Re: [RFC][PATCH 0/7] PM: Asynchronous suspend and resume (updated)

[Rafael J. Wysocki]

On Tuesday 18 August 2009, Alan Stern wrote:
> On Mon, 17 Aug 2009, Rafael J. Wysocki wrote:
> 
> > On Wednesday 12 August 2009, Rafael J. Wysocki wrote:
> > > Hi,
> > > 
> > > The following patches introduce a mechanism allowing us to execute device
> > > drivers' suspend and resume callbacks asynchronously during system sleep
> > > transitions, such as suspend to RAM.  The idea is explained in the [1/1] patch
> > > message.
> > 
> > Changes:
> > 
> > * Added [1/7] that fixes kerneldoc comments in drivers/base/power/main.c
> >   (this is a 2.6.32 candidate).
> > 
> > * Added [2/7] adding a framework for representing PM link (idea described
> >   in the patch message).
> > 
> > * [3/7] is the async resume patch (idea described in the patch message).
> > 
> > * [4/7] is the async suspend patch.
> > 
> > * [5/7] - [7/7] set async_suspend for devices in a few selected subsystems.
> > 
> > The patches have been tested on HP nx6325.
> > 
> > Comments welcome.
> 
> I'm not sure about the design of these things.  How much do we care 
> about wasting memory?

Not much.

> Your scheme allocates six pointers for every dependency, plus four pointers
> for every device.

Yes, it does.

> It's possible to reduce this considerably, especially if the parent-child
> dependencies aren't stored explicitly.

Yes, at the expense of increased complexity and reduced performance.

> If you decide to keep this scheme, you should make pm_link_add() check 
> for duplicate dependencies before adding them.

That's correct.

> Also, I think a better approach to the async execution would not
> require adding a struct completion to each device and making each async
> thread wait for the completion to be signalled.  Instead, have a single
> master thread (i.e., the thread doing the suspend) monitor the
> dependencies and have it farm the devices out to async threads as they
> become ready to be suspended or resumed.

Do you mean that the master thread should check the dependencies
_before_ executing, for example, __device_resume() and execute it
asynchronously only if they are already satisfied?  In that case we might lose
the opportunity to save some time.

For example, assume devices A and B depend on C. Say that normally, A would be
handled before B, so if C hasn't finished yet, the A's callback will be
executed synchronously.  Now, if both A and B take time T to complete the
callback and C finishes dT after we've called A synchronously, we'll lose the
chance to save T - dT by handling A and B in parallel.

The master thread might chose another device for asynchronous execution, but
then it should revisit A and B and that still is going to be suboptimal
time-wise in some specific situations (eg. A and B are the last two devices to
handle).

> Finally, devices that don't have async_suspend set should implicitly 
> depend on everything that comes after them (for suspend) or before them 
> (for resume) in the device list.

They do, through dpm_list.

Thanks,
Rafael
--
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/