From mboxrd@z Thu Jan 1 00:00:00 1970 From: Pavel Emelyanov Subject: [RFC][PATCH 0/3] Make loopback device weight twice as little. Date: Fri, 04 Apr 2008 18:09:05 +0400 Message-ID: <47F63681.1030603@openvz.org> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit To: Linux Netdev List Return-path: Received: from sacred.ru ([62.205.161.221]:33460 "EHLO sacred.ru" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751987AbYDDNpa (ORCPT ); Fri, 4 Apr 2008 09:45:30 -0400 Received: from [10.30.3.76] (swsoft-msk-nat.sw.ru [195.214.232.10]) (authenticated bits=0) by sacred.ru (8.13.8/8.13.8) with ESMTP id m34DjM8v007736 (version=TLSv1/SSLv3 cipher=DHE-RSA-AES256-SHA bits=256 verify=NO) for ; Fri, 4 Apr 2008 17:45:22 +0400 Sender: netdev-owner@vger.kernel.org List-ID: I noticed, that on a 32bit box with a non-debug config the sizeof(struct net_device) is slightly more than 1024. This means, that for example loopback device (with sizeof_priv == 0) is allocated from the size-2048 cache and thus we have ~1000 wasted bytes. I know, that this is not that much, all the more so on most of the setups the lo device is single :) but with the net-namespaced kernel each namespace has its own loopback device, so this problem becomes more relevant. I also guess, that the net_device struct allows many ways to get shrunk, but I noticed, that is has ~20 pointers on functions for different purposes, like init, hard_start_xmit, neigh_setup, etc. The proposal is to _move_ all these callbacks on a new structure called net_device_ops and put a pointer on it on the net_device. I've checked how much will the net_device weigh after this and it turned out, that this difference (80 bytes) is enough to get the 1024-sized net_device and thus a halved loopback. To make this change smooth, I propose the following plan (however, three patches I sent will demonstrate the idea better). * First - declare an empty nd_default_ops and make all new devices point on this new ops temporarily. * Then replace each call to dev->xxx() with the dev->nd_ops->xxx() one step by step. The ns_ops->xxx() will be the nd_default_ops->xxx then and they all will (temporarily) look like static void nd_default_xxx(struct net_device *dev, ...) { if (dev->xxx) dev->xxx(dev, ...) } * After this patch switch all the drivers (lots of work :( but it can be done driver-by-driver, i.e. without HUGE patches) from on-device ops to nd_ops pointer. * Finally - remove all the ops from the net_device and the default stubs described above. Does this idea worth being developed further? Thanks, Pavel