lkml.org 
[lkml]   [2021]   [Mar]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: "struct perf_sample_data" alignment

* Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:

> > This seems to be it... (completely untested)
> >
> > ---
> > diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> > index 3f7f89ea5e51..918a296d2ca2 100644
> > --- a/include/linux/perf_event.h
> > +++ b/include/linux/perf_event.h
> > @@ -1032,7 +1032,9 @@ struct perf_sample_data {
> > u64 cgroup;
> > u64 data_page_size;
> > u64 code_page_size;
> > -} ____cacheline_aligned;
> > +};
> > +
> > +typedef struct perf_sample_data perf_sample_data_t ____cacheline_aligned;
> >
> > /* default value for data source */
> > #define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
> > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> > index b0c45d923f0f..f32c623abef6 100644
> > --- a/kernel/trace/bpf_trace.c
> > +++ b/kernel/trace/bpf_trace.c
> > @@ -923,7 +923,7 @@ __bpf_perf_event_output(struct pt_regs *regs, struct bpf_map *map,
> > * bpf_perf_event_output
> > */
> > struct bpf_trace_sample_data {
> > - struct perf_sample_data sds[3];
> > + perf_sample_data_t sds[3];
>
> bpf side doesn't care about about cacheline aligned.
> No need to add new typedef just for that.

So this structure is not supposed to be exposed to any ABI anywhere.

I did a (non-exhaustive) search of tooling, and there doesn't appear
to be any accidental exposure.

The in-kernel ABI interaction appears to be the following:

- In __perf_event_header_size() we only use fields within
perf_sample_data to size the header. Alignment won't change any of
the output.

- Ditto in perf_event__id_header_size().

I.e. I think we should just zap it per the patch below (untested).

Thanks,

Ingo

============>

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 3f7f89ea5e51..d75e03ff31ea 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -1032,7 +1032,7 @@ struct perf_sample_data {
u64 cgroup;
u64 data_page_size;
u64 code_page_size;
-} ____cacheline_aligned;
+};

/* default value for data source */
#define PERF_MEM_NA (PERF_MEM_S(OP, NA) |\
\
 
 \ /
  Last update: 2021-03-06 14:17    [W:0.048 / U:2.108 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site