[RFC] clk: introduce critical clocks

[Lee Jones]

On Thu, 30 Apr 2015, Sascha Hauer wrote:

> Some clocks are so critical to the system that they must never be turned
> off unless explicitly requested. Normally unused clocks get disabled in
> the clk_disable_unused initcall. Currently there are two ways for
> keeping clocks enabled even if they are unused, both with their own
> drawbacks:
> 
> - The CLK_IGNORE_UNUSED flag. Clocks with this flags are not disabled
>   during clk_disable_unused. The problem with this is that these clocks
>   are skipped during the clk_disable_unused initcall, but can get
>   disabled due to normal clk_enable/disable at any time, not necessarily
>   even on operations on the critical clocks themselves but on operations
>   on their ancestors, siblings or children.
> 
> - call clk_prepare_enable right after registration in the clock
>   driver. This works properly, but due to the increased enable counter
>   the clock will never get disabled, even not when a proper user
>   comes along who knows when the clock can safely be disabled.
> 
> This patch solves this by introducing two new API calls:
> 
> clk_prepare_enable_critical() will call clk_prepare_enable() on a clock
> and additionally set a 'critical' flag on the clock. This call is
> intended for clock providers which provide a critical clock.
> 
> clk_disable_unprepare_critical() is intended for consumers of a critical
> clock.  Consumers should first call clk_prepare_enable() on a clock to
> get their own prepare/enable reference and call
> clk_disable_unprepare_critical() afterwards.  This will clear the
> 'critical' flag from the clock and decrease the prepare/enable counter.
> The ownership of the clock is now transferred to the consumer. Calling
> clk_disable_unprepare_critical() on a clock which doesn't have the
> 'critical' flag set is just a no-op, so consumers can safely call this
> on their clocks.
> 
> To make this more versatile a clk flag could be introduced which could
> be set during registration so that clk providers would not have to make
> this API call on each critical clock. I haven't implemented this yet
> because it turned out not to be trivial. clk_prepare_enable can only
> be called when the clock is not orphaned which may not be the case
> during registration time. I can implement this if this way of handling
> critical clocks is considered suitable for mainline.

I like the premise of this set and I think it'll work well with my
patchset.  So are you two okay with me fixing up my patchset to
encompass it, or did you have other plans?

> Signed-off-by: Sascha Hauer <s.hauer at pengutronix.de>
> ---
>  drivers/clk/clk.c            | 46 ++++++++++++++++++++++++++++++++++++++++++++
>  include/linux/clk-provider.h |  2 ++
>  include/linux/clk.h          | 10 ++++++++++
>  3 files changed, 58 insertions(+)
> 
> diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c
> index 459ce9d..d57d4fd 100644
> --- a/drivers/clk/clk.c
> +++ b/drivers/clk/clk.c
> @@ -64,6 +64,7 @@ struct clk_core {
>  	unsigned long		flags;
>  	unsigned int		enable_count;
>  	unsigned int		prepare_count;
> +	bool			enable_critical;
>  	unsigned long		accuracy;
>  	int			phase;
>  	struct hlist_head	children;
> @@ -940,7 +941,13 @@ void clk_unprepare(struct clk *clk)
>  		return;
>  
>  	clk_prepare_lock();
> +
> +	if (clk->core->enable_critical)
> +		goto out;
> +
>  	clk_core_unprepare(clk->core);
> +
> +out:
>  	clk_prepare_unlock();
>  }
>  EXPORT_SYMBOL_GPL(clk_unprepare);
> @@ -995,7 +1002,9 @@ int clk_prepare(struct clk *clk)
>  		return 0;
>  
>  	clk_prepare_lock();
> +
>  	ret = clk_core_prepare(clk->core);
> +
>  	clk_prepare_unlock();
>  
>  	return ret;
> @@ -2250,6 +2259,43 @@ bool clk_is_match(const struct clk *p, const struct clk *q)
>  }
>  EXPORT_SYMBOL_GPL(clk_is_match);
>  
> +int clk_prepare_enable_critical(struct clk *clk)
> +{
> +	if (!clk)
> +		return 0;
> +
> +	clk_prepare_lock();
> +	if (clk->core->enable_critical) {
> +		clk_prepare_unlock();
> +		return -EBUSY;
> +	}
> +
> +	clk->core->enable_critical = true;
> +
> +	clk_prepare_unlock();
> +
> +	return clk_prepare_enable(clk);
> +
> +}
> +
> +void clk_disable_unprepare_critical(struct clk *clk)
> +{
> +	if (!clk)
> +		return;
> +
> +	clk_prepare_lock();
> +	if (!clk->core->enable_critical) {
> +		clk_prepare_unlock();
> +		return;
> +	}
> +
> +	clk->core->enable_critical = false;
> +
> +	clk_prepare_unlock();
> +
> +	clk_disable_unprepare(clk);
> +}
> +
>  /**
>   * __clk_init - initialize the data structures in a struct clk
>   * @dev:	device initializing this clk, placeholder for now
> diff --git a/include/linux/clk-provider.h b/include/linux/clk-provider.h
> index df69531..8727c12 100644
> --- a/include/linux/clk-provider.h
> +++ b/include/linux/clk-provider.h
> @@ -694,5 +694,7 @@ struct dentry *clk_debugfs_add_file(struct clk_hw *hw, char *name, umode_t mode,
>  				void *data, const struct file_operations *fops);
>  #endif
>  
> +int clk_prepare_enable_critical(struct clk *clk);
> +
>  #endif /* CONFIG_COMMON_CLK */
>  #endif /* CLK_PROVIDER_H */
> diff --git a/include/linux/clk.h b/include/linux/clk.h
> index 68c16a6..b259e36 100644
> --- a/include/linux/clk.h
> +++ b/include/linux/clk.h
> @@ -138,6 +138,16 @@ int clk_get_phase(struct clk *clk);
>   */
>  bool clk_is_match(const struct clk *p, const struct clk *q);
>  
> +/**
> + * clk_disable_unprepare_critical - remove critical flag from clock
> + * @clk: clock source
> + *
> + * This removes the critical flag from a clock and disables it. The
> + * consumer should have enabled the clock by itself now if it wants
> + * to keep the clock enabled.
> + */
> +void clk_disable_unprepare_critical(struct clk *clk);
> +
>  #else
>  
>  static inline long clk_get_accuracy(struct clk *clk)

-- 
Lee Jones
Linaro STMicroelectronics Landing Team Lead
Linaro.org │ Open source software for ARM SoCs
Follow Linaro: Facebook | Twitter | Blog