[1/2] clk: Introduce CLK_ASSUME_ENABLED_WHEN_UNUSED

Commit Message

Bjorn Andersson Dec. 3, 2021, 3:54 a.m. UTC

Some clock implementations doesn't provide means of implementing
is_enabled(), but still requires to be explicitly disabled when found
unused as part of clk_disable_unused().

One such set of clocks are Qualcomm's display RCGs. These can be enabled
and disabled automatically by the hardware, so it's not possible to
reliably query their configuration. Further more, these clocks need to
be disabled when unused, to allow them to be "parked" onto a safe
parent. Failure to disable the RCG results in the hardware locking up as
clk_disable_unused() traverses up the tree and turns off its source
clocks.

Add a new flag, CLK_ASSUME_ENABLED_BOOT, which clock drivers can use to
signal that these clocks should be disabled even if they don't implement
the is_enabled() ops.

Signed-off-by: Bjorn Andersson <[e-mail]>
---
 drivers/clk/clk.c            | 2 +-
 include/linux/clk-provider.h | 2 ++
 2 files changed, 3 insertions(+), 1 deletion(-)

Comments

Vinod Koul Dec. 7, 2021, 5:04 a.m. UTC | #1

On 02-12-21, 19:54, Bjorn Andersson wrote:
> Some clock implementations doesn't provide means of implementing
> is_enabled(), but still requires to be explicitly disabled when found
> unused as part of clk_disable_unused().
> 
> One such set of clocks are Qualcomm's display RCGs. These can be enabled
> and disabled automatically by the hardware, so it's not possible to
> reliably query their configuration. Further more, these clocks need to
> be disabled when unused, to allow them to be "parked" onto a safe
> parent. Failure to disable the RCG results in the hardware locking up as
> clk_disable_unused() traverses up the tree and turns off its source
> clocks.
> 
> Add a new flag, CLK_ASSUME_ENABLED_BOOT, which clock drivers can use to
> signal that these clocks should be disabled even if they don't implement
> the is_enabled() ops.

I think the name is bit long, but can't think of anything better. Btw
the explanation in log is _very_ good. I think we need to capture this
in the header file below as well.

Reviewed-by: Vinod Koul <[e-mail]>

> Signed-off-by: Bjorn Andersson <[e-mail]>
> ---
>  drivers/clk/clk.c            | 2 +-
>  include/linux/clk-provider.h | 2 ++
>  2 files changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c
> index f467d63bbf1e..e0bb53cbd4c8 100644
> --- a/drivers/clk/clk.c
> +++ b/drivers/clk/clk.c
> @@ -1265,7 +1265,7 @@ static void __init clk_disable_unused_subtree(struct clk_core *core)
>  	 * sequence.  call .disable_unused if available, otherwise fall
>  	 * back to .disable
>  	 */
> -	if (clk_core_is_enabled(core)) {
> +	if (clk_core_is_enabled(core) || core->flags & CLK_ASSUME_ENABLED_WHEN_UNUSED) {
>  		trace_clk_disable(core);
>  		if (core->ops->disable_unused)
>  			core->ops->disable_unused(core->hw);
> diff --git a/include/linux/clk-provider.h b/include/linux/clk-provider.h
> index f59c875271a0..7661cce31fa1 100644
> --- a/include/linux/clk-provider.h
> +++ b/include/linux/clk-provider.h
> @@ -32,6 +32,8 @@
>  #define CLK_OPS_PARENT_ENABLE	BIT(12)
>  /* duty cycle call may be forwarded to the parent clock */
>  #define CLK_DUTY_CYCLE_PARENT	BIT(13)
> +/* assume clock is enabled if found unused in late init */
> +#define CLK_ASSUME_ENABLED_WHEN_UNUSED	BIT(14)
>  
>  struct clk;
>  struct clk_hw;
> -- 
> 2.33.1

Stephen Boyd Dec. 9, 2021, 8:56 a.m. UTC | #2

Quoting Bjorn Andersson (2021-12-02 19:54:35)
> Some clock implementations doesn't provide means of implementing
> is_enabled(), but still requires to be explicitly disabled when found
> unused as part of clk_disable_unused().
> 
> One such set of clocks are Qualcomm's display RCGs. These can be enabled
> and disabled automatically by the hardware, so it's not possible to
> reliably query their configuration. Further more, these clocks need to
> be disabled when unused, to allow them to be "parked" onto a safe
> parent. Failure to disable the RCG results in the hardware locking up as
> clk_disable_unused() traverses up the tree and turns off its source
> clocks.
> 
> Add a new flag, CLK_ASSUME_ENABLED_BOOT, which clock drivers can use to
> signal that these clocks should be disabled even if they don't implement
> the is_enabled() ops.
> 
> Signed-off-by: Bjorn Andersson <[e-mail]>
> ---

I'm inclined to remove the disable unused logic. It is the main cause of
problems in the clk framework and with android pushing everyone to use
modules it's become a more broken design in need of an actual fix.  The
best approach is probably to just rip it out and start over, kicking off
the process for someone to fix the power regression of any clks that are
left enabled at boot. Or we can take the regulator approach and delay
disabling for 30 seconds and keep it around.

I'd prefer we take the approach of parking clks at init instead as
Dmitry proposed. It will break continuous splash screen but I don't
think that's being used anyway?

Patch

diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c
index f467d63bbf1e..e0bb53cbd4c8 100644
--- a/drivers/clk/clk.c
+++ b/drivers/clk/clk.c
@@ -1265,7 +1265,7 @@  static void __init clk_disable_unused_subtree(struct clk_core *core)
 	 * sequence.  call .disable_unused if available, otherwise fall
 	 * back to .disable
 	 */
-	if (clk_core_is_enabled(core)) {
+	if (clk_core_is_enabled(core) || core->flags & CLK_ASSUME_ENABLED_WHEN_UNUSED) {
 		trace_clk_disable(core);
 		if (core->ops->disable_unused)
 			core->ops->disable_unused(core->hw);
diff --git a/include/linux/clk-provider.h b/include/linux/clk-provider.h
index f59c875271a0..7661cce31fa1 100644
--- a/include/linux/clk-provider.h
+++ b/include/linux/clk-provider.h
@@ -32,6 +32,8 @@ 
 #define CLK_OPS_PARENT_ENABLE	BIT(12)
 /* duty cycle call may be forwarded to the parent clock */
 #define CLK_DUTY_CYCLE_PARENT	BIT(13)
+/* assume clock is enabled if found unused in late init */
+#define CLK_ASSUME_ENABLED_WHEN_UNUSED	BIT(14)
 
 struct clk;
 struct clk_hw;