[PATCH net-next] ndisc: use netdev_alloc_skb

[PATCH net-next] ndisc: use netdev_alloc_skb

[Stephen Hemminger]

IPv6 neighbor discovery should use netdev_alloc_skb which adds
extra space for headroom. The old code would allocate packet w/o any
additional headroom and force drivers to reallocate and copy the
packet in some cases.

Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>


--- a/net/ipv6/ndisc.c	2014-06-03 08:06:35.549942475 -0700
+++ b/net/ipv6/ndisc.c	2014-06-11 08:54:31.729764741 -0700
@@ -375,7 +375,7 @@ static struct sk_buff *ndisc_alloc_skb(s
 	struct sock *sk = dev_net(dev)->ipv6.ndisc_sk;
 	struct sk_buff *skb;
 
-	skb = alloc_skb(hlen + sizeof(struct ipv6hdr) + len + tlen, GFP_ATOMIC);
+	skb = netdev_alloc_skb(dev, hlen + sizeof(struct ipv6hdr) + len + tlen);
 	if (!skb) {
 		ND_PRINTK(0, err, "ndisc: %s failed to allocate an skb\n",
 			  __func__);
@@ -383,7 +383,6 @@ static struct sk_buff *ndisc_alloc_skb(s
 	}
 
 	skb->protocol = htons(ETH_P_IPV6);
-	skb->dev = dev;
 
 	skb_reserve(skb, hlen + sizeof(struct ipv6hdr));
 	skb_reset_transport_header(skb);

Re: [PATCH net-next] ndisc: use netdev_alloc_skb

[Eric Dumazet]

On Wed, 2014-06-11 at 11:31 -0700, Stephen Hemminger wrote:
> IPv6 neighbor discovery should use netdev_alloc_skb which adds
> extra space for headroom. The old code would allocate packet w/o any
> additional headroom and force drivers to reallocate and copy the
> packet in some cases.
> 
> Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
> 
> 
> --- a/net/ipv6/ndisc.c	2014-06-03 08:06:35.549942475 -0700
> +++ b/net/ipv6/ndisc.c	2014-06-11 08:54:31.729764741 -0700
> @@ -375,7 +375,7 @@ static struct sk_buff *ndisc_alloc_skb(s
>  	struct sock *sk = dev_net(dev)->ipv6.ndisc_sk;
>  	struct sk_buff *skb;
>  
> -	skb = alloc_skb(hlen + sizeof(struct ipv6hdr) + len + tlen, GFP_ATOMIC);
> +	skb = netdev_alloc_skb(dev, hlen + sizeof(struct ipv6hdr) + len + tlen);
>  	if (!skb) {

This looks strange.

netdev_alloc_skb() is mostly used for RX.

Perhaps you need to increase hlen with the needed room, because
NET_SKB_PAD might be only 32, and its a bit hidden.

ie use LL_RESERVED_SPACE_EXTRA() instead of LL_RESERVED_SPACE(), to make
clear what the intentions are.

Re: [PATCH net-next] ndisc: use netdev_alloc_skb

[Stephen Hemminger]

On Wed, 11 Jun 2014 12:58:22 -0700
Eric Dumazet <eric.dumazet@gmail.com> wrote:

> On Wed, 2014-06-11 at 11:31 -0700, Stephen Hemminger wrote:
> > IPv6 neighbor discovery should use netdev_alloc_skb which adds
> > extra space for headroom. The old code would allocate packet w/o any
> > additional headroom and force drivers to reallocate and copy the
> > packet in some cases.
> > 
> > Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
> > 
> > 
> > --- a/net/ipv6/ndisc.c	2014-06-03 08:06:35.549942475 -0700
> > +++ b/net/ipv6/ndisc.c	2014-06-11 08:54:31.729764741 -0700
> > @@ -375,7 +375,7 @@ static struct sk_buff *ndisc_alloc_skb(s
> >  	struct sock *sk = dev_net(dev)->ipv6.ndisc_sk;
> >  	struct sk_buff *skb;
> >  
> > -	skb = alloc_skb(hlen + sizeof(struct ipv6hdr) + len + tlen, GFP_ATOMIC);
> > +	skb = netdev_alloc_skb(dev, hlen + sizeof(struct ipv6hdr) + len + tlen);
> >  	if (!skb) {
> 
> This looks strange.
> 
> netdev_alloc_skb() is mostly used for RX.
> 
> Perhaps you need to increase hlen with the needed room, because
> NET_SKB_PAD might be only 32, and its a bit hidden.
> 
> ie use LL_RESERVED_SPACE_EXTRA() instead of LL_RESERVED_SPACE(), to make
> clear what the intentions are.

Never mind, I just saw a packet w/o space show up.
It was because of bad driver hard header len