On Thu, Oct 06, 2005 at 12:38:56AM +0900, YAMAMOTO Takashi wrote:
> > When I tried it behaved just like if the previous mapping was removed
> > when the new mapping was added (trap in the kernel when trying to access the
> > old address). Now it's possible I did make a mistake, but anyway avoiding
> > entering extra mapping when possible is probably better, performance-wise.
> > With a NetBSD guest, i've seen up to 4 packets in the same page in a row,
> > 2 is common. Theorically it can be up to 16 (PAGE_SIZE / MBYTES).
> 
> btw, shouldn't "xneti" member belong to xni_pkt rather than xni_page?

Ha right, we can have several domains on one page.
Initially I did the hash list per interface, but this wouldn't work well
for a domain with multiple interfaces. So I changed it to be global, but
forgot to move xneti.

> i think that it's better to have xennetback_tx_free() take xni_pkt
> rather than xni_page.

Ha, so we could acknowledge the packets sooner, that's your point ?
My idea was to defer packets ack as much as possible, to have a larger
batch of packets to process in the frontend, and reduce the number of
events sent to the frontend.

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
     NetBSD: 26 ans d'experience feront toujours la difference
--