Re: [PATCH] i2c: i2c-tegra: Move clk_prepare/clk_set_rate to probe

[Stephen Warren <swarren-3lzwWm7+Weoh9ZMKESR00Q@public.gmane.org>]

On 08/15/2014 12:02 PM, Peter De Schrijver wrote:
> On Fri, Aug 15, 2014 at 06:18:15PM +0200, Stephen Warren wrote:
>> On 08/15/2014 03:47 AM, Mikko Perttunen wrote:
>>> Currently the i2c-tegra bus driver prepares, enables
>>> and set_rates its clocks separately for each transfer.
>>> This causes locking problems when doing I2C transfers
>>> from clock notifiers; see
>>> http://lists.infradead.org/pipermail/linux-arm-kernel/2014-July/268653.html
>>>
>>> This patch moves clk_prepare/unprepare and clk_set_rate calls to
>>> the probe function, leaving only clk_enable/disable to be
>>> done on each transfer. This solves the locking issue.
>>
>>> diff --git a/drivers/i2c/busses/i2c-tegra.c b/drivers/i2c/busses/i2c-tegra.c
>>
>>> @@ -380,34 +380,33 @@ static inline int tegra_i2c_clock_enable(struct tegra_i2c_dev *i2c_dev)
>>>    {
>>>    	int ret;
>>>    	if (!i2c_dev->hw->has_single_clk_source) {
>>> -		ret = clk_prepare_enable(i2c_dev->fast_clk);
>>> +		ret = clk_enable(i2c_dev->fast_clk);
>>
>> Here, both the prepare and enable wrap just the I2C transfer, ...
>>
>>> @@ -428,9 +427,6 @@ static int tegra_i2c_init(struct tegra_i2c_dev *i2c_dev)
>>>    	i2c_writel(i2c_dev, val, I2C_CNFG);
>>>    	i2c_writel(i2c_dev, 0, I2C_INT_MASK);
>>>
>>> -	clk_multiplier *= (i2c_dev->hw->clk_divisor_std_fast_mode + 1);
>>> -	clk_set_rate(i2c_dev->div_clk, i2c_dev->bus_clk_rate * clk_multiplier);
>>
>> ... whereas the rate is set up when the controller is initialized, i.e.
>> much earlier.
>>
>>> @@ -777,17 +774,39 @@ static int tegra_i2c_probe(struct platform_device *pdev)
>>
>>> +	if (!i2c_dev->hw->has_single_clk_source) {
>>> +		ret = clk_prepare(i2c_dev->fast_clk);
>>> +		if (ret < 0) {
>>> +			dev_err(i2c_dev->dev, "Clock prepare failed %d\n", ret);
>>> +			return ret;
>>> +		}
>>> +	}
>>> +
>>> +	ret = clk_prepare(i2c_dev->div_clk);
>>> +	if (ret < 0) {
>>> +		dev_err(i2c_dev->dev, "Clock prepare failed %d\n", ret);
>>> +		goto unprepare_fast_clk;
>>> +	}
>>> +
>>> +	clk_multiplier *= (i2c_dev->hw->clk_divisor_std_fast_mode + 1);
>>> +	ret = clk_set_rate(i2c_dev->div_clk,
>>> +			   i2c_dev->bus_clk_rate * clk_multiplier);
>>> +	if (ret) {
>>> +		dev_err(i2c_dev->dev, "Clock rate change failed %d\n", ret);
>>> +		goto unprepare_div_clk;
>>> +	}
>>
>> However, the new code sets the clock rate after the clock is prepared. I
>> think the rate should be set first, then the clock prepared. While this
>> likely doesn't apply to the Tegra clock controller, prepare() is allowed
>> to enable the clock if enable() can't be implemented in an atomic
>> fashion (in which case enable/disable would be no-ops), and we should
>> make sure that the driver correctly configures the clock before
>> potentially enabling it.
>>
>> I'm not sure if a similar change to our SPI drivers is possible; after
>> all, the SPI transfer rate can vary per message, so if clk_set_rate()
>> acquires a lock, it seems there's no way to avoid the issue there.
>
> Even for i2c this could be the case I think if you use the highspeed (3.4Mhz)
> mode? From what I remember, a highspeed i2c transaction starts with a lower
> speed preamble to make sure non highspeed slaves don't get confused? Which
> means you could change the bus speed depending on the slave you're addressing.

Since there's no separate chip-select for I2C, I believe all I2C devices 
need to be able to understand the entire transaction, so the I2C bus 
speed is fixed.

At least, that's my understanding between 100KHz and 400KHz I2C. I don't 
know if 3.4MHz I2C introduced something new, although considering that 
slower I2C never had anything about being compatible with fast stuff in 
the spec AFAIK, and such speed-switching would only be useful for 
backwards-compatibility, I don't see how that would work.

>> Luckily, we don't have any SPI-based chips that do anything related to
>> clocks on any of our current boards...
>
> And we don't use SPI to talk to the PMIC, which is the usecase were actually
> run into problems with the locking.

IIRC, the I2C-based clock provider (or consumer?) issue was something 
mentioned (later on?) in the email thread linked by the patch description.