On Tuesday, July 07, 2015 10:55:59 AM Alan Stern wrote: > On Tue, 7 Jul 2015, Rafael J. Wysocki wrote: > > > > > All right, we can make a decision and document it. The following seems > > > > reasonable to me: > > > > > > > > If dev->power.direct_complete is set then the PM core will > > > > assume that dev->power.rpm_status is accurate even when > > > > dev->power.disable_depth > 0. The core will obey the > > > > .direct_complete setting regardless of .disable_depth. > > > > > > > > As a consequence, devices that support system sleep but don't > > > > support runtime PM must _never_ have .direct_complete set. > > > > > > > > On the other hand, if a device (such as a "virtual" device) > > > > requires no callbacks for either system sleep or runtime PM, > > > > then there is no harm in setting .direct_complete. Indeed, > > > > doing so may help speed up an ancestor device's sleep > > > > transition. > > > > > > > > How does that sound? > > > > > > It would be workable I think, but I'd prefer the core to be told directly > > > about devices whose runtime PM status doesn't matter (because nothing > > > changes > > > between "suspended" and "active"), so they may be treated in a special way > > > safely. > > > > > > If we had that information, no special rules other than "that is a device > > > whose runtime PM status doesn't matter, so treat it accordingly" would be > > > necessary. > > > > That said, a situation to consider is when device X is just a software > > device, > > but it has children that correspond to physical hardware. If that is the > > case, > > the usual parent-children rules should apply to X and its children (ie. X > > should > > only be marked as "suspended" if all of its children are suspended) and I > > see > > no reason why the parent-children rules for direct_resume should not apply > > here. > > Yes, this illustrates that in some ways we must not treat "virtual" or > "software" devices specially. Being "virtual" is not the same as > having the ignore_children flag set. > > The change I'm proposing is not related to whether a device is > "virtual". I'm just suggesting that the normal direct_complete rules > should apply even when devices are runtime-PM-disabled. > > This doesn't mean that their runtime PM status doesn't matter. Just > the opposite, in fact -- it means that the PM core should pay attention > to the runtime PM status during a sleep transition even though > disabled_depth > 0.
I seem to have lost the context here, sorry about that. The idea seems to be to rely on the fact that the RPM status for all devices is initially RPM_SUSPENDED and that never changes if runtime PM is never enabled for the device, so in that particular case it would be OK to treat the "power.direct_complete set + RPM status == RPM_SUSPENDED" combination as valid even though runtime PM has never been enabled for the device in question (provided that power.direct_complete will never be set for "real" devices that don't support runtime PM). Is that correct? That seems to be fragile, but I have no strong opinion. Let's do that change if it allows us to make forward progress here. Please feel free to submit a documentation patch along the lines you've suggested. Thanks, Rafael -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
