[1/2] clk: aspeed: add critical clock setting logic
diff mbox series

Message ID 20200115212639.4998-2-jae.hyun.yoo@linux.intel.com
State New
Headers show
Series
  • clk: aspeed/ast2600: Add critical clock setting logic
Related show

Commit Message

Jae Hyun Yoo Jan. 15, 2020, 9:26 p.m. UTC
This commit adds critical clock setting logic that applies
CLK_IS_CRITICAL flag if it detects 'clock-critical' property in
device tree.

Signed-off-by: Jae Hyun Yoo <jae.hyun.yoo@linux.intel.com>
---
 drivers/clk/clk-aspeed.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

Comments

Joel Stanley Jan. 16, 2020, 1:38 a.m. UTC | #1
On Wed, 15 Jan 2020 at 21:25, Jae Hyun Yoo <jae.hyun.yoo@linux.intel.com> wrote:
>
> This commit adds critical clock setting logic that applies
> CLK_IS_CRITICAL flag if it detects 'clock-critical' property in
> device tree.
>
> Signed-off-by: Jae Hyun Yoo <jae.hyun.yoo@linux.intel.com>
> ---
>  drivers/clk/clk-aspeed.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/clk/clk-aspeed.c b/drivers/clk/clk-aspeed.c
> index 411ff5fb2c07..d22eeb574ede 100644
> --- a/drivers/clk/clk-aspeed.c
> +++ b/drivers/clk/clk-aspeed.c
> @@ -541,8 +541,11 @@ static int aspeed_clk_probe(struct platform_device *pdev)
>
>         for (i = 0; i < ARRAY_SIZE(aspeed_gates); i++) {
>                 const struct aspeed_gate_data *gd = &aspeed_gates[i];
> +               unsigned long flags = gd->flags;
>                 u32 gate_flags;
>
> +               of_clk_detect_critical(pdev->dev.of_node, i, &flags);

This makes sense to me.

> +
>                 /* Special case: the USB port 1 clock (bit 14) is always
>                  * working the opposite way from the other ones.
>                  */
> @@ -550,7 +553,7 @@ static int aspeed_clk_probe(struct platform_device *pdev)
>                 hw = aspeed_clk_hw_register_gate(dev,
>                                 gd->name,
>                                 gd->parent_name,
> -                               gd->flags,
> +                               flags,

For completeness should we make this

 flags | gd->flags

>                                 map,
>                                 gd->clock_idx,
>                                 gd->reset_idx,
> --
> 2.17.1
>
Jae Hyun Yoo Jan. 16, 2020, 1:44 a.m. UTC | #2
On 1/15/2020 5:38 PM, Joel Stanley wrote:
> On Wed, 15 Jan 2020 at 21:25, Jae Hyun Yoo <jae.hyun.yoo@linux.intel.com> wrote:
>>
>> This commit adds critical clock setting logic that applies
>> CLK_IS_CRITICAL flag if it detects 'clock-critical' property in
>> device tree.
>>
>> Signed-off-by: Jae Hyun Yoo <jae.hyun.yoo@linux.intel.com>
>> ---
>>   drivers/clk/clk-aspeed.c | 5 ++++-
>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/clk/clk-aspeed.c b/drivers/clk/clk-aspeed.c
>> index 411ff5fb2c07..d22eeb574ede 100644
>> --- a/drivers/clk/clk-aspeed.c
>> +++ b/drivers/clk/clk-aspeed.c
>> @@ -541,8 +541,11 @@ static int aspeed_clk_probe(struct platform_device *pdev)
>>
>>          for (i = 0; i < ARRAY_SIZE(aspeed_gates); i++) {
>>                  const struct aspeed_gate_data *gd = &aspeed_gates[i];
>> +               unsigned long flags = gd->flags;
>>                  u32 gate_flags;
>>
>> +               of_clk_detect_critical(pdev->dev.of_node, i, &flags);
> 
> This makes sense to me.
> 
>> +
>>                  /* Special case: the USB port 1 clock (bit 14) is always
>>                   * working the opposite way from the other ones.
>>                   */
>> @@ -550,7 +553,7 @@ static int aspeed_clk_probe(struct platform_device *pdev)
>>                  hw = aspeed_clk_hw_register_gate(dev,
>>                                  gd->name,
>>                                  gd->parent_name,
>> -                               gd->flags,
>> +                               flags,
> 
> For completeness should we make this
> 
>   flags | gd->flags

Not needed. of_clk_detect_critical uses OR operation inside.

>>                                  map,
>>                                  gd->clock_idx,
>>                                  gd->reset_idx,
>> --
>> 2.17.1
>>
Paul Menzel Jan. 16, 2020, 9:57 a.m. UTC | #3
Dear Jae,


On 2020-01-15 22:26, Jae Hyun Yoo wrote:
> This commit adds critical clock setting logic that applies
> CLK_IS_CRITICAL flag if it detects 'clock-critical' property in
> device tree.

Tested how?

> Signed-off-by: Jae Hyun Yoo <jae.hyun.yoo@linux.intel.com>
> ---
>  drivers/clk/clk-aspeed.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/clk/clk-aspeed.c b/drivers/clk/clk-aspeed.c
> index 411ff5fb2c07..d22eeb574ede 100644
> --- a/drivers/clk/clk-aspeed.c
> +++ b/drivers/clk/clk-aspeed.c
> @@ -541,8 +541,11 @@ static int aspeed_clk_probe(struct platform_device *pdev)
>  
>  	for (i = 0; i < ARRAY_SIZE(aspeed_gates); i++) {
>  		const struct aspeed_gate_data *gd = &aspeed_gates[i];
> +		unsigned long flags = gd->flags;
>  		u32 gate_flags;
>  
> +		of_clk_detect_critical(pdev->dev.of_node, i, &flags);
> +

The function description in `drivers/clk/clk.c` has the warning below.

>  * Do not use this function. It exists only for legacy Device Tree
>  * bindings, such as the one-clock-per-node style that are outdated.
>  * Those bindings typically put all clock data into .dts and the Linux
>  * driver has no clock data, thus making it impossible to set this flag
>  * correctly from the driver. Only those drivers may call
>  * of_clk_detect_critical from their setup functions.

Will this still work?

>  		/* Special case: the USB port 1 clock (bit 14) is always
>  		 * working the opposite way from the other ones.
>  		 */
> @@ -550,7 +553,7 @@ static int aspeed_clk_probe(struct platform_device *pdev)
>  		hw = aspeed_clk_hw_register_gate(dev,
>  				gd->name,
>  				gd->parent_name,
> -				gd->flags,
> +				flags,
>  				map,
>  				gd->clock_idx,
>  				gd->reset_idx,


Kind regards,

Paul
Jae Hyun Yoo Jan. 16, 2020, 6:41 p.m. UTC | #4
Dear Paul,

On 1/16/2020 1:57 AM, Paul Menzel wrote:
> Dear Jae,
> 
> 
> On 2020-01-15 22:26, Jae Hyun Yoo wrote:
>> This commit adds critical clock setting logic that applies
>> CLK_IS_CRITICAL flag if it detects 'clock-critical' property in
>> device tree.
> 
> Tested how?

I added in the cover letter how I tested it. For an example, BCLK
can have the flag if I add below setting into one of
'aspeed-bmc-*.dts' files.

&syscon {
	clock-critical = <ASPEED_CLK_GATE_BCLK>;
};

>> Signed-off-by: Jae Hyun Yoo <jae.hyun.yoo@linux.intel.com>
>> ---
>>   drivers/clk/clk-aspeed.c | 5 ++++-
>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/clk/clk-aspeed.c b/drivers/clk/clk-aspeed.c
>> index 411ff5fb2c07..d22eeb574ede 100644
>> --- a/drivers/clk/clk-aspeed.c
>> +++ b/drivers/clk/clk-aspeed.c
>> @@ -541,8 +541,11 @@ static int aspeed_clk_probe(struct platform_device *pdev)
>>   
>>   	for (i = 0; i < ARRAY_SIZE(aspeed_gates); i++) {
>>   		const struct aspeed_gate_data *gd = &aspeed_gates[i];
>> +		unsigned long flags = gd->flags;
>>   		u32 gate_flags;
>>   
>> +		of_clk_detect_critical(pdev->dev.of_node, i, &flags);
>> +
> 
> The function description in `drivers/clk/clk.c` has the warning below.
> 
>>   * Do not use this function. It exists only for legacy Device Tree
>>   * bindings, such as the one-clock-per-node style that are outdated.
>>   * Those bindings typically put all clock data into .dts and the Linux
>>   * driver has no clock data, thus making it impossible to set this flag
>>   * correctly from the driver. Only those drivers may call
>>   * of_clk_detect_critical from their setup functions.
> 
> Will this still work?

At least, it still works now and still useful for this case. Actually, I
made this change as an alternative way of
https://www.spinics.net/lists/linux-clk/msg44836.html
because not all Aspeed BMC systems enable BCLK as a critical clock, so
it's for providing more flexible way of critical clock setting for
various hardware configurations.

If the function is deprecated and is going to be removed soon, would it
be acceptable if I add the 'critical-clock' parsing code into this
driver module instead of using the function?

Best Regards,

Jae

>>   		/* Special case: the USB port 1 clock (bit 14) is always
>>   		 * working the opposite way from the other ones.
>>   		 */
>> @@ -550,7 +553,7 @@ static int aspeed_clk_probe(struct platform_device *pdev)
>>   		hw = aspeed_clk_hw_register_gate(dev,
>>   				gd->name,
>>   				gd->parent_name,
>> -				gd->flags,
>> +				flags,
>>   				map,
>>   				gd->clock_idx,
>>   				gd->reset_idx,
> 
> 
> Kind regards,
> 
> Paul
>
Stephen Boyd Jan. 30, 2020, 5:42 p.m. UTC | #5
Quoting Jae Hyun Yoo (2020-01-15 13:26:38)
> This commit adds critical clock setting logic that applies
> CLK_IS_CRITICAL flag if it detects 'clock-critical' property in
> device tree.

Yes that is what the patch does. The commit text is supposed to explain
_why_ the patch is important. Please read "The canonical patch format"
from Documentation/process/submitting-patches.rst to understand what is
expected.

> 
> Signed-off-by: Jae Hyun Yoo <jae.hyun.yoo@linux.intel.com>
> ---
>  drivers/clk/clk-aspeed.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/clk/clk-aspeed.c b/drivers/clk/clk-aspeed.c
> index 411ff5fb2c07..d22eeb574ede 100644
> --- a/drivers/clk/clk-aspeed.c
> +++ b/drivers/clk/clk-aspeed.c
> @@ -541,8 +541,11 @@ static int aspeed_clk_probe(struct platform_device *pdev)
>  
>         for (i = 0; i < ARRAY_SIZE(aspeed_gates); i++) {
>                 const struct aspeed_gate_data *gd = &aspeed_gates[i];
> +               unsigned long flags = gd->flags;
>                 u32 gate_flags;
>  
> +               of_clk_detect_critical(pdev->dev.of_node, i, &flags);
> +

Do you need clks to be critical, but only sometimes? What clks need to
be critical? Why aren't there drivers for those clks that turn them on
as necessary?

There was a lengthy discussion years ago on the list about this function
and how it's not supposed to be used in newer code. Maybe we need to
revisit that discussion and conclude that sometimes we actually do need
clks to be turned on and kept on because we'll never have a driver for
them in the kernel. Similar to how pinctrl has pin hogs.
Jae Hyun Yoo Jan. 30, 2020, 8:59 p.m. UTC | #6
Hi Stephen,

On 1/30/2020 9:42 AM, Stephen Boyd wrote:
> Quoting Jae Hyun Yoo (2020-01-15 13:26:38)
>> This commit adds critical clock setting logic that applies
>> CLK_IS_CRITICAL flag if it detects 'clock-critical' property in
>> device tree.
> 
> Yes that is what the patch does. The commit text is supposed to explain
> _why_ the patch is important. Please read "The canonical patch format"
> from Documentation/process/submitting-patches.rst to understand what is
> expected.

I see. I'll add more detailed summary.

>>
>> Signed-off-by: Jae Hyun Yoo <jae.hyun.yoo@linux.intel.com>
>> ---
>>   drivers/clk/clk-aspeed.c | 5 ++++-
>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/clk/clk-aspeed.c b/drivers/clk/clk-aspeed.c
>> index 411ff5fb2c07..d22eeb574ede 100644
>> --- a/drivers/clk/clk-aspeed.c
>> +++ b/drivers/clk/clk-aspeed.c
>> @@ -541,8 +541,11 @@ static int aspeed_clk_probe(struct platform_device *pdev)
>>   
>>          for (i = 0; i < ARRAY_SIZE(aspeed_gates); i++) {
>>                  const struct aspeed_gate_data *gd = &aspeed_gates[i];
>> +               unsigned long flags = gd->flags;
>>                  u32 gate_flags;
>>   
>> +               of_clk_detect_critical(pdev->dev.of_node, i, &flags);
>> +
> 
> Do you need clks to be critical, but only sometimes? What clks need to
> be critical? Why aren't there drivers for those clks that turn them on
> as necessary?
> 
> There was a lengthy discussion years ago on the list about this function
> and how it's not supposed to be used in newer code. Maybe we need to
> revisit that discussion and conclude that sometimes we actually do need
> clks to be turned on and kept on because we'll never have a driver for
> them in the kernel. Similar to how pinctrl has pin hogs.

Yes, I need to make BCLK as a critical only for specific platforms. BCLK
is for controllers that are connected through PCI or PCIe bus to the
host machine such as VGA controller, 2D graphics engine and P2A bridge
in this SoC.

I'm currently trying to enable VGA controller which is actually
independent from Aspeed BMC SoC. It means that the VGA can be reset
only when either host PCI bus reset or host power-on reset is asserted.
Basically, VGA hardware module is controlled by the host machine not by
the Aspeed BMC SoC.

I submitted this patch as an alternative solution of
https://www.spinics.net/lists/linux-clk/msg44836.html because there
could be use cases that intentionally disable the VGA controller
depend on hardware design. So it'd be helpful for reducing power
consumption and for allocating more generic memory space instead of
allocating dedicated VGA shared memory if we can flexibly config the
BCLK.

I think, we don't need to add VGA driver just for enabling the clock
because the VGA controller is actually controlled by host machine as I
explained above so I made this patch set instead.

I agree with you that we need to revisit the discussion.

Thanks,

Jae

Patch
diff mbox series

diff --git a/drivers/clk/clk-aspeed.c b/drivers/clk/clk-aspeed.c
index 411ff5fb2c07..d22eeb574ede 100644
--- a/drivers/clk/clk-aspeed.c
+++ b/drivers/clk/clk-aspeed.c
@@ -541,8 +541,11 @@  static int aspeed_clk_probe(struct platform_device *pdev)
 
 	for (i = 0; i < ARRAY_SIZE(aspeed_gates); i++) {
 		const struct aspeed_gate_data *gd = &aspeed_gates[i];
+		unsigned long flags = gd->flags;
 		u32 gate_flags;
 
+		of_clk_detect_critical(pdev->dev.of_node, i, &flags);
+
 		/* Special case: the USB port 1 clock (bit 14) is always
 		 * working the opposite way from the other ones.
 		 */
@@ -550,7 +553,7 @@  static int aspeed_clk_probe(struct platform_device *pdev)
 		hw = aspeed_clk_hw_register_gate(dev,
 				gd->name,
 				gd->parent_name,
-				gd->flags,
+				flags,
 				map,
 				gd->clock_idx,
 				gd->reset_idx,