> http://johannes.sipsolutions.net/patches/kernel/all/2008-05-04-11%3a23/02=
7-skb-alloc-stackdump.patch

This was broken when cloning, this one works:

http://johannes.sipsolutions.net/patches/kernel/all/2008-05-04-12%3a19/027-=
skb-alloc-stackdump.patch

Now I see the problem. I increased the LL_MAX_HEADER constant, but all
code uses dev->hard_header_len to allocate the headroom (via
LL_RESERVED_SPACE), e.g. packet.c:

packet_sendmsg_spkt:
        skb =3D sock_wmalloc(sk, len + LL_RESERVED_SPACE(dev), 0, GFP_KERNE=
L); =20
  =20
[...]
        /* FIXME: Save some space for broken drivers that write a
         * hard header at transmission time by themselves. PPP is the
         * notable one here. This should really be fixed at the driver leve=
l.
         */
        skb_reserve(skb, LL_RESERVED_SPACE(dev));

This one even complains about "broken drivers" (like PPP, but wireless
code behaves like this too). This is getting really really really
frustrating. All kinds of comments all over tell you how this all is
wrong but NEVER actually tell you how to do it correctly!

As far as I understand, I cannot change dev->hard_header_len because I
want/need an ethernet header and not more. But then I don't get enough
headroom.

Oddly enough, I even get a warning from tcp_connect() although that
actually does
        /* Reserve space for headers. */
        skb_reserve(buff, MAX_TCP_HEADER);

which should definitely be sufficient. Maybe the packet gets cloned
somewhere? On the other hand, IPv4 raw's =EF=BB=BFraw_send_hdrinc reserves
LL_RESERVED_SPACE(rt->u.dst.dev) like af_packet....

johannes