[2/2] clk: allow reentrant calls into the clk framework

Reentrancy into the clock framework is necessary for clock operations
that result in nested calls to the clk api.  A common example is a clock
that is prepared via an i2c transaction, such as a clock inside of a
discrete audio chip or a power management IC.  The i2c subsystem itself
will use the clk api resulting in a deadlock:

clk_prepare(audio_clk)
	i2c_transfer(..)
		clk_prepare(i2c_controller_clk)

The ability to reenter the clock framework prevents this deadlock.

Other use cases exist such as allowing .set_rate callbacks to call
clk_set_parent to achieve the best rate, or to save power in certain
configurations.  Yet another example is performing pinctrl operations
from a clk_ops callback.  Calls into the pinctrl subsystem may call
clk_{un}prepare on an unrelated clock.  Allowing for nested calls to
reenter the clock framework enables both of these use cases.

Reentrancy is implemented by two global pointers that track the owner
currently holding a global lock.  One pointer tracks the owner during
sleepable, mutex-protected operations and the other one tracks the owner
during non-interruptible, spinlock-protected operations.

When the clk framework is entered we try to hold the global lock.  If it
is held we compare the current task against the current owner; a match
implies a nested call and we reenter.  If the values do not match then
we block on the lock until it is released.

Signed-off-by: Mike Turquette <mturquette@linaro.org>
Cc: Rajagopal Venkat <rajagopal.venkat@linaro.org>
Cc: David Brown <davidb@codeaurora.org>
Cc: Ulf Hansson <ulf.hansson@linaro.org>
Tested-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
---
Changes since v5:
 * fixed up typo in changelog

Changes since v4:
 * remove uneccesary atomic operations
 * remove casting bugs
 * place reentrancy logic into locking helper functions
 * improve debugging with reference counting and WARNs

 drivers/clk/clk.c |   44 ++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 42 insertions(+), 2 deletions(-)

On 28 March 2013 21:59, Mike Turquette <mturquette@linaro.org> wrote:
> Reentrancy into the clock framework is necessary for clock operations
> that result in nested calls to the clk api.  A common example is a clock
> that is prepared via an i2c transaction, such as a clock inside of a
> discrete audio chip or a power management IC.  The i2c subsystem itself
> will use the clk api resulting in a deadlock:
>
> clk_prepare(audio_clk)
>         i2c_transfer(..)
>                 clk_prepare(i2c_controller_clk)
>
> The ability to reenter the clock framework prevents this deadlock.
>
> Other use cases exist such as allowing .set_rate callbacks to call
> clk_set_parent to achieve the best rate, or to save power in certain
> configurations.  Yet another example is performing pinctrl operations
> from a clk_ops callback.  Calls into the pinctrl subsystem may call
> clk_{un}prepare on an unrelated clock.  Allowing for nested calls to
> reenter the clock framework enables both of these use cases.
>
> Reentrancy is implemented by two global pointers that track the owner
> currently holding a global lock.  One pointer tracks the owner during
> sleepable, mutex-protected operations and the other one tracks the owner
> during non-interruptible, spinlock-protected operations.
>
> When the clk framework is entered we try to hold the global lock.  If it
> is held we compare the current task against the current owner; a match
> implies a nested call and we reenter.  If the values do not match then
> we block on the lock until it is released.
>
> Signed-off-by: Mike Turquette <mturquette@linaro.org>
> Cc: Rajagopal Venkat <rajagopal.venkat@linaro.org>
> Cc: David Brown <davidb@codeaurora.org>
> Cc: Ulf Hansson <ulf.hansson@linaro.org>
> Tested-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
> Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
> ---
> Changes since v5:
>  * fixed up typo in changelog
>
> Changes since v4:
>  * remove uneccesary atomic operations
>  * remove casting bugs
>  * place reentrancy logic into locking helper functions
>  * improve debugging with reference counting and WARNs
>
>  drivers/clk/clk.c |   44 ++++++++++++++++++++++++++++++++++++++++++--
>  1 file changed, 42 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c
> index 0b5d612..0230c9d 100644
> --- a/drivers/clk/clk.c
> +++ b/drivers/clk/clk.c
> @@ -19,10 +19,17 @@
>  #include <linux/of.h>
>  #include <linux/device.h>
>  #include <linux/init.h>
> +#include <linux/sched.h>
>
>  static DEFINE_SPINLOCK(enable_lock);
>  static DEFINE_MUTEX(prepare_lock);
>
> +static struct task_struct *prepare_owner;
> +static struct task_struct *enable_owner;
> +
> +static int prepare_refcnt;
> +static int enable_refcnt;
> +
>  static HLIST_HEAD(clk_root_list);
>  static HLIST_HEAD(clk_orphan_list);
>  static LIST_HEAD(clk_notifier_list);
> @@ -30,23 +37,56 @@ static LIST_HEAD(clk_notifier_list);
>  /***           locking             ***/
>  static void clk_prepare_lock(void)
>  {
> -       mutex_lock(&prepare_lock);
> +       if (!mutex_trylock(&prepare_lock)) {
> +               if (prepare_owner == current) {
> +                       prepare_refcnt++;
> +                       return;
> +               }
> +               mutex_lock(&prepare_lock);
> +       }
> +       WARN_ON_ONCE(prepare_owner != NULL);
> +       WARN_ON_ONCE(prepare_refcnt != 0);
> +       prepare_owner = current;
> +       prepare_refcnt = 1;
>  }
>
>  static void clk_prepare_unlock(void)
>  {
> +       WARN_ON_ONCE(prepare_owner != current);
> +       WARN_ON_ONCE(prepare_refcnt == 0);
> +
> +       if (--prepare_refcnt)
> +               return;
> +       prepare_owner = NULL;
>         mutex_unlock(&prepare_lock);
>  }
>
>  static unsigned long clk_enable_lock(void)
>  {
>         unsigned long flags;
> -       spin_lock_irqsave(&enable_lock, flags);
> +
> +       if (!spin_trylock_irqsave(&enable_lock, flags)) {
> +               if (enable_owner == current) {
> +                       enable_refcnt++;
> +                       return flags;
> +               }
> +               spin_lock_irqsave(&enable_lock, flags);
> +       }
> +       WARN_ON_ONCE(enable_owner != NULL);
> +       WARN_ON_ONCE(enable_refcnt != 0);
> +       enable_owner = current;
> +       enable_refcnt = 1;
>         return flags;
>  }
>
>  static void clk_enable_unlock(unsigned long flags)
>  {
> +       WARN_ON_ONCE(enable_owner != current);
> +       WARN_ON_ONCE(enable_refcnt == 0);
> +
> +       if (--enable_refcnt)
> +               return;
> +       enable_owner = NULL;
>         spin_unlock_irqrestore(&enable_lock, flags);
>  }
>
> --
> 1.7.10.4
>

Great piece of work!

Reviewed-by: Ulf Hansson <ulf.hansson@linaro.org>