Re: [musl] [PATCH] ldso: add basic ifunc support

[Rich Felker <dalias@libc.org>]

On Tue, Aug 23, 2022 at 10:03:32PM +0200, Szabolcs Nagy wrote:
> * Ben Noordhuis <info@bnoordhuis.nl> [2022-08-23 20:41:57 +0200]:
> > Not perfect yet because it doesn't handle resolver functions that need
> > relocations themselves but basic functionality works. That is, this
> > works:
> > 
> >     typedef void (*Func)(void);
> >     void f(void) __attribute__((ifunc("g")));
> >     Func g(void) { return h; }
> >     void h(void) { abort(); }
> > 
> > But returning an extern function does not:
> > 
> >     typedef void (*Func)(void);
> >     void f(void) __attribute__((ifunc("g")));
> >     Func g(void) { return abort; } // segfaults
> 
> there are a lot more problems than that.
> 
> the patch only works if f is hidden or binds locally in some other
> way (e.g. in an executable) otherwise references to f will generate
> symbolic relocations not IRELATIVE.
> 
> resolving symbolic relocation for ifunc symbols requires the ldso
> to look at the symbol type (STT_GNU_IFUNC) and do the indirect call
> based on that.
> 
> the main issue is that code is executed before the module is fully
> relocated (in fact it may execute code in other libs before those are
> relocated because symbol lookup can process not yet relcated libs)
> which is problematic if the resolver function itself depends on
> relocations (e.g. return h only works if constructing the address
> of h does not require dynamic relocs or those relocs happen to get
> resolved before the IRELATIVE, glibc tries a bit to shuffle ifunc
> relocs to the end of reloc processing, it was never fixed completely
> iirc because copy relocs and circular deps made things complicated).
> 
> and of course static linking is not addressed (where there are
> further ordering issues since it is target dependent if ifunc has
> to be resolved before or after TLS is set up).

Yes, there are multiple reasons ifunc is not just a missing feature
but rejected as an anti-feature. It has issues with the order
relocations take place in and exposes that as a contract with the
application, especially if (like musl) you do not do lazy binding,
which slightly relaxes the ordering requirement. It does not have a
well-defined or definable contract for what you can actually do from
the resovler functions, isn't compatible with static linking, etc.
Everything that can be done badly with ifunc can be done just as well,
fully portably, with function pointers initialized to a generic
version of the function replaced by ctors, or to a self-initializing
version of the function that replaces itself on first call (modulo
some technicalities about how you do the atomics, but this is
basically solvable with relaxed-mode atomics or call_once). In a
normal dynamic-linked context, this is just as fast as an ifunc call
too -- you just load the function pointer from a PC-relative address
outside the GOT rather than a PC-relative address in the GOT (and
you get to bypass the PLT thunk too).

Rich