On Fri, Nov 29, 2019 at 11:22:08AM +0100, Rafael J. Wysocki wrote: > On Fri, Nov 29, 2019 at 11:08 AM Thierry Reding > wrote: > > > > On Thu, Nov 28, 2019 at 11:20:01PM +0100, Rafael J. Wysocki wrote: > > > On Thursday, November 28, 2019 11:03:57 PM CET Rafael J. Wysocki wrote: > > > > On Thursday, November 28, 2019 5:50:26 PM CET Thierry Reding wrote: > > > > > > > > > > --0F1p//8PRICkK4MW > > > > > Content-Type: text/plain; charset=us-ascii > > > > > Content-Disposition: inline > > > > > Content-Transfer-Encoding: quoted-printable > > > > > > > > > > On Thu, Nov 28, 2019 at 05:14:51PM +0100, Rafael J. Wysocki wrote: > > > > > > On Thu, Nov 28, 2019 at 5:03 PM Thierry Reding = > > > > > wrote: > > > > > > > > > > > > > > From: Thierry Reding > > > > > > > > > > > > > > Currently the driver PM core will automatically acquire a runtime PM > > > > > > > reference for devices before system sleep is entered. This is needed > > > > > > > to avoid potential issues related to devices' parents getting put to > > > > > > > runtime suspend at the wrong time and causing problems with their > > > > > > > children. > > > > > >=20 > > > > > > Not only for that. > > > > > >=20 > > > > > > > In some cases drivers are carefully written to avoid such issues and > > > > > > > the default behaviour can be changed to allow runtime PM to operate > > > > > > > regularly during system sleep. > > > > > >=20 > > > > > > But this change breaks quite a few assumptions in the core too, so no, > > > > > > it can't be made. > > > > > > > > > > Anything in particular that I can look at? I'm not seeing any issues > > > > > when I test this, which could of course mean that I'm just getting > > > > > lucky. > > > > > > > > There are races and such that you may never hit during casual testing. > > > > > > > > > One thing that irritated me is that I think this used to work. I do > > > > > recall testing suspend/resume a few years ago and devices would get > > > > > properly runtime suspended/resumed. > > > > > > > > Not true at all. > > > > > > > > The PM core has always taken PM-runtime references on all devices pretty much > > > > since when PM-runtime was introduced. > > > > > > > > > I did some digging but couldn't > > > > > find anything that would have had an impact on this. > > > > > > > > > > Given that this is completely opt-in feature, why are you categorically > > > > > NAK'ing this? > > > > > > > > The general problem is that if any device has been touched by system-wide > > > > suspend code, it should not be subject to PM-runtime any more until the > > > > subsequent system-wide resume is able to undo whatever the suspend did. > > > > > > > > Moreover, if a device is runtime-suspended, the system-wide suspend code > > > > may mishandle it, in general. That's why PM-runtime suspend is not allowed > > > > during system-wide transitions at all. And it has always been like that. > > > > > > > > For a specific platform you may be able to overcome these limitations if > > > > you are careful enough, but certainly they are there in general and surely > > > > you cannot prevent people from using your opt-in just because they think > > > > that they know what they are doing. > > > > > > BTW, what if user space prevents PM-runtime from suspending devices by writing > > > "on" to their "control" files? > > > > > > System-wide suspend is (of course) still expected to work in that case, so how > > > exactly would you overcome that? > > > > I suppose one way to overcome that would be to make it an error to write > > "on" to the "control" files for these devices. > > Seeing suggestions like this in messages from seasoned kernel > developers is seriously disappointing. :-/ > > > Currently doing this is likely going to break display support on Tegra, > > so this would be a good idea in this case anyway. > > PM-runtime has always allowed user space to prevent devices from being > suspended and it seems that this has not been taken into account by > Tegra display support developers at all. > > > Again, I could avoid all of these issues by avoiding runtime PM in this driver, > > I don't quite see the connection here. > > Preventing a device from suspending should never be a functional > problem. It may be an energy-efficiency problem, but that's something > for user space to consider before writing "on" to a device's control > file. That's really a question of how you define suspension. In the case of display drivers we have the somewhat unfortunate situation that in most SoCs the display "device" is actually represented by a collection of different devices. On Tegra specifically, for example, you have a couple of display controllers, then some "encoders" that take pixel streams from the display controllers and encode them into some wire format like LVDS, HDMI, DSI or DP. Prohibiting suspension of any of the individual devices causes problems because it effectively makes the whole composite display device not suspendable. Doing so in turn usually means that you can't change the display configuration anymore because devices need to be powered up and down in order to change the configuration. I consider powering up and down the devices a form of suspension. Hence it seemed natural to implement using runtime PM. It sounds to me like userspace preventing runtime PM is problematic in most scenarios that involve composite devices because it makes all of the interactions between the devices a bit complicated. > > but I would end up reimplementing some of the same concepts. I'd > > rather use something that's supported by the PM core and that might be > > useful to other drivers than reinvent the wheel. > > Which doesn't have to be by using PM-runtime suspend for the handling > of system-wide suspend, at least in my view. Well, runtime PM is very convenient for this, though. It would allow the same code paths to be used in all cases. Thierry