fiddling with the active mask.

I just realized it's even more obvious for dl_cpu_busy(): IIUC it is meant
to prevent the removal of a CPU if it would lead to a DL overflow - it
works now because the active mask is modified before it gets called, but
here it breaks because it's called before the sched_domain rebuild.

Perhaps re-computing the root domain capacity sum at every dl_bw_cpus()
call would be simpler. It's a bit more work, but then we already have a
for_each_cpu_*() loop, and we only rely on the masks being correct.

>
>                       Luca