[Intel-gfx] [PATCH i-g-t] i915/gem_exec_balancer: Race SUBMIT_FENCE against semaphores

[Chris Wilson]

The scheduler may use a semaphore between engines to serialise requests,
and in doing so submit the request before its signalers are ready. This
dependency must also be copied across any SUBMIT_FENCE so that a
bonded-pair will not execute ahead of any of its implicit dependencies
-- it too must wait for the same semaphores.

Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
---
 tests/i915/gem_exec_balancer.c | 104 +++++++++++++++++++++++++++++++++
 1 file changed, 104 insertions(+)

diff --git a/tests/i915/gem_exec_balancer.c b/tests/i915/gem_exec_balancer.c
index e5e83508f..5f6e3d568 100644
--- a/tests/i915/gem_exec_balancer.c
+++ b/tests/i915/gem_exec_balancer.c
@@ -959,6 +959,107 @@ static void bonded_chain(int i915)
 	gem_context_destroy(i915, ctx);
 }
 
+static void __bonded_sema(int i915, uint32_t ctx,
+			  const struct i915_engine_class_instance *siblings,
+			  unsigned int count)
+{
+	const int priorities[] = { -1023, 0, 1023 };
+	struct drm_i915_gem_exec_object2 batch = {
+		.handle = batch_create(i915),
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(&batch),
+		.buffer_count = 1,
+		.rsvd1 = ctx,
+	};
+	igt_spin_t *spin;
+
+	for (int i = 0; i < ARRAY_SIZE(priorities); i++) {
+		/* A: spin forever on seperate render engine */
+		spin = igt_spin_new(i915,
+				    .flags = (IGT_SPIN_POLL_RUN |
+					      IGT_SPIN_FENCE_OUT));
+		igt_spin_busywait_until_started(spin);
+
+		/*
+		 * Note we replace the timelines between each execbuf, so
+		 * that any pair of requests on the same engine could be
+		 * re-ordered by the scheduler -- if the dependency tracking
+		 * is subpar.
+		 */
+
+		/* B: waits for A (using a semaphore) on engine 1 */
+		set_load_balancer(i915, ctx, siblings, count, NULL);
+		execbuf.rsvd2 = spin->out_fence;
+		execbuf.flags = I915_EXEC_FENCE_IN | I915_EXEC_FENCE_OUT;
+		execbuf.flags |= 1;
+		gem_execbuf_wr(i915, &execbuf);
+
+		/* B': run in parallel with B on engine 2 */
+		if (priorities[i] > 0)
+			gem_context_set_priority(i915, ctx, priorities[i]);
+		set_load_balancer(i915, ctx, siblings, count, NULL);
+		execbuf.flags = I915_EXEC_FENCE_SUBMIT | I915_EXEC_FENCE_OUT;
+		execbuf.flags |= 2;
+		execbuf.rsvd2 >>= 32;
+		gem_execbuf_wr(i915, &execbuf);
+		gem_context_set_priority(i915, ctx, 0);
+
+		/* Wait for any magic timeslicing or preemptions... */
+		igt_assert_eq(sync_fence_wait(execbuf.rsvd2 >> 32, 1000),
+			      -ETIME);
+
+		igt_debugfs_dump(i915, "i915_engine_info");
+
+		/*
+		 * ... which should not have happened, so everything is still
+		 * waiting on the spinner
+		 */
+		igt_assert_eq(sync_fence_status(spin->out_fence), 0);
+		igt_assert_eq(sync_fence_status(execbuf.rsvd2 & 0xffffffff), 0);
+		igt_assert_eq(sync_fence_status(execbuf.rsvd2 >> 32), 0);
+
+		igt_spin_free(i915, spin);
+		gem_sync(i915, batch.handle);
+
+		igt_assert_eq(sync_fence_status(execbuf.rsvd2 & 0xffffffff), 1);
+		igt_assert_eq(sync_fence_status(execbuf.rsvd2 >> 32), 1);
+
+		close(execbuf.rsvd2);
+		close(execbuf.rsvd2 >> 32);
+	}
+
+	gem_close(i915, batch.handle);
+}
+
+static void bonded_semaphore(int i915)
+{
+	uint32_t ctx;
+
+	/*
+	 * Given batches A, B and B', where B and B' are a bonded pair, with
+	 * B' depending on B with a submit fence and B depending on A as
+	 * an ordinary fence; prove B' cannot complete before A, with the
+	 * difference here (wrt bonded_chain) that A is on another engine and
+	 * so A, B and B' are expected to be inflight concurrently.
+	 */
+	igt_require(gem_scheduler_has_semaphores(i915));
+
+	ctx = gem_context_create(i915);
+
+	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)
+			__bonded_sema(i915, ctx, siblings, count);
+		free(siblings);
+	}
+
+	gem_context_destroy(i915, ctx);
+}
+
 static void indices(int i915)
 {
 	I915_DEFINE_CONTEXT_PARAM_ENGINES(engines, I915_EXEC_RING_MASK + 1);
@@ -1685,6 +1786,9 @@ igt_main
 	igt_subtest("bonded-chain")
 		bonded_chain(i915);
 
+	igt_subtest("bonded-semaphore")
+		bonded_semaphore(i915);
+
 	igt_fixture {
 		igt_stop_hang_detector();
 	}
-- 
2.24.0