Re: Fragment ID wrap workaround (read-only, untested).

+	 * other end detect ID wrap. */
+	if (skb->sk) {
+		unsigned int slack;
+		struct inet_opt *inet = inet_sk(skb->sk);
+
+		slack = (left % mtu);
+		if (slack)
+			/* Shift by 8 bytes per id wrap. */
+			len = mtu - (slack % ((inet->id >> 16) << 3));

I'm pretty sure inet->id is wrong here. You would like the counter
in the inet peer entry. inet_sk(sk)->id is just used for the pseudo
counter in TCP that only works around VJ compression bugs. It's never used
for real fragmentation.

-Andi