From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755573Ab1FTVAG (ORCPT ); Mon, 20 Jun 2011 17:00:06 -0400 Received: from iolanthe.rowland.org ([192.131.102.54]:40893 "HELO iolanthe.rowland.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with SMTP id S1755307Ab1FTVAD (ORCPT ); Mon, 20 Jun 2011 17:00:03 -0400 Date: Mon, 20 Jun 2011 17:00:03 -0400 (EDT) From: Alan Stern X-X-Sender: stern@iolanthe.rowland.org To: "Rafael J. Wysocki" cc: Linux PM mailing list , LKML , Jesse Barnes , "linux-pci@vger.kernel.org" Subject: Re: [PATCH] PCI / PM: Block races between runtime PM and system sleep In-Reply-To: <201106202142.45802.rjw@sisk.pl> Message-ID: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Mon, 20 Jun 2011, Rafael J. Wysocki wrote: > > Furthermore, since we're going to disable runtime PM as soon as the > > suspend callback returns anyway, why not increment usage_count before > > invoking the callback? This will prevent runtime suspends from > > occurring while the callback runs, so no changes will be needed in the > > PCI or USB subsystems. > > The PCI case is different from the USB one. PCI needs to resume devices > before calling their drivers' .suspend() callbacks, so it does that in > .prepare(). If the core acquired a reference to every device before executing > the subsystem .suspend(), then pm_runtime_resume() could be moved from > pci_pm_prepare() to pci_prepare_suspend(), but then additionally it would > have to be called from pci_pm_freeze() and pci_pm_poweroff(). It simply is > more efficient to call it once from pci_pm_prepare(), but then PCI needs to > take the reference by itself. Ah, okay. The PCI part makes sense then. > > It also will prevent Kevin from calling pm_runtime_suspend from within > > his suspend callbacks, but you have already determined that subsystems > > and drivers should never do that in any case. > > Then reverting commit e8665002477f0278f84f898145b1f141ba26ee26 would be > even better. :-) See below. > > As I see it, we never want a suspend or suspend_noirq callback to call > > pm_runtime_suspend(). However it's okay for the suspend callback to > > invoke pm_runtime_resume(), as long as this is all done in subsystem > > code. > > First off, I don't really see a reason for a subsystem to call > pm_runtime_resume() from its .suspend_noirq() callback. I was referring to .suspend(), not .suspend_noirq(). > Now, if > pm_runtime_resume() is to be called concurrently with the subsystem's > .suspend_noirq() callback, I'd rather won't let that happen. :-) Me too. But I see no reason to prevent pm_runtime_resume() from being called by .suspend(). > > And in between the prepare and suspend callbacks, runtime PM should be > > more or less fully functional, right? For most devices it will never > > be triggered, because it has to run in process context and both > > userspace and pm_wq are frozen. It may trigger for devices marked as > > IRQ-safe, though. > > It also may trigger for drivers using non-freezable workqueues and calling > runtime PM synchronously from there. Right. So we shouldn't ignore this window. > > Maybe the barrier should be moved into __device_suspend(). > > I _really_ think that the initial approach, i.e. before commit > e8665002477f0278f84f898145b1f141ba26ee26, made the most sense. It didn't > cover the "pm_runtime_resume() called during system suspend" case, but > it did cover everything else. But it prevented runtime PM from working during the window between .prepare() and .suspend(), and also between .resume() and .complete(). If a subsystem like PCI wants to rule out runtime PM during those windows, then fine -- it can do whatever it wants. But the PM core shouldn't do this. > So, I think there are serious technical arguments for reverting that commit. > > I think we went really far trying to avoid that, but I'm not sure I want to go > any further. What I'm suggesting is to revert the commit but at the same time, move the get_noresume() into __device_suspend() and the put_sync() into device_resume(). Alan Stern