On 28/05/2020 21:31, Chris Wilson wrote:
Randomly submit a paired spinner and its cancellation as a bonded
(submit fence) pair. Apply congestion to the engine with more bonded
pairs to see if the execution order fails. If we prevent a cancellation
from running, then the spinner will remain spinning forever.
v2: Test both immediate submission and fenced submission
Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>
---
tests/i915/gem_exec_balancer.c | 172 +++++++++++++++++++++++++++++++++
1 file changed, 172 insertions(+)
diff --git a/tests/i915/gem_exec_balancer.c b/tests/i915/gem_exec_balancer.c
index 80ae82416..04b14dd3a 100644
--- a/tests/i915/gem_exec_balancer.c
+++ b/tests/i915/gem_exec_balancer.c
@@ -1154,6 +1154,175 @@ static void bonded_semaphore(int i915)
gem_context_destroy(i915, ctx);
}
+static void __bonded_dual(int i915,
+ const struct i915_engine_class_instance *siblings,
+ unsigned int count,
+ unsigned int flags,
+ unsigned long *out)
+#define BD_FENCE 0x1
+#define BD_HOSTILE 0x2
+#define BD_MANY 0x4
+{
+ struct drm_i915_gem_exec_object2 batch = {};
+ struct drm_i915_gem_execbuffer2 execbuf = {
+ .buffers_ptr = to_user_pointer(&batch),
+ .buffer_count = 1,
+ };
+ unsigned long cycles = 0;
+ unsigned int spinner;
+ igt_spin_t *a, *b;
+ int timeline;
+ uint32_t A, B;
+
+ srandom(getpid());
+
+ spinner = IGT_SPIN_POLL_RUN;
+ if (flags & BD_HOSTILE)
+ spinner |= IGT_SPIN_NO_PREEMPTION;
+
+ A = gem_context_create(i915);
+ set_load_balancer(i915, A, siblings, count, NULL);
+ a = igt_spin_new(i915, A, .flags = spinner);
+ igt_spin_end(a);
+ gem_sync(i915, a->handle);
+
+ B = gem_context_create(i915);
+ set_load_balancer(i915, B, siblings, count, NULL);
+ b = igt_spin_new(i915, B, .flags = spinner);
+ igt_spin_end(b);
+ gem_sync(i915, b->handle);
+
+ timeline = sw_sync_timeline_create();
+
+ igt_until_timeout(2) {
+ unsigned int master;
+ int fence;
+
+ master = 1;
+ if (flags & BD_MANY)
+ master = rand() % count + 1;
+
+ fence = -1;
+ if (flags & BD_FENCE)
+ fence = sw_sync_timeline_create_fence(timeline,
+ cycles + 1);
+
+ igt_spin_reset(a);
+ a->execbuf.flags = master | I915_EXEC_FENCE_OUT;
+ if (fence != -1) {
+ a->execbuf.rsvd2 = fence;
+ a->execbuf.flags |= I915_EXEC_FENCE_IN;
+ }
+ gem_execbuf_wr(i915, &a->execbuf);
+
+ igt_spin_reset(b);
+ b->execbuf.flags = master | I915_EXEC_FENCE_OUT;
+ if (fence != -1) {
+ b->execbuf.rsvd2 = fence;
+ b->execbuf.flags |= I915_EXEC_FENCE_IN;
+ }
+ gem_execbuf_wr(i915, &b->execbuf);
+
+ if (rand() % 1)
+ igt_swap(a, b);
+
+ batch.handle = create_semaphore_to_spinner(i915, a);
+ execbuf.rsvd1 = a->execbuf.rsvd1;
+ execbuf.rsvd2 = a->execbuf.rsvd2 >> 32;
+ do {
+ execbuf.flags = rand() % count + 1;
+ } while (execbuf.flags == master);
+ execbuf.flags |= I915_EXEC_FENCE_SUBMIT;
+ gem_execbuf(i915, &execbuf);
+ gem_close(i915, batch.handle);
+
+ batch.handle = create_semaphore_to_spinner(i915, b);
+ execbuf.rsvd1 = b->execbuf.rsvd1;
+ execbuf.rsvd2 = b->execbuf.rsvd2 >> 32;
+ do {
+ execbuf.flags = rand() % count + 1;
+ } while (execbuf.flags == master);
+ execbuf.flags |= I915_EXEC_FENCE_SUBMIT;
+ gem_execbuf(i915, &execbuf);
+ gem_close(i915, batch.handle);
+
+ if (fence != -1) {
+ sw_sync_timeline_inc(timeline, 1);
+ close(fence);
+ }
Would it be worth adding another submit pattern: Am + As/Bs, Bm + Bs/As?
A bit awkward to implement, probably would need copy & paste of the
function.
+ close(a->execbuf.rsvd2 >> 32);
+ close(b->execbuf.rsvd2 >> 32);
+
+ gem_sync(i915, a->handle);
+ gem_sync(i915, b->handle);
+
+ cycles++;
+ }
+
+ *out = cycles;
+
+ close(timeline);
+
+ igt_spin_free(i915, a);
+ igt_spin_free(i915, b);
+
+ gem_context_destroy(i915, A);
+ gem_context_destroy(i915, B);
+}
+
+static void bonded_dual(int i915)
+{
+ unsigned long *cycles;
+
+ /*
+ * The purpose of bonded submission is to execute one or more requests
+ * concurrently. However, the very nature of that requires coordinated
+ * submission across multiple engines.
+ */
+ igt_require(gem_scheduler_has_preemption(i915));
+
+ cycles = mmap(0, 4096, PROT_WRITE, MAP_SHARED | MAP_ANON, -1, 0);
+
+ for (int class = 1; class < 32; class++) {
+ struct i915_engine_class_instance *siblings;
+ unsigned int count;
+
+ siblings = list_engines(i915, 1u << class, &count);
+ if (count > 1) {
Count < 2 && continue looks tempting, but up to you.
+ const unsigned int phases[] = {
+ 0,
+ BD_FENCE,
+ BD_MANY,
+ BD_HOSTILE,
+ BD_HOSTILE | BD_FENCE,
+ };
+
+ for (int i = 0; i < ARRAY_SIZE(phases); i++) {
+ memset(cycles, 0, 4096);
+ igt_fork(child, count + 1)
+ __bonded_dual(i915,
+ siblings, count,
+ phases[i],
+ &cycles[child]);
+ igt_waitchildren();
+ gem_quiescent_gpu(i915);
+
+ for (int child = 1; child < count + 1; child++)
+ cycles[0] += cycles[child];
+
+ igt_info("%s %s %s submission, %lu cycles\n",
+ phases[i] & BD_HOSTILE ? "Non-preemptible" : "Preemptible",
+ phases[i] & BD_MANY ? "many-master" : "single-master",
+ phases[i] & BD_FENCE ? "fenced" : "immediate",
+ cycles[0]);
I'd prefix the message with "%u:" class, since the looping is per class.
+ }
+ }
+ free(siblings);
+ }
+
+ munmap(cycles, 4096);
+}
+
static void __bonded_nohang(int i915, uint32_t ctx,
const struct i915_engine_class_instance *siblings,
unsigned int count,
@@ -2284,6 +2453,9 @@ igt_main
igt_subtest("bonded-semaphore")
bonded_semaphore(i915);
+ igt_subtest("bonded-dual")
+ bonded_dual(i915);
+
igt_fixture {
igt_stop_hang_detector();
}
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
