The rzt2h_gpio_get_direction() function is called from
gpiod_get_direction(), which ends up being used within the __setup_irq()
call stack when requesting an interrupt.

__setup_irq() holds a raw_spinlock_t with IRQs disabled, which creates
an atomic context. spinlock_t cannot be used within atomic context
when PREEMPT_RT is enabled, since it may become a sleeping lock.

An "[ BUG: Invalid wait context ]" splat is observed when running with
CONFIG_PROVE_LOCKING enabled, describing exactly the aforementioned call
stack.

__setup_irq() needs to hold a raw_spinlock_t with IRQs disabled to
serialize access against a concurrent hard interrupt.

Switch to raw_spinlock_t to fix this.

Fixes: 829dde3369a9 ("pinctrl: renesas: rzt2h: Add GPIO IRQ chip to handle interrupts")
Signed-off-by: Cosmin Tanislav <cosmin-gabriel.tanislav.xa@renesas.com>
---

V2:
 * reword commit message to better describe why
   rzt2h_gpio_get_direction() needs to use raw_spinlock_t

 drivers/pinctrl/renesas/pinctrl-rzt2h.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/drivers/pinctrl/renesas/pinctrl-rzt2h.c b/drivers/pinctrl/renesas/pinctrl-rzt2h.c
index 9949108a35bb..d810dc0dae95 100644
--- a/drivers/pinctrl/renesas/pinctrl-rzt2h.c
+++ b/drivers/pinctrl/renesas/pinctrl-rzt2h.c
@@ -85,7 +85,7 @@ struct rzt2h_pinctrl {
 	struct gpio_chip		gpio_chip;
 	struct pinctrl_gpio_range	gpio_range;
 	DECLARE_BITMAP(used_irqs, RZT2H_INTERRUPTS_NUM);
-	spinlock_t			lock; /* lock read/write registers */
+	raw_spinlock_t			lock; /* lock read/write registers */
 	struct mutex			mutex; /* serialize adding groups and functions */
 	bool				safety_port_enabled;
 	atomic_t			wakeup_path;
@@ -145,7 +145,7 @@ static void rzt2h_pinctrl_set_pfc_mode(struct rzt2h_pinctrl *pctrl,
 	u64 reg64;
 	u16 reg16;
 
-	guard(spinlock_irqsave)(&pctrl->lock);
+	guard(raw_spinlock_irqsave)(&pctrl->lock);
 
 	/* Set pin to 'Non-use (Hi-Z input protection)'  */
 	reg16 = rzt2h_pinctrl_readw(pctrl, port, PM(port));
@@ -474,7 +474,7 @@ static int rzt2h_gpio_request(struct gpio_chip *chip, unsigned int offset)
 	if (ret)
 		return ret;
 
-	guard(spinlock_irqsave)(&pctrl->lock);
+	guard(raw_spinlock_irqsave)(&pctrl->lock);
 
 	/* Select GPIO mode in PMC Register */
 	rzt2h_pinctrl_set_gpio_en(pctrl, port, bit, true);
@@ -487,7 +487,7 @@ static void rzt2h_gpio_set_direction(struct rzt2h_pinctrl *pctrl, u32 port,
 {
 	u16 reg;
 
-	guard(spinlock_irqsave)(&pctrl->lock);
+	guard(raw_spinlock_irqsave)(&pctrl->lock);
 
 	reg = rzt2h_pinctrl_readw(pctrl, port, PM(port));
 	reg &= ~PM_PIN_MASK(bit);
@@ -509,7 +509,7 @@ static int rzt2h_gpio_get_direction(struct gpio_chip *chip, unsigned int offset)
 	if (ret)
 		return ret;
 
-	guard(spinlock_irqsave)(&pctrl->lock);
+	guard(raw_spinlock_irqsave)(&pctrl->lock);
 
 	if (rzt2h_pinctrl_readb(pctrl, port, PMC(port)) & BIT(bit)) {
 		/*
@@ -547,7 +547,7 @@ static int rzt2h_gpio_set(struct gpio_chip *chip, unsigned int offset,
 	u8 bit = RZT2H_PIN_ID_TO_PIN(offset);
 	u8 reg;
 
-	guard(spinlock_irqsave)(&pctrl->lock);
+	guard(raw_spinlock_irqsave)(&pctrl->lock);
 
 	reg = rzt2h_pinctrl_readb(pctrl, port, P(port));
 	if (value)
@@ -964,7 +964,7 @@ static int rzt2h_pinctrl_probe(struct platform_device *pdev)
 	if (ret)
 		return ret;
 
-	spin_lock_init(&pctrl->lock);
+	raw_spin_lock_init(&pctrl->lock);
 	mutex_init(&pctrl->mutex);
 	platform_set_drvdata(pdev, pctrl);
 
-- 
2.52.0

On 2026-02-05 12:39:30 [+0200], Cosmin Tanislav wrote:
> The rzt2h_gpio_get_direction() function is called from
> gpiod_get_direction(), which ends up being used within the __setup_irq()
> call stack when requesting an interrupt.
> 
> __setup_irq() holds a raw_spinlock_t with IRQs disabled, which creates
> an atomic context. spinlock_t cannot be used within atomic context
> when PREEMPT_RT is enabled, since it may become a sleeping lock.
> 
> An "[ BUG: Invalid wait context ]" splat is observed when running with
> CONFIG_PROVE_LOCKING enabled, describing exactly the aforementioned call
> stack.
> 
> __setup_irq() needs to hold a raw_spinlock_t with IRQs disabled to
> serialize access against a concurrent hard interrupt.
> 
> Switch to raw_spinlock_t to fix this.

I don't like the reasoning here because it looks like "lockdep
complained lets switch the locks and everything is fine now".

It is required to make the suggested change because the lock is used
in hardirq context and only while accessing the HW's register.

I just don't want that a lockdep splat becomes a green card for these
kind of changes without understanding the consequences.

> Fixes: 829dde3369a9 ("pinctrl: renesas: rzt2h: Add GPIO IRQ chip to handle interrupts")
> Signed-off-by: Cosmin Tanislav <cosmin-gabriel.tanislav.xa@renesas.com>

Sebastian