From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S261656AbUKSWpl (ORCPT ); Fri, 19 Nov 2004 17:45:41 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S261657AbUKSWdh (ORCPT ); Fri, 19 Nov 2004 17:33:37 -0500 Received: from omx1-ext.sgi.com ([192.48.179.11]:23687 "EHLO omx1.americas.sgi.com") by vger.kernel.org with ESMTP id S261669AbUKSWdO (ORCPT ); Fri, 19 Nov 2004 17:33:14 -0500 Date: Fri, 19 Nov 2004 15:18:35 -0600 From: Jack Steiner To: linux-kernel@vger.kernel.org Subject: SLIT and IO-only nodes Message-ID: <20041119211835.GA21349@sgi.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline User-Agent: Mutt/1.4.1i Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org An SGI Altix SSI system consists of a collection of nodes connected via a high speed interconnect. Nodes come in several flavors: - memory, IO and cpus - memory & cpus - memory only - IO only - (other combinations don't exist yet) The first 2 types of nodes are typical. You can think of the last 2 types of nodes as nodes that have been partially depopulated. We need to describe all these nodes in the SLIT table. For example, when allocating memory on a memory-only node, knowing the distance to the node is important. When assigning cpus to service interrupts for IO nodes or when creating driver memory structures for devices on IO nodes, it is important to use the nearest node that has cpus & memory. On IA64, memory-only nodes are described in the SRAT, have a proximity domain number, NIDs, and appear in the SLIT. (ie., we don't have a problem with memory-only nodes). However, IO-only nodes (AFAICT) cannot be described in the SLIT. The SLIT is indexed by proximity_domain_number (PXM). Currently, there is no SRAT entry for IO-only nodes. These nodes do not appear in the SLIT. It would seem that a new ACPI table is needed for IO-only nodes. The SRAT would describe the node & identify the IO buses that are attached to the node. I think this would give us what we need. Before I start digging into the ACPI spec, has anyone already addressed this problem? Is this the right approach to take to solve the problem? -- Thanks Jack Steiner (steiner@sgi.com) 651-683-5302 Principal Engineer SGI - Silicon Graphics, Inc.