(Resending as text-only - sorry)
Bringing this item back up again.
I am not suggesting that the application of any particular policy
appears within the kernel or userspace or a secondary policy engine.
In general I am also against codifying policy within drivers.
I am interested seeing the ACPI notifier mechanism expanded to allow
AC/DC state changes propagate to other kernel drivers without requiring
a userspace in between.
I can continue to come up with real scenarios that would possibly
require kernel-to-kernel notification, but would rather focus this
discussion of the pure technical issues associated with adding the
notifier to the AC/DC ACPI subsystem.
Remember it is a one line patch.
Regards,
Matthew
-------- Original Message --------
Subject: Re: [PATCH][ACPI] AC/DC notifier
From: Willy Tarreau <[email protected]>
To: Pavel Machek <[email protected]>
Cc: "Matthew Garrett" <[email protected]>, "Tippett, Matthew"
<[email protected]>, "Langsdorf, Mark" <[email protected]>,
[email protected], [email protected],
[email protected], "Li, Samuel" <[email protected]>
Date: 08/16/2009 03:40 AM
>
> On Fri, Aug 14, 2009 at 06:32:33PM +0200, Pavel Machek wrote:
> > On Wed 2009-08-12 01:55:32, Matthew Garrett wrote:
> > > On Tue, Aug 11, 2009 at 08:51:49PM -0400, Tippett, Matthew wrote:
> > >
> > > > From a graphics perspective (your area of expertise), this
> will allow KMS
> > > > drivers to do some more intelligent actions based on the
> ac/dc state.
> > > > Some examples of this could be improving the power
> consumption of the
> > > > graphics hardware through adapting clock memory/engine
> settings for
> > > > reduced power consumption, reducing refresh rate of the
> display to reduce
> > > > scanout memory access, adjusting backlight brightness, etc.
> > >
> > > Right. As you say, my concern is that most of this should belong in
> > > userspace. Where we risk hardware damage there's an obvious
> argument for
> > > doing this in kernel, but we should ensure that that's limited to
> > > whatever coarse-grain handling is absolutely required rather than
> doing
> > > things like touching display brightness.
> >
> > Yep... Some may want to save power even when AC is online -- like when
> > running on UPS. Some may want max performmance even on battery.
>
> Wholeheartly agreed. IMHO, there's absolutely no relation between power
> source and the expected performance. It's really frustrating when your
> laptop becomes a snail on battery, as well as it's annoying to hear it
> sound like a hairdryer when plugged to mains. This should only be the
> user's choice. Mine automatically adjusts its frequency on demand,
> regardless of the power source, which provides me with the best
> experience. I think that all the tricks used to save power when running
> on battery were invented by laptop makers to artificially show longer
> lasting eventhough the machine sometimes becomes barely usable. For
> instance, some of them dim the backlight so that you can't read anything
> in full light, so you need a power prolongator to use them outside !
>
> Also, with the new trend of laptops making use of huge power-hungry 3D
> graphic chips which suck all the juice out of your battery in less than
> two hours doing nothing, you'd better run at full speed when on battery
> to save energy for CPU-bound tasks, because eventhough the CPU eats more
> power, you significantly reduce the run time, thus the static consumption
> (GPU, backlight, hard disk, ...).
>
> Willy
>
>
On Tue 2009-10-06 10:53:22, Tippett, Matthew wrote:
> (Resending as text-only - sorry)
>
> Bringing this item back up again.
>
> I am not suggesting that the application of any particular policy
> appears within the kernel or userspace or a secondary policy engine.
> In general I am also against codifying policy within drivers.
>
> I am interested seeing the ACPI notifier mechanism expanded to allow
> AC/DC state changes propagate to other kernel drivers without requiring
> a userspace in between.
>
> I can continue to come up with real scenarios that would possibly
> require kernel-to-kernel notification, but would rather focus this
> discussion of the pure technical issues associated with adding the
> notifier to the AC/DC ACPI subsystem.
Please do. So far you did not show valid use for such notifier.
(Ok, I know of one. Old amd64 notebooks had cpufreq scaling enabled,
with battery unable to supply enough current to feed the CPU at
highest cpufreq setting. At that point, scaling cpufreq down at unplug
is correctness issue, and AC/DC notifier in kernel makes
sense.)
So... what do you want to use it for?
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
On Wed, Oct 7, 2009 at 5:31 PM, Pavel Machek <[email protected]> wrote:
> On Tue 2009-10-06 10:53:22, Tippett, Matthew wrote:
>> (Resending as text-only - sorry)
>>
>> Bringing this item back up again.
>>
>> I am not suggesting that the application of any particular policy
>> appears within the kernel or userspace or a secondary policy engine.
>> In general I am also against codifying policy within drivers.
>>
>> I am interested seeing the ACPI notifier mechanism expanded to allow
>> AC/DC state changes propagate to other kernel drivers without requiring
>> a userspace in between.
>>
>> I can continue to come up with real scenarios that would possibly
>> require kernel-to-kernel notification, but would rather focus this
>> discussion of the pure technical issues associated with adding the
>> notifier to the AC/DC ACPI subsystem.
>
> Please do. So far you did not show valid use for such notifier.
>
> (Ok, I know of one. Old amd64 notebooks had cpufreq scaling enabled,
> with battery unable to supply enough current to feed the CPU at
> highest cpufreq setting. At that point, scaling cpufreq down at unplug
> is correctness issue, and AC/DC notifier in kernel makes
> sense.)
>
> So... what do you want to use it for?
I'm not sure we have a open driver use case for this, if not I suggest
we remove it and only add it when a user is added to the kernel.
Dave.
On Wed, Oct 07, 2009 at 06:16:29PM +1000, Dave Airlie wrote:
> I'm not sure we have a open driver use case for this, if not I suggest
> we remove it and only add it when a user is added to the kernel.
We'll need it once we add scaling support to radeon, so I don't see the
harm with it being merged now.
--
Matthew Garrett | [email protected]
(Resending in text-only sorry for the noise to the individual recipients)
Comments below.
-------- Original Message --------
Subject: Re: [PATCH][ACPI] AC/DC notifier
From: Pavel Machek <[email protected]>
To: Tippett, Matthew <[email protected]>
Cc: "Willy Tarreau" <[email protected]>, "Matthew Garrett" <[email protected]>,
"Langsdorf, Mark" <[email protected]>, [email protected],
[email protected], [email protected], "Li, Samuel"
<[email protected]>
Date: 10/07/2009 03:31 AM
>
> On Tue 2009-10-06 10:53:22, Tippett, Matthew wrote:
>
> Please do. So far you did not show valid use for such notifier.
>
> (Ok, I know of one. Old amd64 notebooks had cpufreq scaling enabled,
> with battery unable to supply enough current to feed the CPU at
> highest cpufreq setting. At that point, scaling cpufreq down at unplug
> is correctness issue, and AC/DC notifier in kernel makes
> sense.)
>
> So... what do you want to use it for?
>
We have a general requirement from OEMs and consequently our shared
Windows/Linux components that the AC/DC state is accurately known.
The concrete examples of use include at least the following.
1) Automatic frequency scaling has an AC-mode and a DC-mode in
Powerplay tables in the GPU BIOS ensures that the highest permitted
clocks fit the system design. This allows at least
i) system level thermals and power consumption to be managed
(eg: you shouldn't have the clocks up high if the system fan has been
asked to slow down).
ii) protection of hardware high clocks with a low-current
battery is a bad idea.
2) The pixel clock can only drive certain modes with certain engine
and memory clocks, in DC-mode you will have lower clocks and
consequently you will need to change the pixel clock and hence the mode
to something that fits within the budget, otherwise the 3D or display
engine may not be able to get the bandwidth required to operate effectively.
3) OEM OS equivalency. Some OEMs care that Linux has the same
thermal, power and performance characteristics as other operating
systems. This requires that the software act in a similar way and
reduces OEM/ODM/IHV/OSV validation and deployment costs.
In particular 1 and 2 are very relevant for KMS based drivers.
1 is most likely relevant for other ACPI/SBIOS related hardware
components that rely on co-operating components doing what is expected
by the system design. If a device doesn't change it's behaviour in sync
with other devices when AC/DC changes, then bad (or at the very least
annoying) things may happen. When there is a SBIOS/ACPI/Driver stack in
place, the driver should honor the system design as well.
3 is only really relevant for the commercial side of things, but is a
real issue that takes up more engineering effort than the first 2 - my
cross to bear so to speak.
Regards,
Matthew