From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S261681AbUL3Ro2 (ORCPT ); Thu, 30 Dec 2004 12:44:28 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S261682AbUL3Ro2 (ORCPT ); Thu, 30 Dec 2004 12:44:28 -0500 Received: from [61.49.235.157] ([61.49.235.157]:16883 "EHLO adam.yggdrasil.com") by vger.kernel.org with ESMTP id S261681AbUL3RoW (ORCPT ); Thu, 30 Dec 2004 12:44:22 -0500 Date: Thu, 30 Dec 2004 09:33:32 -0800 From: "Adam J. Richter" Message-Id: <200412301733.iBUHXW719876@adam.yggdrasil.com> To: maneesh@in.ibm.com Subject: Re: [Patch] Do not allocate sysfs_dirent.s_children for non-directories Cc: akpm@osdl.org, chrisw@osdl.org, greg@kroah.com, linux-kernel@vger.kernel.org, viro@parcelfarce.linux.theplanet.co.uk Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Thu, 30 Dec 2004 19:05:29 +0530, Maneesh Soni wrote: >On Thu, Dec 02, 2004 at 07:20:35PM -0800, Adam J. Richter wrote: >> The following patch, against a heavily hacked 2.6.10-rc2-bk15 >> sysfs tree, removes the s_children field from sysfs_dirent and >> creates a new structure just for directories named sysfs_dir, which >> embeds a sysfs_dirent and also adds s_children. Directories allocate >> a sysfs_dir; non-directories allocate a sysfs_dirent. There are >> two separate kmem caches for the different data types. >> >> Not allocating s_children from each non-directory saves >> two pointers (8 bytes) for each of the 2573 non-directory nodes >> in my sysfs tree, or about 20kB on unswappable memory, but >> having another kmem cache probably wastes an average of half >> a page in memory fragmentation and then there is are few >> bytes from the new code and the additional kmem_cache_t >> structure, so I would guess it probably saves about 16kB in >> practice. >> >> In the future, I hope to make a similar change for symbolic >> links. >> >> By the way, this patch will also make it easier for me to >> try to unpin sysfs directories because there are a few other >> fields specific to directories that I would want to store >> in sysfs_dir. >> >Apart from a couple of diff'ing related comments, I feel the code looks >some what complicated. I think we can directly link sysfs_dir to >directory dentries and sysfs_dirent to non-directory dentries instead >of always linking sysfs_dirent to d_fsdata. To differentiate between the >two types of structures linked to dentry's d_fsdata field, we can use >S_ISDIR(dentry->d_inode->i_mode). This will avoid using container_of() and >the dentry_to_sysfs_dir() conversions. >Most of the places we may not need to find what type of struct d_fsdata points >to. Like in sysfs_make_dirent(), first param has to be sysfs_dir as the parent >dentry corresponds to a sysfs directory. >In sysfs_lookup() also, we know parent dentry corresponds to sysfs directory >so dentry's d_fsdata will point to sysfs_dir instead of sysfs_dirent. In the future, I want to make a change so that attributes in a named struct attribute_group do not each have a struct dirent. In that case, it is possible that the attribute_group will only need a struct sysfs_dirent, not a struct sysfs_dir. __ ______________ Adam J. Richter \ / adam@yggdrasil.com | g g d r a s i l