Skip to contentSkip to conversation

perf: arm_pmuv3: Avoid assigning fixed cycle counter with threshold

PATCH In review Rob Herring (Arm) 3 participants 6 messages Jun 11, 2024 – Jun 24, 2024 Arm PMU profiling and debugging

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf: arm_pmuv3: Avoid assigning fixed cycle counter with threshold

If the user has requested a counting threshold for the CPU cycles event, then the fixed cycle counter can't be assigned as it lacks threshold support. Currently, the thresholds will work or not randomly depending on which counter the event is assigned.

While using thresholds for CPU cycles doesn't make much sense, it can be useful for testing purposes.

Fixes 816c26754447 ("arm64: perf: Add support for event counting threshold")
---
 drivers/perf/arm_pmuv3.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/perf/arm_pmuv3.c +2 -1
index 23fa6c5da82c..2612be29ee23 100644
@@ -939,9 +939,10 @@ static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc,
 	struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
 	struct hw_perf_event *hwc = &event->hw;
 	unsigned long evtype = hwc->config_base & ARMV8_PMU_EVTYPE_EVENT;
+	bool has_threshold = !!(hwc->config_base & ARMV8_PMU_EVTYPE_TH);
 
 	/* Always prefer to place a cycle counter into the cycle counter. */
-	if (evtype == ARMV8_PMUV3_PERFCTR_CPU_CYCLES) {
+	if ((evtype == ARMV8_PMUV3_PERFCTR_CPU_CYCLES) && !has_threshold) {
 		if (!test_and_set_bit(ARMV8_IDX_CYCLE_COUNTER, cpuc->used_mask))
 			return ARMV8_IDX_CYCLE_COUNTER;
 		else if (armv8pmu_event_is_64bit(event) &&
Signature · 1 line
2.43.0
#1 lore raw
Quoted 19 lines · Rob Herring (Arm)

If the user has requested a counting threshold for the CPU cycles event, then the fixed cycle counter can't be assigned as it lacks threshold support. Currently, the thresholds will work or not randomly depending on which counter the event is assigned.

While using thresholds for CPU cycles doesn't make much sense, it can be useful for testing purposes.

Fixes 816c26754447 ("arm64: perf: Add support for event counting threshold")
---
 drivers/perf/arm_pmuv3.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/perf/arm_pmuv3.c +0 -0
index 23fa6c5da82c..2612be29ee23 100644

drivers/perf/arm_pmuv3.c static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc, · line 942

Quoted 4 lines
drivers/perf/arm_pmuv3.c +1 -0
@@ -939,9 +939,10 @@ static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc,
 	struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
 	struct hw_perf_event *hwc = &event->hw;
 	unsigned long evtype = hwc->config_base & ARMV8_PMU_EVTYPE_EVENT;
+	bool has_threshold = !!(hwc->config_base & ARMV8_PMU_EVTYPE_TH);

I was going to say doesn't it need to be (ARMV8_PMU_EVTYPE_TH | ARMV8_PMU_EVTYPE_TC) for it to give the same results as the hardware. But then I saw we only enable it if TH != 0, even if TC is set. And now I'm wondering if I inadvertently disabled a useful combination of options.

The Arm ARM says it's only completely disabled when both TC and TH are 0.

Quoted 7 lines
drivers/perf/arm_pmuv3.c +1 -1
 	/* Always prefer to place a cycle counter into the cycle counter. */
-	if (evtype == ARMV8_PMUV3_PERFCTR_CPU_CYCLES) {
+	if ((evtype == ARMV8_PMUV3_PERFCTR_CPU_CYCLES) && !has_threshold) {
 		if (!test_and_set_bit(ARMV8_IDX_CYCLE_COUNTER, cpuc->used_mask))
 			return ARMV8_IDX_CYCLE_COUNTER;
 		else if (armv8pmu_event_is_64bit(event) &&
maintainer
#4 lore raw
In reply to Rob Herring (Arm)
Quoted 19 lines · Rob Herring (Arm)

If the user has requested a counting threshold for the CPU cycles event, then the fixed cycle counter can't be assigned as it lacks threshold support. Currently, the thresholds will work or not randomly depending on which counter the event is assigned.

While using thresholds for CPU cycles doesn't make much sense, it can be useful for testing purposes.

Fixes 816c26754447 ("arm64: perf: Add support for event counting threshold")
---
 drivers/perf/arm_pmuv3.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/perf/arm_pmuv3.c +0 -0
index 23fa6c5da82c..2612be29ee23 100644

drivers/perf/arm_pmuv3.c static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc, · line 942

Quoted 4 lines
drivers/perf/arm_pmuv3.c +1 -0
@@ -939,9 +939,10 @@ static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc,
 	struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
 	struct hw_perf_event *hwc = &event->hw;
 	unsigned long evtype = hwc->config_base & ARMV8_PMU_EVTYPE_EVENT;
+	bool has_threshold = !!(hwc->config_base & ARMV8_PMU_EVTYPE_TH);

Just a nit, but I don't think you need the '!!' here.

Will

#2 lore raw
In reply to James Clark
Quoted 32 lines · James Clark
Quoted 19 lines · Rob Herring (Arm)

If the user has requested a counting threshold for the CPU cycles event, then the fixed cycle counter can't be assigned as it lacks threshold support. Currently, the thresholds will work or not randomly depending on which counter the event is assigned.

While using thresholds for CPU cycles doesn't make much sense, it can be useful for testing purposes.

Fixes 816c26754447 ("arm64: perf: Add support for event counting threshold")
---
 drivers/perf/arm_pmuv3.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/perf/arm_pmuv3.c +0 -0
index 23fa6c5da82c..2612be29ee23 100644

drivers/perf/arm_pmuv3.c static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc, · line 942

Quoted 4 lines
drivers/perf/arm_pmuv3.c +1 -0
@@ -939,9 +939,10 @@ static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc,
 	struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
 	struct hw_perf_event *hwc = &event->hw;
 	unsigned long evtype = hwc->config_base & ARMV8_PMU_EVTYPE_EVENT;
+	bool has_threshold = !!(hwc->config_base & ARMV8_PMU_EVTYPE_TH);

I was going to say doesn't it need to be (ARMV8_PMU_EVTYPE_TH | ARMV8_PMU_EVTYPE_TC) for it to give the same results as the hardware. But then I saw we only enable it if TH != 0, even if TC is set. And now I'm wondering if I inadvertently disabled a useful combination of options.

Quoted 2 lines

The Arm ARM says it's only completely disabled when both TC and TH are 0.

If it's easy it might be worth adding a helper function for has_threshold() that's used in both places. That way if or when this issue gets fixed up it doesn't break here.

Quoted 7 lines
Quoted 7 lines
drivers/perf/arm_pmuv3.c +1 -1
 	/* Always prefer to place a cycle counter into the cycle counter. */
-	if (evtype == ARMV8_PMUV3_PERFCTR_CPU_CYCLES) {
+	if ((evtype == ARMV8_PMUV3_PERFCTR_CPU_CYCLES) && !has_threshold) {
 		if (!test_and_set_bit(ARMV8_IDX_CYCLE_COUNTER, cpuc->used_mask))
 			return ARMV8_IDX_CYCLE_COUNTER;
 		else if (armv8pmu_event_is_64bit(event) &&
#3 lore raw
Quoted 36 lines · James Clark
Quoted 32 lines · James Clark
Quoted 19 lines · Rob Herring (Arm)

If the user has requested a counting threshold for the CPU cycles event, then the fixed cycle counter can't be assigned as it lacks threshold support. Currently, the thresholds will work or not randomly depending on which counter the event is assigned.

While using thresholds for CPU cycles doesn't make much sense, it can be useful for testing purposes.

Fixes 816c26754447 ("arm64: perf: Add support for event counting threshold")
---
 drivers/perf/arm_pmuv3.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/perf/arm_pmuv3.c +0 -0
index 23fa6c5da82c..2612be29ee23 100644

drivers/perf/arm_pmuv3.c static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc, · line 942

Quoted 4 lines
drivers/perf/arm_pmuv3.c +1 -0
@@ -939,9 +939,10 @@ static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc,
     struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
     struct hw_perf_event *hwc = &event->hw;
     unsigned long evtype = hwc->config_base & ARMV8_PMU_EVTYPE_EVENT;
+    bool has_threshold = !!(hwc->config_base & ARMV8_PMU_EVTYPE_TH);

I was going to say doesn't it need to be (ARMV8_PMU_EVTYPE_TH | ARMV8_PMU_EVTYPE_TC) for it to give the same results as the hardware. But then I saw we only enable it if TH != 0, even if TC is set. And now I'm wondering if I inadvertently disabled a useful combination of options.

Quoted 6 lines
Quoted 2 lines

The Arm ARM says it's only completely disabled when both TC and TH are 0.

If it's easy it might be worth adding a helper function for has_threshold() that's used in both places. That way if or when this issue gets fixed up it doesn't break here.

The other place being in armv8pmu_set_event_filter()? A helper doesn't help there because that looks at the attr value, not config_base.

Rob

#5 lore raw
In reply to Will Deacon
Quoted 16 lines · Will Deacon
Quoted 14 lines · Rob Herring (Arm)

If the user has requested a counting threshold for the CPU cycles event, then the fixed cycle counter can't be assigned as it lacks threshold support. Currently, the thresholds will work or not randomly depending on which counter the event is assigned.

While using thresholds for CPU cycles doesn't make much sense, it can be useful for testing purposes.

Fixes 816c26754447 ("arm64: perf: Add support for event counting threshold")
---
 drivers/perf/arm_pmuv3.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
Quoted 11 lines

drivers/perf/arm_pmuv3.c static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc, · line 942

Quoted 9 lines
drivers/perf/arm_pmuv3.c +1 -0
index 23fa6c5da82c..2612be29ee23 100644
@@ -939,9 +939,10 @@ static int armv8pmu_get_event_idx(struct pmu_hw_events *cpuc,
      struct arm_pmu *cpu_pmu = to_arm_pmu(event->pmu);
      struct hw_perf_event *hwc = &event->hw;
      unsigned long evtype = hwc->config_base & ARMV8_PMU_EVTYPE_EVENT;
+     bool has_threshold = !!(hwc->config_base & ARMV8_PMU_EVTYPE_TH);

Just a nit, but I don't think you need the '!!' here.

Right, I guess since bool is a first class type in C9X we don't have to worry about truncation. Old habits...

Rob