Re: [PATCH net-next 2/8] perf, bpf: allow bpf programs attach to tracepoints

[Steven Rostedt <rostedt@goodmis.org>]

On Mon, 18 Apr 2016 18:15:04 -0700
Alexei Starovoitov <ast@fb.com> wrote:

> On 4/18/16 3:16 PM, Steven Rostedt wrote:

> Yes. That what I referred to in below 'a struct to pass args'...
> But, fine, will try to optimize the size further.
> Frankly much bigger .text savings will come from combining
> trace_event_raw_event_*() with perf_trace_*()
> Especially if you're ok with copying tp args into perf's percpu
> buffer first and then copying into ftrace's ring buffer.
> Then we can half the number of such auto-generated functions.

I'm only fine with that when we filter. Otherwise we just lost all the
benefits of zero copy in the first place.

> 
> >> Passing more args or creating a struct to pass args only going to
> >> hurt performance without much reduction in .text size.
> >> tinyfication folks will disable tracepoints anyway.
> >> Note that the most common case is bpf returning 0 and not even
> >> calling perf_trace_buf_submit() which is already slow due
> >> to so many args passed on stack.
> >> This stuff is called million times a second, so every instruction
> >> counts.  
> >
> > Note, that doesn't matter if you are bloating the kernel for the 99.9%
> > of those that don't use bpf.
> >
> > Please remember this! Us tracing folks are second class citizens! If
> > there's a way to speed up tracing by 10%, but in doing so we cause
> > mainline to be hurt by over 1%, we shouldn't be doing it. Tracing and
> > hooks on tracepoints are really not used by many people. Don't fall
> > into Linus's category of "my code is the most important". That's
> > especially true for tracing.  
> 
> tracing was indeed not used that often in the past, but
> bpf+tracing completely changed the picture. It's no longer just
> debugging. It is the first class citizen that runs 24/7 in production
> and its performance and lowest overhead are crucial.

Still, 99.9% of users don't use it.

-- Steve