Re: [RFC][PATCH -v2 4/4] locking/mutex: Add lock handoff to avoid starvation

[Waiman Long <waiman.long@hpe.com>]

On 08/25/2016 06:23 PM, Peter Zijlstra wrote:
> On Thu, Aug 25, 2016 at 06:00:19PM -0400, Waiman Long wrote:
>> On 08/25/2016 02:37 PM, Peter Zijlstra wrote:
>>> @@ -468,9 +496,12 @@ void __sched mutex_unlock(struct mutex *
>>>   	DEBUG_LOCKS_WARN_ON(__mutex_owner(lock) != current);
>>>   #endif
>>>
>>> -	owner = atomic_long_fetch_and(MUTEX_FLAGS,&lock->owner);
>>> +	owner = atomic_long_read(&lock->owner);
>>> +	if (!(owner&   MUTEX_FLAG_HANDOFF))
>>> +		owner = atomic_long_fetch_and(MUTEX_FLAGS,&lock->owner);
>>> +
>>>   	if (__owner_flags(owner))
>>> -		__mutex_unlock_slowpath(lock);
>>> +		__mutex_unlock_slowpath(lock, owner);
>>>   }
>>>   EXPORT_SYMBOL(mutex_unlock);
>> I don't think the race condition is fixed when we don't make sure that lock
>> handoff only happens from current=>new. The problem is due to the fact that
>> the MUTEX_FLAG_HANDOFF check in the unlock fastpath isn't serialized by the
>> wait_lock. As a result, it is possible that the owner is NULL while the
>> HANDOFF bit is set. Or an optimistic spinner may have stolen the lock in the
>> interim.
> You're right.. I'll try again tomorrow.

Actually, it is also OK to handoff from NULL=>new, but nothing else is 
appropriate.

Cheers,
Longman