RE: [PATCH] cpufreq: intel_pstate: Fix cpuinfo_cur_freq after performance governor changes

[Huaisheng HS1 Ye]

Hi Srinivas,

Oh, I see. Originally I thought this function "arch_freq_get_on_cpu" would have chance to expand to other platforms in the future. Because I found that it appears at cpufreq.c as __weak. 
But if it is sure that this function only works for x86 all the time, I think it doesn't matter about its position within show_cpuinfo_cur_freq.

Thanks
Huaisheng

> 
> Hi Huaisheng,
> 
> On Tue, 2017-07-25 at 07:03 +0000, Huaisheng HS1 Ye wrote:
> > Hi Srinivas,
> > Your idea is great, but your patch at cpufreq.c will force all
> > platforms to use scaling_cur_freq as first choice when userspace wants
> > to access cpuinfo_cur_freq. It is ok for intel x86 platfrom but hard
> > to say with other platforms.
> arch_freq_get_on_cpu is only implemented on x86, for other platforms it will
> not change behavior. I didn't understand your comment about first choice.
> 
> Thanks,
> Srinivas
> 
> 
> > I modified it like that, it looks more reasonable. How about that?
> >
> > Hi Rafael,
> > Deleting "get" function pointer within intel_pstate would lead to
> > sysfs interface cpuinfo_cur_freq disappearing, because of
> > cpufreq_add_dev_interface will check "cpufreq_driver->get" for it.
> > Perhaps just return 0 with in intel_pstate_get would be a workaround
> > for this issue, how about it?
> >
> > I have tested this patch based on Purley platform, both Hardware and
> > Software P-states works correct, we could get accurate and same
> > frequency from cpuinfo_cur_freq and scaling_cur_freq.
> >
> > Âdrivers/cpufreq/cpufreq.cÂÂÂÂÂÂ| 4 ++++
> > Âdrivers/cpufreq/intel_pstate.c | 8 +++++---
> > Â2 files changed, 9 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
> > index 9bf97a3..922f9d9 100644
> > --- a/drivers/cpufreq/cpufreq.c
> > +++ b/drivers/cpufreq/cpufreq.c
> > @@ -694,6 +694,10 @@ static ssize_t show_cpuinfo_cur_freq(struct
> > cpufreq_policy *policy,
> > Â	if (cur_freq)
> > Â		return sprintf(buf, "%u\n", cur_freq);
> >
> > +	cur_freq = arch_freq_get_on_cpu(policy->cpu);
> > +	if (cur_freq)
> > +		return sprintf(buf, "%u\n", cur_freq);
> > +
> > Â	return sprintf(buf, "<unknown>\n");
> > Â}
> >
> > diff --git a/drivers/cpufreq/intel_pstate.c
> > b/drivers/cpufreq/intel_pstate.c index 6cd5035..33e6c10 100644
> > --- a/drivers/cpufreq/intel_pstate.c
> > +++ b/drivers/cpufreq/intel_pstate.c
> > @@ -1924,9 +1924,11 @@ static int intel_pstate_init_cpu(unsigned int
> > cpunum)
> >
> > Âstatic unsigned int intel_pstate_get(unsigned int cpu_num)
> > Â{
> > -	struct cpudata *cpu = all_cpu_data[cpu_num];
> > -
> > -	return cpu ? get_avg_frequency(cpu) : 0;
> > +	/*
> > +	Â* Use frequency from scaling_cur_freq, reserve this
> > function
> > +	Â* for existing of sysfs cpuinfo_cur_freq.
> > +	Â*/
> > +	return 0;
> > Â}
> >
> > Âstatic void intel_pstate_set_update_util_hook(unsigned int cpu_num)
> >
> >
> > >
> > > On Tue, 2017-07-25 at 01:46 +0000, Huaisheng HS1 Ye wrote:
> > > >
> > > > Hi Rafael,
> > > >
> > > > If you delete "get" function implement within intel_pstate, the
> > > > sysfs interface cpuinfo_cur_freq will display <unknown> all the
> > > > time.
> > > cpuinfo_cur_freq by definition should show actual frequency HW
> > > frequency.
> > > Unless I missed something. So Len Brown's patch should also take
> > > care of this to get from arch specific function is available.
> > > So in addition to Rafael's change, what about this?
> > >
> > >
> > > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
> > > index
> > > 9bf97a3..29ec687 100644
> > > --- a/drivers/cpufreq/cpufreq.c
> > > +++ b/drivers/cpufreq/cpufreq.c
> > > @@ -689,8 +689,13 @@ store_one(scaling_max_freq, max);
> > > Âstatic ssize_t show_cpuinfo_cur_freq(struct cpufreq_policy *policy,
> > > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂchar *buf)
> > > Â{
> > > -ÂÂÂÂÂÂÂunsigned int cur_freq = __cpufreq_get(policy);
> > > +ÂÂÂÂÂÂÂunsigned int cur_freq;
> > >
> > > +ÂÂÂÂÂÂÂcur_freq = arch_freq_get_on_cpu(policy->cpu);
> > > +ÂÂÂÂÂÂÂif (cur_freq)
> > > +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂreturn sprintf(buf, "%u\n", cur_freq);
> > > +
> > > +ÂÂÂÂÂÂÂcur_freq = __cpufreq_get(policy);
> > > ÂÂÂÂÂÂÂÂif (cur_freq)
> > > ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂreturn sprintf(buf, "%u\n", cur_freq);
> > >
> > >
> > >
> > > Thanks,
> > > Srinivas
> > >
> > > >
> > > > To be honest, at the beginning I have consider this way like you
> > > > patched, but based two reasons below, it is conservative for us to
> > > > do that.
> > > >
> > > > 1. I am worried about whether it would lead to confusion for
> > > > customers or Linux OS venders who are accustomed to
> > > > cpuinfo_cur_freq.
> > > > 2. This is the first time for me to offer patch to intel_pstate,
> > > > not sure whether it could be accepted by you.
> > > >
> > > > >
> > > > >
> > > > > On Monday, July 24, 2017 03:32:47 PM Huaisheng HS1 Ye wrote:
> > > > > >
> > > > > >
> > > > > > Hi Rafael,
> > > > > > Thanks for your reply.
> > > > > >
> > > > > > >
> > > > > > >
> > > > > > > On Monday, July 24, 2017 05:43:14 AM Huaisheng HS1 Ye
> > > > > > > wrote:
> > > > > > > >
> > > > > > > >
> > > > > > > > After commit 82b4e03e01bc (intel_pstate: skip scheduler
> > > > > > > > hook when in "performance" mode) Software P-state control
> > > > > > > > modes couldn't get dynamic value during performance mode,
> > > > > > > Please explain what you mean here.
> > > > > > >
> > > > > > commit 82b4e03e01bc (intel_pstate: skip scheduler hook when in
> > > > > > "performance" mode) disables intel_pstate_set_update_util_hook
> > > > > > when current policy is performance within function
> > > > > > intel_pstate_set_policy.
> > > > > > It leads to Software P-states couldn't update sysfs interface
> > > > > > cpuinfo_cur_freq's value during performance mode, because of
> > > > > > pstate_funcs.update_util couldn't set for the given CPU.
> > > > > >
> > > > > > >
> > > > > > >
> > > > > > > I guess you carried out some tests and the results were not
> > > > > > > as expected, so what was the test?
> > > > > > Exactly, we check the sysfs interface cpuinfo_cur_freq and the
> > > > > > output of cpupower frequency-info both with performance mode.
> > > > > OK, so what about the change below:
> > > > >
> > > > > ---
> > > > > Âdrivers/cpufreq/intel_pstate.c |ÂÂÂÂ8 --------
> > > > > Â1 file changed, 8 deletions(-)
> > > > >
> > > > > Index: linux-pm/drivers/cpufreq/intel_pstate.c
> > > > >
> > >
> ==============================================================
> > > >
> > > > >
> > > > > =====
> > > > > --- linux-pm.orig/drivers/cpufreq/intel_pstate.c
> > > > > +++ linux-pm/drivers/cpufreq/intel_pstate.c
> > > > > @@ -1674,13 +1674,6 @@ static int intel_pstate_init_cpu(unsigne
> > > > > Â	return 0;
> > > > > Â}
> > > > >
> > > > > -static unsigned int intel_pstate_get(unsigned int cpu_num) -{
> > > > > -	struct cpudata *cpu = all_cpu_data[cpu_num];
> > > > > -
> > > > > -	return cpu ? get_avg_frequency(cpu) : 0;
> > > > > -}
> > > > > -
> > > > > Âstatic void intel_pstate_set_update_util_hook(unsigned int
> > > > > cpu_num)ÂÂ{
> > > > > Â	struct cpudata *cpu = all_cpu_data[cpu_num]; @@
> > > > > -1921,7
> > > > > +1914,6 @@
> > > > > static struct cpufreq_driver intel_pstat
> > > > > Â	.setpolicy	= intel_pstate_set_policy,
> > > > > Â	.suspend	= intel_pstate_hwp_save_state,
> > > > > Â	.resume		= intel_pstate_resume,
> > > > > -	.get		= intel_pstate_get,
> > > > > Â	.init		= intel_pstate_cpu_init,
> > > > > Â	.exit		= intel_pstate_cpu_exit,
> > > > > Â	.stop_cpu	= intel_pstate_stop_cpu,