On Mon, May 24, 2021 at 12:31 PM John Fastabend
<john.fastabend@xxxxxxxxx> wrote:
>
> Andrii Nakryiko wrote:
> > On Mon, May 24, 2021 at 8:05 AM Lorenz Bauer <lmb@xxxxxxxxxxxxxx> wrote:
> > >
> > > Hi Andrii,
> > >
> > > A user of bpf2go [1] recently ran into the problem of PT_REGS not
> > > being defined after including bpf_tracing.h. It turns out this is
> > > because we by default compile for bpfel / bpfeb so the logic in the
> > > header file doesn't kick in. I originally filed [2] as a quick fix,
> > > but looking at the issue some more made me wonder how to fit this into
> > > bpf2go better.
> > >
> > > Basically, the convention of bpf2go is that the compiled BPF is
> > > checked into the source code repository to facilitate distributing BPF
> > > as part of Go packages (as opposed to libbpf tooling which doesn't
> > > include generated source). To make this portable, bpf2go by default
> > > generates both bpfel and bpfeb variants of the C.
> > >
> > > However, code using bpf_tracing.h is inherently non-portable since the
> > > fields of struct pt_regs differ in name between platforms. It seems
> > > like this forces one to compile to each possible __TARGET_ARCH
> > > separately. If that is correct, could we extend CO-RE somehow to cover
> > > this as well?
> >
> > If there are enums/types/fields that we can use to reliably detect the
> > platform, then yes, we can have a new set of helpers that would do
> > this with CO-RE. Someone will need to investigate how to do that for
> > all the platforms we have. It's all about finding something that's
> > already in the kernel and can server as a reliably indicator of a
> > target architecture.
> >
> > >
> > > If that isn't possible, I want to avoid compiling and shipping BPF for
> > > each possible __TARGET_ARCH_xxx by default. Instead I would like to
> > > achieve:
> > > * Code that doesn't use bpf_tracing.h is distributed in bpfel and bpfeb variants
> > > * Code that uses bpf_tracing.h has to explicitly opt into the
> > > supported platforms via a flag to bpf2go
> > >
> > > The latter point is because the go tooling has to know the target arch
> > > to be able to generate the correct Go wrappers. How would you feel
> > > about adding something like the following to bpf_tracing.h:
> >
> > Well, obviously I'm not a fan of even more magic #defines. But I think
> > we can achieve a similar effect with a more "lazy" approach. I.e., if
> > user tries to use PT_REGS_xxx macros but doesn't specify the platform
> > -- only then it gets compilation errors. There is stuff in
> > bpf_tracing.h that doesn't need pt_regs, so we can't just outright do
> > #error unconditinally. But we can do something like this:
> >
> > #else /* !bpf_target_defined */
> >
> > #define PT_REGS_PARM1(x) _Pragma("GCC error \"blah blah something
> > user-facing\"")
> >
> > ... and so on for all macros
> >
> > #endif
> >
> > Thoughts?
>
> I don't use bpf_tracing.h, but...
>
> Can we do this similar to feature discovery and simply have the
> user space tell us the platform? I at least do this fairly
> frequently so have infra in place on my side for user space to
> push down feature flags/fields. One of those could be platform then
> we just need helpers,
>
> get_pt_regs_parm() {
> if (bpf_target_defined == is_x86)
> ...
> else if (bpf_target_defined == is_foo)
> ...
> else {
> hard_load_error()
> }
> }
>
> I think we are suggesting the same thing? Then user would need to
> have bpf_target_Defined set but that should be OK and the other
> conditions should all look like dead code.
It's almost what I propose, except I suggest to get rid of the need
for a user to specify the arch they expect/need, and use CO-RE to
detect this. What you propose would work, but it's usability is worse
than the CO-RE-based variant and it requires to hard-code struct
pt_regs exact definitions for each platform in bpf_tracing.h, which
sucks. And it would be like a third way to achieve the same thing
(with a different set of tradeoffs).
>
> >
> >
> > While on the topic, I've noticed that we added BPF_SEQ_PRINTF and
> > BPF_SNPRINTF into bpf_tracing.h, which seems like not the best fit
> > (definitely not for BPF_SNPRINTF). Florent, can you please help moving
> > them into bpf_helpers.h, as it's really more generic functionality
> > rather than low-level tracing primitives. I think it was put here
> > because we needed ___bpf_narg macros, but I'd rather copy/paste them
> > into bpf_helpers.h (they won't change at all) and put
> > BPF_SNPRINTF/BPF_SEQ_PRINTF into bpf_helpers.h.
> >
> > >
> > > --- a/tools/lib/bpf/bpf_tracing.h
> > > +++ b/tools/lib/bpf/bpf_tracing.h
> > > @@ -25,6 +25,9 @@
> > > #define bpf_target_sparc
> > > #define bpf_target_defined
> > > #else
> > > + #if defined(BPF_REQUIRE_EXPLICIT_TARGET_ARCH)
> > > + #error BPF_REQUIRE_EXPLICIT_TARGET_ARCH set and no
> > > target arch defined
> > > + #endif
> > > #undef bpf_target_defined
> > > #endif
> > >
> > > bpf2go would always define BPF_REQUIRE_EXPLICIT_TARGET_ARCH. If the
> > > user included bpf_tracing.h they get this error. They can then add
> > > -target amd64, etc. and the tooling then defines __TARGET_ARCH_x86_64
> > >
> > > 1: https://pkg.go.dev/github.com/cilium/ebpf/cmd/bpf2go
> > > 2: https://github.com/cilium/ebpf/issues/305
> > >
> > > --
> > > Lorenz Bauer | Systems Engineer
> > > 6th Floor, County Hall/The Riverside Building, SE1 7PB, UK
> > >
> > > www.cloudflare.com
>
>
