Everhart, Glenn
From:	Jerry Leichter [leichter@smarts.com]
Sent:	Tuesday, August 25, 1998 11:05 AM
To:	Info-VAX@Mvb.Saic.Com
Cc:	mcclelland@nesbit.enet.dec.com
Subject:	Re: Help required on Paging and segmentation in VMS ?
Every other response has denied that VMS has segments.  I'm going to
disagree, though slightly.

The VAX (and VMS, in its use of the VAX), uses segmentation in a very
basic form.  This shows up in the division of VM into P0, P1, and S0
regions - *each of which has a separate page table*, with its own
length.  If this were not the case, to keep the stack at the top of P1
and user data at the bottom of P0 would require a single, contiguous (in
S0) page table mapping the entire 2 gig of P0+P1, which would quickly
fill S0.  Since the bottom of S0 is contiguous with P1, this particular
consideration doesn't apply; however, having the S0 table separate makes
it very cheap (0 cost) for VMS to be mapped into each process
identically (thus avoiding all kinds of complications in interrupt-level
code).

Is this really "segmentation"?  If you can define the term, we can
perhaps answer the question!  The Multics model of segmentation (and the
rather similar models in Burroughs machines) is more general and
powerful than the truncated version in the Intel 286, and the much more
reasonable and complete versions in later x86 models.  What's present in
the VAX scheme is one of the main ideas in segmentation:  Segments can
be independently allocated, and can grow without impacting each other. 
What's missing is segment-relative addressing, which gives you a true
2-D address space; and, of course, the fixed, very small, number of
segments supported.  Perhaps that makes "regions" a better term than
"segments", but "regions" is much less well-defined even than "segment".

On Alpha, so much of the VM architecture is software-defined that it's
not reasonable to talk about whether the *hardware* supports segments. 
VMS on Alpha makes use of the hardware in a way that avoids the second
of the two ideas missing on the VAX:  With multi-level page tables, it
can do independent mapping for any number of non-contiguous regions/
segment/what have you.  This has relatively little impact on the
programming interface - it just means that some operations that were
very expensive on the VAX (allocating a region well beyond the current
top of P0 would require creating page table entries for all the unused
space "in between") are now cheap.  (The difference in cost does have a
significant practical effect on how the programming interface will get
*used*, of course.)

The Unix VM model, which grew out of the VAX model on which it was
originally developed, started off with the same three regions/segments,
though under different names:  data/text; stack; system.  With the
addition of shared memory and memory mapping calls, things got more
complex, and the implementations tend to have separate regions/segments
for individual shared memory/mmap'ed VM ranges.  Early implementations
of these services were inflexible and limited.  Modern Unixes tend to
use a style not very different from what VMS uses.

To go back to the original (exam) question:  As a former teacher of OS
courses, I wouldn't have posed the question this way.  Perhaps it might
make sense within the context of a course that carefully defined such
terms as "segmentation" for the purpose of the course; but I would never
try to do such a thing, since it would only be confusing to students in
any other context, where "segmentation" is used for so wide a variety of
related, but by no means identical, things.
							-- Jerry