[igt-dev] [Intel-gfx] [PATCH i-g-t] tests/perf_pmu: Test busyness reporting in face of GPU hangs

[Tvrtko Ursulin]

On 19/02/2018 19:21, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2018-02-19 19:12:51)
>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>
>> Verify that the reported busyness is in line with what would we expect
>> from a batch which causes a hang and gets kicked out from the engine.
>>
>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> ---
>>   tests/perf_pmu.c | 41 +++++++++++++++++++++++++++++++++++------
>>   1 file changed, 35 insertions(+), 6 deletions(-)
>>
>> diff --git a/tests/perf_pmu.c b/tests/perf_pmu.c
>> index 7fab73e22c2d..90b6ec4db32d 100644
>> --- a/tests/perf_pmu.c
>> +++ b/tests/perf_pmu.c
>> @@ -168,6 +168,7 @@ static unsigned int e2ring(int gem_fd, const struct intel_execution_engine2 *e)
>>   #define TEST_TRAILING_IDLE (4)
>>   #define TEST_RUNTIME_PM (8)
>>   #define FLAG_LONG (16)
>> +#define FLAG_HANG (32)
>>   
>>   static void end_spin(int fd, igt_spin_t *spin, unsigned int flags)
>>   {
>> @@ -186,11 +187,15 @@ static void end_spin(int fd, igt_spin_t *spin, unsigned int flags)
>>   static void
>>   single(int gem_fd, const struct intel_execution_engine2 *e, unsigned int flags)
>>   {
>> +       const unsigned int hang_us = 10e6;
>>          unsigned long slept;
>>          igt_spin_t *spin;
>> -       uint64_t val;
>> +       uint64_t val[2], ts[2];
>>          int fd;
>>   
>> +       if (flags & FLAG_HANG)
>> +               gem_quiescent_gpu(gem_fd);
>> +
>>          fd = open_pmu(I915_PMU_ENGINE_BUSY(e->class, e->instance));
>>   
>>          if (flags & TEST_BUSY)
>> @@ -198,17 +203,36 @@ single(int gem_fd, const struct intel_execution_engine2 *e, unsigned int flags)
>>          else
>>                  spin = NULL;
>>   
>> -       val = pmu_read_single(fd);
>> -       slept = measured_usleep(batch_duration_ns / 1000);
>> +       val[0] = __pmu_read_single(fd, &ts[0]);
>> +       slept = measured_usleep(flags & FLAG_HANG ?
>> +                               hang_us : batch_duration_ns / 1000);
>>          if (flags & TEST_TRAILING_IDLE)
>>                  end_spin(gem_fd, spin, flags);
>> -       val = pmu_read_single(fd) - val;
>> +       val[1] = pmu_read_single(fd);
>>   
>>          end_spin(gem_fd, spin, FLAG_SYNC);
>>          igt_spin_batch_free(gem_fd, spin);
>> -       close(fd);
>>   
>> -       assert_within_epsilon(val, flags & TEST_BUSY ? slept : 0.f, tolerance);
>> +       if ((flags & TEST_BUSY) && (flags & FLAG_HANG)) {
>> +               val[1] = __pmu_read_single(fd, &ts[1]);
>> +               close(fd);
>> +               igt_info("sampled with hang %.3fms / %.3fms\n",
>> +                        (val[1] - val[0]) / 1e6, (ts[1] - ts[0]) / 1e6);
>> +               /* Check that some busyness was reported. */
>> +               igt_assert(val[1] - val[0] > 0);
>> +               /*
>> +                * But not more than some reasonable value before which we
>> +                * expected the spinner to be kicked out.
>> +                */
> 
> So 120s? And even that carries internal knowledge from across the ages.
> 
> I don't think this is a sensible test. What would be reasonable is
> something like
> 
> 	spinner()
> 	val[0] = pmu()
> 	sleep()
> 	igt_force_gpu_reset()
> 	val[1] = pmu();
> 	d_busy = val[1] - val[0]
> 	sleep()
> 	val[2] = pmu()
> 	d_idle = val[2] - val[1];
> 
> Then d_busy should be d_ts, and d_idle should be 0. i.e. the
> igt_force_gpu_reset() is just an indirect igt_spin_batch_end().

Yeah I am not claiming the test is great. I threw it together quickly 
when I suspected something is going bad. Just want to get some results 
overnight so I can despair tomorrow.

Regards,

Tvrtko