Re: [PATCH] elf: Always set l in _dl_init_paths (Bug 23462).

[Carlos O'Donell <carlos at redhat dot com>]

On 10/16/19 5:54 AM, Florian Weimer wrote:
> * Carlos O'Donell:
> 
>> diff --git a/elf/dl-load.c b/elf/dl-load.c
>> index 438793e53d..903f8af13a 100644
>> --- a/elf/dl-load.c
>> +++ b/elf/dl-load.c
>> @@ -748,50 +748,47 @@ _dl_init_paths (const char *llp)
>>    max_dirnamelen = SYSTEM_DIRS_MAX_LEN;
>>    *aelem = NULL;
>>  
>> -#ifdef SHARED
>> -  /* This points to the map of the main object.  */
>> +  /* This points to the map of the main object.  It is always non-NULL
>> +     since we have purposely made the dynamic and static cases look the
>> +     same.  */
>>    l = GL(dl_ns)[LM_ID_BASE]._ns_loaded;
>> -  if (l != NULL)
>> +  assert (l->l_type != lt_loaded);
>> +
>> +  if (l->l_info[DT_RUNPATH])
>> +    {
>> +      /* Allocate room for the search path and fill in information
>> +	 from RUNPATH.  */
>> +      decompose_rpath (&l->l_runpath_dirs,
>> +		       (const void *) (D_PTR (l, l_info[DT_STRTAB])
>> +				       + l->l_info[DT_RUNPATH]->d_un.d_val),
>> +		       l, "RUNPATH");
>> +      /* During rtld init the memory is allocated by the stub malloc,
>> +	 prevent any attempt to free it by the normal malloc.  */
>> +      l->l_runpath_dirs.malloced = 0;
>> +
>> +      /* The RPATH is ignored.  */
>> +      l->l_rpath_dirs.dirs = (void *) -1;
>> +    }
> 
> Does this change enable DT_RUNPATH/DT_RPATH for statically linked
> binaries?  Should that receive a NEWS entry and a test?

It does enable DT_RUNPATH/DT_RPATH for statically linked binaries,
and we should probably add a test for that?

This is not a consequence I actively thought too deeply about because
it fell under the "make static binaries behave like dynamic binaries"
which was within our design goals. However, now that you mention it,
it will be a semantic change if anyone has ever built a static
binary with DT_RUNPATH/DT_RPATH, the behaviour will now be changed.

Let me add a NEWS entry for this.

If you think the semantic change here is wrong, please call that out.

-- 
Cheers,
Carlos.