> 
> Well, I can hardly speak for core.  But from what little I know of
> their reasons, I think they're unwilling to commit an un"clean" fix,
> for some definition of "clean".  I can to some extent understand this,
> since if the misbehavior goes away the incentive to come up with a good
> fix also largely goes away.  But in the case of bounce buffers, I don't
> think there is any clean fix, because it's an ugly hardware bug.  In
> the case of the VM shadow object leak, there probably is no clean fix
> short of redesigning the whole VM subsystem....
> 
The shadow problem saga was a long one on FreeBSD.  The best thing
that I can say is that any quick-fix will eventually breakdown.  We
implemented a "complete" and UGLY fix, and it was an "exercise" in
data structures.  

1) Make the swap pager PAGE addressible, so that you can collapse
objects on disk by doing meta-data frobbing.
2) Create a shadow chain in reverse across the shadow list.

Those two things, plus logic, will give you most of the shadow fix.

You do NOT need to rewrite the VM system to fix the problem.  Work-arounds
won't solve the problem either.

John
dyson@freebsd.org