On 14 November 2013 16:59, Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote: > On Thu, 14 Nov 2013, Ulf Hansson wrote: > >> > The reason for doing things the way they are is to avoid races between >> > system PM callbacks and runtime PM callbacks. We don't want to have a >> > runtime_resume routine powering up a device at the same time as a >> > suspend routine is powering it down. >> >> I believe I understand why the PM core prevent this. Some subsystems >> and drivers depends on this behaviour, but I wonder if we should >> consider this as common cases or more like special cases? >> >> For those drivers that potentially could rely on runtime PM (and it's >> callbacks) to inactivate the device during system suspend, and thus >> don't have issues with races you describe above, the current behaviour >> in the PM core adds unnecessary complexity for these drivers to handle >> system suspend. > > What unnecessary complexity? > > Bear in mind that drivers _cannot_ rely on runtime PM to inactivate a > device during system suspend. The user can always prevent a device > from going into runtime suspend by writing "on" to the > /sys/.../power/control file. Good that you brought this up. From my point of view I think the sysfs for runtime PM could be debated whether it should exist at all, at least in it's current form. Anyway, if userspace decides to prevent runtime_suspend, I guess it will have take the consequences for it as well. :-) > >> Drivers/buses/power_domains that handles runtime PM enabled devices, >> will, as you also have pointed out earlier, need to implement the >> .suspend callback (suspend_late or suspend_noirq might work as well) >> and the corresponding .resume callback. This just for the sake of >> making sure all their runtime enabled devices gets fully inactivated >> during system suspend. >> >> I wonder if we somehow should move the responsibility of preventing >> "runtime suspend" as a part of the system suspend, down to the driver >> to decide instead of the PM core? > > No, I don't think so. It would mean that every driver's suspend > routine would have to do a pm_runtime_get of some sort. It's better to > just leave the single runtime_get call in the PM core. In the end I suppose what case that is most common should win. :-) At the same time, we should go slowly forward and it is probably not realistic to convert each and every driver in one go. So as a way forward, I am thinking of a similar approach as you suggested with the "generic" suspend_late. But instead add a new runtime PM API, which intent is to let drivers to specify for PM core, if it should care to prevent the runtime suspend from happen during system system - or not. Could that work? Kind regards Ulf Hansson > > Alan Stern > -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
