Re: [RFC] Deadlock via recursive wakeup via RCU with threadirqs

[Peter Zijlstra]

On Fri, Jun 28, 2019 at 03:01:36PM -0500, Scott Wood wrote:
> On Fri, 2019-06-28 at 16:15 +0200, Peter Zijlstra wrote:
> > On Thu, Jun 27, 2019 at 01:36:12PM -0700, Paul E. McKenney wrote:
> > > On Thu, Jun 27, 2019 at 03:17:27PM -0500, Scott Wood wrote:
> > > > On Thu, 2019-06-27 at 11:41 -0700, Paul E. McKenney wrote:
> > > > > Of course, unconditionally refusing to do the wakeup might not be
> > > > > happy
> > > > > thing for NO_HZ_FULL kernels that don't implement IRQ work.
> > > > 
> > > > Couldn't smp_send_reschedule() be used instead?
> > > 
> > > Good point.  If current -rcu doesn't fix things for Sebastian's case,
> > > that would be well worth looking at.  But there must be some reason
> > > why Peter Zijlstra didn't suggest it when he instead suggested using
> > > the IRQ work approach.
> > > 
> > > Peter, thoughts?
> > 
> > I've not exactly kept up with the thread; but irq_work allows you to run
> > some actual code on the remote CPU which is often useful and it is only
> > a little more expensive than smp_send_reschedule().
> > 
> > Also, just smp_send_reschedule() doesn't really do anything without
> > first poking TIF_NEED_RESCHED (or other scheduler state) and if you want
> > to do both, there's other helpers you should use, like resched_cpu().
> 
> resched_cpu() will not send an IPI to the current CPU[1].

Correct, smp_send_reschedule() might not work for self, not all hardware
can self-IPI.

> Plus, the RCU
> code needs to set need_resched even in cases where it doesn't need to send
> the IPI.  And worst of all, resched_cpu() takes the rq lock which is the
> deadlock scenario we're trying to avoid.
> 
> -Scott
> 
> [1] Which makes me nervous about latency if there are any wakeups with irqs
> disabled, without a preempt_enable() after irqs are enabled again, and not
> inside an interrupt.

All good points; and those are all solved with irq_work. That provides a
'clean' IRQ context.