From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752138AbXB1LOK (ORCPT ); Wed, 28 Feb 2007 06:14:10 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1752143AbXB1LOK (ORCPT ); Wed, 28 Feb 2007 06:14:10 -0500 Received: from pfx2.jmh.fr ([194.153.89.55]:40118 "EHLO pfx2.jmh.fr" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752138AbXB1LOJ (ORCPT ); Wed, 28 Feb 2007 06:14:09 -0500 From: Eric Dumazet To: Davide Libenzi Subject: Re: [patch] epoll reduced (to 1) number of passes over the ready set ... Date: Wed, 28 Feb 2007 12:13:59 +0100 User-Agent: KMail/1.9.5 Cc: Linux Kernel Mailing List , Linus Torvalds , Andrew Morton References: <200702271015.33123.dada1@cosmosbay.com> In-Reply-To: MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: 7bit Content-Disposition: inline Message-Id: <200702281213.59695.dada1@cosmosbay.com> Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Tuesday 27 February 2007 17:03, Davide Libenzi wrote: > On Tue, 27 Feb 2007, Eric Dumazet wrote: > > On Tuesday 27 February 2007 03:32, Davide Libenzi wrote: > > > Epoll is doing multiple passes over the ready set at the moment, > > > because of the constraints over the f_op->poll() call. Looking at the > > > code again, I noticed that we already hold the epoll semaphore in read, > > > and this (together with other locking conditions that hold while doing > > > an epoll_wait()) can lead to a smarter way to "ship" events to > > > userspace (in a single pass). I added more (even) more comments to the > > > code to explain the conditions why certain operations are safe. > > > This is a stress application that can be used to test the new code. It > > > spwans multiple thread and call epoll_wait() and epoll_ctl() from many > > > threads. Stress tested on my dual Opteron 254 w/out any problems. > > > > Davide, > > > > This is really cool, because the size of epitem would fit now in 128 > > bytes instead of 192 (on x86_64 platforms). So we also reduce memory > > usage. > > Yeah, I forgot to mention that I removed the txlink member. I am pretty sure you can also remove revents member from epitem. It would greatly benefit to 32bits platforms, because resulting size would be 64 bytes instead of 68 (so a 50 % reduction because of 64 bytes alignment) Eric