Re: [PATCH 2/5] cgroup/cpuset: Include offline CPUs when tasks' cpumasks in top_cpuset are updated

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




On 3/15/23 06:06, Michal Koutný wrote:
On Tue, Mar 14, 2023 at 03:02:53PM -0400, Waiman Long <longman@xxxxxxxxxx> wrote:
+			cpumask_andnot(new_cpus, possible_mask, cs->subparts_cpus);
+		} else {
+			cpumask_and(new_cpus, cs->effective_cpus, possible_mask);
+		}
I'm wrapping my head around this slightly.
1) I'd suggest swapping args in of cpumask_and() to have possible_mask
     consistently first.
I don't quite understand what you meant by "swapping args". It is effective
new_cpus = cs->effective_cpus ∩ possible_mask. What is the point of swapping
cs->effective_cpus and possible_mask.
No effect except better readability (possible_mask comes first in the
other branch above too, left hand arg as the "base" that's modified).
OK, I got it now. I will swap it as suggested.

2) Then I'm wondering whether two branches are truly different when
     effective_cpus := cpus_allowed - subparts_cpus
     top_cpuset.cpus_allowed == possible_mask        (1)
effective_cpus may not be equal "cpus_allowed - subparts_cpus" if some of
the CPUs are offline as effective_cpus contains only online CPUs.
subparts_cpu can include offline cpus too. That is why I choose that
expression. I will add a comment to clarify that.
I see now that it returns offlined cpus to top cpuset's tasks.

IOW, can you see a difference in what affinities are set to eligible
top_cpuset tasks before and after this patch upon CPU hotplug?
(Hm, (1) holds only in v2. So is this a fix for v1 only?)
This is due to the fact that cpu hotplug code currently doesn't update the
cpu affinity of tasks in the top cpuset. Tasks not in the top cpuset can
rely on the hotplug code to update the cpu affinity appropriately.
Oh, I mistook this for hotplug changing behavior but it's actually for
updating top_cpuset when its children becomes a partition root.

	IIUC, top cpuset + hotplug has been treated specially because
	hotplug must have taken care of affinity regardless of cpuset.
	v1 allowed modification of top cpuset's mask (not sure if that
	worked), v2 won't allow direct top cpuset's mask modificiation
	but indirectly via partition root children.

So this is a continuation for 3fb906e7fabb ("cgroup/cpuset: Don't filter
offline CPUs in cpuset_cpus_allowed() for top cpuset tasks") to ensure
hotplug offline/online cycle won't overwrite top_cpuset tasks'
affinities (when partition change during offlined period).
This looks correct in this regard then.
(I wish it were simpler but that's for a different/broader top
cpuset+hotplug approach.)

You can't change the content of "cpuset.cpus" in the top cpuset (both v1 & v2). I believe the CPU hotplug does not attempt to update the cpumask of tasks in the top cpuset mainly due to potential locking issue as hotplug is triggered by hardware event. Partition change, however, is a userspace event. So there shouldn't be any locking implication other than the fact per-cpu kthreads should not have their cpumasks changed.

To be consistent with commit 3fb906e7fabb ("cgroup/cpuset: Don't filter offline CPUs in cpuset_cpus_allowed() for top cpuset tasks"), similar logic needs to be applied in the later case.

Cheers,
Longman




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Security]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]     [Monitors]

  Powered by Linux