Re: [PATCH 4/6 v8] sched/fair: Add push task mechanism for fair

[K Prateek Nayak]

On 12/5/2025 6:26 PM, Peter Zijlstra wrote:
> On Fri, Dec 05, 2025 at 06:19:07PM +0530, K Prateek Nayak wrote:
>> On 12/5/2025 2:29 PM, Peter Zijlstra wrote:
>>>>> Why not use move_queued_task() ?
>>>>
>>>> double_lock_balance() can fail and prevent being blocked waiting for
>>>> new rq whereas move_queued_task() will wait, won't it ?
>>>>
>>>> Do you think move_queued_task() would be better ?
>>>
>>> No, double_lock_balance() never fails, the return value indicates if the
>>> currently held rq-lock, (the first argument) was unlocked while
>>> attaining both -- this is required when the first rq is a higher address
>>> than the second.
>>>
>>> double_lock_balance() also puts the wait-time and hold time of the
>>> second inside the hold time of the first, which gets you a quadric term
>>> in the rq hold times IIRC. Something that's best avoided.
>>>
>>> move_queued_task() OTOH takes the task off the runqueue you already hold
>>> locked, drops this lock, acquires the second, puts the task there, and
>>> returns with the dst rq locked.
>>
>> So I was experimenting with:
>>
>>     deactivate_task(rq, p, 0);
>>     set_task_cpu(p, target_cpu);
>>     __ttwu_queue_wakelist(p, target_cpu, 0);
>>
>> and nothing has screamed at me yet during the benchmark runs.
>> Would this be any good instead of the whole lock juggling?
> 
> This will get schedstats and any class with ->task_woken confused I
> think. The IPI handler (sched_ttwu_pending / ttwu_do_activate) is
> currently only geared towards doing the remote bit of ttwu.
> 
> This is fixable of course.

I believe a simple enough check could be:

  task_on_rq_migrating() -> queued  (vs)  blocked -> queued

during activation.

> 
>> Since this CPU is found to be going overloaded, pushing via an
>> IPI vs taking the overhead ourselves seems to make more sense
>> to me from EAS standpoint.
> 
> The performance characteristics here are very platform dependent.
> Sometimes raising an IPI can be very expensive in and of itself.

Ack! But I assume those platforms already disable TTWU_QUEUE?
Or maybe the ttwu_queue_cond() takes care of it indirectly. We
can add some guards if necessary and use the same fallback as
PREEMPT_RT.

-- 
Thanks and Regards,
Prateek