Skip to contentSkip to conversation

patches that failed to cherry-pick on drm-intel-fixes for 4.16-rc1

PATCH Applied Rodrigo Vivi 5 participants 15 messages Feb 13, 2018 – Feb 14, 2018

Subsystems (3)
lore
Share & tools
lore
#1 lore raw

PATCH 1/1 drm/i915/breadcrumbs: Ignore unsubmitted signalers

When a request is preempted, it is unsubmitted from the HW queue and removed from the active list of breadcrumbs. In the process, this however triggers the signaler and it may see the clear rbtree with the old, and still valid, seqno, or it may match the cleared seqno with the now zero rq->global_seqno. This confuses the signaler into action and signaling the fence.

Fixes d6a2289d9d6b ("drm/i915: Remove the preempted request from the execution queue")
Cc Tvrtko Ursulin <[email protected]>
Cc Joonas Lahtinen <[email protected]>
Cc <[email protected]> # v4.12+
(cherry picked from commit fd10e2ce9905030d922e179a8047a4d50daffd8e)
---
 drivers/gpu/drm/i915/intel_breadcrumbs.c | 29 ++++++++++-------------------
 1 file changed, 10 insertions(+), 19 deletions(-)
drivers/gpu/drm/i915/intel_breadcrumbs.c +10 -19
index bd40fea16b4f..f54ddda9fdad 100644
@@ -594,29 +594,16 @@ void intel_engine_remove_wait(struct intel_engine_cs *engine,
 	spin_unlock_irq(&b->rb_lock);
 }
 
-static bool signal_valid(const struct drm_i915_gem_request *request)
-{
-	return intel_wait_check_request(&request->signaling.wait, request);
-}
-
 static bool signal_complete(const struct drm_i915_gem_request *request)
 {
 	if (!request)
 		return false;
 
-	/* If another process served as the bottom-half it may have already
-	 * signalled that this wait is already completed.
-	 */
-	if (intel_wait_complete(&request->signaling.wait))
-		return signal_valid(request);
-
-	/* Carefully check if the request is complete, giving time for the
+	/*
+	 * Carefully check if the request is complete, giving time for the
 	 * seqno to be visible or if the GPU hung.
 	 */
-	if (__i915_request_irq_complete(request))
-		return true;
-
-	return false;
+	return __i915_request_irq_complete(request);
 }
 
 static struct drm_i915_gem_request *to_signaler(struct rb_node *rb)
@@ -659,9 +646,13 @@ static int intel_breadcrumbs_signaler(void *arg)
 			request = i915_gem_request_get_rcu(request);
 		rcu_read_unlock();
 		if (signal_complete(request)) {
-			local_bh_disable();
-			dma_fence_signal(&request->fence);
-			local_bh_enable(); /* kick start the tasklets */
+			if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
+				      &request->fence.flags)) {
+				local_bh_disable();
+				dma_fence_signal(&request->fence);
+				GEM_BUG_ON(!i915_gem_request_completed(request));
+				local_bh_enable(); /* kick start the tasklets */
+			}
 
 			spin_lock_irq(&b->rb_lock);
 
Signature · 1 line
2.16.1
author maintainer Applied
#2 lore raw

Re: [PATCH] drm/i915/breadcrumbs: Ignore unsubmitted signalers

Quoted 7 lines · Chris Wilson

When a request is preempted, it is unsubmitted from the HW queue and removed from the active list of breadcrumbs. In the process, this however triggers the signaler and it may see the clear rbtree with the old, and still valid, seqno, or it may match the cleared seqno with the now zero rq->global_seqno. This confuses the signaler into action and signaling the fence.

Quoted 8 lines
Fixes d6a2289d9d6b ("drm/i915: Remove the preempted request from the execution queue")
Cc Tvrtko Ursulin <[email protected]>
Cc Joonas Lahtinen <[email protected]>
Cc <[email protected]> # v4.12+
(cherry picked from commit fd10e2ce9905030d922e179a8047a4d50daffd8e)

applied to fixes. Thanks

Quoted 61 lines
---
 drivers/gpu/drm/i915/intel_breadcrumbs.c | 29 ++++++++++-------------------
 1 file changed, 10 insertions(+), 19 deletions(-)
drivers/gpu/drm/i915/intel_breadcrumbs.c +10 -19
index bd40fea16b4f..f54ddda9fdad 100644
@@ -594,29 +594,16 @@ void intel_engine_remove_wait(struct intel_engine_cs *engine,
 	spin_unlock_irq(&b->rb_lock);
 }
 
-static bool signal_valid(const struct drm_i915_gem_request *request)
-{
-	return intel_wait_check_request(&request->signaling.wait, request);
-}
-
 static bool signal_complete(const struct drm_i915_gem_request *request)
 {
 	if (!request)
 		return false;
 
-	/* If another process served as the bottom-half it may have already
-	 * signalled that this wait is already completed.
-	 */
-	if (intel_wait_complete(&request->signaling.wait))
-		return signal_valid(request);
-
-	/* Carefully check if the request is complete, giving time for the
+	/*
+	 * Carefully check if the request is complete, giving time for the
 	 * seqno to be visible or if the GPU hung.
 	 */
-	if (__i915_request_irq_complete(request))
-		return true;
-
-	return false;
+	return __i915_request_irq_complete(request);
 }
 
 static struct drm_i915_gem_request *to_signaler(struct rb_node *rb)
@@ -659,9 +646,13 @@ static int intel_breadcrumbs_signaler(void *arg)
 			request = i915_gem_request_get_rcu(request);
 		rcu_read_unlock();
 		if (signal_complete(request)) {
-			local_bh_disable();
-			dma_fence_signal(&request->fence);
-			local_bh_enable(); /* kick start the tasklets */
+			if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
+				      &request->fence.flags)) {
+				local_bh_disable();
+				dma_fence_signal(&request->fence);
+				GEM_BUG_ON(!i915_gem_request_completed(request));
+				local_bh_enable(); /* kick start the tasklets */
+			}
 
 			spin_lock_irq(&b->rb_lock);
 
Signature · 1 line
2.16.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

maintainer
#3 lore raw
In reply to Rodrigo Vivi
Quoted 7 lines · Rodrigo Vivi

These patches failed to cherry-pick on drm-intel-fixes

Please provide a backport or advise how to proceed.

fd10e2ce9905 ("drm/i915/breadcrumbs: Ignore unsubmitted signalers") 1fe699e30113 ("drm/i915/pmu: Fix sleep under atomic in RC6 readout")

I'm building at the moment in order to test and then I'll send some patches. Plural because, since drm-intel-fixes includes:

commit 4900727d35bb20028f9bd83146ec4bf78afffe30

Author: Chris Wilson <[email protected]>
Date:   Thu Jan 11 07:30:31 2018 +0000

drm/i915/pmu: Reconstruct active state on starting busy-stats

We need to include a chain of two more fixups:

commit 99e48bf98dd036090b480a12c39e8b971731247e

Author: Chris Wilson <[email protected]>
Date:   Mon Jan 15 09:20:41 2018 +0000

drm/i915: Lock out execlist tasklet while peeking inside for busy-stats

commit b2f78cda260bc6a1a2d382b1d85a29e69b5b3724

Author: Tvrtko Ursulin <[email protected]>
Date:   Mon Feb 5 09:34:48 2018 +0000

drm/i915/pmu: Fix PMU enable vs execlists tasklet race

And then the referenced fix, plus another fixup:

commit 1fe699e30113ed6f6e853ff44710d256072ea627

Author: Tvrtko Ursulin <[email protected]>
Date:   Tue Feb 6 18:33:11 2018 +0000

drm/i915/pmu: Fix sleep under atomic in RC6 readout

commit 05273c950a3c93c5f96be8807eaf24f2cc9f1c1e

Author: Chris Wilson <[email protected]>
Date:   Wed Feb 7 16:04:28 2018 +0000

drm/i915/pmu: Fix building without CONFIG_PM

I'll send them when I test my rebases.

Regards,

Tvrtko
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

maintainer
#4 lore raw
Quoted 44 lines · Tvrtko Ursulin
Quoted 7 lines · Rodrigo Vivi

These patches failed to cherry-pick on drm-intel-fixes

Please provide a backport or advise how to proceed.

fd10e2ce9905 ("drm/i915/breadcrumbs: Ignore unsubmitted signalers") 1fe699e30113 ("drm/i915/pmu: Fix sleep under atomic in RC6 readout")

I'm building at the moment in order to test and then I'll send some patches. Plural because, since drm-intel-fixes includes:

commit 4900727d35bb20028f9bd83146ec4bf78afffe30

Author: Chris Wilson <[email protected]>
Date:   Thu Jan 11 07:30:31 2018 +0000

drm/i915/pmu: Reconstruct active state on starting busy-stats

We need to include a chain of two more fixups:

commit 99e48bf98dd036090b480a12c39e8b971731247e

Author: Chris Wilson <[email protected]>
Date:   Mon Jan 15 09:20:41 2018 +0000

drm/i915: Lock out execlist tasklet while peeking inside for busy-stats

commit b2f78cda260bc6a1a2d382b1d85a29e69b5b3724

Author: Tvrtko Ursulin <[email protected]>
Date:   Mon Feb 5 09:34:48 2018 +0000

drm/i915/pmu: Fix PMU enable vs execlists tasklet race

And then the referenced fix, plus another fixup:

commit 1fe699e30113ed6f6e853ff44710d256072ea627

Author: Tvrtko Ursulin <[email protected]>
Date:   Tue Feb 6 18:33:11 2018 +0000

drm/i915/pmu: Fix sleep under atomic in RC6 readout

Quoted 8 lines

commit 05273c950a3c93c5f96be8807eaf24f2cc9f1c1e

Author: Chris Wilson <[email protected]>
Date:   Wed Feb 7 16:04:28 2018 +0000

drm/i915/pmu: Fix building without CONFIG_PM

I'll send them when I test my rebases.

Please add the original dinq commit references in the format

(cherry picked from commit fd10e2ce9905030d922e179a8047a4d50daffd8e)

to the backports.

Thanks,
Jani.

Signature · 5 lines
Jani Nikula, Intel Open Source Technology Center
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer
#5 lore raw

PATCH 1/4 drm/i915: Lock out execlist tasklet while peeking inside for busy-stats

From: Chris Wilson <[email protected]>

In order to prevent a race condition where we may end up overaccounting the active state and leaving the busy-stats believing the GPU is 100% busy, lock out the tasklet while we reconstruct the busy state. There is no direct spinlock guard for the execlists->port[], so we need to utilise tasklet_disable() as a synchronous barrier to prevent it, the only writer to execlists->port[], from running at the same time as the enable.

Fixes 4900727d35bb ("drm/i915/pmu: Reconstruct active state on starting busy-stats")
Cc Tvrtko Ursulin <[email protected]>
(cherry picked from commit 99e48bf98dd036090b480a12c39e8b971731247e)
---
 drivers/gpu/drm/i915/intel_engine_cs.c | 20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)
drivers/gpu/drm/i915/intel_engine_cs.c +12 -8
index acc661aa9c0c..fa960cfd2764 100644
@@ -1945,16 +1945,22 @@ intel_engine_lookup_user(struct drm_i915_private *i915, u8 class, u8 instance)
  */
 int intel_enable_engine_stats(struct intel_engine_cs *engine)
 {
+	struct intel_engine_execlists *execlists = &engine->execlists;
 	unsigned long flags;
+	int err = 0;
 
 	if (!intel_engine_supports_stats(engine))
 		return -ENODEV;
 
+	tasklet_disable(&execlists->tasklet);
 	spin_lock_irqsave(&engine->stats.lock, flags);
-	if (engine->stats.enabled == ~0)
-		goto busy;
+
+	if (unlikely(engine->stats.enabled == ~0)) {
+		err = -EBUSY;
+		goto unlock;
+	}
+
 	if (engine->stats.enabled++ == 0) {
-		struct intel_engine_execlists *execlists = &engine->execlists;
 		const struct execlist_port *port = execlists->port;
 		unsigned int num_ports = execlists_num_ports(execlists);
 
@@ -1969,14 +1975,12 @@ int intel_enable_engine_stats(struct intel_engine_cs *engine)
 		if (engine->stats.active)
 			engine->stats.start = engine->stats.enabled_at;
 	}
-	spin_unlock_irqrestore(&engine->stats.lock, flags);
 
-	return 0;
-
-busy:
+unlock:
 	spin_unlock_irqrestore(&engine->stats.lock, flags);
+	tasklet_enable(&execlists->tasklet);
 
-	return -EBUSY;
+	return err;
 }
 
 static ktime_t __intel_engine_get_busy_time(struct intel_engine_cs *engine)
Signature · 6 lines
2.14.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer
#6 lore raw

PATCH 2/4 drm/i915/pmu: Fix PMU enable vs execlists tasklet race

From: Tvrtko Ursulin <[email protected]>

Commit 99e48bf98dd0 ("drm/i915: Lock out execlist tasklet while peeking inside for busy-stats") added a tasklet_disable call in busy stats enabling, but we failed to understand that the PMU enable callback runs as an hard IRQ (IPI).

Consequence of this is that the PMU enable callback can interrupt the execlists tasklet, and will then deadlock when it calls intel_engine_stats_enable->tasklet_disable.

To fix this, I realized it is possible to move the engine stats enablement and disablement to PMU event init and destroy hooks. This allows for much simpler implementation since those hooks run in normal context (can sleep).

v2: Extract engine_event_destroy. (Chris Wilson)

Fixes 99e48bf98dd0 ("drm/i915: Lock out execlist tasklet while peeking inside for busy-stats")

Testcase: igt/perf_pmu/enable-race-*

Cc Chris Wilson <[email protected]>
Cc Tvrtko Ursulin <[email protected]>
Cc Jani Nikula <[email protected]>
Cc Joonas Lahtinen <[email protected]>
Cc Rodrigo Vivi <[email protected]>
(cherry picked from commit b2f78cda260bc6a1a2d382b1d85a29e69b5b3724)
---
 drivers/gpu/drm/i915/i915_pmu.c         | 125 +++++++++++++-------------------
 drivers/gpu/drm/i915/intel_ringbuffer.h |  14 ----
 2 files changed, 52 insertions(+), 87 deletions(-)
drivers/gpu/drm/i915/i915_pmu.c +52 -73
index 55a8a1e29424..337eaa6ede52 100644
@@ -285,26 +285,41 @@ static u64 count_interrupts(struct drm_i915_private *i915)
 	return sum;
 }
 
-static void i915_pmu_event_destroy(struct perf_event *event)
+static void engine_event_destroy(struct perf_event *event)
 {
-	WARN_ON(event->parent);
+	struct drm_i915_private *i915 =
+		container_of(event->pmu, typeof(*i915), pmu.base);
+	struct intel_engine_cs *engine;
+
+	engine = intel_engine_lookup_user(i915,
+					  engine_event_class(event),
+					  engine_event_instance(event));
+	if (WARN_ON_ONCE(!engine))
+		return;
+
+	if (engine_event_sample(event) == I915_SAMPLE_BUSY &&
+	    intel_engine_supports_stats(engine))
+		intel_disable_engine_stats(engine);
 }
 
-static int engine_event_init(struct perf_event *event)
+static void i915_pmu_event_destroy(struct perf_event *event)
 {
-	struct drm_i915_private *i915 =
-		container_of(event->pmu, typeof(*i915), pmu.base);
+	WARN_ON(event->parent);
 
-	if (!intel_engine_lookup_user(i915, engine_event_class(event),
-				      engine_event_instance(event)))
-		return -ENODEV;
+	if (is_engine_event(event))
+		engine_event_destroy(event);
+}
 
-	switch (engine_event_sample(event)) {
+static int
+engine_event_status(struct intel_engine_cs *engine,
+		    enum drm_i915_pmu_engine_sample sample)
+{
+	switch (sample) {
 	case I915_SAMPLE_BUSY:
 	case I915_SAMPLE_WAIT:
 		break;
 	case I915_SAMPLE_SEMA:
-		if (INTEL_GEN(i915) < 6)
+		if (INTEL_GEN(engine->i915) < 6)
 			return -ENODEV;
 		break;
 	default:
@@ -314,6 +329,30 @@ static int engine_event_init(struct perf_event *event)
 	return 0;
 }
 
+static int engine_event_init(struct perf_event *event)
+{
+	struct drm_i915_private *i915 =
+		container_of(event->pmu, typeof(*i915), pmu.base);
+	struct intel_engine_cs *engine;
+	u8 sample;
+	int ret;
+
+	engine = intel_engine_lookup_user(i915, engine_event_class(event),
+					  engine_event_instance(event));
+	if (!engine)
+		return -ENODEV;
+
+	sample = engine_event_sample(event);
+	ret = engine_event_status(engine, sample);
+	if (ret)
+		return ret;
+
+	if (sample == I915_SAMPLE_BUSY && intel_engine_supports_stats(engine))
+		ret = intel_enable_engine_stats(engine);
+
+	return ret;
+}
+
 static int i915_pmu_event_init(struct perf_event *event)
 {
 	struct drm_i915_private *i915 =
@@ -387,7 +426,7 @@ static u64 __i915_pmu_event_read(struct perf_event *event)
 		if (WARN_ON_ONCE(!engine)) {
 			/* Do nothing */
 		} else if (sample == I915_SAMPLE_BUSY &&
-			   engine->pmu.busy_stats) {
+			   intel_engine_supports_stats(engine)) {
 			val = ktime_to_ns(intel_engine_get_busy_time(engine));
 		} else {
 			val = engine->pmu.sample[sample].cur;
@@ -442,12 +481,6 @@ static void i915_pmu_event_read(struct perf_event *event)
 	local64_add(new - prev, &event->count);
 }
 
-static bool engine_needs_busy_stats(struct intel_engine_cs *engine)
-{
-	return intel_engine_supports_stats(engine) &&
-	       (engine->pmu.enable & BIT(I915_SAMPLE_BUSY));
-}
-
 static void i915_pmu_enable(struct perf_event *event)
 {
 	struct drm_i915_private *i915 =
@@ -487,21 +520,7 @@ static void i915_pmu_enable(struct perf_event *event)
 
 		GEM_BUG_ON(sample >= I915_PMU_SAMPLE_BITS);
 		GEM_BUG_ON(engine->pmu.enable_count[sample] == ~0);
-		if (engine->pmu.enable_count[sample]++ == 0) {
-			/*
-			 * Enable engine busy stats tracking if needed or
-			 * alternatively cancel the scheduled disable.
-			 *
-			 * If the delayed disable was pending, cancel it and
-			 * in this case do not enable since it already is.
-			 */
-			if (engine_needs_busy_stats(engine) &&
-			    !engine->pmu.busy_stats) {
-				engine->pmu.busy_stats = true;
-				if (!cancel_delayed_work(&engine->pmu.disable_busy_stats))
-					intel_enable_engine_stats(engine);
-			}
-		}
+		engine->pmu.enable_count[sample]++;
 	}
 
 	/*
@@ -514,14 +533,6 @@ static void i915_pmu_enable(struct perf_event *event)
 	spin_unlock_irqrestore(&i915->pmu.lock, flags);
 }
 
-static void __disable_busy_stats(struct work_struct *work)
-{
-	struct intel_engine_cs *engine =
-	       container_of(work, typeof(*engine), pmu.disable_busy_stats.work);
-
-	intel_disable_engine_stats(engine);
-}
-
 static void i915_pmu_disable(struct perf_event *event)
 {
 	struct drm_i915_private *i915 =
@@ -545,26 +556,8 @@ 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 (--engine->pmu.enable_count[sample] == 0) {
+		if (--engine->pmu.enable_count[sample] == 0)
 			engine->pmu.enable &= ~BIT(sample);
-			if (!engine_needs_busy_stats(engine) &&
-			    engine->pmu.busy_stats) {
-				engine->pmu.busy_stats = false;
-				/*
-				 * We request a delayed disable to handle the
-				 * rapid on/off cycles on events, which can
-				 * happen when tools like perf stat start, in a
-				 * nicer way.
-				 *
-				 * In addition, this also helps with busy stats
-				 * accuracy with background CPU offline/online
-				 * migration events.
-				 */
-				queue_delayed_work(system_wq,
-						   &engine->pmu.disable_busy_stats,
-						   round_jiffies_up_relative(HZ));
-			}
-		}
 	}
 
 	GEM_BUG_ON(bit >= I915_PMU_MASK_BITS);
@@ -797,8 +790,6 @@ static void i915_pmu_unregister_cpuhp_state(struct drm_i915_private *i915)
 
 void i915_pmu_register(struct drm_i915_private *i915)
 {
-	struct intel_engine_cs *engine;
-	enum intel_engine_id id;
 	int ret;
 
 	if (INTEL_GEN(i915) <= 2) {
@@ -820,10 +811,6 @@ void i915_pmu_register(struct drm_i915_private *i915)
 	hrtimer_init(&i915->pmu.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 	i915->pmu.timer.function = i915_sample;
 
-	for_each_engine(engine, i915, id)
-		INIT_DELAYED_WORK(&engine->pmu.disable_busy_stats,
-				  __disable_busy_stats);
-
 	ret = perf_pmu_register(&i915->pmu.base, "i915", -1);
 	if (ret)
 		goto err;
@@ -843,9 +830,6 @@ void i915_pmu_register(struct drm_i915_private *i915)
 
 void i915_pmu_unregister(struct drm_i915_private *i915)
 {
-	struct intel_engine_cs *engine;
-	enum intel_engine_id id;
-
 	if (!i915->pmu.base.event_init)
 		return;
 
@@ -853,11 +837,6 @@ void i915_pmu_unregister(struct drm_i915_private *i915)
 
 	hrtimer_cancel(&i915->pmu.timer);
 
-	for_each_engine(engine, i915, id) {
-		GEM_BUG_ON(engine->pmu.busy_stats);
-		flush_delayed_work(&engine->pmu.disable_busy_stats);
-	}
-
 	i915_pmu_unregister_cpuhp_state(i915);
 
 	perf_pmu_unregister(&i915->pmu.base);
drivers/gpu/drm/i915/intel_ringbuffer.h +0 -14
index c5ff203e42d6..a0e7a6c2a57c 100644
@@ -366,20 +366,6 @@ struct intel_engine_cs {
 		 */
 #define I915_ENGINE_SAMPLE_MAX (I915_SAMPLE_SEMA + 1)
 		struct i915_pmu_sample sample[I915_ENGINE_SAMPLE_MAX];
-		/**
-		 * @busy_stats: Has enablement of engine stats tracking been
-		 * 		requested.
-		 */
-		bool busy_stats;
-		/**
-		 * @disable_busy_stats: Work item for busy stats disabling.
-		 *
-		 * Same as with @enable_busy_stats action, with the difference
-		 * that we delay it in case there are rapid enable-disable
-		 * actions, which can happen during tool startup (like perf
-		 * stat).
-		 */
-		struct delayed_work disable_busy_stats;
 	} pmu;
 
 	/*
Signature · 6 lines
2.14.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer
#7 lore raw

PATCH 3/4 drm/i915/pmu: Fix sleep under atomic in RC6 readout

In reply to Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

We are not allowed to call intel_runtime_pm_get from the PMU counter read callback since the former can sleep, and the latter is running under IRQ context.

To workaround this, we record the last known RC6 and while runtime suspended estimate its increase by querying the runtime PM core timestamps.

Downside of this approach is that we can temporarily lose a chunk of RC6 time, from the last PMU read-out to runtime suspend entry, but that will eventually catch up, once device comes back online and in the presence of PMU queries.

Also, we have to be careful not to overshoot the RC6 estimate, so once resumed after a period of approximation, we only update the counter once it catches up. With the observation that RC6 is increasing while the device is suspended, this should not pose a problem and can only cause slight inaccuracies due clock base differences.

v2: Simplify by estimating on top of PM core counters. (Imre)

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=104943

Fixes 6060b6aec03c ("drm/i915/pmu: Add RC6 residency metrics")

Testcase: igt/perf_pmu/rc6-runtime-pm

Cc Tvrtko Ursulin <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Imre Deak <[email protected]>
Cc Jani Nikula <[email protected]>
Cc Joonas Lahtinen <[email protected]>
Cc Rodrigo Vivi <[email protected]>
Cc David Airlie <[email protected]>
(cherry picked from commit 1fe699e30113ed6f6e853ff44710d256072ea627)
---
 drivers/gpu/drm/i915/i915_pmu.c | 93 ++++++++++++++++++++++++++++++++++-------
 drivers/gpu/drm/i915/i915_pmu.h |  6 +++
 2 files changed, 84 insertions(+), 15 deletions(-)
drivers/gpu/drm/i915/i915_pmu.c +78 -15
index 337eaa6ede52..e13859aaa2a3 100644
@@ -409,7 +409,81 @@ static int i915_pmu_event_init(struct perf_event *event)
 	return 0;
 }
 
-static u64 __i915_pmu_event_read(struct perf_event *event)
+static u64 get_rc6(struct drm_i915_private *i915, bool locked)
+{
+	unsigned long flags;
+	u64 val;
+
+	if (intel_runtime_pm_get_if_in_use(i915)) {
+		val = intel_rc6_residency_ns(i915, IS_VALLEYVIEW(i915) ?
+						   VLV_GT_RENDER_RC6 :
+						   GEN6_GT_GFX_RC6);
+
+		if (HAS_RC6p(i915))
+			val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
+
+		if (HAS_RC6pp(i915))
+			val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
+
+		intel_runtime_pm_put(i915);
+
+		/*
+		 * If we are coming back from being runtime suspended we must
+		 * be careful not to report a larger value than returned
+		 * previously.
+		 */
+
+		if (!locked)
+			spin_lock_irqsave(&i915->pmu.lock, flags);
+
+		if (val >= i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur) {
+			i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur = 0;
+			i915->pmu.sample[__I915_SAMPLE_RC6].cur = val;
+		} else {
+			val = i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur;
+		}
+
+		if (!locked)
+			spin_unlock_irqrestore(&i915->pmu.lock, flags);
+	} else {
+		struct pci_dev *pdev = i915->drm.pdev;
+		struct device *kdev = &pdev->dev;
+		unsigned long flags2;
+
+		/*
+		 * We are runtime suspended.
+		 *
+		 * Report the delta from when the device was suspended to now,
+		 * on top of the last known real value, as the approximated RC6
+		 * counter value.
+		 */
+		if (!locked)
+			spin_lock_irqsave(&i915->pmu.lock, flags);
+
+		spin_lock_irqsave(&kdev->power.lock, flags2);
+
+		if (!i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur)
+			i915->pmu.suspended_jiffies_last =
+						kdev->power.suspended_jiffies;
+
+		val = kdev->power.suspended_jiffies -
+		      i915->pmu.suspended_jiffies_last;
+		val += jiffies - kdev->power.accounting_timestamp;
+
+		spin_unlock_irqrestore(&kdev->power.lock, flags2);
+
+		val = jiffies_to_nsecs(val);
+		val += i915->pmu.sample[__I915_SAMPLE_RC6].cur;
+		i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur = val;
+
+		if (!locked)
+			spin_unlock_irqrestore(&i915->pmu.lock, flags);
+	}
+
+	return val;
+}
+
+static u64 __i915_pmu_event_read(struct perf_event *event, bool locked)
 {
 	struct drm_i915_private *i915 =
 		container_of(event->pmu, typeof(*i915), pmu.base);
@@ -447,18 +521,7 @@ static u64 __i915_pmu_event_read(struct perf_event *event)
 			val = count_interrupts(i915);
 			break;
 		case I915_PMU_RC6_RESIDENCY:
-			intel_runtime_pm_get(i915);
-			val = intel_rc6_residency_ns(i915,
-						     IS_VALLEYVIEW(i915) ?
-						     VLV_GT_RENDER_RC6 :
-						     GEN6_GT_GFX_RC6);
-			if (HAS_RC6p(i915))
-				val += intel_rc6_residency_ns(i915,
-							      GEN6_GT_GFX_RC6p);
-			if (HAS_RC6pp(i915))
-				val += intel_rc6_residency_ns(i915,
-							      GEN6_GT_GFX_RC6pp);
-			intel_runtime_pm_put(i915);
+			val = get_rc6(i915, locked);
 			break;
 		}
 	}
@@ -473,7 +536,7 @@ static void i915_pmu_event_read(struct perf_event *event)
 
 again:
 	prev = local64_read(&hwc->prev_count);
-	new = __i915_pmu_event_read(event);
+	new = __i915_pmu_event_read(event, false);
 
 	if (local64_cmpxchg(&hwc->prev_count, prev, new) != prev)
 		goto again;
@@ -528,7 +591,7 @@ static void i915_pmu_enable(struct perf_event *event)
 	 * for all listeners. Even when the event was already enabled and has
 	 * an existing non-zero value.
 	 */
-	local64_set(&event->hw.prev_count, __i915_pmu_event_read(event));
+	local64_set(&event->hw.prev_count, __i915_pmu_event_read(event, true));
 
 	spin_unlock_irqrestore(&i915->pmu.lock, flags);
 }
drivers/gpu/drm/i915/i915_pmu.h +6 -0
index 40c154d13565..bb62df15afa4 100644
@@ -27,6 +27,8 @@
 enum {
 	__I915_SAMPLE_FREQ_ACT = 0,
 	__I915_SAMPLE_FREQ_REQ,
+	__I915_SAMPLE_RC6,
+	__I915_SAMPLE_RC6_ESTIMATED,
 	__I915_NUM_PMU_SAMPLERS
 };
 
@@ -94,6 +96,10 @@ struct i915_pmu {
 	 * struct intel_engine_cs.
 	 */
 	struct i915_pmu_sample sample[__I915_NUM_PMU_SAMPLERS];
+	/**
+	 * @suspended_jiffies_last: Cached suspend time from PM core.
+	 */
+	unsigned long suspended_jiffies_last;
 };
 
 #ifdef CONFIG_PERF_EVENTS
Signature · 6 lines
2.14.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer
#8 lore raw

PATCH 4/4 drm/i915/pmu: Fix building without CONFIG_PM

In reply to Tvrtko Ursulin

From: Chris Wilson <[email protected]>

As we peek inside struct device to query members guarded by CONFIG_PM, so must be the code.

Fixes 1fe699e30113 ("drm/i915/pmu: Fix sleep under atomic in RC6 readout")
Cc Tvrtko Ursulin <[email protected]>
(cherry picked from commit 05273c950a3c93c5f96be8807eaf24f2cc9f1c1e)
---
 drivers/gpu/drm/i915/i915_pmu.c | 33 +++++++++++++++++++++++----------
 1 file changed, 23 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_pmu.c +23 -10
index e13859aaa2a3..0e9b98c32b62 100644
@@ -409,22 +409,32 @@ static int i915_pmu_event_init(struct perf_event *event)
 	return 0;
 }
 
-static u64 get_rc6(struct drm_i915_private *i915, bool locked)
+static u64 __get_rc6(struct drm_i915_private *i915)
 {
-	unsigned long flags;
 	u64 val;
 
-	if (intel_runtime_pm_get_if_in_use(i915)) {
-		val = intel_rc6_residency_ns(i915, IS_VALLEYVIEW(i915) ?
-						   VLV_GT_RENDER_RC6 :
-						   GEN6_GT_GFX_RC6);
+	val = intel_rc6_residency_ns(i915,
+				     IS_VALLEYVIEW(i915) ?
+				     VLV_GT_RENDER_RC6 :
+				     GEN6_GT_GFX_RC6);
 
-		if (HAS_RC6p(i915))
-			val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
+	if (HAS_RC6p(i915))
+		val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
+
+	if (HAS_RC6pp(i915))
+		val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
+
+	return val;
+}
 
-		if (HAS_RC6pp(i915))
-			val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
+static u64 get_rc6(struct drm_i915_private *i915, bool locked)
+{
+#if IS_ENABLED(CONFIG_PM)
+	unsigned long flags;
+	u64 val;
 
+	if (intel_runtime_pm_get_if_in_use(i915)) {
+		val = __get_rc6(i915);
 		intel_runtime_pm_put(i915);
 
 		/*
@@ -481,6 +491,9 @@ static u64 get_rc6(struct drm_i915_private *i915, bool locked)
 	}
 
 	return val;
+#else
+	return __get_rc6(i915);
+#endif
 }
 
 static u64 __i915_pmu_event_read(struct perf_event *event, bool locked)
Signature · 6 lines
2.14.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
author maintainer
#9 lore raw

Re: [PATCH 4/4] drm/i915/pmu: Fix building without CONFIG_PM

Quoted 5 lines · Tvrtko Ursulin

From: Chris Wilson <[email protected]>

As we peek inside struct device to query members guarded by CONFIG_PM, so must be the code.

Quoted 7 lines
Fixes 1fe699e30113 ("drm/i915/pmu: Fix sleep under atomic in RC6 readout")
Cc Tvrtko Ursulin <[email protected]>
(cherry picked from commit 05273c950a3c93c5f96be8807eaf24f2cc9f1c1e)

All 4 patches applied to fixes.
Thanks for backporting and for adding cherry-picked message.

Quoted 68 lines
---
 drivers/gpu/drm/i915/i915_pmu.c | 33 +++++++++++++++++++++++----------
 1 file changed, 23 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_pmu.c +23 -10
index e13859aaa2a3..0e9b98c32b62 100644
@@ -409,22 +409,32 @@ static int i915_pmu_event_init(struct perf_event *event)
 	return 0;
 }
 
-static u64 get_rc6(struct drm_i915_private *i915, bool locked)
+static u64 __get_rc6(struct drm_i915_private *i915)
 {
-	unsigned long flags;
 	u64 val;
 
-	if (intel_runtime_pm_get_if_in_use(i915)) {
-		val = intel_rc6_residency_ns(i915, IS_VALLEYVIEW(i915) ?
-						   VLV_GT_RENDER_RC6 :
-						   GEN6_GT_GFX_RC6);
+	val = intel_rc6_residency_ns(i915,
+				     IS_VALLEYVIEW(i915) ?
+				     VLV_GT_RENDER_RC6 :
+				     GEN6_GT_GFX_RC6);
 
-		if (HAS_RC6p(i915))
-			val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
+	if (HAS_RC6p(i915))
+		val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
+
+	if (HAS_RC6pp(i915))
+		val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
+
+	return val;
+}
 
-		if (HAS_RC6pp(i915))
-			val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
+static u64 get_rc6(struct drm_i915_private *i915, bool locked)
+{
+#if IS_ENABLED(CONFIG_PM)
+	unsigned long flags;
+	u64 val;
 
+	if (intel_runtime_pm_get_if_in_use(i915)) {
+		val = __get_rc6(i915);
 		intel_runtime_pm_put(i915);
 
 		/*
@@ -481,6 +491,9 @@ static u64 get_rc6(struct drm_i915_private *i915, bool locked)
 	}
 
 	return val;
+#else
+	return __get_rc6(i915);
+#endif
 }
 
 static u64 __i915_pmu_event_read(struct perf_event *event, bool locked)
Signature · 6 lines
2.14.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx

maintainer
#10 lore raw

PATCH 1/4 drm/i915: Lock out execlist tasklet while peeking inside for busy-stats

In reply to Rodrigo Vivi

From: Chris Wilson <[email protected]>

In order to prevent a race condition where we may end up overaccounting the active state and leaving the busy-stats believing the GPU is 100% busy, lock out the tasklet while we reconstruct the busy state. There is no direct spinlock guard for the execlists->port[], so we need to utilise tasklet_disable() as a synchronous barrier to prevent it, the only writer to execlists->port[], from running at the same time as the enable.

Fixes 4900727d35bb ("drm/i915/pmu: Reconstruct active state on starting busy-stats")
Cc Tvrtko Ursulin <[email protected]>
---
 drivers/gpu/drm/i915/intel_engine_cs.c | 20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)
drivers/gpu/drm/i915/intel_engine_cs.c +12 -8
index acc661aa9c0c..fa960cfd2764 100644
@@ -1945,16 +1945,22 @@ intel_engine_lookup_user(struct drm_i915_private *i915, u8 class, u8 instance)
  */
 int intel_enable_engine_stats(struct intel_engine_cs *engine)
 {
+	struct intel_engine_execlists *execlists = &engine->execlists;
 	unsigned long flags;
+	int err = 0;
 
 	if (!intel_engine_supports_stats(engine))
 		return -ENODEV;
 
+	tasklet_disable(&execlists->tasklet);
 	spin_lock_irqsave(&engine->stats.lock, flags);
-	if (engine->stats.enabled == ~0)
-		goto busy;
+
+	if (unlikely(engine->stats.enabled == ~0)) {
+		err = -EBUSY;
+		goto unlock;
+	}
+
 	if (engine->stats.enabled++ == 0) {
-		struct intel_engine_execlists *execlists = &engine->execlists;
 		const struct execlist_port *port = execlists->port;
 		unsigned int num_ports = execlists_num_ports(execlists);
 
@@ -1969,14 +1975,12 @@ int intel_enable_engine_stats(struct intel_engine_cs *engine)
 		if (engine->stats.active)
 			engine->stats.start = engine->stats.enabled_at;
 	}
-	spin_unlock_irqrestore(&engine->stats.lock, flags);
 
-	return 0;
-
-busy:
+unlock:
 	spin_unlock_irqrestore(&engine->stats.lock, flags);
+	tasklet_enable(&execlists->tasklet);
 
-	return -EBUSY;
+	return err;
 }
 
 static ktime_t __intel_engine_get_busy_time(struct intel_engine_cs *engine)
Signature · 6 lines
2.14.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer
#11 lore raw

PATCH 2/4 drm/i915/pmu: Fix PMU enable vs execlists tasklet race

From: Tvrtko Ursulin <[email protected]>

Commit 99e48bf98dd0 ("drm/i915: Lock out execlist tasklet while peeking inside for busy-stats") added a tasklet_disable call in busy stats enabling, but we failed to understand that the PMU enable callback runs as an hard IRQ (IPI).

Consequence of this is that the PMU enable callback can interrupt the execlists tasklet, and will then deadlock when it calls intel_engine_stats_enable->tasklet_disable.

To fix this, I realized it is possible to move the engine stats enablement and disablement to PMU event init and destroy hooks. This allows for much simpler implementation since those hooks run in normal context (can sleep).

v2: Extract engine_event_destroy. (Chris Wilson)

Fixes 99e48bf98dd0 ("drm/i915: Lock out execlist tasklet while peeking inside for busy-stats")

Testcase: igt/perf_pmu/enable-race-*

---
 drivers/gpu/drm/i915/i915_pmu.c         | 125 +++++++++++++-------------------
 drivers/gpu/drm/i915/intel_ringbuffer.h |  14 ----
 2 files changed, 52 insertions(+), 87 deletions(-)
drivers/gpu/drm/i915/i915_pmu.c +52 -73
index 55a8a1e29424..337eaa6ede52 100644
@@ -285,26 +285,41 @@ static u64 count_interrupts(struct drm_i915_private *i915)
 	return sum;
 }
 
-static void i915_pmu_event_destroy(struct perf_event *event)
+static void engine_event_destroy(struct perf_event *event)
 {
-	WARN_ON(event->parent);
+	struct drm_i915_private *i915 =
+		container_of(event->pmu, typeof(*i915), pmu.base);
+	struct intel_engine_cs *engine;
+
+	engine = intel_engine_lookup_user(i915,
+					  engine_event_class(event),
+					  engine_event_instance(event));
+	if (WARN_ON_ONCE(!engine))
+		return;
+
+	if (engine_event_sample(event) == I915_SAMPLE_BUSY &&
+	    intel_engine_supports_stats(engine))
+		intel_disable_engine_stats(engine);
 }
 
-static int engine_event_init(struct perf_event *event)
+static void i915_pmu_event_destroy(struct perf_event *event)
 {
-	struct drm_i915_private *i915 =
-		container_of(event->pmu, typeof(*i915), pmu.base);
+	WARN_ON(event->parent);
 
-	if (!intel_engine_lookup_user(i915, engine_event_class(event),
-				      engine_event_instance(event)))
-		return -ENODEV;
+	if (is_engine_event(event))
+		engine_event_destroy(event);
+}
 
-	switch (engine_event_sample(event)) {
+static int
+engine_event_status(struct intel_engine_cs *engine,
+		    enum drm_i915_pmu_engine_sample sample)
+{
+	switch (sample) {
 	case I915_SAMPLE_BUSY:
 	case I915_SAMPLE_WAIT:
 		break;
 	case I915_SAMPLE_SEMA:
-		if (INTEL_GEN(i915) < 6)
+		if (INTEL_GEN(engine->i915) < 6)
 			return -ENODEV;
 		break;
 	default:
@@ -314,6 +329,30 @@ static int engine_event_init(struct perf_event *event)
 	return 0;
 }
 
+static int engine_event_init(struct perf_event *event)
+{
+	struct drm_i915_private *i915 =
+		container_of(event->pmu, typeof(*i915), pmu.base);
+	struct intel_engine_cs *engine;
+	u8 sample;
+	int ret;
+
+	engine = intel_engine_lookup_user(i915, engine_event_class(event),
+					  engine_event_instance(event));
+	if (!engine)
+		return -ENODEV;
+
+	sample = engine_event_sample(event);
+	ret = engine_event_status(engine, sample);
+	if (ret)
+		return ret;
+
+	if (sample == I915_SAMPLE_BUSY && intel_engine_supports_stats(engine))
+		ret = intel_enable_engine_stats(engine);
+
+	return ret;
+}
+
 static int i915_pmu_event_init(struct perf_event *event)
 {
 	struct drm_i915_private *i915 =
@@ -387,7 +426,7 @@ static u64 __i915_pmu_event_read(struct perf_event *event)
 		if (WARN_ON_ONCE(!engine)) {
 			/* Do nothing */
 		} else if (sample == I915_SAMPLE_BUSY &&
-			   engine->pmu.busy_stats) {
+			   intel_engine_supports_stats(engine)) {
 			val = ktime_to_ns(intel_engine_get_busy_time(engine));
 		} else {
 			val = engine->pmu.sample[sample].cur;
@@ -442,12 +481,6 @@ static void i915_pmu_event_read(struct perf_event *event)
 	local64_add(new - prev, &event->count);
 }
 
-static bool engine_needs_busy_stats(struct intel_engine_cs *engine)
-{
-	return intel_engine_supports_stats(engine) &&
-	       (engine->pmu.enable & BIT(I915_SAMPLE_BUSY));
-}
-
 static void i915_pmu_enable(struct perf_event *event)
 {
 	struct drm_i915_private *i915 =
@@ -487,21 +520,7 @@ static void i915_pmu_enable(struct perf_event *event)
 
 		GEM_BUG_ON(sample >= I915_PMU_SAMPLE_BITS);
 		GEM_BUG_ON(engine->pmu.enable_count[sample] == ~0);
-		if (engine->pmu.enable_count[sample]++ == 0) {
-			/*
-			 * Enable engine busy stats tracking if needed or
-			 * alternatively cancel the scheduled disable.
-			 *
-			 * If the delayed disable was pending, cancel it and
-			 * in this case do not enable since it already is.
-			 */
-			if (engine_needs_busy_stats(engine) &&
-			    !engine->pmu.busy_stats) {
-				engine->pmu.busy_stats = true;
-				if (!cancel_delayed_work(&engine->pmu.disable_busy_stats))
-					intel_enable_engine_stats(engine);
-			}
-		}
+		engine->pmu.enable_count[sample]++;
 	}
 
 	/*
@@ -514,14 +533,6 @@ static void i915_pmu_enable(struct perf_event *event)
 	spin_unlock_irqrestore(&i915->pmu.lock, flags);
 }
 
-static void __disable_busy_stats(struct work_struct *work)
-{
-	struct intel_engine_cs *engine =
-	       container_of(work, typeof(*engine), pmu.disable_busy_stats.work);
-
-	intel_disable_engine_stats(engine);
-}
-
 static void i915_pmu_disable(struct perf_event *event)
 {
 	struct drm_i915_private *i915 =
@@ -545,26 +556,8 @@ 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 (--engine->pmu.enable_count[sample] == 0) {
+		if (--engine->pmu.enable_count[sample] == 0)
 			engine->pmu.enable &= ~BIT(sample);
-			if (!engine_needs_busy_stats(engine) &&
-			    engine->pmu.busy_stats) {
-				engine->pmu.busy_stats = false;
-				/*
-				 * We request a delayed disable to handle the
-				 * rapid on/off cycles on events, which can
-				 * happen when tools like perf stat start, in a
-				 * nicer way.
-				 *
-				 * In addition, this also helps with busy stats
-				 * accuracy with background CPU offline/online
-				 * migration events.
-				 */
-				queue_delayed_work(system_wq,
-						   &engine->pmu.disable_busy_stats,
-						   round_jiffies_up_relative(HZ));
-			}
-		}
 	}
 
 	GEM_BUG_ON(bit >= I915_PMU_MASK_BITS);
@@ -797,8 +790,6 @@ static void i915_pmu_unregister_cpuhp_state(struct drm_i915_private *i915)
 
 void i915_pmu_register(struct drm_i915_private *i915)
 {
-	struct intel_engine_cs *engine;
-	enum intel_engine_id id;
 	int ret;
 
 	if (INTEL_GEN(i915) <= 2) {
@@ -820,10 +811,6 @@ void i915_pmu_register(struct drm_i915_private *i915)
 	hrtimer_init(&i915->pmu.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 	i915->pmu.timer.function = i915_sample;
 
-	for_each_engine(engine, i915, id)
-		INIT_DELAYED_WORK(&engine->pmu.disable_busy_stats,
-				  __disable_busy_stats);
-
 	ret = perf_pmu_register(&i915->pmu.base, "i915", -1);
 	if (ret)
 		goto err;
@@ -843,9 +830,6 @@ void i915_pmu_register(struct drm_i915_private *i915)
 
 void i915_pmu_unregister(struct drm_i915_private *i915)
 {
-	struct intel_engine_cs *engine;
-	enum intel_engine_id id;
-
 	if (!i915->pmu.base.event_init)
 		return;
 
@@ -853,11 +837,6 @@ void i915_pmu_unregister(struct drm_i915_private *i915)
 
 	hrtimer_cancel(&i915->pmu.timer);
 
-	for_each_engine(engine, i915, id) {
-		GEM_BUG_ON(engine->pmu.busy_stats);
-		flush_delayed_work(&engine->pmu.disable_busy_stats);
-	}
-
 	i915_pmu_unregister_cpuhp_state(i915);
 
 	perf_pmu_unregister(&i915->pmu.base);
drivers/gpu/drm/i915/intel_ringbuffer.h +0 -14
index c5ff203e42d6..a0e7a6c2a57c 100644
@@ -366,20 +366,6 @@ struct intel_engine_cs {
 		 */
 #define I915_ENGINE_SAMPLE_MAX (I915_SAMPLE_SEMA + 1)
 		struct i915_pmu_sample sample[I915_ENGINE_SAMPLE_MAX];
-		/**
-		 * @busy_stats: Has enablement of engine stats tracking been
-		 * 		requested.
-		 */
-		bool busy_stats;
-		/**
-		 * @disable_busy_stats: Work item for busy stats disabling.
-		 *
-		 * Same as with @enable_busy_stats action, with the difference
-		 * that we delay it in case there are rapid enable-disable
-		 * actions, which can happen during tool startup (like perf
-		 * stat).
-		 */
-		struct delayed_work disable_busy_stats;
 	} pmu;
 
 	/*
Signature · 6 lines
2.14.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer
#12 lore raw

PATCH 3/4 drm/i915/pmu: Fix sleep under atomic in RC6 readout

In reply to Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

We are not allowed to call intel_runtime_pm_get from the PMU counter read callback since the former can sleep, and the latter is running under IRQ context.

To workaround this, we record the last known RC6 and while runtime suspended estimate its increase by querying the runtime PM core timestamps.

Downside of this approach is that we can temporarily lose a chunk of RC6 time, from the last PMU read-out to runtime suspend entry, but that will eventually catch up, once device comes back online and in the presence of PMU queries.

Also, we have to be careful not to overshoot the RC6 estimate, so once resumed after a period of approximation, we only update the counter once it catches up. With the observation that RC6 is increasing while the device is suspended, this should not pose a problem and can only cause slight inaccuracies due clock base differences.

v2: Simplify by estimating on top of PM core counters. (Imre)

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=104943

Fixes 6060b6aec03c ("drm/i915/pmu: Add RC6 residency metrics")

Testcase: igt/perf_pmu/rc6-runtime-pm

---
 drivers/gpu/drm/i915/i915_pmu.c | 93 ++++++++++++++++++++++++++++++++++-------
 drivers/gpu/drm/i915/i915_pmu.h |  6 +++
 2 files changed, 84 insertions(+), 15 deletions(-)
drivers/gpu/drm/i915/i915_pmu.c +78 -15
index 337eaa6ede52..e13859aaa2a3 100644
@@ -409,7 +409,81 @@ static int i915_pmu_event_init(struct perf_event *event)
 	return 0;
 }
 
-static u64 __i915_pmu_event_read(struct perf_event *event)
+static u64 get_rc6(struct drm_i915_private *i915, bool locked)
+{
+	unsigned long flags;
+	u64 val;
+
+	if (intel_runtime_pm_get_if_in_use(i915)) {
+		val = intel_rc6_residency_ns(i915, IS_VALLEYVIEW(i915) ?
+						   VLV_GT_RENDER_RC6 :
+						   GEN6_GT_GFX_RC6);
+
+		if (HAS_RC6p(i915))
+			val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
+
+		if (HAS_RC6pp(i915))
+			val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
+
+		intel_runtime_pm_put(i915);
+
+		/*
+		 * If we are coming back from being runtime suspended we must
+		 * be careful not to report a larger value than returned
+		 * previously.
+		 */
+
+		if (!locked)
+			spin_lock_irqsave(&i915->pmu.lock, flags);
+
+		if (val >= i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur) {
+			i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur = 0;
+			i915->pmu.sample[__I915_SAMPLE_RC6].cur = val;
+		} else {
+			val = i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur;
+		}
+
+		if (!locked)
+			spin_unlock_irqrestore(&i915->pmu.lock, flags);
+	} else {
+		struct pci_dev *pdev = i915->drm.pdev;
+		struct device *kdev = &pdev->dev;
+		unsigned long flags2;
+
+		/*
+		 * We are runtime suspended.
+		 *
+		 * Report the delta from when the device was suspended to now,
+		 * on top of the last known real value, as the approximated RC6
+		 * counter value.
+		 */
+		if (!locked)
+			spin_lock_irqsave(&i915->pmu.lock, flags);
+
+		spin_lock_irqsave(&kdev->power.lock, flags2);
+
+		if (!i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur)
+			i915->pmu.suspended_jiffies_last =
+						kdev->power.suspended_jiffies;
+
+		val = kdev->power.suspended_jiffies -
+		      i915->pmu.suspended_jiffies_last;
+		val += jiffies - kdev->power.accounting_timestamp;
+
+		spin_unlock_irqrestore(&kdev->power.lock, flags2);
+
+		val = jiffies_to_nsecs(val);
+		val += i915->pmu.sample[__I915_SAMPLE_RC6].cur;
+		i915->pmu.sample[__I915_SAMPLE_RC6_ESTIMATED].cur = val;
+
+		if (!locked)
+			spin_unlock_irqrestore(&i915->pmu.lock, flags);
+	}
+
+	return val;
+}
+
+static u64 __i915_pmu_event_read(struct perf_event *event, bool locked)
 {
 	struct drm_i915_private *i915 =
 		container_of(event->pmu, typeof(*i915), pmu.base);
@@ -447,18 +521,7 @@ static u64 __i915_pmu_event_read(struct perf_event *event)
 			val = count_interrupts(i915);
 			break;
 		case I915_PMU_RC6_RESIDENCY:
-			intel_runtime_pm_get(i915);
-			val = intel_rc6_residency_ns(i915,
-						     IS_VALLEYVIEW(i915) ?
-						     VLV_GT_RENDER_RC6 :
-						     GEN6_GT_GFX_RC6);
-			if (HAS_RC6p(i915))
-				val += intel_rc6_residency_ns(i915,
-							      GEN6_GT_GFX_RC6p);
-			if (HAS_RC6pp(i915))
-				val += intel_rc6_residency_ns(i915,
-							      GEN6_GT_GFX_RC6pp);
-			intel_runtime_pm_put(i915);
+			val = get_rc6(i915, locked);
 			break;
 		}
 	}
@@ -473,7 +536,7 @@ static void i915_pmu_event_read(struct perf_event *event)
 
 again:
 	prev = local64_read(&hwc->prev_count);
-	new = __i915_pmu_event_read(event);
+	new = __i915_pmu_event_read(event, false);
 
 	if (local64_cmpxchg(&hwc->prev_count, prev, new) != prev)
 		goto again;
@@ -528,7 +591,7 @@ static void i915_pmu_enable(struct perf_event *event)
 	 * for all listeners. Even when the event was already enabled and has
 	 * an existing non-zero value.
 	 */
-	local64_set(&event->hw.prev_count, __i915_pmu_event_read(event));
+	local64_set(&event->hw.prev_count, __i915_pmu_event_read(event, true));
 
 	spin_unlock_irqrestore(&i915->pmu.lock, flags);
 }
drivers/gpu/drm/i915/i915_pmu.h +6 -0
index 40c154d13565..bb62df15afa4 100644
@@ -27,6 +27,8 @@
 enum {
 	__I915_SAMPLE_FREQ_ACT = 0,
 	__I915_SAMPLE_FREQ_REQ,
+	__I915_SAMPLE_RC6,
+	__I915_SAMPLE_RC6_ESTIMATED,
 	__I915_NUM_PMU_SAMPLERS
 };
 
@@ -94,6 +96,10 @@ struct i915_pmu {
 	 * struct intel_engine_cs.
 	 */
 	struct i915_pmu_sample sample[__I915_NUM_PMU_SAMPLERS];
+	/**
+	 * @suspended_jiffies_last: Cached suspend time from PM core.
+	 */
+	unsigned long suspended_jiffies_last;
 };
 
 #ifdef CONFIG_PERF_EVENTS
Signature · 6 lines
2.14.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
maintainer
#13 lore raw

PATCH 4/4 drm/i915/pmu: Fix building without CONFIG_PM

In reply to Tvrtko Ursulin

From: Chris Wilson <[email protected]>

As we peek inside struct device to query members guarded by CONFIG_PM, so must be the code.

Fixes 1fe699e30113 ("drm/i915/pmu: Fix sleep under atomic in RC6 readout")
Cc Tvrtko Ursulin <[email protected]>
---
 drivers/gpu/drm/i915/i915_pmu.c | 33 +++++++++++++++++++++++----------
 1 file changed, 23 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_pmu.c +23 -10
index e13859aaa2a3..0e9b98c32b62 100644
@@ -409,22 +409,32 @@ static int i915_pmu_event_init(struct perf_event *event)
 	return 0;
 }
 
-static u64 get_rc6(struct drm_i915_private *i915, bool locked)
+static u64 __get_rc6(struct drm_i915_private *i915)
 {
-	unsigned long flags;
 	u64 val;
 
-	if (intel_runtime_pm_get_if_in_use(i915)) {
-		val = intel_rc6_residency_ns(i915, IS_VALLEYVIEW(i915) ?
-						   VLV_GT_RENDER_RC6 :
-						   GEN6_GT_GFX_RC6);
+	val = intel_rc6_residency_ns(i915,
+				     IS_VALLEYVIEW(i915) ?
+				     VLV_GT_RENDER_RC6 :
+				     GEN6_GT_GFX_RC6);
 
-		if (HAS_RC6p(i915))
-			val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
+	if (HAS_RC6p(i915))
+		val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6p);
+
+	if (HAS_RC6pp(i915))
+		val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
+
+	return val;
+}
 
-		if (HAS_RC6pp(i915))
-			val += intel_rc6_residency_ns(i915, GEN6_GT_GFX_RC6pp);
+static u64 get_rc6(struct drm_i915_private *i915, bool locked)
+{
+#if IS_ENABLED(CONFIG_PM)
+	unsigned long flags;
+	u64 val;
 
+	if (intel_runtime_pm_get_if_in_use(i915)) {
+		val = __get_rc6(i915);
 		intel_runtime_pm_put(i915);
 
 		/*
@@ -481,6 +491,9 @@ static u64 get_rc6(struct drm_i915_private *i915, bool locked)
 	}
 
 	return val;
+#else
+	return __get_rc6(i915);
+#endif
 }
 
 static u64 __i915_pmu_event_read(struct perf_event *event, bool locked)
Signature · 6 lines
2.14.1

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
#14 lore raw

✗ Fi.CI.BAT: failure for drm/i915/breadcrumbs: Ignore unsubmitted signalers (rev2)

In reply to Rodrigo Vivi
== Series Details ==

== Series Details ==

Series: drm/i915/breadcrumbs: Ignore unsubmitted signalers (rev2)
URL : https://patchwork.freedesktop.org/series/37724/
State : failure

== Summary ==

Applying: drm/i915/breadcrumbs: Ignore unsubmitted signalers
error: Failed to merge in the changes.
Using index info to reconstruct a base tree...
M drivers/gpu/drm/i915/intel_breadcrumbs.c
Falling back to patching base and 3-way merge...
Auto-merging drivers/gpu/drm/i915/intel_breadcrumbs.c
CONFLICT (content): Merge conflict in drivers/gpu/drm/i915/intel_breadcrumbs.c Patch failed at 0001 drm/i915/breadcrumbs: Ignore unsubmitted signalers The copy of the patch that failed is found in: .git/rebase-apply/patch When you have resolved this problem, run "git am --continue".
If you prefer to skip this patch, run "git am --skip" instead.
To restore the original branch and stop patching, run "git am --abort".

_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx