Al Viro discovered in net/ipv[46]/raw.c raw_sendmsg() we have
                if (!inet->hdrincl)
                        raw_probe_proto_opt(&fl, msg);
                err = ip_route_output_flow(&rt, &fl, sk, !(msg-.....

So call sendmsg with sk->sk_protocol == IPPROTO_ICMP, msg->msg_iovlen = 1 and
msg->msg_iov[0].iov_base pointing to kernel space (.iov_len > 1).  Then in
raw_probe_proto_opt() we have type == msg->msg_iov[0].iov_base code ==
msg->msg_iov[0].iov_base + 1 when we hit 
         get_user(fl->fl_icmp_type, type);
         __get_user(fl->fl_icmp_code, code);

Note that failure of the first call is ignored and we happily do the second
call, an unchecked __get_user() which generates a memory read on the arbitrary
user-supplied address.

sendmsg() will fail if iovec points to kernel memory, but it may be possible to
trick ip_route_output_flow() into returning an error depending on the value of
 fl->fl_icmp_code.  Iff we can do that, the attacker can obtain information
about the value read from arbitrary kernel address by looking at the error
returned by sendmsg(2).

On some architectures (including x86) a 16 bit read at a user-supplied address
could access iomem (no separate mmu context) and therefore mess with hardware
state leading to a DoS.

"A local unprivileged user may be able to use this flaw to leak information
about the contents of kernel memory, or on some architectures cause a denial of
service by manipulating hardware state"