Re: [PATCH v23 6/9] sched: Handle blocked-waiter migration (and return migration)

[K Prateek Nayak <kprateek.nayak@amd.com>]

Hello John,

On 11/20/2025 12:03 PM, John Stultz wrote:
> On Wed, Nov 19, 2025 at 6:55 PM K Prateek Nayak <kprateek.nayak@amd.com> wrote:
>> On 11/20/2025 7:30 AM, John Stultz wrote:
>>>> Ok, so you're suggesting maybe putting the
>>>>     if (task_on_rq_migrating(owner))
>>>> case ahead of the
>>>>     if (owner_cpu != this_cpu)
>>>> check?
>>>>
>>>> Let me give that a whirl and see how it does.
>>>
>>> That said, thinking another second on it, I also realize once we
>>> decide to proxy_migrate, there is always the chance the owner gets
>>> migrated somewhere else. So we can check task_on_rq_migrating() but
>>> then right after we check that it might be migrated, and we can't
>>> really prevent this.  And in that case, doing the proxy-migration to
>>> the wrong place will be ok, as that cpu will then bounce the tasks to
>>> the owner's new cpu.
>>>
>>> Hopefully this would be rare though. :)
>>
>> Ack! I was just thinking of some extreme scenarios. We can probably
>> think about it if and when we run into a problem with it :)
>>
>> That said, once we decide to move the first donors to owner's CPU
>> should we task some care to retain the owner on the same CPU as much
>> as possible - take it out of the purview of load balancing and only
>> move it if the owner is no long runnable on that CPU as a result of
>> affinity changes?
> 
> Eh, I'm hesitant to muck with the balancing effects on the lock
> owners.  If it's better for them to move around, then the donor chain
> should follow along (which will happen naturally).

So assume the case where you have the owner and a bunch of blocked
donor on the same rq. This rq appears the busiest to the load balancer.

Load balancer go thorough the task list and find that almost
everything is blocked on the owner. Then it arrives at the owner in a
preempted state (queued; not running) and thinks this is a good enough
task to move to reduce imbalance.

Now, this triggers a whole chain migration at pick for all the blocked
donors to the new CPU. Seems wasteful (although again this is a very
unlikely scenario to not be on_cpu with so many donors on the CPU)

> 
> Now, the donors (or really any blocked tasks on the rq) we avoid
> adding to any movable lists, since we only want proxy-migrations to
> move those tasks (so they stick to the owner).
-- 
Thanks and Regards,
Prateek