On 03/23/2016 10:56 AM, Sekhar Nori wrote:

On Thursday 17 March 2016 07:56 AM, David Lechner wrote:

quoted

The da850 family of processors has an async3 clock domain that can be
muxed to either pll0_sysclk2 or pll1_sysclk2. Now that the davinci clocks
have a set_parent callback, we can use this to control the async3 mux
instead of a stand-alone function.

This adds a new async3_clk and sets the appropriate child clocks. The
default is use to pll1_sysclk2 since it is not affected by processor
frequency scaling.

Signed-off-by: David Lechner <david@lechnology.com>
---

quoted

+static int da850_async3_set_parent(struct clk *clk, struct clk *parent)
+{
+	u32 __iomem *cfgchip3;
+	u32 val;
+
+	/*
+	 * Can't use DA8XX_SYSCFG0_VIRT() here since this can be called before
+	 * da8xx_syscfg0_base is initialized.
+	 */
+	cfgchip3 = ioremap(DA8XX_SYSCFG0_BASE + DA8XX_CFGCHIP3_REG, 4);

Is this just a theoretical possibility or have you seen this happen? I
would like to see if there are ways of avoiding this rather than throw
away usage of DA8XX_SYSCFG0_VIRT()

Yes, it will not boot without this. The problem comes from the fact that 
clocks are setup in davinci_common_init() which is called before 
da8xx_syscfg0_base = ioremap(DA8XX_SYSCFG0_BASE, SZ_4K) in da850_init() 
(and da830_init()). I also tried moving the ioremap() before 
davinci_common_init(), but davinci_common_init() sets up the iomem, so 
that doesn't work either.

So, if you want to use DA8XX_SYSCFG0_VIRT() here, the clock init would 
have to be split out from davinci_common_init() which would affect all 
davinci devices.