[Intel-gfx] [CI i-g-t 1/2] tests/gem_exec_latency: New subtests for checking submission from RT tasks

[Chris Wilson]

Quoting Tvrtko Ursulin (2018-04-17 09:41:11)
> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> 
> We want to make sure RT tasks which use a lot of CPU times can submit
> batch buffers with roughly the same latency (and certainly not worse)
> compared to normal tasks.
> 
> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> Cc: Chris Wilson <chris at chris-wilson.co.uk>
> ---
> +/*
> + * Test whether RT thread which hogs the CPU a lot can submit work with
> + * reasonable latency.
> + */
> +static void
> +rthog_latency_on_ring(int fd, unsigned int ring, const char *name)
> +{
> +       const char *passname[] = { "warmup", "normal", "rt" };
> +       struct rt_pkt res[3];
> +       unsigned int i;
> +       int link[2];
> +       int ret;
> +
> +       igt_require(gem_can_store_dword(fd, ring));
> +
> +       igt_assert(pipe(link) == 0);
> +
> +       memset(res, 0, sizeof(res));
> +
> +        igt_fork(child, 1) {
> +               unsigned int pass = 0; /* Three passes: warmup, normal, rt. */
> +
> +               do {
> +                       struct rt_pkt msg = { };
> +                       igt_spin_t *spin;
> +
> +                       igt_mean_init(&msg.mean);
> +
> +                       if (pass == 2) {
> +                               struct sched_param rt =
> +                                               { .sched_priority = 99 };
> +
> +                               ret = sched_setscheduler(0,
> +                                                       SCHED_FIFO | SCHED_RESET_ON_FORK,
> +                                                       &rt);
> +                               if (ret) {
> +                                       igt_warn("Failed to set scheduling policy!\n");
> +                                       msg.status = RT_FAIL;
> +                                       write(link[1], &msg, sizeof(msg));
> +                                       exit(1);
> +                               }
> +                       }
> +
> +                       spin = __igt_spin_batch_new_poll(fd, 0, ring);
> +                       if (!spin) {
> +                               igt_warn("Failed to create spinner! (%s)\n",
> +                                        passname[pass]);
> +                               msg.status = RT_FAIL;
> +                               write(link[1], &msg, sizeof(msg));
> +                               exit(1);
> +                       }
> +                       igt_spin_busywait_until_running(spin);
> +
> +                       igt_until_timeout(pass > 0 ? 5 : 2) {
> +                               double t;
> +
> +                               igt_spin_batch_end(spin);
> +                               gem_sync(fd, spin->handle);
> +
> +                               __rearm_spin_batch(spin);
> +                               __submit_spin_batch(fd, spin, ring);
> +
> +                               __spin_wait(&msg, spin, &t);
> +                               if (msg.status != RT_OK) {
> +                                       igt_warn("Wait timeout! (%s)\n",
> +                                                passname[pass]);
> +                                       write(link[1], &msg, sizeof(msg));
> +                                       exit(1);
> +                               }
> +
> +                               if (t > msg.max)
> +                                       msg.max = t;
> +
> +                               igt_mean_add(&msg.mean, t);
> +                       }
> +
> +                       igt_spin_batch_free(fd, spin);
> +
> +                       igt_info("%10s: mean=%.2fus variance=%.2fus max=%.2fus (n=%lu)\n",
> +                                passname[pass],
> +                                igt_mean_get(&msg.mean) * 1e6,
> +                                igt_mean_get_variance(&msg.mean) * 1e6,
> +                                msg.max * 1e6,
> +                                msg.mean.count);
> +
> +                       write(link[1], &msg, sizeof(msg));
> +               } while (++pass < 3);
> +
> +               exit(0);
> +       }
> +
> +       for (i = 0; i < 3; i++) {
> +               ret = read(link[0], &res[i], sizeof(res[0]));
> +               igt_assert_eq(ret, sizeof(res[0]));
> +
> +               igt_assert_eq(res[i].status, RT_OK);
> +       }
> +
> +       close(link[0]);
> +       close(link[1]);
> +
> +       igt_waitchildren();
> +
> +       /*
> +        * Check that the submission latency variance for a task with RT
> +        * priority is no larger than three times the same of a normal task.
> +        */
> +       igt_assert(igt_mean_get_variance(&res[2].mean) <
> +                  igt_mean_get_variance(&res[1].mean) * 3);
> +}

Fwiw, I think we should go full on ministat here, use igt_stats_t to
record every result then print the overlapping histograms so we can see
the distribution.

I was pondering a simple metric to warn about long tails. I was
wondering if the sum of values above the median would be interesting,
but really it boils down to wanting 99% to be within a smidgen of each
other.
-Chris