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

From: Peter Zijlstra
Date: Mon Jul 01 2019 - 10:01:25 EST


On Mon, Jul 01, 2019 at 05:23:05AM -0700, Paul E. McKenney wrote:
> On Mon, Jul 01, 2019 at 12:24:42PM +0200, Sebastian Andrzej Siewior wrote:
> > On 2019-07-01 11:42:15 [+0200], Peter Zijlstra wrote:
> > > I'm not sure if smp_send_reschedule() can be used as self-IPI, some
> > > hardware doesn't particularly like that IIRC. That is, hardware might
> > > only have interfaces to IPI _other_ CPUs, but not self.
> > >
> > > The normal scheduler code takes care to not call smp_send_reschedule()
> > > to self.
> >
> > and irq_work:
> > 471ba0e686cb1 ("irq_work: Do not raise an IPI when queueing work on the local CPU")
>
> OK, so it looks like I will need to use something else. But thank you
> for calling my attention to this commit.

I think that commit is worded slight confusing -- sorry I should've paid
more attention.

irq_work _does_ work locally, and arch_irq_work_raise() must self-IPI,
otherwise everything is horribly broken.

But what happened, was that irq_work_queue() and irq_work_queue_on(.cpu
= smp_processor_id()) wasn't using the same code, and the latter would
try to self-IPI through arch_send_call_function_single_ipi().

Nick fixed that so that irq_work_queue() and irq_work_queue_on(.cpu =
smp_processor_id() now both use arch_raise_irq_work() and remote stuff
uses arch_send_call_function_single_ipi().