[PATCH] Protect _dl_profile_fixup data-dependency order [BZ #23690]

[John David Anglin]

On 2018-09-19 10:00 PM, Carlos O'Donell wrote:
> On 09/19/2018 09:59 PM, John David Anglin wrote:
>> On 2018-09-19 8:16 PM, Carlos O'Donell wrote:
>>>>  Â Â  DL_FIXUP_VALUE_TYPE *resultp = &reloc_result->addr;
>>>> +Â  DL_FIXUP_VALUE_TYPE value = atomic_load_acquire(resultp);
>>> You are potentially requiring an atomic load of a structure whose size can be
>>> an arbitrary size depending on machine and ABI design (64-bit fore 32-bit hppa,
>>> and 128-bit for 64-bit ia64). These architectures might not have such wide
>>> atomic operations, and ...
>> We have implemented 64-bit atomic loads and stores for 32-bit hppa. They are not well tested but
>> they might work.  They use floating point loads and stores, and kernel helper.  The code is pretty horrific :-(
> We only need to use the fdesc->ip as the guard, so we don't really need the 64-bit
> atomic, but other algorithms like the new pthread condvars can use them effectively
> to accelerate and avoid 2 lws kernel helper calls and instead use 1 lws kernel helper
> 64-bit atomic.
Regarding using fdesc->ip as the guard, The gp is loaded both before and 
after the ip on hppa.
For example, $$dyncall loads gp before the branch.  This could be 
changed at the cost of one
instruction.  Stubs load gp after ip.  I don't think this is easy to change.

Dave

-- 
John David Anglin  dave.anglin@bell.net