drm/i915/pmu: Suspend sampling when GPU is idle

If only a subset of events is enabled we can afford to suspend
the sampling timer when the GPU is idle and so save some cycles
and power.

v2: Rebase and limit timer even more.
v3: Rebase.
v4: Rebase.
v5: Skip action if perf PMU failed to register.
v6: Checkpatch cleanup.
v7:
 * Add a common helper to start the timer if needed. (Chris Wilson)
 * Add comment explaining bitwise logic in pmu_needs_timer.
v8: Fix some comments styles. (Chris Wilson)
v9: Rebase.
v10: Move function declarations to i915_pmu.h.
v11: Rename functions to i915_pmu_gt_(un)parked. (Chris Wilson)

Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20171121181852.16128-3-tvrtko.ursulin@linux.intel.com

drivers/gpu/drm/i915/i915_gem.c

@@ -3372,6 +3372,7 @@ i915_gem_idle_work_handler(struct work_struct *work)
 
 	intel_engines_park(dev_priv);
 	i915_gem_timelines_mark_idle(dev_priv);
+	i915_pmu_gt_parked(dev_priv);
 
 	GEM_BUG_ON(!dev_priv->gt.awake);
 	dev_priv->gt.awake = false;

drivers/gpu/drm/i915/i915_gem_request.c

@@ -258,6 +258,7 @@ static void mark_busy(struct drm_i915_private *i915)
 	i915_update_gfx_val(i915);
 	if (INTEL_GEN(i915) >= 6)
 		gen6_rps_busy(i915);
+	i915_pmu_gt_unparked(i915);
 
 	intel_engines_unpark(i915);
 

drivers/gpu/drm/i915/i915_pmu.c

@@ -90,6 +90,75 @@ static unsigned int event_enabled_bit(struct perf_event *event)
 	return config_enabled_bit(event->attr.config);
 }
 
+static bool pmu_needs_timer(struct drm_i915_private *i915, bool gpu_active)
+{
+	u64 enable;
+
+	/*
+	 * Only some counters need the sampling timer.
+	 *
+	 * We start with a bitmask of all currently enabled events.
+	 */
+	enable = i915->pmu.enable;
+
+	/*
+	 * Mask out all the ones which do not need the timer, or in
+	 * other words keep all the ones that could need the timer.
+	 */
+	enable &= config_enabled_mask(I915_PMU_ACTUAL_FREQUENCY) |
+		  config_enabled_mask(I915_PMU_REQUESTED_FREQUENCY) |
+		  ENGINE_SAMPLE_MASK;
+
+	/*
+	 * When the GPU is idle per-engine counters do not need to be
+	 * running so clear those bits out.
+	 */
+	if (!gpu_active)
+		enable &= ~ENGINE_SAMPLE_MASK;
+
+	/*
+	 * If some bits remain it means we need the sampling timer running.
+	 */
+	return enable;
+}
+
+void i915_pmu_gt_parked(struct drm_i915_private *i915)
+{
+	if (!i915->pmu.base.event_init)
+		return;
+
+	spin_lock_irq(&i915->pmu.lock);
+	/*
+	 * Signal sampling timer to stop if only engine events are enabled and
+	 * GPU went idle.
+	 */
+	i915->pmu.timer_enabled = pmu_needs_timer(i915, false);
+	spin_unlock_irq(&i915->pmu.lock);
+}
+
+static void __i915_pmu_maybe_start_timer(struct drm_i915_private *i915)
+{
+	if (!i915->pmu.timer_enabled && pmu_needs_timer(i915, true)) {
+		i915->pmu.timer_enabled = true;
+		hrtimer_start_range_ns(&i915->pmu.timer,
+				       ns_to_ktime(PERIOD), 0,
+				       HRTIMER_MODE_REL_PINNED);
+	}
+}
+
+void i915_pmu_gt_unparked(struct drm_i915_private *i915)
+{
+	if (!i915->pmu.base.event_init)
+		return;
+
+	spin_lock_irq(&i915->pmu.lock);
+	/*
+	 * Re-enable sampling timer when GPU goes active.
+	 */
+	__i915_pmu_maybe_start_timer(i915);
+	spin_unlock_irq(&i915->pmu.lock);
+}
+
 static bool grab_forcewake(struct drm_i915_private *i915, bool fw)
 {
 	if (!fw)
@@ -187,7 +256,7 @@ static enum hrtimer_restart i915_sample(struct hrtimer *hrtimer)
 	struct drm_i915_private *i915 =
 		container_of(hrtimer, struct drm_i915_private, pmu.timer);
 
-	if (i915->pmu.enable == 0)
+	if (!READ_ONCE(i915->pmu.timer_enabled))
 		return HRTIMER_NORESTART;
 
 	engines_sample(i915);
@@ -339,14 +408,6 @@ static void i915_pmu_enable(struct perf_event *event)
 
 	spin_lock_irqsave(&i915->pmu.lock, flags);
 
-	/*
-	 * Start the sampling timer when enabling the first event.
-	 */
-	if (i915->pmu.enable == 0)
-		hrtimer_start_range_ns(&i915->pmu.timer,
-				       ns_to_ktime(PERIOD), 0,
-				       HRTIMER_MODE_REL_PINNED);
-
 	/*
 	 * Update the bitmask of enabled events and increment
 	 * the event reference counter.
@@ -356,6 +417,11 @@ static void i915_pmu_enable(struct perf_event *event)
 	i915->pmu.enable |= BIT_ULL(bit);
 	i915->pmu.enable_count[bit]++;
 
+	/*
+	 * Start the sampling timer if needed and not already enabled.
+	 */
+	__i915_pmu_maybe_start_timer(i915);
+
 	/*
 	 * For per-engine events the bitmask and reference counting
 	 * is stored per engine.
@@ -418,8 +484,10 @@ static void i915_pmu_disable(struct perf_event *event)
 	 * Decrement the reference count and clear the enabled
 	 * bitmask when the last listener on an event goes away.
 	 */
-	if (--i915->pmu.enable_count[bit] == 0)
+	if (--i915->pmu.enable_count[bit] == 0) {
 		i915->pmu.enable &= ~BIT_ULL(bit);
+		i915->pmu.timer_enabled &= pmu_needs_timer(i915, true);
+	}
 
 	spin_unlock_irqrestore(&i915->pmu.lock, flags);
 }

drivers/gpu/drm/i915/i915_pmu.h

@@ -82,6 +82,10 @@ struct i915_pmu {
 	 * are using the PMU API.
 	 */
 	unsigned int enable_count[I915_PMU_MASK_BITS];
+	/**
+	 * @timer_enabled: Should the internal sampling timer be running.
+	 */
+	bool timer_enabled;
 	/**
 	 * @sample: Current and previous (raw) counters for sampling events.
 	 *
@@ -96,9 +100,13 @@ struct i915_pmu {
 #ifdef CONFIG_PERF_EVENTS
 void i915_pmu_register(struct drm_i915_private *i915);
 void i915_pmu_unregister(struct drm_i915_private *i915);
+void i915_pmu_gt_parked(struct drm_i915_private *i915);
+void i915_pmu_gt_unparked(struct drm_i915_private *i915);
 #else
 static inline void i915_pmu_register(struct drm_i915_private *i915) {}
 static inline void i915_pmu_unregister(struct drm_i915_private *i915) {}
+static inline void i915_pmu_gt_parked(struct drm_i915_private *i915) {}
+static inline void i915_pmu_gt_unparked(struct drm_i915_private *i915) {}
 #endif
 
 #endif