On 10/19/2015 06:28 PM, York Sun wrote:
>
>
> On 10/19/2015 05:36 PM, Stephen Boyd wrote:
>> On 10/20, York Sun wrote:
>>> Sorry for top posting. I am on outlook web.
>>>
>>> We have no problem with device tree. Let's put it aside.
>>>
>>> When device is not used, the platform data is used to hold the data, filled
>>> by a platform device before probing the clock. For my case, the platform
>>> device is a PCIe device. It is a multifunction device with I2C controller
>>> on it. The pseudo code looks like
>>>
>>> struct i2c_board_info si5338_info[NUM_SI5338_CHIPS] = {
>>> {
>>> .type = "si5338",
>>> .platform_data = &si5338_pdata[0],
>>> },
>>> {
>>> .type = "si5338",
>>> .platform_data = &si5338_pdata[1],
>>> },
>>> {
>>> .type = "si5338",
>>> .platform_data = &si5338_pdata[2],
>>> },
>>> {
>>> .type = "si5338",
>>> .platform_data = &si5338_pdata[3],
>>> },
>>> };
>>> clk = clk_register_fixed_rate(&pdev->dev, "ref25", NULL, CLK_IS_ROOT,
>>> 25000000);
>>> for (i = 0; i < NUM_SI5338_CHIPS; i++) {
>>> si5338_pdata[i].clk_xtal = clk;
>>> adap = i2c_get_adapter(private->i2c_adp->nr + 1 + i);
>>> private->i2c_client[i] = i2c_new_probed_device(adap, &si5338_info[i],
>>> i2c_si5338_addr, NULL);
>>> }
>>>
>>> You can see, when the fixed-rate clock is registered, the device id of
>>> si5338 is unknown yet. (I am using one 25MHz clock to feed multiple si5338,
>>> where maybe I should create multiple 25MHz clocks, but that's another
>>> discussion.)
>>>
>>> I hope I have made it clear.
>>
>> Yes. It would be great if we could modify the i2c framework to
>> let us create an i2c device but not call device_register() until a later
>> time. So something like this could be done in the platform
>> driver:
>>
>> for (i = 0; i < NUM_SI5338_CHIPS; i++) {
>> adap = i2c_get_adapter(private->i2c_adp->nr + 1 + i);
>> private->i2c_client[i] = i2c_new_device_unregistered(adap,
>> &si5338_info[i], i2c_si5338_addr);
>> clkdev_create(clk, NULL, dev_name(private->i2c_client[i]->dev));
>> device_register(private->i2c_client[i]->dev);
>> }
>>
>> Then in the si5338 driver we call devm_clk_get(i2c->dev, NULL)
>> and we get the xtal clock.
>>
>
> Then what do we do before we have this i2c_new_device_unregistered()? Does my
> proposal make sense?
>
Stephen,
How about this? I will use devm_clk_get() when parsing device tree, and still
pass clk pointer if using platform data. In this way, I don't have to call
clk_put() in the driver to release the parent clocks. The platform device driver
is responsible to get the clock and put the clock. It seems reasonable to me. If
you don't disagree, I will prepare v7 patch for review.
York
--
To unsubscribe from this list: send the line "unsubscribe linux-clk" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
