[5/7] elf: Enable relro for static build

  The code is similar to the one at rtld.c, where its check for the
PT_GNU_RELRO header values from program headers and call
_dl_protected_relro with the updated l_relro_{addr,size} values.

Checked on x86_64-linux-gnu, i686-linux-gnu, powerpc64le-linux-gnu,
aarch64-linux-gnu, s390x-linux-gnu, and sparc64-linux-gnu.
---
 elf/dl-support.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
  

* Adhemerval Zanella:

> The code is similar to the one at rtld.c, where its check for the
> PT_GNU_RELRO header values from program headers and call
> _dl_protected_relro with the updated l_relro_{addr,size} values.

This is not the actual code that does RELRO in most cases, it's only
used with prelink.  _dl_relocate_object is what is used.

> diff --git a/elf/dl-support.c b/elf/dl-support.c
> index 5526d5ee6e..bdb5c2ae91 100644
> --- a/elf/dl-support.c
> +++ b/elf/dl-support.c
> @@ -367,14 +367,24 @@ _dl_non_dynamic_init (void)
>    if (_dl_platform != NULL)
>      _dl_platformlen = strlen (_dl_platform);
>  
> -  /* Scan for a program header telling us the stack is nonexecutable.  */
>    if (_dl_phdr != NULL)
> -    for (uint_fast16_t i = 0; i < _dl_phnum; ++i)
> -      if (_dl_phdr[i].p_type == PT_GNU_STACK)
> +    for (const ElfW(Phdr) *ph = _dl_phdr; ph < &_dl_phdr[_dl_phnum]; ++ph)
> +      switch (ph->p_type)
>  	{
> -	  _dl_stack_flags = _dl_phdr[i].p_flags;
> +	/* Check if the stack is nonexecutable.  */
> +	case PT_GNU_STACK:
> +	  _dl_stack_flags = ph->p_flags;
> +	  break;
> +
> +	case PT_GNU_RELRO:
> +	  _dl_main_map.l_relro_addr = ph->p_vaddr;
> +	  _dl_main_map.l_relro_size = ph->p_memsz;
>  	  break;
>  	}
> +
> +  /* Setup relro on the binary itself.  */
> +  if (_dl_main_map.l_relro_size)
> +    _dl_protect_relro (&_dl_main_map);

Please use an explicit comparison with != 0.

I have a test case for this which I can post.  Somewhat bizarrely,
full RELRO for statically linked binaries requires linking with -z now.
  

On 01/12/2019 06:55, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>> The code is similar to the one at rtld.c, where its check for the
>> PT_GNU_RELRO header values from program headers and call
>> _dl_protected_relro with the updated l_relro_{addr,size} values.
> 
> This is not the actual code that does RELRO in most cases, it's only
> used with prelink.  _dl_relocate_object is what is used.

Ack, I changed the commit message to:

The code is similar to the one at elf/dl-reloc.c, where it checks for
the l_relro_size from the link_map (obtained from PT_GNU_RELRO header
from program headers) and calls_dl_protected_relro.

> 
>> diff --git a/elf/dl-support.c b/elf/dl-support.c
>> index 5526d5ee6e..bdb5c2ae91 100644
>> --- a/elf/dl-support.c
>> +++ b/elf/dl-support.c
>> @@ -367,14 +367,24 @@ _dl_non_dynamic_init (void)
>>    if (_dl_platform != NULL)
>>      _dl_platformlen = strlen (_dl_platform);
>>  
>> -  /* Scan for a program header telling us the stack is nonexecutable.  */
>>    if (_dl_phdr != NULL)
>> -    for (uint_fast16_t i = 0; i < _dl_phnum; ++i)
>> -      if (_dl_phdr[i].p_type == PT_GNU_STACK)
>> +    for (const ElfW(Phdr) *ph = _dl_phdr; ph < &_dl_phdr[_dl_phnum]; ++ph)
>> +      switch (ph->p_type)
>>  	{
>> -	  _dl_stack_flags = _dl_phdr[i].p_flags;
>> +	/* Check if the stack is nonexecutable.  */
>> +	case PT_GNU_STACK:
>> +	  _dl_stack_flags = ph->p_flags;
>> +	  break;
>> +
>> +	case PT_GNU_RELRO:
>> +	  _dl_main_map.l_relro_addr = ph->p_vaddr;
>> +	  _dl_main_map.l_relro_size = ph->p_memsz;
>>  	  break;
>>  	}
>> +
>> +  /* Setup relro on the binary itself.  */
>> +  if (_dl_main_map.l_relro_size)
>> +    _dl_protect_relro (&_dl_main_map);
> 
> Please use an explicit comparison with != 0.

Ack.

> 
> I have a test case for this which I can post.  

Sure, I can attach on the patch itself.

> Somewhat bizarrely,
> full RELRO for statically linked binaries requires linking with -z now.
> 

My understanding it is arch-specific and also depends on how bintuils was
build.  For instance, with my system ld (GNU ld (GNU Binutils for Ubuntu) 2.30)
seemed to be built with DEFAULT_LD_Z_RELRO (set by --enable-relro) which sets 
relro by default.  With this binutils I could only disable relro by explicit 
add norelro, the -z {lazy,now} did not change the GNU_RELRO header creation.

Also the code in ld/emultempl/elf.em does seem to select different linker
scripts for both link_info.relro and (link_info.flags & DF_BIND_NOW), however
it does no have a special case for 
link_info.relro and !(link_info.flags & DF_BIND_NOW). I don't see how -relro
is requiring -z now on ld code.
  

* Adhemerval Zanella:

>> Somewhat bizarrely, full RELRO for statically linked binaries
>> requires linking with -z now.

> My understanding it is arch-specific and also depends on how
> bintuils was build.  For instance, with my system ld (GNU ld (GNU
> Binutils for Ubuntu) 2.30) seemed to be built with
> DEFAULT_LD_Z_RELRO (set by --enable-relro) which sets relro by
> default.  With this binutils I could only disable relro by explicit
> add norelro, the -z {lazy,now} did not change the GNU_RELRO header
> creation.

Whether -z relro gives you full RELRO depends somewhat on the
architecture and what relocations can be eliminated from the static
link.  Objects built with -fPIC tend to leave relocations behind,
though, and to protect them, you need -z now.
  

On 02/12/2019 15:25, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>>> Somewhat bizarrely, full RELRO for statically linked binaries
>>> requires linking with -z now.
> 
>> My understanding it is arch-specific and also depends on how
>> bintuils was build.  For instance, with my system ld (GNU ld (GNU
>> Binutils for Ubuntu) 2.30) seemed to be built with
>> DEFAULT_LD_Z_RELRO (set by --enable-relro) which sets relro by
>> default.  With this binutils I could only disable relro by explicit
>> add norelro, the -z {lazy,now} did not change the GNU_RELRO header
>> creation.
> 
> Whether -z relro gives you full RELRO depends somewhat on the
> architecture and what relocations can be eliminated from the static
> link.  Objects built with -fPIC tend to leave relocations behind,
> though, and to protect them, you need -z now.
> 

I was investigating in fact if binutils is requiring the -z now
to actually enable full RELRO, but it seems that there is no
option to actually set 'full RELRO' besides the combination
of -z now and -z relro.

And I think it is worth to check for static PIE as well. At least
for partial relro, .data.rel.ro seems to protect the required
data.

About testing, I am not sure what kind of coverage we are aiming
here. My initial approach would to check if a write on a variable
set to .data.rel.so does trigger a SEGSEGV signal and check with
some different combinations (-z now, -z lazy, static, dynamic,
pie, nopie). Do you have something more elaborated in mind?
  

* Adhemerval Zanella:

> About testing, I am not sure what kind of coverage we are aiming
> here. My initial approach would to check if a write on a variable
> set to .data.rel.so does trigger a SEGSEGV signal and check with
> some different combinations (-z now, -z lazy, static, dynamic,
> pie, nopie). Do you have something more elaborated in mind?

Yes, that's what my tests do.  I will post them once we've stabilized
the dynamic linker again.
  

On 02/12/2019 16:02, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>> About testing, I am not sure what kind of coverage we are aiming
>> here. My initial approach would to check if a write on a variable
>> set to .data.rel.so does trigger a SEGSEGV signal and check with
>> some different combinations (-z now, -z lazy, static, dynamic,
>> pie, nopie). Do you have something more elaborated in mind?
> 
> Yes, that's what my tests do.  I will post them once we've stabilized
> the dynamic linker again.
> 

I will update the patch with the tests I have done so far.