percpu irq APIs and perf

[marc.zyngier@arm.com (Marc Zyngier)]

Hi Vinnet,

On 10/12/15 09:25, Vineet Gupta wrote:
> Hi Marc / Daniel / Jason,
> 
> I had a couple of questions about percpu irq API, hopefully you can help answer.
> 
> On ARM, how do u handle requesting per cpu IRQs - specifically usage
> of request_percpu_irq() / enable_percpu_irq() API.
> It seems, for using them, we obviously need to explicitly set irq as
> percpu and as a consequence explicitly enable autoen (since former
> disables that). See arch/arc/kernel/irq.c: arc_request_percpu_irq()
> called by ARC per cpu timer setup.

Indeed. The interrupt controller code flags these interrupts as being
per-cpu, and we do rely on each CPU performing an enable_percpu_irq().

So the way the whole thing flows is as such:
- Interrupt controller (GIC) flags the PPIs (Private Peripheral
Interrupt) as per-CPU (hwirq 16 to 31 are replicated per CPU) very early
in the boot process
- request_percpu_irq() only occurs once, usually on the boot CPU (but
that's not a requirement)
- each CPU executes enable_percpu_irq(), which touches per-CPU
registers. This usually involves a CPU notifier to enable/disable the
interrupt when hotplug is on.


>     if (!cpu)  {
>            irq_set_percpu_devid()   <--- disables AUTOEN
>            irq_modify_status(IRQ_NOAUTOEN)  <-- to undo side-effect of above
>            request_percpu_irq
>     }
>     enable_percpu_irq
> 
> I don't see pattern in general for drivers/clocksource/ and/or
> arm_arch_timer.c for PPI case.

You can have a look at arch/arm/smp/smp_twd.c which is probably less
cryptic.

> Further there is an ordering requirement as in request_percpu_irq()
> needs to be called only for the first calling core, and
> enable_percpu_irq() on each one. If enable is done ahead of request
> it obviously fails.

Yup.

> For ARC I've historically used a wrapper arc_request_percpu_irq()
> [pseudo code above] - which has an inherent assumption (now realize
> fragile) that it will be called on core0 first thus guaranteeing the
> ordering above. This is true for timer, IPI etc but not for other
> late probed peripherals - specially perf.
> 
> Infact ARC perf probe open codes on_each_cpu() to ensure irq request
> is done locally first.
> 
> But this all falls apart, when perf probe happens on coreX (not
> core0), causing enable to be called ahead of request anyways. This is
> what I'm running into now.
> 
> I think the solution is to call request_percpu_irq() on whatever core
> hits first and call enable_percpu_irq() from a cpu up notifier. But I
> think the notifier won't run on boot cpu ?  Or is there a better way
> to clean up all this mess.

I think that's pretty much it.

See drivers/perf/arm_pmu.c::cpu_pmu_request_irq() for example.

> FWIW, I see this issue on 3.18 kernel but not latest 4.4-rcX because
> in 3.18 arc perf probe invariably happens on coreX (due to init task
> migration right after clocksource switch - something which doesn't
> happen on 4.4 likely due to recent timer core changes).

Hope this helps,

	M.
-- 
Jazz is not dead. It just smells funny...