lkml.org 
[lkml]   [2009]   [Jul]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [RFC][PATCH] PM: Introduce core framework for run-time PM of I/O devices (rev. 8)
Date
On Tuesday 07 July 2009, Magnus Damm wrote:
> Hi Rafael,
>
> On Mon, Jul 6, 2009 at 9:52 AM, Rafael J. Wysocki<rjw@sisk.pl> wrote:
> > Hi,
> >
> > There's a rev. 8 of the run-time PM framework patch.
> >
> > Highlights:
> > * I did my best to follow the design we've recently discussed.
> > * pm_runtime_[get|put]() and the sync versions call
> > pm_[request|runtime]_[resume|idle](), because I don't see much point
> > manipulating the usage counter alone.
> > * pm_runtime_disable() carries out a (synchronous) wake-up if there's a
> > resume request pending.
> >
> > Comments welcome.
>
> I've now jumped from v5 to v8 and I feel that the code is getting
> cleaner and cleaner. Very nice.

That's mostly thanks to Alan.

> My intention was to post a SuperH prototype last week, but I got side
> tracked with other stuff. And today I ran into some problems related
> to probe() that I'd like to ask about right away. At this point I've
> got a few device drivers converted and some simple bus
> runtime_suspend()/runtime_resume() code that stop and start clocks.
>
> Issue 1:
> ------------
> Device drivers which do not perform any hardware access in probe()
> work fine. During software setup in probe() the runtime pm code is
> initialized with the following:
>
> + pm_suspend_ignore_children(&dev->dev, true);
> + pm_runtime_set_suspended(&dev->dev);
> + pm_runtime_enable(&dev->dev);
>
> Before accessing hardware I perform:
> + pm_runtime_resume(pd->dev);
>
> When done with the hardware I do:
> + pm_runtime_suspend(pd->dev);
>
> Not so complicated. Am I supposed to initialize something else as well?
>
> All good with the code above, but there seem to be some issue with how
> usage_count is counted up and down and when runtime_disabled is set:
>
> 1. pm_runtime_init(): usage_count = 1, runtime_disabled = true
> 2. driver_probe_device(): pm_runtime_get_sync()
> 3. pm_runtime_get_sync(): usage_count = 2
> 4. device driver probe(): pm_runtime_enable()
> 5. pm_runtime_enable(): usage_count = 1
> 6. driver_probe_device(): pm_runtime_put()
> 7. pm_runtime_put(): usage_count = 0
>
> I expect runtime_disabled = false in 7. Modifying the get/put calls to
> do enable/disable may work around the issue, but that's probably not
> what you guys want.

Sure, that's my mistake. I should have used a separate counter for
disable/enable, but I thought usage_counter would be sufficient. Will fix.

> Issue 2:
> ------------
> I cannot get any bus ->runtime_resume() callbacks from probe(). This
> also seems related to usage_count and pm_runtime_get_sync() in
> driver_probe_device(). Basically, from probe(), calling
> pm_runtime_resume() after pm_runtime_set_suspended() results in error
> and not in a ->runtime_resume() callback. Some device drives access
> hardware in probe(), so the ->runtime_resume() callback is needed at
> that point to turn on clocks before the hardware can be accessed.

I think the problem is that pm_runtime_get_sync() in driver_probe_device()
calls ->runtime_resume(), so the device is active from the core's point of
view when you call pm_runtime_resume() from probe().

Hmm. OK, perhaps we should just increment usage_count in
driver_device_probe() to prevent suspends from happening at that time, without
calling ->runtime_resume() so that the driver can do it by itself. I'll do
that in the next version.

> Random thought:
> -------------------------
> The runtime_pm_get() and runtime_pm_put() look very nice. I assume
> that inteface is supposed to be used by bus code. I wonder if it would
> be cleaner to use a similar counter based interface from the driver
> instead of the pm_runtime_idle()/suspend()/resume()...
>
> Let me know what you think!

In fact I thought drivers could also use pm_runtime_[get|put]() and the 'sync'
versions. At least, I don't see why not at the moment (well, I'm a bit tired
right now ...).

However, I'm now thinking it should work like this:

* pm_runtime_get() increments usage_count and if it was zero before the
incrementation, it calls pm_request_resume() (pm_runtime_resume() is called
by the 'sync' version).

* pm_runtime_put() decrements usage_count and if it's zero after the
decrementation, it calls pm_request_idle() (pm_runtime_idle() is called by
the 'sync' version).

* The 'suspend' callbacks won't succeed for usage_count > 0.

This way we would avoid calling the 'suspend' and 'idle' functions each time
unnecessarily, but then usage_count would have to be modified under the
spinlock only.

Best,
Rafael


\
 
 \ /
  Last update: 2009-07-08 00:11    [W:0.191 / U:0.112 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site