Re: [PATCH 7/10 v6] PM / Domains: Don't stop wakeup devices during system sleep transitions

[Rafael J. Wysocki]

On Friday, July 01, 2011, Kevin Hilman wrote:
> "Rafael J. Wysocki" <rjw@xxxxxxx> writes:
> 
> > On Friday, July 01, 2011, Kevin Hilman wrote:
> >> "Rafael J. Wysocki" <rjw@xxxxxxx> writes:
> >> 
> >> > On Thursday, June 30, 2011, Kevin Hilman wrote:
> >> >> "Rafael J. Wysocki" <rjw@xxxxxxx> writes:
> >> >> 
> >> >> > From: Rafael J. Wysocki <rjw@xxxxxxx>
> >> >> >
> >> >> > Devices that are set up to wake up the system from sleep states
> >> >> > should not be stopped and power should not be removed from them
> >> >> > when the system goes into a sleep state.  
> >> >> 
> >> >> I don't think this belongs in the generic layer since the two
> >> >> assumptions above are not generally true on embedded systems, and would
> >> >> result in rather significant power consumption unnecessarily.
> >> >
> >> > As to whether or not this belongs to the generic layer, I don't quite agree
> >> > (see below), but the changelog seems to be a bit inaccurate.
> >> >
> >> >> First, whether the device should be stopped on device_may_wakeup():
> >> >> b
> >> >> Some IP blocks (at least on OMAP) have "asynchronous" wakeups.  Meaning
> >> >> that they can generate wakeups even when they're not clocked (a.k.a
> >> >> stopped).  So in this case, even after a ->stop_device (which clock
> >> >> gates the IP), it can still generate wakeups.
> >> >> 
> >> >> Second, whether the device should be powered off if device_may_wakeup():
> >> >> 
> >> >> Embedded SoCs have other ways to wakeup than device-level wakeups.
> >> >> 
> >> >> For example, on OMAP, every pad on the SoC can be configured as a wakeup
> >> >> source So, for example, you could completely power down the UART IP
> >> >> blocks (and the enclosing power domain), configure the UART RX pad as a
> >> >> wakeup source, and still wakeup the system on UART activity.  The OMAP
> >> >> docs call these IO pad wakeups.
> >> >> 
> >> >> On OMAP in fact, this is the common, default behavior when we enable
> >> >> "off-mode" in idle and/or suspend, since most of the IPs are powered off
> >> >> but can still wake up the system.
> >> >> 
> >> >> So in summary, even if device_may_wakeup() is true, many devices (with
> >> >> additional SoC magic) can still generate wakeups even when stopped and
> >> >> powered off.
> >> >
> >> > Well, on the other hand, on some SoCs there are devices that can't be
> >> > powered off (or "declocked") if they are supposed to generate wakeups.
> >> 
> >> Correct.
> >> 
> >> > Also, I'm sure there are cases in which wakeups can be generated for devices
> >> > with their clocks off, but only if power is present.  
> >> 
> >> Yes.
> >> 
> >> > So there are multiple
> >> > cases, but not so many overall.  So, IMO, it makes sense to handle that at
> >> > the generic level, although not necessarily in such a simplistic way.
> >> >
> >> > Now, at this point, I want to do something very simple, which I think is
> >> > done by this patch. 
> >> >
> >> > Is this optimal power comsumption-wise for every potential
> >> > user of the framework?  
> >> 
> >> Well, sub-optimal would be an understatement.  I would consider this a
> >> major regression since if we were to use this for OMAP, we would never
> >> hit the full-chip low-power states if *any* device had wakeups enabled,
> >> whereas today we can.
> >> 
> >> > No, but certainly for some it's sufficient.  Is it
> >> > going to work in general?  I think it is.
> >> >
> >> > Of course, there's the question how to handle that more accurately and I have
> >> > some ideas.  If you have any, please let me know.
> >> >
> >> > In the meantime, I'm going to modify the changelog so that it's clear that
> >> > it's a "first approximation" thing, like in the patch below.
> >> >
> >> > Thanks,
> >> > Rafael
> >> >
> >> > ---
> >> > From: Rafael J. Wysocki <rjw@xxxxxxx>
> >> > Subject: PM / Domains: Don't stop wakeup devices during system sleep transitions
> >> >
> >> > There is the problem how to handle devices set up to wake up the
> >> > system from sleep states during system-wide power transitions.
> >> > In some cases, those devices can be turned off entirely, because the
> >> > wakeup signals will be generated on their behalf anyway.  In some
> >> > other cases, they will generate wakeup signals if their clocks are
> >> > stopped, but only if power is not removed from them.  Finally, in
> >> > some cases, they can only generate wakeup signals if power is not
> >> > removed from them and their clocks are enabled.
> >> 
> >> That's a good summary.
> >> 
> >> > In the future, it will be necessary to take all of the above
> >> > situations into account, but for starters it is possible to use
> >> > the observation that if all wakeup devices are treated like the
> >> > last group (i.e. their clocks are enabled and power in not removed
> >> > from them during system suspend transitions), they all will be able
> >> > to generate wakeups, although power consumption in the resulting
> >> > system sleep state may not be optimal in some cases.
> >> 
> >> I'm not opposed to this kind of check happening.  I'm only opposed to it
> >> happening in this "generic" layer because..., well, it's not generic.
> >> 
> >> Not only is it not generic, it would be a major regression in power
> >> consumption for anyone moving to this layer that has the various
> >> different wakeup capabilities already described.
> >> 
> >> The decision of whether or not to clock gate and/or power gate based on
> >> wakeup capabilies has to be made somewhere (and in fact is already made
> >> by existing code.)  But IMO, that decision should only be made where
> >> wakeup capabilies are known, so that sensible decisions (for power
> >> management) can be made.
> >> 
> >> Until there is a way in the generic code to distinguish between the
> >> various ways a device can wakeup, this decision should be left up to the
> >> code that knows how.
> >
> > OK, so I suppose your suggestion is to drop the patch and let the
> > .stop_device() and .power_off() PM domain callbacks to hand that, is this
> > correct?
> 
> Correct.
> 
> Initially I was thinking only about .power_off(), but you'd probably
> want this at .stop_device() too.  In order to do that, probably want
> .stop_device() to be able to return an error code such that an error
> would prevent .power_off().

I've just sent a reply to that. :-) I'll reproduce it below for easier
reference:

Neither .stop_device(), nor .power_off() can make such decisions,
because they are used for both runtime PM and system suspend, so they
shouldn't do system suspend-specific checks.

So the only way forward I can see is to add a special PM domain callback,
say .active_wakeup(), that will return "true" if the device is to be left
active when wakeup-enabled.  So the check you don't like will become
something like:

if (device_may_wakeup(dev) && genpd->active_wakeup
    && genpd->active_wakeup(dev))
        return 0;

Would that be better?

Rafael
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/