Re: [PATCH v4 07/16] locking/rwsem: Implement lock handoff to prevent lock starvation

[Waiman Long <longman@redhat.com>]

On 04/17/2019 04:05 AM, Peter Zijlstra wrote:
> On Tue, Apr 16, 2019 at 02:16:11PM -0400, Waiman Long wrote:
>
>>>> @@ -608,56 +687,63 @@ __rwsem_down_write_failed_common(struct rw_semaphore *sem, int state)
>>>>  	 */
>>>>  	waiter.task = current;
>>>>  	waiter.type = RWSEM_WAITING_FOR_WRITE;
>>>> +	waiter.timeout = jiffies + RWSEM_WAIT_TIMEOUT;
>>>>  
>>>>  	raw_spin_lock_irq(&sem->wait_lock);
>>>>  
>>>>  	/* account for this before adding a new element to the list */
>>>> +	wstate = list_empty(&sem->wait_list) ? WRITER_FIRST : WRITER_NOT_FIRST;
>>>>  
>>>>  	list_add_tail(&waiter.list, &sem->wait_list);
>>>>  
>>>>  	/* we're now waiting on the lock */
>>>> +	if (wstate == WRITER_NOT_FIRST) {
>>>>  		count = atomic_long_read(&sem->count);
>>>>  
>>>>  		/*
>>>> +		 * If there were already threads queued before us and:
>>>> +		 *  1) there are no no active locks, wake the front
>>>> +		 *     queued process(es) as the handoff bit might be set.
>>>> +		 *  2) there are no active writers and some readers, the lock
>>>> +		 *     must be read owned; so we try to wake any read lock
>>>> +		 *     waiters that were queued ahead of us.
>>>>  		 */
>>>> +		if (!RWSEM_COUNT_LOCKED(count))
>>>> +			__rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
>>>> +		else if (!(count & RWSEM_WRITER_MASK) &&
>>>> +			  (count & RWSEM_READER_MASK))
>>>>  			__rwsem_mark_wake(sem, RWSEM_WAKE_READERS, &wake_q);
>>> Does the above want to be something like:
>>>
>>> 		if (!(count & RWSEM_WRITER_LOCKED)) {
>>> 			__rwsem_mark_wake(sem, (count & RWSEM_READER_MASK) ?
>>> 					       RWSEM_WAKE_READERS :
>>> 					       RWSEM_WAKE_ANY, &wake_q);
>>> 		}
>> Yes.
>>
>>>> +		else
>>>> +			goto wait;
>>>>  
>>>> +		/*
>>>> +		 * The wakeup is normally called _after_ the wait_lock
>>>> +		 * is released, but given that we are proactively waking
>>>> +		 * readers we can deal with the wake_q overhead as it is
>>>> +		 * similar to releasing and taking the wait_lock again
>>>> +		 * for attempting rwsem_try_write_lock().
>>>> +		 */
>>>> +		wake_up_q(&wake_q);
>>> Hurmph.. the reason we do wake_up_q() outside of wait_lock is such that
>>> those tasks don't bounce on wait_lock. Also, it removes a great deal of
>>> hold-time from wait_lock.
>>>
>>> So I'm not sure I buy your argument here.
>>>
>> Actually, we don't want to release the wait_lock, do wake_up_q() and
>> acquire the wait_lock again as the state would have been changed. I
>> didn't change the comment on this patch, but will reword it to discuss that.
> I don't understand, we've queued ourselves, we're on the list, we're not
> first. How would dropping the lock to try and kick waiters before us be
> a problem?
>
> Sure, once we re-acquire the lock we have to re-avaluate @wstate to see
> if we're first now or not, but we need to do that anyway.
>
> So what is wrong with the below?
>
> --- a/include/linux/sched/wake_q.h
> +++ b/include/linux/sched/wake_q.h
> @@ -51,6 +51,11 @@ static inline void wake_q_init(struct wa
>  	head->lastp = &head->first;
>  }
>  
> +static inline bool wake_q_empty(struct wake_q_head *head)
> +{
> +	return head->first == WAKE_Q_TAIL;
> +}
> +
>  extern void wake_q_add(struct wake_q_head *head, struct task_struct *task);
>  extern void wake_q_add_safe(struct wake_q_head *head, struct task_struct *task);
>  extern void wake_up_q(struct wake_q_head *head);
> --- a/kernel/locking/rwsem.c
> +++ b/kernel/locking/rwsem.c
> @@ -700,25 +700,22 @@ __rwsem_down_write_failed_common(struct
>  		 *     must be read owned; so we try to wake any read lock
>  		 *     waiters that were queued ahead of us.
>  		 */
> -		if (!(count & RWSEM_LOCKED_MASK))
> -			__rwsem_mark_wake(sem, RWSEM_WAKE_ANY, &wake_q);
> -		else if (!(count & RWSEM_WRITER_MASK) &&
> -				(count & RWSEM_READER_MASK))
> -			__rwsem_mark_wake(sem, RWSEM_WAKE_READERS, &wake_q);
> -		else
> +		if (count & RWSEM_WRITER_LOCKED)
>  			goto wait;
> -		/*
> -		 * The wakeup is normally called _after_ the wait_lock
> -		 * is released, but given that we are proactively waking
> -		 * readers we can deal with the wake_q overhead as it is
> -		 * similar to releasing and taking the wait_lock again
> -		 * for attempting rwsem_try_write_lock().
> -		 */
> -		wake_up_q(&wake_q);
> -		/*
> -		 * Reinitialize wake_q after use.
> -		 */
> -		wake_q_init(&wake_q);
> +
> +		__rwsem_mark_wake(sem, (count & RWSEM_READER_MASK) ?
> +				RWSEM_WAKE_READERS :
> +				RWSEM_WAKE_ANY, &wake_q);
> +
> +		if (!wake_q_empty(&wake_q)) {
> +			raw_spin_unlock_irq(&sem->wait_lock);
> +			wake_up_q(&wake_q);
> +			/* used again, reinit */
> +			wake_q_init(&wake_q);
> +			raw_spin_lock_irq(&sem->wait_lock);
> +			if (rwsem_waiter_is_first(sem, &waiter))
> +				wstate = WRITER_FIRST;
> +		}
>  	} else {
>  		count = atomic_long_add_return(RWSEM_FLAG_WAITERS, &sem->count);
>  	}

Yes, we can certainly do that. My point is that I haven't changed the
existing logic regarding that wakeup, I only move it around in the
patch. As it is not related to lock handoff, we can do it as a separate
patch.

Cheers,
Longman