Re: [PATCH v2] elf: Implement filtering of symbols historically defined in libpthread

["H.J. Lu" <hjl.tools@gmail.com>]

On Wed, May 5, 2021 at 12:08 PM Florian Weimer <fweimer@redhat.com> wrote:
>
> * H. J. Lu:
>
> > On Wed, May 5, 2021 at 11:48 AM Florian Weimer <fweimer@redhat.com> wrote:
> >>
> >> * H. J. Lu:
> >>
> >> > On Wed, May 5, 2021 at 11:28 AM Florian Weimer <fweimer@redhat.com> wrote:
> >> >>
> >> >> * H. J. Lu:
> >> >>
> >> >> > Why can't we just hide symbols for undefined, weak, unversioned references
> >> >> > from the old binaries?
> >> >>
> >> >> You mean, in general, no matter if the symbol was previously in
> >> >> libpthread or not?
> >> >>
> >> >> Some such symbols had definitions in libc before and where bound.
> >> >
> >> > If they were defined in libc.so to begin with, why are they unversioned?
> >>
> >> I assume some people use stub libraries for linking.
> >
> > Do we really support these binaries?  The symbol versions will be wrong
> > on them.
>
> Well, I guess I have a definitive answer now.  I keep some old binaries
> and their sources around here (along with wrapper scripts to run them on
> current systems, although I haven't checked lately if they still work):
>
>   <https://pagure.io/glibc/glibc-test-binaries/>
>
> gcc-2.7.2.3/i386/root/usr/lib/gcc-lib/i486-linux/2.7.2.3/cc1plus
> contains this in its dynamic symbol table:
>
>    16: 08048c9c    129 FUNC    GLOBAL DEFAULT    UNDEF memcpy
>    17: 08048e8c     62 FUNC    GLOBAL DEFAULT    UNDEF strcmp
>    18: 08048ccc    304 FUNC    WEAK   DEFAULT    UNDEF malloc
>    19: 08048f5c    195 FUNC    WEAK   DEFAULT    UNDEF free
>    20: 08048c7c    136 FUNC    GLOBAL DEFAULT    UNDEF getenv
>    21: 08048dec    565 FUNC    WEAK   DEFAULT    UNDEF realloc
>    22: 08048e4c    124 FUNC    WEAK   DEFAULT    UNDEF fopen
>    23: 08048e7c    192 FUNC    WEAK   DEFAULT    UNDEF fclose
>    24: 08048c8c    186 FUNC    GLOBAL DEFAULT    UNDEF qsort
>    25: 08048d1c     50 FUNC    GLOBAL DEFAULT    UNDEF __sigsetjmp
>    26: 08048c2c     80 FUNC    GLOBAL DEFAULT    UNDEF longjmp
>    27: 08048dac    226 FUNC    WEAK   DEFAULT    UNDEF signal
>    28: 08048cfc     88 FUNC    GLOBAL DEFAULT    UNDEF bzero
>    29: 08048e5c    104 FUNC    GLOBAL DEFAULT    UNDEF memset
>    30: 08048f1c     38 FUNC    GLOBAL DEFAULT    UNDEF atoi
>    31: 08048dbc    159 FUNC    GLOBAL DEFAULT    UNDEF strncmp
>    32: 08048cbc    641 FUNC    WEAK   DEFAULT    UNDEF system
>    33: 08048e9c   1460 FUNC    WEAK   DEFAULT    UNDEF getcwd
>    34: 08048d7c    426 FUNC    GLOBAL DEFAULT    UNDEF strcat
>    35: 08048c3c     39 FUNC    GLOBAL DEFAULT    UNDEF strcpy

This binary was linked against the old glibc which doesn't have
symbol version.  This binary may have other issues with today's glibc.

> I assume an old version of ld copied the weak symbol status from the
> definition in a shared object to the program.  We obviously cannot hide
> these weak symbol references.  So my hunch was right that we have to
> make visibility decisions on a per-symbol basis.
>
> Thanks,
> Florian
>


-- 
H.J.