Message-ID: <20201028211325.vstp37ukcvoilmk3@ast-mbp.dhcp.thefacebook.com>
Date:   Wed, 28 Oct 2020 14:13:25 -0700
From:   Alexei Starovoitov <alexei.starovoitov@...il.com>
To:     Jiri Olsa <jolsa@...hat.com>
Cc:     Steven Rostedt <rostedt@...dmis.org>, Jiri Olsa <jolsa@...nel.org>,
        Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Andrii Nakryiko <andriin@...com>, netdev@...r.kernel.org,
        bpf@...r.kernel.org, Martin KaFai Lau <kafai@...com>,
        Song Liu <songliubraving@...com>, Yonghong Song <yhs@...com>,
        John Fastabend <john.fastabend@...il.com>,
        KP Singh <kpsingh@...omium.org>, Daniel Xu <dxu@...uu.xyz>,
        Jesper Brouer <jbrouer@...hat.com>,
        Toke Høiland-Jørgensen <toke@...hat.com>,
        Viktor Malik <vmalik@...hat.com>
Subject: Re: [RFC bpf-next 00/16] bpf: Speed up trampoline attach

On Tue, Oct 27, 2020 at 03:28:03PM +0100, Jiri Olsa wrote:
> On Mon, Oct 26, 2020 at 09:30:14PM -0700, Alexei Starovoitov wrote:
> > On Thu, Oct 22, 2020 at 10:42:05AM -0400, Steven Rostedt wrote:
> > > On Thu, 22 Oct 2020 16:11:54 +0200
> > > Jiri Olsa <jolsa@...hat.com> wrote:
> > > 
> > > > I understand direct calls as a way that bpf trampolines and ftrace can
> > > > co-exist together - ebpf trampolines need that functionality of accessing
> > > > parameters of a function as if it was called directly and at the same
> > > > point we need to be able attach to any function and to as many functions
> > > > as we want in a fast way
> > > 
> > > I was sold that bpf needed a quick and fast way to get the arguments of a
> > > function, as the only way to do that with ftrace is to save all registers,
> > > which, I was told was too much overhead, as if you only care about
> > > arguments, there's much less that is needed to save.
> > > 
> > > Direct calls wasn't added so that bpf and ftrace could co-exist, it was
> > > that for certain cases, bpf wanted a faster way to access arguments,
> > > because it still worked with ftrace, but the saving of regs was too
> > > strenuous.
> > 
> > Direct calls in ftrace were done so that ftrace and trampoline can co-exist.
> > There is no other use for it.
> > 
> > Jiri,
> > could you please redo your benchmarking hardcoding ftrace_managed=false ?
> > If going through register_ftrace_direct() is indeed so much slower
> > than arch_text_poke() then something gotta give.
> > Either register_ftrace_direct() has to become faster or users
> > have to give up on co-existing of bpf and ftrace.
> > So far not a single user cared about using trampoline and ftrace together.
> > So the latter is certainly an option.
> 
> I tried that, and IIRC it was not much faster, but I don't have details
> on that.. but it should be quick check, I'll do it
> 
> anyway later I realized that for us we need ftrace to stay, so I abandoned
> this idea ;-) and started to check on how to keep them both together and
> just make it faster
> 
> also currently bpf trampolines will not work without ftrace being
> enabled, because ftrace is doing the preparation work during compile,
> and replaces all the fentry calls with nop instructions and the
> replace code depends on those nops...  so if we go this way, we would
> need to make this preparation code generic

I didn't mean that part.
I was talking about register_ftrace_direct() only.
Could you please still do ftrace_managed=false experiment?
Sounds like the time to attach/detach will stay the same?
If so, then don't touch ftrace internals then. What's the point?

Powered by blists - more mailing lists