Re: [RFC] elf: Implement filtering of symbols historically defined in libpthread

[Florian Weimer via Libc-alpha <libc-alpha@sourceware.org>]

* H. J. Lu:

> On Tue, May 4, 2021 at 5:33 AM Florian Weimer <fweimer@redhat.com> wrote:
>>
>> * H. J. Lu:
>>
>> > On Tue, May 4, 2021 at 5:27 AM Florian Weimer <fweimer@redhat.com> wrote:
>> >>
>> >> * H. J. Lu:
>> >>
>> >> >> The patch attempts to detect old main programs by looking for the
>> >> >> GLIBC_2.34 symbol version.  Since we added __libc_start_main@@GLIBC_2.34
>> >> >> (which is called from our version of _start), all standard main programs
>> >> >> linked with glibc 2.34 or later will have this symbol version.
>> >>
>> >> > Can we invent a symbol or version to detect the older binaries?
>> >> > If not, can GNU property, ABI note, .... help here?
>> >>
>> >> I think we have all we need due to __libc_start_main@@GLIBC_2.34.  It
>> >> was an unrelated change, but it helps here as well.
>> >
>> > So your patch isn't required?  Can you add some tests to verify it?
>>
>> No, the patch uses the absence of GLIBC_2.34 for detecting old binaries.
>>
>> We cannot use symbol versions in other ways because one key
>> characteristic of weak references and underlinking is the lack of
>> version information on the symbol itself.
>
> Can we add support for binary testcases like this, even if it can only
> run on a single target?  It shouldn't be too hard.

Yes, we can verify the binding status of weak symbols with a smaller
assembler program that was linked against a stub libc.so.6 library that
only contains the soname and no symbols.

As the code is architecture-agnostic, testing on e.g. x86-64 should be
sufficient.

We cannot easily verify the behavior of real-world binaries because
re-linking them with current binutils probably changes behavior.

Thanks,
Florian