[PATCH 0/2] ARM: stm32: add clock for pwrcfg syscon

The STM32F4/7 pwrcfg syscon was missing its clock, making it impossible
to use after clk_disable_unused(). Simply adding the clock creates a
circular dependency, because the syscon is used by the clock driver.
This series resolves this dependency and then adds the clock.

I tested this on a STM32F746 and verified that syscon writes now
succeed even after clk_disable_unused().

Ben Wolsieffer (2):
  clk: stm32: initialize syscon after clocks are registered
  ARM: dts: stm32: add pwrcfg clock for stm32f4/7

 arch/arm/boot/dts/st/stm32f429.dtsi |  1 +
 arch/arm/boot/dts/st/stm32f746.dtsi |  1 +
 drivers/clk/clk-stm32f4.c           | 12 ++++++------
 3 files changed, 8 insertions(+), 6 deletions(-)

-- 
2.42.0

The stm32-power-config syscon (PWR peripheral) is used in this driver
and the STM32 RTC driver to enable write access to backup domain
registers. The syscon's clock has a gate controlled by this clock
driver, but this clock is currently not registered in the device tree.
This only happens to work currently because all relevant clock setup and
RTC initialization happens before clk_disabled_unused(). After this
point, all syscon register writes are ignored.

If we simply add the syscon clock in the device tree, we end up with a
circular dependency because the clock has not been registered at the
point this driver requests the syscon.

This patch avoids this circular dependency by moving the syscon lookup
after the clocks are registered. This does appear to create a possible
race condition where someone could attempt to perform an operation on a
backup domain clock before the syscon has been initialized. This would
result in the operation having no effect because backup domain writes
could not be enabled. I'm not sure if this is a problem or if there is
a way to avoid it.

Signed-off-by: Ben Wolsieffer <redacted>
---
 drivers/clk/clk-stm32f4.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/drivers/clk/clk-stm32f4.c b/drivers/clk/clk-stm32f4.c
index 07c13ebe327d..a88e762d2b5e 100644
--- a/drivers/clk/clk-stm32f4.c
+++ b/drivers/clk/clk-stm32f4.c

@@ -1697,12 +1697,6 @@ static void __init stm32f4_rcc_init(struct device_node *np)
 		return;
 	}
 
-	pdrm = syscon_regmap_lookup_by_phandle(np, "st,syscfg");
-	if (IS_ERR(pdrm)) {
-		pdrm = NULL;
-		pr_warn("%s: Unable to get syscfg\n", __func__);
-	}
-
 	match = of_match_node(stm32f4_of_match, np);
 	if (WARN_ON(!match))
 		return;

@@ -1894,6 +1888,12 @@ static void __init stm32f4_rcc_init(struct device_node *np)
 
 	of_clk_add_hw_provider(np, stm32f4_rcc_lookup_clk, NULL);
 
+	pdrm = syscon_regmap_lookup_by_phandle(np, "st,syscfg");
+	if (IS_ERR(pdrm)) {
+		pdrm = NULL;
+		pr_warn("%s: Unable to get syscfg\n", __func__);
+	}
+
 	return;
 fail:
 	kfree(clks);

-- 
2.42.0

Now that the circular dependency between syscon and clock initialization
has been resolved, we can add the clock that drives the pwrcfg syscon to
the device tree.

Signed-off-by: Ben Wolsieffer <redacted>
---
 arch/arm/boot/dts/st/stm32f429.dtsi | 1 +
 arch/arm/boot/dts/st/stm32f746.dtsi | 1 +
 2 files changed, 2 insertions(+)

diff --git a/arch/arm/boot/dts/st/stm32f429.dtsi b/arch/arm/boot/dts/st/stm32f429.dtsi
index 8efcda9ef8ae..7c9a9133cc86 100644
--- a/arch/arm/boot/dts/st/stm32f429.dtsi
+++ b/arch/arm/boot/dts/st/stm32f429.dtsi

@@ -665,6 +665,7 @@ spi6: spi@40015400 {
 		pwrcfg: power-config@40007000 {
 			compatible = "st,stm32-power-config", "syscon";
 			reg = <0x40007000 0x400>;
+			clocks = <&rcc 0 STM32F4_APB1_CLOCK(PWR)>;
 		};
 
 		ltdc: display-controller@40016800 {

diff --git a/arch/arm/boot/dts/st/stm32f746.dtsi b/arch/arm/boot/dts/st/stm32f746.dtsi
index d1802efd067c..cc8177466a51 100644
--- a/arch/arm/boot/dts/st/stm32f746.dtsi
+++ b/arch/arm/boot/dts/st/stm32f746.dtsi

@@ -510,6 +510,7 @@ pwm {
 		pwrcfg: power-config@40007000 {
 			compatible = "st,stm32-power-config", "syscon";
 			reg = <0x40007000 0x400>;
+			clocks = <&rcc 0 STM32F7_APB1_CLOCK(PWR)>;
 		};
 
 		crc: crc@40023000 {

-- 
2.42.0

Quoting Ben Wolsieffer (2023-10-02 11:08:53)

The stm32-power-config syscon (PWR peripheral) is used in this driver
and the STM32 RTC driver to enable write access to backup domain
registers. The syscon's clock has a gate controlled by this clock
driver, but this clock is currently not registered in the device tree.
This only happens to work currently because all relevant clock setup and
RTC initialization happens before clk_disabled_unused(). After this
point, all syscon register writes are ignored.

Seems like we should mark those clks as CLK_IGNORE_UNUSED and add a
comment to that fact.

If we simply add the syscon clock in the device tree, we end up with a
circular dependency because the clock has not been registered at the
point this driver requests the syscon.

This patch avoids this circular dependency by moving the syscon lookup
after the clocks are registered. This does appear to create a possible
race condition where someone could attempt to perform an operation on a
backup domain clock before the syscon has been initialized. This would
result in the operation having no effect because backup domain writes
could not be enabled. I'm not sure if this is a problem or if there is
a way to avoid it.

There's no comment in the code that says the regmap must be set there
instead of earlier. What's to stop someone from tripping over this
problem later? At the least, please add a comment.

On Sun, Dec 17, 2023 at 03:05:01PM -0800, Stephen Boyd wrote:

Quoting Ben Wolsieffer (2023-10-02 11:08:53)

quoted

The stm32-power-config syscon (PWR peripheral) is used in this driver
and the STM32 RTC driver to enable write access to backup domain
registers. The syscon's clock has a gate controlled by this clock
driver, but this clock is currently not registered in the device tree.
This only happens to work currently because all relevant clock setup and
RTC initialization happens before clk_disabled_unused(). After this
point, all syscon register writes are ignored.

Seems like we should mark those clks as CLK_IGNORE_UNUSED and add a
comment to that fact.

That seems like a worse solution than specifying the clock dependency in
the device tree.

quoted

If we simply add the syscon clock in the device tree, we end up with a
circular dependency because the clock has not been registered at the
point this driver requests the syscon.

This patch avoids this circular dependency by moving the syscon lookup
after the clocks are registered. This does appear to create a possible
race condition where someone could attempt to perform an operation on a
backup domain clock before the syscon has been initialized. This would
result in the operation having no effect because backup domain writes
could not be enabled. I'm not sure if this is a problem or if there is
a way to avoid it.

There's no comment in the code that says the regmap must be set there
instead of earlier. What's to stop someone from tripping over this
problem later? At the least, please add a comment.

Yeah, I'll fix that. Do you have any thoughts on the race condition I
described? Should I add some kind of locking to block
enable/disable_power_domain_write_protection() until stm32f4_rcc_init()
attempts to initialize the syscon?

Thank you, Ben

Quoting Ben Wolsieffer (2024-01-12 14:00:04)

On Sun, Dec 17, 2023 at 03:05:01PM -0800, Stephen Boyd wrote:

quoted

Quoting Ben Wolsieffer (2023-10-02 11:08:53)

quoted

The stm32-power-config syscon (PWR peripheral) is used in this driver
and the STM32 RTC driver to enable write access to backup domain
registers. The syscon's clock has a gate controlled by this clock
driver, but this clock is currently not registered in the device tree.
This only happens to work currently because all relevant clock setup and
RTC initialization happens before clk_disabled_unused(). After this
point, all syscon register writes are ignored.

Seems like we should mark those clks as CLK_IGNORE_UNUSED and add a
comment to that fact.

That seems like a worse solution than specifying the clock dependency in
the device tree.

quoted

quoted

If we simply add the syscon clock in the device tree, we end up with a
circular dependency because the clock has not been registered at the
point this driver requests the syscon.

This patch avoids this circular dependency by moving the syscon lookup
after the clocks are registered. This does appear to create a possible
race condition where someone could attempt to perform an operation on a
backup domain clock before the syscon has been initialized. This would
result in the operation having no effect because backup domain writes
could not be enabled. I'm not sure if this is a problem or if there is
a way to avoid it.

There's no comment in the code that says the regmap must be set there
instead of earlier. What's to stop someone from tripping over this
problem later? At the least, please add a comment.

Yeah, I'll fix that. Do you have any thoughts on the race condition I
described? Should I add some kind of locking to block
enable/disable_power_domain_write_protection() until stm32f4_rcc_init()
attempts to initialize the syscon?

Maybe. I don't really know and it's probably because I don't really
understand the problem. Maybe you can solve it by turning on the backup
domain clock manually when the driver probes via direct register writes
and then only publish the syscon once the backup domain clock is
registered?