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

[Peter Zijlstra]

On Wed, Jan 20, 2016 at 12:48:24PM +0800, Huang Rui wrote:
> Hi Peter,
> 
> Thanks so much to your comments.
> 
> On Tue, Jan 19, 2016 at 01:12:50PM +0100, Peter Zijlstra wrote:
> > On Thu, Jan 14, 2016 at 10:50:08AM +0800, Huang Rui wrote:
> > > +struct power_pmu {
> > > +	spinlock_t		lock;
> > 
> > This should be a raw_spinlock_t, as it'll be nested under other
> > raw_spinlock_t's.
> > 
> 
> Do you mean the following spinlock operations are in hardware
> interrupts disabled case, so I need use raw_spinlock_t instead, right?


			mainline		-rt

raw_spinlock_t		spin-waits		spin-waits
spinlock_t		spin-waits		blocks (rt-mutex)
struct mutex		blocks			blocks (rt-mutex)


since these functions are themselves called with raw_spinlock_t held
(perf_event_context::lock for example, but also rq::lock), any lock
nested inside them must also be raw_spinlock_t.

I have a lockdep patch somewhere that checks these ordering things; I
should rebase and post that again.

> Use raw_spin_lock_irqsave/raw_spin_unlock_irqrestore?

pmu::{start,stop,add,del} will be called with IRQs already disabled.

> > > +static int power_cpu_init(int cpu)
> > > +{
> > > +	int i, cu, ret = 0;
> > > +	cpumask_var_t mask, dummy_mask;
> > > +
> > > +	cu = cpu / cores_per_cu;
> > > +
> > > +	if (!zalloc_cpumask_var(&mask, GFP_KERNEL))
> > > +		return -ENOMEM;
> > > +
> > > +	if (!zalloc_cpumask_var(&dummy_mask, GFP_KERNEL)) {
> > > +		ret = -ENOMEM;
> > > +		goto out;
> > > +	}
> > > +
> > > +	for (i = 0; i < cores_per_cu; i++)
> > > +		cpumask_set_cpu(i, mask);
> > > +
> > > +	cpumask_shift_left(mask, mask, cu * cores_per_cu);
> > > +
> > > +	if (!cpumask_and(dummy_mask, mask, &cpu_mask))
> > > +		cpumask_set_cpu(cpu, &cpu_mask);
> > > +
> > > +	free_cpumask_var(dummy_mask);
> > > +out:
> > > +	free_cpumask_var(mask);
> > > +
> > > +	return ret;
> > > +}
> > 
> > > +static int power_cpu_notifier(struct notifier_block *self,
> > > +			      unsigned long action, void *hcpu)
> > > +{
> > > +	unsigned int cpu = (long)hcpu;
> > > +
> > > +	switch (action & ~CPU_TASKS_FROZEN) {
> > > +	case CPU_UP_PREPARE:
> > > +		if (power_cpu_prepare(cpu))
> > > +			return NOTIFY_BAD;
> > > +		break;
> > > +	case CPU_STARTING:
> > > +		if (power_cpu_init(cpu))
> > > +			return NOTIFY_BAD;
> > 
> > this is called with IRQs disabled, which makes those GFP_KERNEL allocs
> > above a pretty bad idea.
> > 
> 
> Right, so should I use GFP_ATOMIC to allocate cpumask here?

One should not use GFP_ATOMIC if at all possible, also no, -rt cannot do
_any_ allocations from this site.

> > Also, note that -rt cannot actually do _any_ allocations/frees from
> > STARTING.
> > 
> > Please move the allocs/frees to PREPARE/ONLINE.
> > 
> 
> How about add two cpumask_var_t at power_pmu structure? Then allocate
> the two cpumask_var_t (pmu->mask, pmu->dummy_mask), and they can be
> also used on power_cpu_init.

That would work.