Re: [PATCH v2 1/7] locking/rwsem: check for active writer/spinner before wakeup

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

On 08/08/2014 01:39 AM, Davidlohr Bueso wrote:
> On Thu, 2014-08-07 at 17:45 -0700, Davidlohr Bueso wrote:
>> On Thu, 2014-08-07 at 18:26 -0400, Waiman Long wrote:
>>> On a highly contended rwsem, spinlock contention due to the slow
>>> rwsem_wake() call can be a significant portion of the total CPU cycles
>>> used. With writer lock stealing and writer optimistic spinning, there
>>> is also a pretty good chance that the lock may have been stolen
>>> before the waker wakes up the waiters. The woken tasks, if any,
>>> will have to go back to sleep again.
>> Good catch! And this applies to mutexes as well. How about something
>> like this:
>>
>> diff --git a/kernel/locking/mutex.c b/kernel/locking/mutex.c
>> index dadbf88..e037588 100644
>> --- a/kernel/locking/mutex.c
>> +++ b/kernel/locking/mutex.c
>> @@ -707,6 +707,20 @@ EXPORT_SYMBOL_GPL(__ww_mutex_lock_interruptible);
>>
>>   #endif
>>
>> +#if defined(CONFIG_DEBUG_MUTEXES) || defined(CONFIG_MUTEX_SPIN_ON_OWNER)
> If DEBUG, we don't clear the owner when unlocking. This can just be
>
> +#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
>
>> +static inline bool mutex_has_owner(struct mutex *lock)
>> +{
>> +	struct task_struct *owner = ACCESS_ONCE(lock->owner);
>> +
>> +	return owner != NULL;
>> +}
>> +#else
>> +static inline bool mutex_has_owner(struct mutex *lock)
>> +{
>> +	return false;
>> +}
>> +#endif
>> +
>>   /*
>>    * Release the lock, slowpath:
>>    */
>> @@ -734,6 +748,15 @@ __mutex_unlock_common_slowpath(struct mutex *lock, int nested)
>>   	mutex_release(&lock->dep_map, nested, _RET_IP_);
>>   	debug_mutex_unlock(lock);
>>
>> +	/*
>> +	 * Abort the wakeup operation if there is an active writer as the
>> +	 * lock was stolen. mutex_unlock() should have cleared the owner field
>> +	 * before calling this function. If that field is now set, there must
>> +	 * be an active writer present.
>> +	 */
>> +	if (mutex_has_owner(lock))
>> +		goto done;
> Err so we actually deadlock here because we do the check with the
> lock->wait_lock held and at the same time another task comes into the
> slowpath of a mutex_lock() call which also tries to take the wait_lock.
> Ending up with hung tasks. Here's a more tested patch against
> peterz-queue, survives aim7 and kernel builds on a 80core box. Thanks.

I couldn't figure out why there will be hang tasks. The logic looks OK 
to me.

>
> 8<---------------------------------------------------------------
> From: Davidlohr Bueso<davidlohr@hp.com>
> Subject: [PATCH] locking/mutex: Do not falsely wake-up tasks
>
> Mutexes lock-stealing functionality allows another task to
> skip its turn in the wait-queue and atomically acquire the lock.
> This is fine and a nice optimization, however, when releasing
> the mutex, we always wakeup the next task in FIFO order. When
> the lock has been stolen this leads to wasting waking up a
> task just to immediately realize it cannot acquire the lock
> and just go back to sleep. This is specially true on highly
> contended mutexes that stress the wait_lock.
>
> Signed-off-by: Davidlohr Bueso<davidlohr@hp.com>
> ---
>   kernel/locking/mutex.c | 32 +++++++++++++++++++++++++++++++-
>   1 file changed, 31 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/locking/mutex.c b/kernel/locking/mutex.c
> index dadbf88..52e1136 100644
> --- a/kernel/locking/mutex.c
> +++ b/kernel/locking/mutex.c
> @@ -383,12 +383,26 @@ done:
>
>   	return false;
>   }
> +
> +static inline bool mutex_has_owner(struct mutex *lock)
> +{
> +	struct task_struct *owner = ACCESS_ONCE(lock->owner);
> +
> +	return owner != NULL;
> +}
> +
>   #else
> +
>   static bool mutex_optimistic_spin(struct mutex *lock,
>   				  struct ww_acquire_ctx *ww_ctx, const bool use_ww_ctx)
>   {
>   	return false;
>   }
> +
> +static inline bool mutex_has_owner(struct mutex *lock)
> +{
> +	return false;
> +}
>   #endif
>
>   __visible __used noinline
> @@ -730,6 +744,23 @@ __mutex_unlock_common_slowpath(struct mutex *lock, int nested)
>   	if (__mutex_slowpath_needs_to_unlock())
>   		atomic_set(&lock->count, 1);
>
> +/*
> + * Skipping the mutex_has_owner() check when DEBUG, allows us to
> + * avoid taking the wait_lock in order to do not call mutex_release()
> + * and debug_mutex_unlock() when !DEBUG. This can otherwise result in
> + * deadlocks when another task enters the lock's slowpath in mutex_lock().
> + */
> +#ifndef CONFIG_DEBUG_MUTEXES
> +	/*
> +	 * Abort the wakeup operation if there is an another mutex owner, as the
> +	 * lock was stolen. mutex_unlock() should have cleared the owner field
> +	 * before calling this function. If that field is now set, another task
> +	 * must have acquired the mutex.
> +	 */
> +	if (mutex_has_owner(lock))
> +		return;
> +#endif
> +
>   	spin_lock_mutex(&lock->wait_lock, flags);
>   	mutex_release(&lock->dep_map, nested, _RET_IP_);
>   	debug_mutex_unlock(lock);
> @@ -744,7 +775,6 @@ __mutex_unlock_common_slowpath(struct mutex *lock, int nested)
>
>   		wake_up_process(waiter->task);
>   	}
> -
>   	spin_unlock_mutex(&lock->wait_lock, flags);
>   }
>

I have 2 issues about this. First of all, the timing windows between 
atomic_set() and mutex_has_owner() check is really small, I doubt it 
will be that effective. Secondly, I think you may need to call 
mutex_release() and debug_mutex_unlock() to make the debugging code 
work, but they seems to be called only under the wait_lock. So I think 
there is more work that need to be done before this patch is ready.

-Longman