ipv6/ndisc: Preserve IPv6 control buffer if protocol error handlers are called

Commit a61bbcf2 ("[NET]: Store skb->timestamp as offset to a base
timestamp") introduces a neighbour control buffer and zeroes it out in
ndisc_rcv(), as ndisc_recv_ns() uses it.

Commit f2776ff0 ("[IPV6]: Fix address/interface handling in UDP and
DCCP, according to the scoping architecture.") introduces the usage of the
IPv6 control buffer in protocol error handlers (e.g. inet6_iif() in
present-day __udp6_lib_err()).

Now, with commit b94f1c09 ("ipv6: Use icmpv6_notify() to propagate
redirect, instead of rt6_redirect()."), we call protocol error handlers
from ndisc_redirect_rcv(), after the control buffer is already stolen and
some parts are already zeroed out. This implies that inet6_iif() on this
path will always return zero.

This gives unexpected results on UDP socket lookup in __udp6_lib_err(), as
we might actually need to match sockets for a given interface.

Instead of always claiming the control buffer in ndisc_rcv(), do that only
when needed.

Fixes: b94f1c09 ("ipv6: Use icmpv6_notify() to propagate redirect, instead of rt6_redirect().")
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Reviewed-by: Sabrina Dubroca <sd@queasysnail.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

net/ipv6/ndisc.c

@@ -1732,10 +1732,9 @@ int ndisc_rcv(struct sk_buff *skb)
 		return 0;
 	}
 
-	memset(NEIGH_CB(skb), 0, sizeof(struct neighbour_cb));
-
 	switch (msg->icmph.icmp6_type) {
 	case NDISC_NEIGHBOUR_SOLICITATION:
+		memset(NEIGH_CB(skb), 0, sizeof(struct neighbour_cb));
 		ndisc_recv_ns(skb);
 		break;