Re: [PATCH v2 5/5] perf/x86/amd/power: Add AMD accumulated power reporting mechanism

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jan 21, 2016 at 03:04:38PM +0800, Huang Rui wrote:
> I just quickly looked at about the spinlock on -rt mode. Because
> realtime linux kernel provides two kinds of spinlock, the original
> spinlock_t will be replaced the one which is able to sleep, actually,
> like mutex. And another one (you mentioned here, raw_spinlock_t) can
> keep on non-sleep behavior, that is the real spinlock.
> 
> And my lock here also will be nested under perf_event_context::lock,
> right?

Yep.

> > I have a lockdep patch somewhere that checks these ordering things; I
> > should rebase and post that again.
> > 
> 
> Can you CC me when you post that patch next time?

Sure.

> > One should not use GFP_ATOMIC if at all possible, also no, -rt cannot do
> > _any_ allocations from this site.
> > 
> 
> OK, that's because allocation might sleep when IRQ disabled. That's
> incorrect.

Right.

Its related to the above, the allocator locks are spinlock_t and as a
consequence of them becoming a blocking lock, spin_lock_irq() will also
no longer disable IRQs.

The CPU_STARTING notifier however will still be called with IRQs
disabled because it is CPU bringup.

So on -rt even GFP_ATOMIC will no longer work here.

> I draft an update diff that based on original patch, please take a
> look.
> 
> 8<--------------------------------------------------------------------------
> 
> diff --git a/arch/x86/kernel/cpu/perf_event_amd_power.c b/arch/x86/kernel/cpu/perf_event_amd_power.c
> index 69ef234..e71d993 100644
> --- a/arch/x86/kernel/cpu/perf_event_amd_power.c
> +++ b/arch/x86/kernel/cpu/perf_event_amd_power.c
> @@ -46,10 +46,17 @@ static unsigned int cu_num;
>  static u64 max_cu_acc_power;
>  
>  struct power_pmu {
> -	spinlock_t		lock;
> +	raw_spinlock_t		lock;
>  	struct list_head	active_list;
>  	struct pmu		*pmu; /* pointer to power_pmu_class */
>  	local64_t		cpu_sw_pwr_ptsc;
> +	/*
> +	 * These two cpumasks is used for avoiding the allocations on
> +	 * CPU_STARTING phase. Because power_cpu_prepare will be
> +	 * called on IRQs disabled status.
> +	 */
> +	cpumask_var_t		mask;
> +	cpumask_var_t		tmp_mask;
>  };
>  
>  static struct pmu pmu_class;
> @@ -126,9 +133,9 @@ static void pmu_event_start(struct perf_event *event, int mode)
>  	struct power_pmu *pmu = __this_cpu_read(amd_power_pmu);
>  	unsigned long flags;
>  
> -	spin_lock_irqsave(&pmu->lock, flags);
> +	raw_spin_lock_irqsave(&pmu->lock, flags);
>  	__pmu_event_start(pmu, event);
> -	spin_unlock_irqrestore(&pmu->lock, flags);
> +	raw_spin_unlock_irqrestore(&pmu->lock, flags);
>  }
>  
>  static void pmu_event_stop(struct perf_event *event, int mode)
> @@ -137,7 +144,7 @@ static void pmu_event_stop(struct perf_event *event, int mode)
>  	struct hw_perf_event *hwc = &event->hw;
>  	unsigned long flags;
>  
> -	spin_lock_irqsave(&pmu->lock, flags);
> +	raw_spin_lock_irqsave(&pmu->lock, flags);
>  
>  	/* mark event as deactivated and stopped */
>  	if (!(hwc->state & PERF_HES_STOPPED)) {
> @@ -155,7 +162,7 @@ static void pmu_event_stop(struct perf_event *event, int mode)
>  		hwc->state |= PERF_HES_UPTODATE;
>  	}
>  
> -	spin_unlock_irqrestore(&pmu->lock, flags);
> +	raw_spin_unlock_irqrestore(&pmu->lock, flags);
>  }
>  
>  static int pmu_event_add(struct perf_event *event, int mode)
> @@ -164,14 +171,14 @@ static int pmu_event_add(struct perf_event *event, int mode)
>  	struct hw_perf_event *hwc = &event->hw;
>  	unsigned long flags;
>  
> -	spin_lock_irqsave(&pmu->lock, flags);
> +	raw_spin_lock_irqsave(&pmu->lock, flags);
>  
>  	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
>  
>  	if (mode & PERF_EF_START)
>  		__pmu_event_start(pmu, event);
>  
> -	spin_unlock_irqrestore(&pmu->lock, flags);
> +	raw_spin_unlock_irqrestore(&pmu->lock, flags);
>  
>  	return 0;
>  }

So for these 4 {start,stop,add,del} you can drop the irqsave/irqrestore
thing as its guaranteed that IRQs will be disabled.

> +	cpumask_clear(pmu->mask);
> +	cpumask_clear(pmu->tmp_mask);
>  
>  	for (i = 0; i < cores_per_cu; i++)
> +		cpumask_set_cpu(i, pmu->mask);
>  
> +	cpumask_shift_left(pmu->mask, pmu->mask, cu * cores_per_cu);

Couldn't you simply use topology_sibling_cpumask(cpu) instead?

>  
>  static int power_cpu_init(int cpu)
>  {
> +	struct power_pmu *pmu = per_cpu(amd_power_pmu, cpu);
> +	int i, cu;
>  
> +	if (pmu)
> +		return 0;
>  
> +	cu = cpu / cores_per_cu;
>  
>  	for (i = 0; i < cores_per_cu; i++)
> +		cpumask_set_cpu(i, pmu->mask);
>  
> +	cpumask_shift_left(pmu->mask, pmu->mask, cu * cores_per_cu);

topology_sibling_cpumask(cpu) again?

>  
> +	if (!cpumask_and(pmu->tmp_mask, pmu->mask, &cpu_mask))
>  		cpumask_set_cpu(cpu, &cpu_mask);
>  
> +	return 0;
>  }
>  
>  static int power_cpu_prepare(int cpu)
>  {
>  	struct power_pmu *pmu = per_cpu(amd_power_pmu, cpu);
>  	int phys_id = topology_physical_package_id(cpu);
> +	int ret = 0;
>  
>  	if (pmu)
>  		return 0;
> @@ -391,7 +380,17 @@ static int power_cpu_prepare(int cpu)
>  	if (!pmu)
>  		return -ENOMEM;
>  
> +	if (!zalloc_cpumask_var(&pmu->mask, GFP_KERNEL)) {
> +		ret = -ENOMEM;
> +		goto out;
> +	}
> +
> +	if (!zalloc_cpumask_var(&pmu->tmp_mask, GFP_KERNEL)) {
> +		ret = -ENOMEM;
> +		goto out1;
> +	}
> +
> +	raw_spin_lock_init(&pmu->lock);
>  
>  	INIT_LIST_HEAD(&pmu->active_list);
>  
> @@ -400,12 +399,21 @@ static int power_cpu_prepare(int cpu)
>  	per_cpu(amd_power_pmu, cpu) = pmu;
>  
>  	return 0;
> +
> +out1:
> +	free_cpumask_var(pmu->mask);
> +out:
> +	kfree(pmu);
> +
> +	return ret;
>  }
>  
>  static void power_cpu_kfree(int cpu)
>  {
>  	struct power_pmu *pmu = per_cpu(amd_power_pmu, cpu);
>  
> +	free_cpumask_var(pmu->mask);
> +	free_cpumask_var(pmu->tmp_mask);
>  	kfree(pmu);
>  }

Yes this should work I think.