[PATCH 2/2] drm/i915: Delay disabling the user interrupt for breadcrumbs

Chris Wilson chris at chris-wilson.co.uk
Fri Feb 24 09:55:21 UTC 2017


A significant cost in setting up a wait is the overhead of enabling the
interrupt. As we disable the interrupt whenever the queue of waiters is
empty, if we are frequently waiting on alternating batches, we end up
re-enabling the interrupt on a frequent basis. We do want to disable the
interrupt during normal operations as under high load it may add several
thousand interrupts/s - we have been known in the past to occupy whole
cores with our interrupt handler after accidentally leaving user
interrupts enabled. As a compromise, leave the interrupt enabled until
the next IRQ, or the system is idle. This gives a small window for a
waiter to keep the interrupt active and not be delayed by having to
re-enable the interrupt.

v2: Restore hangcheck/missed-irq detection for continuations
v3: Be more careful restoring the hangcheck timer after reset
v4: Be more careful restoring the fake irq after reset (if required!)

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/i915_irq.c          |   5 +-
 drivers/gpu/drm/i915/intel_breadcrumbs.c | 135 ++++++++++++++++---------------
 drivers/gpu/drm/i915/intel_ringbuffer.h  |   3 +-
 3 files changed, 76 insertions(+), 67 deletions(-)

diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
index 3c79753edf0e..ba0bb33e12ed 100644
--- a/drivers/gpu/drm/i915/i915_irq.c
+++ b/drivers/gpu/drm/i915/i915_irq.c
@@ -1036,9 +1036,6 @@ static void notify_ring(struct intel_engine_cs *engine)
 	struct drm_i915_gem_request *rq = NULL;
 	struct intel_wait *wait;
 
-	if (!intel_engine_has_waiter(engine))
-		return;
-
 	atomic_inc(&engine->irq_count);
 	set_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);
 
@@ -1061,6 +1058,8 @@ static void notify_ring(struct intel_engine_cs *engine)
 			rq = wait->request;
 
 		wake_up_process(wait->tsk);
+	} else {
+		__intel_engine_disarm_breadcrumbs(engine);
 	}
 	spin_unlock(&engine->breadcrumbs.lock);
 
diff --git a/drivers/gpu/drm/i915/intel_breadcrumbs.c b/drivers/gpu/drm/i915/intel_breadcrumbs.c
index 5f2665aa817c..2f8f2706796e 100644
--- a/drivers/gpu/drm/i915/intel_breadcrumbs.c
+++ b/drivers/gpu/drm/i915/intel_breadcrumbs.c
@@ -35,8 +35,10 @@ static void intel_breadcrumbs_hangcheck(unsigned long data)
 {
 	struct intel_engine_cs *engine = (struct intel_engine_cs *)data;
 	struct intel_breadcrumbs *b = &engine->breadcrumbs;
+	struct intel_wait *wait;
+	unsigned long flags;
 
-	if (!b->irq_enabled)
+	if (!b->irq_armed)
 		return;
 
 	if (b->hangcheck_interrupts != atomic_read(&engine->irq_count)) {
@@ -45,29 +47,22 @@ static void intel_breadcrumbs_hangcheck(unsigned long data)
 		return;
 	}
 
+	spin_lock_irqsave(&b->lock, flags);
+
 	/* If the waiter was currently running, assume it hasn't had a chance
 	 * to process the pending interrupt (e.g, low priority task on a loaded
 	 * system) and wait until it sleeps before declaring a missed interrupt.
 	 */
-	if (!intel_engine_wakeup(engine)) {
+	wait = b->first_wait;
+	if (!wait || !wake_up_process(wait->tsk)) {
 		mod_timer(&b->hangcheck, wait_timeout());
-		return;
+	} else {
+		DRM_DEBUG("Hangcheck timer elapsed... %s idle\n", engine->name);
+		set_bit(engine->id, &engine->i915->gpu_error.missed_irq_rings);
+		mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1);
 	}
 
-	DRM_DEBUG("Hangcheck timer elapsed... %s idle\n", engine->name);
-	set_bit(engine->id, &engine->i915->gpu_error.missed_irq_rings);
-	mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1);
-
-	/* Ensure that even if the GPU hangs, we get woken up.
-	 *
-	 * However, note that if no one is waiting, we never notice
-	 * a gpu hang. Eventually, we will have to wait for a resource
-	 * held by the GPU and so trigger a hangcheck. In the most
-	 * pathological case, this will be upon memory starvation! To
-	 * prevent this, we also queue the hangcheck from the retire
-	 * worker.
-	 */
-	i915_queue_hangcheck(engine->i915);
+	spin_unlock_irqrestore(&b->lock, flags);
 }
 
 static void intel_breadcrumbs_fake_irq(unsigned long data)
@@ -86,6 +81,17 @@ static void intel_breadcrumbs_fake_irq(unsigned long data)
 
 	intel_engine_wakeup(engine);
 	mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1);
+
+	/* Ensure that even if the GPU hangs, we get woken up.
+	 *
+	 * However, note that if no one is waiting, we never notice
+	 * a gpu hang. Eventually, we will have to wait for a resource
+	 * held by the GPU and so trigger a hangcheck. In the most
+	 * pathological case, this will be upon memory starvation! To
+	 * prevent this, we also queue the hangcheck from the retire
+	 * worker.
+	 */
+	i915_queue_hangcheck(engine->i915);
 }
 
 static void irq_enable(struct intel_engine_cs *engine)
@@ -110,6 +116,20 @@ static void irq_disable(struct intel_engine_cs *engine)
 	spin_unlock(&engine->i915->irq_lock);
 }
 
+void __intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine)
+{
+	struct intel_breadcrumbs *b = &engine->breadcrumbs;
+
+	assert_spin_locked(&b->lock);
+
+	if (b->irq_enabled) {
+		irq_disable(engine);
+		b->irq_enabled = false;
+	}
+
+	b->irq_armed = false;
+}
+
 static bool use_fake_irq(const struct intel_breadcrumbs *b)
 {
 	const struct intel_engine_cs *engine =
@@ -127,6 +147,14 @@ static bool use_fake_irq(const struct intel_breadcrumbs *b)
 	return atomic_read(&engine->irq_count) == b->hangcheck_interrupts;
 }
 
+static void enable_fake_irq(struct intel_breadcrumbs *b)
+{
+	if (!b->irq_enabled || use_fake_irq(b))
+		mod_timer(&b->fake_irq, jiffies + 1);
+	else
+		mod_timer(&b->hangcheck, wait_timeout());
+}
+
 static void __intel_breadcrumbs_enable_irq(struct intel_breadcrumbs *b)
 {
 	struct intel_engine_cs *engine =
@@ -134,9 +162,17 @@ static void __intel_breadcrumbs_enable_irq(struct intel_breadcrumbs *b)
 	struct drm_i915_private *i915 = engine->i915;
 
 	assert_spin_locked(&b->lock);
-	if (b->rpm_wakelock)
+	if (b->irq_armed)
 		return;
 
+	/* The breadcrumb irq will be disarmed on the interrupt after the
+	 * waiters are signaled. This gives us a single interrupt window in
+	 * which we can add a new waiter and avoid the cost of re-enabling
+	 * the irq.
+	 */
+	b->irq_armed = true;
+	GEM_BUG_ON(b->irq_enabled);
+
 	if (I915_SELFTEST_ONLY(b->mock)) {
 		/* For our mock objects we want to avoid interaction
 		 * with the real hardware (which is not set up). So
@@ -145,17 +181,15 @@ static void __intel_breadcrumbs_enable_irq(struct intel_breadcrumbs *b)
 		 * implementation to call intel_engine_wakeup()
 		 * itself when it wants to simulate a user interrupt,
 		 */
-		b->rpm_wakelock = true;
 		return;
 	}
 
 	/* Since we are waiting on a request, the GPU should be busy
-	 * and should have its own rpm reference. For completeness,
-	 * record an rpm reference for ourselves to cover the
-	 * interrupt we unmask.
+	 * and should have its own rpm reference. This is tracked
+	 * by i915->gt.awake, we can forgo holding our own wakref
+	 * for the interrupt as before i915->gt.awake is released (when
+	 * the driver is idle) we disarm the breadcrumbs.
 	 */
-	intel_runtime_pm_get_noresume(i915);
-	b->rpm_wakelock = true;
 
 	/* No interrupts? Kick the waiter every jiffie! */
 	if (intel_irqs_enabled(i915)) {
@@ -164,36 +198,7 @@ static void __intel_breadcrumbs_enable_irq(struct intel_breadcrumbs *b)
 		b->irq_enabled = true;
 	}
 
-	if (!b->irq_enabled || use_fake_irq(b)) {
-		mod_timer(&b->fake_irq, jiffies + 1);
-		i915_queue_hangcheck(i915);
-	} else {
-		/* Ensure we never sleep indefinitely */
-		mod_timer(&b->hangcheck, wait_timeout());
-	}
-}
-
-static void __intel_breadcrumbs_disable_irq(struct intel_breadcrumbs *b)
-{
-	struct intel_engine_cs *engine =
-		container_of(b, struct intel_engine_cs, breadcrumbs);
-
-	assert_spin_locked(&b->lock);
-	if (!b->rpm_wakelock)
-		return;
-
-	if (I915_SELFTEST_ONLY(b->mock)) {
-		b->rpm_wakelock = false;
-		return;
-	}
-
-	if (b->irq_enabled) {
-		irq_disable(engine);
-		b->irq_enabled = false;
-	}
-
-	intel_runtime_pm_put(engine->i915);
-	b->rpm_wakelock = false;
+	enable_fake_irq(b);
 }
 
 static inline struct intel_wait *to_wait(struct rb_node *node)
@@ -415,7 +420,6 @@ static void __intel_engine_remove_wait(struct intel_engine_cs *engine,
 			wake_up_process(b->first_wait->tsk);
 		} else {
 			b->first_wait = NULL;
-			__intel_breadcrumbs_disable_irq(b);
 		}
 	} else {
 		GEM_BUG_ON(rb_first(&b->waiters) == &wait->node);
@@ -707,15 +711,20 @@ void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine)
 	cancel_fake_irq(engine);
 	spin_lock_irq(&b->lock);
 
-	__intel_breadcrumbs_disable_irq(b);
-	if (intel_engine_has_waiter(engine)) {
-		__intel_breadcrumbs_enable_irq(b);
-		if (test_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted))
-			wake_up_process(b->first_wait->tsk);
-	} else {
-		/* sanitize the IMR and unmask any auxiliary interrupts */
+	if (b->irq_enabled)
+		irq_enable(engine);
+	else
 		irq_disable(engine);
-	}
+
+	/* The engine is currently idle (we haven't started it yet), there
+	 * is no possibility for a missed interrupt as we enabled the irq
+	 * and so we can clear the immediate wakeup (until a real interrupt
+	 * arrives for the waiter).
+	 */
+	clear_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);
+
+	if (b->irq_armed)
+		enable_fake_irq(b);
 
 	spin_unlock_irq(&b->lock);
 }
diff --git a/drivers/gpu/drm/i915/intel_ringbuffer.h b/drivers/gpu/drm/i915/intel_ringbuffer.h
index c90b60f47f25..c33b39fd9c5c 100644
--- a/drivers/gpu/drm/i915/intel_ringbuffer.h
+++ b/drivers/gpu/drm/i915/intel_ringbuffer.h
@@ -246,8 +246,8 @@ struct intel_engine_cs {
 
 		unsigned int hangcheck_interrupts;
 
+		bool irq_armed : 1;
 		bool irq_enabled : 1;
-		bool rpm_wakelock : 1;
 		I915_SELFTEST_DECLARE(bool mock : 1);
 	} breadcrumbs;
 
@@ -657,6 +657,7 @@ static inline bool intel_engine_wakeup(struct intel_engine_cs *engine)
 	return wakeup;
 }
 
+void __intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine);
 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine);
 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine);
 bool intel_breadcrumbs_busy(struct intel_engine_cs *engine);
-- 
2.11.0



More information about the Intel-gfx-trybot mailing list