From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1750761AbVKEXnE (ORCPT ); Sat, 5 Nov 2005 18:43:04 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1751190AbVKEXnE (ORCPT ); Sat, 5 Nov 2005 18:43:04 -0500 Received: from fsmlabs.com ([168.103.115.128]:55992 "EHLO spamalot.fsmlabs.com") by vger.kernel.org with ESMTP id S1750761AbVKEXnD (ORCPT ); Sat, 5 Nov 2005 18:43:03 -0500 X-ASG-Debug-ID: 1131234180-8858-13-0 X-Barracuda-URL: http://10.0.1.244:8000/cgi-bin/mark.cgi Date: Sat, 5 Nov 2005 15:48:39 -0800 (PST) From: Zwane Mwaikambo To: "Jeffrey V. Merkey" cc: listmonkey@neo.relay-host.net, linux-kernel@vger.kernel.org X-ASG-Orig-Subj: Re: SMP CPU affinity questions Subject: Re: SMP CPU affinity questions In-Reply-To: <4368EBBF.2000302@wolfmountaingroup.com> Message-ID: References: <20051102175022.13637.qmail@neo.relay-host.net> <4368EBBF.2000302@wolfmountaingroup.com> MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII X-Barracuda-Spam-Score: 0.00 X-Barracuda-Spam-Status: No, SCORE=0.00 using global scores of TAG_LEVEL=1000.0 QUARANTINE_LEVEL=5.0 KILL_LEVEL=5.0 tests= X-Barracuda-Spam-Report: Code version 3.02, rules version 3.0.5075 Rule breakdown below pts rule name description ---- ---------------------- -------------------------------------------------- Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Wed, 2 Nov 2005, Jeffrey V. Merkey wrote: > IOApic's support binding of interrupt delivery in intel based platforms, but I > am unaware of tools which force this > setting by default on Linux, but someone else may be able to point you in that > direction. Most folks code APIC ICC delivery to > AV_LOPRI (meaning lowest priority processor gets next interrupt). This is > advantageous for cache coherency since the IRQ code > is probaby still in that processors cache. You may have to modify the kernel. > Linux doesn't allow processors to be shutdown and > reactiviated real time, it just starts them and lets them run, so you don;t > have to worry about the case of migrating interrupts > off pinned APICs. The APIC supports what you are asking for, but I am not > certain anyone implemented anything other > than AV_LOPRI settings by default in the IO APIC code. I would suggest you > look over the IO APIC Code -- this is a lot > of work, BTW. Or you could just put something which writes to /proc/irq/$IRQ/smp_affinity in your initscripts. Cheers, Zwane