On 8/16/22 18:19, Tejun Heo wrote:
Hello,
On Tue, Aug 16, 2022 at 06:11:03PM -0400, Waiman Long wrote:
It is hard to synchronize different subsystems atomically without running
into locking issue. Let me think about what can be done in this case.
I have a hard time seeing why this would be particularly difficult. cpuset
just needs to make the latest cpumask available to sched core in an easily
accessible form and whenever that changes, trigger a set_cpus_allowed call.
There's no need to entangle operations across the whole subsystems. All
that's needed to be communicated is the current cpumask.
Is using a sequence number to check for race with retry good enough?
It seems unnecessarily fragile and complicated to me. If we're gonna change
it, let's change it right.
Thanks for the suggestion. I think I get what you want. I am going to
migrate the cpuset_set_cpus_allowed_ptr() logic into
set_cpus_allowed_ptr() itself. IOW, if user_cpus_ptr is defined, it will
be an additional mask to be applied on top. It does affect all callers
of set_cpus_allowed_ptr() though. I am going to drop this cpuset
specific patch.
BTW, I will be on PTO starting tomorrow until next Tuesday. So I will be
slow in responding to emails.
Cheers,
Longman
Cheers,
Longman
