Re: Choosing between DELAY(useconds) and pause()

On Tue, Sep 27, 2011 at 10:39:44AM -0600, Julian Elischer wrote:
> On 9/27/11 4:12 AM, Gavin Atkinson wrote:
> >On Mon, 2011-09-26 at 09:30 -0400, John Baldwin wrote:
> >>On Friday, September 23, 2011 11:21:06 am Gavin Atkinson wrote:
> >>>On Thu, 2011-09-22 at 20:07 +0200, Hans Petter Selasky wrote:
> >>>>On Thursday 22 September 2011 19:55:23 David Somayajulu wrote:
> >>>>>It appears that the pause() function cannot be used in driver functions
> >>>>>which are invoked early in the boot process. Is there is a kernel api
> >>>>>which a device driver can use to determine whether to use pause() or
> >>>>>DELAY(), for delays which are say greater than 10hz - may be even 1 hz 
> >>>>>?
> >>>>Maybe you want to use something like this:
> >>>>
> >>>>if (cold)
> >>>>  DELAY()
> >>>>else
> >>>>  pause()
> >>>>
> >>>>In your code.
> >>>Note that this still shouldn't be done in your suspend/resume paths, as
> >>>"cold" isn't set there, however there also appears to be no guarantee
> >>>that pause() will ever return (as you could be running after the timer
> >>>has been suspended, or before it resumes).
> >>>
> >>>I'm not sure what the correct answer is for suspend/resume code.
> >>Hmmm, on x86 the timers are explicitly shutdown after the DEVICE_SUSPEND()
> >>pass over the tree and re-enabled before DEVICE_RESUME().  Perhaps this 
> >>has
> >>changed in HEAD though with the eventtimers stuff.  I do think it is best
> >>however, to use DELAY() in the suspend/resume path always regardless.
> >I don't think head is any different from stable/8 in this respect - the
> >same hack patch that fixes suspend/resume for me on head also fixes it
> >on stable/8 (the patch basically fakes "cold" during USB
> >suspend/resume).  See my email to -usb a few months ago:
> >http://docs.FreeBSD.org/cgi/mid.cgi?alpine.LNX.2.00.1106041548370.26975
> >
> >I'd really like some guidance as to the correct solution to this, I have
> >four separate laptops which fail out of the box on 8 and 9, but
> >suspend/resume perfectly with this hack.
> 
> code for timers should have a generally readable state that says if 
> they are useable or not, and we should test that instead of 'cold'

I think that this should be encapsulated into the usable function,
instead of being directly exposed to the driver authors.
With the my lack of imagination, I propose driver_delay() that
would do if (cold) ... inside.