Re: [PATCH v7 2/2] clocksource: add J-Core timer/clocksource driver

From: Rich Felker
Date: Sat Oct 08 2016 - 21:32:09 EST


On Sat, Oct 08, 2016 at 07:03:30PM +0200, Thomas Gleixner wrote:
> On Sat, 8 Oct 2016, Rich Felker wrote:
> > On Sat, Oct 08, 2016 at 01:32:06PM +0200, Thomas Gleixner wrote:
> > > CPU spins and waits for an interrupt to happen
> > >
> > >
> > > <idle>-0 [000] d... 150.841530: rcu_dyntick: End 0 1
> > >
> > > Dropping out of the spin about the time we expect the PIT interrupt to
> > > fire. And an interrupt is the reason why we drop out because the 'need
> > > resched' flag is not set and we end up in:
> >
> > OK, I missed that.
> >
> > > <idle>-0 [000] d.s. 150.841724: tick_irq_enter: update jiffies via irq
> > >
> > > which is called from irq_enter()
> > >
> > > <idle>-0 [000] d.s. 150.841918: tick_do_update_jiffies64: finished do_timer(1)
> > > <idle>-0 [000] d.s. 150.842348: tick_do_update_jiffies64: finished updating jiffies
> > >
> > >
> > > So here we would expect:
> > >
> > > irq_handler_entry: irq=16 name=jcore_pit
> > > ...
> > > irq_handler_exit .....
> > >
> > >
> > > or any other interrupt being invoked. But we see this here:
> >
> > According to the trace the 'irq' is soft ('s'). I couldn't find the
> > code path from the idle loop to softirq but so maybe this is a bug. Is
> > it possible that in some cases the arch irq entry does not get
> > identified as a hard irq or traced and then the subsequent tick code
> > thinks it's running in a softirq and behaves differently? I could add
> > more tracing around irq entry.
>
> No.
>
> irq_enter()
> if (is_idle_task(current) && !in_interrupt()) {
> local_bh_disable();
> tick_irq_enter();
> local_bh_enable();
> }
> __irq_enter()
> preempt_count_add(HARDIRQ_OFFSET);
>
> So the 's' comes from local_bh_disable() and the code does not behave
> special because of that. It's just always that way. Look at all the other
> instances of tick_irq_enter() which do not show that issue.

Thank you -- that was really confusing me. Now that I know there's
actually a hardirq handling I know where to look for the problem.

> > > <idle>-0 [000] d... 150.842603: __tick_nohz_idle_enter: can stop idle tick
> > >
> > > And that's just wrong.
> >
> > Can you explain?
>
> Because you drop out the idle spin due to an interrupt, but no interrupt is
> handled according to the trace. You just go back to sleep and the trace is
> full of this behaviour.

Found the problem. IRQD_IRQ_INPROGRESS is causing the interrupt to be
ignored if the same interrupt is being handled on the other cpu.
Modifying the jcore aic driver to call handle_percpu_irq rather than
handle_simple_irq when the irq was registered as percpu solves the
problem, but I'm actually concerned that handle_simple_irq would also
be wrong in the case of non-percpu irqs if they could be delivered on
either core -- if another irq arrives after the driver is finished
checking for new device status, but before the IRQD_IRQ_INPROGRESS
flag is removed, it seems handle_simple_irq loses the irq. This is not
a problem for the current hardware since non-percpu irqs always arrive
on cpu0, but I'd rather improve the driver to properly handle possible
future hardware that allows irq delivery on any core.

> > > Both CPUs have the same interrupt number for their per cpu PIT interrupt.
> > >
> > > IIRC you said earlier when the percpu interrupt discussion happened, that
> > > your per cpu PITs have distinct interrupt numbers, but that does not seem
> > > the case here.
> >
> > No, I said the opposite. They use the same irq number but they're each
> > wired to their own cpu, and there's no way for them to fire on the
> > wrong cpu.
>
> Your patch does:
>
> > + err = request_irq(pit_irq, jcore_timer_interrupt,
> > + IRQF_TIMER | IRQF_PERCPU,
> > + "jcore_pit", jcore_pit_percpu);
>
> which is the wrong thing to do. You need request_percpu_irq() here, but of
> course with the irq chip you are using, which has every callback as a noop,
> it does not matter at all. So that's not an issue and not the reason for
> this wreckage.

Mentioning this was helpful to get me looking at the right things
tracking from the irq entry point to where the irq was lost/dropped,
so thanks for bringing it up. request_irq ends up working fine still
because IRQF_PERCPU causes __setup_irq to mark the irqdesc/irq_data as
percpu, so that my handle function can treat it differently. Here's
the patch that has it working:

diff --git a/drivers/irqchip/irq-jcore-aic.c b/drivers/irqchip/irq-jcore-aic.c
index 5e5e3bb..b53a8a5 100644
--- a/drivers/irqchip/irq-jcore-aic.c
+++ b/drivers/irqchip/irq-jcore-aic.c
@@ -25,12 +25,20 @@

static struct irq_chip jcore_aic;

+static void handle_jcore_irq(struct irq_desc *desc)
+{
+ if (irq_is_percpu(irq_desc_get_irq(desc)))
+ handle_percpu_irq(desc);
+ else
+ handle_simple_irq(desc);
+}
+
static int jcore_aic_irqdomain_map(struct irq_domain *d, unsigned int irq,
irq_hw_number_t hwirq)
{
struct irq_chip *aic = d->host_data;

- irq_set_chip_and_handler(irq, aic, handle_simple_irq);
+ irq_set_chip_and_handler(irq, aic, handle_jcore_irq);

return 0;
}

After some more testing I'll send this patch or something similar.

> But what you need to figure out is why this happens:
>
> 100.00x hrtimer_start(expires = 100.01)
> 100.00x idle spin
> 100.01x tick_irq_enter()
> 100.01x tick_stop()
>
> i.e. why do you drop out of your idle spin, take the low level interrupt
> entry path which calls irq_enter() -> tick_irq_enter() and then you do not
> handle any interrupt at all and drop back into the idle loop. That's the
> real question and that's a problem in your architecture low level interrupt
> entry code or some hardware related issue. But certainly not a bug in the
> core tick code.
>
> You can come up with another boat load of weird theories about bugs in
> that code, but I won't even have a look _before_ you can explain why the
> above sequence happens and the PIT timer interrupt is not handled.

Apologies for this big distraction for what was ultimately a driver
bug on my side. I was misled by the way it seemed to be a regression,
since the symptoms did not appear in earlier kernel versions for
whatever reason (likely just chance).

Rich