[PATCH] clk: mediatek: Add MT8173 MMPLL change rate support

Heiko Stübner heiko at sntech.de
Tue Jul 7 02:34:17 PDT 2015


Hi James,

Am Dienstag, 7. Juli 2015, 17:28:38 schrieb James Liao:
> On Tue, 2015-07-07 at 10:58 +0200, Heiko Stübner wrote:
> > > +#define PLL(_id, _name, _reg, _pwr_reg, _en_mask, _flags, _pcwbits,	\
> > > +			_pd_reg, _pd_shift, _tuner_reg, _pcw_reg,	\
> > > +			_pcw_shift)					\
> > > +		PLL_B(_id, _name, _reg, _pwr_reg, _en_mask, _flags, _pcwbits, \
> > > +			_pd_reg, _pd_shift, _tuner_reg, _pcw_reg, _pcw_shift, \
> > > +			NULL)
> > > +
> > > +const unsigned long mmpll_div_rate[] = {
> > 
> > static?
> 
> OK. I'll add it.
> 
> > > +	MT8173_PLL_FMAX,
> > > +	1000000000,
> > > +	702000000,
> > > +	253500000,
> > > +	126750000,
> > > +	0,
> > 
> > it's more common to label sentinel entries (the 0 marking the end) with
> > 
> > 	/* sentinel */
> > 
> > instead of value 0.
> 
> OK. I'll add it.
> 
> > If I'm reading the code correctly, this is a mapping divider -> frequency,
> > right? So it may be nice to make this a bit more explicit, like:
> > 
> > struct mtk_pll_div_table {
> > 
> > 	unsigned int    freq;
> > 	unsigned int    div;
> > 
> > };
> > 
> > static const struct mtk_pll_div_table mmpll_div_rate[] = {
> > 
> > 	{ .freq = MT8173_PLL_FMAX, .div = 0 },
> > 	{ .freq = 1000000000, .div = 1 },
> > 	{ .freq = 702000000, .div = 2 },
> > 	{ .freq = 253500000, .div = 3 },
> > 	{ .freq = 126750000, .div = 4 },
> > 	{ /* sentinel */ },
> > 
> > };
> 
> Hmm.. OK. I'll try to use a more readable way to implement it.
> 
> > -------------
> > 
> > > -	u32 con1, pd, val;
> > > +	u32 con1, val;
> > > 
> > >  	int pll_en;
> > > 
> > > -	/* set postdiv */
> > > -	pd = readl(pll->pd_addr);
> > > -	pd &= ~(POSTDIV_MASK << pll->data->pd_shift);
> > > -	pd |= (ffs(postdiv) - 1) << pll->data->pd_shift;
> > > -	writel(pd, pll->pd_addr);
> > > -
> > > 
> > >  	pll_en = readl(pll->base_addr + REG_CON0) & CON0_BASE_EN;
> > > 
> > > -	/* set pcw */
> > > -	val = readl(pll->pcw_addr);
> > > +	/* set postdiv */
> > > +	val = readl(pll->pd_addr);
> > > +	val &= ~(POSTDIV_MASK << pll->data->pd_shift);
> > > +	val |= (ffs(postdiv) - 1) << pll->data->pd_shift;
> > > +
> > > +	/* postdiv and pcw need to set at the same time if on same register 
*/
> > > +	if (pll->pd_addr != pll->pcw_addr) {
> > > +		writel(val, pll->pd_addr);
> > > +		val = readl(pll->pcw_addr);
> > > +	}
> > > 
> > > +	/* set pcw */
> > > 
> > >  	val &= ~GENMASK(pll->data->pcw_shift + pll->data->pcwbits - 1,
> > >  	
> > >  			pll->data->pcw_shift);
> > >  	
> > >  	val |= pcw << pll->data->pcw_shift;
> > 
> > This whole block probably wants to be a separate patch ;-) .
> > 
> > While it may not affect previous pll implementations, it changes how
> > register accesses are handled and should not hide in another patch.
> 
> OK, I'll separate it.
> 
> > > @@ -135,16 +138,26 @@ static void mtk_pll_calc_values(struct mtk_clk_pll
> > > *pll, u32 *pcw, u32 *postdiv, u32 freq, u32 fin)
> > > 
> > >  {
> > >  
> > >  	unsigned long fmin = 1000 * MHZ;
> > > 
> > > +	const unsigned long *div_rate = pll->data->div_rate;
> > > 
> > >  	u64 _pcw;
> > >  	u32 val;
> > >  	
> > >  	if (freq > pll->data->fmax)
> > >  	
> > >  		freq = pll->data->fmax;
> > > 
> > > -	for (val = 0; val < 4; val++) {
> > > +	if (div_rate) {
> > > +		for (val = 1; div_rate[val] != 0; val++) {
> > > +			if (freq > div_rate[val])
> > > +				break;
> > > +		}
> > > +		val--;
> > 
> > if you're changing the table struct, this of course also would need to be
> > adapted.
> > 
> > 
> > Hmm, what I don't understand is, what does MT8173_PLL_FMAX in the table,
> > if
> > you ignore it here all the time?
> > 
> > So the table should probably look more like [when using the concept from
> > above]
> > 
> > static const struct mtk_pll_div_table mmpll_div_rate[] = {
> > 
> > 	{ .freq = 1000000000, .div = 0 },
> > 	{ .freq = 702000000, .div = 1 },
> > 	{ .freq = 253500000, .div = 2 },
> > 	{ .freq = 126750000, .div = 3 },
> > 	{ /* sentinel */ },
> > 
> > };
> 
> The freq-div table describes the maximum frequency of each divider
> setting. Although the first element doesn't used in current
> implementation, I think it's better to keep freq-div table's
> completeness.

the issue I see is, that its value is currently 0 and the code substracts 1. 
So if anything would (accidentially) select MT8173_PLL_FMAX, the u32 val would 
wrap around, as you're subtracting 1 from 0 .


Heiko



More information about the Linux-mediatek mailing list