[PATCH v3 5/5] csu: Move static pie self relocation later [BZ #27072]

[Szabolcs Nagy]

The 01/14/2021 17:19, Szabolcs Nagy via Libc-alpha wrote:
> The 01/14/2021 08:26, H.J. Lu wrote:
> > 
> > We need to make sure that there are no RELATIVE relocations before
> > _dl_relocate_static_pie is called.  The problems with i386 are
> > 
> > 1.  All calls to IFUNC functions must go through PLT.
> > 2.  Calls to hidden functions CANNOT go through PLT in PIE since
> > EBX used in PIE PLT may not be set up for local calls.
> > 
> > I think we should add a new attribute, attribute_hidden_ifunc
> > which should be defined as
> > 
> > 1. __attribute__ ((visibility ("default"))) if in PIE on i386
> > 2. __attribute__ ((visibility ("hidden"))) else
> > 
> > attribute_hidden_ifunc should be used on prototypes of all IFUNC
> > functions.  This is similar to NO_HIDDEN_EXTERN_FUNC_IN_PIE.
> 
> so is it enough to declare ifuncs with such attribute?
> 
> e.g. would it work if memcpy is default visibility
> in PIE libc.a but user code is static linking that
> with non-pie caller?

hm no, i think the only inconsistency that can happen
is if an ifunc function is marked hidden in non-pie
libc.a, but user calls it with default visibility,
but that is not a problem i guess.

maybe this works:

target gives a list of ifunc declarations with explicit
visibility attribute (default vis for i686 PIE) in a
header that is pre-included very early everywhere so later
declarations keep the explicit visibility instead of the
one specified by the gcc pragma. (but i don't yet see how
to have the right prototypes in an early declaration)

> 
> do we have a way to track which functions may be
> defined as ifunc? should we do that manually? or add
> the attribute to every extern function declaration
> within the libc?