> To me the -current code looks wrong, but I must be missing something,
> as the code is used on 64bit archs.  What have I missed?

Hmm.  Perhaps cgd just doesn't get enough fragments.  I know that when *I*
ported the networking stack to a 64-bit machine (the KSR/1), I abandoned
the overlay crap (and discovered that the next release of OSF/1 from OSF
contained exactly that change, indicating that *their* Alpha engineers
reached the same conclusion).

[That particular project was a lot of fun.  I did that work before the
availability of hardware, and didn't want to test under the 1/1000th speed
simulator, so I made identical changes to a near-UNIX system I was running
at home at the time (which used the 4.2 network stacks), plus I added a
modified loopback driver which (a) let a process receive and answer ICMP
ping packets, and (b) dropped one packet in about 5, giving a severe test
of fragment assembly and discard...]