[PATCHv3] PowerPC: Fix a race condition when eliding a lock

[Carlos O'Donell]

On 08/24/2015 03:28 PM, Adhemerval Zanella wrote:
> 
> 
> On 24-08-2015 15:58, Carlos O'Donell wrote:
>> On 08/24/2015 02:11 PM, Tulio Magno Quites Machado Filho wrote:
>>> Changes since v2:
>>>  - Moved part of the source code comments to the commit message.
>>>  - Added O'Donnel's suggestion to the source code comments.
>>>
>>> Changes since v1:
>>>  - Improved commit message.
>>>  - Added new comments in the source code alerting to the concurrency details
>>>    intrinsic to this code.
>>>  - Removed compiler barriers from this patch, which will be treated in another
>>>    patch and will be synchronized with the GCC implementation.
>>
>>> +	  if (__builtin_tbegin (0))					\
>>> +	    {								\
>>> +	      if (is_lock_free)						\
>>> +		{							\
>>> +		  ret = 1;						\
>>> +		  break;						\
>>> +		}							\
>>> +	      __builtin_tabort (_ABORT_LOCK_BUSY);			\
>>> +	    }								\
>>> +	  else								\
>>> +	    if (!__get_new_count(&adapt_count))				\
>>> +	      break;							\
>>
>> Toravld still suggests an atomic_load_relaxed here:
>> https://www.sourceware.org/ml/libc-alpha/2015-08/msg00893.html
>>
>> Is there any technical objection to that request?
>>
>> It does highlight, as he notes, the transactional and non-transactional
>> accesses for that evaluated value.
> 
> No 'technical' objections, only that the transaction idea is exactly to
> avoid the use of atomic accesses altogether.  For relaxed accesses it
> won't change anything, but if the idea is to highlight every access
> within transaction with atomics, it might be case where the algorithm
> will require another semantic (like a acquire or release) and it will
> require to drop the atomic usage to avoid generate subpar code.

I'll respond to Torvald and ask him to explain in more detail.

Cheers,
Carlos.