Re: [PATCH] Add missing skb->dev assignment in Frame Relay RX code

[Stephen Hemminger <shemminger@vyatta.com>]

On Sun, 29 Jun 2008 23:10:11 +0200
Krzysztof Halasa <khc@pm.waw.pl> wrote:

> Stephen Hemminger <shemminger@vyatta.com> writes:
> 
> >> --- a/drivers/net/wan/hdlc_fr.c
> >> +++ b/drivers/net/wan/hdlc_fr.c
> >> @@ -1008,6 +1008,7 @@ static int fr_rx(struct sk_buff *skb)
> >>  		stats->rx_bytes += skb->len;
> >>  		if (pvc->state.becn)
> >>  			stats->rx_compressed++;
> >> +		skb->dev = dev;
> >>  		netif_rx(skb);
> >>  		return NET_RX_SUCCESS;
> >>  	} else {
> >
> > Better to use netdev_alloc_skb for receive buffers instead.
> > --- a/drivers/net/wan/hdlc_fr.c	2008-06-29 13:02:42.000000000 -0700
> > +++ b/drivers/net/wan/hdlc_fr.c	2008-06-29 13:04:01.000000000 -0700
> > @@ -515,7 +515,7 @@ static void fr_lmi_send(struct net_devic
> >  		}
> >  	}
> >  
> > -	skb = dev_alloc_skb(len);
> > +	skb = netdev_alloc_skb(dev, len);
> >  	if (!skb) {
> >  		printk(KERN_WARNING "%s: Memory squeeze on fr_lmi_send()\n",
> >  		       dev->name);
> 
> Well, no, that's another story - the missing assignment is in fr_rx(),
> in the regular receive path (similar to 802.1q case, ethX ->
> ethX.VLAN_ID transition). I.e., we get incoming packet from a hardware
> driver, with skb->dev pointing to hardware hdlcX device, and we change
> it to point to a logical pvcX device.
> 
> fr_lmi_send() is for control messages (generated locally) only, in TX
> path. IOW we alloc skb and send it through hw driver immediately. Not
> related to the bug.
> 
> That said, perhaps I should indeed use netdev_alloc_skb() for those
> control messages? Or is it RX-only?

netdev_alloc_skb does two things:
  1) it sets skb->dev
  2) on some platforms it can choose memory "closer" to the device.
     but this is really a NUMA issue