Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!utgpu!water!watnot!watmath!clyde!rutgers!sri-spam!ames!cit-vax!mangler
From: mangler@cit-vax.UUCP
Newsgroups: comp.unix.questions
Subject: Re: Booting UNIX on 11/785
Message-ID: <2199@cit-vax.Caltech.Edu>
Date: Thu, 2-Apr-87 05:09:05 EST
Article-I.D.: cit-vax.2199
Posted: Thu Apr  2 05:09:05 1987
Date-Received: Sat, 4-Apr-87 11:48:29 EST
References: <5472@brl-adm.ARPA>
Organization: California Institute of Technology
Lines: 31
Summary: botch in determining position of error

In article <5472@brl-adm.ARPA>, rbj@icst-cmr.arpa (Root Boy Jim) writes:
> The problem seems to be that SI's report bad blocks differently.

MBA's have two 16-bit byte counters (in one 32-bit register) to aid
in determining which block got the error.  One tallies the bytes
transferred to/from memory, the other tallies bytes to/from the disk.
The former is what matters when reading, and the latter when writing.
Unfortunately, the SI 9900 only implements the former, and has enough
buffering that the distinction is critical.  When a write encounters
a bad sector, the bytes are counted even though they don't make it
onto the disk.

There is an "SI ECC BUG" fix floating around which looks at the disk
address register instead.  This can lose bytes on large reads that
end with an error.  (We've used this patch a long time, and I'm
amazed at how well it works in practice).

The correct way to do this is to look at the disk address register
and both byte counters, take the minimum of the three, and continue
the transfer from there (with fudging for DCK and SKI).  This works
on SI 9900 and Emulex SC7000 (even with Rev. B proms, which are
incredibly buggy).

> I don't see how you can do this without a "normal" disk, altho one
> guy around here runs a 780 with *only* SI controllers and 9766/9751's.

So does the 780 I'm posting on.  I think we get away with it simply
by having few bad blocks.  The pack on the 9766 was bought error-free
and we use that as the root disk.

Don Speck   speck@vlsi.caltech.edu  {seismo,rutgers,ames}!cit-vax!speck