Re: [PATCH v11 7/8] cgroup/cpuset: Update description of cpuset.cpus.partition in cgroup-v2.rst
From: Tejun Heo
Date: Thu Jun 30 2022 - 18:54:03 EST
Hello,
On Thu, Jun 30, 2022 at 04:32:11PM +0200, Michal Koutný wrote:
> file owner
> parent/ user (mkdir)
> `- cpuset.cpus root
> `- cpuset.cpus.partition root (P)
> `- child_1/ user
> ` cpuset.cpus user (*)
> `- child_2/ user
> ` cpuset.cpus user (*)
>
> The writes to child cpuset.cpus may/may not invalidate parent's (P)
> partition validity (whether a cpu is left to it to host possible tasks).
> child_1 vs child_2 overlap affects only whether the children cgroups are
> a valid partition.
>
> I think you mean: writes to children cpuset.cpus should be allowed,
> possible exclusivity violation should be reported in
> parent/cpuset.cpus.partition.
I see.
> What I thought was OK: prevent (fail) writes to children cpuset.cpus
> that'd violate the exclusivity (or would take the last cpu from parent
> if it's necessary to host a task).
> IMO, it's similar to failed writes to parent/cgroup.subtree_control in a
> delegated subtree if the parent still has some tasks (that'd violate
> internal node constraint).
>
> What I think might still be OK: allow writes to children cpuset.cpus
> that violate exclusivity and report that in children's
> cpuset.cpus.partition. Writes that'd take last cpu from parent should
> still fail (similar to the failing subtree_control writes above).
Yeah, this one.
So, here, one important question is who owns cpuset.cpus.partition
file - is it a konb which is owned by the parent like other resource
control knobs including cpuset.cpus or is it a knob which is owned by
the cgroup itself for selecting its own operation like cgroup.procs or
cgroup.subtree_control.
In the former case, the parent being able to say that "my children
can't overlap" makes sense although I'm not a big fan of the current
interface (again, who owns that knob?). In the latter case, it doesn't
really make sense cuz it'd be declaring "I can't make my children
overlap" - well, then, don't.
Thanks.
--
tejun