Skip to contentSkip to conversation

perf: fix PERF_EVENT_IOC_PERIOD to force-reset the period

PATCH In review Will Deacon 4 participants 7 messages Nov 27, 2013 – Dec 18, 2013

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

PATCH 1/2 perf: fix PERF_EVENT_IOC_PERIOD to force-reset the period

From: Peter Zijlstra <[email protected]>

Vince Weaver reports that, on all architectures apart from ARM, PERF_EVENT_IOC_PERIOD doesn't actually update the period until the next event fires. This is counter-intuitive behaviour and is better dealt with in the core code.

This patch ensures that the period is forcefully reset when dealing with such a request in the core code. A subsequent patch removes the equivalent hack from the ARM back-end.

Cc Vince Weaver <[email protected]>
Cc Peter Zijlstra <[email protected]>

---

Peter -- I didn't want this to get lost, so I picked it up, fixed a typo
and wrote a commit message. Could I add your SoB please?

Cheers,
Will

 kernel/events/core.c | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)
kernel/events/core.c +15 -1
index d724e7757cd1..1c626b923ddf 100644
@@ -3527,7 +3527,7 @@ static void perf_event_for_each(struct perf_event *event,
 static int perf_event_period(struct perf_event *event, u64 __user *arg)
 {
 	struct perf_event_context *ctx = event->ctx;
-	int ret = 0;
+	int ret = 0, active;
 	u64 value;
 
 	if (!is_sampling_event(event))
@@ -3551,6 +3551,20 @@ static int perf_event_period(struct perf_event *event, u64 __user *arg)
 		event->attr.sample_period = value;
 		event->hw.sample_period = value;
 	}
+
+	active = (event->state == PERF_EVENT_STATE_ACTIVE);
+	if (active) {
+		perf_pmu_disable(ctx->pmu);
+		event->pmu->stop(event, PERF_EF_UPDATE);
+	}
+
+	local64_set(&event->hw.period_left, 0);
+
+	if (active) {
+		event->pmu->start(event, PERF_EF_RELOAD);
+		perf_pmu_enable(ctx->pmu);
+	}
+
 unlock:
 	raw_spin_unlock_irq(&ctx->lock);
 
Signature · 1 line
1.8.2.2
#1 lore raw

PATCH 2/2 Revert "ARM: 7556/1: perf: fix updated event period in response to PERF_EVENT_IOC_PERIOD"

This reverts commit 3581fe0ef37ce12ac7a4f74831168352ae848edc.

Updates to the handling of PERF_EVENT_IOC_PERIOD in the core code mean we no longer have to play this horrible game.

Cc Peter Zijlstra <[email protected]>
---
 arch/arm/kernel/perf_event.c | 4 ----
 1 file changed, 4 deletions(-)
arch/arm/kernel/perf_event.c +0 -4
index bc3f2efa0d86..789d846a9184 100644
@@ -99,10 +99,6 @@ int armpmu_event_set_period(struct perf_event *event)
 	s64 period = hwc->sample_period;
 	int ret = 0;
 
-	/* The period may have been changed by PERF_EVENT_IOC_PERIOD */
-	if (unlikely(period != hwc->last_period))
-		left = period - (hwc->last_period - left);
-
 	if (unlikely(left <= -period)) {
 		left = period;
 		local64_set(&hwc->period_left, left);
Signature · 1 line
1.8.2.2
#2 lore raw

[tip:perf/core] Revert "ARM: 7556/1: perf: fix updated event period in response to PERF_EVENT_IOC_PERIOD"

Revert "ARM: 7556/1: perf: fix updated event period in response to PERF_EVENT_IOC_PERIOD"
Commit-ID:  9450d14fb959336803e5209119eb422b667b96aa
Gitweb:     http://git.kernel.org/tip/9450d14fb959336803e5209119eb422b667b96aa
Author:     Will Deacon <[email protected]>
AuthorDate: Wed, 27 Nov 2013 13:54:39 +0000
Committer:  Ingo Molnar <[email protected]>
CommitDate: Tue, 17 Dec 2013 15:21:35 +0100

Revert "ARM: 7556/1: perf: fix updated event period in response to PERF_EVENT_IOC_PERIOD"

This reverts commit 3581fe0ef37ce12ac7a4f74831168352ae848edc.

Fixes to the handling of PERF_EVENT_IOC_PERIOD in the core code mean we no longer have to play this horrible game.

---
 arch/arm/kernel/perf_event.c | 4 ----
 1 file changed, 4 deletions(-)
arch/arm/kernel/perf_event.c +0 -4
index bc3f2ef..789d846 100644
@@ -99,10 +99,6 @@ int armpmu_event_set_period(struct perf_event *event)
 	s64 period = hwc->sample_period;
 	int ret = 0;
 
-	/* The period may have been changed by PERF_EVENT_IOC_PERIOD */
-	if (unlikely(period != hwc->last_period))
-		left = period - (hwc->last_period - left);
-
 	if (unlikely(left <= -period)) {
 		left = period;
 		local64_set(&hwc->period_left, left);
#3 lore raw
In reply to Will Deacon
Quoted 4 lines · Will Deacon

Vince Weaver reports that, on all architectures apart from ARM, PERF_EVENT_IOC_PERIOD doesn't actually update the period until the next event fires. This is counter-intuitive behaviour and is better dealt with in the core code.

To be fair, it was Andreas Sandberg <[email protected]> who noticed this issue and reported it to me (in the context of the manpage being wrong). I just passed it on to the linux-kernel list.

Vince

#4 lore raw
In reply to Will Deacon

Hi Peter,

Do you have any thoughts on the below? It would be good to add your signed-off to the patch, since you wrote it originally.

Cheers,

Will

Quoted 62 lines · Will Deacon

From: Peter Zijlstra <[email protected]>

Vince Weaver reports that, on all architectures apart from ARM, PERF_EVENT_IOC_PERIOD doesn't actually update the period until the next event fires. This is counter-intuitive behaviour and is better dealt with in the core code.

This patch ensures that the period is forcefully reset when dealing with such a request in the core code. A subsequent patch removes the equivalent hack from the ARM back-end.

Cc Vince Weaver <[email protected]>
Cc Peter Zijlstra <[email protected]>

---

Peter -- I didn't want this to get lost, so I picked it up, fixed a typo
and wrote a commit message. Could I add your SoB please?

Cheers,
Will

 kernel/events/core.c | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)
kernel/events/core.c +15 -1
index d724e7757cd1..1c626b923ddf 100644
@@ -3527,7 +3527,7 @@ static void perf_event_for_each(struct perf_event *event,
 static int perf_event_period(struct perf_event *event, u64 __user *arg)
 {
 	struct perf_event_context *ctx = event->ctx;
-	int ret = 0;
+	int ret = 0, active;
 	u64 value;
 
 	if (!is_sampling_event(event))
@@ -3551,6 +3551,20 @@ static int perf_event_period(struct perf_event *event, u64 __user *arg)
 		event->attr.sample_period = value;
 		event->hw.sample_period = value;
 	}
+
+	active = (event->state == PERF_EVENT_STATE_ACTIVE);
+	if (active) {
+		perf_pmu_disable(ctx->pmu);
+		event->pmu->stop(event, PERF_EF_UPDATE);
+	}
+
+	local64_set(&event->hw.period_left, 0);
+
+	if (active) {
+		event->pmu->start(event, PERF_EF_RELOAD);
+		perf_pmu_enable(ctx->pmu);
+	}
+
 unlock:
 	raw_spin_unlock_irq(&ctx->lock);
 
Signature · 1 line
1.8.2.2
maintainer
#5 lore raw
Quoted 5 lines · Will Deacon

Hi Peter,

Do you have any thoughts on the below? It would be good to add your signed-off to the patch, since you wrote it originally.

Picked them up, thanks will!

#6 lore raw

[tip:perf/core] perf: Fix PERF_EVENT_IOC_PERIOD to force-reset the period

In reply to Will Deacon
perf: Fix PERF_EVENT_IOC_PERIOD to force-reset the period
Commit-ID:  bad7192b842c83e580747ca57104dd51fe08c223
Gitweb:     http://git.kernel.org/tip/bad7192b842c83e580747ca57104dd51fe08c223
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Wed, 27 Nov 2013 13:54:38 +0000
Committer:  Ingo Molnar <[email protected]>
CommitDate: Tue, 17 Dec 2013 15:21:33 +0100

perf: Fix PERF_EVENT_IOC_PERIOD to force-reset the period

Vince Weaver reports that, on all architectures apart from ARM, PERF_EVENT_IOC_PERIOD doesn't actually update the period until the next event fires. This is counter-intuitive behaviour and is better dealt with in the core code.

This patch ensures that the period is forcefully reset when dealing with such a request in the core code. A subsequent patch removes the equivalent hack from the ARM back-end.

---
 kernel/events/core.c | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)
kernel/events/core.c +15 -1
index 403b781..89d34f9 100644
@@ -3527,7 +3527,7 @@ static void perf_event_for_each(struct perf_event *event,
 static int perf_event_period(struct perf_event *event, u64 __user *arg)
 {
 	struct perf_event_context *ctx = event->ctx;
-	int ret = 0;
+	int ret = 0, active;
 	u64 value;
 
 	if (!is_sampling_event(event))
@@ -3551,6 +3551,20 @@ static int perf_event_period(struct perf_event *event, u64 __user *arg)
 		event->attr.sample_period = value;
 		event->hw.sample_period = value;
 	}
+
+	active = (event->state == PERF_EVENT_STATE_ACTIVE);
+	if (active) {
+		perf_pmu_disable(ctx->pmu);
+		event->pmu->stop(event, PERF_EF_UPDATE);
+	}
+
+	local64_set(&event->hw.period_left, 0);
+
+	if (active) {
+		event->pmu->start(event, PERF_EF_RELOAD);
+		perf_pmu_enable(ctx->pmu);
+	}
+
 unlock:
 	raw_spin_unlock_irq(&ctx->lock);