On Wed, Oct 31, 2012 at 9:10 PM, Kevin Hilman
<[email protected]> wrote:
> Linus Walleij <[email protected]> writes:
>
> > piece of hardware, this would be the right thing to do,
> > and I think the in-kernel examples are all "simple",
> > e.g. arch/arm/mach-omap2/powerdomain* is all about
> > power domains and nothing else,
>
> FYI... that code isn't the same as PM domain.
This sort of points to a core problem here. Our terminologies are
ambiguous that we cannot understand each others code. As long
as <linux/pm_domain.h> begins:
/*
* pm_domain.h - Definitions and headers related to device power domains.
*
But arguably that should just be patched (I think there are a few
remnants in the code still implying that these things are only about
power).
>
> That code is for the
> *hardware* powerdomains, not the software concept of "PM domain." In
> OMAP, PM domain is implmented at the omap_device level. And omap_device
> is the abstraction of an IP block that knows about all the PM related
> register settings, clocks, HW powerdomain, voltage domain, PM related
> pin-muxing etc. etc. All of these things are abstracted in an
> omap_device, so that the PM domain implementation for OMAP looks rather
> simple (c.f. omap_device_pm_domain in arch/arm/plat-omap/omap_device.c.)
OK following now...
>
> > I think the lesser of two evils is the distributed approach,
>
> The pinctrl examples I've seen mentioned so far are all PM related
>
> (sleep, idle, wakeup, etc.) so to me I think they still belong in
> PM domains (and that's how we handle the PM related pins in OMAP.)
Well, the pinctrl grabbers in these drivers are using these states also
for platforms that do not even select CONFIG_PM. For example
mach-nomadik is quite happy that the PL011 driver is thusly
muxing in its pins. And would require refactoring to use PM
domains.
So basically this requirement comes down to:
- When dealing with a SoC IP block driver
- That need to multiplex pins
- Then your SoC must select CONFIG_PM and
CONFIG_PM_RUNTIME and
CONFIG_PM_GENERIC_DOMAINS and implement
proper domain handling hooks.
Is this correct? And for Mark, Dmitry, does this correspond to
your view?
It's actually something that needs to be acknowledged by the
ARM SoC maintainers, because they will be the ones telling
all subarch maintainers to go implement full PM handling
with these three frameworks whenever an SoC driver want
to handle pins.
Bascially all subsystem maintainers dealing with embedded
SoCs need to be aligned on this as well...
And IIUC not only pins but also silicon block clocks?
I can surely fix these for "my" systems, but it really needs
to be enforced widely or it will be a mess.
>
> > I worry that the per-SoC power domain implementation
> > which will live in arch/arm/mach-* as of today will become
> > the new board file problem, overburdening the arch/* tree.
> > Maybe I'm mistaken as to the size of these things,
> > but just doing ls arch/arm/mach-omap2/powerdomain*
> > makes me start worrying.
>
> Yes, I agree that this means more code/data in arch/arm/mach-*, but
> IMO, that's really where it belongs. It really is SoC integration
> details, and driver should really not know about it.
OK we need feedback from ARM SoC on this.
Yours,
Linus Walleij
Linus Walleij <[email protected]> writes:
> On Wed, Oct 31, 2012 at 9:10 PM, Kevin Hilman
> <[email protected]> wrote:
>
>> Linus Walleij <[email protected]> writes:
>
>
>>
>> > piece of hardware, this would be the right thing to do,
>> > and I think the in-kernel examples are all "simple",
>> > e.g. arch/arm/mach-omap2/powerdomain* is all about
>> > power domains and nothing else,
>>
>> FYI... that code isn't the same as PM domain.
>
>
> This sort of points to a core problem here. Our terminologies are
> ambiguous that we cannot understand each others code. As long
> as <linux/pm_domain.h> begins:
>
> /*
> * pm_domain.h - Definitions and headers related to device power domains.
> *
>
> But arguably that should just be patched (I think there are a few
> remnants in the code still implying that these things are only about
> power).
Agreed. The terminology is confusing, and any situations like this in
the code/comments/docs should be patched.
When PM domains were introduced, I was the first to complain that we
shouldn't use the term power domain so as not to be confused with HW
concepts, so we settled on the term 'PM domain.' Ultimately, it's just
a configurable grouping of devices whose callbacks happen during PM
transitions.
>>
>> That code is for the
>> *hardware* powerdomains, not the software concept of "PM domain." In
>> OMAP, PM domain is implmented at the omap_device level. And omap_device
>> is the abstraction of an IP block that knows about all the PM related
>> register settings, clocks, HW powerdomain, voltage domain, PM related
>> pin-muxing etc. etc. All of these things are abstracted in an
>> omap_device, so that the PM domain implementation for OMAP looks rather
>> simple (c.f. omap_device_pm_domain in arch/arm/plat-omap/omap_device.c.)
>
>
> OK following now...
>
>>
>> > I think the lesser of two evils is the distributed approach,
>>
>> The pinctrl examples I've seen mentioned so far are all PM related
>>
>> (sleep, idle, wakeup, etc.) so to me I think they still belong in
>> PM domains (and that's how we handle the PM related pins in OMAP.)
>
>
> Well, the pinctrl grabbers in these drivers are using these states also
> for platforms that do not even select CONFIG_PM. For example
> mach-nomadik is quite happy that the PL011 driver is thusly
> muxing in its pins. And would require refactoring to use PM
> domains.
If CONFIG_PM is disabled, then is it safe to assume that the pins in
question are probably only done once at init time. I assume during
->probe(). ?
>
> So basically this requirement comes down to:
>
> - When dealing with a SoC IP block driver
>
> - That need to multiplex pins
>
> - Then your SoC must select CONFIG_PM and
> CONFIG_PM_RUNTIME andb
> CONFIG_PM_GENERIC_DOMAINS and implement
> proper domain handling hooks.
>
> Is this correct?
I would say yes. Currently, PM domains are the way to hook SoC-specific
integration details into PM transitions.
However, if what we want/need are only ways to introduce SoC-specific
integration details into non-PM transitions (e.g. probe/remove), maybe
bus notifiers would suffice here. e.g. you'd get a bus notifier when
the device is added/attached and any init-time pinctrl setup could be
done then. This still keeps drivers clean of SoC-specific integration
data/code, and also allows that to happen whether or not PM features are
enabled.
Kevin
On Thu, Nov 1, 2012 at 12:42 PM, Kevin Hilman
<[email protected]> wrote:
> [Me]
>> Well, the pinctrl grabbers in these drivers are using these states also
>> for platforms that do not even select CONFIG_PM. For example
>> mach-nomadik is quite happy that the PL011 driver is thusly
>> muxing in its pins. And would require refactoring to use PM
>> domains.
>
> If CONFIG_PM is disabled, then is it safe to assume that the pins in
> question are probably only done once at init time. I assume during
> ->probe(). ?
Sadly no.
Consider drivers/tty/serial/amba-pl011.c
Many ARM platforms have several instances of PL011, and not all of
them have CONFIG_PM & friends, so it's a good example.
Here the driver will probe and currently fetch a pinctrl handle and
looks up two states: "default", which refers to the situation you
describe, and optionally "sleep" which will put pins into a
low-power state.
The driver will currently put the pins into the "sleep" state when
.shutdown() is called by something (userspace or in-kernel users).
So in the new suggested scheme using runtime PM, this
would have to be replaced by pm_runtime_get[_sync]()
and pm_runtime_put() hints and the current pin handling
deleted, and for each platform using this driver instead
implement a PM domain to do the same thing.
Else you loose this runtime power optimization.
This is what I refer to the all-or-nothing charcter of
runtime PM domains... but maybe it's a good thing,
I haven't quite made my mind up about it.
> (...) if what we want/need are only ways to introduce SoC-specific
> integration details into non-PM transitions (e.g. probe/remove), maybe
> bus notifiers would suffice here. e.g. you'd get a bus notifier when
> the device is added/attached and any init-time pinctrl setup could be
> done then. This still keeps drivers clean of SoC-specific integration
> data/code, and also allows that to happen whether or not PM features are
> enabled.
It doesn't cut it for any of our drivers as shown above,
but it would work for the patch in $SUBJECT.
It sounds like the way silicon clocks are handled on
SH am I right?
Yours,
Linus Walleij