Re: Ottawa [topics]

[Benjamin Herrenschmidt <benh-XVmvHMARGAS8U2dJNN8I7kB+6BGkLq7r@public.gmane.org>]

[-- Attachment #1: Type: text/plain, Size: 6742 bytes --]

On Wed, 2005-03-16 at 13:50 -0800, David Brownell wrote:

> > 
> > Yes, but it's orthogonal to the PM model. A device could even request
> > slower clocks in intermediate power states or whatever, but it's the
> > device driver that calls into the "clock manager" on power state
> > transitions (or even at runtime) to express it's needs.
> 
> Not orthogonal.  Suspending a device will often involve a clock
> transition ... and certain clocking levels will preclude certain
> system PM states.

 .../...

Oh well, then come up with a scheme that can deal with those without
beeing as big as the kernel itself...

> 
> > > That's sort of a "low level" category ... except that it pops up to
> > > higher level concerns.  For example, clocks may be shared (which can
> > > make the "low level" vs "driver state and relationships" distinctions
> > > look nonexistent), and they may interact with system power transitions. 
> > 
> > A bit, but then, the clock net is often even more scary than the power
> > tree. I wouldn't try to mix them up.
> 
> Embedded chips normally couple clock management with power management
> very tightly.  They may have separate power switching controls, but
> that's less widespread.

I am familiar with that, and I still think this shouldn't be directly
dealt by the PM framework. In fact, in most embedded cases, drivers are
specific anyway and can embed the clock net stuff.

>  
> > > The primary mechanism I've heard discussed is having devices monitor
> > > when they haven't been used recently, and using that to enter some
> > > suspend level.  The "not even active" clock gating approach isn't
> > > quite the same.
> > 
> > Yes, drivers deciding themselves, or through sysfs. Part of the model I
> > propose is to let drivers expose a richer set of states to userland as
> > well. Policy for things like handelds can become so complicated that I
> > think it should be moved to userspace.
> 
> When drivers just don't use clocks they don't need (a "policy"), it
> gets simple rather quickly.  Pushing it to userspace would be a way
> to add complexity ... IMO a non-goal.

Gah ? Did you understand what I said ?
> 
> > > > 4.) Userspace - how to notify of power changes, how to allow the user to
> > > > control power management.
> > > > 
> > > > Anything else?
> > > 
> > > Oh, and wakeup event sources.  ACPI has a very indirect approach,
> > > which seems to underlay some assumptions on this list.  There's a
> > > similar thing going on with PCI.  So for example when folk start
> > > to assume that a "wakeup" event is for some reason distinct from
> > > the device's normal IRQ, I must cringe ... anyone assuming that
> > > device or system suspend involves turning off IRQs is erecting
> > > roadblocks against most embedded hardware.
> > 
> > Yes and no ... They must be treated differently for sure. There are a
> > lot of cases where they are actually completely different.
> 
> And as many when they aren't.  Hence my point that assuming they ARE
> is making trouble.  Some of the swsusp discussions seem to bring that
> notion out.

It's all bla bla bla as far as I'm concerned, we should find a model and
see if it actually precludes anything. I think in all cases, the
specifics of clock dependencies can be expressed as power state
dependencies.
> 
> > The problem 
> > we discussed a while ago with USB is an example of the mess, and I still
> > have to investigate in details. Basically, the current USB code upstream
> > will crash the machine on sleep 90% of the times when I have something
> > plugged, while the same set of devices sleeps happily in OS X 100% of
> > the time. So we must be doing something wrong.
> 
> Orthogonal issue; still loks specific to your hardware.

"My" hardware is jsut a fucking standard NEC USB2 chip in most cases and
a lucent cell in Apple stuff in the other. Also "My" hardware is also
the class of laptops that happen to have had working sleep support for
much longer and much more reliably than any x86 out there. You are
breaking that and don't seem to be interested in making that work. THat
makes me thing that working sleep on USB today is just a matter of luck.
 
> > Wether the system IRQs must be masked or not is one thing that is
> > platform specific during suspend. Wether devices must stop their IRQ
> > emission, well... 
> 
> What's a "system IRQ" as opposed to any other kind?  Drivers don't
> generally know or care how their IRQs are routed, they just care
> that they get a callback.

"system irq" means masking the irq at the PIC level (enable/disable irq)
while masking the irq at the device level means stopping irq emission on
the device using device local registers. It's totally different and may
have a different impact on PM depending on how wakeup works.
> 
> > At least with USB, you do free_irq(). That means that 
> > unless your IRQ is shared, it _will_ be masked at the controller level
> > anyway. So if the IRQ is used to wakeup the CPU, that is broken, at
> > least unless the IRQ line from the device (before reaching the IRQ
> > controller) is also triggering some sort of external wakeup mecanism.
> 
> That's a good example of a PCI-specific assumption, the type I
> was saying creates problems.  Non-PCI hardware has no reason to
> free the IRQ.  (In fact, the only reason PCI hardware does it is
> to keep that PPC-ism out of the OHCI code.)

Which we have to keep anyway for the suspend code. And embedded with
bizarre clock stuff will need hooks there too. I don't understand how
you try to get the ppc stuff out on one side and deal with all the
embedded cruft on the other. Masking IRQ in my case is mostly a matter
of sanity. I'm switching the main clock off on the OHCI cell and disable
the cell in the ASIC, I disable the IRQ that I know is not shared to
make sure that there is no "dangling" irq line. I'm fairly sure other
platforms (embedded especially) will need similar tweaks.
>  
> > I tend to think that devices using IRQs for wakeup instead of separate
> > IOs are broken design, but then, we all know that HW folks can't get
> > anything right ;)
> 
> Again, that's a strong PCI bias.

Not at all. I would have said the same before we had PCI PM in the PCI
spec.

>   Having PCI biases like that is
> just creating trouble.  In fact, not all PCI hardware works that
> way either ... just stuff using the PCI PM model.  And the last
> many PCI busess I've looked at have had devices that don't use
> the PCM PM attributes, don't support PCI D3 ...
> 
> - Dave
> 
> 
> 
> > In some cases, we might have to provide some arch hooks to be called by
> > drivers to know what to do ...
> > 
> > Ben.
> > 
> > 
-- 
Benjamin Herrenschmidt <benh-XVmvHMARGAS8U2dJNN8I7kB+6BGkLq7r@public.gmane.org>


[-- Attachment #2: Type: text/plain, Size: 0 bytes --]