perf/core: Increase lost_samples count only for samples

[PATCH] perf/core: Increase lost_samples count only for samples

Posted by Namhyung Kim 3 years, 7 months ago

The event->lost_samples count is intended to count (lost) sample records
but it's also counted for non-sample records like PERF_RECORD_MMAP etc.
This can be a problem when a sampling event tracks those side-band
events together.

As overflow handler for user events only calls perf_output_begin_
{for,back}ward() before writing to the ring buffer, we can pass an
additional flag to indicate that it's writing a sample record.

Fixes: 119a784c8127 ("perf/core: Add a new read format to get a number of lost samples")
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
---
 kernel/events/ring_buffer.c | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)

diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
index 726132039c38..5f38ee4edbdb 100644
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -149,7 +149,7 @@ static __always_inline int
 __perf_output_begin(struct perf_output_handle *handle,
 		    struct perf_sample_data *data,
 		    struct perf_event *event, unsigned int size,
-		    bool backward)
+		    bool backward, bool sample)
 {
 	struct perf_buffer *rb;
 	unsigned long tail, offset, head;
@@ -174,7 +174,8 @@ __perf_output_begin(struct perf_output_handle *handle,
 	if (unlikely(rb->paused)) {
 		if (rb->nr_pages) {
 			local_inc(&rb->lost);
-			atomic64_inc(&event->lost_samples);
+			if (sample)
+				atomic64_inc(&event->lost_samples);
 		}
 		goto out;
 	}
@@ -256,7 +257,8 @@ __perf_output_begin(struct perf_output_handle *handle,
 
 fail:
 	local_inc(&rb->lost);
-	atomic64_inc(&event->lost_samples);
+	if (sample)
+		atomic64_inc(&event->lost_samples);
 	perf_output_put_handle(handle);
 out:
 	rcu_read_unlock();
@@ -268,14 +270,14 @@ int perf_output_begin_forward(struct perf_output_handle *handle,
 			      struct perf_sample_data *data,
 			      struct perf_event *event, unsigned int size)
 {
-	return __perf_output_begin(handle, data, event, size, false);
+	return __perf_output_begin(handle, data, event, size, false, true);
 }
 
 int perf_output_begin_backward(struct perf_output_handle *handle,
 			       struct perf_sample_data *data,
 			       struct perf_event *event, unsigned int size)
 {
-	return __perf_output_begin(handle, data, event, size, true);
+	return __perf_output_begin(handle, data, event, size, true, true);
 }
 
 int perf_output_begin(struct perf_output_handle *handle,
@@ -284,7 +286,7 @@ int perf_output_begin(struct perf_output_handle *handle,
 {
 
 	return __perf_output_begin(handle, data, event, size,
-				   unlikely(is_write_backward(event)));
+				   unlikely(is_write_backward(event)), false);
 }
 
 unsigned int perf_output_copy(struct perf_output_handle *handle,
-- 
2.37.2.789.g6183377224-goog

Re: [PATCH] perf/core: Increase lost_samples count only for samples

Posted by Namhyung Kim 3 years, 6 months ago

Hi Peter,

On Fri, Sep 2, 2022 at 11:19 AM Namhyung Kim <namhyung@kernel.org> wrote:
>
> The event->lost_samples count is intended to count (lost) sample records
> but it's also counted for non-sample records like PERF_RECORD_MMAP etc.
> This can be a problem when a sampling event tracks those side-band
> events together.
>
> As overflow handler for user events only calls perf_output_begin_
> {for,back}ward() before writing to the ring buffer, we can pass an
> additional flag to indicate that it's writing a sample record.

Could you please take a look?

Thanks,
Namhyung


>
> Fixes: 119a784c8127 ("perf/core: Add a new read format to get a number of lost samples")
> Signed-off-by: Namhyung Kim <namhyung@kernel.org>
> ---
>  kernel/events/ring_buffer.c | 14 ++++++++------
>  1 file changed, 8 insertions(+), 6 deletions(-)
>
> diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
> index 726132039c38..5f38ee4edbdb 100644
> --- a/kernel/events/ring_buffer.c
> +++ b/kernel/events/ring_buffer.c
> @@ -149,7 +149,7 @@ static __always_inline int
>  __perf_output_begin(struct perf_output_handle *handle,
>                     struct perf_sample_data *data,
>                     struct perf_event *event, unsigned int size,
> -                   bool backward)
> +                   bool backward, bool sample)
>  {
>         struct perf_buffer *rb;
>         unsigned long tail, offset, head;
> @@ -174,7 +174,8 @@ __perf_output_begin(struct perf_output_handle *handle,
>         if (unlikely(rb->paused)) {
>                 if (rb->nr_pages) {
>                         local_inc(&rb->lost);
> -                       atomic64_inc(&event->lost_samples);
> +                       if (sample)
> +                               atomic64_inc(&event->lost_samples);
>                 }
>                 goto out;
>         }
> @@ -256,7 +257,8 @@ __perf_output_begin(struct perf_output_handle *handle,
>
>  fail:
>         local_inc(&rb->lost);
> -       atomic64_inc(&event->lost_samples);
> +       if (sample)
> +               atomic64_inc(&event->lost_samples);
>         perf_output_put_handle(handle);
>  out:
>         rcu_read_unlock();
> @@ -268,14 +270,14 @@ int perf_output_begin_forward(struct perf_output_handle *handle,
>                               struct perf_sample_data *data,
>                               struct perf_event *event, unsigned int size)
>  {
> -       return __perf_output_begin(handle, data, event, size, false);
> +       return __perf_output_begin(handle, data, event, size, false, true);
>  }
>
>  int perf_output_begin_backward(struct perf_output_handle *handle,
>                                struct perf_sample_data *data,
>                                struct perf_event *event, unsigned int size)
>  {
> -       return __perf_output_begin(handle, data, event, size, true);
> +       return __perf_output_begin(handle, data, event, size, true, true);
>  }
>
>  int perf_output_begin(struct perf_output_handle *handle,
> @@ -284,7 +286,7 @@ int perf_output_begin(struct perf_output_handle *handle,
>  {
>
>         return __perf_output_begin(handle, data, event, size,
> -                                  unlikely(is_write_backward(event)));
> +                                  unlikely(is_write_backward(event)), false);
>  }
>
>  unsigned int perf_output_copy(struct perf_output_handle *handle,
> --
> 2.37.2.789.g6183377224-goog
>