[linux-pm] Runtime PM discussion notes

[Alan Stern]

On Thu, 23 Jun 2011, Paul Walmsley wrote:

> Hi
> 
> a few thoughts here:
> 
> On Wed, 15 Jun 2011, Rafael J. Wysocki wrote:
> 
> > On Tuesday, June 14, 2011, Magnus Damm wrote:
> >
> > > As for freezing user space, yes, I agree. The other feature including
> > > a different set of wakeup sources, not so sure why you would want to
> > > have that. I can't think of any good use case for that, from my point
> > > of view system-wide suspend is more like the limping cousin of the
> > > full-blown kernel.
> > 
> > Well, the freezing of user space by itself doesn't buy you anything
> > power management-wise, you pretty much need to do the other things too to
> > really save energy this way. 
> 
> Freezing user space stops CPU-hogging processes from running.  When the 
> runqueue is empty, the scheduler can go idle.  This in turn allows the 
> CPU(s) to enter low power sleep states via CPUIdle.

Very true.  At the moment, isn't it possible for the userspace 
ioctl PM interface to freeze processes without going all the way 
through to a system sleep?

> As I understand it, this is really the basis for the modern Android 
> use-case for wakelocks and opportunistic suspend.  They prevent 
> non-power-privileged userspace applications from keeping the system from 
> entering a low-power state.  (Secondarily, suspend also prevents kernel 
> code from running, e.g., timers set by filesystems, the networking stack, 
> etc.; but problems here should be fixed in the offending kernel code, 
> rather than hacked around, since some of the users of those timers could 
> be important)

I think you have oversimplified things a little, but never mind...

> > Anyway, the value of system suspend is to allow the user to tell the system
> > "OK, now I'm not going to use you for a while and I don't want the USB mouse
> > to wake you up", so that the system can go into a state in which it draws
> > minimum energy.
> > 
> > Of course, you can argue that the system may detect when it's not used and
> > put itself into that state, which is somewhat correct, but the system can
> > _never_ know two things by itself: (1) that it's not going to be used _in_
> > _advance_
> 
> But suspend users don't know this either, since they can't predict when 
> the next external wakeup can happen.

But they do know (or should know) that they don't intend to use the 
system in the near future.  It might be good to have a separate way to 
express that idea to the kernel.

> Having some kind of indication that the user is done using a device for 
> the time being is definitely useful.  But that's separate from system 
> suspend.  System suspend is just one of many actions that the system might 
> choose to take when the user gives that indication.
> 
> Consider something like a screensaver.  Usually, the longer latency the 
> wakeup operation is, the longer the screensaver waits before deciding to, 
> say, power the monitor off, or put the system into suspend (as we 
> discussed that Mac OS X does at LinuxCon Boston).  Similarly, if there is 
> some background task that really needs to run, an intelligent system will 
> not allow a power economization mode to prevent that from happening.
> 
> > and (2) that it is supposed to ignore wakeup signals from certain
> > sources. 
> 
> Isn't this configurable now without system suspend with the wakeup sources 
> code?

The wakeup code does two things: It tells the system which devices 
should be enabled for wakeup when the system sleeps (as opposed to 
cpuidle, when all devices should be able to generate interrupts), and 
it tells the system when a wakeup event has just occurred so that the 
system won't allow itself to go to sleep before the event can be 
processed.

> > As I said, support for system suspend allows the user to do more with the
> > system and you really can't implement that functionality differently.  You
> > may choose not to implement it at all, but why should you?
> 
> The individual components of system suspend are definitely useful:
> 
> 1. preventing userspace processes from keeping the system from entering 
>    idle
> 2. preventing userspace processes from waking the system back up

It's not clear what this means.  How can a user process wake the system 
up?  Are you referring to user timers?

Do we have a way (or want a way) to distinguish between user timers and
kernel timers?

> 3. preventing kernel code from keeping the system from entering idle

Relatively few kernel threads are freezable.  The ones that are tend to 
be involved in device management.  Did you have something specific in 
mind?

> 4. preventing kernel code from waking the system back up
> 5. requesting that drivers abort what they are currently doing

Not exactly "abort".  More like "stop processing their I/O queues".

> The main issues are that these:
> 
> 1. are all bundled up into one operation
> 
> 2. target the entire universe of entities that they manage, rather than 
> selectively targeting non-power-privileged entities

That's an interesting way of thinking about it.  Thanks for bringing 
it up.

Alan Stern