[Intel-gfx] [PATCH] drm/i915/pmu: Inspect runtime PM state more carefully while estimating RC6

[Tvrtko Ursulin]

On 10/04/2018 10:57, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2018-04-10 10:23:28)
>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>
>> While thinking about sporadic failures of perf_pmu/rc6-runtime-pm* tests
>> on some CI machines I have concluded that: a) the PMU readout of RC6 can
>> race against runtime PM transitions, and b) there are other reasons than
>> being runtime suspended which can cause intel_runtime_pm_get_if_in_use to
>> fail.
>>
>> Therefore when estimating RC6 the code needs to assert we are indeed in
>> suspended state and if not the best we can do is return the last known RC6
>> value.
>>
>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> Fixes: 1fe699e30113 ("drm/i915/pmu: Fix sleep under atomic in RC6 readout")
>> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105010
>> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> Cc: Chris Wilson <chris at chris-wilson.co.uk>
>> Cc: Imre Deak <imre.deak at intel.com>
>> ---
>> I was able to trigger state != RPM_SUSPENDED on the shards, but not yet
>> the actual estimation overaccounting. As such this fix is based partially
>> on speculation that it will fix the sporadic perf_pmu/rc6* failures.
>> Nevertheless I think it is correct to add this check regardless.
>> ---
>>   drivers/gpu/drm/i915/i915_pmu.c | 24 ++++++++++++++++++++++++
>>   1 file changed, 24 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/i915/i915_pmu.c b/drivers/gpu/drm/i915/i915_pmu.c
>> index bd7e695fc663..e92a9571db77 100644
>> --- a/drivers/gpu/drm/i915/i915_pmu.c
>> +++ b/drivers/gpu/drm/i915/i915_pmu.c
>> @@ -473,6 +473,30 @@ static u64 get_rc6(struct drm_i915_private *i915)
>>                  spin_lock_irqsave(&i915->pmu.lock, flags);
>>                  spin_lock(&kdev->power.lock);
>>   
>> +               /*
>> +                * After the above branch intel_runtime_pm_get_if_in_use failed
>> +                * to get the runtime PM reference we cannot assume we are in
>> +                * runtime suspend since we can either: a) race with coming out
>> +                * of it before we took the power.lock, or b) there are other
>> +                * states than suspended which can bring us here.
>> +                *
>> +                * We need to double-check that we are indeed currently runtime
>> +                * suspended and if not we cannot do better than report the last
>> +                * known RC6 value.
>> +                */
>> +               if (kdev->power.runtime_status != RPM_SUSPENDED) {
>> +                       spin_unlock(&kdev->power.lock);
>> +
>> +                       if (i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur)
>> +                               val = i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur;
>> +                       else
>> +                               val = i915->pmu.sample[__I915_SAMPLE_RC6].cur;
> 
> If rpm awake, but having lost the race to read the regs, report the last
> known value.
> 
> This is because we don't know if another thread is in the other branch,
> and so we will have one updating the estimate while it being compared
> against.

No, the race is intel_runtime_pm_get_if_in_use telling us the device is 
not active, but a) that doesn't mean it is suspended, and b) it doesn't 
mean it is still suspended after the check.

Also PMU internal state is serialized by the spinlock so there is no 
inconsistency there.
> But I'm not understanding the failure -- why is the estimate bad? At the
> very least we still ensure that it is monotonic? Is it just the jitter
> you are worrying about? (If the estimate is bad here, isn't it always
> bad?)

As far as I have seen failures from CI are all estimate being too large. 
(no jitter and no going backwards)

What I suspect is going bad in either case, is that we must not add the 
delta from current jiffies to internal runtime pm counters if state is 
not suspended. If we do that we are accounting an unknown period of time 
as suspended time and that would explain the over-estimation.

In other words we are only allowed to estimate if the current state is 
definitely suspended. If it is anything else we need to report either 
the last estimated value, or the last real value, depending what is more 
recent.

I've done a CI run which definitely shows we can end up in this path 
when state is not suspended.

>> +
>> +                       spin_unlock_irqrestore(&i915->pmu.lock, flags);
>> +
>> +                       return val;
>> +               }
> 
> I'd prefer moving the RPM_SUSPENDED code into an else branch to avoid
> another unlock/early return here. (It just fits into 80cols, so no
> excuses ;)

Okay makes sense.

Regards,

Tvrtko