From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754616AbZBDCtL (ORCPT ); Tue, 3 Feb 2009 21:49:11 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1751880AbZBDCs4 (ORCPT ); Tue, 3 Feb 2009 21:48:56 -0500 Received: from gw1.cosmosbay.com ([212.99.114.194]:51306 "EHLO gw1.cosmosbay.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751695AbZBDCsz convert rfc822-to-8bit (ORCPT ); Tue, 3 Feb 2009 21:48:55 -0500 Message-ID: <498901E7.4050405@cosmosbay.com> Date: Wed, 04 Feb 2009 03:48:07 +0100 From: Eric Dumazet User-Agent: Thunderbird 2.0.0.19 (Windows/20081209) MIME-Version: 1.0 To: Davide Libenzi CC: Andrew Morton , Jonathan Corbet , Linux Kernel Mailing List , andi@firstfloor.org, oleg@redhat.com, viro@ZenIV.linux.org.uk, David Miller , hch@lst.de, linux-api@vger.kernel.org, mpm@selenic.com, Alan Cox Subject: Re: [PATCH 2/4] Convert epoll to a bitlock References: <1233598811-6871-1-git-send-email-corbet@lwn.net> <1233598811-6871-3-git-send-email-corbet@lwn.net> <20090203133942.2ecec281.akpm@linux-foundation.org> <4988BD4E.8080206@cosmosbay.com> In-Reply-To: Content-Type: text/plain; charset=ISO-8859-15 Content-Transfer-Encoding: 8BIT X-Greylist: Sender IP whitelisted, not delayed by milter-greylist-1.6 (gw1.cosmosbay.com [0.0.0.0]); Wed, 04 Feb 2009 03:48:08 +0100 (CET) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Davide Libenzi a écrit : > On Tue, 3 Feb 2009, Eric Dumazet wrote: > >> Andrew Morton a écrit : >>> On Mon, 2 Feb 2009 11:20:09 -0700 >>> Jonathan Corbet wrote: >>> >>>> Matt Mackall suggested converting epoll's ep_lock to a bitlock as a way of >>>> saving space in struct file. This patch makes that change. >>> hrm. bit_spin_lock() makes people upset (large penguiny people). iirc >>> it doesn't have all the correct/well-understood memory/compiler >>> ordering semantics which spinlocks have. And lockdep doesn't know about >>> it. >>> >> In a previous attempt (2005), I suggested using a single global lock. >> >> http://search.luky.org/linux-kernel.2005/msg50862.html >> >> Probably an array of hashed spinlocks would be more than enough. > > That could be done, although I'm not sure it's worth going that way to > save 4 bytes. The effective saving rate is not even 4/sizeof(struct file) > since struct file never comes alone, and when you allocate a struct file > you always carry more allocations behind (at least for the cases where you > tend to have a lot of them around, so size would matter). > The add/remove path in epoll is not a super-hot one, so it could be done. > I dunno how this change matter with the patchset though. Back in 2005, I saved 4 bytes per file, and because of HWCACHE alignment, sizeof(struct file) shrinked by 64 bytes. With more than 1.000.000 sockets opened on a busy server, it saved 64 MB of ram. At that time, this mattered (8GB of ram), but in 2009, 64 MB is so small I dont care anymore about sizeof(struct file) AFAIK, I just checked on x86_64 and got : sizeof(struct file)=0xc0 , so thats perfect :) (Only thing I still do is to move private_data in the first cache line of struct file, because it speedups a lot socket operation, when dealing with 1.000.000 sockets : one cache line miss avoided per socket syscall) diff --git a/include/linux/fs.h b/include/linux/fs.h index 6022f44..03b2227 100644 --- a/include/linux/fs.h +++ b/include/linux/fs.h @@ -842,6 +842,8 @@ struct file { #define f_dentry f_path.dentry #define f_vfsmnt f_path.mnt const struct file_operations *f_op; + /* needed for tty driver, and maybe others */ + void *private_data; atomic_long_t f_count; unsigned int f_flags; fmode_t f_mode; @@ -854,8 +856,6 @@ struct file { #ifdef CONFIG_SECURITY void *f_security; #endif - /* needed for tty driver, and maybe others */ - void *private_data; #ifdef CONFIG_EPOLL /* Used by fs/eventpoll.c to link all the hooks to this file */