On 31/03/15 18:08, Mark Rutland wrote:
> Hi Daniel,
>
> I'd very much like to see us converge on a solution for this soon. The
> existing hack is getting in the way of other rework of the arm/arm64
> perf code.
I'd quite like to see this patch sorted out too (mostly because one o my
"go to" devices still crashes horribly whenever I try to do any
profiling on a mainline kernel).
> I think the approach this patch takes should work, but there are some
> parts that can be cleaned up (hopefully mostly cosmetic). Unfortunately
> I don't seem to have a relevant platform for testing on.
>
> [...]
>
>> diff --git a/arch/arm/include/asm/pmu.h b/arch/arm/include/asm/pmu.h
>> index b1596bd59129..dfef7904b790 100644
>> --- a/arch/arm/include/asm/pmu.h
>> +++ b/arch/arm/include/asm/pmu.h
>> @@ -87,6 +87,14 @@ struct pmu_hw_events {
>> * already have to allocate this struct per cpu.
>> */
>> struct arm_pmu *percpu_pmu;
>> +
>> +#ifdef CONFIG_SMP
>> + /*
>> + * This is used to schedule workaround logic on platforms where all
>> + * the PMUs are attached to a single SPI.
>> + */
>> + struct irq_work work;
>> +#endif
>> };
>>
>> struct arm_pmu {
>> @@ -117,6 +125,10 @@ struct arm_pmu {
>> struct platform_device *plat_device;
>> struct pmu_hw_events __percpu *hw_events;
>> struct notifier_block hotplug_nb;
>> +#ifdef CONFIG_SMP
>> + int muxed_spi_workaround_irq;
>
> There's nothing workaround specific about this IRQ; it's just the only
> IRQ.
>
> I think we should just pre-parse all the IRQs into a list at probe time,
> regardless of SMP or the workaround. Then we can just grab the first
> (and only) interrupt from the list in the workaround paths, and
> otherwise just iterate over the list.
What other uses do you anticipate for such a list? I don't really
understand why we would create a list when we only ever consume the
first entry.
If there is a use for such information then why keep it as a list.
Wouldn't the code be simpler if we add it as a field in the percpu data
structure?
For PPIs all values would be the same, for SPIs all different, for
broken SoCs all after [0] would be 0.
>> + atomic_t remaining_irq_work;
>
> Perhaps remaining_work_irqs? That would make it clear that this is a
> counter rather than a boolean or enumeration. We could s/work/fake/ or
> something to that effect.
I guess the thing we are actually counting is CPUs so
remaining_cpus_with_irq_work would therefore be extremely descriptive.
>> @@ -317,14 +315,9 @@ static irqreturn_t armpmu_dispatch_irq(int irq, void *dev)
>> * dereference.
>> */
>> armpmu = *(void **)dev;
>> - plat_device = armpmu->plat_device;
>> - plat = dev_get_platdata(&plat_device->dev);
>>
>> start_clock = sched_clock();
>> - if (plat && plat->handle_irq)
>> - ret = plat->handle_irq(irq, armpmu, armpmu->handle_irq);
>> - else
>> - ret = armpmu->handle_irq(irq, armpmu);
>> + ret = armpmu->handle_irq(irq, armpmu);
>> finish_clock = sched_clock();
>>
>> perf_sample_event_took(finish_clock - start_clock);
>
> It's nice to see the plat stuff disappearing!
>
>> diff --git a/arch/arm/kernel/perf_event_cpu.c b/arch/arm/kernel/perf_event_cpu.c
>> index 61b53c46edfa..d5bbd79abd4c 100644
>> --- a/arch/arm/kernel/perf_event_cpu.c
>> +++ b/arch/arm/kernel/perf_event_cpu.c
>> @@ -59,6 +59,116 @@ int perf_num_counters(void)
>> }
>> EXPORT_SYMBOL_GPL(perf_num_counters);
>>
>> +#ifdef CONFIG_SMP
>> +
>> +static cpumask_t down_prepare_cpu_mask;
>> +static DEFINE_SPINLOCK(down_prepare_cpu_lock);
>
> I think the names here are a little misleading, because we care about
> the whole window from CPU_DOWN_PREPARE to CPU_DEAD (or DOWN_FAILED). I
> think these would be clearer with s/down_prepare_cpu/dying_cpu/ (though
> admittedly that could also be confused with CPU_DYING).
Fine. I'll change this.
>> +
>> +/*
>> + * Workaround logic that is distributed to all cores if the PMU has only
>> + * a single IRQ and the CPU receiving that IRQ cannot handle it. Its
>> + * job is to try to service the interrupt on the current CPU. It will
>> + * also enable the IRQ again if all the other CPUs have already tried to
>> + * service it.
>> + */
>> +static void cpu_pmu_do_percpu_work(struct irq_work *w)
>
> Perhaps do_muxed_irq_work?
Ok.
>> +{
>> + struct pmu_hw_events *hw_events =
>> + container_of(w, struct pmu_hw_events, work);
>> + struct arm_pmu *cpu_pmu = hw_events->percpu_pmu;
>> +
>> + /* Ignore the return code, we can do nothing useful with it */
>> + (void) cpu_pmu->handle_irq(0, cpu_pmu);
>
> Nit: no space after a cast please.
>
> Do we need the void cast here? Does your toolchain complain?
I've always considered casting to void to be an executable comment... In
this case I added it out of habit rather than because the toolchain
complained.
>> +
>> + if (atomic_dec_and_test(&cpu_pmu->remaining_irq_work))
>> + enable_irq(cpu_pmu->muxed_spi_workaround_irq);
>
> I'm a little uneasy about calling enable_irq here, given we're in IRQ
> context. While it doesn't look we'd be wired up through an irqchip with
> irq_bus_lock or irq_bus_sync_unlock, I'm not sure how safe it is to rely
> on that being the only thing that matters.
>
> It would be nice to hear from someone familiar with the IRQ code on that
> respect.
Sure. I've reviewed this as far as I can.
Probably the most significant difference between enable and disable
paths is the retrigger logic. I can't see anything there that would
cause a problem.
>> +}
>> +
>> +/*
>> + * Workaround for systems where all PMU interrupts are targeting a
>> + * single SPI.
>> + *
>> + * The workaround will disable the interrupt and distribute irqwork to all
>> + * the other processors in the system. Hopefully one of them will clear the
>> + * interrupt...
>> + *
>> + * The workaround is only deployed when all PMU interrupts are aimed
>> + * at a single core. As a result the workaround is never re-entered
>> + * making it safe for us to use static data to maintain state.
>> + */
>> +static void cpu_pmu_deploy_muxed_spi_workaround(struct arm_pmu *cpu_pmu)
>
> Perhaps distribute_muxed_irq?
Could do. Personally I prefer to be clear that this is deploying a
workaround (i.e. implying it does something odd on a small number of
SoCs) rather than this code being part of the normal case.
>> +{
>> + static cpumask_t irqwork_mask;
>> + int cpu;
>> +
>> + disable_irq_nosync(cpu_pmu->muxed_spi_workaround_irq);
>> + spin_lock(&down_prepare_cpu_lock);
>> +
>> + /*
>> + * Combining cpu_online_mask and down_prepare_cpu_mask gives
>> + * us the CPUs that are currently online and cannot die until
>> + * we release down_prepare_cpu_lock.
>> + */
>> + cpumask_andnot(&irqwork_mask, cpu_online_mask, &down_prepare_cpu_mask);
>> + cpumask_clear_cpu(smp_processor_id(), &irqwork_mask);
>> + atomic_add(cpumask_weight(&irqwork_mask), &cpu_pmu->remaining_irq_work);
>> +
>> + for_each_cpu(cpu, &irqwork_mask) {
>> + struct pmu_hw_events *hw_events =
>> + per_cpu_ptr(cpu_pmu->hw_events, cpu);
>> +
>> + if (!irq_work_queue_on(&hw_events->work, cpu))
>> + if (atomic_dec_and_test(&cpu_pmu->remaining_irq_work))
>> + enable_irq(cpu_pmu->muxed_spi_workaround_irq);
>> + }
>> +
>> + spin_unlock(&down_prepare_cpu_lock);
>> +}
>
> I think this works, given the notifier logic and hotplug_cfd flushing
> any pending irq_work items.
Certainly its a lot better than the benign race thing I was trying to do
before!
>> +
>> +/*
>> + * Called when the main interrupt handler cannot determine the source
>> + * of interrupt. It will deploy a workaround if we are running on an SMP
>> + * platform with only a single muxed SPI.
>> + */
>> +static irqreturn_t cpu_pmu_handle_irq_none(int irq_num, struct arm_pmu *cpu_pmu)
>> +{
>> + if (irq_num != cpu_pmu->muxed_spi_workaround_irq)
>> + return IRQ_NONE;
>
> This is somewhat opaque.
>
> I'd rather just have a flag to determine when we need to do any special
> handling for the muxed case (or better, swizzle the irq handler to a
> wrapper that pings the other CPUs and calls the usual handler).
Using a different handler for the workaround makes sense.
IIRC I avoided that previously because the way the code is currently
split between perf_event.c and perf_event_cpu.c didn't make that very
natural.
I'll look at this again.
> [...]
>
>> +static int cpu_pmu_muxed_spi_workaround_init(struct arm_pmu *cpu_pmu)
>> +{
>> + struct platform_device *pmu_device = cpu_pmu->plat_device;
>> +
>> + atomic_set(&cpu_pmu->remaining_irq_work, 0);
>
> Then we can move this atomic_set into the usual init path for SMP and
> get rid of these init/term functions.
Ok.
> [...]
>
>> diff --git a/arch/arm/kernel/perf_event_v7.c b/arch/arm/kernel/perf_event_v7.c
>> index 8993770c47de..0dd914c10803 100644
>> --- a/arch/arm/kernel/perf_event_v7.c
>> +++ b/arch/arm/kernel/perf_event_v7.c
>> @@ -792,7 +792,7 @@ static irqreturn_t armv7pmu_handle_irq(int irq_num, void *dev)
>> * Did an overflow occur?
>> */
>> if (!armv7_pmnc_has_overflowed(pmnc))
>> - return IRQ_NONE;
>> + return cpu_pmu_handle_irq_none(irq_num, cpu_pmu);
>
> Won't this leave samples skewed towards the CPU the interrupt is affine
> to? If you're counting something like cycles with a short enough period
> (and therefore effectively always have something to handle on the local
> CPU), we might never ping the other CPUs.
Do we really care about fidelity of results when servicing a perf
interrupt causes another perf interrupt to happen ;-) ? At this point
perf has already stopped profiling anything useful.
> I think we always need to ping the other CPUs, regardless of whether
> there was something to handle on this CPU.
Ok.
If we make this unconditional then we can, as you suggest above
somewhere, redistribute the interrupt before trying to handle it. The
faster we can interrupt the other CPUs the higher the fidelity of the
profile (less time for PC to change between perf event triggering and
CPU reacting).
Daniel.
> >> struct arm_pmu {
> >> @@ -117,6 +125,10 @@ struct arm_pmu {
> >> struct platform_device *plat_device;
> >> struct pmu_hw_events __percpu *hw_events;
> >> struct notifier_block hotplug_nb;
> >> +#ifdef CONFIG_SMP
> >> + int muxed_spi_workaround_irq;
> >
> > There's nothing workaround specific about this IRQ; it's just the only
> > IRQ.
> >
> > I think we should just pre-parse all the IRQs into a list at probe time,
> > regardless of SMP or the workaround. Then we can just grab the first
> > (and only) interrupt from the list in the workaround paths, and
> > otherwise just iterate over the list.
>
> What other uses do you anticipate for such a list? I don't really
> understand why we would create a list when we only ever consume the
> first entry.
It would go parallel to the affinity property Will recently added, and
would allow us to separate the DT parsing from the request/free paths,
making those far easier to understand (and likely a little faster too).
> If there is a use for such information then why keep it as a list.
> Wouldn't the code be simpler if we add it as a field in the percpu data
> structure?
>
> For PPIs all values would be the same, for SPIs all different, for
> broken SoCs all after [0] would be 0.
That's a nice idea. That would make the affinity info implicit, and
could be far neater.
> >> + atomic_t remaining_irq_work;
> >
> > Perhaps remaining_work_irqs? That would make it clear that this is a
> > counter rather than a boolean or enumeration. We could s/work/fake/ or
> > something to that effect.
>
> I guess the thing we are actually counting is CPUs so
> remaining_cpus_with_irq_work would therefore be extremely descriptive.
It would be nice if we could come up with something that was a little
more concise. Perhaps cpus_with_work would be good enough?
> >> +}
> >> +
> >> +/*
> >> + * Workaround for systems where all PMU interrupts are targeting a
> >> + * single SPI.
> >> + *
> >> + * The workaround will disable the interrupt and distribute irqwork to all
> >> + * the other processors in the system. Hopefully one of them will clear the
> >> + * interrupt...
> >> + *
> >> + * The workaround is only deployed when all PMU interrupts are aimed
> >> + * at a single core. As a result the workaround is never re-entered
> >> + * making it safe for us to use static data to maintain state.
> >> + */
> >> +static void cpu_pmu_deploy_muxed_spi_workaround(struct arm_pmu *cpu_pmu)
> >
> > Perhaps distribute_muxed_irq?
>
> Could do. Personally I prefer to be clear that this is deploying a
> workaround (i.e. implying it does something odd on a small number of
> SoCs) rather than this code being part of the normal case.
I'd prefer the more succinct name. We already have the comment stating
that this is a workaround.
[...]
> >> +/*
> >> + * Called when the main interrupt handler cannot determine the source
> >> + * of interrupt. It will deploy a workaround if we are running on an SMP
> >> + * platform with only a single muxed SPI.
> >> + */
> >> +static irqreturn_t cpu_pmu_handle_irq_none(int irq_num, struct arm_pmu *cpu_pmu)
> >> +{
> >> + if (irq_num != cpu_pmu->muxed_spi_workaround_irq)
> >> + return IRQ_NONE;
> >
> > This is somewhat opaque.
> >
> > I'd rather just have a flag to determine when we need to do any special
> > handling for the muxed case (or better, swizzle the irq handler to a
> > wrapper that pings the other CPUs and calls the usual handler).
>
> Using a different handler for the workaround makes sense.
>
> IIRC I avoided that previously because the way the code is currently
> split between perf_event.c and perf_event_cpu.c didn't make that very
> natural.
>
> I'll look at this again.
The split between the two is an unfortunate legacy from when it looked
like system PMUs could reuse the code. We recently decoupled the CCI PMU
from arm_pmu so that we could fuse the two layers and get rid of a load
of indirection.
I have some patches doing that. I'll try to rework them next week given
it sounds like they could be a useful base for this.
> > [...]
> >
> >> diff --git a/arch/arm/kernel/perf_event_v7.c b/arch/arm/kernel/perf_event_v7.c
> >> index 8993770c47de..0dd914c10803 100644
> >> --- a/arch/arm/kernel/perf_event_v7.c
> >> +++ b/arch/arm/kernel/perf_event_v7.c
> >> @@ -792,7 +792,7 @@ static irqreturn_t armv7pmu_handle_irq(int irq_num, void *dev)
> >> * Did an overflow occur?
> >> */
> >> if (!armv7_pmnc_has_overflowed(pmnc))
> >> - return IRQ_NONE;
> >> + return cpu_pmu_handle_irq_none(irq_num, cpu_pmu);
> >
> > Won't this leave samples skewed towards the CPU the interrupt is affine
> > to? If you're counting something like cycles with a short enough period
> > (and therefore effectively always have something to handle on the local
> > CPU), we might never ping the other CPUs.
>
> Do we really care about fidelity of results when servicing a perf
> interrupt causes another perf interrupt to happen ;-) ? At this point
> perf has already stopped profiling anything useful.
Fair point. I'll have to consider that for a bit...
> > I think we always need to ping the other CPUs, regardless of whether
> > there was something to handle on this CPU.
>
> Ok.
>
> If we make this unconditional then we can, as you suggest above
> somewhere, redistribute the interrupt before trying to handle it. The
> faster we can interrupt the other CPUs the higher the fidelity of the
> profile (less time for PC to change between perf event triggering and
> CPU reacting).
Yup!
Mark.