On Tue, 1 Dec 1998 jaharkes_at_cs.cmu.edu wrote:
> The RVM data partition is copied into the server process memory image. It is 
> the way RVM is implemented. We might be able to use mmap to avoid using both 
> disk- and swap-space, but it is not yet possible to memory map a raw partition
> on linux. And we would have to make doubly sure it gives the same consistency 
> guarantees.

KeyKOS did this, actually.  It actually implemented something like RVM
for the whole system, checkpointing the whole system state every 30
seconds or so, so when rebooted, it could restore the state of the
whole system to where it was no more than 30 seconds before it lost
power.

In order to support DBMSs and similar programs, KeyKOS had a "journal"
call that would write some range of pages to disk before the next
scheduled checkpoint, and return when those pages were safely written
to disk.  So in order to provide durability guarantees about data, you
would just "journal" the pages containing those data.

Linux's mmap() has a similar system call: msync().

> How many programs can you `kill-9' and restart without losing information?

Linux, generally.  (How often do you lose data when you unexpectedly
power it down?)  Also Oracle, Sybase, etc.  IMHO, it's an absolute
requirement for a filesystem.

It sounds like you need three things:
- mmap() for raw disk partitions
- working msync() for raw disk partitions
- disks that let you disable any hardware lazy-write caching, and Linux
  support for such (maybe this already exists?)

> As not all parts of the RVM mapping are used very intensively, most of the
> memory occupied by RVM will get swapped out.

As you pointed out, it were better for it never to be swapped in.

mmap() poses some potential problems for middle-sized and large
filesystems, I'd think; 32 bits is an awfully small address space for
today's disks.  I can buy a 6GB disk for probably $300.

Kragen

-- 
<kragen@pobox.com>       Kragen Sitaker     <http://www.pobox.com/~kragen/>
It frightens the daylights out of me whenever I hear people proclaim that
the less knowledge our children have access to, the better.  
-- Duane Lindstrom, at <URL:http://www.examiner.com/skink/skinkmail.html>