Our customer encountered stuck NFS writes for blocks starting at specific
offsets w.r.t. page boundary caused by networking stack sending packets via
UFO enabled device with wrong checksum. The problem can be reproduced by
composing a long UDP datagram from multiple parts using MSG_MORE flag:

  sendto(sd, buff, 1000, MSG_MORE, ...);
  sendto(sd, buff, 1000, MSG_MORE, ...);
  sendto(sd, buff, 3000, 0, ...);

Assume this packet is to be routed via a device with MTU 1500 and
NETIF_F_UFO enabled. When second sendto() gets into __ip_append_data(),
this condition is tested (among others) to decide whether to call
ip_ufo_append_data():

  ((length + fragheaderlen) > mtu) || (skb && skb_is_gso(skb))

At the moment, we already have skb with 1028 bytes of data which is not
marked for GSO so that the test is false (fragheaderlen is usually 20).
Thus we append second 1000 bytes to this skb without invoking UFO. Third
sendto(), however, has sufficient length to trigger the UFO path so that we
end up with non-UFO skb followed by a UFO one. Later on, udp_send_skb()
uses udp_csum() to calculate the checksum but that assumes all fragments
have correct checksum in skb->csum which is not true for UFO fragments.

When checking against MTU, we need to add skb->len to length of new segment
if we already have a partially filled skb and fragheaderlen only if there
isn't one.

In the IPv6 case, skb can only be null if this is the first segment so that
we have to use headersize (length of the first IPv6 header) rather than
fragheaderlen (length of IPv6 header of further fragments) for skb == NULL.

Fixes: e89e9cf5 ("[IPv4/IPv6]: UFO Scatter-gather approach")
Fixes: e4c5e13a ("ipv6: Should use consistent conditional judgement for
	ip6 fragment between __ip6_append_data and ip6_finish_output")
Signed-off-by: Michal Kubecek <mkubecek@suse.cz>
Acked-by: Vlad Yasevich <vyasevic@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>