Re: [PATCH v11 06/12] pwm: imx27: Use 64-bit division macro and function

From: Arnd Bergmann
Date: Fri Apr 03 2020 - 15:13:40 EST


On Fri, Apr 3, 2020 at 7:37 PM Guru Das Srinagesh <gurus@xxxxxxxxxxxxxx> wrote:
>
> On Thu, Apr 02, 2020 at 11:16:22PM +0200, Arnd Bergmann wrote:
> > This looks correct, but very expensive, and you don't really have to
> > go this far, given that c1 is guaranteed to be a 32-bit number, and
> > you divide by a constant in the end.
> >
> > Why not do something like
> >
> > #define SHIFT 41 /* arbitrarily picked, not too big, not too small */
> > #define MUL 2199 /* 2^SHIFT / NSEC_PER_SEC */
> > period_cycles = clk_get_rate(imx->clk_per) * ((state->period * MUL) >> SHIFT);
>
> I have two concerns with this:
>
> 1. This actually results in the division by 1000010575.5125057 instead
> of NSECS_PER_SEC whereas both the existing as well as the proposed logic
> divide exactly by NSECS_PER_SEC.
> 2. What method shall be used to pick the SHIFT value? How is this to be
> chosen?

I picked one that would result in a fairly accurate rounding and did not
seem too likely to overflow.

> Also, this seems sort of similar to my initial attempt at this
> problem, where period was being pre-divided prior to the multiplication,
> which was (rightly) NACKed.
>
> c *= div_u64(state->period, 1000000000);

Ah, right, I did make the same mistake here. What I actually indended
was to do two separate shifts that add up to 41. I suppose there is also still
the fundamental problem that without bounds checking the calculation can
always overflow for large inputs.

Arnd