[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH 1/3] xen: timers: don't miss a timer event because of stop_timer()

On Wed, 2017-09-27 at 02:20 -0600, Jan Beulich wrote:
> > > > On 26.09.17 at 20:11, <dario.faggioli@xxxxxxxxxx> wrote:
> > it means there is an event that *is* in progress right now (i.e.,
> > we're
> > stopping the timer on the path that goes from the interrupt that
> > raised
> > TIMER_SOFTIRQ, and the timer softirq handler).
> > 
> > So, basically, I'm trying to achieve exactly what you call
> > reasonable:
> > servicing an event which is in progress. :-)
> I'm afraid I don't understand - if the processing of the timer is
> already in progress, why would you need to raise another
> softirq? Wouldn't it suffice to simply skip deactivate_timer() then?
Ah, yes! In the use case I'm trying to fix, that's indeed not
necessary, as it has already been risen.

Basically, as it's harmless to re-raise it, I thought to do so, with
the aim of making it easier to understand what the code was trying to
achieve. But now I actually agree with you that the effect is quite the
opposite. :-(

I will get rid of the re-raising, and explain the situation better in
the both the comment and the changelog.

> This raising of the softirq is what made me imply that, perhaps
> due to some minimal skew e.g. resulting from rounding, you
> assume the interrupt did not fire yet, which I'd then call the
> timer event not being in progress (yet).
Sure, I totally see it now, and I also totally agree.

> In the end what I think I'm missing is a clear description of an
> actual
> case where the current behavior causes breakage (plus of course
> an explanation why the new behavior is unlikely to cause issues with
> existing users).
So, the problem is that the handler of the RCU idle_timer, introduced
by 2b936ea7b716dc1a13c ("xen: RCU: avoid busy waiting until the end of
grace period."), never runs.

And that is because the following happens:
- the CPU wants to go idle
- sched_tick_suspend()
- the CPU goes idle
  ... ... ...
- RCU_idle_timer's IRQ arrives
- the CPU wakes up
- raise_softirq(TIMER_SOFTIRQ)
- sched_tick_resume()
          remove_entry(RCU_idle_timer) // timer out of heap/list
- do_softirq() (we are inside idle_loop())
- softirq_handlers[TIMER_SOFTIRQ]()
- timer_softirq_action()
    // but the timer is not in the heap/list!

Now, as far as the purpose of that patch goes, we're fine. In fact,
there, we "only" needed to avoid that a certain CPU (because of its
role in the grace period) would sleep too long/indefinitely. And the
fact that the CPU does wake up, because of the timer interrupt, is

However, it's been asked to try to make the logic a bit more clever,
basically for preventing RCU_idle_timer from firing too often. And
that's actually what this series is doing. But now, in order to achieve
that, I do need the timer handler to run.

About the (lack of) breakage of existing use cases. Well, hard to tell
like this, but I'd say that, if, right now, we are not missing any
timer event handling, it means that this situation --where you stop the
timer in between raise_softirq(TIMER_SOFTIRQ) and
softirq_handler[TIMER_SOFTIRQ]()-- never happens.

Inspecting the code, in fact, I can't spot any stop_timer() happening
within that window, which I think it means we're fine (IOW,
RCU_idle_timer is the first, and for now only, timer that is stopped on
the CPU wake-up-from-idle path).

It is important (in the new version of this patch) for deactivation to
be skipped only in stop_timer(), and not, e.g., in kill_timer(). As, if
someone, in future, will want to kill and free the timer during the
window, then in that case the handler *must* not run.

Makes sense?

Thanks and Regards,
<<This happens because I choose it to happen!>> (Raistlin Majere)
Dario Faggioli, Ph.D, http://about.me/dario.faggioli
Senior Software Engineer, Citrix Systems R&D Ltd., Cambridge (UK)

Attachment: signature.asc
Description: This is a digitally signed message part

Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.