[v3,5/5] drm/i915: Speed up idle detection by kicking the tasklets

Commit Message

Chris Wilson May 9, 2018, 12:04 p.m. UTC

We rely on ksoftirqd to run in a timely fashion in order to drain the
execlists queue. Quite frequently, it does not. In some cases we may see
latencies of over 200ms triggering our idle timeouts and forcing us to
declare the driver wedged!

Thus we can speed up idle detection by bypassing ksoftirqd in these
cases and flush our tasklet to confirm if we are indeed still waiting
for the ELSP to drain.

v2: Put the execlists.first check back; it is required for handling
reset!
v3: Follow Mika's suggestion to try and limit kicking the tasklet to
only when we expect it to make a difference, i.e. in catch up after a CS
interrupt, and not just execute it everytime as that is likely just to
cover over our own bugs.

References: https://bugs.freedesktop.org/show_bug.cgi?id=106373
Signed-off-by: Chris Wilson <[e-mail]>
Cc: Tvrtko Ursulin <[e-mail]>
Cc: Mika Kuoppala <[e-mail]>
Reviewed-by: Mika Kuoppala <[e-mail]>
---
 drivers/gpu/drm/i915/intel_engine_cs.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

Patch

diff --git a/drivers/gpu/drm/i915/intel_engine_cs.c b/drivers/gpu/drm/i915/intel_engine_cs.c
index 3c8a0c3245f3..f975091c5498 100644
--- a/drivers/gpu/drm/i915/intel_engine_cs.c
+++ b/drivers/gpu/drm/i915/intel_engine_cs.c
@@ -944,11 +944,20 @@  bool intel_engine_is_idle(struct intel_engine_cs *engine)
 	if (I915_SELFTEST_ONLY(engine->breadcrumbs.mock))
 		return true;
 
+	/*
+	 * ksoftirqd has notorious latency that may cause us to
+	 * timeout while waiting for the engine to idle as we wait for
+	 * ksoftirqd to run the execlists tasklet to drain the ELSP.
+	 * If we are expecting a context switch from the GPU, check now.
+	 */
+	if (test_bit(ENGINE_IRQ_EXECLIST, &engine->irq_posted))
+		i915_tasklet_try(&engine->execlists.tasklet);
+
 	/* Waiting to drain ELSP? */
 	if (READ_ONCE(engine->execlists.active))
 		return false;
 
-	/* ELSP is empty, but there are ready requests? */
+	/* ELSP is empty, but there are ready requests? E.g. after reset */
 	if (READ_ONCE(engine->execlists.first))
 		return false;