lkml.org 
[lkml]   [2018]   [Jan]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH V4] cpufreq: intel_pstate: allow trace in passive mode
Date
On Saturday, January 6, 2018 5:40:34 PM CET Doug Smythies wrote:
> On 2018.01.05 14:52 Rafael J. Wysocki wrote:
> > On Fri, Jan 5, 2018 at 11:14 PM, Doug Smythies <doug.smythies@gmail.com> wrote:
>
> >> Allow use of the trace_pstate_sample trace function
> >> when the intel_pstate driver is in passive mode.
> >> Since the core_busy and scaled_busy fields are not
> >> used, and it might be desirable to know which path
> >> through the driver was used, either intel_cpufreq_target
> >> or intel_cpufreq_fast_switch, re-task the core_busy
> >> field as a flag indicator.
> >>
> >> The user can then use the intel_pstate_tracer.py utility
> >> to summarize and plot the trace.
> >>
> >> Sometimes, in passive mode, the driver is not called for
> >> many tens or even hundreds of seconds. The user
> >> needs to understand, and not be confused by, this limitation.
> >
> > The description of the changes between different versions should go
> > under the Signed-off-by: tag, separated by an extra "---" from it.
>
> O.K. sorry.
>
> > Also please see a couple of cosmetic comments below.
> >
> >> V4: Only execute the trace specific overhead code if trace
> >> is enabled. Suggested by Srinivas Pandruvada.
> >>
> >> V3: Move largely duplicate code to a subroutine.
> >> Suggested by Rafael J. Wysocki.
> >>
> >> V2: prepare for resend. Rebase to current kernel, 4.15-rc3.
> >> Signed-off-by: Doug Smythies <dsmythies@telus.net>
> >> ---
> >> drivers/cpufreq/intel_pstate.c | 35 +++++++++++++++++++++++++++++++++--
> >> 1 file changed, 33 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
> >> index 93a0e88..53bb953 100644
> >> --- a/drivers/cpufreq/intel_pstate.c
> >> +++ b/drivers/cpufreq/intel_pstate.c
> >> @@ -1943,13 +1943,40 @@ static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
> >> return 0;
> >> }
> >>
> >> +static void intel_cpufreq_trace(struct cpudata *cpu, int fast, int from)
> >
> > Please use "bool" for "fast" and I'd call it "fast_switch".
>
> O.K. thanks.
>
> >> +{
> >> + struct sample *sample;
> >> + u64 time;
> >> +
> >> + time = ktime_get();
> >
> > It is pointless to evaluate ktime_get() if
> > trace_pstate_sample_enabled() returns "false".
>
> Of course, thanks.
>
> >> + if (trace_pstate_sample_enabled()) {
> >> + if (intel_pstate_sample(cpu, time)) {
> >
> > And the extra indentation here is not very useful, so I'd write it as
> >
> > if (!trace_pstate_sample_enabled())
> > return;
> >
> > if (!intel_pstate_sample(cpu, ktime_get()))
> > return;
> >
> > (note that you don't need the "time" variable any more with this).
>
> That is much better, Thanks.
>
> >> + sample = &cpu->sample;
> >> + /* In passvie mode the trace core_busy field is
> >
> > "passive" (typo)
> >
> >> + * re-assigned to indicate if the driver call
> >> + * was via the normal or fast switch path.
> >> + * The scaled_busy field is not used, set to 0.
> >> + */
> >> + trace_pstate_sample(fast,
> >> + 0,
> >> + from,
> >> + cpu->pstate.current_pstate,
> >> + sample->mperf,
> >> + sample->aperf,
> >> + sample->tsc,
> >> + get_avg_frequency(cpu),
> >> + fp_toint(cpu->iowait_boost * 100));
> >> + }
> >> + }
> >> +}
> >> +
> >> static int intel_cpufreq_target(struct cpufreq_policy *policy,
> >> unsigned int target_freq,
> >> unsigned int relation)
> >> {
> >> struct cpudata *cpu = all_cpu_data[policy->cpu];
> >> struct cpufreq_freqs freqs;
> >> - int target_pstate;
> >> + int target_pstate, from;
> >
> > I would call the new variable "old_pstate" or "orig_pstate" (so that
> > it is visibly clear that it represents a P-state).
>
> O.K.
> I used "from" because that is what Dirk called it in the trace buffer stuff.
>
> >>
> >> update_turbo_state();
> >>
> >> @@ -1969,12 +1996,14 @@ static int intel_cpufreq_target(struct cpufreq_policy *policy,
> >> break;
> >> }
> >> target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
> >> + from = cpu->pstate.current_pstate;
> >> if (target_pstate != cpu->pstate.current_pstate) {
> >> cpu->pstate.current_pstate = target_pstate;
> >> wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
> >> pstate_funcs.get_val(cpu, target_pstate));
> >> }
> >> freqs.new = target_pstate * cpu->pstate.scaling;
> >> + intel_cpufreq_trace(cpu, 0, from);
> >> cpufreq_freq_transition_end(policy, &freqs, false);
> >>
> >> return 0;
> >> @@ -1984,13 +2013,15 @@ static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
> >> unsigned int target_freq)
> >> {
> >> struct cpudata *cpu = all_cpu_data[policy->cpu];
> >> - int target_pstate;
> >> + int target_pstate, from;
> >>
> >> update_turbo_state();
> >>
> >> target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
> >> target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
> >> + from = cpu->pstate.current_pstate;
> >> intel_pstate_update_pstate(cpu, target_pstate);
> >> + intel_cpufreq_trace(cpu, 100, from);
> >
> > Why are you passing 100 here? Anything different from 0 should
> > suffice, 1 in particular. And I'd pass "false" or "true" (they will
> > be converted to 0 and 1 for output anyway).
>
> Well, I wanted to just re-use the existing graphs generated by
> tools/power/x86/intel_pstate_tracer/intel_pstate_tracer.py
> and so wanted to pass 0 or 100% to it. On purpose, those graphs
> do not autoscale on the y-axis.
>
> When investigating, the graphs can be used as a way to determine
> where to look in more detail at the raw csv files.

OK, but that would require a comment at least.

I would #ifdef symbols for that, like INTEL_PSTATE_TRACE_FAST_SWITCH
and INTEL_PSTATE_TRACE_TARGET or similar and define them as 100 and 0,
respectively.

I would call the corresponding function argument "trace_type" or
similar (it should be u32 too) and I would explain in a comment
why INTEL_PSTATE_TRACE_FAST_SWITCH is 100.

Thanks,
Rafael

\
 
 \ /
  Last update: 2018-01-14 23:16    [W:0.160 / U:0.412 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site