[PATCH] gpio: mvebu: use raw spinlock for MMIO regmaps

[PATCH] gpio: mvebu: use raw spinlock for MMIO regmaps

[Rosen Penev]

The mask/unmask handlers run with the generic irqchip's raw spinlock
held and call regmap_read()/regmap_write() on the GPIO regmaps. By
default regmap protects itself with a plain spinlock, which becomes a
sleeping rt_mutex on CONFIG_PREEMPT_RT; taking it while holding the raw
irqchip lock trips might_sleep(), and PROVE_RAW_LOCK_NESTING flags the
same nesting on non-RT builds.

Configure both GPIO regmaps with .use_raw_spinlock = true so the MMIO
accesses stay atomic in hardirq and raw-locked context.

Assisted-by: LLM
Signed-off-by: Rosen Penev <rosenp@gmail.com>
---
 drivers/gpio/gpio-mvebu.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/gpio/gpio-mvebu.c b/drivers/gpio/gpio-mvebu.c
index 03e977c2a629..674feeb31774 100644
--- a/drivers/gpio/gpio-mvebu.c
+++ b/drivers/gpio/gpio-mvebu.c
@@ -598,11 +598,18 @@ static void mvebu_gpio_irq_handler(struct irq_desc *desc)
 	chained_irq_exit(chip, desc);
 }
 
+/*
+ * These regmaps are accessed from the mask/unmask handlers while the
+ * irqchip's raw spinlock is held, so they must use a raw spinlock
+ * internally. A plain spinlock becomes a sleeping rt_mutex on
+ * CONFIG_PREEMPT_RT and would trip might_sleep() from that context.
+ */
 static const struct regmap_config mvebu_gpio_regmap_config = {
 	.name = "gpio",
 	.reg_bits = 32,
 	.reg_stride = 4,
 	.val_bits = 32,
+	.use_raw_spinlock = true,
 };
 
 static const struct regmap_config mvebu_gpio_percpu_regmap_config = {
@@ -610,6 +617,7 @@ static const struct regmap_config mvebu_gpio_percpu_regmap_config = {
 	.reg_bits = 32,
 	.reg_stride = 4,
 	.val_bits = 32,
+	.use_raw_spinlock = true,
 };
 
 static const struct regmap_config mvebu_pwm_regmap_config = {
-- 
2.55.0

Re: [PATCH] gpio: mvebu: use raw spinlock for MMIO regmaps

[Sebastian Andrzej Siewior]

On 2026-09-15 15:26:09 [-0700], Rosen Penev wrote:
> The mask/unmask handlers run with the generic irqchip's raw spinlock
> held and call regmap_read()/regmap_write() on the GPIO regmaps. By
> default regmap protects itself with a plain spinlock, which becomes a
> sleeping rt_mutex on CONFIG_PREEMPT_RT; taking it while holding the raw
  sleeping lock

> irqchip lock trips might_sleep(), and PROVE_RAW_LOCK_NESTING flags the
> same nesting on non-RT builds.
> 
> Configure both GPIO regmaps with .use_raw_spinlock = true so the MMIO
> accesses stay atomic in hardirq and raw-locked context.
> 
> Assisted-by: LLM
> Signed-off-by: Rosen Penev <rosenp@gmail.com>
> ---

other than the comments,

Reviewed-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>

>  drivers/gpio/gpio-mvebu.c | 8 ++++++++
>  1 file changed, 8 insertions(+)
> 
> diff --git a/drivers/gpio/gpio-mvebu.c b/drivers/gpio/gpio-mvebu.c
> index 03e977c2a629..674feeb31774 100644
> --- a/drivers/gpio/gpio-mvebu.c
> +++ b/drivers/gpio/gpio-mvebu.c
> @@ -598,11 +598,18 @@ static void mvebu_gpio_irq_handler(struct irq_desc *desc)
>  	chained_irq_exit(chip, desc);
>  }
>  
> +/*
> + * These regmaps are accessed from the mask/unmask handlers while the
> + * irqchip's raw spinlock is held, so they must use a raw spinlock
> + * internally. A plain spinlock becomes a sleeping rt_mutex on
> + * CONFIG_PREEMPT_RT and would trip might_sleep() from that context.
> + */

IRQ delivery path must hold raw_spinlock_t. This is not special to this
driver but must be done in every gpio chip. I wouldn't add this comment.

>  static const struct regmap_config mvebu_gpio_regmap_config = {
>  	.name = "gpio",
>  	.reg_bits = 32,
>  	.reg_stride = 4,
>  	.val_bits = 32,
> +	.use_raw_spinlock = true,
>  };
>  
>  static const struct regmap_config mvebu_gpio_percpu_regmap_config = {
> @@ -610,6 +617,7 @@ static const struct regmap_config mvebu_gpio_percpu_regmap_config = {
>  	.reg_bits = 32,
>  	.reg_stride = 4,
>  	.val_bits = 32,
> +	.use_raw_spinlock = true,
>  };
>  
>  static const struct regmap_config mvebu_pwm_regmap_config = {
> -- 
> 2.55.0

Sebastian