Re: [PATCH v8 11/21] clk: tegra: clk-dfll: Add suspend and resume support

[Sowjanya Komatineni <skomatineni@nvidia.com>]

On 8/9/19 11:00 AM, Dmitry Osipenko wrote:
> 09.08.2019 19:39, Sowjanya Komatineni пишет:
>> On 8/9/19 5:23 AM, Dmitry Osipenko wrote:
>>> 09.08.2019 2:46, Sowjanya Komatineni пишет:
>>>> This patch implements DFLL suspend and resume operation.
>>>>
>>>> During system suspend entry, CPU clock will switch CPU to safe
>>>> clock source of PLLP and disables DFLL clock output.
>>>>
>>>> DFLL driver suspend confirms DFLL disable state and errors out on
>>>> being active.
>>>>
>>>> DFLL is re-initialized during the DFLL driver resume as it goes
>>>> through complete reset during suspend entry.
>>>>
>>>> Signed-off-by: Sowjanya Komatineni <skomatineni@nvidia.com>
>>>> ---
>>>>    drivers/clk/tegra/clk-dfll.c               | 56 ++++++++++++++++++++++++++++++
>>>>    drivers/clk/tegra/clk-dfll.h               |  2 ++
>>>>    drivers/clk/tegra/clk-tegra124-dfll-fcpu.c |  1 +
>>>>    3 files changed, 59 insertions(+)
>>>>
>>>> diff --git a/drivers/clk/tegra/clk-dfll.c b/drivers/clk/tegra/clk-dfll.c
>>>> index f8688c2ddf1a..eb298a5d7be9 100644
>>>> --- a/drivers/clk/tegra/clk-dfll.c
>>>> +++ b/drivers/clk/tegra/clk-dfll.c
>>>> @@ -1487,6 +1487,7 @@ static int dfll_init(struct tegra_dfll *td)
>>>>        td->last_unrounded_rate = 0;
>>>>          pm_runtime_enable(td->dev);
>>>> +    pm_runtime_irq_safe(td->dev);
>>>>        pm_runtime_get_sync(td->dev);
>>>>          dfll_set_mode(td, DFLL_DISABLED);
>>>> @@ -1513,6 +1514,61 @@ static int dfll_init(struct tegra_dfll *td)
>>>>        return ret;
>>>>    }
>>>>    +/**
>>>> + * tegra_dfll_suspend - check DFLL is disabled
>>>> + * @dev: DFLL device *
>>>> + *
>>>> + * DFLL clock should be disabled by the CPUFreq driver. So, make
>>>> + * sure it is disabled and disable all clocks needed by the DFLL.
>>>> + */
>>>> +int tegra_dfll_suspend(struct device *dev)
>>>> +{
>>>> +    struct tegra_dfll *td = dev_get_drvdata(dev);
>>>> +
>>>> +    if (dfll_is_running(td)) {
>>>> +        dev_err(td->dev, "dfll is enabled while shouldn't be\n");
>>>> +        return -EBUSY;
>>>> +    }
>>>> +
>>>> +    reset_control_assert(td->dvco_rst);
>>>> +
>>>> +    return 0;
>>>> +}
>>>> +EXPORT_SYMBOL(tegra_dfll_suspend);
>>>> +
>>>> +/**
>>>> + * tegra_dfll_resume - reinitialize DFLL on resume
>>>> + * @dev: DFLL instance
>>>> + *
>>>> + * DFLL is disabled and reset during suspend and resume.
>>>> + * So, reinitialize the DFLL IP block back for use.
>>>> + * DFLL clock is enabled later in closed loop mode by CPUFreq
>>>> + * driver before switching its clock source to DFLL output.
>>>> + */
>>>> +int tegra_dfll_resume(struct device *dev)
>>>> +{
>>>> +    struct tegra_dfll *td = dev_get_drvdata(dev);
>>>> +
>>>> +    reset_control_deassert(td->dvco_rst);
>>> This doesn't look right because I assume that DFLL resetting is
>>> synchronous and thus clk should be enabled in order for reset to
>>> propagate inside hardware.
>>>
>>>> +    pm_runtime_get_sync(td->dev);
>>> Hence it will be better to remove the above reset_control_deassert() and
>>> add here:
>>>
>>>      reset_control_reset(td->dvco_rst);
>> By the time dfll resume happens, dfll controller clock will already be enabled.
>>
>> so doing reset de-assert before pm_runtime seems ok.
> I don't see what enables the DFLL clock because it should be enabled by the CPUFreq driver
> on resume from suspend and resume happens after resuming of the DFLL driver.

dvco_rst is part of peripheral clocks and all peripheral clocks are 
restored by clk-tegra210 driver which happens before dfll driver resume.

So dfll rst thru part of peripheral clock enable is set prior to dfll 
reset deassertion