Re: Problem with endianess of pahole BTF output for vmlinux

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Wed, Sep 09, 2020 at 11:02:24AM +0200, Ilya Leoshkevich escreveu:
> On Tue, 2020-09-08 at 13:18 -0700, Andrii Nakryiko wrote:
> > On Mon, Sep 7, 2020 at 9:02 AM Ilya Leoshkevich <iii@linux.ibm.com> wrote:
> > > On Sat, 2020-09-05 at 21:16 -0700, Tony Ambardar wrote:

> > > > I'm using GCC 8.4.0, binutils 2.34 and pahole 1.17, compiling on
> > > > an Ubuntu/x86_64 host and targeting both little- and big-endian
> > > > mips running on malta/qemu. When cross-compiling Linux 5.4.x LTS
> > > > and testing bpftool/BTF functionality on the target, I encounter
> > > > errors on big-endian targets:

> > > > > root@OpenWrt:/# bpftool btf dump file /sys/kernel/btf/vmlinux
> > > > > libbpf: failed to get EHDR from /sys/kernel/btf/vmlinux Error:
> > > > > failed to load BTF from /sys/kernel/btf/vmlinux: No error
> > > > > information
>
> > > > After investigating, the problem appears to be that "pahole -J"
> > > > running on the x86_64 little-endian host will always generate
> > > > raw BTF of native endianness (based on BTF magic), which causes
> > > > the error above on big-endian targets.

> > > > Is this expected? Is DEBUG_INFO_BTF supported in general when
> > > > cross-compiling? How does one generate BTF encoded for the
> > > > target endianness with pahole?

The BTF loader has support for endianness, its just the encoder that
doesn't :-\

I.e. pahole can grok a big endian BTF payload on a little endian machine
and vice-versa, just can't cross-build BTF payloads ATM.

> > Yes, it's expected, unfortunately. Right now cross-compiling to a
> > different endianness isn't supported. You can cross-compile only if
> > target endianness matches host endianness.

I agree that having this in libbpf is better, it should be done as part
of producing the result of the deduplication phase.

> > > > Thanks for any feedback or suggestions,
> > > > Tony
> > > 
> > > We have the same problem on s390, and I'm not aware of any solution
> > > at
> > > the moment. It would be great if we could figure out how to resolve
> > > this.
> > 
> > I'm working on extending BTF APIs in libbpf at the moment. Switching
> > endianness would be rather easy once all that is done. With these new
> > APIs it will be possible to switch pahole to use libbpf APIs to
> > produce BTF output and support arbitrary endianness as well. Right
> > now, I'd rather avoid implementing this in pahole, libbpf is a much
> > better place for this (and will require ongoing updates if/when we
> > introduce new types and fields to BTF).

Right, we could do it right after btf_dedup() and before
btf_elf__write(), doing the same process as in btf_loader.c, i.e.
checking if the ELF target arch is different in endianness and doing the
reverse of the loader.

> > Hope this plan works for you guys.
> 
> That sounds really good to me, thanks!

- Arnaldo