Re: [PATCH] clk: divider: Use DIV_ROUND_CLOSEST

["Sören Brinkmann" <soren.brinkmann@xilinx.com>]

On Tue, Mar 26, 2013 at 06:37:03PM -0700, Mike Turquette wrote:
> Quoting Sören Brinkmann (2013-03-26 15:45:22)
> > On Thu, Mar 21, 2013 at 10:15:31AM +0100, Uwe Kleine-König wrote:
> > > Hello,
> > > 
> > > On Wed, Mar 20, 2013 at 07:50:51PM +0100, Sascha Hauer wrote:
> > > > On Wed, Mar 20, 2013 at 09:32:51AM -0700, Sören Brinkmann wrote:
> > > > > If the caller
> > > > > doesn't like the returned frequency he can request a different one.
> > > > > And he's eventually happy with the return value he calls
> > > > > clk_set_rate() requesting the frequency clk_round_rate() returned.
> > > > > Always rounding down seems a bit odd to me.
> > > > > 
> > > > > Another issue with the current implmentation:
> > > > > clk_divider_round_rate() calls clk_divider_bestdiv(), which uses the ROUND_UP macro, returning a rather low frequency.
> > > > 
> > > > And that is correct. clk_divider_bestdiv is used to calculate the
> > > > maximum parent frequency for which a given divider value does not
> > > > exceed the desired rate.
> > > The reason for that is that the (more?) usual constraint is like: This
> > > mmc card can handle up to 100 MHz. Or this i2c device can handle up to
> > > this and that frequency. Of course there are different constraints, e.g.
> > > for a UART if the target baud speed is 38400 you better run at 38402
> > > than at 19201.
> > > 
> > > I wonder if it depends on the clock if you want "best approximation <=
> > > requested value" or "best approximation" or on the caller. In the former
> > > case a flag for the clock would be the right thing (as suggested in this
> > > thread). If however it's the caller of round_rate who knows better which
> > > rounding is preferred than better extend the clk API.
> > > 
> > > Extending the API could just be a convenience function that doesn't
> > > affect the implementations of the clk API. E.g.:
> > > 
> > >       long clk_round_rate_nearest(struct clk *clk, unsigned long rate)
> > >       {
> > >               long lower_limit = clk_round_rate(clk, rate);
> > >               long upper_limit = clk_round_rate(clk, rate + (rate - lower_limit));
> > > 
> > >               if (rate - lower_limit < upper_limit - rate)
> > >                       return lower_limit;
> > >               else
> > >                       return upper_limit;
> > >       }
> > > 
> > I guess both approaches may work. Anybody has a preference?
> > 
> 
> A dedicated function like the one Uwe defined is better than adding
> subtlety to the existing clk_round_rate via a flag in a clock driver.
I looked at my problem again.

A new API function is probably fine for UART, ethernet drivers and
similar. Although, compared to a flag it would add some redundant
rounding, since clk_set_rate() implicitly also rounds the rate.
	clk_set_rate()
		clk_calc_new_rates()
			clk_round_rate()
But that is true for every driver which doesn't blindly call
clk_set_rate() and checks upfront through clk_round_rate() what
the actual frequency would look like.

So, do we agree to add this additional clk_round_rate_nearest()
function?
And if, should I just make Uwe's proposal another patch, additionally to
the other clk-divider change I'm working on?
Or Uwe, do you prefer to submit it yourself?


For my original problem, though, this is only part of a solution. It
appeared to be a rounding issue, but the actual root cause is the loss
of resolution when OPPs are converted to a frequency table for cpufreq.
I'm not sure how this can be resolved, yet.


	Sören