Re: [RFC/PATCH] sched: Support moving kthreads into cpuset cgroups

[Tejun Heo <tj@kernel.org>]

Hello,

On Tue, May 06, 2025 at 11:35:32AM -0700, Xi Wang wrote:
> In theory we should be able to manage kernel tasks with cpuset
> cgroups just like user tasks, would be a flexible way to limit
> interferences to real-time and other sensitive workloads. This is
> however not supported today: When setting cpu affinity for kthreads,
> kernel code uses a simpler control path that directly lead to
> __set_cpus_allowed_ptr or __ktread_bind_mask. Neither honors cpuset
> restrictions.
> 
> This patch adds cpuset support for kernel tasks by merging userspace
> and kernel cpu affinity control paths and applying the same
> restrictions to kthreads.
> 
> The PF_NO_SETAFFINITY flag is still supported for tasks that have to
> run with certain cpu affinities. Kernel ensures kthreads with this
> flag have their affinities locked and they stay in the root cpuset:
> 
> If userspace moves kthreadd out of the root cpuset (see example
> below), a newly forked kthread will be in a non root cgroup as well.
> If PF_NO_SETAFFINITY is detected for the kthread, it will move itself
> into the root cpuset before the threadfn is called. This does depend
> on the kthread create -> kthread bind -> wake up sequence.

Can you describe the use cases in detail? This is not in line with the
overall direction. e.g. We're making cpuset work with housekeeping mechanism
and tell workqueue which CPUs can be used for unbound execution and kthreads
which are closely tied to userspace activities are spawned into the same
cgroups as the user thread and subject to usual resource control.

There are a lot of risks in subjecting arbitrary kthreads to all cgroup
resource controls and just allowing cpuset doesn't seem like a great idea.
Integration through housekeeping makes a lot more sense to me. Note that
even for just cpuset thread level control doesn't really work that well. All
kthreads are forked by kthreadd. If you move the kthreadd into a cgroup, all
kthreads includling kworkers for all workqueues will be spawned there. The
granularity of control isn't much better than going through housekeeping.

Thanks.

-- 
tejun