Hi,

On Sun, May 22, 2016 at 12:43:48AM +0800, Chen-Yu Tsai wrote:
> > +static int ccu_phase_set_phase(struct clk_hw *hw, int degrees)
> > +{
> > +       struct ccu_phase *phase = hw_to_ccu_phase(hw);
> > +       struct clk_hw *parent, *pparent;
> > +       unsigned int parent_rate, pparent_rate;
> 
> grandparent(_rate) would be easier to understand.

Ack.

> 
> > +       unsigned long flags;
> > +       u32 reg;
> > +       u8 delay;
> > +
> > +       /* Get our parent clock, it's the one that can adjust its rate */
> > +       parent = clk_hw_get_parent(hw);
> > +       if (!parent)
> > +               return -EINVAL;
> > +
> > +       /* And its rate */
> > +       parent_rate = clk_hw_get_rate(parent);
> > +       if (!parent_rate)
> > +               return -EINVAL;
> > +
> > +       /* Now, get our parent's parent (most likely some PLL) */
> > +       pparent = clk_hw_get_parent(parent);
> > +       if (!pparent)
> > +               return -EINVAL;
> > +
> > +       /* And its rate */
> > +       pparent_rate = clk_hw_get_rate(pparent);
> > +       if (!pparent_rate)
> > +               return -EINVAL;
> > +
> > +       if (degrees != 180) {
> > +               u16 step, parent_div;
> > +
> > +               /* Get our parent divider */
> > +               parent_div = pparent_rate / parent_rate;
> > +
> > +               /*
> > +                * We can only outphase the clocks by multiple of the
> > +                * PLL's period.
> > +                *
> > +                * Since our parent clock is only a divider, and the
> > +                * formula to get the outphasing in degrees is deg =
> > +                * 360 * delta / period
> > +                *
> > +                * If we simplify this formula, we can see that the
> > +                * only thing that we're concerned about is the number
> > +                * of period we want to outphase our clock from, and
> > +                * the divider set by our parent clock.
> > +                */
> > +               step = DIV_ROUND_CLOSEST(360, parent_div);
> > +               delay = DIV_ROUND_CLOSEST(degrees, step);
> 
> Doesn't this mean some delay values are impossible to set?
> 
> For instance, for PLL = 600 MHz and this clock = 50 MHz, div would be 12,
> and a step would be 30 degrees. This means we can't ask for a delay of 6,
> which is 180 degrees.
> 
> For PLL = 600 MHz, clock = 100 MHz, div would be 6, and a step is 60
> degrees. Therefor we can't ask for a delay of 3.

You don't ask for a delay, you ask for an outphasing in degrees. In
the hardware, in the register 0 means an outphasing of 180 degrees
(and this has been confirmed by Allwinner a while back). In the two
cases you point out, we would have two ways of achieving the same
thing, we prefer one over another, but I don't see how it's
problematic.

It's also a direct copy of the current code we have, which didn't
raise any objection, or had any known bugs.

> > +struct ccu_phase {
> > +       u8                      shift;
> > +       u8                      width;
> 
> Not sure why you used struct ccu_factor in the divider table clock,
> but separate fields directly in ccu_phase here.

Because this is not meant for the same thing. ccu_factor is probably
going to go away anyway because of the dividers consolidation.

Thanks!
Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com