[Intel-gfx] [PATCH] drm/i915: use udelay for very short delays

Ville Syrjälä ville.syrjala at linux.intel.com
Thu Dec 15 11:51:43 UTC 2016


On Thu, Dec 15, 2016 at 10:34:00AM +0000, Nicholas Mc Guire wrote:
> On Thu, Dec 15, 2016 at 12:20:01PM +0200, Jani Nikula wrote:
> > On Thu, 15 Dec 2016, Ville Syrjälä <ville.syrjala at linux.intel.com> wrote:
> > > On Thu, Dec 15, 2016 at 11:52:34AM +0200, Jani Nikula wrote:
> > >> On Thu, 15 Dec 2016, Nicholas Mc Guire <der.herr at hofr.at> wrote:
> > >> > On Thu, Dec 15, 2016 at 11:08:49AM +0200, Jani Nikula wrote:
> > >> >> On Thu, 15 Dec 2016, Nicholas Mc Guire <hofrat at osadl.org> wrote:
> > >> >> > Even on fast systems a 2 microsecond delay is most likely more efficient
> > >> >> > as a busy-wait loop. The overhead of a hrtimer does not seem warranted -
> > >> >> > change this to a udelay(2).
> > >> >> 
> > >> >> Similar concerns as in [1]. We don't need the accuracy of udelay() here,
> > >> >> so this boils down to which is the better use of CPU. We could probably
> > >> >> relax the max delay more if that was helpful. But I'm not immediately
> > >> >> sold on "is most likely more efficient" which sounds like a gut feeling.
> > >> >> 
> > >> >> I'm sorry it's not clear in my other reply that I do appreciate
> > >> >> addressing incorrect/silly use of usleep_range(); I'm just not (yet)
> > >> >> convinced udelay() is the answer.
> > >> >
> > >> > if the delay is not critical and all that is needed 
> > >> > is an assurance that it is greater than X us then 
> > >> > usleep_range is fine with a relaxed limit. 
> > >> > So from what you wrote my patch proposal is wrong - the
> > >> > udelay() is not the way to got.
> > >> > My intent is to get rid of very small usleep_range() cases
> > >> > so if usleep_range(20,50) causes no issues with this driver
> > >> > and does not induce any performance penalty then that would 
> > >> > be the way to go I think.
> > >> 
> > >> Okay, so I looked at the code, and I looked at our spec, and I looked at
> > >> the MIPI D-PHY spec, and I cried a little.
> > >> 
> > >> Long story short, I think usleep_range(10, 50) will be fine.
> > >
> > > Note that I really want to see a comment next to every delay like this
> > > documenting the actual hardware requirement, if the delay used by the
> > > code doesn't 100% match it.
> > 
> > Our spec says, "Wait for 2us for ULPS to complete". That's a simplistic
> > view wrt D-PHY, and our code doesn't even match the spec. Hence the
> > tears. Want to propose a wording for the comment so we can apply this
> > change, without going for a full rewrite of the sequence?
> >
> is that suitable or am I overdoing it ?
> 
> -               usleep_range(2, 3);
> +               /* delay for at least 2us - relaxed to 10-50 to allow
> +                * hrtimer subsystem to optimize uncritical timer handling
> +                */

That's entirely too verbose IMO, and the reason for using usleep_range()
is pretty obvious without spelling it out.

All we really want to know is what the spec says is the minimum
acceptable delay.

> +               usleep_range(10, 50);
> 
> thx!
> hofrat 

-- 
Ville Syrjälä
Intel OTC


More information about the Intel-gfx mailing list