On Wed, Aug 25, 2021 at 05:03:30PM +0200, Jernej Škrabec wrote:
> Dne sreda, 25. avgust 2021 ob 16:50:27 CEST je Maxime Ripard napisal(a):
> > Hi,
> > 
> > On Fri, Aug 20, 2021 at 06:34:38AM +0200, Jernej Škrabec wrote:
> > > > > +static void __init sun50i_r329_r_ccu_setup(struct device_node *node)
> > > > > +{
> > > > > +	void __iomem *reg;
> > > > > +	u32 val;
> > > > > +	int i;
> > > > > +
> > > > > +	reg = of_io_request_and_map(node, 0, of_node_full_name(node));
> > > > > +	if (IS_ERR(reg)) {
> > > > > +		pr_err("%pOF: Could not map clock registers\n", node);
> > > > > +		return;
> > > > > +	}
> > > > > +
> > > > > +	/* Enable the lock bits and the output enable bits on all PLLs */
> > > > > +	for (i = 0; i < ARRAY_SIZE(pll_regs); i++) {
> > > > > +		val = readl(reg + pll_regs[i]);
> > > > > +		val |= BIT(29) | BIT(27);
> > > > > +		writel(val, reg + pll_regs[i]);
> > > > > +	}
> > > > > +
> > > > > +	/*
> > > > > +	 * Force the I/O dividers of PLL-AUDIO1 to reset default value
> > > > > +	 *
> > > > > +	 * See the comment before pll-audio1 definition for the reason.
> > > > > +	 */
> > > > > +
> > > > > +	val = readl(reg + SUN50I_R329_PLL_AUDIO1_REG);
> > > > > +	val &= ~BIT(1);
> > > > > +	val |= BIT(0);
> > > > > +	writel(val, reg + SUN50I_R329_PLL_AUDIO1_REG);
> > > > > +
> > > > > +	i = sunxi_ccu_probe(node, reg, &sun50i_r329_r_ccu_desc);
> > > > > +	if (i)
> > > > > +		pr_err("%pOF: probing clocks fails: %d\n", node, i);
> > > > > +}
> > > > > +
> > > > > +CLK_OF_DECLARE(sun50i_r329_r_ccu, "allwinner,sun50i-r329-r-ccu",
> > > > > +	       sun50i_r329_r_ccu_setup);
> > > > 
> > > > Please make this a platform driver. There is no particular reason why it
> > > > needs to be an early OF clock provider.
> > > 
> > > Why? It's good to have it as early clock provider. It has no dependencies
> > > and other drivers that depends on it, like IR, can be deferred, if this
> > > is loaded later.
> > 
> > No, Samuel is right, we should make them regular drivers as much as we
> > can.
> > 
> > The reason we had CLK_OF_DECLARE in the first place is that timers
> > usually have a parent clock, and you need the timers before the device
> > model is set up.
> > 
> > Fortunately for us, since the A20, the architected timers don't require
> > a parent clock from us, and we can thus boot up fine.
> 
> There are other timers. A lot of SoCs, newer than A20 (like H6), have High 
> Speed Timer, which requires parent clock to be enabled. We just choose not to 
> add node for it to DT, even if it's there and driver already exists.

Yeah, I know. The thing is, we just need one timer in order to boot to
the point where the DM is there. We can totally have a timer driver
probing just like any other driver, through the DM, later on.

Maxime