On Wednesday, November 8, 2017 1:28:24 AM CET Florian Fainelli wrote: > On 11/07/2017 04:23 PM, Rafael J. Wysocki wrote: > > On Tuesday, November 7, 2017 5:00:06 PM CET Tony Lindgren wrote: > >> * Florian Fainelli <f.faine...@gmail.com> [171104 17:21]: > >>> > >>> > >>> On 11/04/2017 05:25 AM, Rafael J. Wysocki wrote: > >>>> On Friday, November 3, 2017 6:33:53 PM CET Tony Lindgren wrote: > >>>>> * Florian Fainelli <f.faine...@gmail.com> [171103 17:04]: > >>>>>> On 11/03/2017 09:11 AM, Tony Lindgren wrote: > >>>>>> The pinctrl provider is losing its state, hence these two patches. > >>>>> > >>>>> OK > >>>>> > >>>>>>> Anyways, the context lost flag should be managed in the PM core for > >>>>>>> the device, so adding linux-pm and Rafael to Cc. > >>>>>> > >>>>>> I don't think it's that simple but sure, why not. > >>>>> > >>>>> Just having bool context_lost in struct dev_pm_info would probably > >>>>> be enough to allow drivers to deal with it. This flag could then > >>>>> be set for a device by power domain related code that knows if > >>>>> context got lost. > >>>> > >>>> Something like: if the driver sees "context_lost" set, it should restore > >>>> the context to the device from memory? > >>> > >>> That is what is being proposed here, except that the actual mechanism > >>> where this matters needs to be in the core pinctrl code, otherwise the > >>> state (context) is not restored due to a check that attempts not to > >>> (re)apply a previous state. > >>> > >>>> > >>>> But the it would also need to save the context beforehand, so why not to > >>>> restore it unconditionally on resume? > >>> > >>> That's what my original attempts did here: > >>> > >>> https://patchwork.kernel.org/patch/9598969/ > >>> > >>> but Linus rightfully requested this to be done differently, hence this > >>> attempt now to solve it in a slightly more flexible way based on DT > >>> properties. > >> > >> For runtime PM, restoring the state constantly is unnecessary and not > >> good for battery life. The logic can be just: > >> > >> 1. Device driver runtime PM suspend saves the state when needed > >> > >> 2. Device driver runtime PM resume checks if context_lost was set by > >> the bus or power domain code > >> > >> 3. If context was lost, device driver restores the state, or in some > >> cases may need re-run the driver register init related parts > >> to bring the driver back up, then clears the context_lost flag > >> > >> How about something like the following patch? So far only compile > >> tested with CONFIG_PM enabled. If that looks like the way to go, > >> I'll test it properly and add some comments for the functions and > >> post a proper patch :) > > > > Honestly, I'm not sure. > > > > I'd rather have a context_lost flag to start with and see how/if > > drivers will use that before adding any common infra for handling > > this. > > I am afraid we are being slightly side tracked here on this context_loss > flag, because the crux of the problem is not whether a driver knows or > not when it loses state, it's more than the pinctrl core refuses to > re-apply the same state upon resumption even when the consumer driver > tells it to, because it has not seen a transition (consider it as a > stale software cache of the state), and there is only one state defined. > > I don't particularly care how its gets solved, at the generic device > driver model or at the pinctrl level, but I think the pinctrl code needs > to change in that regard no matter what we do, because right now, if you > call pinctrl_select_state() in your driver's resume function, and there > is only one state defined, nothing happens, that's a problem.
I see, thanks for explaining that clearly. Then it looks like that there needs to be a way for the "cached" state to be invalidated and the question is what that way should be and how it is going to be triggered. Thanks, Rafael
