perf/core: Trim dyn_size if raw data is absent

[PATCH] perf/core: Trim dyn_size if raw data is absent

Posted by Yabin Cui 1 week, 4 days ago

Currently, perf_tp_event() always allocates space for raw sample data,
even when the PERF_SAMPLE_RAW flag is not set. This leads to unused
spaces within generated sample records.

This patch reduces dyn_size when PERF_SAMPLE_RAW is not present,
ensuring sample records use only the necessary amount of space.

Fixes: 0a9081cf0a11 ("perf/core: Add perf_sample_save_raw_data() helper")
Signed-off-by: Yabin Cui <yabinc@google.com>
---
 kernel/events/core.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 724e6d7e128f..d68ecdc264d3 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -7688,6 +7688,10 @@ void perf_prepare_sample(struct perf_sample_data *data,
 		data->raw = NULL;
 		data->dyn_size += sizeof(u64);
 		data->sample_flags |= PERF_SAMPLE_RAW;
+	} else if ((data->sample_flags & ~sample_type) & PERF_SAMPLE_RAW) {
+		data->dyn_size -= data->raw->size + sizeof(u32);
+		data->raw = NULL;
+		data->sample_flags &= ~PERF_SAMPLE_RAW;
 	}
 
 	if (filtered_sample_type & PERF_SAMPLE_BRANCH_STACK) {
-- 
2.44.0.769.g3c40516874-goog

Re: [PATCH] perf/core: Trim dyn_size if raw data is absent

Posted by Namhyung Kim 1 week ago

Hello Yabin,

CC-ing the bpf list.

On Thu, Apr 25, 2024 at 3:05 PM Yabin Cui <yabinc@google.com> wrote:
>
> Currently, perf_tp_event() always allocates space for raw sample data,
> even when the PERF_SAMPLE_RAW flag is not set. This leads to unused
> spaces within generated sample records.
>
> This patch reduces dyn_size when PERF_SAMPLE_RAW is not present,
> ensuring sample records use only the necessary amount of space.

Right, it seems bpf-output and tracepoint events set the flags without
checking PERF_SAMPLE_RAW.  Can you fix the callsites instead?
Or we can add perf_event argument to perf_sample_save_raw_data()
and check the flag inside.

We might reject the output data when it's not opened with the flag.
But I'm afraid it might break some existing BPF programs.

Thanks,
Namhyung

>
> Fixes: 0a9081cf0a11 ("perf/core: Add perf_sample_save_raw_data() helper")
> Signed-off-by: Yabin Cui <yabinc@google.com>
> ---
>  kernel/events/core.c | 4 ++++
>  1 file changed, 4 insertions(+)
>
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index 724e6d7e128f..d68ecdc264d3 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -7688,6 +7688,10 @@ void perf_prepare_sample(struct perf_sample_data *data,
>                 data->raw = NULL;
>                 data->dyn_size += sizeof(u64);
>                 data->sample_flags |= PERF_SAMPLE_RAW;
> +       } else if ((data->sample_flags & ~sample_type) & PERF_SAMPLE_RAW) {
> +               data->dyn_size -= data->raw->size + sizeof(u32);
> +               data->raw = NULL;
> +               data->sample_flags &= ~PERF_SAMPLE_RAW;
>         }
>
>         if (filtered_sample_type & PERF_SAMPLE_BRANCH_STACK) {
> --
> 2.44.0.769.g3c40516874-goog
>

Re: [PATCH] perf/core: Trim dyn_size if raw data is absent

Posted by Yabin Cui 5 days, 11 hours ago

Hi Namyung,

Thanks for reviewing the patch! Fixing the callsites is a better idea.
I have sent a v2 patch with name [PATCH v2] perf/core: Save raw sample
data conditionally based on sample type.
Rejecting tracepoint events without PERF_SAMPLE_RAW will break my use
case in Android. So I hope we don't do that.

Thanks,
Yabin






On Mon, Apr 29, 2024 at 2:59 PM Namhyung Kim <namhyung@kernel.org> wrote:
>
> Hello Yabin,
>
> CC-ing the bpf list.
>
> On Thu, Apr 25, 2024 at 3:05 PM Yabin Cui <yabinc@google.com> wrote:
> >
> > Currently, perf_tp_event() always allocates space for raw sample data,
> > even when the PERF_SAMPLE_RAW flag is not set. This leads to unused
> > spaces within generated sample records.
> >
> > This patch reduces dyn_size when PERF_SAMPLE_RAW is not present,
> > ensuring sample records use only the necessary amount of space.
>
> Right, it seems bpf-output and tracepoint events set the flags without
> checking PERF_SAMPLE_RAW.  Can you fix the callsites instead?
> Or we can add perf_event argument to perf_sample_save_raw_data()
> and check the flag inside.
>
> We might reject the output data when it's not opened with the flag.
> But I'm afraid it might break some existing BPF programs.
>
> Thanks,
> Namhyung
>
> >
> > Fixes: 0a9081cf0a11 ("perf/core: Add perf_sample_save_raw_data() helper")
> > Signed-off-by: Yabin Cui <yabinc@google.com>
> > ---
> >  kernel/events/core.c | 4 ++++
> >  1 file changed, 4 insertions(+)
> >
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index 724e6d7e128f..d68ecdc264d3 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -7688,6 +7688,10 @@ void perf_prepare_sample(struct perf_sample_data *data,
> >                 data->raw = NULL;
> >                 data->dyn_size += sizeof(u64);
> >                 data->sample_flags |= PERF_SAMPLE_RAW;
> > +       } else if ((data->sample_flags & ~sample_type) & PERF_SAMPLE_RAW) {
> > +               data->dyn_size -= data->raw->size + sizeof(u32);
> > +               data->raw = NULL;
> > +               data->sample_flags &= ~PERF_SAMPLE_RAW;
> >         }
> >
> >         if (filtered_sample_type & PERF_SAMPLE_BRANCH_STACK) {
> > --
> > 2.44.0.769.g3c40516874-goog
> >