From mboxrd@z Thu Jan  1 00:00:00 1970
From: Grant Grundler <iod00d@hp.com>
Date: Fri, 02 Sep 2005 16:48:28 +0000
Subject: Re: [PATCH 2.6.13] IOCHK interface for I/O error handling/detecting (for ia64)
Message-Id: <20050902164828.GA10587@esmail.cup.hp.com>
List-Id: <linux-ia64.vger.kernel.org>
References: <431694DB.90400@jp.fujitsu.com> <20050901172917.I10072@chenjesu.americas.sgi.com>
In-Reply-To: <20050901172917.I10072@chenjesu.americas.sgi.com>
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-Transfer-Encoding: 7bit
To: Brent Casavant <bcasavan@sgi.com>
Cc: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>, linux-ia64@vger.kernel.org, Linux Kernel list <linux-kernel@vger.kernel.org>

On Thu, Sep 01, 2005 at 05:45:54PM -0500, Brent Casavant wrote:
...
> The first is serialization of all I/O reads and writes.  This will
> be a severe problem on systems with large numbers of PCI buses, the
> very type of system that stands the most to gain in reliability from
> these efforts.  At a minimum any locking should be done on a per-bus
> basis.

The lock could be per "error domain" - that would require some
arch specific support though to define the scope of the "error domain".

> The second is the raw performance penalty from acquiring and dropping
> a lock with every read and write.  This will be a substantial amount
> of activity for any I/O-intensive system, heck even for moderate I/O
> levels.

Sorry - I think this is BS.

Please run mmio_test on your box and share the results.
mmio_test is available here:
	svn co http://svn.gnumonks.org/trunk/mmio_test/

Then we can talk about the cost of spinlocks vs cost of MMIO access.

thanks,
grant