Date:   Wed, 18 Oct 2017 15:00:38 +0200
From:   "Rafael J. Wysocki" <rjw@...ysocki.net>
To:     Ulf Hansson <ulf.hansson@...aro.org>
Cc:     Linux PM <linux-pm@...r.kernel.org>,
        Bjorn Helgaas <bhelgaas@...gle.com>,
        Alan Stern <stern@...land.harvard.edu>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Linux ACPI <linux-acpi@...r.kernel.org>,
        Linux PCI <linux-pci@...r.kernel.org>,
        Linux Documentation <linux-doc@...r.kernel.org>,
        Mika Westerberg <mika.westerberg@...ux.intel.com>,
        Andy Shevchenko <andriy.shevchenko@...ux.intel.com>,
        Kevin Hilman <khilman@...nel.org>,
        Wolfram Sang <wsa@...-dreams.de>,
        "linux-i2c@...r.kernel.org" <linux-i2c@...r.kernel.org>,
        Lee Jones <lee.jones@...aro.org>
Subject: Re: [PATCH 0/12] PM / sleep: Driver flags for system suspend/resume

On Wednesday, October 18, 2017 1:57:52 PM CEST Ulf Hansson wrote:
> On 18 October 2017 at 02:39, Rafael J. Wysocki <rjw@...ysocki.net> wrote:
> > On Tuesday, October 17, 2017 9:41:16 PM CEST Ulf Hansson wrote:
> >
> > [cut]
> >
> >> >
> >> >> deploying this and from a middle layer point of view, all the trivial
> >> >> cases supports this.
> >> >
> >> > These functions are wrong, however, because they attempt to reuse the
> >> > whole callback *path* instead of just reusing driver callbacks.  The
> >> > *only* reason why it all "works" is because there are no middle layer
> >> > callbacks involved in that now.
> >> >
> >> > If you changed them to reuse driver callbacks only today, nothing would break
> >> > AFAICS.
> >>
> >> Yes, it would.
> >>
> >> First, for example, the amba bus is responsible for the amba bus
> >> clock, but relies on drivers to gate/ungate it during system sleep. In
> >> case the amba drivers don't use the pm_runtime_force_suspend|resume(),
> >> it will explicitly have to start manage the clock during system sleep
> >> themselves. Leading to open coding.
> >
> > Well, I suspected that something like this would surface. ;-)
> >
> > Are there any major reasons why the appended patch (obviously untested) won't
> > work, then?
> 
> Let me comment on the code, instead of here...
> 
> ...just realized your second reply, so let me reply to that instead
> regarding the patch.
> 
> >
> >> Second, it will introduce a regression in behavior for all users of
> >> pm_runtime_force_suspend|resume(), especially during system resume as
> >> the driver may then end up resuming the device even in case it isn't
> >> needed.
> >
> > How so?
> >
> > I'm talking about a change like in the appended patch, where
> > pm_runtime_force_* simply invoke driver callbacks directly.  What is
> > skipped there is middle-layer stuff which is empty anyway in all cases
> > except for AMBA (if that's all what is lurking below the surface), so
> > I don't quite see how the failure will happen.
> 
> I am afraid changing pm_runtime_force* to only call driver callbacks
> may become fragile. Let me elaborate.
> 
> The reason why pm_runtime_force_* needs to respects the hierarchy of
> the RPM callbacks, is because otherwise it can't safely update the
> runtime PM status of the device.

I'm not sure I follow this requirement.  Why is that so?

> And updating the runtime PM status of
> the device is required to manage the optimized behavior during system
> resume (avoiding to unnecessary resume devices).

Well, OK.  The runtime PM status of the device after system resume should
better reflect its physical state.

[The physical state of the device may not be under the control of the
kernel in some cases, like in S3 resume on some systems that reset
devices in the firmware and so on, but let's set that aside.]

However, for the runtime PM status of the device may still reflect its state
if, say, a ->resume_early of the middle layer is called during resume along
with a driver's ->runtime_resume.  That still can produce the right state
of the device and all depends on the middle layer.

On the other hand, as I said before, using a middle-layer ->runtime_suspend
during a system sleep transition may be outright incorrect, say if device
wakeup settings need to be adjusted by the middle layer (which is the
case for some of them).

Of course, if the middle layer expects the driver to point its
system-wide PM callbacks to pm_runtime_force_*, then that's how it goes,
but the drivers working with this particular middle layer generally
won't work with other middle layers and may interact incorrectly
with parents and/or children using the other middle layers.

I guess the problem boils down to having a common set of expectations
on the driver side and on the middle layer side allowing different
combinations of these to work together.

> Besides the AMBA case, I also realized that we are dealing with PM
> clocks in the genpd case. For this, genpd relies on the that runtime
> PM status of the device properly reflects the state of the HW, during
> system-wide PM.
> 
> In other words, if the driver would change the runtime PM status of
> the device, without respecting the hierarchy of the runtime PM
> callbacks, it would lead to that genpd starts taking wrong decisions
> while managing the PM clocks during system-wide PM. So in case you
> intend to change pm_runtime_force_* this needs to be addressed too.

I've just looked at the genpd code and quite frankly I'm not sure how this
works, but I'll figure this out. :-)

> >
> >> I believe I have explained why, also several times by now -
> >> and that's also how far you could take the i2c designware driver at
> >> this point.
> >>
> >> That said, I assume the second part may be addressed in this series,
> >> if these drivers convert to use the "driver PM flags", right?
> >>
> >> However, what about the first case? Is some open coding needed or your
> >> think the amba driver can instruct the amba bus via the "driver PM
> >> flags"?
> >
> > With the appended patch applied things should work for AMBA like for
> > any other bus type implementing PM, so I don't see why not.
> >
> >> >
> >> >> Like the spi bus, i2c bus, amba bus, platform
> >> >> bus, genpd, etc. There are no changes needed to continue to support
> >> >> this option, if you see what I mean.
> >> >
> >> > For the time being, nothing changes in that respect, but eventually I'd
> >> > prefer the pm_runtime_force_* things to go away, frankly.
> >>
> >> Okay, thanks for that clear statement!
> >>
> >> >
> >> >> So, when you say that re-using runtime PM callbacks for system-wide PM
> >> >> isn't going to happen, can you please elaborate what you mean?
> >> >
> >> > I didn't mean "reusing runtime PM callbacks for system-wide PM" overall, but
> >> > reusing *middle-layer* runtime PM callbacks for system-wide PM.  That is the
> >> > bogus part.
> >>
> >> I think we have discussed this several times, but the arguments you
> >> have put forward, explaining *why* haven't yet convinced me.
> >
> > Well, sorry about that.  I would like to be able to explain my point to you so
> > that you understand my perspective, but if that's not working, that's not a
> > sufficient reason for me to give up.
> >
> > I'm just refusing to maintain code that I don't agree with in the long run.
> >
> >> In principle what you have been saying is that it's a "layering
> >> violation" to use pm_runtime_force_suspend|resume() from driver's
> >> system sleep callbacks, but on the other hand you think using
> >> pm_runtime_get*  and friends is okay!?
> >
> > Not unconditionally, which would be fair to mention.
> >
> > Only if it is called in ->prepare or as the first thing in a ->suspend
> > callback.  Later than that is broken too in principle.
> >
> >> That makes little sense to me, because it's the same "layering
> >> violation" that is done for both cases.
> >
> > The "layering violation" is all about things possibly occurring in a
> > wrong order.  For example, say a middle-layer ->runtime_suspend is
> > called via pm_runtime_force_suspend() which in turn is called from
> > middle-layer ->suspend_late as a driver callback.  If the ->runtime_suspend
> > does anything significat to the device, then executing the remaining part of
> > ->suspend_late will almost cetainly break things, more or less.
> >
> > That is not a concern with a middle-layer ->runtime_resume running
> > *before* a middle-layer ->suspend (or any subsequent callbacks) does
> > anything significant to the device.
> >
> > Is there anything in the above which is not clear enough?
> >
> >> Moreover, you have been explaining that re-using runtime PM callbacks
> >> for PCI doesn't work. Then my question is, why should a limitation of
> >> the PCI subsystem put constraints on the behavior for all other
> >> subsystems/middle-layers?
> >
> > Because they aren't just PCI subsystem limitations only.  The need to handle
> > wakeup setup differently for runtime PM and system sleep is not PCI-specific.
> > The need to handle suspend and hibernation differently isn't too.
> >
> > Those things may be more obvious in PCI, but they are generic rather than
> > special.
> 
> Absolutely agree about the different wake-up settings. However, these
> issues can be addressed also when using pm_runtime_force_*, at least
> in general, but then not for PCI.

Well, not for the ACPI PM domain too.

In general, not if the wakeup settings are adjusted by the middle layer.

> Regarding hibernation, honestly that's not really my area of
> expertise. Although, I assume the middle-layer and driver can treat
> that as a separate case, so if it's not suitable to use
> pm_runtime_force* for that case, then they shouldn't do it.

Well, agreed.

In some simple cases, though, driver callbacks can be reused for hibernation
too, so it would be good to have a common way to do that too, IMO.

> >
> > Also, quite so often other middle layers interact with PCI directly or
> > indirectly (eg. a platform device may be a child or a consumer of a PCI
> > device) and some optimizations need to take that into account (eg. parents
> > generally need to be accessible when their childres are resumed and so on).
> 
> A device's parent becomes informed when changing the runtime PM status
> of the device via pm_runtime_force_suspend|resume(), as those calls
> pm_runtime_set_suspended|active().

This requires the parent driver or middle layer to look at the reference
counter and understand it the same way as pm_runtime_force_*.

> In case that isn't that sufficient, what else is needed? Perhaps you can
> point me to an example so I can understand better?

Say you want to leave the parent suspended after system resume, but the
child drivers use pm_runtime_force_suspend|resume().  The parent would then
need to use pm_runtime_force_suspend|resume() too, no?
 
> For a PCI consumer device those will of course have to play by the rules of PCI.
> 
> >
> > Moreover, the majority of the "other subsystems/middle-layers" you've talked
> > about so far don't provide any PM callbacks to be invoked by pm_runtime_force_*,
> > so question is how representative they really are.
> 
> That's the point. We know pm_runtime_force_* works nicely for the
> trivial middle-layer cases.

In which cases the middle-layer callbacks don't exist, so it's just like
reusing driver callbacks directly. :-)

> For the more complex cases, we need something additional/different.

Something different.

But overall, as I said, this is about common expectations.

Today, some middle layers expect drivers to point their callback pointers
to the same routine in order to resue it (PCI, ACPI bus type), some of them
expect pm_runtime_force_suspend|resume() to be used (AMBA, maybe genpd),
and some of them have no expectations at all.

There needs to be a common ground in that area for drivers to be able to
work with different middle layers.

> >
> >> >
> >> > Quoting again:
> >> >
> >> > "If you are a middle layer, your role is basically to do PM for a certain
> >> > group of devices.  Thus you cannot really do the same in ->suspend or
> >> > ->suspend_early and in ->runtime_suspend (because the former generally need to
> >> > take device_may_wakeup() into account and the latter doesn't) and you shouldn't
> >> > really do the same in ->suspend and ->freeze (becuase the latter shouldn't
> >> > change the device's power state) and so on."
> >> >
> >> > I have said for multiple times that re-using *driver* callbacks actually makes
> >> > sense and the series is for doing that easier in general among other things.
> >> >
> >> >> I assume you mean that the PM core won't be involved to support this,
> >> >> but is that it?
> >> >>
> >> >> Do you also mean that *all* users of pm_runtime_force_suspend|resume()
> >> >> must convert to this new thing, using "driver PM flags", so in the end
> >> >> you want to remove pm_runtime_force_suspend|resume()?
> >> >>  - Then if so, you must of course consider all cases for how
> >> >> pm_runtime_force_suspend|resume() are being deployed currently, else
> >> >> existing users can't convert to the "driver PM flags" thing. Have you
> >> >> done that in this series?
> >> >
> >> > Let me turn this around.
> >> >
> >> > The majority of cases in which pm_runtime_force_* are used *should* be
> >> > addressable using the flags introduced here.  Some case in which
> >> > pm_runtime_force_* cannot be used should be addressable by these flags
> >> > as well.
> >>
> >> That's sounds really great!
> >>
> >> >
> >> > There may be some cases in which pm_runtime_force_* are used that may
> >> > require something more, but I'm not going to worry about that right now.
> >>
> >> This approach concerns me, because if we in the end realizes that
> >> pm_runtime_force_suspend|resume() will be too hard to get rid of, then
> >> this series just add yet another generic way of trying to optimize the
> >> system sleep path for runtime PM enabled devices.
> >
> > Which also works for PCI and the ACPI PM domain and that's sort of valuable
> > anyway, isn't it?
> 
> Indeed it is! I am definitely open to improve the situation for ACPI and PCI.
> 
> Seems like I may have given the wrong impression about that.
> 
> >
> > For the record, I don't think it will be too hard to get rid of
> > pm_runtime_force_suspend|resume(), although that may take quite some time.
> >
> >> So then we would end up having to support the "direct_complete" path,
> >> the "driver PM flags" and cases where
> >> pm_runtime_force_suspend|resume() is used. No, that just isn't good
> >> enough to me. That will just lead to similar scenarios as we had in
> >> the i2c designware driver.
> >
> > Frankly, this sounds like staging for indefinite blocking of changes in
> > this area on non-technical grounds.  I hope that it isn't the case ...
> >
> >> If we decide to go with these new "driver PM flags", I want to make
> >> sure, as long as possible, that we can remove both the
> >> "direct_complete" path support from the PM core as well as removing
> >> the pm_runtime_force_suspend|resume() helpers.
> >
> > We'll see.
> >
> >> >
> >> > I'll take care of that when I'll be removing pm_runtime_force_*, which I'm
> >> > not doing here.
> >>
> >> Of course I am fine with that we postpone doing the actual converting
> >> of drivers etc from this series, although as stated above, let's sure
> >> we *can* do it by using the "driver PM flags".
> >
> > There clearly are use cases that benefit from this series and I don't see
> > any alternatives covering them, including both direct-complete and the
> > pm_runtime_force* approach, so I'm not buying this "let's make sure
> > it can cover all possible use cases that exist" argumentation.
> 
> Alright, let me re-phrase my take on this.
> 
> Because you stated that you plan to remove pm_runtime_force_*
> eventually, then I think you need to put up some valid reasons of why
> (I consider that done), but more importantly, you need to offer an
> alternative solution that can replace it. Else such that statement can
> easily become wrong interpreted. My point is, the "driver PM flags" do
> *not* offers a full alternative solution, it may do in the future or
> it may not.
> 
> So, to conclude from my side, I don't have any major objections to
> going forward with the "driver PM flags", especially with the goal of
> improving the situation for PCI and ACPI. Down the road, we can then
> *try* to make it replace pm_runtime_force_* and the "direct_complete
> path".
> 
> Hopefully that makes it more clear.

Yes, it does, thank you!

Rafael

Powered by blists - more mailing lists