[PATCH] backends,bpf: add proper relocation support

[Mark Wielaard]

Hi,

I added Richard to the CC, who added the original BPF support.
Who might remember where the R_BPF_MAP_FD comes from (see at the end).

On Fri, 2018-06-15 at 15:40 -0700, Yonghong Song wrote:
> Due to libdw does not have proper BPF relocation support,
> the pahole cannot display filenames correctly for objects
> with default llvm options. So we have to invent
> a special option "llc -march=bpf -mattr=dwarfris" to
> prevent llvm from generating cross-section dwarf relocation
> records (https://reviews.llvm.org/rL326505).
> The pahole related discussion is in linux netdev
> mailing list (http://lists.openwall.net/netdev/2018/06/15/38, etc.)
> 
> We would like to add proper BPF relocation support
> to libdw so eventually we could retire the special llc bpf
> flag "-mattr=dwarfris".

Yes. elfutils/libdwfl only does "simple relocations", but that is all
you need anyway.

Do you have a test file (binary for something simple/trivial generated
by llc -march=bpf that contains at least one reloc). I looked at your
implementation and I am sure it works correctly. But having a small
testfile is always a plus.

> The bpf relocations are defined in
> llvm_repo:include/llvm/BinaryFormat/ELFRelocs/BPF.def:
>   ELF_RELOC(R_BPF_NONE,        0)
>   ELF_RELOC(R_BPF_64_64,       1)
>   ELF_RELOC(R_BPF_64_32,      10)
> 
> Removed the relocation type R_BPF_MAP_FD whoes name does not
> confirm to llvm definition and replaced it with R_BPF_64_64.
> The BPF object is just a relocatible object, not an executable or
> a shared library, so assign ELF type to REL only in bpf_reloc.def.
>
> Tested locally with building pahole with this patch and
> pahole is able to display structures in bpf object file properly.

Patch looks good. Thanks. I'll add a ChangeLog entry because that is
what we still do.

> Signed-off-by: Yonghong Song <yhs@fb.com>
> ---
>  backends/Makefile.am   |  2 +-
>  backends/bpf_init.c    |  1 +
>  backends/bpf_reloc.def |  3 ++-
>  backends/bpf_symbol.c  | 54
> ++++++++++++++++++++++++++++++++++++++++++++++++++
>  libelf/elf.h           |  3 ++-
>  5 files changed, 60 insertions(+), 3 deletions(-)
>  create mode 100644 backends/bpf_symbol.c
> [...]
> diff --git a/libelf/elf.h b/libelf/elf.h
> index f7748983..940e88dd 100644
> --- a/libelf/elf.h
> +++ b/libelf/elf.h
> @@ -3848,7 +3848,8 @@ enum
>  /* BPF specific declarations.  */
>  
>  #define R_BPF_NONE		0	/* No reloc */
> -#define R_BPF_MAP_FD		1	/* Map fd to pointer */
> +#define R_BPF_64_64		1
> +#define R_BPF_64_32		10

We should sync this with glibc. This file really is a copy of elf/elf.h
in glibc, which we periodically sync. It would be good if all projects
agree on the constants.

I would like to understand where the R_BPF_MAP_FD comes from. But I
assume it was a typo for BPF_PSEUDO_MAP_FD from bpf.h (which has the
same constant number 1).

I'll sent a patch to libc-alpha@sourceware.org unless you beat me to
it.

Thanks,

Mark