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

[Chris Wilson]

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().
-Chris