>If possible a better approach might be to use the host route allocated to the
>outer tunnel endpoint (oops, another can of worms if you look at the code,
>and subject to the recursion bug discussed here some time ago) and use the
>route mtu. I don't know how that could be updated when PMTUD changes it, but
>if it would be possible, we'd get the optimal MTU for the gre interface.

	PMTUD is designed with assumption that link MTUs are stable (do not
	change over time).  I think it causes trouble if we decide MTU
	dynamically.

itojun