Re: [PATCH] cgroup: Disable preemption for cgrp->freezer.freeze_seq when CONFIG_PREEMPT_RT=y.

[Waiman Long <llong@redhat.com>]

On 10/2/25 4:28 AM, Michal Koutný wrote:
> Hello.
>
> Thanks for looking into this Kuniyuki.
>
> On Thu, Oct 02, 2025 at 05:22:07AM +0000, Kuniyuki Iwashima <kuniyu@google.com> wrote:
>> The writer side is under spin_lock_irq(), but the section is still
>> preemptible with CONFIG_PREEMPT_RT=y.
> I see similar construction in other places, e.g.
> 	mems_allowed_seq in set_mems_allowed
> 	period_seqcount in ioc_start_period
> 	pidmap_lock_seq in alloc_pid/pidfs_add_pid
> (where their outer lock becomes preemptible on PREEMPT_RT.)
>
>> Let's wrap the section with preempt_{disable,enable}_nested().
> Is it better to wrap them all (for CONFIG_PREEMPT_RT=y) or should they
> become seqlock_t on CONFIG_PREEMPT_RT=y?

Changing it to seqlock_t will introduce another lock that need to be 
handled. Alternatively, we can use seqcount_spinlock_t for freeze_seq 
for the current case and associate css_set_lock with the sequence count 
in seqcount_spinlock_init(). That should properly handle the case for 
PREEMPT_RT kernel.

Cheers,
Longman