[PATCH] drm/i915/pmu: Match frequencies reported by PMU and sysfs
Tvrtko Ursulin
tvrtko.ursulin at linux.intel.com
Tue Oct 4 13:00:22 UTC 2022
On 04/10/2022 10:29, Tvrtko Ursulin wrote:
>
> On 03/10/2022 20:24, Ashutosh Dixit wrote:
>> PMU and sysfs use different wakeref's to "interpret" zero freq. Sysfs
>> uses
>> runtime PM wakeref (see intel_rps_read_punit_req and
>> intel_rps_read_actual_frequency). PMU uses the GT parked/unparked
>> wakeref. In general the GT wakeref is held for less time that the runtime
>> PM wakeref which causes PMU to report a lower average freq than the
>> average
>> freq obtained from sampling sysfs.
>>
>> To resolve this, use the same freq functions (and wakeref's) in PMU as
>> those used in sysfs.
>>
>> Bug: https://gitlab.freedesktop.org/drm/intel/-/issues/7025
>> Reported-by: Ashwin Kumar Kulkarni <ashwin.kumar.kulkarni at intel.com>
>> Cc: Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com>
>> Signed-off-by: Ashutosh Dixit <ashutosh.dixit at intel.com>
>> ---
>> drivers/gpu/drm/i915/i915_pmu.c | 27 ++-------------------------
>> 1 file changed, 2 insertions(+), 25 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/i915/i915_pmu.c
>> b/drivers/gpu/drm/i915/i915_pmu.c
>> index 958b37123bf1..eda03f264792 100644
>> --- a/drivers/gpu/drm/i915/i915_pmu.c
>> +++ b/drivers/gpu/drm/i915/i915_pmu.c
>> @@ -371,37 +371,16 @@ static void
>> frequency_sample(struct intel_gt *gt, unsigned int period_ns)
>> {
>> struct drm_i915_private *i915 = gt->i915;
>> - struct intel_uncore *uncore = gt->uncore;
>> struct i915_pmu *pmu = &i915->pmu;
>> struct intel_rps *rps = >->rps;
>> if (!frequency_sampling_enabled(pmu))
>> return;
>> - /* Report 0/0 (actual/requested) frequency while parked. */
>> - if (!intel_gt_pm_get_if_awake(gt))
>> - return;
>> -
>> if (pmu->enable & config_mask(I915_PMU_ACTUAL_FREQUENCY)) {
>> - u32 val;
>> -
>> - /*
>> - * We take a quick peek here without using forcewake
>> - * so that we don't perturb the system under observation
>> - * (forcewake => !rc6 => increased power use). We expect
>> - * that if the read fails because it is outside of the
>> - * mmio power well, then it will return 0 -- in which
>> - * case we assume the system is running at the intended
>> - * frequency. Fortunately, the read should rarely fail!
>> - */
>> - val = intel_uncore_read_fw(uncore, GEN6_RPSTAT1);
>> - if (val)
>> - val = intel_rps_get_cagf(rps, val);
>> - else
>> - val = rps->cur_freq;
>> -
>> add_sample_mult(&pmu->sample[__I915_SAMPLE_FREQ_ACT],
>> - intel_gpu_freq(rps, val), period_ns / 1000);
>> + intel_rps_read_actual_frequency(rps),
>> + period_ns / 1000);
>> }
>> if (pmu->enable & config_mask(I915_PMU_REQUESTED_FREQUENCY)) {
>
> What is software tracking of requested frequency showing when GT is
> parked or runtime suspended? With this change sampling would be outside
> any such checks so we need to be sure reported value makes sense.
>
> Although more important open is around what is actually correct.
>
> For instance how does the patch affect RC6 and power? I don't know how
> power management of different blocks is wired up, so personally I would
> only be able to look at it empirically. In other words what I am asking
> is this - if we changed from skipping obtaining forcewake even when
> unparked, to obtaining forcewake if not runtime suspended - what
> hardware blocks does that power up and how it affects RC6 and power? Can
> it affect actual frequency or not? (Will "something" power up the clocks
> just because we will be getting forcewake?)
>
> Or maybe question simplified - does 200Hz polling on existing sysfs
> actual frequency field disturbs the system under some circumstances?
> (Increases power and decreases RC6.) If it does then that would be a
> problem. We want a solution which shows the real data, but where the act
> of monitoring itself does not change it too much. If it doesn't then
> it's okay.
>
> Could you somehow investigate on these topics? Maybe log RAPL GPU power
> while polling on sysfs, versus getting the actual frequency from the
> existing PMU implementation and see if that shows anything? Or actually
> simpler - RAPL GPU power for current PMU intel_gpu_top versus this
> patch? On idle(-ish) desktop workloads perhaps? Power and frequency
> graphed for both.
Another thought - considering that bspec says for 0xa01c "This register
reflects real-time values and thus does not have a pre-determined
default value out of reset" - could it be that it also does not reflect
a real value when GPU is not executing anything (so zero), just happens
to be not runtime suspended? That would mean sysfs reads could maybe
show last known value? Just a thought to check.
I've also tried on my Alderlake desktop:
1)
while true; do cat gt_act_freq_mhz >/dev/null; sleep 0.005; done
This costs ~120mW of GPU power and ~20% decrease in RC6.
2)
intel_gpu_top -l -s 5 >/dev/null
This costs no power or RC6.
I have also never observed sysfs to show below min freq. This was with
no desktop so it's possible this register indeed does not reflect the
real situation when things are idle.
So I think it is possible sysfs value is the misleading one.
Regards,
Tvrtko
More information about the dri-devel
mailing list