Re: Weak references, --as-needed, and pthreads

[Florian Weimer <fweimer at redhat dot com>]

On 06/13/2018 06:57 PM, Cary Coutant wrote:
> > > (7) When ld.so binds a symbol with no version information, it does not
> > > bind to the default version; instead it binds to the older version
> > > (not sure if that's by chance or it it deliberately picks the first or
> > > oldest version).
> >
> > In general, it really must pick the base version for the symbol because the
> > binary could have been produced before symbol versioning was introduced for
> > this particular library.
> >
> > In particular, a reference (weak or strong) to any glibc symbol really must
> > have an attached symbol version, especially if a weak function symbol is
> > actually used to call the function, otherwise the binary is invalid and
> > might fail to work with any future glibc update.
>
> Makes sense. But what do I do if the library with the definition is
> not needed? Should I keep the version info anyway?

Yes, and it should emit a weak version, so that loading of the library 
with the weak reference is not rejected if the referenced library does 
not provide that symbol version.

> Alan's comment from the original patch said:
>
> > Not quite so simple as just testing
> > whether the reference is strong before setting is_needed, since we
> > can't have symbol versions coming from a non-loaded shared lib.  If
> > you do, ld.so segfaults.
>
> But that was five years ago -- is that still a problem with ld.so?

I'm not aware of any segfaults, and I couldn't find any defaults.  Weak 
versions are however a bit obscure, and perhaps they are not implemented 
correctly.  I'm not even sure if we have any tests for this 
functionality; it's kind of hard to spot this by reading sources.

> I'm thinking that if I do end up discarding the version info because
> it was derived from a non-needed library, I should at least print a
> warning.

I'm still not sure that discarding a version that is known to be needed 
at run time is the right thing to do.

Thanks,
Florian