Re: [RFC bpf-next 13/16] libbpf: Add trampoline batch attach support

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Mon, Oct 26, 2020 at 04:15:48PM -0700, Andrii Nakryiko wrote:
> On Sun, Oct 25, 2020 at 12:12 PM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> >
> > On Fri, Oct 23, 2020 at 01:09:26PM -0700, Andrii Nakryiko wrote:
> > > On Thu, Oct 22, 2020 at 2:03 AM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> > > >
> > > > Adding trampoline batch attach support so it's possible to use
> > > > batch mode to load tracing programs.
> > > >
> > > > Adding trampoline_attach_batch bool to struct bpf_object_open_opts.
> > > > When set to true the bpf_object__attach_skeleton will try to load
> > > > all tracing programs via batch mode.
> > > >
> > > > Signed-off-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> > > > ---
> > >
> > > Assuming we go with the current kernel API for batch-attach, why can't
> > > libbpf just detect kernel support for it and just use it always,
> > > without requiring users to opt into anything?
> >
> > yea, it's rfc ;-) I wanted some simple usage of the
> > interface so it's obvious how it works
> >
> > if we'll end up with some batch interface I agree
> > we should use it as you suggested
> >
> > >
> > > But I'm also confused a bit how this is supposed to be used with BPF
> > > skeleton. You use case described in a cover letter (bpftrace glob
> > > attach, right?) would have a single BPF program attached to many
> > > different functions. While here you are trying to collect different
> > > programs and attach each one to its respective kernel function. Do you
> > > expect users to have hundreds of BPF programs in their skeletons? If
> > > not, I don't really see why adding this complexity. What am I missing?
> >
> > AFAIU when you use trampoline program you declare the attach point
> > at the load time, so you actually can't use same program for different
> > kernel functions - which would be great speed up actually, because
> > that's where the rest of the cycles in bpftrace is spent (in that cover
> > letter example) - load/verifier check of all those programs
> 
> Ah, I see, you are right. And yes, I agree, it would be nice to not
> have to clone the BPF program many times to attach to fentry/fexit, if
> the program itself doesn't really change.
> 
> >
> > it's different for kprobe where you hook single kprobe via multiple
> > kprobe perf events to different kernel function
> >
> > >
> > > Now it also seems weird to me for the kernel API to allow attaching
> > > many-to-many BPF programs-to-attach points. One BPF program-to-many
> > > attach points seems like a more sane and common requirement, no?
> >
> > right, but that's the consequence of what I wrote above
> 
> Well, maybe we should get rid of that limitation first ;)
>

I see this as 2 different things.. even when this would be
possible - attach single program to multiple trampolines,
you still need to install those trampolines via ftrace and
you'll end up in this discussion ;-)

jirka




[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux