Re: [linux-pm] Power Management framework proposal

[david]

On Sun, 22 Jul 2007, Igor Stoppa wrote:

> On Sun, 2007-07-22 at 01:58 -0700, ext david@xxxxxxx wrote:
> > On Sun, 22 Jul 2007, Igor Stoppa wrote:
>
> [snip]
>
> > > Could you elaborate on how your proposal is incompatible with enhancing
> > > the clock framework?
> >
> > It's not that I think it's incompatible with any existing powersaving
> > tools (in fact I hope it's not)
> >
> > it's that I think that this (or something similar) could be made to cover
> > all thevarious power options instead of CPU's having one interface, ACPI
> > capable drivers having another, embeded devices presenting a third, etc
> >
> > this was triggered by the mess of different function calls for different
> > purposes that are used for the suspend functions where you have a bunch of
> > different functions that are each supposed to be called at a specific time
> > from a specific mode during the suspend process. with all these different
> > functions driver writes tend to not bother implementing any of them, and
> > it seems like there is a fairly steady stream of new functions that end up
> > being needed. the initial intent was to just change this into a generic
> > set of calls that every driver writer would implement the minimum set of,
> > and make it trivially extensable to future capabilities of hardware.
>
> Every now and then there is some attempt to find One solution to bind
> them all: x86, SoC, ACPI ... you name it.

this is another one. I'd be happy to get pointers to prior ones to learn 
from.

> Unfortunately, while it's true that there are significant similarities,
> there are also notable differencies; as far as i know the USB subsystem
> is the one that gets closer to what we have in the embedded arena, since
> it can have complex cases of parent-child powering and wakeup.

this API is not trying to represent the parent-child hierarchy. as far as 
I know that's documented in sysfs (or is supposed to be). this is just an 
attempt to make it so that as you are going through the hierarchy you 
don't have to use vastly different API's to control the different 
functions.

I suspect that most (if not all) of the previous One Solutions have tried 
to completely handle all the details of their original case, and then 
branch out to the other cases.

this attempt is working from the other direction. the user of this API 
doesn't care how something is done, it just wants to know what's possible 
and how to tell the system to switch modes.

other then just me searching through the lists, do you have a pointer to 
some of the differences between the different types that are seen as being 
so large that they can't be unified?

> > while I was describing the issues to my roomates over dinner I realized
> > that the same type of functions are needed for the CPU clocks.
> >
> > if you have an accepted framework in place there that can do what I
> > described, please consider extending it to cover other types of devices
> > and drivers.
>
> That is not part of the fw: the fw simply expresses parent-child clock
> distribution and keeps usecounts so that unused clocks are automatically
> gated.
>
> The actual clock tree description is platform/arch/board specific and
> doesn't affect the framework. You can just roll your own version for x86
> by providing a description of the methods used to switch on/off every
> individual clock on your board.
>
> So what you are asking for is that somebody writes an x86 version of the
> clock fw.

this is more then just setting the clocks on everything (although setting 
clocks seems like it fits well into the model) becouse some power modes 
are not easily represented just as clocks.

> As for latencies, well, only few clocks really have significant impact.
> Most notably the main system oscillator. Everything else has 0 latency
> since it ends up in opening/closing a clock gate.
>
> Powering device on/off will certainly introduce more latency, but either
> the powering is supported by the hw, to make it quick or it has to go
> through most, if not all of he usual initialisation sequence; in that
> case it probably makes sense to avoid controlling it from kernelspace,
> since it will be slow and won't require dedcisions made with us
> precision.

and many devices support both a quick almost-off mode and a slow 
almost-off mode (as well as a completely off mode), with the slow mode 
eating less power, but takeing longer to wake up from. that's the reason 
for providing the matrix to let the program makeing the decision decide if 
it's worth the time delays to get the power savings

as I note in anther message, this SPI isn't intended to be strictly 
kernelspace or strictly userspace. for the ondemand speed governer you are 
changing the settings quickly and so probably want to do so in the kernel, 
however some people may be satisfied with slower controls and so could 
have them in userspace (an extreme example of this would be turning off 
wireless cards that aren't in use to save power and improve security)


> > I think you are passing too much
> > info up the chain to the part makeing the decision (that part doesn't need
> > to know the details of the voltage/freq choices, the %power/%capability
> > numbers I suggested are in many ways more what they are making decision
> > son anyway)
>
> I don't think you have got it right: the only info being passed is the
> standard cpufreq list of frequencies; everything else is part of the
> cpufreq driver.

to make the decisions the software makeing the decision needs to know how 
much power would be used at each freq setting.

> > in the slideshow you list in the sequence of changing the cpu speed to pre
> > and post notify drivers. what exactly are the drivers expected to do with
> > the notification? are you asking them to pause and then re-initialize for
> > the new power level?
>
> It's just a  notification. The drivers are supposed to know how to deal
> with it.
> In OMAP2 the major concern is that the external memory cannot be
> accessed since it is on a bus that is being re-clocked:
> - the dma controllers must be paused
> - the other cores (dsp) must not access sdram
> - the onenand driver needs to adjust its timing parameters

in my proposal this would require one or more 'pause' modes (more then 
one if you need to pause at different power settings fro some reason) for 
the first notification, and then you would set them to the mode you want 
them in at the second notification point (which is probably going to be 
the mode they were in before)

> [snip]
>
> > > To make any proposal that has some chance of being accepted, you have to
> > > compare it against the existing solution, explaining:
> > >
> > > -what it is bringing in terms of new functionalities
> > > -how it is different
> >
> > it unifies all power/performance trade-offs (including power on/off) into
> > a single API, but decouples that API from the implementation details of
> > exactly what the technical details of the different modes are and how the
> > changes are made.
>
> It always looks great at this level of abstraction, but then usually
> what is discovered later is that _a lot_ of extra complexity is
> introduced, in order to cover every case on every platform that is
> intended to be supported.

which is why I posted this for comments.

what are the cases that require extra info.can that extra info be as 
simple as a set of flags for the mode (or possibly for the transition 
matrix).

for your clock example you need a flag that says 'this requires everything 
connected to this be paused'

for suspend other low power modes you need to be able to say 'contents of 
things below this point will be lost when you go into this mode' so that 
the decision makeing software knows that it needs to save the contents of 
memory before switching to a mode that stops the dram refresh. I don't 
have any idea at the moment for how to prvide a common interface for 
actually saving or restoring the contents, that is outside the scope of 
this API

the ACPI people will need a flag for 'this device can generate wakeup 
signals in this mode'

but this API would just provide this info to the decision makeing code, 
that code would have to antually enforce the limits

> > for some subsystems this would be little more then renameing existing
> > fucntions, for others it would be converting several indepndant functions
> > into one, discoverable api
>
> if you check cpufreq, you will find out that it already covers the
> multiple cores case (but nothing prevents from using the same logic on
> something that is not really a cpu) and also has some simple concept of
> latency for frequency transition, concept that could be enhanced to
> handle latencies that are depending on the current operating point and
> target operating point.

does it provide a full matrix of latencies, or just mode 1->mode2=x, 
mode2->mode3=y so mode1->mode3=x+y?

> > > -why the current implementation cannot simply be enhanced
> >
> > which current implementation should be enhanced? and with the massive
> > broadening of functionality should it retain the same name, or should it
> > get renamed to something more generic?
>
> cpufreq could be renamed to anything that makes sense, but i see _no_
> massive broadening of functionality.

what I'm talking about would provide an API to devices that you are 
ignoring becouse they should be managed from userspace.

> > the cpufreq implementation is very close to what I'm proposing, it would
> > need to get broadend to cover other devices (like disk drives, wireless
> > cards, etc), is this really the right thing to do or should the more
> > generic API go in for external use and then the existing cpufreq be called
> > from the set_mode() call?
>
> No, that doesn't make sense, as general approach.
> You want to manage from kernel only those parts of the system where the
> latency is so low that userspace wouldn't be able to keep up.
>
> Your examples (wireless, disk drive) can be easily controlled from
> userspace, with a timeout.

absoutly, and they should be (at least most of the time). this was not 
intended as a kernelspace only api. it is intended to be available to both 
kernelspace and userspace.

> In both cases there are significant delays (change of rotation speed /
> sync with the access point).

correct, and these delays should be reflected in the transition cost 
matrix

> All this is hand waiving unless it is backed up by numbers.
> Real cases are required in order to establish a list of priorities for
> latency/power consumption.

this isn't attempting to establish a list of priorites, simply to give the 
software that is trying to establish such a list the info to make it's 
decisions, and the interface to use to issue the resulting instructions.

> Afterward, a valid solution that can address such cases can be sketched.

with this API you should be able to create a very trivial power manager 
that can know nothing about the system other then the info found in this 
API and the hirarchy of devices, but can transition the system between 
three easily explained modes.

A. full power operation
B. off
C. as low a power mode as is available on the hardware without having to 
save the contents of something somewhere else.

Thanks for your time in replying to me on this topic.

David Lang
-
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/