drm/i915/gt: Flush other retirees inside intel_gt_retire_requests()

Our goal in wait_for_idle (intel_gt_retire_requests) is to the current
workload *and* their idle barriers. This requires us to notice the late
arrival of those, which is done by inspecting the list of active
timelines. However, if a concurrent retirer is running that new timeline
may not be added until after we drop the lock -- so flush concurrent
retirers before we take the lock and inspect the list.

Closes: https://gitlab.freedesktop.org/drm/intel/issues/878


Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191223211008.2371613-1-chris@chris-wilson.co.uk

drivers/gpu/drm/i915/gt/intel_engine.h

@@ -282,7 +282,7 @@ static inline void __intel_engine_reset(struct intel_engine_cs *engine,
 
 bool intel_engines_are_idle(struct intel_gt *gt);
 bool intel_engine_is_idle(struct intel_engine_cs *engine);
-bool intel_engine_flush_submission(struct intel_engine_cs *engine);
+void intel_engine_flush_submission(struct intel_engine_cs *engine);
 
 void intel_engines_reset_default_submission(struct intel_gt *gt);
 

drivers/gpu/drm/i915/gt/intel_engine_cs.c

@@ -1047,10 +1047,9 @@ static bool ring_is_idle(struct intel_engine_cs *engine)
 	return idle;
 }
 
-bool intel_engine_flush_submission(struct intel_engine_cs *engine)
+void intel_engine_flush_submission(struct intel_engine_cs *engine)
 {
 	struct tasklet_struct *t = &engine->execlists.tasklet;
-	bool active = tasklet_is_locked(t);
 
 	if (__tasklet_is_scheduled(t)) {
 		local_bh_disable();
@@ -1061,13 +1060,10 @@ bool intel_engine_flush_submission(struct intel_engine_cs *engine)
 			tasklet_unlock(t);
 		}
 		local_bh_enable();
-		active = true;
 	}
 
 	/* Otherwise flush the tasklet if it was running on another cpu */
 	tasklet_unlock_wait(t);
-
-	return active;
 }
 
 /**

drivers/gpu/drm/i915/gt/intel_gt_requests.c

@@ -26,21 +26,18 @@ static bool retire_requests(struct intel_timeline *tl)
 	return !i915_active_fence_isset(&tl->last_request);
 }
 
-static bool flush_submission(struct intel_gt *gt)
+static void flush_submission(struct intel_gt *gt)
 {
 	struct intel_engine_cs *engine;
 	enum intel_engine_id id;
-	bool active = false;
 
 	if (!intel_gt_pm_is_awake(gt))
-		return false;
+		return;
 
 	for_each_engine(engine, gt, id) {
-		active |= intel_engine_flush_submission(engine);
-		active |= flush_work(&engine->retire_work);
+		intel_engine_flush_submission(engine);
+		flush_work(&engine->retire_work);
 	}
-
-	return active;
 }
 
 static void engine_retire(struct work_struct *work)
@@ -126,7 +123,6 @@ long intel_gt_retire_requests_timeout(struct intel_gt *gt, long timeout)
 		timeout = -timeout, interruptible = false;
 
 	flush_submission(gt); /* kick the ksoftirqd tasklets */
-
 	spin_lock(&timelines->lock);
 	list_for_each_entry_safe(tl, tn, &timelines->active_list, link) {
 		if (!mutex_trylock(&tl->mutex)) {
@@ -153,6 +149,7 @@ long intel_gt_retire_requests_timeout(struct intel_gt *gt, long timeout)
 
 		active_count += !retire_requests(tl);
 
+		flush_submission(gt); /* sync with concurrent retirees */
 		spin_lock(&timelines->lock);
 
 		/* Resume iteration after dropping lock */
@@ -173,9 +170,6 @@ long intel_gt_retire_requests_timeout(struct intel_gt *gt, long timeout)
 	list_for_each_entry_safe(tl, tn, &free, link)
 		__intel_timeline_free(&tl->kref);
 
-	if (flush_submission(gt))
-		active_count++;
-
 	return active_count ? timeout : 0;
 }