Re: [PATCH v2 bpf-next 04/17] libbpf: mark BPF subprogs with hidden visibility as static for BPF verifier

[Yonghong Song <yhs@fb.com>]

On 4/22/21 11:09 AM, Andrii Nakryiko wrote:
> On Wed, Apr 21, 2021 at 10:43 PM Yonghong Song <yhs@fb.com> wrote:
>>
>>
>>
>> On 4/16/21 1:23 PM, Andrii Nakryiko wrote:
>>> Define __hidden helper macro in bpf_helpers.h, which is a short-hand for
>>> __attribute__((visibility("hidden"))). Add libbpf support to mark BPF
>>> subprograms marked with __hidden as static in BTF information to enforce BPF
>>> verifier's static function validation algorithm, which takes more information
>>> (caller's context) into account during a subprogram validation.
>>>
>>> Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
>>> ---
>>>    tools/lib/bpf/bpf_helpers.h     |  8 ++++++
>>>    tools/lib/bpf/btf.c             |  5 ----
>>>    tools/lib/bpf/libbpf.c          | 45 ++++++++++++++++++++++++++++++++-
>>>    tools/lib/bpf/libbpf_internal.h |  6 +++++
>>>    4 files changed, 58 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/tools/lib/bpf/bpf_helpers.h b/tools/lib/bpf/bpf_helpers.h
>>> index 75c7581b304c..9720dc0b4605 100644
>>> --- a/tools/lib/bpf/bpf_helpers.h
>>> +++ b/tools/lib/bpf/bpf_helpers.h
>>> @@ -47,6 +47,14 @@
>>>    #define __weak __attribute__((weak))
>>>    #endif
>>>
>>> +/*
>>> + * Use __hidden attribute to mark a non-static BPF subprogram effectively
>>> + * static for BPF verifier's verification algorithm purposes, allowing more
>>> + * extensive and permissive BPF verification process, taking into account
>>> + * subprogram's caller context.
>>> + */
>>> +#define __hidden __attribute__((visibility("hidden")))
>>
>> To prevent potential external __hidden macro definition conflict, how
>> about
>>
>> #ifdef __hidden
>> #undef __hidden
>> #define __hidden __attribute__((visibility("hidden")))
>> #endif
>>
> 
> We do force #undef only with __always_inline because of the bad
> definition in linux/stddef.h And we check #ifndef for __weak, because
> __weak is defined in kernel headers. This is not really the case for
> __hidden, the only definition is in
> tools/lib/traceevent/event-parse-local.h, which I don't think we
> should worry about in BPF context. So I wanted to keep it simple and
> fix only if that really causes some real conflicts.
> 
> And keep in mind that in BPF code bpf_helpers.h is usually included as
> one of the first few headers anyways.

That is fine. Conflict of __hidden is a low risk and we can deal with it
later if needed.

> 
> 
>>> +
>>>    /* When utilizing vmlinux.h with BPF CO-RE, user BPF programs can't include
>>>     * any system-level headers (such as stddef.h, linux/version.h, etc), and
>>>     * commonly-used macros like NULL and KERNEL_VERSION aren't available through
> 
> [...]
> 
>>> @@ -698,6 +700,15 @@ bpf_object__add_programs(struct bpf_object *obj, Elf_Data *sec_data,
>>>                if (err)
>>>                        return err;
>>>
>>> +             /* if function is a global/weak symbol, but has hidden
>>> +              * visibility (or any non-default one), mark its BTF FUNC as
>>> +              * static to enable more permissive BPF verification mode with
>>> +              * more outside context available to BPF verifier
>>> +              */
>>> +             if (GELF_ST_BIND(sym.st_info) != STB_LOCAL
>>> +                 && GELF_ST_VISIBILITY(sym.st_other) != STV_DEFAULT)
>>
>> Maybe we should check GELF_ST_VISIBILITY(sym.st_other) == STV_HIDDEN
>> instead?
> 
> It felt like only STV_DEFAULT should be "exported", semantically
> speaking. Everything else would be treated as if it was static, except
> that C rules require that function has to be global. Do you think
> there is some danger to do it this way?
> 
> Currently static linker doesn't do anything special for STV_INTERNAL
> and STV_PROTECTED, so we could just disable those. Do you prefer that?

Yes, let us just deal with STV_DEFAULT and STV_HIDDEN. We already
specialized STV_HIDDEN, so we should not treat STV_INTERNAL/PROTECTED
as what they mean in ELF standard, so let us disable them for now.

> 
> I just felt that there is no risk of regression if we do this for
> non-STV_DEFAULT generically.
> 
> 
>>
>>> +                     prog->mark_btf_static = true;
>>> +
>>>                nr_progs++;
>>>                obj->nr_programs = nr_progs;
>>>
> 
> [...]
>