Virtualization - Apr 2014 - [PATCH v9 04/19] qspinlock: Extract out the exchange of tail code word

On Thu, Apr 17, 2014 at 05:28:17PM -0400, Waiman Long
wrote:
> On 04/17/2014 11:49 AM, Peter Zijlstra wrote:
> >On Thu, Apr 17, 2014 at 11:03:56AM -0400, Waiman Long wrote:
> >>@@ -192,36 +220,25 @@ void queue_spin_lock_slowpath(struct
qspinlock *lock, u32 val)
> >>  	node->next = NULL;
> >>
> >>  	/*
> >>+	 * We touched a (possibly) cold cacheline; attempt the trylock
once
> >>+	 * more in the hope someone let go while we weren't watching
as long
> >>+	 * as no one was queuing.
> >>  	 */
> >>+	if (!(val&  _Q_TAIL_MASK)&& 
queue_spin_trylock(lock))
> >>+		goto release;
> >But you just did a potentially very expensive op; @val isn't
> >representative anymore!
> 
> That is not true. I pass in a pointer to val to trylock_pending() (the
> pointer thing) so that it will store the latest value that it reads from
the
> lock back into val. I did miss one in the PV qspinlock exit loop. I will
add
> it back when I do the next version.
But you did that read _before_ you touched a cold cacheline, that's 100s
of cycles. Whatever value you read back then is now complete nonsense.

On 04/18/2014 04:15 AM, Peter Zijlstra wrote:
> On Thu, Apr 17, 2014 at 05:28:17PM -0400, Waiman Long wrote:
>> On 04/17/2014 11:49 AM, Peter Zijlstra wrote:
>>> On Thu, Apr 17, 2014 at 11:03:56AM -0400, Waiman Long wrote:
>>>> @@ -192,36 +220,25 @@ void queue_spin_lock_slowpath(struct
qspinlock *lock, u32 val)
>>>>   	node->next = NULL;
>>>>
>>>>   	/*
>>>> +	 * We touched a (possibly) cold cacheline; attempt the
trylock once
>>>> +	 * more in the hope someone let go while we weren't
watching as long
>>>> +	 * as no one was queuing.
>>>>   	 */
>>>> +	if (!(val&   _Q_TAIL_MASK)&&  
queue_spin_trylock(lock))
>>>> +		goto release;
>>> But you just did a potentially very expensive op; @val isn't
>>> representative anymore!
>> That is not true. I pass in a pointer to val to trylock_pending() (the
>> pointer thing) so that it will store the latest value that it reads
from the
>> lock back into val. I did miss one in the PV qspinlock exit loop. I
will add
>> it back when I do the next version.
> But you did that read _before_ you touched a cold cacheline, that's
100s
> of cycles. Whatever value you read back then is now complete nonsense.
For spin_lock(), the lock cacheline is touched by a cmpxchg(). It can 
takes 100s of cycles whether it is hot or cold.

I will take the precheck out, it is not such a big deal anyway.

-Longman

On Fri, Apr 18, 2014 at 01:32:47PM -0400, Waiman Long
wrote:
> On 04/18/2014 04:15 AM, Peter Zijlstra wrote:
> >On Thu, Apr 17, 2014 at 05:28:17PM -0400, Waiman Long wrote:
> >>On 04/17/2014 11:49 AM, Peter Zijlstra wrote:
> >>>On Thu, Apr 17, 2014 at 11:03:56AM -0400, Waiman Long wrote:
> >>>>@@ -192,36 +220,25 @@ void queue_spin_lock_slowpath(struct
qspinlock *lock, u32 val)
> >>>>  	node->next = NULL;
> >>>>
> >>>>  	/*
> >>>>+	 * We touched a (possibly) cold cacheline; attempt the
trylock once
> >>>>+	 * more in the hope someone let go while we weren't
watching as long
> >>>>+	 * as no one was queuing.
> >>>>  	 */
> >>>>+	if (!(val&   _Q_TAIL_MASK)&&  
queue_spin_trylock(lock))
> >>>>+		goto release;
> >>>But you just did a potentially very expensive op; @val
isn't
> >>>representative anymore!
> >>That is not true. I pass in a pointer to val to trylock_pending()
(the
> >>pointer thing) so that it will store the latest value that it reads
from the
> >>lock back into val. I did miss one in the PV qspinlock exit loop. I
will add
> >>it back when I do the next version.
> >But you did that read _before_ you touched a cold cacheline, that's
100s
> >of cycles. Whatever value you read back then is now complete nonsense.
> 
> For spin_lock(), the lock cacheline is touched by a cmpxchg(). It can takes
> 100s of cycles whether it is hot or cold.
Its not the lock cacheline, you just touched the per-cpu node cacheline
for the first time, setting up the node.