>It's been a while since I looked at it, but Thor and I went to some
>effort a while ago to make sure that rnd(4) and arc4random() were
>at least initialiased before being used.
>
>There is certainly an issue of a shortage of good entropy early in
>the boot process, so the "quality" of randomness should be questioned,
>but for machines with rnd(4) there should not be collisions,
>especially if there is a cycle counter. For machines without rnd(4),
>it is possible, but all generators on such machines are suspect.

	i plan to commit sys/netinet/ip_id.c from openbsd (and its IPv6
	variant) very soon.  it will solve predictable fragment ID problem.
	so we'll go back to randomized ip_id very soon.

itojun