From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Rafael J. Wysocki" Subject: Re: [PATCH 0/12] PM / sleep: Driver flags for system suspend/resume Date: Mon, 16 Oct 2017 23:50:51 +0200 Message-ID: References: <3806130.B2KCK0tvef@aspire.rjw.lan> <20171016070850.GA26934@kroah.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Return-path: Received: from mail-oi0-f44.google.com ([209.85.218.44]:53697 "EHLO mail-oi0-f44.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932309AbdJPVuw (ORCPT ); Mon, 16 Oct 2017 17:50:52 -0400 In-Reply-To: <20171016070850.GA26934@kroah.com> Sender: linux-acpi-owner@vger.kernel.org List-Id: linux-acpi@vger.kernel.org To: Greg Kroah-Hartman Cc: "Rafael J. Wysocki" , Linux PM , Bjorn Helgaas , Alan Stern , LKML , Linux ACPI , Linux PCI , Linux Documentation , Mika Westerberg , Ulf Hansson , Andy Shevchenko , Kevin Hilman , Wolfram Sang , linux-i2c , Lee Jones On Mon, Oct 16, 2017 at 9:08 AM, Greg Kroah-Hartman wrote: > On Mon, Oct 16, 2017 at 03:12:35AM +0200, Rafael J. Wysocki wrote: >> Hi All, >> >> Well, this took more time than expected, as I tried to cover everything I had >> in mind regarding PM flags for drivers. >> >> This work was triggered by attempts to fix and optimize PM in the >> i2c-designware-platdev driver that ended up with adding a couple of >> flags to the driver's internal data structures for the tracking of >> device state (https://marc.info/?l=linux-acpi&m=150629646805636&w=2). >> That approach is sort of suboptimal, though, because other drivers will >> probably want to do similar things and if all of them need to use internal >> flags for that, quite a bit of code duplication may ensue at least. >> >> That can be avoided in a couple of ways and one of them is to provide a means >> for drivers to tell the core what to do and to make the core take care of it >> if told to do so. Hence, the idea to use driver flags for system-wide PM >> that was briefly discussed during the LPC in LA last month. >> >> One of the flags considered at that time was to possibly cause the core >> to reuse the runtime PM callback path of a device for system suspend/resume. >> Admittedly, that idea didn't look too bad to me until I had started to try to >> implement it and I got to the PCI bus type's hibernation callbacks. Then, I >> moved the patch I was working on to /dev/null right away. I mean it. >> >> No, this is not going to happen. No way. >> >> Moreover, that experience made me realize that the whole *idea* of using the >> runtime PM callback path for system-wide PM was actually totally bogus (sorry >> Ulf). >> >> The whole point of having different callbacks pointers for different types of >> device transitions is because it may be necessary to do different things in >> those callbacks in general. Now, if you consider runtime PM and system >> suspend/resume *only* and from a driver perspective, then yes, in some cases >> the same pair of callback routines may be used for all suspend-like and >> resume-like transitions of the device, but if you add hibernation to the mix, >> then it is not so clear any more unless the callbacks don't actually do any >> power management at all, but simply quiesce the device's activity and then >> activate it again. Namely, changing power states of devices during the >> hibernation's "freeze" and "thaw" transitions rarely makes sense at all and >> the "restore" transition needs to be able to cope with uninitialized devices >> (in fact, it should be prepared to cope with devices in *any* state), so >> runtime PM is hardly suitable for them. Still, if a *driver* choses to not >> do any real PM in its PM callbacks and leaves that to a middle layer (quite >> a few drivers do that), then it possibly can use one pair of callbacks in all >> cases and be happy, but middle layers pretty much have to use different >> callback routines for different transitions. >> >> If you are a middle layer, your role is basically to do PM for a certain >> group of devices. Thus you cannot really do the same in ->suspend or >> ->suspend_early and in ->runtime_suspend (because the former generally need to >> take device_may_wakeup() into account and the latter doesn't) and you shouldn't >> really do the same in ->suspend and ->freeze (becuase the latter shouldn't >> change the device's power state) and so on. To put it bluntly, trying >> to use the ->runtime_suspend callback of a middle layer for anything other >> than runtime suspend is complete and utter nonsense. At the same time, the >> ->runtime_resume callback of a middle layer may be reused to some extent, >> but even that doesn't cover the "thaw" transitions during hibernation. >> >> What can work (and this is the only strategy that can work AFAICS) is to >> point different callback pointers *in* *a* *driver* to the same routine >> if the driver wants to reuse that code. That actually will work for PCI >> and USB drivers today, at least most of the time, but unfortunately there >> are problems with it for, say, platform devices. >> >> The first problem is the requirement to track the status of the device >> (suspended vs not suspended) in the callbacks, because the system-wide PM >> code in the PM core doesn't do that. The runtime PM framework does it, so >> this means adding some extra code which isn't necessary for runtime PM to >> the callback routines and that is not particularly nice. >> >> The second problem is that, if the driver wants to do anything in its >> ->suspend callback, it generally has to prevent runtime suspend of the >> device from taking place in parallel with that, which is quite cumbersome. >> Usually, that is taken care of by resuming the device from runtime suspend >> upfront, but generally doing that is wasteful (there may be no real need to >> resume the device except for the fact that the code is designed this way). >> >> On top of the above, there are optimizations to be made, like leaving certain >> devices in suspend after system resume to avoid wasting time on waiting for >> them to resume before user space can run again and similar. >> >> This patch series focuses on addressing those problems so as to make it >> easier to reuse callback routines by pointing different callback pointers >> to them in device drivers. The flags introduced here are to instruct the >> PM core and middle layers (whatever they are) on how the driver wants the >> device to be handled and then the driver has to provide callbacks to match >> these instructions and the rest should be taken care of by the code above it. >> >> The flags are introduced one by one to avoid making too many changes in >> one go and to allow things to be explained better (hopefully). They mostly >> are mutually independent with some clearly documented exceptions. >> >> The first three patches in the series are about an issue with the >> direct-complete optimization introduced some time ago in which some middle >> layers decide on whether or not to do the optimization without asking the >> drivers. And, as it turns out, in some cases the drivers actually know >> better, so the new flags introduced by these patches are here for these >> drivers (and the DPM_FLAG_NEVER_SKIP one is really to avoid having to define >> ->prepare callbacks always returning zero). >> >> The really interesting things start to happen in patches [4-9/12] which make it >> possible to avoid resuming devices from runtime suspend upfront during system >> suspend at least in some cases (and when direct-complete is not applied to the >> devices in question), but please refer to the changelogs for details. >> >> The i2d-designware-platdev driver is used as the primary example in the series >> and the patches modifying it are based on some previous changes currently in >> linux-next AFAICS (the same applies to the intel-lpss driver), but these >> patches can wait until everything is properly merged. They are included here >> mostly as illustration. >> >> Overall, the series is based on the linux-next branch of the linux-pm.git tree >> with some extra patches on top of it and all of the names of new entities >> introduced in it are negotiable. > > Thanks for the great explaination, I was wondering how your proposal > discussed at Plumbers was going to work out in the end :) > > The patch series looks good to me (minor questions already sent on the > patches), Cool. :-) > but what does this mean for drivers? Do they now have to do a > lot of work to take advantage of this, like you did for the > i2d-designware-platdev driver? Or will things continue to work as-is > and it's only an opt-in type thing where the bus/driver wants to take > advantage of it? It's envisioned as an opt-in thing mostly, except for the flags introduced by patch [01/12] that may be needed to address existing issues. It is not strictly necessary to set any of the other flags, but I guess some use cases may benefit quite a bit from setting them. :-) Thanks, Rafael