From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755068AbXDZUzx (ORCPT ); Thu, 26 Apr 2007 16:55:53 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1755067AbXDZUzx (ORCPT ); Thu, 26 Apr 2007 16:55:53 -0400 Received: from nigel.suspend2.net ([203.171.70.205]:49694 "EHLO nigel.suspend2.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1755058AbXDZUzv (ORCPT ); Thu, 26 Apr 2007 16:55:51 -0400 Subject: Re: suspend2 merge (was Re: [Suspend2-devel] Re: CFS and suspend2: hang in atomic copy) From: Nigel Cunningham Reply-To: nigel@nigel.suspend2.net To: "Rafael J. Wysocki" Cc: Pekka Enberg , Pavel Machek , Dumitru Ciobarcianu , Ingo Molnar , Linus Torvalds , Christian Hesse , Nick Piggin , Mike Galbraith , "linux-kernel@vger.kernel.org" , Con Kolivas , "suspend2-devel@lists.suspend2.net" , Andrew Morton , Thomas Gleixner , Arjan van de Ven In-Reply-To: <200704262237.32665.rjw@sisk.pl> References: <200704262128.31547.rjw@sisk.pl> <1177618612.4737.50.camel@nigel.suspend2.net> <200704262237.32665.rjw@sisk.pl> Content-Type: multipart/signed; micalg=pgp-sha1; protocol="application/pgp-signature"; boundary="=-q/PCwF43u5dJO6x1mOU3" Date: Fri, 27 Apr 2007 06:55:49 +1000 Message-Id: <1177620949.4737.61.camel@nigel.suspend2.net> Mime-Version: 1.0 X-Mailer: Evolution 2.10.1 Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org --=-q/PCwF43u5dJO6x1mOU3 Content-Type: text/plain Content-Transfer-Encoding: quoted-printable Hi. On Thu, 2007-04-26 at 22:37 +0200, Rafael J. Wysocki wrote: > On Thursday, 26 April 2007 22:16, Nigel Cunningham wrote: > > Hi. > >=20 > > On Thu, 2007-04-26 at 21:28 +0200, Rafael J. Wysocki wrote: > > > On Thursday, 26 April 2007 18:10, Pekka Enberg wrote: > > > >=20 > > > > On 4/26/2007, "Rafael J. Wysocki" wrote: > > > > > In principle, we could add suspend2 as an alternative (in analogy= with the I/O > > > > > schedulers, for example), but I think for this purpose it should = be reviewed > > > > > properly. > > > >=20 > > > > Yeah, this makes sense. > > > >=20 > > > > On 4/26/2007, "Rafael J. Wysocki" wrote: > > > > > There also is a real problem with how it uses the LRU pages. It = _seems_ to > > > > > work, but at least to me it seems to be potentially dangerous. > > > >=20 > > > > I am new to suspend2 so can you please explain what exactly is dang= erous > > > > about it? > > >=20 > > > After freezing tasks, it first saves the contents of the LRU pages, f= reezes > > > devices and then uses the LRU pages for storing the suspend image (if= more > > > memory is needed, it's allocated, but that's irrelevant here). Now, = we have no > > > warranty that the LRU pages are not updated after we've saved their c= ontents > > > (first potential problem here). > > >=20 > > > After the image has been created, we have to unfreeze devices and sav= e the > > > image. Now, we have no warranty that no one will be writing to the L= RU pages > > > that we have used to store the image, for whatever reasons known to h= im, so the > > > image can potentially get corrupted while it's being saved. > > >=20 > > > In principle, device drivers can do this and there are some kernel th= reads that > > > also can do this (we don't freeze them, because they're needed for th= e image > > > saving). > > >=20 > > > The design is conceptually really really complicated and it makes str= ong > > > assumptions about the behavior of different subsystems. While these > > > assumptions _may_ be satisfied right now, we'd have to ensure the sat= isfaction > > > of them in the future if suspend2 were merged. > >=20 > > That's a good description of the issue, although I think _may_ and > > _seems_ are stating things a bit more pessimistically than is > > necessary.=20 >=20 > I've used them to express my personal concerns. >=20 > > You see, we need to remember that the pages which are saved separately > > are LRU pages. Because userspace is frozen, their contents are going to > > be static. The only possibilities for modifying them come from timer > > routines, improperly frozen filesystems and device drivers. >=20 > And kernel threads that we don't freeze deliberately. Currently, these a= re > all worker threads, dm-related kernel threads and some others. >=20 > > We have code to check that the LRU isn't changing, and I've only seen > > one report of modifications to about 20 LRU pages. I haven't had the > > time yet to chase down the cause, but hope to do so soon. >=20 > I didn't say that would be common. If it had been, you'd have seen probl= ems > with it. To me the problem is the lack of warranty that it won't happen. >=20 > > The general scheme has been working for four or five years - if there > > was a fundamental issue, we would have found it by now. > >=20 > > The scheme isn't complicated. >=20 > Conceptually, it is complicated just because you're using the LRU. Well, I'm willing to look at other ideas. I actually selected LRU because it was simple. Prior to this, we did have a try at just iterating over the pages of frozen processes, but it didn't yield enough pages to be viable. I wouldn't be surprised if hunting down the cause of these changing pages leads to doing the opposite - starting with LRU pages and then removing all the ones owned by processes. (Am I right in thinking the remainder would be anonymous pages? I must learn more mm inards :>). Nigel --=-q/PCwF43u5dJO6x1mOU3 Content-Type: application/pgp-signature; name=signature.asc Content-Description: This is a digitally signed message part -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.6 (GNU/Linux) iD8DBQBGMRHVN0y+n1M3mo0RAtNiAKCic5vE3WXbxmpfgmaxZxaKz38VhACguiDv T7nl/BC7GAM1fxHYXlAajCs= =jZB8 -----END PGP SIGNATURE----- --=-q/PCwF43u5dJO6x1mOU3--