Re: [PATCH] cgroup: don't queue css_release_work if one already pending

[Tejun Heo]

On Mon, May 23, 2022 at 12:00:06PM -0700, Tadeusz Struk wrote:
> On 5/20/22 09:42, Michal Koutný wrote:
> > On Fri, May 20, 2022 at 09:38:12AM -0700, Tadeusz Struk <tadeusz.struk@xxxxxxxxxx> wrote:
> > > as this is a result of enqueuing the same css->destroy_work onto the same WQ,
> > > one on the rcu path and one on the css_release path.
> > > I will prototype it today and test with syzbot.
> > 
> > In my understanding, you'd need two independent work_structs in a css,
> > not two separate workqueues to put the single entry on.
> 
> Yes, separating the css_killed_ref and css_release paths with two separate work_structs
> fixes the two syzbot list corruption issues [1]&[2].
> I tested it on mainline v5.18.0 and v5.10.117
> In case of [2] the mainline triggers an issue, but it is unrelated to list corruption.

Can you try holding an extra ref in the killed path?

Thanks.

-- 
tejun