From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1750990AbWFDAtE (ORCPT ); Sat, 3 Jun 2006 20:49:04 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1751082AbWFDAtE (ORCPT ); Sat, 3 Jun 2006 20:49:04 -0400 Received: from terminus.zytor.com ([192.83.249.54]:64460 "EHLO terminus.zytor.com") by vger.kernel.org with ESMTP id S1750990AbWFDAtD (ORCPT ); Sat, 3 Jun 2006 20:49:03 -0400 Message-ID: <44822DF4.5090504@zytor.com> Date: Sat, 03 Jun 2006 17:48:52 -0700 From: "H. Peter Anvin" User-Agent: Thunderbird 1.5.0.2 (X11/20060501) MIME-Version: 1.0 To: Robert Hancock CC: linux-kernel Subject: Re: memcpy_toio on i386 using byte writes even when n%2==0 References: <6ined-4gY-17@gated-at.bofh.it> <6ined-4gY-21@gated-at.bofh.it> <6inee-4gY-23@gated-at.bofh.it> <6ined-4gY-15@gated-at.bofh.it> <6inxv-4U2-17@gated-at.bofh.it> <448220CE.4030709@shaw.ca> In-Reply-To: <448220CE.4030709@shaw.ca> Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org Robert Hancock wrote: > H. Peter Anvin wrote: >> For something that generates I/O transactions, it's imperative to >> generate the smallest possible number of transactions. Furthermore, >> smaller than dword transactions aren't burstable, except at the >> beginning and end of a burst. > > Well, theoretically for writes they could be, if the memory region was > prefetchable and the PCI chipset supported byte merge. It certainly > isn't optimal however. > If so, then merging doesn't matter either way. -hpa