Re: [PATCH] hurd: Fix early rtld access to errno

[Samuel Thibault <samuel.thibault@ens-lyon.org>]

Samuel Thibault, on dim. 18 mars 2018 02:53:32 +0100, wrote:
> 	* sysdeps/generic/dl-sysdep.h (RTLD_EXTERN_ERRNO): Define macro to 0.
> 	* sysdeps/mach/hurd/i386/dl-sysdep.h: Include <dl-sysdep.h>.
> 	[IS_IN(rtld)] (RTLD_EXTERN_ERRNO): Define macro to 1.
> 	[!IS_IN(rtld)] (RTLD_EXTERN_ERRNO): Define macro to 0.
> 	* include/errno.h [IS_IN(rtld) && !defined RTLD_EXTERN_ERRNO]: Error
> 	out.
> 	[!IS_IN(rtld)] (RTLD_EXTERN_ERRNO): Define to 0.
> 	[RTLD_EXTERN_ERRNO]: Do not use TLS access to errno.

All that being said, I'm surprised: AIUI this problem would be the
same on all platforms as soon as one would set RTLD_PRIVATE_ERRNO to
0.

Trying to set RTLD_PRIVATE_ERRNO to 0 on Linux x86_64 for instance,
ld.so crashes in

assert (info[DT_FLAGS] == NULL
	  || (info[DT_FLAGS]->d_un.d_val & ~DF_BIND_NOW) == 0);

since __thread access produces DF_STATIC_TLS. Ignoring that check just
pushes the crash to open_verify()'s __set_errno call, on fs:(%eax) or
fs:0 access (depending whether initial-exec TLS model is enabled or
not).  I guess that at that point Linux x86_64 hasn't initialized %fs
yet.  I tried to just disable the use of __thread access to errno in
rtld with

> @@ -20,7 +24,7 @@
>  #  define errno rtld_errno
>  extern int rtld_errno attribute_hidden;
>  
> -# elif IS_IN_LIB
> +# elif IS_IN_LIB && !IS_IN(rtld)
>  
>  #  include <tls.h>
>  

But that's only pushing the issue to the __errno_location()
function. In sysdeps/mach/hurd/errno-loc.c I made the rtld version of
__errno_location() a weak function which avoids accessing TLS by using
a static variable, that function being overriden by libc as soon as it
gets loaded.  I'm actually wondering why not all archs would need this
when RTLD_PRIVATE_ERRNO is 0.

Samuel