[Intel-gfx] [PATCH i-g-t] i915/gem_exec_balancer: Kick tasklets harder

[Chris Wilson]

Tasklet scheduling can unfortunately be quite predicated. Where we are
only caring about workload distribution and not initial execution
latency (see kick_kthreads), make that kick more precise.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
---
 lib/igt_debugfs.h              |  7 +++++++
 tests/i915/gem_exec_balancer.c | 28 +++++++++++++++++-----------
 2 files changed, 24 insertions(+), 11 deletions(-)

diff --git a/lib/igt_debugfs.h b/lib/igt_debugfs.h
index 7d1a6175e..768341387 100644
--- a/lib/igt_debugfs.h
+++ b/lib/igt_debugfs.h
@@ -181,6 +181,13 @@ void igt_require_hpd_storm_ctl(int fd);
  * which will flush any RCU callbacks and deferred tasks.
  */
 #define DROP_RCU 0x200
+/**
+ * DROP_TASKLETS:
+ *
+ * Flushes the softirq tasklets on each CPU, such as used by the execution
+ * threads.
+ */
+#define DROP_TASKLETS 0x400
 /**
  * DROP_ALL:
  *
diff --git a/tests/i915/gem_exec_balancer.c b/tests/i915/gem_exec_balancer.c
index e9a893e90..9ec5a146e 100644
--- a/tests/i915/gem_exec_balancer.c
+++ b/tests/i915/gem_exec_balancer.c
@@ -24,6 +24,7 @@
 #include <sched.h>
 
 #include "igt.h"
+#include "igt_debugfs.h"
 #include "igt_gt.h"
 #include "igt_perf.h"
 #include "i915/gem_ring.h"
@@ -405,17 +406,20 @@ static void invalid_bonds(int i915)
 	gem_context_destroy(i915, p.ctx_id);
 }
 
-static void kick_kthreads(void)
+static void kick_kthreads(int i915)
 {
-	usleep(20 * 1000); /* 20ms should be enough for ksoftirqd! */
+	if (igt_drop_caches_has(i915, DROP_TASKLETS))
+		igt_drop_caches_set(i915, DROP_TASKLETS);
+	else
+		usleep(20 * 1000); /* 20ms should be enough for ksoftirqd! */
 }
 
-static double measure_load(int pmu, int period_us)
+static double measure_load(int i915, int pmu, int period_us)
 {
 	uint64_t data[2];
 	uint64_t d_t, d_v;
 
-	kick_kthreads();
+	kick_kthreads(i915);
 
 	igt_assert_eq(read(pmu, data, sizeof(data)), sizeof(data));
 	d_v = -data[0];
@@ -430,13 +434,14 @@ static double measure_load(int pmu, int period_us)
 	return d_v / (double)d_t;
 }
 
-static double measure_min_load(int pmu, unsigned int num, int period_us)
+static double
+measure_min_load(int i915, int pmu, unsigned int num, int period_us)
 {
 	uint64_t data[2 + num];
 	uint64_t d_t, d_v[num];
 	uint64_t min = -1, max = 0;
 
-	kick_kthreads();
+	kick_kthreads(i915);
 
 	igt_assert_eq(read(pmu, data, sizeof(data)), sizeof(data));
 	for (unsigned int n = 0; n < num; n++)
@@ -464,12 +469,13 @@ static double measure_min_load(int pmu, unsigned int num, int period_us)
 	return min / (double)d_t;
 }
 
-static void measure_all_load(int pmu, double *v, unsigned int num, int period_us)
+static void
+measure_all_load(int i915, int pmu, double *v, unsigned int num, int period_us)
 {
 	uint64_t data[2 + num];
 	uint64_t d_t, d_v[num];
 
-	kick_kthreads();
+	kick_kthreads(i915);
 
 	igt_assert_eq(read(pmu, data, sizeof(data)), sizeof(data));
 	for (unsigned int n = 0; n < num; n++)
@@ -527,7 +533,7 @@ static void check_individual_engine(int i915,
 						  ci[idx].engine_instance));
 
 	spin = igt_spin_new(i915, .ctx = ctx, .engine = idx + 1);
-	load = measure_load(pmu, 10000);
+	load = measure_load(i915, pmu, 10000);
 	igt_spin_free(i915, spin);
 
 	close(pmu);
@@ -676,7 +682,7 @@ static void bonded(int i915, unsigned int flags)
 				igt_spin_free(i915, plug);
 			}
 
-			measure_all_load(pmu[0], v, limit + 1, 10000);
+			measure_all_load(i915, pmu[0], v, limit + 1, 10000);
 			igt_spin_free(i915, spin);
 
 			igt_assert_f(v[bond] > 0.90,
@@ -1512,7 +1518,7 @@ static void full(int i915, unsigned int flags)
 			close(fence);
 		}
 
-		load = measure_min_load(pmu[0], count, 10000);
+		load = measure_min_load(i915, pmu[0], count, 10000);
 		igt_spin_free(i915, spin);
 
 		close(pmu[0]);
-- 
2.26.0