All of lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH] tests/perf_pmu: Fix busy-double-start for GuC backend
@ 2026-10-05 23:33 Umesh Nerlige Ramappa
  2026-10-06  0:28 ` ✓ Xe.CI.BAT: success for tests/perf_pmu: Fix busy-double-start for GuC backend (rev2) Patchwork
                   ` (3 more replies)
  0 siblings, 4 replies; 5+ messages in thread
From: Umesh Nerlige Ramappa @ 2026-10-05 23:33 UTC (permalink / raw)
  To: igt-dev

On GuC-based platforms, backend could switch work at a very fast rate
determined by "timeslice_duration_ms". For a default value of 1 ms, the
switching latencies could add up really quickly over the test duration.

While this works fine, workaround and other factors that may kick in for
some platforms at context switches add to the context switch latencies.

Choose a reasonable timeslice for this test so that expected and actual
busyness is within acceptable threshold. A 10ms timeslice works well for
this specific use case.

Closes: https://gitlab.freedesktop.org/drm/i915/kernel/-/work_items/4349
Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
---
 tests/intel/perf_pmu.c | 44 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 44 insertions(+)

diff --git a/tests/intel/perf_pmu.c b/tests/intel/perf_pmu.c
index 3a7f3410961f..b20f9bba1f9a 100644
--- a/tests/intel/perf_pmu.c
+++ b/tests/intel/perf_pmu.c
@@ -405,6 +405,33 @@ busy_start(int gem_fd, const intel_ctx_t *ctx,
 	gem_quiescent_gpu(gem_fd);
 }
 
+static void set_timeslice(int cs, unsigned int value)
+{
+	unsigned int delay;
+
+	igt_debug("setting timeslice to %u ms\n", value);
+	igt_assert_lte(0, igt_sysfs_printf(cs, "timeslice_duration_ms", "%u", value));
+	igt_sysfs_scanf(cs, "timeslice_duration_ms", "%u", &delay);
+	igt_assert_eq(delay, value);
+}
+
+static int timeslice_fd(int fd, const struct intel_execution_engine2 *e)
+{
+	int sys, cs = -1;
+	char buf[32];
+
+	sys = igt_sysfs_open(fd);
+	igt_require(sys != -1);
+
+	snprintf(buf, sizeof(buf), "engine/%s", e->name);
+	cs = openat(sys, buf, O_RDONLY);
+	igt_require(cs != -1);
+
+	close(sys);
+
+	return cs;
+}
+
 /*
  * This test has a potentially low rate of catching the issue it is trying to
  * catch. Or in other words, quite high rate of false negative successes. We
@@ -420,6 +447,18 @@ busy_double_start(int gem_fd, const intel_ctx_t *ctx,
 	const intel_ctx_t *tmp_ctx;
 	int fd;
 	uint64_t ahnd = get_reloc_ahnd(gem_fd, ctx->id), ahndN;
+	unsigned int saved, cs;
+
+	/*
+	 * Set a larger timeslice to minimize the engine busyness differences
+	 * due to context switch times
+	 */
+	if (gem_using_guc_submission(gem_fd)) {
+		cs = timeslice_fd(gem_fd, e);
+		igt_assert(igt_sysfs_scanf(cs, "timeslice_duration_ms", "%u", &saved) == 1);
+		igt_debug("initial timeslice %u ms\n", saved);
+		set_timeslice(cs, 10);
+	}
 
 	tmp_ctx = intel_ctx_create(gem_fd, &ctx->cfg);
 	ahndN = get_reloc_ahnd(gem_fd, tmp_ctx->id);
@@ -474,6 +513,11 @@ busy_double_start(int gem_fd, const intel_ctx_t *ctx,
 	put_ahnd(ahnd);
 	put_ahnd(ahndN);
 
+	if (gem_using_guc_submission(gem_fd)) {
+		set_timeslice(cs, saved);
+		close(cs);
+	}
+
 	assert_within_epsilon(val, ts[1] - ts[0], tolerance);
 	igt_assert_eq(val2, 0);
 
-- 
2.55.0


^ permalink raw reply related	[flat|nested] 5+ messages in thread
end of thread, other threads:[~2026-10-06  9:43 UTC | newest]

Thread overview: 5+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-10-05 23:33 [PATCH] tests/perf_pmu: Fix busy-double-start for GuC backend Umesh Nerlige Ramappa
2026-10-06  0:28 ` ✓ Xe.CI.BAT: success for tests/perf_pmu: Fix busy-double-start for GuC backend (rev2) Patchwork
2026-10-06  0:39 ` ✓ i915.CI.BAT: " Patchwork
2026-10-06  8:12 ` ✗ i915.CI.Full: failure " Patchwork
2026-10-06  9:42 ` ✗ Xe.CI.FULL: " Patchwork

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.