[PATCH 0/3] clk: divider: three exactness fixes (and a rant)

[Uwe Kleine-König]

Hello Stephen,

On Mon, Mar 09, 2015 at 12:05:34PM -0700, Stephen Boyd wrote:
> On 03/09/15 02:58, Philipp Zabel wrote:
> > Am Freitag, den 06.03.2015, 11:40 -0800 schrieb Stephen Boyd:
> >> On 03/06/15 11:28, Uwe Kleine-König wrote:
> >>> Hello Mike,
> >>>
> >>> On Fri, Mar 06, 2015 at 10:57:30AM -0800, Mike Turquette wrote:
> >>>> Quoting Uwe Kleine-König (2015-02-21 02:40:22)
> >>>>> Hello,
> >>>>>
> >>>>> TLDR: only apply patch 1 and rip of CLK_DIVIDER_ROUND_CLOSEST.
> >>>>>
> >>>>> I stared at clk-divider.c for some time now given Sascha's failing test
> >>>>> case. I found a fix for the failure (which happens to be what Sascha
> >>>>> suspected).
> >>>>>
> >>>>> The other two patches fix problems only present when handling dividers
> >>>>> that have CLK_DIVIDER_ROUND_CLOSEST set. Note that these are still
> >>>>> heavily broken however. So having a 4bit-divider and a parent clk of
> >>>>> 10000 (as in Sascha's test case) requesting
> >>>>>
> >>>>>         clk_set_rate(clk, 666)
> >>>>>
> >>>>> sets the rate to 625 (div=15) instead of 667 (div=16). The reason is the
> >>>>> choice of parent_rate in clk_divider_bestdiv's loop is wrong for
> >>>>> CLK_DIVIDER_ROUND_CLOSEST (with and without patch 1). A fix here is
> >>>>> non-trivial and for sure more than one rate must be tested here. This is
> >>>>> complicated by the fact that clk_round_rate might return a value bigger
> >>>>> than the requested rate which convinces me (once more) that it's a bad
> >>>>> idea to allow that. Even if this was fixed for .round_rate,
> >>>>> clk_divider_set_rate is still broken because it also uses
> >>>>>
> >>>>>         div = DIV_ROUND_UP(parent_rate, rate);
> >>>>>
> >>>>> to calculate the (pretended) best divider to get near rate.
> >>>>>
> >>>>> Note this makes at least two reasons to remove support for
> >>>>> CLK_DIVIDER_ROUND_CLOSEST!
> >>>>>
> >>>>> Instead I'd favour creating a function
> >>>>>
> >>>>>         clk_round_rate_nearest
> >>>>>
> >>>>> as was suggested some time ago by Soren Brinkmann and me[1] that doesn't
> >>>> Uwe,
> >>>>
> >>>> Thanks for the fixes. I'm thinking of taking all three for 4.0. I also
> >>>> agree on clk_round_rate_nearest (along with a _ceil and _floor version
> >>>> as well). That's something we can do for 4.1 probably.
> >>> I'd say that we make round_rate the _floor version. I guess in most
> >>> cases that already does the right thing. Also I think 4.1 is very
> >>> ambitious, so my suggestion for 4.1 is:
> >>>
> >>>  - add a WARN_ON_ONCE to clk_round_rate catching calls that return a
> >>>    value bigger than requested.
> >>>  - implement clk_round_rate_nearest using clk_round_rate and the
> >>>    assumption that it returns a value that is <= the requested rate.
> >>>    I think without that there are too many special cases to handle and
> >>>    probably not even a reliable way to determine the nearest rate.
> >>>  - while we're at it tightening the requirements for clk_round_rate
> >>>    let's also specify the expected rounding. I'd vote for the
> >>>    mathematical rounding, that is
> >>>
> >>>    	clk_round_rate(someclk, 333)
> >>>
> >>>    explicitly is allowed to return a rate bigger than 333 as long as it
> >>>    is less than 333.5.
> >>>
> >>> At one point while developing patch 1 I had the dividers fixed for the
> >>> rounding issue. I think I still have that patch somewhere so can post it
> >>> as RFC.
> >>>
> >> Why do we need clk_round_rate_nearest? We have rate constraints now so
> >> drivers should be moving towards requesting a rate that's within a
> >> tolerance they can accept if they even care to enforce anything like
> >> that. Eventually clk_round_rate() returning a value smaller or larger
> >> than what it's called with won't matter as long as what the
> >> implementation does fits within the constraints that consumers specify.
> >> It may even be possible to remove clk_round_rate() as a consumer API.
> > If I have to provide a panel pixel clock I usually want to get a rate as
> > close as possible to the specified typical rate, but within the
> > specified limits.
> >
> > Assume a panel with 70 MHz ideal pixel clock and a valid range of 60 MHz
> > to 80 MHz. If the clock supply supports two frequencies within that
> > range, 60 MHz and 72 MHz, I'd prefer 72 MHz to be chosen over 60 Mhz.
> >
> >
> 
> Hm.. Maybe we should tweak the arguments to clk_set_range() to have a
> "typical" rate? So instead of the current API:
> 
>   int clk_set_rate_range(struct clk *clk, unsigned long min, unsigned
> long max)

Looking at the current implementation of clk_set_rate_range I wonder how
it should be used. The "usual" code flow goes through

	ret = clk_core_set_rate_nolock(clk->core, clk->core->req_rate);

without modifying clk->core->req_rate first. So after

	clk_set_rate(clk, 50);

a call to

	clk_set_rate_range(clk, 100, 200);

returns -EINVAL (at least if determine_rate isn't implemented). Is
this intended? This even happens if the clk could provide a rate between
100 and 200. Also if I first call set_rate_range and then set_rate, the
clk's members for min and max are unaffected by the latter call, right?

> we should have
> 
>  int clk_set_rate_range(struct clk *clk, unsigned long min, unsigned
> long typical, unsigned long max)
> 
> with the semantics that we'll set the rate within the min,max
> constraints and try to get the rate as close to the typical rate as
> possible? That would match quite a few datasheets out there that specify
> these triplets.
And according to which metric do you determine "close"? Consider a clock
that can provide 900 Hz and 1110 Hz (but nothing in between). Which one
do you consider closer if 1000 Hz are requested?
On the first look 900 Hz looks better because

	abs(1000 - 900) < abs(1000 - 1110)

but if you look at the respective cycle times we get:

	1 / 900 Hz = 1.111 ms
	1 / 1110 Hz = 0.901 ms

so 1110 Hz is better according to the metric

	abs(1 / 1000 - 1 / 1110) < abs(1 / 1000 - 1 / 900)

. Maybe even other metrics might be sensible for certain scenarios[1]?
That convinces me that it's better to be able to query a clock for
possible rates (.round_rate) and get the logic to determine "close" into
a single helper (well, one per metric) in the clk core. This is much
better because then you have a single complex function instead of one
per clk implementation.

And then to keep the complexity at a sane level it would be great if you
could assume e.g. that .round_rate always rounds down.

Best regards
Uwe

[1] I first thought about minimizing abs(1 - rate / optrate), but that
    is equivalent to minimizing abs(optrate - rate).

-- 
Pengutronix e.K.                           | Uwe Kleine-König            |
Industrial Linux Solutions                 | http://www.pengutronix.de/  |