Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.2 9/5/84; site gondor.UUCP
Path: utzoo!watmath!clyde!cbosgd!ukma!psuvm.bitnet!psuvax1!gondor!steve
From: steve@gondor.UUCP (Stephen J. Williams)
Newsgroups: net.arch
Subject: Re: Single VA space/TLB flushing
Message-ID: <2106@gondor.UUCP>
Date: Fri, 18-Apr-86 20:11:12 EST
Article-I.D.: gondor.2106
Posted: Fri Apr 18 20:11:12 1986
Date-Received: Wed, 23-Apr-86 20:40:17 EST
References: <282@pyramid.UUCP> <5100054@ccvaxa>
Reply-To: steve@gondor.UUCP (Stephen J. Williams)
Organization: Pennsylvania State Univ.
Lines: 25

>
>What you've said is true of present systems, which tag entries with
>process numbers so that they don't have to physically flush the TLB.
>It helps that interrupt routines do not completely flush the TLB, but
>anything reasonably long-lived, like another process, eventually replace
>all of the previous process's TLB entries WHEN SOME OF THEM COULD HAVE
>BEEN REUSED - for shared objects like system space, if they were at the same
>virtual address.

this pushing out of entries will happen ANYHOW.  You have TLB entries
for lots of small segments, or for one big segment.  The effect is the
same, replace-wise.

>What I mean by a single virtual address space is one where all objects have
>unique virtual addresses, where the address ranges of processes do not
>collide, where all objects could at least potentially see each other if it
>were not for the protection system. There'd be no need for a tag then,
>because there is only one translation going on in the system at a time.

I still think this is unrealistic.  The internal workings of MVS demon-
strate the problems that you are trying to solve with bigness.  Bigness
is not an adequate solution.


--Scal