Re: some odd library loading errors

[Rich Felker <dalias@libc.org>]

On Tue, Nov 17, 2015 at 05:34:00PM +0100, Szabolcs Nagy wrote:
> * Rich Felker <dalias@libc.org> [2015-11-17 10:55:45 -0500]:
> > On Tue, Nov 17, 2015 at 04:45:16PM +0100, Szabolcs Nagy wrote:
> > > 
> > > i guess --disable-gnu-indirect-function could be the default when
> > > gcc targets musl (assuming musl will not implement ifuncs anytime soon).
> > 
> > I think that's a reasonable assumption. Based on what I've seen
> > lately, rather than finding solutions to the problems we knew about
> > already, the rabbit hole keeps going deeper...
> > 
> > If you have reason to believe otherwise could you explain how we might
> > reasonably support ifunc?
> > 
> 
> well we can add support for ifunc by..
> 
> calling the ifunc resolver during reloc processing when STT_GNU_IFUNC
> symbol or R_*_IRELATIVE reloc is found and in case of static linking
> the relocs between __rel_iplt_{start,end} should be processed somehow.

This is probably a significant code size burden for static linking
unless there's some way to avoid linking it when ifunc is not used.
There's also the static-pie case you haven't considered, which would
require its own separate implementation. Presumably that would be
processing R_*_IRELATIVE since symbolic relocations are not permitted
here.

> the only ugliness is that passing arguments to the resolver is arch
> specific.. and that there is no guarantee what the resolver might do
> (but that's the same with glibc and it is something the user should
> worry about)

Well that's ugly too.

> i think even if musl does the reloc processing in different order
> than glibc, this might work in practice and if the resolver crashes
> we can say that it invoked ub.

It seems to me that order issues, except for circular dependencies,
are less of an issue when you have a lazy resolver; if A gets resolved
first but A needs B, then calling B invokes the lazy resolver and all
is fine. Since musl doesn't (and won't) have lazy resolver, the whole
setup seems much more fragile.

Then of course there's just the general issue of having arbitrary code
run in callbacks called from the dynamic linker with things in
inconsistent state. Without a formal spec for what ifunc resolvers are
actually allowed to do, we can easily fall into a trap of "having to"
support whatever happened to work on one particular version of the
implementation.

Rich