S.J. Borrill wrote:

> a) the vastly speeded up pmap stuff that Neil kept talking about

It's done. A hunk of it has been merged: The closing of the pmap_enter
whole, the acceleration of pmap_copy_page/pmap_zero_page, rewrite of
pmap_remove, the removal of unnecessary cache flushes from the context
switch, the removal of unnecessary write buffer drains from most CPU
functions (and the addition of one to the bus_dma code) have all been
done.

Compared to a kernel of six months ago, you should notice a pretty big
performance increase running scripts, using programs that fork etc.

The one last remaining thing to be added is an improvement to the speed
of execs, but I don't think Mark liked the way I did it, and wants to do
it in a more structured way instead. This changed the number of entire
cache invalidates in an exec from 12 to 2 :-)

But things should have improved a fair bit already. Try running an
lmbench compared to a kernel six months ago.

-- 
Neil A. Carson