Re: [PATCH v2 4/9] cgroup: rstat: add WARN_ON_ONCE() if flushing outside task context

[Yosry Ahmed]

On Mon, Apr 3, 2023 at 1:38 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
>
> On Fri 31-03-23 12:03:47, Yosry Ahmed wrote:
> > On Fri, Mar 31, 2023 at 4:02 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > >
> > > On Thu 30-03-23 01:53:38, Yosry Ahmed wrote:
> > > [...]
> > > > Maybe we can add a primitive like might_sleep() for this, just food for thought.
> > >
> > > I do not think it is the correct to abuse might_sleep if the function
> > > itself doesn't sleep. If it does might_sleep is already involved.
> >
> > Oh, sorry if I wasn't clear, I did not mean to reuse might_sleep() --
> > I meant introducing a new similar debug primitive that shouts if irqs
> > are disabled.
>
> This is circling back to original concerns about arbitrary decision to
> care about IRQs. Is this really any different from spin locks or preempt
> disabled critical sections preventing any scheduling and potentially
> triggereing soft lockups?

Not really, I am sure there are other code paths that may cause
similar problems. At least we can start annotating them so that we
don't regress them (e.g. by introducing a caller that disables irqs --
converting them into hard lockups).

> --
> Michal Hocko
> SUSE Labs