Intel-GFX Archive on lore.kernel.org
 help / color / mirror / Atom feed
* [Intel-gfx] [PATCH] drm/i915: Improve the start alignment of bonded pairs
@ 2020-03-05  8:28 Chris Wilson
  2020-03-05 11:44 ` Mika Kuoppala
                   ` (3 more replies)
  0 siblings, 4 replies; 5+ messages in thread
From: Chris Wilson @ 2020-03-05  8:28 UTC (permalink / raw)
  To: intel-gfx

Always wait on the start of the signaler request to reduce the problem
of dequeueing the bonded pair too early -- we want both payloads to
start at the same time, with no latency, and yet still allow others to
make full use of the slack in the system.

Remindme: add testcases for starting the bonded pair too early due to an
infinite spin before the signaler, and via a semaphore.

Testcase: XXX
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
---
 drivers/gpu/drm/i915/i915_request.c | 33 ++++++++++++++++++++++++-----
 1 file changed, 28 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index 4bfe68edfc81..c0a0089111a1 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -1127,14 +1127,37 @@ __i915_request_await_execution(struct i915_request *to,
 					  &from->fence))
 		return 0;
 
-	/* Ensure both start together [after all semaphores in signal] */
-	if (intel_engine_has_semaphores(to->engine))
-		err = __emit_semaphore_wait(to, from, from->fence.seqno - 1);
-	else
-		err = i915_request_await_start(to, from);
+	/*
+	 * Wait until the start of this request.
+	 *
+	 * The execution cb fires when we submit the request to HW. But in
+	 * many cases this may be long before the request itself is ready to
+	 * run (consider that we submit 2 requests for the same context, where
+	 * the request of interest is behind an indefinite spinner). So we hook
+	 * up to both to reduce our queues tidy and execution lag minimised in
+	 * the worst case, though we hope that the await_start is elided.
+	 */
+	err = i915_request_await_start(to, from);
 	if (err < 0)
 		return err;
 
+	/*
+	 * Ensure both start together [after all semaphores in signal]
+	 *
+	 * Now that we are queued to the HW at roughly the same time (thanks
+	 * to the execute cb) and are ready to run at roughly the same time
+	 * (thanks to the await start), our signaler may still be indefinitely
+	 * delayed by waiting on a semaphore from a remote engine. If our
+	 * signaler depends on a sempahore, so indirectly do we, and we do not
+	 * want to start our payload until our signaler also starts theirs.
+	 * So we wait.
+	 */
+	if (intel_engine_has_semaphores(to->engine) && from->sched.semaphores) {
+		err = __emit_semaphore_wait(to, from, from->fence.seqno - 1);
+		if (err < 0)
+			return err;
+	}
+
 	/* Couple the dependency tree for PI on this exposed to->fence */
 	if (to->engine->schedule) {
 		err = i915_sched_node_add_dependency(&to->sched, &from->sched);
-- 
2.25.1

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

^ permalink raw reply related	[flat|nested] 5+ messages in thread
end of thread, other threads:[~2020-03-06 11:45 UTC | newest]

Thread overview: 5+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2020-03-05  8:28 [Intel-gfx] [PATCH] drm/i915: Improve the start alignment of bonded pairs Chris Wilson
2020-03-05 11:44 ` Mika Kuoppala
2020-03-05 20:20 ` [Intel-gfx] ✗ Fi.CI.DOCS: warning for " Patchwork
2020-03-05 20:30 ` [Intel-gfx] ✓ Fi.CI.BAT: success " Patchwork
2020-03-06 11:45 ` [Intel-gfx] ✓ Fi.CI.IGT: " Patchwork

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox