[Intel-gfx] [PATCH] drm/i915/gt: Track engine round-trip times

[Chris Wilson]

Knowing the round trip time of an engine is useful for tracking the
health of the system as well as providing a metric for the baseline
responsiveness of the engine. We can use the latter metric for
automatically tuning our waits in selftests and when idling so we don't
confuse a slower system with a dead one.

Upon idling the engine, we send one last pulse to switch the context
away from precious user state to the volatile kernel context. We know
the engine is idle at this point, and the pulse is non-preemptable, so
this provides us with a good measurement of the round trip time. A
secondary effect is that by installing an interrupt onto the pulse, we
can flush the engine immediately upon completion, curtailing the
background flush and entering powersaving immediately.

References: 7e34f4e4aad3 ("drm/i915/gen8+: Add RC6 CTX corruption WA")
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
Cc: Stuart Summers <stuart.summers at intel.com>
---
 drivers/gpu/drm/i915/gt/intel_engine_cs.c    |  2 ++
 drivers/gpu/drm/i915/gt/intel_engine_pm.c    | 34 +++++++++++++++++++-
 drivers/gpu/drm/i915/gt/intel_engine_types.h |  6 ++++
 3 files changed, 41 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/i915/gt/intel_engine_cs.c b/drivers/gpu/drm/i915/gt/intel_engine_cs.c
index b9613d044393..2d11db13dc89 100644
--- a/drivers/gpu/drm/i915/gt/intel_engine_cs.c
+++ b/drivers/gpu/drm/i915/gt/intel_engine_cs.c
@@ -334,6 +334,7 @@ static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id)
 	/* Nothing to do here, execute in order of dependencies */
 	engine->schedule = NULL;
 
+	ewma_delay_init(&engine->delay);
 	seqlock_init(&engine->stats.lock);
 
 	ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
@@ -1477,6 +1478,7 @@ void intel_engine_dump(struct intel_engine_cs *engine,
 		drm_printf(m, "*** WEDGED ***\n");
 
 	drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
+	drm_printf(m, "\tDelay: %luus\n", ewma_delay_read(&engine->delay));
 
 	rcu_read_lock();
 	rq = READ_ONCE(engine->heartbeat.systole);
diff --git a/drivers/gpu/drm/i915/gt/intel_engine_pm.c b/drivers/gpu/drm/i915/gt/intel_engine_pm.c
index 3c0f490ff2c7..1175f9a63882 100644
--- a/drivers/gpu/drm/i915/gt/intel_engine_pm.c
+++ b/drivers/gpu/drm/i915/gt/intel_engine_pm.c
@@ -73,6 +73,23 @@ static inline void __timeline_mark_unlock(struct intel_context *ce,
 
 #endif /* !IS_ENABLED(CONFIG_LOCKDEP) */
 
+struct duration_cb {
+	struct dma_fence_cb cb;
+	ktime_t emitted;
+};
+
+static void duration_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
+{
+	struct duration_cb *dcb = container_of(cb, typeof(*dcb), cb);
+	struct intel_engine_cs *engine = to_request(fence)->engine;
+
+	ewma_delay_add(&engine->delay,
+		       ktime_us_delta(ktime_get(), dcb->emitted));
+
+	/* Kick retire for quicker powersaving (soft-rc6). */
+	mod_delayed_work(system_wq, &engine->gt->requests.retire_work, 0);
+}
+
 static bool switch_to_kernel_context(struct intel_engine_cs *engine)
 {
 	struct i915_request *rq;
@@ -114,7 +131,22 @@ static bool switch_to_kernel_context(struct intel_engine_cs *engine)
 
 	/* Install ourselves as a preemption barrier */
 	rq->sched.attr.priority = I915_PRIORITY_BARRIER;
-	__i915_request_commit(rq);
+	if (likely(!__i915_request_commit(rq))) { /* engine should be idle! */
+		struct duration_cb *dcb;
+
+		BUILD_BUG_ON(sizeof(*dcb) > sizeof(rq->submitq));
+		dcb = (struct duration_cb *)&rq->submitq;
+
+		/*
+		 * Use an interrupt for precise measurement of duration,
+		 * otherwise we rely on someone else retiring all the requests
+		 * which may delay the signaling (i.e. we will likely wait
+		 * until the background request retirement running every
+		 * second or two).
+		 */
+		dma_fence_add_callback(&rq->fence, &dcb->cb, duration_cb);
+		dcb->emitted = ktime_get();
+	}
 
 	/* Release our exclusive hold on the engine */
 	__intel_wakeref_defer_park(&engine->wakeref);
diff --git a/drivers/gpu/drm/i915/gt/intel_engine_types.h b/drivers/gpu/drm/i915/gt/intel_engine_types.h
index 758f0e8ec672..c6a607d9cf6a 100644
--- a/drivers/gpu/drm/i915/gt/intel_engine_types.h
+++ b/drivers/gpu/drm/i915/gt/intel_engine_types.h
@@ -7,6 +7,7 @@
 #ifndef __INTEL_ENGINE_TYPES__
 #define __INTEL_ENGINE_TYPES__
 
+#include <linux/average.h>
 #include <linux/hashtable.h>
 #include <linux/irq_work.h>
 #include <linux/kref.h>
@@ -119,6 +120,9 @@ enum intel_engine_id {
 #define INVALID_ENGINE ((enum intel_engine_id)-1)
 };
 
+/* A simple estimator for the round-trip responsive time of an engine */
+DECLARE_EWMA(delay, 6, 4)
+
 struct st_preempt_hang {
 	struct completion completion;
 	unsigned int count;
@@ -316,6 +320,8 @@ struct intel_engine_cs {
 		struct intel_timeline *timeline;
 	} legacy;
 
+	struct ewma_delay delay;
+
 	/* Rather than have every client wait upon all user interrupts,
 	 * with the herd waking after every interrupt and each doing the
 	 * heavyweight seqno dance, we delegate the task (of being the
-- 
2.24.0