Re: [PATCH 4.19 01/34] cgroup/cpuset: Fix a race between cpuset_attach() and cpu hotplug

["Michal Koutný" <mkoutny@suse.com>]

Hello.

On Mon, Feb 28, 2022 at 06:24:07PM +0100, Greg Kroah-Hartman <gregkh@linuxfoundation.org> wrote:
> [...]
>      cpuset_attach()				cpu hotplug
>     ---------------------------            ----------------------
>     down_write(cpuset_rwsem)
>     guarantee_online_cpus() // (load cpus_attach)
> 					sched_cpu_deactivate
> 					  set_cpu_active()
> 					  // will change cpu_active_mask
>     set_cpus_allowed_ptr(cpus_attach)
>       __set_cpus_allowed_ptr_locked()
>        // (if the intersection of cpus_attach and
>          cpu_active_mask is empty, will return -EINVAL)
>     up_write(cpuset_rwsem)
> [...]
> --- a/kernel/cgroup/cpuset.c
> +++ b/kernel/cgroup/cpuset.c
> @@ -1528,6 +1528,7 @@ static void cpuset_attach(struct cgroup_
>  	cgroup_taskset_first(tset, &css);
>  	cs = css_cs(css);
>  
> +	cpus_read_lock();
>  	mutex_lock(&cpuset_mutex);

This backport (and possible older kernels) looks suspicious since it comes
before commit d74b27d63a8b ("cgroup/cpuset: Change cpuset_rwsem and
hotplug lock order") v5.4-rc1~176^2~30 when the locking order was:
cpuset lock, cpus lock.

At the same time it also comes before commit 710da3c8ea7d ("sched/core:
Prevent race condition between cpuset and __sched_setscheduler()")
v5.4-rc1~176^2~27 when neither __sched_setscheduler() cared and this
race is similar. (The swapped locking may still conflict with
rebuild_sched_domains() before d74b27d63a8b.)


Michal