From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932458AbXCGQbQ (ORCPT ); Wed, 7 Mar 2007 11:31:16 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S932744AbXCGQbQ (ORCPT ); Wed, 7 Mar 2007 11:31:16 -0500 Received: from mga03.intel.com ([143.182.124.21]:24772 "EHLO mga03.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932458AbXCGQbP (ORCPT ); Wed, 7 Mar 2007 11:31:15 -0500 X-ExtLoop1: 1 X-IronPort-AV: i="4.14,261,1170662400"; d="scan'208"; a="191740398:sNHT23577358" Message-ID: <45EEE8CF.1060803@intel.com> Date: Wed, 07 Mar 2007 08:31:11 -0800 From: "Kok, Auke" User-Agent: Mail/News 1.5.0.9 (X11/20061228) MIME-Version: 1.0 To: "Eric W. Biederman" CC: "Kok, Auke" , Ingo Molnar , Jeff Garzik , Linus Torvalds , "Michael S. Tsirkin" , Pavel Machek , Jens Axboe , Adrian Bunk , Andrew Morton , Linux Kernel Mailing List , Thomas Gleixner , linux-pm@lists.osdl.org, Michal Piotrowski Subject: Re: SATA resume slowness, e1000 MSI warning References: <20070227103021.GA2250@kernel.dk> <20070227103407.GA17819@elte.hu> <20070227105922.GD2250@kernel.dk> <20070227111515.GA4271@kernel.dk> <20070301093450.GA8508@elte.hu> <20070302100704.GB2293@elf.ucw.cz> <20070305084257.GA4464@mellanox.co.il> <20070305101120.GA23032@elte.hu> <45ECFC5F.7000102@garzik.org> <45ED0BBF.1050000@intel.com> <20070306090444.GA25409@elte.hu> <45ED8A12.5040803@intel.com> In-Reply-To: Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit X-OriginalArrivalTime: 07 Mar 2007 16:31:11.0777 (UTC) FILETIME=[047AC110:01C760D6] Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org Eric W. Biederman wrote: > "Kok, Auke" writes: > >> Ingo Molnar wrote: >>> * Kok, Auke wrote: >>> >>>>>> BUG: at drivers/pci/msi.c:611 pci_enable_msi() >>>>> I would poke Eric Biederman(sp?) about this one. Maybe its even solved by >>>>> the MSI-enable-related patch he posted in the past 24-48 hours. >>>> I tried the 3-patch series "[PATCH 0/3] Basic msi bug fixes.." and they fix >>>> this problem for me. Were you expecting the OOPS in the first place? [...] >>> the bug was the warning message (a WARN_ON()) above - not an oops. So that >>> warning message is gone in your testing? >> yes. > > Sorry for the slow delay. I was out of town for my brothers wedding the last few > days. > > I wasn't exactly expecting the WARN_ON to trigger. What I fixed was > an inconsistency in handling our state bits. Fixing that > inconsistency appears to have fixed the e1000 usage scenario mostly by > accident. > > The basic issue is that pci_save_state saves the current msi state > along with other registers, and then the e1000 driver goes and > disables the msi irq after we have saved the irq state as on. > > My code notices that the msi irq was disabled before restore time, so > it skips the restore. However we now have a leak of the msi saved cap > because we are not freeing it. > > This leaves with some basic questions. > - Does it make sense for suspend/resume methods to request/free irqs? > - Does it make sense for suspend/resume methods to allocate/free msi irqs? > - Do we want pci_save/restore_cap to save/restore msi state? > > The path of least resistance is to just free the extra state and we > are good. I'm just not quite certain that is sane and it has been a > long day. we used to have a lengthy e1000_pci_save|restore_state in our code, which is now gone, so I'm all for that. A separate pci_save_pxie|msi(x)_state for every driver seems completely unnecessary. I can't think of a use case where saving+restoring everything hurts. That's what you want I presume. We currently free all irq's and msi before going into suspend in e1000, and I think that is probably a good thing, somehow I can think of bad things happening if we dont, but I admit that I haven't tried it without alloc/free. We do this in e100 as well and it works. Another motivation would be to leave this up to the driver: if the driver chooses to free/alloc interrupts because it makes sense, you probably would want to keep that choice available. Devices that don't need this can skip the alloc/free, but leave the choice open for others. hth Auke