Re: [PATCH 0/2] Run callback of device_prepare/complete consistently

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Saturday, June 08, 2013 10:37:18 AM Yanmin Zhang wrote:
> On Sat, 2013-06-08 at 03:52 +0200, Rafael J. Wysocki wrote:
> > On Saturday, June 08, 2013 09:36:03 AM Yanmin Zhang wrote:
> > > On Sat, 2013-06-08 at 03:30 +0200, Rafael J. Wysocki wrote:
> > > > On Friday, June 07, 2013 06:16:25 PM Greg KH wrote:
> > > > > On Sat, Jun 08, 2013 at 08:42:12AM +0800, Yanmin Zhang wrote:
> > > > > > On Fri, 2013-06-07 at 12:36 +0200, Rafael J. Wysocki wrote:
> > > > > > > On Friday, June 07, 2013 04:20:30 PM shuox.liu@intel.com wrote:
> > > > > > > > dpm_run_callback is used in other stages of power states changing.
> > > > > > > > It provides debug info message and time measurement when call these
> > > > > > > > callback. We also want to benefit ->prepare and ->complete.
> > > > > > > > 
> > > > > > > > [PATCH 1/2] PM: use dpm_run_callback in device_prepare
> > > > > > > > [PATCH 2/2] PM: add dpm_run_callback_void and use it in device_complete
> > > > > > > 
> > > > > > > Is this an "Oh, why don't we do that?" series, or is it useful for anything
> > > > > > > in practice?  I'm asking, because we haven't added that stuff to start with
> > > > > > > since we didn't see why it would be useful to anyone.
> > > > > > > 
> > > > > > > And while patch [1/2] reduces the code size (by 1 line), so I can see some
> > > > > > > (tiny) benefit from applying it, patch [2/2] adds more code and is there any
> > > > > > > paractical reason?
> > > > > > Sometimes, suspend-to-ram path spends too much time (either suspend slowly
> > > > > > or wakeup slowly) and we need optimize it.
> > > > > > With the 2 patches, we could collect initcall_debug printk info and manually
> > > > > > check what prepare/complete callbacks consume too much time.
> > > > > 
> > > > > But initcall information is for initialization stuff, not suspend/resume
> > > > > things, right?  Doesn't the existing tools for parsing this choke if it
> > > > > sees the information at suspend/resume time?
> > > > 
> > > > We've been using that for suspend/resume for quite some time too, but not
> > > > for the prepare/complete phases (because we still believe that's not really
> > > > useful for them). 
> > > > 
> > > > Well, I'll be handling patches changing code under drivers/base/power,
> > > > I promise. :-)
> > > > 
> > > > I've been doing that for quite a few years now ...
> > > Yes, indeed. Power is one of the most important features on embedded devices.
> > > Lots of smart phones don't really go through the full cycles of suspend-to-ram.
> > > We are following the full steps of the suspend.
> > 
> > But if you go through the code, you'll see that alomost no drivers actually
> > implemet .prepare() and .complete().  Some subsystems do, but they really don't
> > take too much time to execute.  Which means that your patches with
> > initcall_debug will add quite a pile of useless garbage to the kernel log
> Does that mean we need add more log levels around such info instead of just having or
> not having?

Since we don't have any code in the tree that causes problems those patches are
supposed to catch, I don't see why we need them in the tree.  Would it be
viable to keep them out of the tree for the time being and re-submit once
there is real need?

Rafael


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.