From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Rafael J. Wysocki" Subject: Re: [PATCH] driver core: Ensure proper suspend/resume ordering Date: Thu, 17 Sep 2015 02:07:48 +0200 Message-ID: <4548634.pC6GYZYGWm@vostro.rjw.lan> References: Mime-Version: 1.0 Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: 7Bit Return-path: Received: from v094114.home.net.pl ([79.96.170.134]:45507 "HELO v094114.home.net.pl" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with SMTP id S1752836AbbIPXjr (ORCPT ); Wed, 16 Sep 2015 19:39:47 -0400 In-Reply-To: Sender: linux-pm-owner@vger.kernel.org List-Id: linux-pm@vger.kernel.org To: Alan Stern Cc: Grygorii Strashko , Thierry Reding , Greg Kroah-Hartman , "Rafael J. Wysocki" , Tomeu Vizoso , linux-pm@vger.kernel.org, linux-kernel@vger.kernel.org On Wednesday, September 16, 2015 03:27:55 PM Alan Stern wrote: > On Wed, 16 Sep 2015, Grygorii Strashko wrote: > > > I think, It should prohibited to probe devices during suspend/hibernation. > > And solution introduced in this patch might help to fix it - > > in general, we could do : > > - add sync point on suspend enter: wait_for_device_probe() and > > - prohibit probing: move all devices which will request probing into > > deferred_probe list > > - one suspend exit: allow probing and do driver_deferred_probe_trigger > > That could work; it's a good idea. > > > I'd like to mention here that this patch will work only > > if dmp_list will be filled according device creation order ("parent<-child" dependencies) > > *AND* according device's probing order ("supplier<-consumer"). > > So, if there is the case when Parent device can be probed AFTER its children > > - it will not work, because "parent<-child" dependencies will not be tracked > > any more :( Sry, I could not even imagine that such crazy case exist :'( > > If we avoid moving devices to the end of the dpm_list when they already > have children, then we should be okay, right? > > > Are there any other subsystems with the same behavior like PCI? > > I don't know. > > > If not - probably, it could be fixed in PCI subsystem using device_pm_move_after() or > > device_move() in PCIe ports probe. > > if yes - ... maybe we can scan/re-check and reorder dpm_list on suspend enter and > > restore ("parent<-child" dependencies). > > > Truth is that smth. need to be done 100%. Personally, I was hit by this issue also, > > and it cost me 3 hours of debugging and I came up with the same patch as > > Bill Huang, then spent some time trying to understand what is wrong with PCI > > - finally, I've just changed the order of my devices in DT :) > > > > Also, I think, it will be good to have this patch in -next to collect more feedbacks. > > I like the idea of forcing all probes during a sleep transition to be > deferred. We could carry them out just before unfreezing the user > threads. That combined with the change mentioned above ought to be > worth testing. Agreed. Thanks, Rafael