2021-05-16 10:20:14

by Alexandru Ardelean

[permalink] [raw]
Subject: [PATCH] gpio: gpio-stmpe: fully use convert probe to device-managed

The driver doesn't look like it can be built as a kmod, so leaks cannot
happen via a rmmod mechanism.
The remove hook was removed via commit 3b52bb960ec6 ("gpio: stmpe: make
it explicitly non-modular").

The IRQ is registered via devm_request_threaded_irq(), making the driver
only partially device-managed.

In any case all resources should be made device-managed, mostly as a good
practice. That way at least the unwinding on error is happening in reverse
order (as the probe).

This change also removes platform_set_drvdata() since the information is
never retrieved to be used in the driver.

Signed-off-by: Alexandru Ardelean <[email protected]>
---

I'm not sure if this should be marked with a Fixes tag.
But if so, it should probably be for commit 3b52bb960ec6 (also mentioned in
the comment above).

drivers/gpio/gpio-stmpe.c | 32 +++++++++++++-------------------
1 file changed, 13 insertions(+), 19 deletions(-)

diff --git a/drivers/gpio/gpio-stmpe.c b/drivers/gpio/gpio-stmpe.c
index b94ef8181428..dd4d58b4ae49 100644
--- a/drivers/gpio/gpio-stmpe.c
+++ b/drivers/gpio/gpio-stmpe.c
@@ -449,6 +449,11 @@ static void stmpe_init_irq_valid_mask(struct gpio_chip *gc,
}
}

+static void stmpe_gpio_disable(void *stmpe)
+{
+ stmpe_disable(stmpe, STMPE_BLOCK_GPIO);
+}
+
static int stmpe_gpio_probe(struct platform_device *pdev)
{
struct stmpe *stmpe = dev_get_drvdata(pdev->dev.parent);
@@ -461,7 +466,7 @@ static int stmpe_gpio_probe(struct platform_device *pdev)
return -EINVAL;
}

- stmpe_gpio = kzalloc(sizeof(*stmpe_gpio), GFP_KERNEL);
+ stmpe_gpio = devm_kzalloc(&pdev->dev, sizeof(*stmpe_gpio), GFP_KERNEL);
if (!stmpe_gpio)
return -ENOMEM;

@@ -489,7 +494,11 @@ static int stmpe_gpio_probe(struct platform_device *pdev)

ret = stmpe_enable(stmpe, STMPE_BLOCK_GPIO);
if (ret)
- goto out_free;
+ return ret;
+
+ ret = devm_add_action_or_reset(&pdev->dev, stmpe_gpio_disable, stmpe);
+ if (ret)
+ return ret;

if (irq > 0) {
struct gpio_irq_chip *girq;
@@ -499,7 +508,7 @@ static int stmpe_gpio_probe(struct platform_device *pdev)
"stmpe-gpio", stmpe_gpio);
if (ret) {
dev_err(&pdev->dev, "unable to get irq: %d\n", ret);
- goto out_disable;
+ return ret;
}

girq = &stmpe_gpio->chip.irq;
@@ -514,22 +523,7 @@ static int stmpe_gpio_probe(struct platform_device *pdev)
girq->init_valid_mask = stmpe_init_irq_valid_mask;
}

- ret = gpiochip_add_data(&stmpe_gpio->chip, stmpe_gpio);
- if (ret) {
- dev_err(&pdev->dev, "unable to add gpiochip: %d\n", ret);
- goto out_disable;
- }
-
- platform_set_drvdata(pdev, stmpe_gpio);
-
- return 0;
-
-out_disable:
- stmpe_disable(stmpe, STMPE_BLOCK_GPIO);
- gpiochip_remove(&stmpe_gpio->chip);
-out_free:
- kfree(stmpe_gpio);
- return ret;
+ return devm_gpiochip_add_data(&pdev->dev, &stmpe_gpio->chip, stmpe_gpio);
}

static struct platform_driver stmpe_gpio_driver = {
--
2.31.1



2021-05-19 23:56:28

by Linus Walleij

[permalink] [raw]
Subject: Re: [PATCH] gpio: gpio-stmpe: fully use convert probe to device-managed

On Sun, May 16, 2021 at 8:14 AM Alexandru Ardelean
<[email protected]> wrote:

> The driver doesn't look like it can be built as a kmod, so leaks cannot
> happen via a rmmod mechanism.
> The remove hook was removed via commit 3b52bb960ec6 ("gpio: stmpe: make
> it explicitly non-modular").
>
> The IRQ is registered via devm_request_threaded_irq(), making the driver
> only partially device-managed.
>
> In any case all resources should be made device-managed, mostly as a good
> practice. That way at least the unwinding on error is happening in reverse
> order (as the probe).
>
> This change also removes platform_set_drvdata() since the information is
> never retrieved to be used in the driver.
>
> Signed-off-by: Alexandru Ardelean <[email protected]>

Nice!
Reviewed-by: Linus Walleij <[email protected]>

Yours,
Linus Walleij

2021-05-20 11:52:11

by Alexandru Ardelean

[permalink] [raw]
Subject: Re: [PATCH] gpio: gpio-stmpe: fully use convert probe to device-managed

On Thu, 20 May 2021 at 09:40, Andy Shevchenko <[email protected]> wrote:
>
>
>
> On Sunday, May 16, 2021, Alexandru Ardelean <[email protected]> wrote:
>>
>> The driver doesn't look like it can be built as a kmod, so leaks cannot
>> happen via a rmmod mechanism.
>> The remove hook was removed via commit 3b52bb960ec6 ("gpio: stmpe: make
>> it explicitly non-modular").
>>
>> The IRQ is registered via devm_request_threaded_irq(), making the driver
>> only partially device-managed.
>>
>> In any case all resources should be made device-managed, mostly as a good
>> practice. That way at least the unwinding on error is happening in reverse
>> order (as the probe).
>>
>> This change also removes platform_set_drvdata() since the information is
>> never retrieved to be used in the driver.
>
>
> Any driver can be unbind from device thru sysfs. The exception is when they (device drivers) specifically disable that.

Oh, I see.
Thanks for the info :)

>
>>
>> Signed-off-by: Alexandru Ardelean <[email protected]>
>> ---
>>
>> I'm not sure if this should be marked with a Fixes tag.
>> But if so, it should probably be for commit 3b52bb960ec6 (also mentioned in
>> the comment above).
>>
>> drivers/gpio/gpio-stmpe.c | 32 +++++++++++++-------------------
>> 1 file changed, 13 insertions(+), 19 deletions(-)
>>
>> diff --git a/drivers/gpio/gpio-stmpe.c b/drivers/gpio/gpio-stmpe.c
>> index b94ef8181428..dd4d58b4ae49 100644
>> --- a/drivers/gpio/gpio-stmpe.c
>> +++ b/drivers/gpio/gpio-stmpe.c
>> @@ -449,6 +449,11 @@ static void stmpe_init_irq_valid_mask(struct gpio_chip *gc,
>> }
>> }
>>
>> +static void stmpe_gpio_disable(void *stmpe)
>> +{
>> + stmpe_disable(stmpe, STMPE_BLOCK_GPIO);
>> +}
>> +
>> static int stmpe_gpio_probe(struct platform_device *pdev)
>> {
>> struct stmpe *stmpe = dev_get_drvdata(pdev->dev.parent);
>> @@ -461,7 +466,7 @@ static int stmpe_gpio_probe(struct platform_device *pdev)
>> return -EINVAL;
>> }
>>
>> - stmpe_gpio = kzalloc(sizeof(*stmpe_gpio), GFP_KERNEL);
>> + stmpe_gpio = devm_kzalloc(&pdev->dev, sizeof(*stmpe_gpio), GFP_KERNEL);
>> if (!stmpe_gpio)
>> return -ENOMEM;
>>
>> @@ -489,7 +494,11 @@ static int stmpe_gpio_probe(struct platform_device *pdev)
>>
>> ret = stmpe_enable(stmpe, STMPE_BLOCK_GPIO);
>> if (ret)
>> - goto out_free;
>> + return ret;
>> +
>> + ret = devm_add_action_or_reset(&pdev->dev, stmpe_gpio_disable, stmpe);
>> + if (ret)
>> + return ret;
>>
>> if (irq > 0) {
>> struct gpio_irq_chip *girq;
>> @@ -499,7 +508,7 @@ static int stmpe_gpio_probe(struct platform_device *pdev)
>> "stmpe-gpio", stmpe_gpio);
>> if (ret) {
>> dev_err(&pdev->dev, "unable to get irq: %d\n", ret);
>> - goto out_disable;
>> + return ret;
>> }
>>
>> girq = &stmpe_gpio->chip.irq;
>> @@ -514,22 +523,7 @@ static int stmpe_gpio_probe(struct platform_device *pdev)
>> girq->init_valid_mask = stmpe_init_irq_valid_mask;
>> }
>>
>> - ret = gpiochip_add_data(&stmpe_gpio->chip, stmpe_gpio);
>> - if (ret) {
>> - dev_err(&pdev->dev, "unable to add gpiochip: %d\n", ret);
>> - goto out_disable;
>> - }
>> -
>> - platform_set_drvdata(pdev, stmpe_gpio);
>> -
>> - return 0;
>> -
>> -out_disable:
>> - stmpe_disable(stmpe, STMPE_BLOCK_GPIO);
>> - gpiochip_remove(&stmpe_gpio->chip);
>> -out_free:
>> - kfree(stmpe_gpio);
>> - return ret;
>> + return devm_gpiochip_add_data(&pdev->dev, &stmpe_gpio->chip, stmpe_gpio);
>> }
>>
>> static struct platform_driver stmpe_gpio_driver = {
>> --
>> 2.31.1
>>
>
>
> --
> With Best Regards,
> Andy Shevchenko
>
>

2021-05-21 12:55:17

by Bartosz Golaszewski

[permalink] [raw]
Subject: Re: [PATCH] gpio: gpio-stmpe: fully use convert probe to device-managed

On Sun, May 16, 2021 at 8:14 AM Alexandru Ardelean
<[email protected]> wrote:
>
> The driver doesn't look like it can be built as a kmod, so leaks cannot
> happen via a rmmod mechanism.
> The remove hook was removed via commit 3b52bb960ec6 ("gpio: stmpe: make
> it explicitly non-modular").
>
> The IRQ is registered via devm_request_threaded_irq(), making the driver
> only partially device-managed.
>
> In any case all resources should be made device-managed, mostly as a good
> practice. That way at least the unwinding on error is happening in reverse
> order (as the probe).
>
> This change also removes platform_set_drvdata() since the information is
> never retrieved to be used in the driver.
>
> Signed-off-by: Alexandru Ardelean <[email protected]>

I applied the patch and tweaked the commit message because as Andy
pointed out - this driver can be unbound over sysfs.

Bart