Re: [PATCH] rseq: Protect event mask against membarrier IPI

[Mathieu Desnoyers]

On 2025-08-23 08:31, Thomas Gleixner wrote:
> On Wed, Aug 20 2025 at 09:23, Mathieu Desnoyers wrote:
> > On 2025-08-13 11:02, Thomas Gleixner wrote:
> > We should also update this comment in include/linux/sched.h:
> >
> >           /*
> >            * RmW on rseq_event_mask must be performed atomically
> >            * with respect to preemption.
> >            */
> >           unsigned long rseq_event_mask;
> >
> > to e.g.:
> >
> > /*
> >    * RmW on rseq_event_mask must be performed atomically
> >    * with respect to preemption and membarrier IPIs.
> >    */
>
> True.
>
> > > +
> > >    /*
> > >     * Map the event mask on the user-space ABI enum rseq_cs_flags
> > >     * for direct mask checks.
> > > @@ -41,9 +47,8 @@ static inline void rseq_handle_notify_re
> > >    static inline void rseq_signal_deliver(struct ksignal *ksig,
> > >    				       struct pt_regs *regs)
> > >    {
> > > -	preempt_disable();
> > > -	__set_bit(RSEQ_EVENT_SIGNAL_BIT, &current->rseq_event_mask);
> > > -	preempt_enable();
> > > +	scoped_guard(RSEQ_EVENT_GUARD)
> > > +		__set_bit(RSEQ_EVENT_SIGNAL_BIT, &current->rseq_event_mask);
> >
> > Then we have more to worry about interaction of the following
> > rseq events with membarrier IPI:
> >
> > - rseq_preempt, rseq_migrate, rseq_signal_deliver.
> >
> > Both rseq_preempt and rseq_migrate are documented as only being required
> > to be called with preempt off, not irq off.
>
> They are always invoked from the middle of the scheduler with interrupts
> disabled, so just the documentation is wrong.

OK

> > Can we just move the event sources requiring preempt-off to their own
> > word, and use a separate word for membarrier IPI instead ? This would
> > allow us to partition the problem into two distinct states each
> > protected by their respective mechanism.
>
> signal delivery can just use set_bit() which is atomic vs. the IPI no?

Good point!

> But as I pointed out in the other series, we don't need that whole zoo
> of event bits at all. There is absolutely no point.
>
> signal delivery does not need to set an event in the first place. It can
> just unconditionally invoke this stuff, no?

Yes, as long as it can take a page fault from there.

Thanks,

Mathieu

> Thanks,
>
>          tglx


--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com