Re: [RFC PATCH 0/5] Make eBPF programs output data to perf event

[Alexei Starovoitov <ast@plumgrid.com>]

On 7/1/15 4:58 AM, Peter Zijlstra wrote:
>
> But why create a separate trace buffer, it should go into the regular
> perf buffer.

+1

I think
+static char __percpu *perf_extra_trace_buf[PERF_NR_CONTEXTS];
is redundant.
It adds quite a bit of unnecessary complexity to the whole patch set.

Also the call to bpf_output_sample() is not effective unless program
returns 1. It's a confusing user interface.

Also you cannot ever do:
  	BPF_FUNC_probe_read,
+	BPF_FUNC_output_sample,
  	BPF_FUNC_ktime_get_ns,
new functions must be added to the end.

Why not just do:
perf_trace_buf_prepare() + perf_trace_buf_submit() from the helper?
No changes to current code.
No need to call __get_data_size() and other overhead.
The helper can be called multiple times from the same program.
imo much cleaner.

Also how about calling this helper:
bpf_trace_buf_submit(void *stack_ptr, int size) ?
bpf_output_sample, I think, is odd name. It's not a sample.
May be other name?