Re: [PATCH 2/3] powerpc/sysfs: Show idle_purr and idle_spurr for every CPU

[Naveen N. Rao]

Gautham R Shenoy wrote:
> Hi Naveen,
>
> On Thu, Dec 05, 2019 at 10:23:58PM +0530, Naveen N. Rao wrote:
> > >diff --git a/arch/powerpc/kernel/sysfs.c b/arch/powerpc/kernel/sysfs.c
> > >index 80a676d..42ade55 100644
> > >--- a/arch/powerpc/kernel/sysfs.c
> > >+++ b/arch/powerpc/kernel/sysfs.c
> > >@@ -1044,6 +1044,36 @@ static ssize_t show_physical_id(struct device *dev,
> > > }
> > > static DEVICE_ATTR(physical_id, 0444, show_physical_id, NULL);
> > >
> > >+static ssize_t idle_purr_show(struct device *dev,
> > >+			      struct device_attribute *attr, char *buf)
> > >+{
> > >+	struct cpu *cpu = container_of(dev, struct cpu, dev);
> > >+	unsigned int cpuid = cpu->dev.id;
> > >+	struct lppaca *cpu_lppaca_ptr = paca_ptrs[cpuid]->lppaca_ptr;
> > >+	u64 idle_purr_cycles = be64_to_cpu(cpu_lppaca_ptr->wait_state_cycles);
> > >+
> > >+	return sprintf(buf, "%llx\n", idle_purr_cycles);
> > >+}
> > >+static DEVICE_ATTR_RO(idle_purr);
> > >+
> > >+DECLARE_PER_CPU(u64, idle_spurr_cycles);
> > >+static ssize_t idle_spurr_show(struct device *dev,
> > >+			       struct device_attribute *attr, char *buf)
> > >+{
> > >+	struct cpu *cpu = container_of(dev, struct cpu, dev);
> > >+	unsigned int cpuid = cpu->dev.id;
> > >+	u64 *idle_spurr_cycles_ptr = per_cpu_ptr(&idle_spurr_cycles, cpuid);
> >
> > Is it possible for a user to read stale values if a particular cpu is in an
> > extended cede? Is it possible to use smp_call_function_single() to force the
> > cpu out of idle?
>
> Yes, if the CPU whose idle_spurr cycle is being read is still in idle,
> then we will miss reporting the delta spurr cycles for this last
> idle-duration. Yes, we can use an smp_call_function_single(), though
> that will introduce IPI noise. How often will idle_[s]purr be read ?

Since it is possible for a cpu to go into extended cede for multiple 
seconds during which time it is possible to mis-report utilization, I 
think it is better to ensure that the sysfs interface for idle_[s]purr 
report the proper values through use of IPI.

With repect to lparstat, the read interval is user-specified and just 
gets passed onto sleep().

- Naveen