Re: Progress since 1.1.9

[Rich Felker <dalias@libc.org>]

On Thu, May 28, 2015 at 09:47:34PM +0200, Szabolcs Nagy wrote:
> * Rich Felker <dalias@libc.org> [2015-05-28 13:12:41 -0400]:
> 
> > On Thu, May 28, 2015 at 01:45:15PM +0200, Szabolcs Nagy wrote:
> > > * Rich Felker <dalias@libc.org> [2015-05-27 19:13:08 -0400]:
> > > > The one roadmap item I don't have any progress on is what to do with
> > > > the libgcc_s symbol version mess, and I think we should probably just
> > > > hold off until the next release cycle for that now.
> > > 
> > > in case anyone wonders about the issue:
> > > 
> > > libgcc_s has a symbolic reference to an internal deprecated
> > > symbol with no default version on x86 (in gcc trunk).
> > > (so no @version only @@version which makes it invisible to
> > > musl and ld).
> > > 
> > > the deprecated function is a constructor in libgcc_s that
> > > initializes an unused struct in libgcc_s.
> > > 
> > > the only reason the symbol got versioned this way is because
> > > they want old binaries to work and remove the symbol from
> > > libgcc_s for new binaries, while still using the same symbol
> > > name in new binaries but with different abi behaviour: now
> > > it's only available in libgcc.a to fix their ifunc hack for
> > > multi-versioning.
> > > 
> > > since it's a symbol internal to libgcc and the semantics
> > > of the symbol changed they could have just used a new name
> > > and left the old one alone (so old stuff is guaranteed to not
> > > break including musl and new binaries use the new symbol with
> > > new semantics).
> > 
> > Yes but if the new symbol is not linkable like they made it, then old
> > musl-linked binaries depending on getting it from libgcc_s would fail
> > to find it at dynamic-link time and error out. I don't know whether
> > this can happen without use of the multiversioning feature, though,
> > which would not have worked with musl anyway since we don't have
> > ifunc.
> 
> old binaries don't reference the new name unless new name == old name.
> 
> there is no reason to make the old name unlinkable in libgcc_s.so if
> the new symbol has new name.  (old binaries only reference the old
> name which is available in libgcc_s.so and libgcc.a like before, new
> binaries only reference the new name that is only in libgcc.a).
> (this is what my patch does).

Yes, that would have been preferable, but it looks hard to get gcc to
change this... :(

> (it was possible to reference __cpu_indicator_init in libgcc_s.so without
> using ifunc based multi-versioning: the x86 specific __builtin_cpu_init
> can be called directly to make __builtin_cpu_is and __builtin_cpu_supports
> work, but i think only c++ code used -lgcc_s).

Bleh.

> > > so i still think my proposed libgcc patch makes more sense
> > > than versioning:
> > > https://gcc.gnu.org/ml/gcc-patches/2015-05/msg00899.html
> > > 
> > > gcc is unlikely to fix this, but this is a nonsense usecase
> > > (and there is no other known use of versioning that is broken
> > > with musl.. in theory other libs may deprecate symbols in a
> > > similar way while still keeping symbolic references to them,
> > > but such use was not yet observed).
> > 
> > Are you sure? have you tried building C++ programs with gcc 3.x or 4.2
> > then using a libstdc++ from recent gcc? Unless the soname is
> > different, I suspect they're using symbol versions to make it "work"
> > and it will probably break catastrophically.
> 
> ok that may cause problems (but won't make the dynamic linking fail:
> all versioned symbols in libstdc++ has default version).

If so, it's worse than making dynamic linking fail -- it presumably
causes silent runtime memory corruption due to using wrong data
structures.

> but mixing libs linked with different versions of libstdc++ is
> broken anyway if any two libs happen to use the same symbol with
> different versions.

I agree completely that symbol versioning is a horrible idea and does
not properly solve the problem it was invented to solve. (And this
problem can be addressed much better, though still not perfectly, with
macro-based, inline-function-based, or asm-label-based redirects in
header files.) But I believe the way gcc is doing things _does_ work
reliably when you do not have C++ types crossing DSO interface
boundaries, and that's a fairly large portion of real-world usage.
(And when C++ types do cross DSO boundaries, it's usually
intra-project DSOs, not system-wide ones in the public lib dirs,
because people generally know C++ ABI is continually broken for public
library interfaces.)

> > > in musl-gcc it can be worked around by preloading a noop
> > > __cpu_indicator_init and in a musl based gcc it can be
> > > patched out.
> > 
> > But it sounds like that patch will be unacceptable for upstream. There
> > are alternatives we could do, like providing in musl a list of useless
> > symbols to ignore (resolve to a nop func or a dummy data symbol) if
> > they're not found, but these are all a bit hackish and I worry more
> > stuff with symbol versioning will come back to bite us in the future.
> > The other option would be to teach gcc and binutils that musl does not
> > support symbol versions, but I worry that might break even more things
> > that expect versions to be available on Linux but that work fine with
> > musl's current dummying-out of version matching.
> 
> if things use symbol versioning then musl will have problems.
> 
> but is that really widespread?

I recall ALSA libs were using it; this is why we added the hack that
we have now, always using the version ld would choose instead of the
first-found version.

Rich