Re: bpftool gen object doesn't handle GCC built BPF ELF files

["Jose E. Marchesi" <jose.marchesi@oracle.com>]

> On Wed, Jul 6, 2022 at 10:13 AM James Hilliard
> <james.hilliard1@gmail.com> wrote:
>>
>> Note I'm testing with the following patches:
>> https://lore.kernel.org/bpf/20220706111839.1247911-1-james.hilliard1@gmail.com/
>> https://lore.kernel.org/bpf/20220706140623.2917858-1-james.hilliard1@gmail.com/
>>
>> It would appear there's some compatibility issues with bpftool gen and
>> GCC, not sure what side though is wrong here:
>> /home/buildroot/buildroot/output/per-package/systemd/host/sbin/bpftool
>> gen object src/core/bpf/restrict_ifaces/restrict-ifaces.bpf.o
>> src/core/bpf/restrict_ifaces/restrict-ifaces.bpf.unstripped.o
>> libbpf: failed to find BTF info for global/extern symbol 'sd_restrictif_i'
>> Error: failed to link
>> 'src/core/bpf/restrict_ifaces/restrict-ifaces.bpf.unstripped.o':
>> Unknown error -2 (-2)
>>
>> Relevant difference seems to be this:
>> GCC:
>> [55] FUNC 'sd_restrictif_i' type_id=47 linkage=static
>> Clang:
>> [27] FUNC 'sd_restrictif_i' type_id=26 linkage=global

For functions GCC generates a BTF_KIND_FUNC entry, which has no linkage
information, or so we thought: I just looked at bpftool/btf.c and I
found the linkage info for function types is expected to be encoded in
the vlen field of BTF_KIND_FUNC entries (why not adding a btf_func
instead???) which is surprising to say the least.

We are changing GCC to encode the linkage info in vlen for these types.
Thanks for reporting this.

> GCC is wrong, clearly. This function is global ([0]) and libbpf
> expects it to be marked as such in BTF.
>
> https://github.com/systemd/systemd/blob/main/src/core/bpf/restrict_ifaces/restrict-ifaces.bpf.c#L42-L50
>
>
>> GCC:
>>
>> [1] INT 'signed char' size=1 bits_offset=0 nr_bits=8 encoding=UNKN
>> [2] INT 'unsigned char' size=1 bits_offset=0 nr_bits=8 encoding=CHAR
>> [3] TYPEDEF '__u8' type_id=2
>> [4] CONST '(anon)' type_id=3
>
> [...]