diff mbox

[v13,3/6] clk: Make clk API return per-user struct clk instances

Message ID 54D3CD6A.1010209@codeaurora.org (mailing list archive)
State New, archived
Headers show

Commit Message

Stephen Boyd Feb. 5, 2015, 8:07 p.m. UTC
On 02/05/15 11:44, Sylwester Nawrocki wrote:
> Hi Tomeu,
>
> On 23/01/15 12:03, Tomeu Vizoso wrote:
>>  int __clk_get(struct clk *clk)
>>  {
>> -	if (clk) {
>> -		if (!try_module_get(clk->owner))
>> +	struct clk_core *core = !clk ? NULL : clk->core;
>> +
>> +	if (core) {
>> +		if (!try_module_get(core->owner))
>>  			return 0;
>>  
>> -		kref_get(&clk->ref);
>> +		kref_get(&core->ref);
>>  	}
>>  	return 1;
>>  }
>>  
>> -void __clk_put(struct clk *clk)
>> +static void clk_core_put(struct clk_core *core)
>>  {
>>  	struct module *owner;
>>  
>> -	if (!clk || WARN_ON_ONCE(IS_ERR(clk)))
>> -		return;
>> +	owner = core->owner;
>>  
>>  	clk_prepare_lock();
>> -	owner = clk->owner;
>> -	kref_put(&clk->ref, __clk_release);
>> +	kref_put(&core->ref, __clk_release);
>>  	clk_prepare_unlock();
>>  
>>  	module_put(owner);
>>  }
>>  
>> +void __clk_put(struct clk *clk)
>> +{
>> +	if (!clk || WARN_ON_ONCE(IS_ERR(clk)))
>> +		return;
>> +
>> +	clk_core_put(clk->core);
>> +	kfree(clk);
> Why do we have kfree() here? clk_get() doesn't allocate the data structure 
> being freed here. What happens if we do clk_get(), clk_put(), clk_get() 
> on same clock?
>
> I suspect __clk_free_clk() should be called in __clk_release() callback
> instead, but then there is an issue of safely getting reference to
> struct clk from struct clk_core pointer.
>
> I tested linux-next on Odroid U3 and booting fails with oopses as below.
> There is no problems when the above kfree() is commented out.
>
>

Ah now I get it. You meant to say that of_clk_get_by_clkspec() doesn't
return an allocated clk pointer. Let's fix that.

----8<----

From: Stephen Boyd <sboyd@codeaurora.org>
Subject: [PATCH] clkdev: Always allocate a struct clk in OF functions

of_clk_get_by_clkspec() returns a struct clk pointer but it
doesn't create a new handle for the consumers. Instead it just
returns whatever the OF clk provider hands out. Let's create a
per-user handle here so that clk_put() can properly unlink it and
free it when the consumer is done.

Fixes: 035a61c314eb "clk: Make clk API return per-user struct clk instances"
Reported-by: Sylwester Nawrocki <s.nawrocki@samsung.com>
Signed-off-by: Stephen Boyd <sboyd@codeaurora.org>
---
 drivers/clk/clkdev.c | 44 +++++++++++++++++++++++---------------------
 1 file changed, 23 insertions(+), 21 deletions(-)

Comments

Stephen Boyd Feb. 5, 2015, 10:14 p.m. UTC | #1
On 02/05/15 12:07, Stephen Boyd wrote:
> On 02/05/15 11:44, Sylwester Nawrocki wrote:
>> Hi Tomeu,
>>
>> On 23/01/15 12:03, Tomeu Vizoso wrote:
>>>  int __clk_get(struct clk *clk)
>>>  {
>>> -	if (clk) {
>>> -		if (!try_module_get(clk->owner))
>>> +	struct clk_core *core = !clk ? NULL : clk->core;
>>> +
>>> +	if (core) {
>>> +		if (!try_module_get(core->owner))
>>>  			return 0;
>>>  
>>> -		kref_get(&clk->ref);
>>> +		kref_get(&core->ref);
>>>  	}
>>>  	return 1;
>>>  }
>>>  
>>> -void __clk_put(struct clk *clk)
>>> +static void clk_core_put(struct clk_core *core)
>>>  {
>>>  	struct module *owner;
>>>  
>>> -	if (!clk || WARN_ON_ONCE(IS_ERR(clk)))
>>> -		return;
>>> +	owner = core->owner;
>>>  
>>>  	clk_prepare_lock();
>>> -	owner = clk->owner;
>>> -	kref_put(&clk->ref, __clk_release);
>>> +	kref_put(&core->ref, __clk_release);
>>>  	clk_prepare_unlock();
>>>  
>>>  	module_put(owner);
>>>  }
>>>  
>>> +void __clk_put(struct clk *clk)
>>> +{
>>> +	if (!clk || WARN_ON_ONCE(IS_ERR(clk)))
>>> +		return;
>>> +
>>> +	clk_core_put(clk->core);
>>> +	kfree(clk);
>> Why do we have kfree() here? clk_get() doesn't allocate the data structure 
>> being freed here. What happens if we do clk_get(), clk_put(), clk_get() 
>> on same clock?
>>
>> I suspect __clk_free_clk() should be called in __clk_release() callback
>> instead, but then there is an issue of safely getting reference to
>> struct clk from struct clk_core pointer.
>>
>> I tested linux-next on Odroid U3 and booting fails with oopses as below.
>> There is no problems when the above kfree() is commented out.
>>
>>
> Ah now I get it. You meant to say that of_clk_get_by_clkspec() doesn't
> return an allocated clk pointer. Let's fix that.
>
> ----8<----
>
> From: Stephen Boyd <sboyd@codeaurora.org>
> Subject: [PATCH] clkdev: Always allocate a struct clk in OF functions
>
> of_clk_get_by_clkspec() returns a struct clk pointer but it
> doesn't create a new handle for the consumers. Instead it just
> returns whatever the OF clk provider hands out. Let's create a
> per-user handle here so that clk_put() can properly unlink it and
> free it when the consumer is done.
>
> Fixes: 035a61c314eb "clk: Make clk API return per-user struct clk instances"
> Reported-by: Sylwester Nawrocki <s.nawrocki@samsung.com>
> Signed-off-by: Stephen Boyd <sboyd@codeaurora.org>
> ---
>  drivers/clk/clkdev.c | 44 +++++++++++++++++++++++---------------------
>  1 file changed, 23 insertions(+), 21 deletions(-)
>
> diff --git a/drivers/clk/clkdev.c b/drivers/clk/clkdev.c
> index 29a1ab7af4b8..00d747d09b2a 100644
> --- a/drivers/clk/clkdev.c
> +++ b/drivers/clk/clkdev.c
> @@ -29,15 +29,8 @@ static DEFINE_MUTEX(clocks_mutex);
>  
>  #if defined(CONFIG_OF) && defined(CONFIG_COMMON_CLK)
>  
> -/**
> - * of_clk_get_by_clkspec() - Lookup a clock form a clock provider
> - * @clkspec: pointer to a clock specifier data structure
> - *
> - * This function looks up a struct clk from the registered list of clock
> - * providers, an input is a clock specifier data structure as returned
> - * from the of_parse_phandle_with_args() function call.
> - */
> -struct clk *of_clk_get_by_clkspec(struct of_phandle_args *clkspec)
> +static struct clk *__of_clk_get_by_clkspec(struct of_phandle_args *clkspec,
> +					 const char *dev_id, const char *con_id)
>  {
>  	struct clk *clk;
>  
> @@ -47,6 +40,8 @@ struct clk *of_clk_get_by_clkspec(struct of_phandle_args *clkspec)
>  	of_clk_lock();
>  	clk = __of_clk_get_from_provider(clkspec);
>  
> +	if (!IS_ERR(clk))
> +		clk = __clk_create_clk(__clk_get_hw(clk), dev_id, con_id);
>  	if (!IS_ERR(clk) && !__clk_get(clk))
>  		clk = ERR_PTR(-ENOENT);
>

Actually we can bury the __clk_create_clk() inside
__of_clk_get_from_provider(). We should also move __clk_get() into there
because right now we have a hole where whoever calls
of_clk_get_from_provider() never calls __clk_get() on the clk, leading
to possible badness. v2 coming soon.
Russell King - ARM Linux Feb. 6, 2015, 12:42 a.m. UTC | #2
On Thu, Feb 05, 2015 at 02:14:01PM -0800, Stephen Boyd wrote:
> Actually we can bury the __clk_create_clk() inside
> __of_clk_get_from_provider(). We should also move __clk_get() into there
> because right now we have a hole where whoever calls
> of_clk_get_from_provider() never calls __clk_get() on the clk, leading
> to possible badness. v2 coming soon.

There's some other issues here too...

sound/soc/kirkwood/kirkwood-i2s.c:

        priv->clk = devm_clk_get(&pdev->dev, np ? "internal" : NULL);
...
        priv->extclk = devm_clk_get(&pdev->dev, "extclk");
        if (IS_ERR(priv->extclk)) {
...
	} else {
                if (priv->extclk == priv->clk) {
                        devm_clk_put(&pdev->dev, priv->extclk);
                        priv->extclk = ERR_PTR(-EINVAL);
                } else {
                        dev_info(&pdev->dev, "found external clock\n");
                        clk_prepare_enable(priv->extclk);
                        soc_dai = kirkwood_i2s_dai_extclk;
                }

It should be fine provided your "trick" is only done for DT clocks,
but not for legacy - with legacy, a NULL in the clkdev tables will
match both these requests, hence the need to compare the clk_get()
return value to tell whether we get the same clock.
Stephen Boyd Feb. 6, 2015, 1:35 a.m. UTC | #3
On 02/05/15 16:42, Russell King - ARM Linux wrote:
> On Thu, Feb 05, 2015 at 02:14:01PM -0800, Stephen Boyd wrote:
>> Actually we can bury the __clk_create_clk() inside
>> __of_clk_get_from_provider(). We should also move __clk_get() into there
>> because right now we have a hole where whoever calls
>> of_clk_get_from_provider() never calls __clk_get() on the clk, leading
>> to possible badness. v2 coming soon.
> There's some other issues here too...
>
> sound/soc/kirkwood/kirkwood-i2s.c:
>
>         priv->clk = devm_clk_get(&pdev->dev, np ? "internal" : NULL);
> ...
>         priv->extclk = devm_clk_get(&pdev->dev, "extclk");
>         if (IS_ERR(priv->extclk)) {
> ...
> 	} else {
>                 if (priv->extclk == priv->clk) {
>                         devm_clk_put(&pdev->dev, priv->extclk);
>                         priv->extclk = ERR_PTR(-EINVAL);
>                 } else {
>                         dev_info(&pdev->dev, "found external clock\n");
>                         clk_prepare_enable(priv->extclk);
>                         soc_dai = kirkwood_i2s_dai_extclk;
>                 }
>
> It should be fine provided your "trick" is only done for DT clocks,
> but not for legacy - with legacy, a NULL in the clkdev tables will
> match both these requests, hence the need to compare the clk_get()
> return value to tell whether we get the same clock.
>

Are we still talking about of_clk_get_from_provider()? Or are we talking
about comparing struct clk pointers? From what I can tell this code is
now broken because we made all clk getting functions (there's quite a
few...) return unique pointers every time they're called. It seems that
the driver wants to know if extclk and clk are the same so it can do
something differently in kirkwood_set_rate(). Do we need some sort of
clk_equal(struct clk *a, struct clk *b) function for drivers like this?

Also, even on DT this could fail if the DT author made internal and
extclk map to the same clock provider and cell.
Russell King - ARM Linux Feb. 6, 2015, 1:39 p.m. UTC | #4
On Thu, Feb 05, 2015 at 05:35:28PM -0800, Stephen Boyd wrote:
> On 02/05/15 16:42, Russell King - ARM Linux wrote:
> > On Thu, Feb 05, 2015 at 02:14:01PM -0800, Stephen Boyd wrote:
> >> Actually we can bury the __clk_create_clk() inside
> >> __of_clk_get_from_provider(). We should also move __clk_get() into there
> >> because right now we have a hole where whoever calls
> >> of_clk_get_from_provider() never calls __clk_get() on the clk, leading
> >> to possible badness. v2 coming soon.
> > There's some other issues here too...
> >
> > sound/soc/kirkwood/kirkwood-i2s.c:
> >
> >         priv->clk = devm_clk_get(&pdev->dev, np ? "internal" : NULL);
> > ...
> >         priv->extclk = devm_clk_get(&pdev->dev, "extclk");
> >         if (IS_ERR(priv->extclk)) {
> > ...
> > 	} else {
> >                 if (priv->extclk == priv->clk) {
> >                         devm_clk_put(&pdev->dev, priv->extclk);
> >                         priv->extclk = ERR_PTR(-EINVAL);
> >                 } else {
> >                         dev_info(&pdev->dev, "found external clock\n");
> >                         clk_prepare_enable(priv->extclk);
> >                         soc_dai = kirkwood_i2s_dai_extclk;
> >                 }
> >
> > It should be fine provided your "trick" is only done for DT clocks,
> > but not for legacy - with legacy, a NULL in the clkdev tables will
> > match both these requests, hence the need to compare the clk_get()
> > return value to tell whether we get the same clock.
> >
> 
> Are we still talking about of_clk_get_from_provider()? Or are we talking
> about comparing struct clk pointers?

Comparing struct clk pointers, and the implications of the patch changing
the clk_get() et.al. to be unique struct clk pointers.

> From what I can tell this code is
> now broken because we made all clk getting functions (there's quite a
> few...) return unique pointers every time they're called. It seems that
> the driver wants to know if extclk and clk are the same so it can do
> something differently in kirkwood_set_rate(). Do we need some sort of
> clk_equal(struct clk *a, struct clk *b) function for drivers like this?

Well, the clocks in question are the SoC internal clock (which is more or
less fixed, but has a programmable divider) and an externally supplied
clock, and the IP has a multiplexer on its input which allows us to select
between those two sources.

If it were possible to bind both to the same clock, it wouldn't be a
useful configuration - nothing would be gained from doing so in terms of
available rates.

What the comparison is there for is to catch the case with legacy lookups
where a clkdev lookup entry with a NULL connection ID results in matching
any connection ID passed to clk_get().  If the patch changes this, then
we will have a regression - and this is something which needs fixing
_before_ we do this "return unique clocks".
diff mbox

Patch

diff --git a/drivers/clk/clkdev.c b/drivers/clk/clkdev.c
index 29a1ab7af4b8..00d747d09b2a 100644
--- a/drivers/clk/clkdev.c
+++ b/drivers/clk/clkdev.c
@@ -29,15 +29,8 @@  static DEFINE_MUTEX(clocks_mutex);
 
 #if defined(CONFIG_OF) && defined(CONFIG_COMMON_CLK)
 
-/**
- * of_clk_get_by_clkspec() - Lookup a clock form a clock provider
- * @clkspec: pointer to a clock specifier data structure
- *
- * This function looks up a struct clk from the registered list of clock
- * providers, an input is a clock specifier data structure as returned
- * from the of_parse_phandle_with_args() function call.
- */
-struct clk *of_clk_get_by_clkspec(struct of_phandle_args *clkspec)
+static struct clk *__of_clk_get_by_clkspec(struct of_phandle_args *clkspec,
+					 const char *dev_id, const char *con_id)
 {
 	struct clk *clk;
 
@@ -47,6 +40,8 @@  struct clk *of_clk_get_by_clkspec(struct of_phandle_args *clkspec)
 	of_clk_lock();
 	clk = __of_clk_get_from_provider(clkspec);
 
+	if (!IS_ERR(clk))
+		clk = __clk_create_clk(__clk_get_hw(clk), dev_id, con_id);
 	if (!IS_ERR(clk) && !__clk_get(clk))
 		clk = ERR_PTR(-ENOENT);
 
@@ -54,7 +49,21 @@  struct clk *of_clk_get_by_clkspec(struct of_phandle_args *clkspec)
 	return clk;
 }
 
-static struct clk *__of_clk_get(struct device_node *np, int index)
+/**
+ * of_clk_get_by_clkspec() - Lookup a clock form a clock provider
+ * @clkspec: pointer to a clock specifier data structure
+ *
+ * This function looks up a struct clk from the registered list of clock
+ * providers, an input is a clock specifier data structure as returned
+ * from the of_parse_phandle_with_args() function call.
+ */
+struct clk *of_clk_get_by_clkspec(struct of_phandle_args *clkspec)
+{
+	return __of_clk_get_by_clkspec(clkspec, NULL, __func__);
+}
+
+static struct clk *__of_clk_get(struct device_node *np, int index,
+			       const char *dev_id, const char *con_id)
 {
 	struct of_phandle_args clkspec;
 	struct clk *clk;
@@ -68,7 +77,7 @@  static struct clk *__of_clk_get(struct device_node *np, int index)
 	if (rc)
 		return ERR_PTR(rc);
 
-	clk = of_clk_get_by_clkspec(&clkspec);
+	clk = __of_clk_get_by_clkspec(&clkspec, dev_id, con_id);
 	of_node_put(clkspec.np);
 
 	return clk;
@@ -76,12 +85,7 @@  static struct clk *__of_clk_get(struct device_node *np, int index)
 
 struct clk *of_clk_get(struct device_node *np, int index)
 {
-	struct clk *clk = __of_clk_get(np, index);
-
-	if (!IS_ERR(clk))
-		clk = __clk_create_clk(__clk_get_hw(clk), np->full_name, NULL);
-
-	return clk;
+	return __of_clk_get(np, index, np->full_name, NULL);
 }
 EXPORT_SYMBOL(of_clk_get);
 
@@ -102,12 +106,10 @@  static struct clk *__of_clk_get_by_name(struct device_node *np,
 		 */
 		if (name)
 			index = of_property_match_string(np, "clock-names", name);
-		clk = __of_clk_get(np, index);
+		clk = __of_clk_get(np, index, dev_id, name);
 		if (!IS_ERR(clk)) {
-			clk = __clk_create_clk(__clk_get_hw(clk), dev_id, name);
 			break;
-		}
-		else if (name && index >= 0) {
+		} else if (name && index >= 0) {
 			if (PTR_ERR(clk) != -EPROBE_DEFER)
 				pr_err("ERROR: could not get clock %s:%s(%i)\n",
 					np->full_name, name ? name : "", index);