Skip to contentSkip to conversation

perf/x86/intel: Fix n_metric for the canceled group

PATCH In review [email protected] 4 participants 8 messages Sep 30, 2020 – Oct 7, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf/x86/intel: Fix n_metric for the canceled group

From: Kan Liang <[email protected]>

When a group that has TopDown members is failed to be scheduled, any later TopDown groups will not return valid values.

Here is an example.

A background perf that occupies all the GP counters and the fixed counter 1.
$perf stat -e "{cycles,cycles,cycles,cycles,cycles,cycles,cycles,
cycles,cycles}:D" -a

A user monitors a TopDown group. It works well, because the fixed counter 3 and the PERF_METRICS are available.
$perf stat -x, --topdown -- ./workload
retiring,bad speculation,frontend bound,backend bound, 18.0,16.1,40.4,25.5,

Then the user tries to monitor a group that has TopDown members. Because of the cycles event, the group is failed to be scheduled.
$perf stat -x, -e '{slots,topdown-retiring,topdown-be-bound,

                     topdown-fe-bound,topdown-bad-spec,cycles}'
                     -- ./workload
    <not counted>,,slots,0,0.00,,
    <not counted>,,topdown-retiring,0,0.00,,
    <not counted>,,topdown-be-bound,0,0.00,,
    <not counted>,,topdown-fe-bound,0,0.00,,
    <not counted>,,topdown-bad-spec,0,0.00,,
    <not counted>,,cycles,0,0.00,,

The user tries to monitor a TopDown group again. It doesn't work anymore.
$perf stat -x, --topdown -- ./workload

    ,,,,,

In a txn, cancel_txn() is to truncate the event_list for a canceled group and update the number of events added in this transaction. However, the number of TopDown events added in this transaction is not updated. The kernel will probably fail to add new Topdown events.

Check if the canceled group has Topdown events. If so, subtract the TopDown events from n_metric accordingly.

Fixes 7b2c05a15d29 ("perf/x86/intel: Generic support for hardware TopDown metrics")
---
 arch/x86/events/core.c | 10 ++++++++++
 1 file changed, 10 insertions(+)
arch/x86/events/core.c +10 -0
index 0f3d01562ded..4cb3ccbe2d62 100644
@@ -2017,6 +2017,7 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
 {
 	unsigned int txn_flags;
 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	int i;
 
 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
 
@@ -2031,6 +2032,15 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
 	 */
 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
+
+	/* Subtract Topdown events in the canceled group from n_metric */
+	if (x86_pmu.intel_cap.perf_metrics && cpuc->n_metric) {
+		for (i = 0; i < cpuc->n_txn; i++) {
+			if (is_metric_event(cpuc->event_list[i + cpuc->n_events]))
+				__this_cpu_dec(cpu_hw_events.n_metric);
+		}
+		WARN_ON_ONCE(__this_cpu_read(cpu_hw_events.n_metric) < 0);
+	}
 	perf_pmu_enable(pmu);
 }
 
Signature · 1 line
2.17.1
maintainer
#1 lore raw
Quoted 68 lines · On Wed, Sep 30, 2020 at 07:29:35AM -0700, [email protected] wrote:

From: Kan Liang <[email protected]>

When a group that has TopDown members is failed to be scheduled, any later TopDown groups will not return valid values.

Here is an example.

A background perf that occupies all the GP counters and the fixed counter 1.
$perf stat -e "{cycles,cycles,cycles,cycles,cycles,cycles,cycles,
cycles,cycles}:D" -a

A user monitors a TopDown group. It works well, because the fixed counter 3 and the PERF_METRICS are available.
$perf stat -x, --topdown -- ./workload
retiring,bad speculation,frontend bound,backend bound, 18.0,16.1,40.4,25.5,

Then the user tries to monitor a group that has TopDown members. Because of the cycles event, the group is failed to be scheduled.
$perf stat -x, -e '{slots,topdown-retiring,topdown-be-bound,

                     topdown-fe-bound,topdown-bad-spec,cycles}'
                     -- ./workload
    <not counted>,,slots,0,0.00,,
    <not counted>,,topdown-retiring,0,0.00,,
    <not counted>,,topdown-be-bound,0,0.00,,
    <not counted>,,topdown-fe-bound,0,0.00,,
    <not counted>,,topdown-bad-spec,0,0.00,,
    <not counted>,,cycles,0,0.00,,

The user tries to monitor a TopDown group again. It doesn't work anymore.
$perf stat -x, --topdown -- ./workload

    ,,,,,

In a txn, cancel_txn() is to truncate the event_list for a canceled group and update the number of events added in this transaction. However, the number of TopDown events added in this transaction is not updated. The kernel will probably fail to add new Topdown events.

Check if the canceled group has Topdown events. If so, subtract the TopDown events from n_metric accordingly.

Fixes 7b2c05a15d29 ("perf/x86/intel: Generic support for hardware TopDown metrics")
---
 arch/x86/events/core.c | 10 ++++++++++
 1 file changed, 10 insertions(+)
arch/x86/events/core.c +2 -0
index 0f3d01562ded..4cb3ccbe2d62 100644
@@ -2017,6 +2017,7 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
 {
 	unsigned int txn_flags;
 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	int i;
 
 	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
 
@@ -2031,6 +2032,15 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
 	 */
 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
+

arch/x86/events/core.c static void x86_pmu_cancel_txn(struct pmu *pmu) · line 2045

Quoted 10 lines
arch/x86/events/core.c +8 -0
+	/* Subtract Topdown events in the canceled group from n_metric */
+	if (x86_pmu.intel_cap.perf_metrics && cpuc->n_metric) {
+		for (i = 0; i < cpuc->n_txn; i++) {
+			if (is_metric_event(cpuc->event_list[i + cpuc->n_events]))
+				__this_cpu_dec(cpu_hw_events.n_metric);
+		}
+		WARN_ON_ONCE(__this_cpu_read(cpu_hw_events.n_metric) < 0);
+	}
 	perf_pmu_enable(pmu);
 }

Urgh, I'd much rather we add n_txn_metric. But also, while looking at this, don't we have the same problem with n_pair ?

Something like this perhaps...

---

arch/x86/events/core.c +8 -1
index 757e49755e7c..9b7792c0b6fb 100644
@@ -1066,6 +1066,7 @@ static int add_nr_metric_event(struct cpu_hw_events *cpuc,
 		if (cpuc->n_metric == INTEL_TD_METRIC_NUM)
 			return -EINVAL;
 		cpuc->n_metric++;
+		cpuc->n_txn_metric++;
 	}
 
 	return 0;
@@ -1089,8 +1090,10 @@ static int collect_event(struct cpu_hw_events *cpuc, struct perf_event *event,
 		return -EINVAL;
 
 	cpuc->event_list[n] = event;
-	if (is_counter_pair(&event->hw))
+	if (is_counter_pair(&event->hw)) {
 		cpuc->n_pair++;
+		cpuc->n_txn_pair++;
+	}
 
 	return 0;
 }
@@ -2062,6 +2065,8 @@ static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
 
 	perf_pmu_disable(pmu);
 	__this_cpu_write(cpu_hw_events.n_txn, 0);
+	__this_cpu_write(cpu_hw_events.n_txn_metric, 0);
+	__this_cpu_write(cpu_hw_events.n_txn_pair, 0);
 }
 
 /*
@@ -2087,6 +2092,8 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
 	 */
 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
+	__this_cpu_sub(cpu_hw_events.n_metric, __this_cpu_read(cpu_hw_events.n_txn_metric));
+	__this_cpu_sub(cpu_hw_events.n_pair, __this_cpu_read(cpu_hw_events.n_txn_pair));
 	perf_pmu_enable(pmu);
 }
 
arch/x86/events/perf_event.h +2 -0
index 345442410a4d..6348105b6d30 100644
@@ -235,6 +235,8 @@ struct cpu_hw_events {
 					     they've never been enabled yet */
 	int			n_txn;    /* the # last events in the below arrays;
 					     added in the current transaction */
+	int			n_txn_metric;
+	int			n_txn_pair;
 	int			assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
 	u64			tags[X86_PMC_IDX_MAX];
 
author Tested-by
#2 lore raw
Quoted 84 lines · Peter Zijlstra
Quoted 68 lines · On Wed, Sep 30, 2020 at 07:29:35AM -0700, [email protected] wrote:

From: Kan Liang <[email protected]>

When a group that has TopDown members is failed to be scheduled, any later TopDown groups will not return valid values.

Here is an example.

A background perf that occupies all the GP counters and the fixed counter 1.
$perf stat -e "{cycles,cycles,cycles,cycles,cycles,cycles,cycles,
cycles,cycles}:D" -a

A user monitors a TopDown group. It works well, because the fixed counter 3 and the PERF_METRICS are available.
$perf stat -x, --topdown -- ./workload
retiring,bad speculation,frontend bound,backend bound, 18.0,16.1,40.4,25.5,

Then the user tries to monitor a group that has TopDown members. Because of the cycles event, the group is failed to be scheduled.
$perf stat -x, -e '{slots,topdown-retiring,topdown-be-bound,

                      topdown-fe-bound,topdown-bad-spec,cycles}'
                      -- ./workload
     <not counted>,,slots,0,0.00,,
     <not counted>,,topdown-retiring,0,0.00,,
     <not counted>,,topdown-be-bound,0,0.00,,
     <not counted>,,topdown-fe-bound,0,0.00,,
     <not counted>,,topdown-bad-spec,0,0.00,,
     <not counted>,,cycles,0,0.00,,

The user tries to monitor a TopDown group again. It doesn't work anymore.
$perf stat -x, --topdown -- ./workload

     ,,,,,

In a txn, cancel_txn() is to truncate the event_list for a canceled group and update the number of events added in this transaction. However, the number of TopDown events added in this transaction is not updated. The kernel will probably fail to add new Topdown events.

Check if the canceled group has Topdown events. If so, subtract the TopDown events from n_metric accordingly.

Fixes 7b2c05a15d29 ("perf/x86/intel: Generic support for hardware TopDown metrics")
---
  arch/x86/events/core.c | 10 ++++++++++
  1 file changed, 10 insertions(+)
arch/x86/events/core.c +2 -0
index 0f3d01562ded..4cb3ccbe2d62 100644
@@ -2017,6 +2017,7 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
  {
  	unsigned int txn_flags;
  	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+	int i;
 
  	WARN_ON_ONCE(!cpuc->txn_flags);	/* no txn in flight */
 
@@ -2031,6 +2032,15 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
  	 */
  	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
  	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
+

arch/x86/events/core.c static void x86_pmu_cancel_txn(struct pmu *pmu) · line 2045

Quoted 10 lines
arch/x86/events/core.c +8 -0
+	/* Subtract Topdown events in the canceled group from n_metric */
+	if (x86_pmu.intel_cap.perf_metrics && cpuc->n_metric) {
+		for (i = 0; i < cpuc->n_txn; i++) {
+			if (is_metric_event(cpuc->event_list[i + cpuc->n_events]))
+				__this_cpu_dec(cpu_hw_events.n_metric);
+		}
+		WARN_ON_ONCE(__this_cpu_read(cpu_hw_events.n_metric) < 0);
+	}
  	perf_pmu_enable(pmu);
  }

Urgh, I'd much rather we add n_txn_metric. But also, while looking at this, don't we have the same problem with n_pair ?

Quoted 2 lines

Something like this perhaps...

Sure. For the perf metric, the below patch fixes the problem.

Thanks,
Kan

Quoted 57 lines

---

arch/x86/events/core.c +8 -1
index 757e49755e7c..9b7792c0b6fb 100644
@@ -1066,6 +1066,7 @@ static int add_nr_metric_event(struct cpu_hw_events *cpuc,
  		if (cpuc->n_metric == INTEL_TD_METRIC_NUM)
  			return -EINVAL;
  		cpuc->n_metric++;
+		cpuc->n_txn_metric++;
  	}
 
  	return 0;
@@ -1089,8 +1090,10 @@ static int collect_event(struct cpu_hw_events *cpuc, struct perf_event *event,
  		return -EINVAL;
 
  	cpuc->event_list[n] = event;
-	if (is_counter_pair(&event->hw))
+	if (is_counter_pair(&event->hw)) {
  		cpuc->n_pair++;
+		cpuc->n_txn_pair++;
+	}
 
  	return 0;
  }
@@ -2062,6 +2065,8 @@ static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
 
  	perf_pmu_disable(pmu);
  	__this_cpu_write(cpu_hw_events.n_txn, 0);
+	__this_cpu_write(cpu_hw_events.n_txn_metric, 0);
+	__this_cpu_write(cpu_hw_events.n_txn_pair, 0);
  }
 
  /*
@@ -2087,6 +2092,8 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
  	 */
  	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
  	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
+	__this_cpu_sub(cpu_hw_events.n_metric, __this_cpu_read(cpu_hw_events.n_txn_metric));
+	__this_cpu_sub(cpu_hw_events.n_pair, __this_cpu_read(cpu_hw_events.n_txn_pair));
  	perf_pmu_enable(pmu);
  }
 
arch/x86/events/perf_event.h +2 -0
index 345442410a4d..6348105b6d30 100644
@@ -235,6 +235,8 @@ struct cpu_hw_events {
  					     they've never been enabled yet */
  	int			n_txn;    /* the # last events in the below arrays;
  					     added in the current transaction */
+	int			n_txn_metric;
+	int			n_txn_pair;
  	int			assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
  	u64			tags[X86_PMC_IDX_MAX];
 
Tested-by
#3 lore raw
Quoted 55 lines · Kan
Quoted 54 lines · Peter Zijlstra
Quoted 40 lines · On Wed, Sep 30, 2020 at 07:29:35AM -0700, [email protected] wrote:

From: Kan Liang <[email protected]>

When a group that has TopDown members is failed to be scheduled, any later TopDown groups will not return valid values.

Here is an example.

A background perf that occupies all the GP counters and the fixed counter 1.
  $perf stat -e "{cycles,cycles,cycles,cycles,cycles,cycles,cycles, cycles,cycles}:D" -a

A user monitors a TopDown group. It works well, because the fixed counter 3 and the PERF_METRICS are available.
  $perf stat -x, --topdown -- ./workload
    retiring,bad speculation,frontend bound,backend bound, 18.0,16.1,40.4,25.5,

Then the user tries to monitor a group that has TopDown members. Because of the cycles event, the group is failed to be scheduled. $perf stat -x, -e '{slots,topdown-retiring,topdown-be-bound, topdown-fe-bound,topdown-bad-spec,cycles}' -- ./workload
     <not counted>,,slots,0,0.00,,
     <not counted>,,topdown-retiring,0,0.00,,
     <not counted>,,topdown-be-bound,0,0.00,,
     <not counted>,,topdown-fe-bound,0,0.00,,
     <not counted>,,topdown-bad-spec,0,0.00,,
     <not counted>,,cycles,0,0.00,,

The user tries to monitor a TopDown group again. It doesn't work anymore. $perf stat -x, --topdown -- ./workload

     ,,,,,

In a txn, cancel_txn() is to truncate the event_list for a canceled group and update the number of events added in this transaction. However, the number of TopDown events added in this transaction is not updated. The kernel will probably fail to add new Topdown events.

Quoted 8 lines

Check if the canceled group has Topdown events. If so, subtract the TopDown events from n_metric accordingly.

Fixes 7b2c05a15d29 ("perf/x86/intel: Generic support for hardware TopDown metrics")

---

Urgh, I'd much rather we add n_txn_metric. But also, while looking at this, don't we have the same problem with n_pair ?

Quoted 6 lines
Quoted 2 lines

Something like this perhaps...

Sure. For the perf metric, the below patch fixes the problem.

Excerpt from test script:

sudo perf stat -e "{cycles,cycles,cycles,cycles}:D" -a sleep 10 &

# should succeed:
sudo perf stat -e "{fp_ret_sse_avx_ops.all}:D" -a workload
# should fail:
sudo perf stat -e "{fp_ret_sse_avx_ops.all,fp_ret_sse_avx_ops.all,cycles}:D" -a workload
# previously failed, now succeeds with this patch:
sudo perf stat -e "{fp_ret_sse_avx_ops.all}:D" -a workload

Thanks both,

Kim

maintainer
#4 lore raw

PATCH 1/1 perf/x86: Fix n_pair for cancelled txn

Quoted 1 line · Kim Phillips

---
Subject: perf/x86: Fix n_pair for cancelled txn
From: Peter Zijlstra <[email protected]>
Date: Mon Oct 5 10:09:06 CEST 2020

Kan reported that n_metric gets corrupted for cancelled transactions; a similar issue exists for n_pair for AMD's Large Increment thing.

The problem was confirmed and confirmed fixed by Kim using:

sudo perf stat -e "{cycles,cycles,cycles,cycles}:D" -a sleep 10 &

  # should succeed:
  sudo perf stat -e "{fp_ret_sse_avx_ops.all}:D" -a workload
  # should fail:
  sudo perf stat -e "{fp_ret_sse_avx_ops.all,fp_ret_sse_avx_ops.all,cycles}:D" -a workload
  # previously failed, now succeeds with this patch:
  sudo perf stat -e "{fp_ret_sse_avx_ops.all}:D" -a workload
Fixes 5738891229a2 ("perf/x86/amd: Add support for Large Increment per Cycle Events")
---
 arch/x86/events/core.c       |    6 +++++-
 arch/x86/events/perf_event.h |    1 +
 2 files changed, 6 insertions(+), 1 deletion(-)
arch/x86/events/core.c +5 -1
@@ -1089,8 +1089,10 @@ static int collect_event(struct cpu_hw_e
 		return -EINVAL;
 
 	cpuc->event_list[n] = event;
-	if (is_counter_pair(&event->hw))
+	if (is_counter_pair(&event->hw)) {
 		cpuc->n_pair++;
+		cpuc->n_txn_pair++;
+	}
 
 	return 0;
 }
@@ -2062,6 +2064,7 @@ static void x86_pmu_start_txn(struct pmu
 
 	perf_pmu_disable(pmu);
 	__this_cpu_write(cpu_hw_events.n_txn, 0);
+	__this_cpu_write(cpu_hw_events.n_txn_pair, 0);
 }
 
 /*
@@ -2087,6 +2090,7 @@ static void x86_pmu_cancel_txn(struct pm
 	 */
 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
+	__this_cpu_sub(cpu_hw_events.n_pair, __this_cpu_read(cpu_hw_events.n_txn_pair));
 	perf_pmu_enable(pmu);
 }
 
arch/x86/events/perf_event.h +1 -0
@@ -235,6 +235,7 @@ struct cpu_hw_events {
 					     they've never been enabled yet */
 	int			n_txn;    /* the # last events in the below arrays;
 					     added in the current transaction */
+	int			n_txn_pair;
 	int			assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
 	u64			tags[X86_PMC_IDX_MAX];
 
maintainer
#6 lore raw

PATCH 1/1 perf/x86: Fix n_metric for cancelled txn

In reply to Liang, Kan
Quoted 1 line · Kan

---
Subject: perf/x86: Fix n_metric for cancelled txn
From: Peter Zijlstra <[email protected]>
Date: Mon Oct 5 10:10:24 CEST 2020

When a group that has TopDown members is failed to be scheduled, any later TopDown groups will not return valid values.

Here is an example.

A background perf that occupies all the GP counters and the fixed counter 1.
$perf stat -e "{cycles,cycles,cycles,cycles,cycles,cycles,cycles,
cycles,cycles}:D" -a

A user monitors a TopDown group. It works well, because the fixed counter 3 and the PERF_METRICS are available.
$perf stat -x, --topdown -- ./workload
retiring,bad speculation,frontend bound,backend bound, 18.0,16.1,40.4,25.5,

Then the user tries to monitor a group that has TopDown members. Because of the cycles event, the group is failed to be scheduled.
$perf stat -x, -e '{slots,topdown-retiring,topdown-be-bound,

                     topdown-fe-bound,topdown-bad-spec,cycles}'
                     -- ./workload
    <not counted>,,slots,0,0.00,,
    <not counted>,,topdown-retiring,0,0.00,,
    <not counted>,,topdown-be-bound,0,0.00,,
    <not counted>,,topdown-fe-bound,0,0.00,,
    <not counted>,,topdown-bad-spec,0,0.00,,
    <not counted>,,cycles,0,0.00,,

The user tries to monitor a TopDown group again. It doesn't work anymore.
$perf stat -x, --topdown -- ./workload

    ,,,,,

In a txn, cancel_txn() is to truncate the event_list for a canceled group and update the number of events added in this transaction. However, the number of TopDown events added in this transaction is not updated. The kernel will probably fail to add new Topdown events.

Fixes 7b2c05a15d29 ("perf/x86/intel: Generic support for hardware TopDown metrics")
---
 arch/x86/events/core.c       |    3 +++
 arch/x86/events/perf_event.h |    1 +
 2 files changed, 4 insertions(+)
arch/x86/events/core.c +3 -0
@@ -1066,6 +1066,7 @@ static int add_nr_metric_event(struct cp
 		if (cpuc->n_metric == INTEL_TD_METRIC_NUM)
 			return -EINVAL;
 		cpuc->n_metric++;
+		cpuc->n_txn_metric++;
 	}
 
 	return 0;
@@ -2065,6 +2066,7 @@ static void x86_pmu_start_txn(struct pmu
 	perf_pmu_disable(pmu);
 	__this_cpu_write(cpu_hw_events.n_txn, 0);
 	__this_cpu_write(cpu_hw_events.n_txn_pair, 0);
+	__this_cpu_write(cpu_hw_events.n_txn_metric, 0);
 }
 
 /*
@@ -2091,6 +2093,7 @@ static void x86_pmu_cancel_txn(struct pm
 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
 	__this_cpu_sub(cpu_hw_events.n_pair, __this_cpu_read(cpu_hw_events.n_txn_pair));
+	__this_cpu_sub(cpu_hw_events.n_metric, __this_cpu_read(cpu_hw_events.n_txn_metric));
 	perf_pmu_enable(pmu);
 }
 
arch/x86/events/perf_event.h +1 -0
@@ -236,6 +236,7 @@ struct cpu_hw_events {
 	int			n_txn;    /* the # last events in the below arrays;
 					     added in the current transaction */
 	int			n_txn_pair;
+	int			n_txn_metric;
 	int			assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
 	u64			tags[X86_PMC_IDX_MAX];
 
merge bot Applied
#5 lore raw

[tip: perf/core] perf/x86: Fix n_pair for cancelled txn

In reply to Peter Zijlstra
The following commit has been merged into the perf/core branch of tip:

The following commit has been merged into the perf/core branch of tip:

Commit-ID:     871a93b0aad65a7f44ee25f2d17932ef6d559850
Gitweb:        https://git.kernel.org/tip/871a93b0aad65a7f44ee25f2d17932ef6d559850
Author:        Peter Zijlstra <[email protected]>
AuthorDate:    Mon, 05 Oct 2020 10:09:06 +02:00
Committer:     Peter Zijlstra <[email protected]>
CommitterDate: Tue, 06 Oct 2020 15:18:17 +02:00

perf/x86: Fix n_pair for cancelled txn

Kan reported that n_metric gets corrupted for cancelled transactions; a similar issue exists for n_pair for AMD's Large Increment thing.

The problem was confirmed and confirmed fixed by Kim using:

sudo perf stat -e "{cycles,cycles,cycles,cycles}:D" -a sleep 10 &

  # should succeed:
  sudo perf stat -e "{fp_ret_sse_avx_ops.all}:D" -a workload
  # should fail:
  sudo perf stat -e "{fp_ret_sse_avx_ops.all,fp_ret_sse_avx_ops.all,cycles}:D" -a workload
  # previously failed, now succeeds with this patch:
  sudo perf stat -e "{fp_ret_sse_avx_ops.all}:D" -a workload
Fixes 5738891229a2 ("perf/x86/amd: Add support for Large Increment per Cycle Events")
---
 arch/x86/events/core.c       | 6 +++++-
 arch/x86/events/perf_event.h | 1 +
 2 files changed, 6 insertions(+), 1 deletion(-)
arch/x86/events/core.c +5 -1
index cb5cfef..a7248a3 100644
@@ -1064,8 +1064,10 @@ static int collect_event(struct cpu_hw_events *cpuc, struct perf_event *event,
 		return -EINVAL;
 
 	cpuc->event_list[n] = event;
-	if (is_counter_pair(&event->hw))
+	if (is_counter_pair(&event->hw)) {
 		cpuc->n_pair++;
+		cpuc->n_txn_pair++;
+	}
 
 	return 0;
 }
@@ -2006,6 +2008,7 @@ static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
 
 	perf_pmu_disable(pmu);
 	__this_cpu_write(cpu_hw_events.n_txn, 0);
+	__this_cpu_write(cpu_hw_events.n_txn_pair, 0);
 }
 
 /*
@@ -2031,6 +2034,7 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
 	 */
 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
+	__this_cpu_sub(cpu_hw_events.n_pair, __this_cpu_read(cpu_hw_events.n_txn_pair));
 	perf_pmu_enable(pmu);
 }
 
arch/x86/events/perf_event.h +1 -0
index 3454424..93e56d7 100644
@@ -235,6 +235,7 @@ struct cpu_hw_events {
 					     they've never been enabled yet */
 	int			n_txn;    /* the # last events in the below arrays;
 					     added in the current transaction */
+	int			n_txn_pair;
 	int			assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
 	u64			tags[X86_PMC_IDX_MAX];
 
merge bot Applied
#7 lore raw

[tip: perf/core] perf/x86: Fix n_metric for cancelled txn

In reply to Peter Zijlstra
The following commit has been merged into the perf/core branch of tip:

The following commit has been merged into the perf/core branch of tip:

Commit-ID:     3dbde69575637658d2094ee4416c21bc22eb89fe
Gitweb:        https://git.kernel.org/tip/3dbde69575637658d2094ee4416c21bc22eb89fe
Author:        Peter Zijlstra <[email protected]>
AuthorDate:    Mon, 05 Oct 2020 10:10:24 +02:00
Committer:     Peter Zijlstra <[email protected]>
CommitterDate: Tue, 06 Oct 2020 15:18:17 +02:00

perf/x86: Fix n_metric for cancelled txn

When a group that has TopDown members is failed to be scheduled, any later TopDown groups will not return valid values.

Here is an example.

A background perf that occupies all the GP counters and the fixed counter 1.
$perf stat -e "{cycles,cycles,cycles,cycles,cycles,cycles,cycles,
cycles,cycles}:D" -a

A user monitors a TopDown group. It works well, because the fixed counter 3 and the PERF_METRICS are available.
$perf stat -x, --topdown -- ./workload
retiring,bad speculation,frontend bound,backend bound, 18.0,16.1,40.4,25.5,

Then the user tries to monitor a group that has TopDown members. Because of the cycles event, the group is failed to be scheduled.
$perf stat -x, -e '{slots,topdown-retiring,topdown-be-bound,

                     topdown-fe-bound,topdown-bad-spec,cycles}'
                     -- ./workload
    <not counted>,,slots,0,0.00,,
    <not counted>,,topdown-retiring,0,0.00,,
    <not counted>,,topdown-be-bound,0,0.00,,
    <not counted>,,topdown-fe-bound,0,0.00,,
    <not counted>,,topdown-bad-spec,0,0.00,,
    <not counted>,,cycles,0,0.00,,

The user tries to monitor a TopDown group again. It doesn't work anymore.
$perf stat -x, --topdown -- ./workload

    ,,,,,

In a txn, cancel_txn() is to truncate the event_list for a canceled group and update the number of events added in this transaction. However, the number of TopDown events added in this transaction is not updated. The kernel will probably fail to add new Topdown events.

Fixes 7b2c05a15d29 ("perf/x86/intel: Generic support for hardware TopDown metrics")
---
 arch/x86/events/core.c       | 3 +++
 arch/x86/events/perf_event.h | 1 +
 2 files changed, 4 insertions(+)
arch/x86/events/core.c +3 -0
index a7248a3..7b802a7 100644
@@ -1041,6 +1041,7 @@ static int add_nr_metric_event(struct cpu_hw_events *cpuc,
 		if (cpuc->n_metric == INTEL_TD_METRIC_NUM)
 			return -EINVAL;
 		cpuc->n_metric++;
+		cpuc->n_txn_metric++;
 	}
 
 	return 0;
@@ -2009,6 +2010,7 @@ static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
 	perf_pmu_disable(pmu);
 	__this_cpu_write(cpu_hw_events.n_txn, 0);
 	__this_cpu_write(cpu_hw_events.n_txn_pair, 0);
+	__this_cpu_write(cpu_hw_events.n_txn_metric, 0);
 }
 
 /*
@@ -2035,6 +2037,7 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
 	__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
 	__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
 	__this_cpu_sub(cpu_hw_events.n_pair, __this_cpu_read(cpu_hw_events.n_txn_pair));
+	__this_cpu_sub(cpu_hw_events.n_metric, __this_cpu_read(cpu_hw_events.n_txn_metric));
 	perf_pmu_enable(pmu);
 }
 
arch/x86/events/perf_event.h +1 -0
index 93e56d7..ee2b9b9 100644
@@ -236,6 +236,7 @@ struct cpu_hw_events {
 	int			n_txn;    /* the # last events in the below arrays;
 					     added in the current transaction */
 	int			n_txn_pair;
+	int			n_txn_metric;
 	int			assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
 	u64			tags[X86_PMC_IDX_MAX];