Re: [PATCH v2] thermal/drivers/rcar_thermal: disable the work before masking the interrupt

From: Niklas Söderlund

Date: Thu Nov 19 2026 - 01:22:21 EST


Hi Yaohan,

Thanks for your work.

On 2026-11-19 10:29:03 +0800, Yaohan Li wrote:
> rcar_thermal_remove() masks the interrupt sources and then cancels the
> delayed work of each sensor. rcar_thermal_irq() drops common->lock before
> it walks the sensors and queues priv->work for those whose status changed,
> so a handler that read the status before the mask was written can still
> queue the work after cancel_delayed_work_sync() has returned. The work
> then runs after ->remove() has finished and devres has freed common, priv
> and the thermal zone.
>
> Waiting for the handler alone is not enough either: rcar_thermal_work()
> re-enables its interrupt source once it has updated the temperature, so a
> work that is still running when the removal starts undoes the mask that
> ->remove() has just written.
>
> Disable the work before the source is masked. That makes the
> queue_delayed_work() calls in rcar_thermal_irq() no-ops and guarantees
> that the re-enable in rcar_thermal_work() has already happened, so the
> mask that follows cannot be undone.
>
> The interrupt line is shared, so waiting for the handler once does not keep
> it from running again later; unregister it in ->remove() instead of only
> calling synchronize_irq(). devm_free_irq() waits for a handler that is
> running right now as well, and the slot is marked unused so that the probe
> failure path, which also ends here, cannot release it twice.
>
> Every sensor that the removal walks has been through INIT_DELAYED_WORK()
> before it is linked into common->head, so no uninitialised work is touched.
> The interrupt numbers are recorded in an array whose unused slots stay
> negative, instead of being counted separately.
>
> Fixes: 697ee786f15d ("thermal: rcar_thermal: Prevent doing work after unbind")
> Signed-off-by: Yaohan Li <shunsheng1218@xxxxxxxxx>
>
> ---
> Prepared against v7.4-rc1-1969-g666539eecdf4.
>
> Compile tested only: arm64, aarch64-linux-gnu-gcc 11.4, CONFIG_RCAR_THERMAL=y.
> No hardware was available, so the removal path was not exercised and no race
> injection was run.

Ohh, could you test this please. It's a rather large change and while it
looks correct it could behave differently then expected.

>
> AI assistance: prepared with Claude (Anthropic) as a coding assistant; the
> author reviewed the reasoning and takes responsibility for the content.
>
> Changes in v2:
> - Drop the shutdown flag and the separate IRQ count. Niklas asked for less
> state, so the removal now disables the delayed work before the source is
> masked and unregisters the shared IRQ instead of waiting for it once.
> - Keep the interrupt numbers in an array whose unused slots stay negative.
> ---
> drivers/thermal/renesas/rcar_thermal.c | 37 ++++++++++++++++++++++++--
> 1 file changed, 35 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/thermal/renesas/rcar_thermal.c b/drivers/thermal/renesas/rcar_thermal.c
> index af6d09a75bb6..3310ed4a8bda 100644
> --- a/drivers/thermal/renesas/rcar_thermal.c
> +++ b/drivers/thermal/renesas/rcar_thermal.c
> @@ -39,11 +39,15 @@
> /* THSSR */
> #define CTEMP 0x3f
>
> +/* .nirqs of every supported chip is at most this */

.nirqs is not really needed here. Just say something along the line of
max number of supported interrupts or some such.

> +#define RCAR_THERMAL_MAX_IRQ 2
> +
> struct rcar_thermal_common {
> void __iomem *base;
> struct device *dev;
> struct list_head head;
> spinlock_t lock;
> + int irqs[RCAR_THERMAL_MAX_IRQ];
> };
>
> struct rcar_thermal_chip {
> @@ -376,10 +380,34 @@ static void rcar_thermal_remove(struct platform_device *pdev)
> struct rcar_thermal_common *common = platform_get_drvdata(pdev);
> struct device *dev = &pdev->dev;
> struct rcar_thermal_priv *priv;
> + int i;

unsigned.

>
> - rcar_thermal_for_each_priv(priv, common) {
> + /*
> + * Stop the work from being queued again and wait for a running one to
> + * finish before the interrupt source is masked: rcar_thermal_work()
> + * re-enables its source when it is done, so it has to be gone before
> + * the source is disabled for good.

I think you can make this less verbose and more to the point.

> + */
> + rcar_thermal_for_each_priv(priv, common)
> + disable_delayed_work_sync(&priv->work);
> +
> + rcar_thermal_for_each_priv(priv, common)
> rcar_thermal_irq_disable(priv);
> - cancel_delayed_work_sync(&priv->work);
> +
> + /*
> + * The handler is shared, so waiting for it once is not enough: another
> + * user of the line can invoke it again after that, when the zones and

line?

> + * the sensors it walks are already gone. Unregister it here; that
> + * waits for a handler that is running right now as well.

I don't understand this.

> + */
> + for (i = 0; i < RCAR_THERMAL_MAX_IRQ; i++) {

Please declare the loop variable in the for () statement to reduce
scope.

> + if (common->irqs[i] < 0)
> + continue;
> + devm_free_irq(dev, common->irqs[i], common);
> + common->irqs[i] = -ENOENT;
> + }
> +
> + rcar_thermal_for_each_priv(priv, common) {
> if (priv->chip->use_of_thermal)
> thermal_remove_hwmon_sysfs(priv->zone);
> else
> @@ -413,6 +441,9 @@ static int rcar_thermal_probe(struct platform_device *pdev)
> spin_lock_init(&common->lock);
> common->dev = dev;
>
> + for (i = 0; i < RCAR_THERMAL_MAX_IRQ; i++)

Here you can't declare the loop variable as the function already
declares it at the function level. This is likely because this function
pre-dates when it was OK to declare it directly in the loop definition.

No need to update old tested code en-mass, but lets use the better way
in new code.

> + common->irqs[i] = -ENOENT;
> +
> pm_runtime_enable(dev);
> pm_runtime_get_sync(dev);
>
> @@ -449,6 +480,8 @@ static int rcar_thermal_probe(struct platform_device *pdev)
> if (ret)
> goto error_unregister;
>
> + common->irqs[i] = irq;
> +
> /* update ENR bits */
> if (chip->irq_per_ch)
> enr_bits |= 1 << i;
> --
> 2.34.1
>

--
Kind Regards,
Niklas Söderlund