Re: [Qemu-devel] [PATCH] linux-user: manage SOCK_PACKET socket type.

[Peter Maydell]

On 27 October 2015 at 03:09, Laurent Vivier <address@hidden> wrote:
> Le 26/10/2015 15:40, Peter Maydell a écrit :
>> This confuses me. The packet(7) manpage suggests there are two flavours
>> of packet socket:
>>  (1) legacy AF_INET + SOCK_PACKET
>>  (2) new style AF_PACKET + SOCK_RAW / SOCK_DGRAM
>>
>> but this comment suggests it's trying to handle AF_PACKET + SOCK_PACKET ?
>
> In fact, I've started not from the man page, but from a non working dhcp
> client, originally with a m68k target and etch-m68k distro, and I've met
> again this problem on a ppc target and jessie distro.

> kernel 4.3.0-rc3, net/packet/af_packet.c:
>
>    2961
>    2962 static int packet_bind_spkt(struct socket *sock, struct sockaddr
> *uaddr,
>    2963                             int addr_len)
>    2964 {
>    2965         struct sock *sk = sock->sk;
>    2966         char name[15];
>    2967         struct net_device *dev;
>    2968         int err = -ENODEV;
>    2969
>    2970         /*
>    2971          *      Check legality
>    2972          */
>    2973
>    2974         if (addr_len != sizeof(struct sockaddr))
>    2975                 return -EINVAL;
>    2976         strlcpy(name, uaddr->sa_data, sizeof(name));
>    2977
>    2978         dev = dev_get_by_name(sock_net(sk), name);
>    2979         if (dev)
>    2980                 err = packet_do_bind(sk, dev, pkt_sk(sk)->num);
>    2981         return err;
>    2982 }
> ...
>    4246 static const struct proto_ops packet_ops_spkt = {
>    4247         .family =       PF_PACKET,
> ...
>    4250         .bind =         packet_bind_spkt,
> ...
>    3022
>    3023 static int packet_create(struct net *net, struct socket *sock,
> int protocol,
>    3024                          int kern)
> ...
>    3045         if (sock->type == SOCK_PACKET)
>    3046                 sock->ops = &packet_ops_spkt;
> ...

Yes, but also:
http://lxr.free-electrons.com/source/net/socket.c#L1109

1114         if (family == PF_INET && type == SOCK_PACKET) {
1115                 static int warned;
1116                 if (!warned) {
1117                         warned = 1;
1118                         pr_info("%s uses obsolete (PF_INET,SOCK_PACKET)\n",
1119                                 current->comm);
1120                 }
1121                 family = PF_PACKET;
1122         }

So I think my conclusion is:
 * Original 2.0 kernels used PF_INET + SOCK_PACKET
 * When the non-legacy stuff was added and this compat warning
   came in, the (accidental?) result was that PF_PACKET + SOCK_PACKET
   gave a warning, PF_PACKET + SOCK_PACKET gave legacy behaviour
   without a warning, and PF_PACKET + SOCK_RAW/SOCK_DGRAM gave
   the new interface
 * Some userspace programs responded not by updating to the new
   non-legacy interface, but by moving to PF_PACKET + SOCK_PACKET
   which just suppresses the warning

So we should special case both PF_INET + SOCK_PACKET and
PF_PACKET + SOCK_PACKET (but not any other PF_* + SOCK_PACKET).

thanks
-- PMM