Skip to contentSkip to conversation

perf/x86/uncore: Overflow handling enhancements

PATCH In review Sandipan Das 3 participants 10 messages Apr 18, 2025

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

v2 cover perf/x86/uncore: Overflow handling enhancements

Uncore counters on AMD processors either roll over or saturate on overflow and the amd-uncore driver has no way of knowing if multiple overflows have occurred between two successive pmu->read() requests for an event. This makes the user-visible counter values inaccurate. Solve this by periodically initiating pmu->read() in order to keep prev_count up-to-date. The approach follows Intel's precedent in handling uncore counters.

Previous versions can be found at:
v1: https://lore.kernel.org/all/[email protected]/

Changes in v2:
- Add a Fixes tag for the first patch.
- Change the hrtimer mode for both the Intel and the AMD uncore drivers based on Peter's suggestion.
- Fix an issue in the UMC workaround where prev_count is not zeroed out after a counter is reset.

Sandipan Das (5):
perf/x86/amd/uncore: Remove unused member from amd_uncore_ctx perf/x86/intel/uncore: Use HRTIMER_MODE_HARD for detecting overflows perf/x86/amd/uncore: Use hrtimer for handling overflows perf/x86/amd/uncore: Add parameter to configure hrtimer perf/x86/amd/uncore: Prevent UMC counters from saturating

 arch/x86/events/amd/uncore.c   | 103 ++++++++++++++++++++++++++++++++-
 arch/x86/events/intel/uncore.c |  12 +---
 2 files changed, 103 insertions(+), 12 deletions(-)
Signature · 1 line
2.43.0
#1 lore raw

PATCH v2 1/5 perf/x86/amd/uncore: Remove unused member from amd_uncore_ctx

Remove unused member "node" from struct amd_uncore_ctx.

Fixes d6389d3ccc13 ("perf/x86/amd/uncore: Refactor uncore management")
---
 arch/x86/events/amd/uncore.c | 1 -
 1 file changed, 1 deletion(-)
arch/x86/events/amd/uncore.c +0 -1
index 49c26ce2b115..010024f09f2c 100644
@@ -38,7 +38,6 @@ struct amd_uncore_ctx {
 	int refcnt;
 	int cpu;
 	struct perf_event **events;
-	struct hlist_node node;
 };
 
 struct amd_uncore_pmu {
Signature · 1 line
2.43.0
#2 lore raw

PATCH v2 2/5 perf/x86/intel/uncore: Use HRTIMER_MODE_HARD for detecting overflows

In reply to Sandipan Das

Hrtimer handlers can be deferred to softirq context and affect timely detection of counter overflows. Hence switch to HRTIMER_MODE_HARD.

Disabling and re-enabling IRQs in the hrtimer handler is not required as pmu->start() and pmu->stop() can no longer intervene while updating event->hw.prev_count.

---
 arch/x86/events/intel/uncore.c | 12 ++----------
 1 file changed, 2 insertions(+), 10 deletions(-)
arch/x86/events/intel/uncore.c +2 -10
index a34e50fc4a8f..5811e172f721 100644
@@ -305,17 +305,11 @@ static enum hrtimer_restart uncore_pmu_hrtimer(struct hrtimer *hrtimer)
 {
 	struct intel_uncore_box *box;
 	struct perf_event *event;
-	unsigned long flags;
 	int bit;
 
 	box = container_of(hrtimer, struct intel_uncore_box, hrtimer);
 	if (!box->n_active || box->cpu != smp_processor_id())
 		return HRTIMER_NORESTART;
-	/*
-	 * disable local interrupt to prevent uncore_pmu_event_start/stop
-	 * to interrupt the update process
-	 */
-	local_irq_save(flags);
 
 	/*
 	 * handle boxes with an active event list as opposed to active
@@ -328,8 +322,6 @@ static enum hrtimer_restart uncore_pmu_hrtimer(struct hrtimer *hrtimer)
 	for_each_set_bit(bit, box->active_mask, UNCORE_PMC_IDX_MAX)
 		uncore_perf_event_update(box, box->events[bit]);
 
-	local_irq_restore(flags);
-
 	hrtimer_forward_now(hrtimer, ns_to_ktime(box->hrtimer_duration));
 	return HRTIMER_RESTART;
 }
@@ -337,7 +329,7 @@ static enum hrtimer_restart uncore_pmu_hrtimer(struct hrtimer *hrtimer)
 void uncore_pmu_start_hrtimer(struct intel_uncore_box *box)
 {
 	hrtimer_start(&box->hrtimer, ns_to_ktime(box->hrtimer_duration),
-		      HRTIMER_MODE_REL_PINNED);
+		      HRTIMER_MODE_REL_PINNED_HARD);
 }
 
 void uncore_pmu_cancel_hrtimer(struct intel_uncore_box *box)
@@ -347,7 +339,7 @@ void uncore_pmu_cancel_hrtimer(struct intel_uncore_box *box)
 
 static void uncore_pmu_init_hrtimer(struct intel_uncore_box *box)
 {
-	hrtimer_setup(&box->hrtimer, uncore_pmu_hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+	hrtimer_setup(&box->hrtimer, uncore_pmu_hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
 }
 
 static struct intel_uncore_box *uncore_alloc_box(struct intel_uncore_type *type,
Signature · 1 line
2.43.0
#3 lore raw

PATCH v2 3/5 perf/x86/amd/uncore: Use hrtimer for handling overflows

In reply to Sandipan Das

Uncore counters do not provide mechanisms like interrupts to report overflows and the accumulated user-visible count is incorrect if there is more than one overflow between two successive read requests for the same event because the value of prev_count goes out-of-date for calculating the correct delta.

To avoid this, start a hrtimer to periodically initiate a pmu->read() of the active counters for keeping prev_count up-to-date. It should be noted that the hrtimer duration should be lesser than the shortest time it takes for a counter to overflow for this approach to be effective.

---
 arch/x86/events/amd/uncore.c | 63 ++++++++++++++++++++++++++++++++++++
 1 file changed, 63 insertions(+)
arch/x86/events/amd/uncore.c +63 -0
index 010024f09f2c..e09bfbb4a4cd 100644
@@ -21,6 +21,7 @@
 #define NUM_COUNTERS_NB		4
 #define NUM_COUNTERS_L2		4
 #define NUM_COUNTERS_L3		6
+#define NUM_COUNTERS_MAX	64
 
 #define RDPMC_BASE_NB		6
 #define RDPMC_BASE_LLC		10
@@ -38,6 +39,10 @@ struct amd_uncore_ctx {
 	int refcnt;
 	int cpu;
 	struct perf_event **events;
+	unsigned long active_mask[BITS_TO_LONGS(NUM_COUNTERS_MAX)];
+	int nr_active;
+	struct hrtimer hrtimer;
+	u64 hrtimer_duration;
 };
 
 struct amd_uncore_pmu {
@@ -87,6 +92,42 @@ static struct amd_uncore_pmu *event_to_amd_uncore_pmu(struct perf_event *event)
 	return container_of(event->pmu, struct amd_uncore_pmu, pmu);
 }
 
+static enum hrtimer_restart amd_uncore_hrtimer(struct hrtimer *hrtimer)
+{
+	struct amd_uncore_ctx *ctx;
+	struct perf_event *event;
+	int bit;
+
+	ctx = container_of(hrtimer, struct amd_uncore_ctx, hrtimer);
+
+	if (!ctx->nr_active || ctx->cpu != smp_processor_id())
+		return HRTIMER_NORESTART;
+
+	for_each_set_bit(bit, ctx->active_mask, NUM_COUNTERS_MAX) {
+		event = ctx->events[bit];
+		event->pmu->read(event);
+	}
+
+	hrtimer_forward_now(hrtimer, ns_to_ktime(ctx->hrtimer_duration));
+	return HRTIMER_RESTART;
+}
+
+static void amd_uncore_start_hrtimer(struct amd_uncore_ctx *ctx)
+{
+	hrtimer_start(&ctx->hrtimer, ns_to_ktime(ctx->hrtimer_duration),
+		      HRTIMER_MODE_REL_PINNED_HARD);
+}
+
+static void amd_uncore_cancel_hrtimer(struct amd_uncore_ctx *ctx)
+{
+	hrtimer_cancel(&ctx->hrtimer);
+}
+
+static void amd_uncore_init_hrtimer(struct amd_uncore_ctx *ctx)
+{
+	hrtimer_setup(&ctx->hrtimer, amd_uncore_hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
+}
+
 static void amd_uncore_read(struct perf_event *event)
 {
 	struct hw_perf_event *hwc = &event->hw;
@@ -117,18 +158,26 @@ static void amd_uncore_read(struct perf_event *event)
 
 static void amd_uncore_start(struct perf_event *event, int flags)
 {
+	struct amd_uncore_pmu *pmu = event_to_amd_uncore_pmu(event);
+	struct amd_uncore_ctx *ctx = *per_cpu_ptr(pmu->ctx, event->cpu);
 	struct hw_perf_event *hwc = &event->hw;
 
+	if (!ctx->nr_active++)
+		amd_uncore_start_hrtimer(ctx);
+
 	if (flags & PERF_EF_RELOAD)
 		wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
 
 	hwc->state = 0;
+	__set_bit(hwc->idx, ctx->active_mask);
 	wrmsrl(hwc->config_base, (hwc->config | ARCH_PERFMON_EVENTSEL_ENABLE));
 	perf_event_update_userpage(event);
 }
 
 static void amd_uncore_stop(struct perf_event *event, int flags)
 {
+	struct amd_uncore_pmu *pmu = event_to_amd_uncore_pmu(event);
+	struct amd_uncore_ctx *ctx = *per_cpu_ptr(pmu->ctx, event->cpu);
 	struct hw_perf_event *hwc = &event->hw;
 
 	wrmsrl(hwc->config_base, hwc->config);
@@ -138,6 +187,11 @@ static void amd_uncore_stop(struct perf_event *event, int flags)
 		event->pmu->read(event);
 		hwc->state |= PERF_HES_UPTODATE;
 	}
+
+	if (!--ctx->nr_active)
+		amd_uncore_cancel_hrtimer(ctx);
+
+	__clear_bit(hwc->idx, ctx->active_mask);
 }
 
 static int amd_uncore_add(struct perf_event *event, int flags)
@@ -490,6 +544,9 @@ static int amd_uncore_ctx_init(struct amd_uncore *uncore, unsigned int cpu)
 				goto fail;
 			}
 
+			amd_uncore_init_hrtimer(curr);
+			curr->hrtimer_duration = 60LL * NSEC_PER_SEC;
+
 			cpumask_set_cpu(cpu, &pmu->active_mask);
 		}
 
@@ -879,12 +936,18 @@ static int amd_uncore_umc_event_init(struct perf_event *event)
 
 static void amd_uncore_umc_start(struct perf_event *event, int flags)
 {
+	struct amd_uncore_pmu *pmu = event_to_amd_uncore_pmu(event);
+	struct amd_uncore_ctx *ctx = *per_cpu_ptr(pmu->ctx, event->cpu);
 	struct hw_perf_event *hwc = &event->hw;
 
+	if (!ctx->nr_active++)
+		amd_uncore_start_hrtimer(ctx);
+
 	if (flags & PERF_EF_RELOAD)
 		wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
 
 	hwc->state = 0;
+	__set_bit(hwc->idx, ctx->active_mask);
 	wrmsrl(hwc->config_base, (hwc->config | AMD64_PERFMON_V2_ENABLE_UMC));
 	perf_event_update_userpage(event);
 }
Signature · 1 line
2.43.0
#6 lore raw

PATCH v2 4/5 perf/x86/amd/uncore: Add parameter to configure hrtimer

In reply to Sandipan Das

Introduce a module parameter for configuring the hrtimer duration in milliseconds. The default duration is 60000 milliseconds and the intent is to allow users to customize it to suit jitter tolerances. It should be noted that a longer duration will reduce jitter but affect accuracy if the programmed events cause the counters to overflow multiple times in a single interval.

---
 arch/x86/events/amd/uncore.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
arch/x86/events/amd/uncore.c +5 -1
index e09bfbb4a4cd..70e0af36c378 100644
@@ -87,6 +87,10 @@ struct amd_uncore {
 
 static struct amd_uncore uncores[UNCORE_TYPE_MAX];
 
+/* Interval for hrtimer, defaults to 60000 milliseconds */
+static unsigned int update_interval = 60 * MSEC_PER_SEC;
+module_param(update_interval, uint, 0444);
+
 static struct amd_uncore_pmu *event_to_amd_uncore_pmu(struct perf_event *event)
 {
 	return container_of(event->pmu, struct amd_uncore_pmu, pmu);
@@ -545,7 +549,7 @@ static int amd_uncore_ctx_init(struct amd_uncore *uncore, unsigned int cpu)
 			}
 
 			amd_uncore_init_hrtimer(curr);
-			curr->hrtimer_duration = 60LL * NSEC_PER_SEC;
+			curr->hrtimer_duration = (u64)update_interval * NSEC_PER_MSEC;
 
 			cpumask_set_cpu(cpu, &pmu->active_mask);
 		}
Signature · 1 line
2.43.0
#7 lore raw

PATCH v2 5/5 perf/x86/amd/uncore: Prevent UMC counters from saturating

In reply to Sandipan Das

Unlike L3 and DF counters, UMC counters (PERF_CTRs) set the Overflow bit (bit 48) and saturate on overflow. A subsequent pmu->read() of the event reports an incorrect accumulated count as there is no difference between the previous and the current values of the counter.

To avoid this, inspect the current counter value and proactively reset the corresponding PERF_CTR register on every pmu->read(). Combined with the periodic reads initiated by the hrtimer, the counters never get a chance saturate but the resolution reduces to 47 bits.

Fixes 25e56847821f ("perf/x86/amd/uncore: Add memory controller support")
---
 arch/x86/events/amd/uncore.c | 35 ++++++++++++++++++++++++++++++++++-
 1 file changed, 34 insertions(+), 1 deletion(-)
arch/x86/events/amd/uncore.c +34 -1
index 70e0af36c378..d328de166481 100644
@@ -956,6 +956,39 @@ static void amd_uncore_umc_start(struct perf_event *event, int flags)
 	perf_event_update_userpage(event);
 }
 
+static void amd_uncore_umc_read(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	u64 prev, new, shift;
+	s64 delta;
+
+	shift = COUNTER_SHIFT + 1;
+	prev = local64_read(&hwc->prev_count);
+
+	/*
+	 * UMC counters do not have RDPMC assignments. Read counts directly
+	 * from the corresponding PERF_CTR.
+	 */
+	rdmsrl(hwc->event_base, new);
+
+	/*
+	 * Unlike the other uncore counters, UMC counters saturate and set the
+	 * Overflow bit (bit 48) on overflow. Since they do not roll over,
+	 * proactively reset the corresponding PERF_CTR when bit 47 is set so
+	 * that the counter never gets a chance to saturate.
+	 */
+	if (new & BIT_ULL(63 - COUNTER_SHIFT)) {
+		wrmsrl(hwc->event_base, 0);
+		local64_set(&hwc->prev_count, 0);
+	} else {
+		local64_set(&hwc->prev_count, new);
+	}
+
+	delta = (new << shift) - (prev << shift);
+	delta >>= shift;
+	local64_add(delta, &event->count);
+}
+
 static
 void amd_uncore_umc_ctx_scan(struct amd_uncore *uncore, unsigned int cpu)
 {
@@ -1034,7 +1067,7 @@ int amd_uncore_umc_ctx_init(struct amd_uncore *uncore, unsigned int cpu)
 				.del		= amd_uncore_del,
 				.start		= amd_uncore_umc_start,
 				.stop		= amd_uncore_stop,
-				.read		= amd_uncore_read,
+				.read		= amd_uncore_umc_read,
 				.capabilities	= PERF_PMU_CAP_NO_EXCLUDE | PERF_PMU_CAP_NO_INTERRUPT,
 				.module		= THIS_MODULE,
 			};
Signature · 1 line
2.43.0
#4 lore raw

Re: [PATCH v2 3/5] perf/x86/amd/uncore: Use hrtimer for handling overflows

In reply to Sandipan Das
Quoted 7 lines · Sandipan Das

Uncore counters do not provide mechanisms like interrupts to report overflows and the accumulated user-visible count is incorrect if there is more than one overflow between two successive read requests for the same event because the value of prev_count goes out-of-date for calculating the correct delta.

Quoted 5 lines

To avoid this, start a hrtimer to periodically initiate a pmu->read() of the active counters for keeping prev_count up-to-date. It should be noted that the hrtimer duration should be lesser than the shortest time it takes for a counter to overflow for this approach to be effective.

The problem I see is that the number of uncore PMU varies a lot based on the CPU model, in particular due to the L3 PMU.
Is there a timer armed per CCX or only a global one that will generate IPI to all other CPUs?

Quoted 142 lines
---
 arch/x86/events/amd/uncore.c | 63 ++++++++++++++++++++++++++++++++++++
 1 file changed, 63 insertions(+)
arch/x86/events/amd/uncore.c +63 -1
index 010024f09f2c..e09bfbb4a4cd 100644
@@ -21,6 +21,7 @@
 #define NUM_COUNTERS_NB                4
 #define NUM_COUNTERS_L2                4
 #define NUM_COUNTERS_L3                6
+#define NUM_COUNTERS_MAX       64
 
 #define RDPMC_BASE_NB          6
 #define RDPMC_BASE_LLC         10
@@ -38,6 +39,10 @@ struct amd_uncore_ctx {
        int refcnt;
        int cpu;
        struct perf_event **events;
+       unsigned long active_mask[BITS_TO_LONGS(NUM_COUNTERS_MAX)];
+       int nr_active;
+       struct hrtimer hrtimer;
+       u64 hrtimer_duration;
 };
 
 struct amd_uncore_pmu {
@@ -87,6 +92,42 @@ static struct amd_uncore_pmu *event_to_amd_uncore_pmu(struct perf_event *event)
        return container_of(event->pmu, struct amd_uncore_pmu, pmu);
 }
 
+static enum hrtimer_restart amd_uncore_hrtimer(struct hrtimer *hrtimer)
+{
+       struct amd_uncore_ctx *ctx;
+       struct perf_event *event;
+       int bit;
+
+       ctx = container_of(hrtimer, struct amd_uncore_ctx, hrtimer);
+
+       if (!ctx->nr_active || ctx->cpu != smp_processor_id())
+               return HRTIMER_NORESTART;
+
+       for_each_set_bit(bit, ctx->active_mask, NUM_COUNTERS_MAX) {
+               event = ctx->events[bit];
+               event->pmu->read(event);
+       }
+
+       hrtimer_forward_now(hrtimer, ns_to_ktime(ctx->hrtimer_duration));
+       return HRTIMER_RESTART;
+}
+
+static void amd_uncore_start_hrtimer(struct amd_uncore_ctx *ctx)
+{
+       hrtimer_start(&ctx->hrtimer, ns_to_ktime(ctx->hrtimer_duration),
+                     HRTIMER_MODE_REL_PINNED_HARD);
+}
+
+static void amd_uncore_cancel_hrtimer(struct amd_uncore_ctx *ctx)
+{
+       hrtimer_cancel(&ctx->hrtimer);
+}
+
+static void amd_uncore_init_hrtimer(struct amd_uncore_ctx *ctx)
+{
+       hrtimer_setup(&ctx->hrtimer, amd_uncore_hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
+}
+
 static void amd_uncore_read(struct perf_event *event)
 {
        struct hw_perf_event *hwc = &event->hw;
@@ -117,18 +158,26 @@ static void amd_uncore_read(struct perf_event *event)
 
 static void amd_uncore_start(struct perf_event *event, int flags)
 {
+       struct amd_uncore_pmu *pmu = event_to_amd_uncore_pmu(event);
+       struct amd_uncore_ctx *ctx = *per_cpu_ptr(pmu->ctx, event->cpu);
        struct hw_perf_event *hwc = &event->hw;
 
+       if (!ctx->nr_active++)
+               amd_uncore_start_hrtimer(ctx);
+
        if (flags & PERF_EF_RELOAD)
                wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
 
        hwc->state = 0;
+       __set_bit(hwc->idx, ctx->active_mask);
        wrmsrl(hwc->config_base, (hwc->config | ARCH_PERFMON_EVENTSEL_ENABLE));
        perf_event_update_userpage(event);
 }
 
 static void amd_uncore_stop(struct perf_event *event, int flags)
 {
+       struct amd_uncore_pmu *pmu = event_to_amd_uncore_pmu(event);
+       struct amd_uncore_ctx *ctx = *per_cpu_ptr(pmu->ctx, event->cpu);
        struct hw_perf_event *hwc = &event->hw;
 
        wrmsrl(hwc->config_base, hwc->config);
@@ -138,6 +187,11 @@ static void amd_uncore_stop(struct perf_event *event, int flags)
                event->pmu->read(event);
                hwc->state |= PERF_HES_UPTODATE;
        }
+
+       if (!--ctx->nr_active)
+               amd_uncore_cancel_hrtimer(ctx);
+
+       __clear_bit(hwc->idx, ctx->active_mask);
 }
 
 static int amd_uncore_add(struct perf_event *event, int flags)
@@ -490,6 +544,9 @@ static int amd_uncore_ctx_init(struct amd_uncore *uncore, unsigned int cpu)
                                goto fail;
                        }
 
+                       amd_uncore_init_hrtimer(curr);
+                       curr->hrtimer_duration = 60LL * NSEC_PER_SEC;
+
                        cpumask_set_cpu(cpu, &pmu->active_mask);
                }
 
@@ -879,12 +936,18 @@ static int amd_uncore_umc_event_init(struct perf_event *event)
 
 static void amd_uncore_umc_start(struct perf_event *event, int flags)
 {
+       struct amd_uncore_pmu *pmu = event_to_amd_uncore_pmu(event);
+       struct amd_uncore_ctx *ctx = *per_cpu_ptr(pmu->ctx, event->cpu);
        struct hw_perf_event *hwc = &event->hw;
 
+       if (!ctx->nr_active++)
+               amd_uncore_start_hrtimer(ctx);
+
        if (flags & PERF_EF_RELOAD)
                wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
 
        hwc->state = 0;
+       __set_bit(hwc->idx, ctx->active_mask);
        wrmsrl(hwc->config_base, (hwc->config | AMD64_PERFMON_V2_ENABLE_UMC));
        perf_event_update_userpage(event);
 }
--

2.43.0

#5 lore raw

Re: [PATCH v2 3/5] perf/x86/amd/uncore: Use hrtimer for handling overflows

Quoted 13 lines · Stephane Eranian
Quoted 7 lines · Sandipan Das

Uncore counters do not provide mechanisms like interrupts to report overflows and the accumulated user-visible count is incorrect if there is more than one overflow between two successive read requests for the same event because the value of prev_count goes out-of-date for calculating the correct delta.

Quoted 5 lines

To avoid this, start a hrtimer to periodically initiate a pmu->read() of the active counters for keeping prev_count up-to-date. It should be noted that the hrtimer duration should be lesser than the shortest time it takes for a counter to overflow for this approach to be effective.

Quoted 5 lines

The problem I see is that the number of uncore PMU varies a lot based on the CPU model, in particular due to the L3 PMU.
Is there a timer armed per CCX or only a global one that will generate IPI to all other CPUs?

For L3 PMU, its on a per-CCX basis.

Quoted 142 lines
Quoted 142 lines
---
 arch/x86/events/amd/uncore.c | 63 ++++++++++++++++++++++++++++++++++++
 1 file changed, 63 insertions(+)
arch/x86/events/amd/uncore.c +63 -1
index 010024f09f2c..e09bfbb4a4cd 100644
@@ -21,6 +21,7 @@
 #define NUM_COUNTERS_NB                4
 #define NUM_COUNTERS_L2                4
 #define NUM_COUNTERS_L3                6
+#define NUM_COUNTERS_MAX       64
 
 #define RDPMC_BASE_NB          6
 #define RDPMC_BASE_LLC         10
@@ -38,6 +39,10 @@ struct amd_uncore_ctx {
        int refcnt;
        int cpu;
        struct perf_event **events;
+       unsigned long active_mask[BITS_TO_LONGS(NUM_COUNTERS_MAX)];
+       int nr_active;
+       struct hrtimer hrtimer;
+       u64 hrtimer_duration;
 };
 
 struct amd_uncore_pmu {
@@ -87,6 +92,42 @@ static struct amd_uncore_pmu *event_to_amd_uncore_pmu(struct perf_event *event)
        return container_of(event->pmu, struct amd_uncore_pmu, pmu);
 }
 
+static enum hrtimer_restart amd_uncore_hrtimer(struct hrtimer *hrtimer)
+{
+       struct amd_uncore_ctx *ctx;
+       struct perf_event *event;
+       int bit;
+
+       ctx = container_of(hrtimer, struct amd_uncore_ctx, hrtimer);
+
+       if (!ctx->nr_active || ctx->cpu != smp_processor_id())
+               return HRTIMER_NORESTART;
+
+       for_each_set_bit(bit, ctx->active_mask, NUM_COUNTERS_MAX) {
+               event = ctx->events[bit];
+               event->pmu->read(event);
+       }
+
+       hrtimer_forward_now(hrtimer, ns_to_ktime(ctx->hrtimer_duration));
+       return HRTIMER_RESTART;
+}
+
+static void amd_uncore_start_hrtimer(struct amd_uncore_ctx *ctx)
+{
+       hrtimer_start(&ctx->hrtimer, ns_to_ktime(ctx->hrtimer_duration),
+                     HRTIMER_MODE_REL_PINNED_HARD);
+}
+
+static void amd_uncore_cancel_hrtimer(struct amd_uncore_ctx *ctx)
+{
+       hrtimer_cancel(&ctx->hrtimer);
+}
+
+static void amd_uncore_init_hrtimer(struct amd_uncore_ctx *ctx)
+{
+       hrtimer_setup(&ctx->hrtimer, amd_uncore_hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
+}
+
 static void amd_uncore_read(struct perf_event *event)
 {
        struct hw_perf_event *hwc = &event->hw;
@@ -117,18 +158,26 @@ static void amd_uncore_read(struct perf_event *event)
 
 static void amd_uncore_start(struct perf_event *event, int flags)
 {
+       struct amd_uncore_pmu *pmu = event_to_amd_uncore_pmu(event);
+       struct amd_uncore_ctx *ctx = *per_cpu_ptr(pmu->ctx, event->cpu);
        struct hw_perf_event *hwc = &event->hw;
 
+       if (!ctx->nr_active++)
+               amd_uncore_start_hrtimer(ctx);
+
        if (flags & PERF_EF_RELOAD)
                wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
 
        hwc->state = 0;
+       __set_bit(hwc->idx, ctx->active_mask);
        wrmsrl(hwc->config_base, (hwc->config | ARCH_PERFMON_EVENTSEL_ENABLE));
        perf_event_update_userpage(event);
 }
 
 static void amd_uncore_stop(struct perf_event *event, int flags)
 {
+       struct amd_uncore_pmu *pmu = event_to_amd_uncore_pmu(event);
+       struct amd_uncore_ctx *ctx = *per_cpu_ptr(pmu->ctx, event->cpu);
        struct hw_perf_event *hwc = &event->hw;
 
        wrmsrl(hwc->config_base, hwc->config);
@@ -138,6 +187,11 @@ static void amd_uncore_stop(struct perf_event *event, int flags)
                event->pmu->read(event);
                hwc->state |= PERF_HES_UPTODATE;
        }
+
+       if (!--ctx->nr_active)
+               amd_uncore_cancel_hrtimer(ctx);
+
+       __clear_bit(hwc->idx, ctx->active_mask);
 }
 
 static int amd_uncore_add(struct perf_event *event, int flags)
@@ -490,6 +544,9 @@ static int amd_uncore_ctx_init(struct amd_uncore *uncore, unsigned int cpu)
                                goto fail;
                        }
 
+                       amd_uncore_init_hrtimer(curr);
+                       curr->hrtimer_duration = 60LL * NSEC_PER_SEC;
+
                        cpumask_set_cpu(cpu, &pmu->active_mask);
                }
 
@@ -879,12 +936,18 @@ static int amd_uncore_umc_event_init(struct perf_event *event)
 
 static void amd_uncore_umc_start(struct perf_event *event, int flags)
 {
+       struct amd_uncore_pmu *pmu = event_to_amd_uncore_pmu(event);
+       struct amd_uncore_ctx *ctx = *per_cpu_ptr(pmu->ctx, event->cpu);
        struct hw_perf_event *hwc = &event->hw;
 
+       if (!ctx->nr_active++)
+               amd_uncore_start_hrtimer(ctx);
+
        if (flags & PERF_EF_RELOAD)
                wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
 
        hwc->state = 0;
+       __set_bit(hwc->idx, ctx->active_mask);
        wrmsrl(hwc->config_base, (hwc->config | AMD64_PERFMON_V2_ENABLE_UMC));
        perf_event_update_userpage(event);
 }
--

2.43.0

#8 lore raw
In reply to Sandipan Das
Quoted 6 lines · Sandipan Das

Sandipan Das (5):
perf/x86/amd/uncore: Remove unused member from amd_uncore_ctx perf/x86/intel/uncore: Use HRTIMER_MODE_HARD for detecting overflows perf/x86/amd/uncore: Use hrtimer for handling overflows perf/x86/amd/uncore: Add parameter to configure hrtimer perf/x86/amd/uncore: Prevent UMC counters from saturating

Could you please fix your mailer to not mutiliate Cc: lines?

All these email addresses have real names, I suppose they weren't just written in in such a fashion?

Thanks,

Ingo

#9 lore raw
Quoted 12 lines · Ingo Molnar
Quoted 6 lines · Sandipan Das

Sandipan Das (5):
perf/x86/amd/uncore: Remove unused member from amd_uncore_ctx perf/x86/intel/uncore: Use HRTIMER_MODE_HARD for detecting overflows perf/x86/amd/uncore: Use hrtimer for handling overflows perf/x86/amd/uncore: Add parameter to configure hrtimer perf/x86/amd/uncore: Prevent UMC counters from saturating

Could you please fix your mailer to not mutiliate Cc: lines?

My bad. Will fix this.

Quoted 21 lines
Quoted 3 lines

All these email addresses have real names, I suppose they weren't just written in in such a fashion?

And this too.