On Thu, Oct 07, 2010 at 11:36:07AM +0800, Yong Shen wrote:
Hi Sascha,
Thanks for your thorough comments. I have already received comments from Arnd before yours arrived. So some of the commends you two provided are common. I acknowledge most of your opinions, except for two, I have some explanations.
- */
- reg = __raw_readl(MXC_CCM_CACRR);
- reg &= ~MXC_CCM_CACRR_ARM_PODF_MASK;
- reg |= cpu_wp_tbl[wp].cpu_podf << MXC_CCM_CACRR_ARM_PODF_OFFSET;
- __raw_writel(reg, MXC_CCM_CACRR);
We have a simple divider here. Why do we need a reference to struct cpu_wp then? This code could look much simpler.
(side note: I'm aware that the original Freescale code is also able to change the cpu frequency using the pll instead of the divider, but is this really necessary?)
Using wp_tbl is because that it also contains information like regulator voltage.
The clock code does not handle the regulators, not even in the fsl kernel.
Since the regulator driver for imx51 have not been upstreamed, you don't see any regulator operation here. Also, like you mentioned, there are two ways to change cpufreq, by post divider or pll change. And post divider change is a simpler way and also the only way for babbage, since cpu freq and ddr freq are all root from the same pll on babbage and they will interfere each other.
I understand what you have to do, but the way you did really looks strange. You introduce get_cpu_wp() to get the complete array of workpoints. In the cpufreq driver you have this complete array, pick the desired workpoint, pass the frequency to the clock layer which in turn calls get_cpu_wp() to get the array and loop around it to get the workpoint from the frequency again. Addionally you maintain a pointer to what the clock code thinks the current workpoint is.
How about making the clock code agnostic of such workpoints and calculate the pll values and dividers for a given frequency based on the frequency. If that's to complicated you could maintain a table in the clock code. If that's not flexible enough you could pass a pointer to this table to mx51_clocks_init. This array could carry information relevant only to the clock code and only relevant to the i.MX51, In the fsl kernel struct cpu_wp is a catch-all struct for all i.MXs. The i.MX51 uses the pdf/mfi/mfd/mfn/cpu_podf fields, the i.MX35 uses the pll_reg and pdr0_reg fields and who knows what the next i.MX needs.
+static struct cpufreq_frequency_table imx_freq_table[4];
Three frequencies should be enough for everyone, right? This should be dynamically allocated like in other cpufreq drivers.
Yes, we only support 3 work points, which is for sure. Actually, we only support 2 points on most mx51 chips. I supposed it is ok to use static array here.
Please just add dynamic allocation, then we are safe for any potential future use.
Sascha