From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1756564Ab0EXHoI (ORCPT ); Mon, 24 May 2010 03:44:08 -0400 Received: from cantor.suse.de ([195.135.220.2]:38667 "EHLO mx1.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753014Ab0EXHoF (ORCPT ); Mon, 24 May 2010 03:44:05 -0400 Date: Mon, 24 May 2010 17:43:36 +1000 From: Nick Piggin To: Andrew Morton Cc: Manfred Spraul , linux-kernel@vger.kernel.org Subject: Re: [patch 3/3] ipc: increase IPCMNI_MAX Message-ID: <20100524074335.GX2516@laptop> References: <20100520065911.GG2516@laptop> <20100520070741.GI2516@laptop> <20100521133136.d0e282df.akpm@linux-foundation.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20100521133136.d0e282df.akpm@linux-foundation.org> User-Agent: Mutt/1.5.20 (2009-06-14) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Fri, May 21, 2010 at 01:31:36PM -0700, Andrew Morton wrote: > On Thu, 20 May 2010 17:07:41 +1000 > Nick Piggin wrote: > > > Just wondering whether there is a good reason to have a full 16 bits of > > sequence in ipc ids? 32K indexes is pretty easy to overflow, if only in > > stress tests for now. I was doing some big aim7 stress testing, which > > required this patch, but it's not exactly a realistic workload :) > > > > But the sequence seems like it just helps slightly with buggy apps, and > > if the app is buggy then it can by definition mess up its own ids > > anyway? So I don't see that such amount of seq is required. > > > > Index: linux-2.6/ipc/util.h > > =================================================================== > > --- linux-2.6.orig/ipc/util.h > > +++ linux-2.6/ipc/util.h > > @@ -14,7 +14,16 @@ > > #include > > > > /* IPCMNI_MAX should be <= MAX_INT, absolute limit for ipc arrays */ > > -#define IPCMNI_MAX_SHIFT 15 > > +/* > > + * IPC ids consist of an index into the idr, which allocates from the bottom > > + * up, and a sequence number which is continually incremented. Valid indexes > > + * are from 0..IPCMNI_MAX (or further constrained by sysctls or other limits). > > + * The sequence number prevents ids from being reused quickly. The sequence > > + * number resides in the top part of the 'int' after IPCMNI_MAX. > > + * > > + * Increasing IPCMNI_MAX reduces the sequence wrap interval. > > + */ > > +#define IPCMNI_MAX_SHIFT 20 > > #define IPCMNI_MAX (1 << IPCMNI_MAX_SHIFT) > > > > #define SEQ_SHIFT IPCMNI_MAX_SHIFT > > Some anaylsis of the worst-case memory consumption would be mollifying. OK. > I took the absence of Signed-off-by:'s to mean "rfc" and wandered away. Yes. They should probably go through Manfred to you anyway.