[RFC PATCH 5/5] arm: perf: Use FIQ to handle PMU events.

[daniel.thompson@linaro.org (Daniel Thompson)]

On 19/01/15 17:48, Russell King - ARM Linux wrote:
> On Tue, Jan 13, 2015 at 04:35:31PM +0000, Daniel Thompson wrote:
>> +/*
>> + * This handler is called *unconditionally* from the default NMI/FIQ
>> + * handler. The irq may not be anything to do with us so the main
>> + * job of this function is to figure out if the irq passed in is ours
>> + * or not.
>> + */
>> +void cpu_pmu_handle_fiq(int irq)
>> +{
>> +	int cpu = smp_processor_id();
> 
> This can be either debug_smp_processor_id() or raw_smp_processor_id().
> raw_smp_processor_id() is fine from FIQ contexts, as seems to be
> debug_smp_processor_id(), but only because we guarantee that
> irqs_disabled() in there will be true.

Curiously I was looking at exactly this yesterday (because I was
intrigued why the NMI-safe bits of kgdb use raw_smp_processor_id() but
the x86 arch_trigger_all_cpu_backtrace() implementation uses
smp_processor_id()).

Given the comments make clear smp_processor_id() is the preferred
variant except for false positives I concluded I would continue with
smp_processor_id() for any code I write hanging off the default FIQ
handler. No objections?


>> +
>> +	if (irq != get_cpu_var(cpu_pmu_irqs))
>> +		return;
> 
> get_cpu_var() needs put_cpu_var() to undo its effects.  get_cpu_var()
> calls preempt_disable(), which calls into lockdep...  I think we
> determined that was fine last time we went digging?

Yes. We reviewed lockdep from the point-of-view of RCU and found that
lockdep disabled most of itself when in_nmi() is true.

> put_cpu_var()
> would call preempt_enable() which I'd hope would be safe in FIQ/NMI
> contexts?

Yes.

preempt_count_add/sub form part of the work done by nmi_enter() and
nmi_exit().

However this code gets no benefit from calling get_cpu_var(). I think it
would be better to switch it to this_cpu_ptr.


>> +
>> +	(void)armpmu_dispatch_irq(irq,
>> +				  get_cpu_ptr(&cpu_pmu->hw_events->percpu_pmu));
> 
> Again, get_cpu_xxx() needs to be balanced with a put_cpu_xxx().
> 
>