Subject: Re: pmap tweak sanity check
To: Jason R Thorpe <email@example.com>
From: Richard Earnshaw <firstname.lastname@example.org>
Date: 11/12/2002 21:25:19
> On Tue, Nov 12, 2002 at 04:15:25PM -0000, Chris Gilbert wrote:
> > An optimisation I've just thought of, is it worth adding flags to the
> > pvh_attrs field of the vm_page_md stuff that indicates that all entries in
> > this vm_page list are readonly and therefore all cacheable? that way when
> > we add a new readonly pv to the list we don't have to walk the list, we
> > just add it in.
> I'd use counters -- a flag might require a traversal of the list again
> to update the flag. A counter would not.
> Maybe use:
> - kernel-writable
> - user-writable
> - read-only
> as the counters? This way you only need to incr/decr the counters, and
> check them for != 0. Counters should be unsigned, and sprinkle some
> KDASSERT()s in there to detect underflow.
I was pondering whether a 2-bit counter of the number of pv entries would
actually be useful. Basically the counter would tell us if we have 1 2 or
many (=3) entries on the list and we could switch on that -- that would
take the common cases out of the need for entering a load of loops (in
fact it might well be possible to eliminate the vac-me-harder calls
altogether for the non-many cases).