On Thu, Jan 18, 2018 at 07:31:56PM +0000, Al Viro wrote:

* SO_RCVTIMEO/SO_SNDTIMEO handling in compat [sg]etsockopt()
* passing SIOC{ADD,DEL}TUNNEL down (ipmr_del_tunnel(),ipmr_new_tunnel(),
  addrconf_set_dstaddr())
* SIOCGSTAMP/SIOCGSTAMPNS in compat ioctls
* SIOCADDRT/SIOCDELRT in compat ioctls
* kernel_[gs]etsockopt()
* ipv6_renew_options_kern()

I don't know if all of that stuff can be realistically done without set_fs().
kernel_setsockopt(), in particular, is unpleasant...

Speaking of weird indirect calls: in net/packet/af_packet.c:packet_ioctl() we
have this:
#ifdef CONFIG_INET
        case SIOCADDRT:
        case SIOCDELRT:
        case SIOCDARP:
        case SIOCGARP:
        case SIOCSARP:
        case SIOCGIFADDR:
        case SIOCSIFADDR:
        case SIOCGIFBRDADDR:
        case SIOCSIFBRDADDR:
        case SIOCGIFNETMASK:
        case SIOCSIFNETMASK:
        case SIOCGIFDSTADDR:
        case SIOCSIFDSTADDR:
        case SIOCSIFFLAGS:
                return inet_dgram_ops.ioctl(sock, cmd, arg);
#endif

That's inet_ioctl(sock, cmd, arg) disguised by indirect.  AFAICS,
that line dates back to 2.1.89; back then inet_dgram_ops had been
exported and inet_ioctl() had been static.  When SCTP went in
they'd exported inet_ioctl() rather than playing that kind of
games.

Is there anything subtle I'm missing here that would make it
wrong to replace that with explicit call of inet_ioctl()?