From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S965915AbXCPU7Y (ORCPT ); Fri, 16 Mar 2007 16:59:24 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S965925AbXCPU7Y (ORCPT ); Fri, 16 Mar 2007 16:59:24 -0400 Received: from netops-testserver-4-out.sgi.com ([192.48.171.29]:51091 "EHLO netops-testserver-4.corp.sgi.com" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S965915AbXCPU7X (ORCPT ); Fri, 16 Mar 2007 16:59:23 -0400 Date: Fri, 16 Mar 2007 13:59:21 -0700 (PDT) From: Christoph Lameter X-X-Sender: clameter@schroedinger.engr.sgi.com To: David Miller cc: andi@firstfloor.org, rostedt@goodmis.org, torvalds@linux-foundation.org, mbligh@mbligh.org, mingo@elte.hu, linux-kernel@vger.kernel.org, akpm@linux-foundation.org, chrisw@sous-sol.org, rusty@rustcorp.com.au, glommer@gmail.com Subject: Re: [PATCH 00/18] Make common x86 arch area for i386 and x86_64 - Take 2 In-Reply-To: <20070316.135545.119242282.davem@davemloft.net> Message-ID: References: <20070316.134718.95058720.davem@davemloft.net> <20070316.135545.119242282.davem@davemloft.net> MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Fri, 16 Mar 2007, David Miller wrote: > From: Christoph Lameter > Date: Fri, 16 Mar 2007 13:52:18 -0700 (PDT) > > > Virtual mmap allows holes in the same way as page tables do. > > I don't want to take expensive TLB misses to lookup a page. Ummm. You are missing key details please look at linux-mm. We will use huge tlb entries for the mmap and its optional. > TLB misses are extremely expensive on some platforms, that's why > I try to make all the critical structures go through the 4MB > locked TLB entry for the kernel image, or failing that the > 4MB linear mappings at PAGE_OFFSET on sparc64. Exactly. That is why 16M TLB entries speed up the kernel by 4-5% using vmmemmap on ia64. Basically we end up with one TLB per node for the vmemmap which always stays in the cache. No additional lookups. Without vmemmap we still need a 16M TLB for the 1-1 kernel mapping. This jsut cuts out the crud from sparsemem.