RE: [PATCH 1/2] NET: Multiple queue network device support

["Waskiewicz Jr, Peter P" <peter.p.waskiewicz.jr@intel.com>]

> -----Original Message-----
> From: Thomas Graf [mailto:tgraf@suug.ch] 
> Sent: Friday, March 09, 2007 6:35 PM
> To: Waskiewicz Jr, Peter P
> Cc: Kok, Auke-jan H; David Miller; Garzik, Jeff; 
> netdev@vger.kernel.org; linux-kernel@vger.kernel.org; 
> Brandeburg, Jesse; Kok, Auke; Ronciak, John
> Subject: Re: [PATCH 1/2] NET: Multiple queue network device support
> 
> * Waskiewicz Jr, Peter P <peter.p.waskiewicz.jr@intel.com> 
> 2007-03-09 15:27
> > That's the entire point of this extra locking.  enqueue() 
> is going to 
> > put an skb into a band somewhere that maps to some queue, 
> and there is 
> > no way to guarantee the skb I retrieve from dequeue() is headed for 
> > the same queue.  Therefore, I need to unlock the queue 
> after I finish 
> > enqueuing, since having that lock makes little sense to dequeue().
> > dequeue() will then grab *a* lock on a queue; it may be the 
> same one 
> > we had during enqueue(), but it may not be.  And the 
> placement of the 
> > unlock of that queue is exactly where it happens in non-multiqueue, 
> > which is right before the hard_start_xmit().
> 
> The lock is already unlocked after dequeue, from your prio_dequeue():
> 
>        if (netif_is_multiqueue(sch->dev)) {
>                queue = q->band2queue[prio];
>                if 
> (spin_trylock(&sch->dev->egress_subqueue[queue].queue_lock)) {
>                        qdisc = q->queues[prio];
>                        skb = qdisc->dequeue(qdisc);
>                        if (skb) {
>                                sch->q.qlen--;
>                                skb->priority = prio;
>                                
> spin_unlock(&sch->dev->egress_subqueue[queue].queue_lock);
>                                return skb;
>                        }
>                        
> spin_unlock(&sch->dev->egress_subqueue[queue].queue_lock);
>        }

Ok, now I see what's wrong.  Taking Dave M.'s recommendation to store
the queue mapping in the skb will let me unlock the queue when dequeue()
returns.  I'll fix this locking issue; thanks for the feedback and
persistent drilling into my thick head.

-PJ Waskiewicz
peter.p.waskiewicz.jr@intel.com