> First, the skb_header_cloned() patch I posted in another reply will
quoted text
> get rid of copying due to clones.  I've included it below for
> completeness.
>=20
> The next problem is to make sure there is enough space available.  And
> all that's needed is some help from the bridging layer and some hooks
> into netdev_alloc_skb().
>=20
> On bridge transmit, it knows the input and output devices, and the
> requirements of LL header space on the transmit side.
>=20
> If the transmit requirements are not met for the received packet, we
> can tag the difference into the input netdev.
>=20
> Using that information we can allocate extra space in
> netdev_alloc_skb(), and do an skb_reserve().
>=20
> The only requirement is that the ethernet driver serving input packets
> uses netdev_alloc_skb().  The most important drivers already do, and
> those which do not are trivially converted.
>=20
> The following along with Johannes's needed_header et al. patch should
> take care of the overhead.
>=20
> If this proves to be a working solution, we can do something similar
> for IPv4 and IPv6 forwarding.  At that point, we should make this
> "adjust input device extra space" a helper function that all of
> these spots can call.

This looks feasible, thanks for looking into it. But isn't, when the
bridge gives us a cloned skb, the complete header the complete data
since it won't be a fraglist skb? And then we copy it all anyway? That's
just for group addressed frames, of course, so still a lot better I
guess. The main point is getting enough headroom.

Also, should there be some sort of timer that resets the rx_alloc_extra
again so that when you bridge it once with p54 (needs heaps of headroom)
you don't suffer forever?

quoted text
> @@ -255,11 +255,12 @@ struct sk_buff *__netdev_alloc_skb(struct net_devic=
e *dev,
quoted text
>  		unsigned int length, gfp_t gfp_mask)
>  {
>  	int node =3D dev->dev.parent ? dev_to_node(dev->dev.parent) : -1;
> +	unsigned int extra =3D dev->rx_alloc_extra + NET_SKB_PAD;
>  	struct sk_buff *skb;
> =20
> -	skb =3D __alloc_skb(length + NET_SKB_PAD, gfp_mask, 0, node);
> +	skb =3D __alloc_skb(length + extra, gfp_mask, 0, node);
>  	if (likely(skb)) {
> -		skb_reserve(skb, NET_SKB_PAD);
> +		skb_reserve(skb, extra);

Doesn't that break alignment though?

johannes