[igt-dev] [PATCH i-g-t 5/7] tests/i915/perf_pmu: Quiesce GPU if measuring idle busyness without spinner

[Umesh Nerlige Ramappa]

On Thu, Mar 30, 2023 at 10:00:28AM +0100, Tvrtko Ursulin wrote:
>
>On 30/03/2023 01:36, Umesh Nerlige Ramappa wrote:
>>The assumption in some tests is that the engines are not busy if no
>>spinners are being run. This is not true in some cases where we see
>>that the render is busy at the start of the test. Quiesce GPU to wait
>>for such work to complete before checking for idle busyness.
>
>Interesting.. hit in CI?

Was failing almost always on MTL in local runs. I think I saw one bug 
report as well.

>
>>Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
>>---
>>  tests/i915/perf_pmu.c | 12 ++++++++----
>>  1 file changed, 8 insertions(+), 4 deletions(-)
>>
>>diff --git a/tests/i915/perf_pmu.c b/tests/i915/perf_pmu.c
>>index 1bea3e57..5185cd9a 100644
>>--- a/tests/i915/perf_pmu.c
>>+++ b/tests/i915/perf_pmu.c
>>@@ -281,10 +281,12 @@ single(int gem_fd, const intel_ctx_t *ctx,
>>  	fd = open_pmu(gem_fd, I915_PMU_ENGINE_BUSY(e->class, e->instance));
>>-	if (flags & TEST_BUSY)
>>+	if (flags & TEST_BUSY) {
>>  		spin = spin_sync(gem_fd, ahnd, ctx, e);
>>-	else
>>+	} else {
>>  		spin = NULL;
>>+		gem_quiescent_gpu(gem_fd);
>>+	}
>
>But I'd probably just do it unconditionally at the start of tests, 
>given that if we are testing 100% busy we want it to be from our own 
>spinner, and not some potential leftover, which will then context 
>switch to our spinner at a random point.

Sure, will do.

Thanks,
Umesh
>
>Regards,
>
>Tvrtko
>
>>  	val = pmu_read_single(fd);
>>  	slept = measured_usleep(batch_duration_ns / 1000);
>>@@ -644,10 +646,12 @@ no_sema(int gem_fd, const intel_ctx_t *ctx,
>>  	fd[1] = open_group(gem_fd, I915_PMU_ENGINE_WAIT(e->class, e->instance),
>>  			   fd[0]);
>>-	if (flags & TEST_BUSY)
>>+	if (flags & TEST_BUSY) {
>>  		spin = spin_sync(gem_fd, ahnd, ctx, e);
>>-	else
>>+	} else {
>>  		spin = NULL;
>>+		gem_quiescent_gpu(gem_fd);
>>+	}
>>  	pmu_read_multi(fd[0], 2, val[0]);
>>  	measured_usleep(batch_duration_ns / 1000);