[PATCH] drm/i915: Only start retire worker when idle

[Chris Wilson <chris@chris-wilson.co.uk>]

The retire worker is a low frequency task that makes sure we retire
outstanding requests if userspace is being lax. We only need to start it
once as it remains active until the GPU is idle, so do a cheap test
before the more expensive queue_work(). A consequence of this is that we
need correct locking in the worker to make the hot path of request
submission cheap. To keep the symmetry and keep hangcheck strictly bound
by the GPU's wakelock, we move the cancel_sync(hangcheck) to the idle
worker before dropping the wakelock.

v2: Guard against RCU fouling the breadcrumbs bottom-half whilst we kick
the waiter.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
References: https://bugs.freedesktop.org/show_bug.cgi?id=88437
---
 drivers/gpu/drm/i915/i915_drv.c      |  2 -
 drivers/gpu/drm/i915/i915_drv.h      |  2 +-
 drivers/gpu/drm/i915/i915_gem.c      | 90 +++++++++++++++++++++++++-----------
 drivers/gpu/drm/i915/intel_display.c | 29 ------------
 4 files changed, 64 insertions(+), 59 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
index ba91f65b6082..0f79ee1d35a2 100644
--- a/drivers/gpu/drm/i915/i915_drv.c
+++ b/drivers/gpu/drm/i915/i915_drv.c
@@ -1472,8 +1472,6 @@ static int intel_runtime_suspend(struct device *device)
 	i915_gem_release_all_mmaps(dev_priv);
 	mutex_unlock(&dev->struct_mutex);
 
-	cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work);
-
 	intel_guc_suspend(dev);
 
 	intel_suspend_gt_powersave(dev);
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index bbaa12e85c68..9350ce1f0bfa 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -3001,7 +3001,7 @@ int __must_check i915_gem_set_seqno(struct drm_device *dev, u32 seqno);
 struct drm_i915_gem_request *
 i915_gem_find_active_request(struct intel_engine_cs *ring);
 
-bool i915_gem_retire_requests(struct drm_device *dev);
+void i915_gem_retire_requests(struct drm_device *dev);
 void i915_gem_retire_requests_ring(struct intel_engine_cs *ring);
 
 static inline u32 i915_reset_counter(struct i915_gpu_error *error)
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
index 5719e8d9e33b..2350e00f2e45 100644
--- a/drivers/gpu/drm/i915/i915_gem.c
+++ b/drivers/gpu/drm/i915/i915_gem.c
@@ -2495,6 +2495,48 @@ i915_gem_get_seqno(struct drm_device *dev, u32 *seqno)
 	return 0;
 }
 
+static void i915_gem_mark_busy(struct drm_i915_private *dev_priv)
+{
+	if (dev_priv->mm.busy)
+		return;
+
+	intel_runtime_pm_get_noresume(dev_priv);
+
+	i915_update_gfx_val(dev_priv);
+	if (INTEL_INFO(dev_priv)->gen >= 6)
+		gen6_rps_busy(dev_priv);
+
+	queue_delayed_work(dev_priv->wq,
+			   &dev_priv->mm.retire_work,
+			   round_jiffies_up_relative(HZ));
+
+	dev_priv->mm.busy = true;
+}
+
+static void kick_waiters(struct drm_i915_private *dev_priv)
+{
+	struct intel_engine_cs *ring;
+	int i;
+
+	rcu_read_lock();
+	for_each_ring(ring, dev_priv, i)
+		intel_engine_wakeup(ring);
+	rcu_read_unlock();
+}
+
+static void i915_gem_mark_idle(struct drm_i915_private *dev_priv)
+{
+	dev_priv->mm.busy = false;
+
+	if (cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work))
+		kick_waiters(dev_priv);
+
+	if (INTEL_INFO(dev_priv)->gen >= 6)
+		gen6_rps_idle(dev_priv);
+
+	intel_runtime_pm_put(dev_priv);
+}
+
 /*
  * NB: This function is not allowed to fail. Doing so would mean the the
  * request is not being tracked for completion but the work itself is
@@ -2575,10 +2617,7 @@ void __i915_add_request(struct drm_i915_gem_request *request,
 
 	trace_i915_gem_request_add(request);
 
-	queue_delayed_work(dev_priv->wq,
-			   &dev_priv->mm.retire_work,
-			   round_jiffies_up_relative(HZ));
-	intel_mark_busy(dev_priv->dev);
+	i915_gem_mark_busy(dev_priv);
 
 	/* Sanity check that the reserved size was large enough. */
 	intel_ring_reserved_space_end(ringbuf);
@@ -2903,7 +2942,7 @@ i915_gem_retire_requests_ring(struct intel_engine_cs *ring)
 	WARN_ON(i915_verify_lists(ring->dev));
 }
 
-bool
+void
 i915_gem_retire_requests(struct drm_device *dev)
 {
 	struct drm_i915_private *dev_priv = dev->dev_private;
@@ -2929,8 +2968,6 @@ i915_gem_retire_requests(struct drm_device *dev)
 		mod_delayed_work(dev_priv->wq,
 				 &dev_priv->mm.idle_work,
 				 msecs_to_jiffies(100));
-
-	return idle;
 }
 
 static void
@@ -2939,17 +2976,21 @@ i915_gem_retire_work_handler(struct work_struct *work)
 	struct drm_i915_private *dev_priv =
 		container_of(work, typeof(*dev_priv), mm.retire_work.work);
 	struct drm_device *dev = dev_priv->dev;
-	bool idle;
 
 	/* Come back later if the device is busy... */
-	idle = false;
 	if (mutex_trylock(&dev->struct_mutex)) {
-		idle = i915_gem_retire_requests(dev);
+		i915_gem_retire_requests(dev);
 		mutex_unlock(&dev->struct_mutex);
 	}
-	if (!idle) {
+
+	/* Keep the retire handler running until we are finally idle.
+	 * We do not need to do this test under locking as in the worst-case
+	 * we queue the retire worker once too often.
+	 */
+	if (READ_ONCE(dev_priv->mm.busy)) {
 		i915_queue_hangcheck(dev_priv);
-		queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work,
+		queue_delayed_work(dev_priv->wq,
+				   &dev_priv->mm.retire_work,
 				   round_jiffies_up_relative(HZ));
 	}
 }
@@ -2963,25 +3004,20 @@ i915_gem_idle_work_handler(struct work_struct *work)
 	struct intel_engine_cs *ring;
 	int i;
 
-	for_each_ring(ring, dev_priv, i)
-		if (!list_empty(&ring->request_list))
-			return;
-
-	/* we probably should sync with hangcheck here, using cancel_work_sync.
-	 * Also locking seems to be fubar here, ring->request_list is protected
-	 * by dev->struct_mutex. */
+	if (!mutex_trylock(&dev->struct_mutex))
+		return;
 
-	intel_mark_idle(dev);
+	for_each_ring(ring, dev_priv, i) {
+		if (!list_empty(&ring->request_list))
+			goto out;
 
-	if (mutex_trylock(&dev->struct_mutex)) {
-		struct intel_engine_cs *ring;
-		int i;
+		i915_gem_batch_pool_fini(&ring->batch_pool);
+	}
 
-		for_each_ring(ring, dev_priv, i)
-			i915_gem_batch_pool_fini(&ring->batch_pool);
+	i915_gem_mark_idle(dev_priv);
 
-		mutex_unlock(&dev->struct_mutex);
-	}
+out:
+	mutex_unlock(&dev->struct_mutex);
 }
 
 /**
diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
index ffcdc2c631e1..3c3766eab5c4 100644
--- a/drivers/gpu/drm/i915/intel_display.c
+++ b/drivers/gpu/drm/i915/intel_display.c
@@ -10810,35 +10810,6 @@ struct drm_display_mode *intel_crtc_mode_get(struct drm_device *dev,
 	return mode;
 }
 
-void intel_mark_busy(struct drm_device *dev)
-{
-	struct drm_i915_private *dev_priv = dev->dev_private;
-
-	if (dev_priv->mm.busy)
-		return;
-
-	intel_runtime_pm_get(dev_priv);
-	i915_update_gfx_val(dev_priv);
-	if (INTEL_INFO(dev)->gen >= 6)
-		gen6_rps_busy(dev_priv);
-	dev_priv->mm.busy = true;
-}
-
-void intel_mark_idle(struct drm_device *dev)
-{
-	struct drm_i915_private *dev_priv = dev->dev_private;
-
-	if (!dev_priv->mm.busy)
-		return;
-
-	dev_priv->mm.busy = false;
-
-	if (INTEL_INFO(dev)->gen >= 6)
-		gen6_rps_idle(dev->dev_private);
-
-	intel_runtime_pm_put(dev_priv);
-}
-
 static void intel_crtc_destroy(struct drm_crtc *crtc)
 {
 	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
-- 
2.6.4

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