diff mbox series

[v4,3/5] perf: Allow adding fixed random jitter to the sampling period

Message ID 20250408171530.140858-4-mark.barnett@arm.com (mailing list archive)
State New
Headers show
Series A mechanism for efficient support for per-function metrics | expand

Commit Message

Mark Barnett April 8, 2025, 5:15 p.m. UTC 
From: Ben Gainey <ben.gainey@arm.com>

This change modifies the core perf overflow handler, adding some small
random jitter to each sample period when the high-frequency sample
period is in use. A new flag is added to perf_event_attr to opt into
this behaviour.

This change follows the discussion in [1], where it is recognized that it
may be possible for certain patterns of execution to end up with biased
results.

[1] https://lore.kernel.org/linux-perf-users/Zc24eLqZycmIg3d2@tassilo/

Signed-off-by: Ben Gainey <ben.gainey@arm.com>
Signed-off-by: Mark Barnett <mark.barnett@arm.com>
---
 include/linux/perf_event.h |  1 +
 kernel/events/core.c       | 26 ++++++++++++++++++++++++++
 2 files changed, 27 insertions(+)

Comments

Ingo Molnar April 9, 2025, 10:54 a.m. UTC | #1 
* mark.barnett@arm.com <mark.barnett@arm.com> wrote:

> @@ -14384,6 +14409,7 @@ static void __init perf_event_init_all_cpus(void)
>  		cpuctx->online = cpumask_test_cpu(cpu, perf_online_mask);
>  		cpuctx->heap_size = ARRAY_SIZE(cpuctx->heap_default);
>  		cpuctx->heap = cpuctx->heap_default;
> +
>  	}
>  }

A stray newline snuck up on you here I think. ;-)

	Ingo
Peter Zijlstra April 9, 2025, 2:24 p.m. UTC | #2 
On Tue, Apr 08, 2025 at 06:15:28PM +0100, mark.barnett@arm.com wrote:
> @@ -10224,6 +10227,19 @@ static int __perf_event_overflow(struct perf_event *event,
>  	 *
>  	 * By ignoring the HF samples, we measure the actual period.
>  	 */
> +
> +	/*
> +	 * Apply optional jitter to the overall sample period
> +	 */
> +	if (hwc->sample_period_state & PERF_SPS_HF_RAND
> +			&& !(hwc->sample_period_state & PERF_SPS_HF_SAMPLE)) {

Coding style nit: when breaking lines, the operator goes on the end of
the preceding line.

> +		struct rnd_state *state = &get_cpu_var(sample_period_jitter_rnd);
> +		u64 rand_period = 1 << event->attr.hf_sample_rand;
> +
> +		sample_period -= rand_period / 2;
> +		sample_period += prandom_u32_state(state) & (rand_period - 1);
> +	}
> +
>  	if (hwc->sample_period_state & PERF_SPS_HF_ON) {
>  		u64 hf_sample_period = event->attr.hf_sample_period;
>
diff mbox series

Patch

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index be006965054e..78a6fd14b412 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -233,6 +233,7 @@  struct hw_perf_event {
 
 #define PERF_SPS_HF_ON			0x00000001
 #define PERF_SPS_HF_SAMPLE		0x00000002
+#define PERF_SPS_HF_RAND		0x00000004
 	u32				sample_period_state;
 
 	/*
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 5752ac7408b1..bc6991a33048 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -56,6 +56,7 @@ 
 #include <linux/buildid.h>
 #include <linux/task_work.h>
 #include <linux/percpu-rwsem.h>
+#include <linux/prandom.h>
 
 #include "internal.h"
 
@@ -472,6 +473,8 @@  static int perf_sample_period_ns __read_mostly	= DEFAULT_SAMPLE_PERIOD_NS;
 static int perf_sample_allowed_ns __read_mostly =
 	DEFAULT_SAMPLE_PERIOD_NS * DEFAULT_CPU_TIME_MAX_PERCENT / 100;
 
+static DEFINE_PER_CPU(struct rnd_state, sample_period_jitter_rnd);
+
 static void update_perf_cpu_limits(void)
 {
 	u64 tmp = perf_sample_period_ns;
@@ -10224,6 +10227,19 @@  static int __perf_event_overflow(struct perf_event *event,
 	 *
 	 * By ignoring the HF samples, we measure the actual period.
 	 */
+
+	/*
+	 * Apply optional jitter to the overall sample period
+	 */
+	if (hwc->sample_period_state & PERF_SPS_HF_RAND
+			&& !(hwc->sample_period_state & PERF_SPS_HF_SAMPLE)) {
+		struct rnd_state *state = &get_cpu_var(sample_period_jitter_rnd);
+		u64 rand_period = 1 << event->attr.hf_sample_rand;
+
+		sample_period -= rand_period / 2;
+		sample_period += prandom_u32_state(state) & (rand_period - 1);
+	}
+
 	if (hwc->sample_period_state & PERF_SPS_HF_ON) {
 		u64 hf_sample_period = event->attr.hf_sample_period;
 
@@ -12756,6 +12772,14 @@  perf_event_alloc(struct perf_event_attr *attr, int cpu,
 	if (attr->hf_sample_period)
 		hwc->sample_period_state |= PERF_SPS_HF_ON;
 
+	if (attr->hf_sample_rand) {
+		/* high-frequency jitter is only valid with a high-freq period */
+		if (!attr->hf_sample_period)
+			return ERR_PTR(-EINVAL);
+
+		hwc->sample_period_state |= PERF_SPS_HF_RAND;
+	}
+
 	/*
 	 * Disallow uncore-task events. Similarly, disallow uncore-cgroup
 	 * events (they don't make sense as the cgroup will be different
@@ -14367,6 +14391,7 @@  static void __init perf_event_init_all_cpus(void)
 	zalloc_cpumask_var(&perf_online_pkg_mask, GFP_KERNEL);
 	zalloc_cpumask_var(&perf_online_sys_mask, GFP_KERNEL);
 
+	prandom_seed_full_state(&sample_period_jitter_rnd);
 
 	for_each_possible_cpu(cpu) {
 		swhash = &per_cpu(swevent_htable, cpu);
@@ -14384,6 +14409,7 @@  static void __init perf_event_init_all_cpus(void)
 		cpuctx->online = cpumask_test_cpu(cpu, perf_online_mask);
 		cpuctx->heap_size = ARRAY_SIZE(cpuctx->heap_default);
 		cpuctx->heap = cpuctx->heap_default;
+
 	}
 }