On receiving a packet too big icmp error we check if our current cached
dst_entry in the socket is still valid. This validation check did not
care about the expiration of the (cached) route.

The error path I traced down:
The socket receives a packet too big mtu notification. It still has a
valid dst_entry and thus issues the ip6_rt_pmtu_update on this dst_entry,
setting RTF_EXPIRE and updates the dst.expiration value (which could
fail because of not up-to-date expiration values, see previous patch).

In some seldom cases we race with a) the ip6_fib gc or b) another routing
lookup which would result in a recreation of the cached rt6_info from its
parent non-cached rt6_info. While copying the rt6_info we reinitialize the
metrics store by copying it over from the parent thus invalidating the
just installed pmtu update (both dsts use the same key to the inetpeer
storage). The dst_entry with the just invalidated metrics data would
just get its RTF_EXPIRES flag cleared and would continue to stay valid
for the socket.

We should have not issued the pmtu update on the already expired dst_entry
in the first placed. By checking the expiration on the dst entry and
doing a relookup in case it is out of date we close the race because
we would install a new rt6_info into the fib before we issue the pmtu
update, thus closing this race.

Not reliably updating the dst.expire value was fixed by the patch "ipv6:
reset dst.expires value when clearing expire flag".
Reported-by: Steinar H. Gunderson <sgunderson@bigfoot.com>
Reported-by: Valentijn Sessink <valentyn@blub.net>
Cc: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org>
Signed-off-by: Hannes Frederic Sowa <hannes@stressinduktion.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Tested-by: Valentijn Sessink <valentyn@blub.net>
Signed-off-by: David S. Miller <davem@davemloft.net>