Hello,

when I did tests (including repeated transferring of large kernels voer ARCnet
to the other test machine) of IPv6 over ARCnet, I found that, for some reason,
link level fragmented IP had a poor performance, while non-fragmented was
better.

After some testing, I finally discovered that cacheops are the culprits.
This is for a generic (in terms of CPU support) kernel only... 
An 68040-only  kernel on the 68040 machine does fine.

(Especially Amiga-Users) Keep that in mind when you get performance problems
on some i/o, especially ARCnet, with recent kernels that you didn't have
earlier.

I can think of some optimizations in the multi-cpu cacheop (and for some
few-cpu cacheop), but will need some time until I can implement and test...
Besides, non of this should be REALLY evil.

Reply-To: set to port-m68k, as the work to be eventually done is not Amiga
specific.

Regards,
	Ignatios