Re: [RFC PATCH 1/5] bpf: Put perf_events check ahead of bpf prog

[Alexei Starovoitov <ast@plumgrid.com>]

On 7/1/15 10:52 PM, Wangnan (F) wrote:
> I'd like to discuss with you about the correctness of our
> understanding. Do you have any strong reason to put BPF filters at such
> an early stage?

the obvious reason is performance.
It is so much faster to run generated
'if (bpf_get_current_pid() != expected_pid) return'
instead of going through __get_data_size,
perf_trace_buf_prepare, store_trace_args,
perf_trace_buf_submit->perf_tp_event_match->filter_match_preds.
bpf is the fastest way to filter out things, so it should be first.

I would argue that even for regular samples (cycle counts and so on),
we should be using this tiny bpf prog to filter by pid.
It's around 5 or so instructions that perf can always use instead
of doing 'common_pid != expected_pid' > filter. Disturbance to the
whole kernel will be much lower. Obviously there are no hooks for
bpf programs in regular samples, but I think it's worth doing.