From: Ben Gainey <ben.gainey@arm.com>

This change modifies perf_event_attr to add a second, alternative
sample period field, and modifies the core perf overflow handling
such that when specified an event will alternate between two sample
periods.

Currently, perf does not provide a  mechanism for decoupling the period
over which counters are counted from the period between samples. This is
problematic for building a tool to measure per-function metrics derived
from a sampled counter group. Ideally such a tool wants a very small
sample window in order to correctly attribute the metrics to a given
function, but prefers a larger sample period that provides representative
coverage without excessive probe effect, triggering throttling, or
generating excessive amounts of data.

By alternating between a long and short sample_period and subsequently
discarding the long samples, tools may decouple the period between
samples that the tool cares about from the window of time over which
interesting counts are collected.

It is expected that typically tools would use this feature with the
cycles or instructions events as an approximation for time, but no
restrictions are applied to which events this can be applied to.

Signed-off-by: Ben Gainey <ben.gainey@arm.com>
Signed-off-by: Mark Barnett <mark.barnett@arm.com>
---
 include/linux/perf_event.h      |  5 +++++
 include/uapi/linux/perf_event.h |  3 +++
 kernel/events/core.c            | 39 ++++++++++++++++++++++++++++++++-
 3 files changed, 46 insertions(+), 1 deletion(-)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 8333f132f4a9..99ba72c8fb6d 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -276,6 +276,11 @@ struct hw_perf_event {
 	 */
 	u64				freq_time_stamp;
 	u64				freq_count_stamp;
+
+	/*
+	 * Indicates that the alternative sample period is used
+	 */
+	bool				using_alt_sample_period;
 #endif
 };
 
diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
index 0524d541d4e3..499a8673df8e 100644
--- a/include/uapi/linux/perf_event.h
+++ b/include/uapi/linux/perf_event.h
@@ -379,6 +379,7 @@ enum perf_event_read_format {
 #define PERF_ATTR_SIZE_VER6	120	/* add: aux_sample_size */
 #define PERF_ATTR_SIZE_VER7	128	/* add: sig_data */
 #define PERF_ATTR_SIZE_VER8	136	/* add: config3 */
+#define PERF_ATTR_SIZE_VER9	144	/* add: alt_sample_period */
 
 /*
  * Hardware event_id to monitor via a performance monitoring event:
@@ -531,6 +532,8 @@ struct perf_event_attr {
 	__u64	sig_data;
 
 	__u64	config3; /* extension of config2 */
+
+	__u64	alt_sample_period;
 };
 
 /*
diff --git a/kernel/events/core.c b/kernel/events/core.c
index bcb09e011e9e..7ec8ec6ba7ef 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -4178,6 +4178,8 @@ static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bo
 	s64 period, sample_period;
 	s64 delta;
 
+	WARN_ON_ONCE(hwc->using_alt_sample_period);
+
 	period = perf_calculate_period(event, nsec, count);
 
 	delta = (s64)(period - hwc->sample_period);
@@ -9894,6 +9896,7 @@ static int __perf_event_overflow(struct perf_event *event,
 				 int throttle, struct perf_sample_data *data,
 				 struct pt_regs *regs)
 {
+	struct hw_perf_event *hwc = &event->hw;
 	int events = atomic_read(&event->event_limit);
 	int ret = 0;
 
@@ -9913,6 +9916,18 @@ static int __perf_event_overflow(struct perf_event *event,
 	    !bpf_overflow_handler(event, data, regs))
 		goto out;
 
+	/*
+	 * Swap the sample period to the alternative period
+	 */
+	if (event->attr.alt_sample_period) {
+		bool using_alt = hwc->using_alt_sample_period;
+		u64 sample_period = (using_alt ? event->attr.sample_period
+					       : event->attr.alt_sample_period);
+
+		hwc->sample_period = sample_period;
+		hwc->using_alt_sample_period = !using_alt;
+	}
+
 	/*
 	 * XXX event_limit might not quite work as expected on inherited
 	 * events
@@ -12335,9 +12350,19 @@ perf_event_alloc(struct perf_event_attr *attr, int cpu,
 	if (attr->freq && attr->sample_freq)
 		hwc->sample_period = 1;
 	hwc->last_period = hwc->sample_period;
-
 	local64_set(&hwc->period_left, hwc->sample_period);
 
+	/*
+	 * alt_sample_period cannot be used with freq
+	 */
+	if (attr->freq && attr->alt_sample_period)
+		goto err_ns;
+
+	if (attr->alt_sample_period) {
+		hwc->sample_period = attr->alt_sample_period;
+		hwc->using_alt_sample_period = true;
+	}
+
 	/*
 	 * We do not support PERF_SAMPLE_READ on inherited events unless
 	 * PERF_SAMPLE_TID is also selected, which allows inherited events to
@@ -12807,9 +12832,21 @@ SYSCALL_DEFINE5(perf_event_open,
 	if (attr.freq) {
 		if (attr.sample_freq > sysctl_perf_event_sample_rate)
 			return -EINVAL;
+		if (attr.alt_sample_period)
+			return -EINVAL;
 	} else {
 		if (attr.sample_period & (1ULL << 63))
 			return -EINVAL;
+		if (attr.alt_sample_period) {
+			if (!attr.sample_period)
+				return -EINVAL;
+			if (attr.alt_sample_period & (1ULL << 63))
+				return -EINVAL;
+			if (attr.alt_sample_period > attr.sample_period)
+				return -EINVAL;
+			if (attr.alt_sample_period == attr.sample_period)
+				attr.alt_sample_period = 0;
+		}
 	}
 
 	/* Only privileged users can get physical addresses */
-- 
2.43.0

On Fri, Mar 07, 2025 at 08:22:44PM +0000, mark.barnett@arm.com wrote:
> From: Ben Gainey <ben.gainey@arm.com>
> 
> This change modifies perf_event_attr to add a second, alternative
> sample period field, and modifies the core perf overflow handling
> such that when specified an event will alternate between two sample
> periods.
> 
> Currently, perf does not provide a  mechanism for decoupling the period
> over which counters are counted from the period between samples. This is
> problematic for building a tool to measure per-function metrics derived
> from a sampled counter group. Ideally such a tool wants a very small
> sample window in order to correctly attribute the metrics to a given
> function, but prefers a larger sample period that provides representative
> coverage without excessive probe effect, triggering throttling, or
> generating excessive amounts of data.
> 
> By alternating between a long and short sample_period and subsequently
> discarding the long samples, tools may decouple the period between
> samples that the tool cares about from the window of time over which
> interesting counts are collected.
> 
> It is expected that typically tools would use this feature with the
> cycles or instructions events as an approximation for time, but no
> restrictions are applied to which events this can be applied to.

So you do add the constraint that 'alt_sample_period < sample_period'
but there is no natural reason for this to be so.

Additionally, this way the total period ends up being 'sample_period +
alt_sample_period'.

Would not a more natural way to express things be:

	p1 = sample_period - alt_sample_period;
	p2 = alt_sample_period;

This way you retain the total period to be 'sample_period' and naturally
get the constraint: 'alt_sample_period < sample_period'.

That is; I'm somewhat confused by the state of things; it doesn't seem
consistent.

(Also note that this alternative form might actually work in combination
with attr.freq set -- although that has a number of 'fun' details I'm
sure).

> Signed-off-by: Ben Gainey <ben.gainey@arm.com>
> Signed-off-by: Mark Barnett <mark.barnett@arm.com>
> ---
>  include/linux/perf_event.h      |  5 +++++
>  include/uapi/linux/perf_event.h |  3 +++
>  kernel/events/core.c            | 39 ++++++++++++++++++++++++++++++++-
>  3 files changed, 46 insertions(+), 1 deletion(-)
> 
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index 8333f132f4a9..99ba72c8fb6d 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -276,6 +276,11 @@ struct hw_perf_event {
>  	 */
>  	u64				freq_time_stamp;
>  	u64				freq_count_stamp;
> +
> +	/*
> +	 * Indicates that the alternative sample period is used
> +	 */
> +	bool				using_alt_sample_period;

There's a 4 byte hole in this structure if you look; also please use a
flag, sizeof(_Bool) is ABI dependent.

>  #endif
>  };
>  
> diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
> index 0524d541d4e3..499a8673df8e 100644
> --- a/include/uapi/linux/perf_event.h
> +++ b/include/uapi/linux/perf_event.h
> @@ -379,6 +379,7 @@ enum perf_event_read_format {
>  #define PERF_ATTR_SIZE_VER6	120	/* add: aux_sample_size */
>  #define PERF_ATTR_SIZE_VER7	128	/* add: sig_data */
>  #define PERF_ATTR_SIZE_VER8	136	/* add: config3 */
> +#define PERF_ATTR_SIZE_VER9	144	/* add: alt_sample_period */
>  
>  /*
>   * Hardware event_id to monitor via a performance monitoring event:
> @@ -531,6 +532,8 @@ struct perf_event_attr {
>  	__u64	sig_data;
>  
>  	__u64	config3; /* extension of config2 */
> +
> +	__u64	alt_sample_period;
>  };
>  
>  /*
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index bcb09e011e9e..7ec8ec6ba7ef 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -4178,6 +4178,8 @@ static void perf_adjust_period(struct perf_event *event, u64 nsec, u64 count, bo
>  	s64 period, sample_period;
>  	s64 delta;
>  
> +	WARN_ON_ONCE(hwc->using_alt_sample_period);

Groan; so that bit keeps flipping in and off, and statistically we'll
warn, but urgh.

>  	period = perf_calculate_period(event, nsec, count);
>  
>  	delta = (s64)(period - hwc->sample_period);
> @@ -9894,6 +9896,7 @@ static int __perf_event_overflow(struct perf_event *event,
>  				 int throttle, struct perf_sample_data *data,
>  				 struct pt_regs *regs)
>  {
> +	struct hw_perf_event *hwc = &event->hw;
>  	int events = atomic_read(&event->event_limit);
>  	int ret = 0;
>  
> @@ -9913,6 +9916,18 @@ static int __perf_event_overflow(struct perf_event *event,
>  	    !bpf_overflow_handler(event, data, regs))
>  		goto out;
>  
> +	/*
> +	 * Swap the sample period to the alternative period
> +	 */
> +	if (event->attr.alt_sample_period) {
> +		bool using_alt = hwc->using_alt_sample_period;
> +		u64 sample_period = (using_alt ? event->attr.sample_period
> +					       : event->attr.alt_sample_period);
> +
> +		hwc->sample_period = sample_period;
> +		hwc->using_alt_sample_period = !using_alt;
> +	}
> +
>  	/*
>  	 * XXX event_limit might not quite work as expected on inherited
>  	 * events
> @@ -12335,9 +12350,19 @@ perf_event_alloc(struct perf_event_attr *attr, int cpu,
>  	if (attr->freq && attr->sample_freq)
>  		hwc->sample_period = 1;
>  	hwc->last_period = hwc->sample_period;
> -
>  	local64_set(&hwc->period_left, hwc->sample_period);
>  
> +	/*
> +	 * alt_sample_period cannot be used with freq
> +	 */
> +	if (attr->freq && attr->alt_sample_period)
> +		goto err_ns;

How can this happen? This case has already been filtered in
perf_event_open() below, no?

Also, this doesn't apply to tip/perf/core, someone went and changed
things...

> +
> +	if (attr->alt_sample_period) {
> +		hwc->sample_period = attr->alt_sample_period;
> +		hwc->using_alt_sample_period = true;
> +	}
> +
>  	/*
>  	 * We do not support PERF_SAMPLE_READ on inherited events unless
>  	 * PERF_SAMPLE_TID is also selected, which allows inherited events to
> @@ -12807,9 +12832,21 @@ SYSCALL_DEFINE5(perf_event_open,
>  	if (attr.freq) {
>  		if (attr.sample_freq > sysctl_perf_event_sample_rate)
>  			return -EINVAL;
> +		if (attr.alt_sample_period)
> +			return -EINVAL;
>  	} else {
>  		if (attr.sample_period & (1ULL << 63))
>  			return -EINVAL;
> +		if (attr.alt_sample_period) {
> +			if (!attr.sample_period)
> +				return -EINVAL;
> +			if (attr.alt_sample_period & (1ULL << 63))
> +				return -EINVAL;
> +			if (attr.alt_sample_period > attr.sample_period)
> +				return -EINVAL;
> +			if (attr.alt_sample_period == attr.sample_period)
> +				attr.alt_sample_period = 0;
> +		}
>  	}
>  
>  	/* Only privileged users can get physical addresses */
> -- 
> 2.43.0
>