Re: [PATCH v2 2/2] powercap/rapl: reduce ipi calls

From: Borislav Petkov
Date: Wed Jan 13 2016 - 16:26:24 EST


On Wed, Jan 13, 2016 at 12:10:03PM -0800, Jacob Pan wrote:
> > > rdmsr_on_cpu(cpu, MSR_IA32_THERM_CONTROL, &l, &h);
> > > if (newstate == DC_DISABLE) {
> > > pr_debug("CPU#%d disabling modulation\n", cpu);
> > > wrmsr_on_cpu(cpu, MSR_IA32_THERM_CONTROL, l &
> > > ~(1<<4), h); } else {
> > > pr_debug("CPU#%d setting duty cycle to %d%%\n",
> > > cpu, ((125 * newstate) / 10));
> > > /* bits 63 - 5 : reserved
> > > * bit 4 : enable/disable
> > > * bits 3-1 : duty cycle
> > > * bit 0 : reserved
> > > */
> > > l = (l & ~14);
> > > l = l | (1<<4) | ((newstate & 0x7)<<1);
> > > wrmsr_on_cpu(cpu, MSR_IA32_THERM_CONTROL, l, h);
> > > }
> >
> > This cannot be converted because you need to do the stuff between the
> > rdmsr_on_cpu() and wrmsr_on_cpu() calls.
> >
> it can be converted if move the below if statement outside read/write
> pair.
> if (newstate == DC_DISABLE) {

You mean something like this (I'm having hard time even figuring out what
goes where):

if (newstate == DC_DISABLE) {
pr_debug("CPU#%d disabling modulation\n", cpu);
rmwmsrl_safe_on_cpu(cpu, MSR_IA32_THERM_CONTROL, (1 << 4), 0);
} else {
pr_debug("CPU#%d setting duty cycle to %d%%\n", cpu, ((125 * newstate) / 10));
rmwmsrl_safe_on_cpu(cpu, MSR_IA32_THERM_CONTROL, 14, (1 << 4) | ((newstate & 0x7)<<1));
}

Now this is *absolutely* unreadable and hard to use. The previous
version at least showed what happens to which bits. This call site will
make everyone go look at the definition of rmwmsrl_safe_on_cpu() and see
what those last two arguments do actually.

And, again, for the n-th time, this still doesn't work if you need to do
other stuff between the rdmsr and wrmsr. So your interface will cover
*some* cases but not all. So people should do rmwmsrl_safe_on_cpu() but
not always - only if they don't need to do stuff between the reads and
the writes.

Hmm, no thanks.

> > > static int sfi_cpufreq_target(struct cpufreq_policy *policy,
> > > unsigned int index) {
> > > ...
> > >
> > > rdmsr_on_cpu(policy->cpu, MSR_IA32_PERF_CTL, &lo, &hi);
> > > lo = (lo & ~INTEL_PERF_CTL_MASK) |
> > > ((u32) sfi_cpufreq_array[next_perf_state].ctrl_val &
> > > INTEL_PERF_CTL_MASK);
> > > wrmsr_on_cpu(policy->cpu, MSR_IA32_PERF_CTL, lo, hi);
> >
> > Ditto.
> >
> > These two examples prove my point, actually.
>
> same here, it is just clear mask and set mask, why not?

Like this?

rmwmsrl_safe_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
INTEL_PERF_CTL_MASK,
(u32)sfi_cpufreq_array[next_perf_state].ctrl_val & INTEL_PERF_CTL_MASK);

Yikes!

So yes, it can work but it is ugly, hard to parse and use, not generic
enough, etc, etc.

So thanks, but no thanks.

--
Regards/Gruss,
Boris.

ECO tip #101: Trim your mails when you reply.