[Intel-gfx] [PATCH] drm/i915/execlists: Prevent GPU death on ELSP[1] promotion to idle context

[Chris Wilson]

In what seems remarkably similar to the w/a required to not reload an
idle context with HEAD==TAIL, it appears we must prevent the HW from
switching to an idle context in ELSP[1], while simultaneously trying to
preempt the HW to run another context and a continuation of the idle
context (which is no longer idle).

Fortunately, we just so happen to have a semaphore in place to prevent
the ring HEAD from proceeding past the end of a request that we can use
to fix the HEAD in position as we reprogram ELSP.

Closes: https://gitlab.freedesktop.org/drm/intel/issues/1501
Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
Cc: Mika Kuoppala <mika.kuoppala at linux.intel.com>
---
 drivers/gpu/drm/i915/gt/intel_lrc.c | 71 ++++++++++++++---------------
 1 file changed, 35 insertions(+), 36 deletions(-)

diff --git a/drivers/gpu/drm/i915/gt/intel_lrc.c b/drivers/gpu/drm/i915/gt/intel_lrc.c
index b12355048501..9e24ff7451a9 100644
--- a/drivers/gpu/drm/i915/gt/intel_lrc.c
+++ b/drivers/gpu/drm/i915/gt/intel_lrc.c
@@ -1854,7 +1854,7 @@ static inline void clear_ports(struct i915_request **ports, int count)
 	memset_p((void **)ports, NULL, count);
 }
 
-static void execlists_dequeue(struct intel_engine_cs *engine)
+static bool execlists_dequeue(struct intel_engine_cs *engine)
 {
 	struct intel_engine_execlists * const execlists = &engine->execlists;
 	struct i915_request **port = execlists->pending;
@@ -1928,13 +1928,6 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 				     execlists->queue_priority_hint);
 			record_preemption(execlists);
 
-			/*
-			 * Don't let the RING_HEAD advance past the breadcrumb
-			 * as we unwind (and until we resubmit) so that we do
-			 * not accidentally tell it to go backwards.
-			 */
-			ring_set_paused(engine, 1);
-
 			/*
 			 * Note that we have not stopped the GPU at this point,
 			 * so we are unwinding the incomplete requests as they
@@ -1954,7 +1947,6 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 				     last->sched.attr.priority,
 				     execlists->queue_priority_hint);
 
-			ring_set_paused(engine, 1);
 			defer_active(engine);
 
 			/*
@@ -1988,7 +1980,7 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 				 * of timeslices, our queue might be.
 				 */
 				start_timeslice(engine);
-				return;
+				return false;
 			}
 		}
 	}
@@ -2021,9 +2013,10 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 			}
 
 			if (last && !can_merge_rq(last, rq)) {
+				/* leave this for another sibling */
 				spin_unlock(&ve->base.active.lock);
 				start_timeslice(engine);
-				return; /* leave this for another sibling */
+				return false;
 			}
 
 			ENGINE_TRACE(engine,
@@ -2193,32 +2186,30 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
 	 * interrupt for secondary ports).
 	 */
 	execlists->queue_priority_hint = queue_prio(execlists);
+	if (!submit)
+		return false;
 
-	if (submit) {
-		*port = execlists_schedule_in(last, port - execlists->pending);
-		execlists->switch_priority_hint =
-			switch_prio(engine, *execlists->pending);
+	*port = execlists_schedule_in(last, port - execlists->pending);
+	execlists->switch_priority_hint =
+		switch_prio(engine, *execlists->pending);
 
-		/*
-		 * Skip if we ended up with exactly the same set of requests,
-		 * e.g. trying to timeslice a pair of ordered contexts
-		 */
-		if (!memcmp(active, execlists->pending,
-			    (port - execlists->pending + 1) * sizeof(*port))) {
-			do
-				execlists_schedule_out(fetch_and_zero(port));
-			while (port-- != execlists->pending);
-
-			goto skip_submit;
-		}
-		clear_ports(port + 1, last_port - port);
+	/*
+	 * Skip if we ended up with exactly the same set of requests,
+	 * e.g. trying to timeslice a pair of ordered contexts
+	 */
+	if (!memcmp(active, execlists->pending,
+		    (port - execlists->pending + 1) * sizeof(*port))) {
+		do
+			execlists_schedule_out(fetch_and_zero(port));
+		while (port-- != execlists->pending);
 
-		execlists_submit_ports(engine);
-		set_preempt_timeout(engine, *active);
-	} else {
-skip_submit:
-		ring_set_paused(engine, 0);
+		return false;
 	}
+	clear_ports(port + 1, last_port - port);
+
+	execlists_submit_ports(engine);
+	set_preempt_timeout(engine, *active);
+	return true;
 }
 
 static void
@@ -2478,7 +2469,16 @@ static void __execlists_submission_tasklet(struct intel_engine_cs *const engine)
 	lockdep_assert_held(&engine->active.lock);
 	if (!READ_ONCE(engine->execlists.pending[0])) {
 		rcu_read_lock(); /* protect peeking at execlists->active */
-		execlists_dequeue(engine);
+
+		/*
+		 * Don't let the RING_HEAD advance past the breadcrumb
+		 * as we unwind (and until we resubmit) so that we do
+		 * not accidentally tell it to go backwards.
+		 */
+		ring_set_paused(engine, 1);
+		if (!execlists_dequeue(engine))
+			ring_set_paused(engine, 0);
+
 		rcu_read_unlock();
 	}
 }
@@ -2816,8 +2816,7 @@ static void execlists_reset(struct intel_engine_cs *engine, const char *msg)
 	ring_set_paused(engine, 1); /* Freeze the current request in place */
 	if (execlists_capture(engine))
 		intel_engine_reset(engine, msg);
-	else
-		ring_set_paused(engine, 0);
+	ring_set_paused(engine, 0);
 
 	tasklet_enable(&engine->execlists.tasklet);
 	clear_and_wake_up_bit(bit, lock);
-- 
2.20.1