Re: [PATCH] bonding: Fix corrupted queue_mapping

[Eric Dumazet <eric.dumazet@gmail.com>]

On Fri, 2012-06-08 at 08:47 +0200, Eric Dumazet wrote:
> On Thu, 2012-06-07 at 23:15 -0700, David Miller wrote:
> > From: Eric Dumazet <eric.dumazet@gmail.com>
> > Date: Fri, 08 Jun 2012 08:11:21 +0200
> > 
> > > On Thu, 2012-06-07 at 23:02 -0700, David Miller wrote:
> > >> Hmmm, isn't that what qdisc_skb_cb is for?  And even private data is
> > >> explicitly allocated:
> > >> 
> > >> >  	unsigned char		data[24];
> > >> 
> > >> there. :-)
> > >> 
> > > 
> > > Yes, but some other layers can use the same trick so it might collide.
> > > 
> > > Inserting the bond field in qdisc_skb_cb (level0) is safer.
> > 
> > Do you suggest that Infiniband does the same thing? :-)
> 
> I wonder if another way to solve this is not letting ndo_select_queue()
> method the responsibility to call skb_set_queue_mapping() itself ?
> 
> (ie removing skb_set_queue_mapping() done in dev_pick_tx())
> 
> bonding would not have to save/restore skb queue mapping ?
> 
> Partial patch : (we have to audit all ndo_select_queue()
> 
> diff --git a/net/core/dev.c b/net/core/dev.c
> index cd09819..c6c92d5 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -2368,6 +2368,7 @@ static struct netdev_queue *dev_pick_tx(struct net_device *dev,
>  
>  	if (dev->real_num_tx_queues == 1)
>  		queue_index = 0;
> +		skb_set_queue_mapping(skb, queue_index);
>  	else if (ops->ndo_select_queue) {
>  		queue_index = ops->ndo_select_queue(dev, skb);
>  		queue_index = dev_cap_txqueue(dev, queue_index);
> @@ -2391,9 +2392,9 @@ static struct netdev_queue *dev_pick_tx(struct net_device *dev,
>  					sk_tx_queue_set(sk, queue_index);
>  			}
>  		}
> +		skb_set_queue_mapping(skb, queue_index);
>  	}
>  
> -	skb_set_queue_mapping(skb, queue_index);
>  	return netdev_get_tx_queue(dev, queue_index);
>  }
>  
> 


I must say I dont understand dev_pick_tx() anymore.

It seems to ignore skb->queue_mapping (unless device provides its own
ndo_select_queue() and this functions is aware of skb->queue_mapping, as
correctly done in ixgbe)

So commit fff3269907897ee (tcp: reflect SYN queue_mapping into SYNACK
packets) works on ixgbe, but probably not on other multiqueue devices.

This sounds like a regression to me.