[PATCH v2] perf/core: Add a new read format to get a number of lost samples

[PATCH v2] perf/core: Add a new read format to get a number of lost samples

[Namhyung Kim]

Sometimes we want to know an accurate number of samples even if it's
lost.  Currenlty PERF_RECORD_LOST is generated for a ring-buffer which
might be shared with other events.  So it's hard to know per-event
lost count.

Add event->lost_samples field and PERF_FORMAT_LOST to retrieve it from
userspace.

Original-patch-by: Jiri Olsa <jolsa@redhat.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
---
 include/linux/perf_event.h      |  2 ++
 include/uapi/linux/perf_event.h |  5 ++++-
 kernel/events/core.c            | 22 +++++++++++++++++++---
 kernel/events/ring_buffer.c     |  5 ++++-
 4 files changed, 29 insertions(+), 5 deletions(-)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index f5a6a2f069ed..189a471fba42 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -756,6 +756,8 @@ struct perf_event {
 	struct pid_namespace		*ns;
 	u64				id;
 
+	atomic64_t			lost_samples;
+
 	u64				(*clock)(void);
 	perf_overflow_handler_t		overflow_handler;
 	void				*overflow_handler_context;
diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
index bf8143505c49..f72008949ff0 100644
--- a/include/uapi/linux/perf_event.h
+++ b/include/uapi/linux/perf_event.h
@@ -299,6 +299,7 @@ enum {
  *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
  *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
  *	  { u64		id;           } && PERF_FORMAT_ID
+ *	  { u64		lost;         } && PERF_FORMAT_LOST
  *	} && !PERF_FORMAT_GROUP
  *
  *	{ u64		nr;
@@ -306,6 +307,7 @@ enum {
  *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
  *	  { u64		value;
  *	    { u64	id;           } && PERF_FORMAT_ID
+ *	    { u64	lost;         } && PERF_FORMAT_LOST
  *	  }		cntr[nr];
  *	} && PERF_FORMAT_GROUP
  * };
@@ -315,8 +317,9 @@ enum perf_event_read_format {
 	PERF_FORMAT_TOTAL_TIME_RUNNING		= 1U << 1,
 	PERF_FORMAT_ID				= 1U << 2,
 	PERF_FORMAT_GROUP			= 1U << 3,
+	PERF_FORMAT_LOST			= 1U << 4,
 
-	PERF_FORMAT_MAX = 1U << 4,		/* non-ABI */
+	PERF_FORMAT_MAX = 1U << 5,		/* non-ABI */
 };
 
 #define PERF_ATTR_SIZE_VER0	64	/* sizeof first published struct */
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 0e125ae2fa92..8708325ee4a2 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -1841,6 +1841,9 @@ static void __perf_event_read_size(struct perf_event *event, int nr_siblings)
 	if (event->attr.read_format & PERF_FORMAT_ID)
 		entry += sizeof(u64);
 
+	if (event->attr.read_format & PERF_FORMAT_LOST)
+		entry += sizeof(u64);
+
 	if (event->attr.read_format & PERF_FORMAT_GROUP) {
 		nr += nr_siblings;
 		size += sizeof(u64);
@@ -5255,11 +5258,15 @@ static int __perf_read_group_add(struct perf_event *leader,
 	values[n++] += perf_event_count(leader);
 	if (read_format & PERF_FORMAT_ID)
 		values[n++] = primary_event_id(leader);
+	if (read_format & PERF_FORMAT_ID)
+		values[n++] = atomic64_read(&leader->lost_samples);
 
 	for_each_sibling_event(sub, leader) {
 		values[n++] += perf_event_count(sub);
 		if (read_format & PERF_FORMAT_ID)
 			values[n++] = primary_event_id(sub);
+		if (read_format & PERF_FORMAT_ID)
+			values[n++] = atomic64_read(&sub->lost_samples);
 	}
 
 	raw_spin_unlock_irqrestore(&ctx->lock, flags);
@@ -5316,7 +5323,7 @@ static int perf_read_one(struct perf_event *event,
 				 u64 read_format, char __user *buf)
 {
 	u64 enabled, running;
-	u64 values[4];
+	u64 values[5];
 	int n = 0;
 
 	values[n++] = __perf_event_read_value(event, &enabled, &running);
@@ -5326,6 +5333,8 @@ static int perf_read_one(struct perf_event *event,
 		values[n++] = running;
 	if (read_format & PERF_FORMAT_ID)
 		values[n++] = primary_event_id(event);
+	if (read_format & PERF_FORMAT_LOST)
+		values[n++] = atomic64_read(&event->lost_samples);
 
 	if (copy_to_user(buf, values, n * sizeof(u64)))
 		return -EFAULT;
@@ -5664,6 +5673,7 @@ static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned lon
 
 		return perf_event_modify_attr(event,  &new_attr);
 	}
+
 	default:
 		return -ENOTTY;
 	}
@@ -6835,7 +6845,7 @@ static void perf_output_read_one(struct perf_output_handle *handle,
 				 u64 enabled, u64 running)
 {
 	u64 read_format = event->attr.read_format;
-	u64 values[4];
+	u64 values[5];
 	int n = 0;
 
 	values[n++] = perf_event_count(event);
@@ -6849,6 +6859,8 @@ static void perf_output_read_one(struct perf_output_handle *handle,
 	}
 	if (read_format & PERF_FORMAT_ID)
 		values[n++] = primary_event_id(event);
+	if (read_format & PERF_FORMAT_LOST)
+		values[n++] = atomic64_read(&event->lost_samples);
 
 	__output_copy(handle, values, n * sizeof(u64));
 }
@@ -6859,7 +6871,7 @@ static void perf_output_read_group(struct perf_output_handle *handle,
 {
 	struct perf_event *leader = event->group_leader, *sub;
 	u64 read_format = event->attr.read_format;
-	u64 values[5];
+	u64 values[6];
 	int n = 0;
 
 	values[n++] = 1 + leader->nr_siblings;
@@ -6877,6 +6889,8 @@ static void perf_output_read_group(struct perf_output_handle *handle,
 	values[n++] = perf_event_count(leader);
 	if (read_format & PERF_FORMAT_ID)
 		values[n++] = primary_event_id(leader);
+	if (read_format & PERF_FORMAT_LOST)
+		values[n++] = atomic64_read(&leader->lost_samples);
 
 	__output_copy(handle, values, n * sizeof(u64));
 
@@ -6890,6 +6904,8 @@ static void perf_output_read_group(struct perf_output_handle *handle,
 		values[n++] = perf_event_count(sub);
 		if (read_format & PERF_FORMAT_ID)
 			values[n++] = primary_event_id(sub);
+		if (read_format & PERF_FORMAT_LOST)
+			values[n++] = atomic64_read(&sub->lost_samples);
 
 		__output_copy(handle, values, n * sizeof(u64));
 	}
diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
index 52868716ec35..727ca8f4caad 100644
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -172,8 +172,10 @@ __perf_output_begin(struct perf_output_handle *handle,
 		goto out;
 
 	if (unlikely(rb->paused)) {
-		if (rb->nr_pages)
+		if (rb->nr_pages) {
 			local_inc(&rb->lost);
+			atomic64_inc(&event->lost_samples);
+		}
 		goto out;
 	}
 
@@ -254,6 +256,7 @@ __perf_output_begin(struct perf_output_handle *handle,
 
 fail:
 	local_inc(&rb->lost);
+	atomic64_inc(&event->lost_samples);
 	perf_output_put_handle(handle);
 out:
 	rcu_read_unlock();
-- 
2.33.0.309.g3052b89438-goog

Re: [PATCH v2] perf/core: Add a new read format to get a number of lost samples

[Jiri Olsa]

On Wed, Sep 08, 2021 at 10:24:20AM -0700, Namhyung Kim wrote:
> Sometimes we want to know an accurate number of samples even if it's
> lost.  Currenlty PERF_RECORD_LOST is generated for a ring-buffer which
> might be shared with other events.  So it's hard to know per-event
> lost count.
> 
> Add event->lost_samples field and PERF_FORMAT_LOST to retrieve it from
> userspace.

hi,
looks good.. will there be some tools/perf change using this?

thanks,
jirka

> 
> Original-patch-by: Jiri Olsa <jolsa@redhat.com>
> Signed-off-by: Namhyung Kim <namhyung@kernel.org>
> ---
>  include/linux/perf_event.h      |  2 ++
>  include/uapi/linux/perf_event.h |  5 ++++-
>  kernel/events/core.c            | 22 +++++++++++++++++++---
>  kernel/events/ring_buffer.c     |  5 ++++-
>  4 files changed, 29 insertions(+), 5 deletions(-)
> 
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index f5a6a2f069ed..189a471fba42 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -756,6 +756,8 @@ struct perf_event {
>  	struct pid_namespace		*ns;
>  	u64				id;
>  
> +	atomic64_t			lost_samples;
> +
>  	u64				(*clock)(void);
>  	perf_overflow_handler_t		overflow_handler;
>  	void				*overflow_handler_context;
> diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
> index bf8143505c49..f72008949ff0 100644
> --- a/include/uapi/linux/perf_event.h
> +++ b/include/uapi/linux/perf_event.h
> @@ -299,6 +299,7 @@ enum {
>   *	  { u64		time_enabled; } && PERF_FORMAT_TOTAL_TIME_ENABLED
>   *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
>   *	  { u64		id;           } && PERF_FORMAT_ID
> + *	  { u64		lost;         } && PERF_FORMAT_LOST
>   *	} && !PERF_FORMAT_GROUP
>   *
>   *	{ u64		nr;
> @@ -306,6 +307,7 @@ enum {
>   *	  { u64		time_running; } && PERF_FORMAT_TOTAL_TIME_RUNNING
>   *	  { u64		value;
>   *	    { u64	id;           } && PERF_FORMAT_ID
> + *	    { u64	lost;         } && PERF_FORMAT_LOST
>   *	  }		cntr[nr];
>   *	} && PERF_FORMAT_GROUP
>   * };
> @@ -315,8 +317,9 @@ enum perf_event_read_format {
>  	PERF_FORMAT_TOTAL_TIME_RUNNING		= 1U << 1,
>  	PERF_FORMAT_ID				= 1U << 2,
>  	PERF_FORMAT_GROUP			= 1U << 3,
> +	PERF_FORMAT_LOST			= 1U << 4,
>  
> -	PERF_FORMAT_MAX = 1U << 4,		/* non-ABI */
> +	PERF_FORMAT_MAX = 1U << 5,		/* non-ABI */
>  };
>  
>  #define PERF_ATTR_SIZE_VER0	64	/* sizeof first published struct */
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 0e125ae2fa92..8708325ee4a2 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -1841,6 +1841,9 @@ static void __perf_event_read_size(struct perf_event *event, int nr_siblings)
>  	if (event->attr.read_format & PERF_FORMAT_ID)
>  		entry += sizeof(u64);
>  
> +	if (event->attr.read_format & PERF_FORMAT_LOST)
> +		entry += sizeof(u64);
> +
>  	if (event->attr.read_format & PERF_FORMAT_GROUP) {
>  		nr += nr_siblings;
>  		size += sizeof(u64);
> @@ -5255,11 +5258,15 @@ static int __perf_read_group_add(struct perf_event *leader,
>  	values[n++] += perf_event_count(leader);
>  	if (read_format & PERF_FORMAT_ID)
>  		values[n++] = primary_event_id(leader);
> +	if (read_format & PERF_FORMAT_ID)
> +		values[n++] = atomic64_read(&leader->lost_samples);
>  
>  	for_each_sibling_event(sub, leader) {
>  		values[n++] += perf_event_count(sub);
>  		if (read_format & PERF_FORMAT_ID)
>  			values[n++] = primary_event_id(sub);
> +		if (read_format & PERF_FORMAT_ID)
> +			values[n++] = atomic64_read(&sub->lost_samples);
>  	}
>  
>  	raw_spin_unlock_irqrestore(&ctx->lock, flags);
> @@ -5316,7 +5323,7 @@ static int perf_read_one(struct perf_event *event,
>  				 u64 read_format, char __user *buf)
>  {
>  	u64 enabled, running;
> -	u64 values[4];
> +	u64 values[5];
>  	int n = 0;
>  
>  	values[n++] = __perf_event_read_value(event, &enabled, &running);
> @@ -5326,6 +5333,8 @@ static int perf_read_one(struct perf_event *event,
>  		values[n++] = running;
>  	if (read_format & PERF_FORMAT_ID)
>  		values[n++] = primary_event_id(event);
> +	if (read_format & PERF_FORMAT_LOST)
> +		values[n++] = atomic64_read(&event->lost_samples);
>  
>  	if (copy_to_user(buf, values, n * sizeof(u64)))
>  		return -EFAULT;
> @@ -5664,6 +5673,7 @@ static long _perf_ioctl(struct perf_event *event, unsigned int cmd, unsigned lon
>  
>  		return perf_event_modify_attr(event,  &new_attr);
>  	}
> +
>  	default:
>  		return -ENOTTY;
>  	}
> @@ -6835,7 +6845,7 @@ static void perf_output_read_one(struct perf_output_handle *handle,
>  				 u64 enabled, u64 running)
>  {
>  	u64 read_format = event->attr.read_format;
> -	u64 values[4];
> +	u64 values[5];
>  	int n = 0;
>  
>  	values[n++] = perf_event_count(event);
> @@ -6849,6 +6859,8 @@ static void perf_output_read_one(struct perf_output_handle *handle,
>  	}
>  	if (read_format & PERF_FORMAT_ID)
>  		values[n++] = primary_event_id(event);
> +	if (read_format & PERF_FORMAT_LOST)
> +		values[n++] = atomic64_read(&event->lost_samples);
>  
>  	__output_copy(handle, values, n * sizeof(u64));
>  }
> @@ -6859,7 +6871,7 @@ static void perf_output_read_group(struct perf_output_handle *handle,
>  {
>  	struct perf_event *leader = event->group_leader, *sub;
>  	u64 read_format = event->attr.read_format;
> -	u64 values[5];
> +	u64 values[6];
>  	int n = 0;
>  
>  	values[n++] = 1 + leader->nr_siblings;
> @@ -6877,6 +6889,8 @@ static void perf_output_read_group(struct perf_output_handle *handle,
>  	values[n++] = perf_event_count(leader);
>  	if (read_format & PERF_FORMAT_ID)
>  		values[n++] = primary_event_id(leader);
> +	if (read_format & PERF_FORMAT_LOST)
> +		values[n++] = atomic64_read(&leader->lost_samples);
>  
>  	__output_copy(handle, values, n * sizeof(u64));
>  
> @@ -6890,6 +6904,8 @@ static void perf_output_read_group(struct perf_output_handle *handle,
>  		values[n++] = perf_event_count(sub);
>  		if (read_format & PERF_FORMAT_ID)
>  			values[n++] = primary_event_id(sub);
> +		if (read_format & PERF_FORMAT_LOST)
> +			values[n++] = atomic64_read(&sub->lost_samples);
>  
>  		__output_copy(handle, values, n * sizeof(u64));
>  	}
> diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
> index 52868716ec35..727ca8f4caad 100644
> --- a/kernel/events/ring_buffer.c
> +++ b/kernel/events/ring_buffer.c
> @@ -172,8 +172,10 @@ __perf_output_begin(struct perf_output_handle *handle,
>  		goto out;
>  
>  	if (unlikely(rb->paused)) {
> -		if (rb->nr_pages)
> +		if (rb->nr_pages) {
>  			local_inc(&rb->lost);
> +			atomic64_inc(&event->lost_samples);
> +		}
>  		goto out;
>  	}
>  
> @@ -254,6 +256,7 @@ __perf_output_begin(struct perf_output_handle *handle,
>  
>  fail:
>  	local_inc(&rb->lost);
> +	atomic64_inc(&event->lost_samples);
>  	perf_output_put_handle(handle);
>  out:
>  	rcu_read_unlock();
> -- 
> 2.33.0.309.g3052b89438-goog
> 

Re: [PATCH v2] perf/core: Add a new read format to get a number of lost samples

[Namhyung Kim]

On Fri, Sep 10, 2021 at 2:42 PM Jiri Olsa <jolsa@redhat.com> wrote:
>
> On Wed, Sep 08, 2021 at 10:24:20AM -0700, Namhyung Kim wrote:
> > Sometimes we want to know an accurate number of samples even if it's
> > lost.  Currenlty PERF_RECORD_LOST is generated for a ring-buffer which
> > might be shared with other events.  So it's hard to know per-event
> > lost count.
> >
> > Add event->lost_samples field and PERF_FORMAT_LOST to retrieve it from
> > userspace.
>
> hi,
> looks good.. will there be some tools/perf change using this?

Sure, I'll work on the userspace part after this is merged.
I'm thinking about adding LOST_SAMPLES records at the end
of the data section as if they came from the kernel.

Thanks,
Namhyung

Re: [PATCH v2] perf/core: Add a new read format to get a number of lost samples

[Namhyung Kim]

Hi Peter,

On Fri, Sep 10, 2021 at 3:40 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> On Fri, Sep 10, 2021 at 2:42 PM Jiri Olsa <jolsa@redhat.com> wrote:
> >
> > On Wed, Sep 08, 2021 at 10:24:20AM -0700, Namhyung Kim wrote:
> > > Sometimes we want to know an accurate number of samples even if it's
> > > lost.  Currenlty PERF_RECORD_LOST is generated for a ring-buffer which
> > > might be shared with other events.  So it's hard to know per-event
> > > lost count.
> > >
> > > Add event->lost_samples field and PERF_FORMAT_LOST to retrieve it from
> > > userspace.
> >
> > hi,
> > looks good.. will there be some tools/perf change using this?
>
> Sure, I'll work on the userspace part after this is merged.
> I'm thinking about adding LOST_SAMPLES records at the end
> of the data section as if they came from the kernel.

Can I get your feedback?

Thanks,
Namhyung

Re: [PATCH v2] perf/core: Add a new read format to get a number of lost samples

[Namhyung Kim]

Ping!

On Wed, Sep 22, 2021 at 12:34 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> Hi Peter,
>
> On Fri, Sep 10, 2021 at 3:40 PM Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > On Fri, Sep 10, 2021 at 2:42 PM Jiri Olsa <jolsa@redhat.com> wrote:
> > >
> > > On Wed, Sep 08, 2021 at 10:24:20AM -0700, Namhyung Kim wrote:
> > > > Sometimes we want to know an accurate number of samples even if it's
> > > > lost.  Currenlty PERF_RECORD_LOST is generated for a ring-buffer which
> > > > might be shared with other events.  So it's hard to know per-event
> > > > lost count.
> > > >
> > > > Add event->lost_samples field and PERF_FORMAT_LOST to retrieve it from
> > > > userspace.
> > >
> > > hi,
> > > looks good.. will there be some tools/perf change using this?
> >
> > Sure, I'll work on the userspace part after this is merged.
> > I'm thinking about adding LOST_SAMPLES records at the end
> > of the data section as if they came from the kernel.
>
> Can I get your feedback?
>
> Thanks,
> Namhyung