Re: pthread_key_create bug?

["writeonce@midipix.org" <writeonce@midipix.org>]

On 01/08/2019 20:34, Markus Wichmann wrote:
> On Tue, Jan 08, 2019 at 09:43:10AM +0100, u-uy74@aetey.se wrote:
> > On Mon, Jan 07, 2019 at 07:00:18PM -0500, Rich Felker wrote:
> > > >  
> > > > +extern hidden weak void __pthread_key_delete_synccall(void (*f)(void *), void *p);
> > 
> > > musl on the basis (perhaps somewhat dubious) that they're an
> > > additional toolchain feature that might cause problems reusing the
> > > code in non-ELF contexts (this may affect midipix; I'm not sure).
> > 
> > Thanks. That's no doubt, the less the reliance on toolchain features,
> > the easier to use, especially in the ways/areas not known in advance.
> > 
> > Rune
> > 
> 
> Well, what happens on midipix with this patch? Worst case scenario is,
> the toolchain doesn't do weak references, and the reference becomes
> strong. So that would leave you no worse than the current situation. Or
> am I missing the point?

Weak/hidden symbols and references as used in the patch are actually supported. Generally speaking, default visibility means that a register (say, rax) is going to have the address of a function's .got entry (.got entries in PE are specific to midipix and are part of the toolchain's customization), and the call will thus take the form of callq *%rax. With hidden visibility, rax would contain the actual function address, and the call would accordingly become callq %rax. Weak references on midipix provide most of everything that ELF provides. With your patch applied, and linking libc.so with pthread_key_delete.lo (and thus also tss_delete.lo) left out, the resuling image contains the following as expected (callq 0 being the bit of interest):

6bc4d0f3:     48 8d 0d 46 c7 00 00    lea    0xc746(%rip),%rcx        # 6bc59840 <clean_dirty_tsd_callback>
6bc4d0fa:     48 89 44 24 20          mov    %rax,0x20(%rsp)
6bc4d0ff:     c7 44 24 28 0b 00 00    movl   $0xb,0x28(%rsp)
6bc4d106:     00 
6bc4d107:     e8 f4 2e 3b 94          callq  0 <__dll__>

To complete the picture, the one weak/hidden trick that's currently not supported on midipix consists in musl's src/errno/__errno_location.c and src/include/errno.h, where you end up with "call ___errno_location" (because of the hidden attribute) in many translation units, yet a strongly defined ___errno_location in none (for which the workaround is to provide a strong ___errno_location() function).


> 
> Ciao,
> Markus
> 

--