Skip to contentSkip to conversation

mmc: pwrseq_simple: fix probe failure when CONFIG_RESET_GPIO is enabled

PATCH In review Sergey Suloev 5 participants 7 messages Feb 20 – Mar 9 MMC, SD and SDIO

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mmc: pwrseq_simple: fix probe failure when CONFIG_RESET_GPIO is enabled

The driver uses a heuristic — checking if exactly one entry exists in 'reset-gpios' — to decide whether to attempt reset controller API usage via devm_reset_control_get_optional_shared(). This is semantically incorrect and causes a permanent probe failure when CONFIG_RESET_GPIO is enabled.

When CONFIG_RESET_GPIO=y, __of_reset_control_get() intercepts the 'reset-gpios' DT property and attempts to register the GPIO as a reset controller via __reset_add_reset_gpio_device(). If the GPIO controller is not yet fully initialized at probe time, this returns -ENOENT instead of -EPROBE_DEFER, and the error path does not respect the 'optional' flag, causing pwrseq_simple to fail permanently rather than defer. This prevents mmc1 from probing on boards that use 'reset-gpios' for WiFi power sequencing, such as BananaPi M2 Magic with AP6212/BCM43430 on Allwinner R16.

The correct indicator of reset controller intent is the explicit presence of a 'resets' property in the device tree node, not the count of 'reset-gpios' entries. A single 'reset-gpios' entry is just as likely to be a plain GPIO as multiple entries.

Replace the ngpio == 1 heuristic with an explicit check for the 'resets' DT property, making the behavior unambiguous and immune to CONFIG_RESET_GPIO interference.

Fixes 73bf4b7381f7 ("mmc: pwrseq_simple: add support for one reset control")
---
 drivers/mmc/core/pwrseq_simple.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/mmc/core/pwrseq_simple.c +1 -3
index 4b47e6c3b04b..780c8818a273 100644
@@ -124,7 +124,6 @@ static int mmc_pwrseq_simple_probe(struct platform_device *pdev)
 {
 	struct mmc_pwrseq_simple *pwrseq;
 	struct device *dev = &pdev->dev;
-	int ngpio;
 
 	pwrseq = devm_kzalloc(dev, sizeof(*pwrseq), GFP_KERNEL);
 	if (!pwrseq)
@@ -134,8 +133,7 @@ static int mmc_pwrseq_simple_probe(struct platform_device *pdev)
 	if (IS_ERR(pwrseq->ext_clk) && PTR_ERR(pwrseq->ext_clk) != -ENOENT)
 		return dev_err_probe(dev, PTR_ERR(pwrseq->ext_clk), "external clock not ready\n");
 
-	ngpio = of_count_phandle_with_args(dev->of_node, "reset-gpios", "#gpio-cells");
-	if (ngpio == 1) {
+	if (device_property_present(dev, "resets")) {
 		pwrseq->reset_ctrl = devm_reset_control_get_optional_shared(dev, NULL);
 		if (IS_ERR(pwrseq->reset_ctrl))
 			return dev_err_probe(dev, PTR_ERR(pwrseq->reset_ctrl),
Signature · 1 line
2.43.0
#1 lore raw

+ gpio and reset maintainers

Quoted 45 lines · Sergey Suloev

The driver uses a heuristic — checking if exactly one entry exists in 'reset-gpios' — to decide whether to attempt reset controller API usage via devm_reset_control_get_optional_shared(). This is semantically incorrect and causes a permanent probe failure when CONFIG_RESET_GPIO is enabled.

When CONFIG_RESET_GPIO=y, __of_reset_control_get() intercepts the 'reset-gpios' DT property and attempts to register the GPIO as a reset controller via __reset_add_reset_gpio_device(). If the GPIO controller is not yet fully initialized at probe time, this returns -ENOENT instead of -EPROBE_DEFER, and the error path does not respect the 'optional' flag, causing pwrseq_simple to fail permanently rather than defer. This prevents mmc1 from probing on boards that use 'reset-gpios' for WiFi power sequencing, such as BananaPi M2 Magic with AP6212/BCM43430 on Allwinner R16.

The correct indicator of reset controller intent is the explicit presence of a 'resets' property in the device tree node, not the count of 'reset-gpios' entries. A single 'reset-gpios' entry is just as likely to be a plain GPIO as multiple entries.

Replace the ngpio == 1 heuristic with an explicit check for the 'resets' DT property, making the behavior unambiguous and immune to CONFIG_RESET_GPIO interference.

Fixes 73bf4b7381f7 ("mmc: pwrseq_simple: add support for one reset control")
---
 drivers/mmc/core/pwrseq_simple.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/mmc/core/pwrseq_simple.c +0 -1
index 4b47e6c3b04b..780c8818a273 100644
@@ -124,7 +124,6 @@ static int mmc_pwrseq_simple_probe(struct platform_device *pdev)
 {
        struct mmc_pwrseq_simple *pwrseq;
        struct device *dev = &pdev->dev;
-       int ngpio;
 
        pwrseq = devm_kzalloc(dev, sizeof(*pwrseq), GFP_KERNEL);
        if (!pwrseq)

drivers/mmc/core/pwrseq_simple.c static int mmc_pwrseq_simple_probe(struct platform_device *pdev) · line 139

Quoted 9 lines
drivers/mmc/core/pwrseq_simple.c +1 -2
@@ -134,8 +133,7 @@ static int mmc_pwrseq_simple_probe(struct platform_device *pdev)
        if (IS_ERR(pwrseq->ext_clk) && PTR_ERR(pwrseq->ext_clk) != -ENOENT)
                return dev_err_probe(dev, PTR_ERR(pwrseq->ext_clk), "external clock not ready\n");
 
-       ngpio = of_count_phandle_with_args(dev->of_node, "reset-gpios", "#gpio-cells");
-       if (ngpio == 1) {
+       if (device_property_present(dev, "resets")) {
                pwrseq->reset_ctrl = devm_reset_control_get_optional_shared(dev, NULL);
                if (IS_ERR(pwrseq->reset_ctrl))
                        return dev_err_probe(dev, PTR_ERR(pwrseq->reset_ctrl),

Indeed the existing code looks a bit weird, but checking for the "resets" property here should also require an update to the DT docs [1].

I have looped in the gpio/reset maintainers to get their opinion on this too.

Kind regards
Uffe

[1] Documentation/devicetree/bindings/mmc/mmc-pwrseq-simple.yaml

#2 lore raw
In reply to Sergey Suloev
Quoted 11 lines · Sergey Suloev

The driver uses a heuristic — checking if exactly one entry exists in 'reset-gpios' — to decide whether to attempt reset controller API usage via devm_reset_control_get_optional_shared(). This is semantically incorrect and causes a permanent probe failure when CONFIG_RESET_GPIO is enabled.

When CONFIG_RESET_GPIO=y, __of_reset_control_get() intercepts the 'reset-gpios' DT property and attempts to register the GPIO as a reset controller via __reset_add_reset_gpio_device(). If the GPIO controller is not yet fully initialized at probe time, this returns -ENOENT instead of -EPROBE_DEFER, and the error path does not respect the 'optional'

If this is what's happening, then it is not the expected behavior[1].

Quoted 10 lines

flag, causing pwrseq_simple to fail permanently rather than defer. This prevents mmc1 from probing on boards that use 'reset-gpios' for WiFi power sequencing, such as BananaPi M2 Magic with AP6212/BCM43430 on Allwinner R16.

The correct indicator of reset controller intent is the explicit presence of a 'resets' property in the device tree node, not the count of 'reset-gpios' entries. A single 'reset-gpios' entry is just as likely to be a plain GPIO as multiple entries.

Am I getting it right? You have a node that has both the "resets" property as well as "reset-gpios" with more than one entry? If that's the case then it should be fixed in reset core, not in a user because there'll surely be another one in no time.

Quoted 33 lines

Replace the ngpio == 1 heuristic with an explicit check for the 'resets' DT property, making the behavior unambiguous and immune to CONFIG_RESET_GPIO interference.

Fixes 73bf4b7381f7 ("mmc: pwrseq_simple: add support for one reset control")
---
 drivers/mmc/core/pwrseq_simple.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/mmc/core/pwrseq_simple.c +1 -4
index 4b47e6c3b04b..780c8818a273 100644
@@ -124,7 +124,6 @@ static int mmc_pwrseq_simple_probe(struct platform_device *pdev)
 {
 	struct mmc_pwrseq_simple *pwrseq;
 	struct device *dev = &pdev->dev;
-	int ngpio;
 
 	pwrseq = devm_kzalloc(dev, sizeof(*pwrseq), GFP_KERNEL);
 	if (!pwrseq)
@@ -134,8 +133,7 @@ static int mmc_pwrseq_simple_probe(struct platform_device *pdev)
 	if (IS_ERR(pwrseq->ext_clk) && PTR_ERR(pwrseq->ext_clk) != -ENOENT)
 		return dev_err_probe(dev, PTR_ERR(pwrseq->ext_clk), "external clock not ready\n");
 
-	ngpio = of_count_phandle_with_args(dev->of_node, "reset-gpios", "#gpio-cells");
-	if (ngpio == 1) {
+	if (device_property_present(dev, "resets")) {
 		pwrseq->reset_ctrl = devm_reset_control_get_optional_shared(dev, NULL);
 		if (IS_ERR(pwrseq->reset_ctrl))
 			return dev_err_probe(dev, PTR_ERR(pwrseq->reset_ctrl),
--
Quoted 4 lines

2.43.0

I think reset core should check if the client node has both "reset" and "reset-gpios". If so - it should assume the client wants to resolve "resets" and only fall-back to "reset-gpios" if there's no "resets" and number of GPIOs is 1.

Philipp, Krzysztof: what do you think?

Bart

[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/reset/core.c#n910

#3 lore raw
Quoted 8 lines · Bartosz Golaszewski
Quoted 6 lines · Sergey Suloev

The driver uses a heuristic — checking if exactly one entry exists in 'reset-gpios' — to decide whether to attempt reset controller API usage via devm_reset_control_get_optional_shared(). This is semantically incorrect and causes a permanent probe failure when CONFIG_RESET_GPIO is enabled.

Quoted 7 lines
Quoted 5 lines

When CONFIG_RESET_GPIO=y, __of_reset_control_get() intercepts the 'reset-gpios' DT property and attempts to register the GPIO as a reset controller via __reset_add_reset_gpio_device(). If the GPIO controller is not yet fully initialized at probe time, this returns -ENOENT instead of -EPROBE_DEFER, and the error path does not respect the 'optional'

If this is what's happening, then it is not the expected behavior[1].

Is this -ENOENT returned because of #gpio-cells != 2 [1]?

[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/reset/core.c#n883

It looks like it [2].

[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm/boot/dts/allwinner/sunxi-bananapi-m2-plus.dtsi#n103

Quoted 5 lines
Quoted 5 lines

flag, causing pwrseq_simple to fail permanently rather than defer. This prevents mmc1 from probing on boards that use 'reset-gpios' for WiFi power sequencing, such as BananaPi M2 Magic with AP6212/BCM43430 on Allwinner R16.

Quoted 8 lines
Quoted 5 lines

The correct indicator of reset controller intent is the explicit presence of a 'resets' property in the device tree node, not the count of 'reset-gpios' entries. A single 'reset-gpios' entry is just as likely to be a plain GPIO as multiple entries.

Am I getting it right? You have a node that has both the "resets" property as well as "reset-gpios" with more than one entry?

According to the mmc-pwrseq-simple binding doc, a "resets" property is not valid at all. Apparently, this is just a "reset-gpios" property with a single GPIO with three cells.

Quoted 29 lines

If that's the case then it should be fixed in reset core, not in a user because there'll surely be another one in no time.

Quoted 26 lines

Replace the ngpio == 1 heuristic with an explicit check for the 'resets' DT property, making the behavior unambiguous and immune to CONFIG_RESET_GPIO interference.

Fixes 73bf4b7381f7 ("mmc: pwrseq_simple: add support for one reset control")
---
 drivers/mmc/core/pwrseq_simple.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/mmc/core/pwrseq_simple.c +0 -1
index 4b47e6c3b04b..780c8818a273 100644
@@ -124,7 +124,6 @@ static int mmc_pwrseq_simple_probe(struct platform_device *pdev)
 {
 	struct mmc_pwrseq_simple *pwrseq;
 	struct device *dev = &pdev->dev;
-	int ngpio;
 
 	pwrseq = devm_kzalloc(dev, sizeof(*pwrseq), GFP_KERNEL);
 	if (!pwrseq)
@@ -134,8 +133,7 @@ static int mmc_pwrseq_simple_probe(struct platform_device *pdev)
 	if (IS_ERR(pwrseq->ext_clk) && PTR_ERR(pwrseq->ext_clk) != -ENOENT)
 		return dev_err_probe(dev, PTR_ERR(pwrseq->ext_clk), "external clock not ready\n");
 
Quoted 3 lines

drivers/mmc/core/pwrseq_simple.c static int mmc_pwrseq_simple_probe(struct platform_device *pdev) · line 136

Quoted 3 lines
drivers/mmc/core/pwrseq_simple.c +1 -2
-	ngpio = of_count_phandle_with_args(dev->of_node, "reset-gpios", "#gpio-cells");
-	if (ngpio == 1) {
+	if (device_property_present(dev, "resets")) {

This effectively turns the following into dead code because the binding does not allow a "resets" property.

Quoted 11 lines
Quoted 8 lines
drivers/mmc/core/pwrseq_simple.c +0 -1
 		pwrseq->reset_ctrl = devm_reset_control_get_optional_shared(dev, NULL);
 		if (IS_ERR(pwrseq->reset_ctrl))
 			return dev_err_probe(dev, PTR_ERR(pwrseq->reset_ctrl),
--

2.43.0

I think reset core should check if the client node has both "reset" and "reset-gpios".

Why check both? If the "resets" property exists, we use it [3]. "reset-gpios" is only checked as a fallback if the "resets" property is not present [4].

[3] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/reset/core.c#n1019
[4] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/reset/core.c#n1031

Quoted 3 lines

If so - it should assume the client wants to resolve "resets"
and only fall-back to "reset-gpios" if there's no "resets" and number of GPIOs is 1.

That is exactly what happens right now.

Quoted 1 line

Philipp, Krzysztof: what do you think?

I think this is a missing feature in reset-gpio.
This driver could fall back to to devm_gpiod_get_array() also if devm_reset_control_get_optional_shared() returns -ENOENT instead of erroring out (maybe -ENOTSUPP would be a better return code for [1]?).

regards
Philipp

#4 lore raw
Quoted 95 lines · Philipp Zabel
Quoted 8 lines · Bartosz Golaszewski
Quoted 6 lines · Sergey Suloev

The driver uses a heuristic — checking if exactly one entry exists in 'reset-gpios' — to decide whether to attempt reset controller API usage via devm_reset_control_get_optional_shared(). This is semantically incorrect and causes a permanent probe failure when CONFIG_RESET_GPIO is enabled.

Quoted 7 lines
Quoted 5 lines

When CONFIG_RESET_GPIO=y, __of_reset_control_get() intercepts the 'reset-gpios' DT property and attempts to register the GPIO as a reset controller via __reset_add_reset_gpio_device(). If the GPIO controller is not yet fully initialized at probe time, this returns -ENOENT instead of -EPROBE_DEFER, and the error path does not respect the 'optional'

If this is what's happening, then it is not the expected behavior[1].

Is this -ENOENT returned because of #gpio-cells != 2 [1]?

[1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/reset/core.c#n883

It looks like it [2].

[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm/boot/dts/allwinner/sunxi-bananapi-m2-plus.dtsi#n103

Quoted 5 lines
Quoted 5 lines

flag, causing pwrseq_simple to fail permanently rather than defer. This prevents mmc1 from probing on boards that use 'reset-gpios' for WiFi power sequencing, such as BananaPi M2 Magic with AP6212/BCM43430 on Allwinner R16.

Quoted 8 lines
Quoted 5 lines

The correct indicator of reset controller intent is the explicit presence of a 'resets' property in the device tree node, not the count of 'reset-gpios' entries. A single 'reset-gpios' entry is just as likely to be a plain GPIO as multiple entries.

Am I getting it right? You have a node that has both the "resets" property as well as "reset-gpios" with more than one entry?

According to the mmc-pwrseq-simple binding doc, a "resets" property is not valid at all. Apparently, this is just a "reset-gpios" property with a single GPIO with three cells.

Quoted 29 lines

If that's the case then it should be fixed in reset core, not in a user because there'll surely be another one in no time.

Quoted 26 lines

Replace the ngpio == 1 heuristic with an explicit check for the 'resets' DT property, making the behavior unambiguous and immune to CONFIG_RESET_GPIO interference.

Fixes 73bf4b7381f7 ("mmc: pwrseq_simple: add support for one reset control")
---
 drivers/mmc/core/pwrseq_simple.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/mmc/core/pwrseq_simple.c +0 -1
index 4b47e6c3b04b..780c8818a273 100644
@@ -124,7 +124,6 @@ static int mmc_pwrseq_simple_probe(struct platform_device *pdev)
 {
 	struct mmc_pwrseq_simple *pwrseq;
 	struct device *dev = &pdev->dev;
-	int ngpio;
 
 	pwrseq = devm_kzalloc(dev, sizeof(*pwrseq), GFP_KERNEL);
 	if (!pwrseq)
@@ -134,8 +133,7 @@ static int mmc_pwrseq_simple_probe(struct platform_device *pdev)
 	if (IS_ERR(pwrseq->ext_clk) && PTR_ERR(pwrseq->ext_clk) != -ENOENT)
 		return dev_err_probe(dev, PTR_ERR(pwrseq->ext_clk), "external clock not ready\n");
 
Quoted 3 lines

drivers/mmc/core/pwrseq_simple.c static int mmc_pwrseq_simple_probe(struct platform_device *pdev) · line 136

Quoted 3 lines
drivers/mmc/core/pwrseq_simple.c +1 -2
-	ngpio = of_count_phandle_with_args(dev->of_node, "reset-gpios", "#gpio-cells");
-	if (ngpio == 1) {
+	if (device_property_present(dev, "resets")) {

This effectively turns the following into dead code because the binding does not allow a "resets" property.

Quoted 11 lines
Quoted 8 lines
drivers/mmc/core/pwrseq_simple.c +0 -1
 		pwrseq->reset_ctrl = devm_reset_control_get_optional_shared(dev, NULL);
 		if (IS_ERR(pwrseq->reset_ctrl))
 			return dev_err_probe(dev, PTR_ERR(pwrseq->reset_ctrl),
--

2.43.0

I think reset core should check if the client node has both "reset" and "reset-gpios".

Why check both? If the "resets" property exists, we use it [3]. "reset-gpios" is only checked as a fallback if the "resets" property is not present [4].

Quoted 9 lines

[3] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/reset/core.c#n1019
[4] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/reset/core.c#n1031

Quoted 3 lines

If so - it should assume the client wants to resolve "resets"
and only fall-back to "reset-gpios" if there's no "resets" and number of GPIOs is 1.

That is exactly what happens right now.

Right.

Quoted 7 lines
Quoted 1 line

Philipp, Krzysztof: what do you think?

I think this is a missing feature in reset-gpio.
This driver could fall back to to devm_gpiod_get_array() also if devm_reset_control_get_optional_shared() returns -ENOENT instead of erroring out (maybe -ENOTSUPP would be a better return code for [1]?).

I did not write the initial version but Krzysztof: is there any reason we only support a single, two-cell GPIO? Would it be problematic to support any kind of reset GPIO setup?

Bartosz

#5 lore raw
Quoted 9 lines · Bartosz Golaszewski
Quoted 8 lines
Quoted 2 lines

Philipp, Krzysztof: what do you think?

I think this is a missing feature in reset-gpio.
This driver could fall back to to devm_gpiod_get_array() also if devm_reset_control_get_optional_shared() returns -ENOENT instead of erroring out (maybe -ENOTSUPP would be a better return code for [1]?).

Quoted 3 lines

I did not write the initial version but Krzysztof: is there any reason we only support a single, two-cell GPIO? Would it be problematic to support any kind of reset GPIO setup?

It was only due to limitation of my initial solution - see even a TODO note in reset/core.c. I think the problem I had is that GPIO specifier without flags (so cells=1) would mess with GPIO_LOOKUP(). And if we have cells=3, where do the flags are? If you assume flags are last argument, that's an implied ABI encoded in the reset core.

Well, heh, implied ABI that 2nd argument are flags is already there, so the code is partially hacky, I agree.

So honestly I believe that GPIO core should handle all this mapping of original reset-gpios property into platform data to reset-gpio driver. That's the answer for your question "would it be problematic".

For reference, original patch:
https://lore.kernel.org/all/[email protected]/

Best regards,
Krzysztof

#6 lore raw
Quoted 24 lines · Krzysztof Kozlowski
Quoted 9 lines · Bartosz Golaszewski
Quoted 8 lines
Quoted 2 lines

Philipp, Krzysztof: what do you think?

I think this is a missing feature in reset-gpio.
This driver could fall back to to devm_gpiod_get_array() also if devm_reset_control_get_optional_shared() returns -ENOENT instead of erroring out (maybe -ENOTSUPP would be a better return code for [1]?).

Quoted 3 lines

I did not write the initial version but Krzysztof: is there any reason we only support a single, two-cell GPIO? Would it be problematic to support any kind of reset GPIO setup?

It was only due to limitation of my initial solution - see even a TODO note in reset/core.c. I think the problem I had is that GPIO specifier without flags (so cells=1) would mess with GPIO_LOOKUP(). And if we have cells=3, where do the flags are? If you assume flags are last argument, that's an implied ABI encoded in the reset core.

Well, heh, implied ABI that 2nd argument are flags is already there, so the code is partially hacky, I agree.

Quoted 7 lines

So honestly I believe that GPIO core should handle all this mapping of original reset-gpios property into platform data to reset-gpio driver. That's the answer for your question "would it be problematic".

For reference, original patch:
https://lore.kernel.org/all/[email protected]/

Philipp just picked up the reset core rework, once it's in linux-next, I'll try to extend the support to accept multiple entries in reset-gpios.

Bart