From owner-freebsd-scsi  Tue Sep 14  8:47:55 1999
Delivered-To: freebsd-scsi@freebsd.org
Received: from arjun.niksun.com (gw.niksun.com [206.20.52.122])
	by hub.freebsd.org (Postfix) with ESMTP id 2395514BFD
	for <freebsd-scsi@freebsd.org>; Tue, 14 Sep 1999 08:47:52 -0700 (PDT)
	(envelope-from ath@niksun.com)
Received: from stiegl.niksun.com (stiegl.niksun.com [10.0.0.44])
	by arjun.niksun.com (8.8.8/8.8.8) with ESMTP id LAA16715;
	Tue, 14 Sep 1999 11:47:50 -0400 (EDT)
Received: (from ath@localhost)
	by stiegl.niksun.com (8.9.2/8.8.7) id LAA06404;
	Tue, 14 Sep 1999 11:47:46 -0400 (EDT)
	(envelope-from ath)
To: Andrew Gallatin <gallatin@cs.duke.edu>
Cc: freebsd-scsi@freebsd.org, gibbs@caspian.plutotech.com,
	anderson@cs.duke.edu, don.lewis@tsc.tdk.com
Subject: Re: data corruption when using aic7890
References: <199909110138.TAA03862@caspian.plutotech.com> <14301.2437.648605.15748@grasshopper.cs.duke.edu>
From: Andrew Heybey <ath@niksun.com>
Date: 14 Sep 1999 11:47:46 -0400
In-Reply-To: Andrew Gallatin's message of "Mon, 13 Sep 1999 11:45:57 -0400 (EDT)"
Message-ID: <85aeqpfqal.fsf@stiegl.niksun.com>
Lines: 42
X-Mailer: Gnus v5.5/XEmacs 20.4 - "Emerald"
Sender: owner-freebsd-scsi@FreeBSD.ORG
Precedence: bulk
X-Loop: FreeBSD.org

Andrew Gallatin <gallatin@cs.duke.edu> writes:

> Justin T. Gibbs writes:
>  > >I would continue to tweak the variable.  I assume you tried setting
>  > >the read threshold to MIN and the write threshold to MAX?  In other
>  > >words, don't start a read from host memory until the FIFO is almost
>  > >empty, and don't start a write to host memory until the FIFO is almost
>  > >full.  The assumption here is that PCI is faster than the SCSI bus
>  > >speed so we'll get the longest bursts this way.
>  > 
>  > I misread the data book.  If you set them both to MAX, then you would
>  > get this behavior.  The read fifo threshold is based on the amount
>  > of empty space in the fifo, not the amount of data in it.
> 
> 
> I've tried all the extremes, with the following results:
> 
> RD_DFTHRSH_MAX | WR_DFTHRSH_MIN:	corruption on read soon after first pass
> RD_DFTHRSH_MIN | WR_DFTHRSH_MIN:	corruption on read soon after first pass
> RD_DFTHRSH_MAX | WR_DFTHRSH_MAX:	immediate corruption (maybe write was corrupted?)
> RD_DFTHRSH_MIN | WR_DFTHRSH_MAX:	immediate corruption (maybe write was corrupted?)
> 
> Before I go marching through all the combinations, is it possible that
> there is a flag set someplace when a fifo overrun/underrun occurs that
> you're not checking?  Are the docs for this board available?
> 
> BTW, in case I wasn't clear originally, I'm not just running the test
> on 2 disks connected to the 7890 controller.  I'm also running our
> sequential read/write test on 2 identical disks connected to an ncr875 
> controller & to an IDE drive connected to the built-in PIIX4.  Eg, I'm 
> beating the snot out of the PCI bus by pushing about 50-70MB/sec of
> data across it from 3 separate PCI devices, so its not unreasonable to 
> expect that a fifo might fill up under these conditions.

Contention on the PCI bus seems to be the key to causing the problem.
In my case it was 10MByte/s of network traffic (with 44K
interrupts/sec) plus reading the disks as fast as possible.  I never
saw the problem with *just* reading disks, only when I added the
network traffic.  In your case you have multiple disk controllers
fighting for the bus.

andrew


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-scsi" in the body of the message