Re: VMM

[Niels Möller]

address@hidden (Neal H. Walfield) writes:

> > And the physical memory
> > server has to export the needed information.
> 
> Which functionality?

I was thinking about page usage information. The process (or its
pager, if they are different) already gets information about page
faults. But most cpu's memory management hardware also keep other
information about each mapped page, such as dirtyness (can be
implemented by the kernel or pager if it isn't done by the hardware)
and usage bits (too expensive to do in software, if the hardware
doesn't support it).

This is information that's available to the kernel code, (running in
supervisor mode) and which seems useful when making decisions on which
pages to throw out. So if that decision making is moved to userspace
(which seems like the right thing to do), then one should also give
userspace some way to get to that information about the pages.

One can use LRU as a simple design testcase: It may suck as a paging
strategy (I don't really know enough about OS research to debate one
way or the other), but if a userspace process can't implement least
recently used for some or all of its pages, then the paging interfaces
and procedures are too limited.

Another design testcase is a userspace LISP kernel or other garbage
collector that wants to use available memory management hardware to
implement an efficient read or write barrier.

/Niels