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

  On 15 Dec 2016, Florian Weimer told this:

> On 12/15/2016 03:15 PM, Nix wrote:
>
>> Possible fix, untested:
>>
>> diff --git a/sysdeps/generic/symbol-hacks.h b/sysdeps/generic/symbol-hacks.h
>> index 36908b5..0679354 100644
>> --- a/sysdeps/generic/symbol-hacks.h
>> +++ b/sysdeps/generic/symbol-hacks.h
>> @@ -7,5 +7,7 @@ asm ("memcpy = __GI_memcpy");
>>
>>  /* -fstack-protector generates calls to __stack_chk_fail, which need
>>     similar adjustments to avoid going through the PLT.  */
>> +#if defined __SSP__ || defined __SSP_ALL__ || defined __SSP_STRONG__
>>  asm ("__stack_chk_fail = __stack_chk_fail_local");
>>  #endif
>> +#endif
>
> The condition looks rather brittle.  What if GCC grows an -fstack-protector-light switch and __SSP_LIGHT__ macro?

We'd need to change configure.ac before that would have an effect in any
case... but it does seem likely that changing this too would be
overlooked.

> I wonder if it's better to add something to $(no-stack-protector) and use that in the conditional.

That was my other option, but the total absence of anything in
configure.ac passing -D made me think twice.

Something like this? (even more untested than the last one, if
possible -- but adds a new possibility: we can now differentiate between
"glibc built without stack protector" and "glibc built with stack
protector but this file doesn't have it" without relying on GCC
predefined macros. The __WITH_ naming scheme is completely arbitrary
and I can change it to anything you prefer.)
  

On 12/15/2016 03:29 PM, Nix wrote:
> On 15 Dec 2016, Florian Weimer told this:
>
>> On 12/15/2016 03:15 PM, Nix wrote:
>>
>>> Possible fix, untested:
>>>
>>> diff --git a/sysdeps/generic/symbol-hacks.h b/sysdeps/generic/symbol-hacks.h
>>> index 36908b5..0679354 100644
>>> --- a/sysdeps/generic/symbol-hacks.h
>>> +++ b/sysdeps/generic/symbol-hacks.h
>>> @@ -7,5 +7,7 @@ asm ("memcpy = __GI_memcpy");
>>>
>>>  /* -fstack-protector generates calls to __stack_chk_fail, which need
>>>     similar adjustments to avoid going through the PLT.  */
>>> +#if defined __SSP__ || defined __SSP_ALL__ || defined __SSP_STRONG__
>>>  asm ("__stack_chk_fail = __stack_chk_fail_local");
>>>  #endif
>>> +#endif
>>
>> The condition looks rather brittle.  What if GCC grows an -fstack-protector-light switch and __SSP_LIGHT__ macro?
>
> We'd need to change configure.ac before that would have an effect in any
> case... but it does seem likely that changing this too would be
> overlooked.

Right.

>> I wonder if it's better to add something to $(no-stack-protector) and use that in the conditional.
>
> That was my other option, but the total absence of anything in
> configure.ac passing -D made me think twice.
>
> Something like this? (even more untested than the last one, if
> possible -- but adds a new possibility: we can now differentiate between
> "glibc built without stack protector" and "glibc built with stack
> protector but this file doesn't have it" without relying on GCC
> predefined macros. The __WITH_ naming scheme is completely arbitrary
> and I can change it to anything you prefer.)

WITH_STACK_PROTECTOR (without the leading underscores) looks okay to me 
because it's only used at build time.  Or you could call it 
STACK_PROTECTOR_LEVEL, to match the other variable.

> diff --git a/configure.ac b/configure.ac
> index 2396c1f..8bb8c2c 100644
> --- a/configure.ac
> +++ b/configure.ac
> @@ -638,18 +638,18 @@ LIBC_TRY_CC_OPTION([$CFLAGS $CPPFLAGS -Werror -fstack-protector-all],
>  stack_protector=
>  no_stack_protector=
>  if test "$libc_cv_ssp" = yes; then
> -  no_stack_protector="-fno-stack-protector"
> +  no_stack_protector="-fno-stack-protector -D__WITH_STACK_PROTECTOR=0"
>    AC_DEFINE(HAVE_CC_NO_STACK_PROTECTOR)
>  fi
>
>  if test "$enable_stack_protector" = yes && test "$libc_cv_ssp" = yes; then
> -  stack_protector="-fstack-protector"
> +  stack_protector="-fstack-protector -D__WITH_STACK_PROTECTOR=1"
>    AC_DEFINE(STACK_PROTECTOR_LEVEL, 1)
>  elif test "$enable_stack_protector" = all && test "$libc_cv_ssp_all" = yes; then
> -  stack_protector="-fstack-protector-all"
> +  stack_protector="-fstack-protector-all -D__WITH_STACK_PROTECTOR=2"
>    AC_DEFINE(STACK_PROTECTOR_LEVEL, 2)
>  elif test "$enable_stack_protector" = strong && test "$libc_cv_ssp_strong" = yes; then
> -  stack_protector="-fstack-protector-strong"
> +  stack_protector="-fstack-protector-strong -D__WITH_STACK_PROTECTOR=3"
>    AC_DEFINE(STACK_PROTECTOR_LEVEL, 3)
>  fi
>  AC_SUBST(libc_cv_ssp)
> diff --git a/sysdeps/generic/symbol-hacks.h b/sysdeps/generic/symbol-hacks.h
> index 36908b5..12b4fe7 100644
> --- a/sysdeps/generic/symbol-hacks.h
> +++ b/sysdeps/generic/symbol-hacks.h
> @@ -7,5 +7,7 @@ asm ("memcpy = __GI_memcpy");
>
>  /* -fstack-protector generates calls to __stack_chk_fail, which need
>     similar adjustments to avoid going through the PLT.  */
> +#if defined __WITH_STACK_PROTECTOR && __WITH_STACK_PROTECTOR > 0
>  asm ("__stack_chk_fail = __stack_chk_fail_local");
>  #endif
> +#endif

The new #if/#endif need to be indented.

Thanks,
Florian