Re: [PATCH 08/12] De-PLTize __stack_chk_fail internal calls within libc.so.

[Florian Weimer <fweimer at redhat dot com>]

On 12/15/2016 09:00 PM, Nix wrote:

> > Could you try this?
> >
> > # if defined STACK_PROTECTOR_LEVEL && STACK_PROTECTOR_LEVEL > 0
> > asm (".hidden __stack_chk_fail_local");
> > asm ("__stack_chk_fail = __stack_chk_fail_local");
> > # endif
>
> No change :( the only reference to __stack_chk_fail is still inside
> stack_chk_fail_local:
>
> Symbols from libc_pic.a[libc-stack_chk_fail_local.os]:
>
> Name                        Value            Class  Type     Size             Line Section
>
> __GI_memcpy                ||GLOBAL|NOTYPE  ||    |UNDEF
> __GI_memmove               ||GLOBAL|NOTYPE  ||    |UNDEF
> __GI_memset                ||GLOBAL|NOTYPE  ||    |UNDEF
> __stack_chk_fail           ||GLOBAL|NOTYPE  ||    |UNDEF
> __stack_chk_fail_local     |0000000000000000|GLOBAL|FUNC    |0000000000000010|    |.text
> libc-stack_chk_fail_local.c|0000000000000000|LOCAL |FILE    |0000000000000000|    |ABS
>
> (And, of course, this code is not affected by your suggestion, because
> it's compiled with -fno-stack-protector -DSTACK_PROTECTOR_LEVEL=0.)

I think this attempt at PLT avoidance within libc.so itself is subtly 
wrong.  We need to mirror more closely what 
libc_hidden_proto/libc_hidden_def does, and perhaps disentangle this 
from the __stack_chk_fail_local definition used in other DSOs.

I think this means removing any definition of a C function definition 
called __stack_chk_fail_local from libc.so, and instead use a strong 
alias from __stack_chk_fail to __stack_chk_fail_local to define the 
symbol.  The alias will not incorporate a PLT reference.  If you look at 
include/libc-symbols.h, strong_alias and hidden_def are quite similar.

It's too late for me to try this today. :-/

Florian