mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
* [PATCH] thermal/drivers/rcar_thermal: wait for IRQs before cancelling work
@ 2026-11-18  7:23 Yaohan Li
  2026-11-18  9:00 ` Niklas Söderlund
  0 siblings, 1 reply; 2+ messages in thread
From: Yaohan Li @ 2026-11-18  7:23 UTC (permalink / raw)
  To: niklas.soderlund, rafael, daniel.lezcano
  Cc: rui.zhang, lukasz.luba, geert+renesas, magnus.damm,
	linux-renesas-soc, linux-pm, linux-kernel, edubezval

rcar_thermal_remove() masks the interrupt sources and cancels the
delayed work of each sensor.  rcar_thermal_irq() reads INTMSK and STR
under common->lock, drops the lock, and only then walks the sensors and
queues priv->work for every sensor whose status changed.

A handler that sampled the status before the mask was written can
therefore reach queue_delayed_work() after cancel_delayed_work_sync()
has returned.  The work stays pending past the end of ->remove(), devres
then frees common, priv and the thermal zone, and rcar_thermal_work()
runs 300 ms later on freed memory.

Waiting for the handler is not enough on its own: rcar_thermal_work()
calls rcar_thermal_irq_enable() once it has updated the temperature, so
a work that is still running when the removal starts re-enables its
source and undoes the mask that ->remove() has just written.

Record the shutdown in rcar_thermal_common under common->lock and make
_rcar_thermal_irq_ctrl() refuse to enable anything once that state is
visible.  The removal then publishes the shutdown, masks all sources,
waits for the handlers with synchronize_irq() and only then cancels the
work, so neither a source nor a work item can be brought back after it
returns.

Fixes: 697ee786f15d ("thermal: rcar_thermal: Prevent doing work after unbind")
Signed-off-by: Yaohan Li <shunsheng1218@gmail.com>
---
Prepared against v7.4-rc1-1969-g666539eecdf4.

AI assistance: the analysis, the change and its review notes were
prepared with Claude (Anthropic) as a coding assistant.  The author
reviewed the reasoning and takes responsibility for the content.

Testing: compile tested only (arm64, gcc 11.4).  No R-Car hardware was
available, so the unbind race was not run.
 drivers/thermal/renesas/rcar_thermal.c | 43 +++++++++++++++++++++++++-
 1 file changed, 42 insertions(+), 1 deletion(-)

diff --git a/drivers/thermal/renesas/rcar_thermal.c b/drivers/thermal/renesas/rcar_thermal.c
index af6d09a75bb6..1d6ee37968e2 100644
--- a/drivers/thermal/renesas/rcar_thermal.c
+++ b/drivers/thermal/renesas/rcar_thermal.c
@@ -39,11 +39,17 @@
 /* THSSR */
 #define CTEMP	0x3f
 
+/* .nirqs of every supported chip is at most this */
+#define RCAR_THERMAL_MAX_IRQ	2
+
 struct rcar_thermal_common {
 	void __iomem *base;
 	struct device *dev;
 	struct list_head head;
 	spinlock_t lock;
+	bool shutdown;
+	int irqs[RCAR_THERMAL_MAX_IRQ];
+	int nirqs;
 };
 
 struct rcar_thermal_chip {
@@ -301,6 +307,15 @@ static void _rcar_thermal_irq_ctrl(struct rcar_thermal_priv *priv, int enable)
 
 	spin_lock_irqsave(&common->lock, flags);
 
+	/*
+	 * rcar_thermal_work() re-enables its source when it finishes, so a
+	 * work that was still running when the device went away would undo
+	 * the mask that rcar_thermal_remove() writes.  Once the shutdown
+	 * state is visible nothing may enable the sources again.
+	 */
+	if (common->shutdown)
+		enable = 0;
+
 	rcar_thermal_common_bset(common, INTMSK, mask, enable ? 0 : mask);
 
 	spin_unlock_irqrestore(&common->lock, flags);
@@ -376,9 +391,33 @@ 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;
+	unsigned long flags;
+	int i;
 
-	rcar_thermal_for_each_priv(priv, common) {
+	/*
+	 * Publish the shutdown state before anything else: a work that is
+	 * still running re-enables its interrupt source on its way out, and
+	 * has to be prevented from undoing the mask written below.
+	 */
+	spin_lock_irqsave(&common->lock, flags);
+	common->shutdown = true;
+	spin_unlock_irqrestore(&common->lock, flags);
+
+	/*
+	 * Mask the sources, then wait for a handler that sampled the status
+	 * before the mask was written.  The handler drops common->lock
+	 * before it queues the work, so without the wait it could queue
+	 * priv->work after cancel_delayed_work_sync() has returned and leave
+	 * the work running on a priv that devres frees once ->remove() has
+	 * finished.
+	 */
+	rcar_thermal_for_each_priv(priv, common)
 		rcar_thermal_irq_disable(priv);
+
+	for (i = 0; i < common->nirqs; i++)
+		synchronize_irq(common->irqs[i]);
+
+	rcar_thermal_for_each_priv(priv, common) {
 		cancel_delayed_work_sync(&priv->work);
 		if (priv->chip->use_of_thermal)
 			thermal_remove_hwmon_sysfs(priv->zone);
@@ -449,6 +488,8 @@ static int rcar_thermal_probe(struct platform_device *pdev)
 		if (ret)
 			goto error_unregister;
 
+		common->irqs[common->nirqs++] = irq;
+
 		/* update ENR bits */
 		if (chip->irq_per_ch)
 			enr_bits |= 1 << i;
-- 
2.34.1


^ permalink raw reply	[flat|nested] 2+ messages in thread

* Re: [PATCH] thermal/drivers/rcar_thermal: wait for IRQs before cancelling work
  2026-11-18  7:23 [PATCH] thermal/drivers/rcar_thermal: wait for IRQs before cancelling work Yaohan Li
@ 2026-11-18  9:00 ` Niklas Söderlund
  0 siblings, 0 replies; 2+ messages in thread
From: Niklas Söderlund @ 2026-11-18  9:00 UTC (permalink / raw)
  To: Yaohan Li
  Cc: rafael, daniel.lezcano, rui.zhang, lukasz.luba, geert+renesas,
	magnus.damm, linux-renesas-soc, linux-pm, linux-kernel,
	edubezval

Hi Yaohan,

Thanks for your work.

On 2026-11-18 15:23:21 +0800, Yaohan Li wrote:
> rcar_thermal_remove() masks the interrupt sources and cancels the
> delayed work of each sensor.  rcar_thermal_irq() reads INTMSK and STR
> under common->lock, drops the lock, and only then walks the sensors and
> queues priv->work for every sensor whose status changed.
> 
> A handler that sampled the status before the mask was written can
> therefore reach queue_delayed_work() after cancel_delayed_work_sync()
> has returned.  The work stays pending past the end of ->remove(), devres
> then frees common, priv and the thermal zone, and rcar_thermal_work()
> runs 300 ms later on freed memory.
> 
> Waiting for the handler is not enough on its own: rcar_thermal_work()
> calls rcar_thermal_irq_enable() once it has updated the temperature, so
> a work that is still running when the removal starts re-enables its
> source and undoes the mask that ->remove() has just written.
> 
> Record the shutdown in rcar_thermal_common under common->lock and make
> _rcar_thermal_irq_ctrl() refuse to enable anything once that state is
> visible.  The removal then publishes the shutdown, masks all sources,
> waits for the handlers with synchronize_irq() and only then cancels the
> work, so neither a source nor a work item can be brought back after it
> returns.
> 
> Fixes: 697ee786f15d ("thermal: rcar_thermal: Prevent doing work after unbind")
> Signed-off-by: Yaohan Li <shunsheng1218@gmail.com>
> ---
> Prepared against v7.4-rc1-1969-g666539eecdf4.
> 
> AI assistance: the analysis, the change and its review notes were
> prepared with Claude (Anthropic) as a coding assistant.  The author
> reviewed the reasoning and takes responsibility for the content.
> 
> Testing: compile tested only (arm64, gcc 11.4).  No R-Car hardware was
> available, so the unbind race was not run.
>  drivers/thermal/renesas/rcar_thermal.c | 43 +++++++++++++++++++++++++-
>  1 file changed, 42 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/thermal/renesas/rcar_thermal.c b/drivers/thermal/renesas/rcar_thermal.c
> index af6d09a75bb6..1d6ee37968e2 100644
> --- a/drivers/thermal/renesas/rcar_thermal.c
> +++ b/drivers/thermal/renesas/rcar_thermal.c
> @@ -39,11 +39,17 @@
>  /* THSSR */
>  #define CTEMP	0x3f
>  
> +/* .nirqs of every supported chip is at most this */
> +#define RCAR_THERMAL_MAX_IRQ	2
> +
>  struct rcar_thermal_common {
>  	void __iomem *base;
>  	struct device *dev;
>  	struct list_head head;
>  	spinlock_t lock;
> +	bool shutdown;
> +	int irqs[RCAR_THERMAL_MAX_IRQ];
> +	int nirqs;
>  };
>  
>  struct rcar_thermal_chip {
> @@ -301,6 +307,15 @@ static void _rcar_thermal_irq_ctrl(struct rcar_thermal_priv *priv, int enable)
>  
>  	spin_lock_irqsave(&common->lock, flags);
>  
> +	/*
> +	 * rcar_thermal_work() re-enables its source when it finishes, so a
> +	 * work that was still running when the device went away would undo
> +	 * the mask that rcar_thermal_remove() writes.  Once the shutdown
> +	 * state is visible nothing may enable the sources again.
> +	 */
> +	if (common->shutdown)
> +		enable = 0;
> +
>  	rcar_thermal_common_bset(common, INTMSK, mask, enable ? 0 : mask);
>  
>  	spin_unlock_irqrestore(&common->lock, flags);
> @@ -376,9 +391,33 @@ 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;
> +	unsigned long flags;
> +	int i;
>  
> -	rcar_thermal_for_each_priv(priv, common) {
> +	/*
> +	 * Publish the shutdown state before anything else: a work that is
> +	 * still running re-enables its interrupt source on its way out, and
> +	 * has to be prevented from undoing the mask written below.
> +	 */
> +	spin_lock_irqsave(&common->lock, flags);
> +	common->shutdown = true;
> +	spin_unlock_irqrestore(&common->lock, flags);

Adding a shutdown flag seems like a lot of complexity for somehting like 
this. Would not being more clever here avoid that? Maybe something like 
this,

    /* No more work will be able to queued hence the irq can not rearm 
     * itself from the top half.
     */
    disable_work_sync()

    /* Disable any new interrupts in case they currently are enabled. */
    rcar_thermal_irq_disable()

    /* Wait for any bottom half interrupt to complete. */
    synchronize_irq()

    Now there can't be any races!

For this flag,

Nacked-by: Niklas Söderlund <niklas.soderlund+renesas@ragnatech.se>

> +
> +	/*
> +	 * Mask the sources, then wait for a handler that sampled the status
> +	 * before the mask was written.  The handler drops common->lock
> +	 * before it queues the work, so without the wait it could queue
> +	 * priv->work after cancel_delayed_work_sync() has returned and leave
> +	 * the work running on a priv that devres frees once ->remove() has
> +	 * finished.
> +	 */
> +	rcar_thermal_for_each_priv(priv, common)
>  		rcar_thermal_irq_disable(priv);
> +
> +	for (i = 0; i < common->nirqs; i++)
> +		synchronize_irq(common->irqs[i]);
> +
> +	rcar_thermal_for_each_priv(priv, common) {
>  		cancel_delayed_work_sync(&priv->work);
>  		if (priv->chip->use_of_thermal)
>  			thermal_remove_hwmon_sysfs(priv->zone);
> @@ -449,6 +488,8 @@ static int rcar_thermal_probe(struct platform_device *pdev)
>  		if (ret)
>  			goto error_unregister;
>  
> +		common->irqs[common->nirqs++] = irq;
> +
>  		/* update ENR bits */
>  		if (chip->irq_per_ch)
>  			enr_bits |= 1 << i;
> -- 
> 2.34.1
> 

-- 
Kind Regards,
Niklas Söderlund

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2026-11-18  9:00 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-18  7:23 [PATCH] thermal/drivers/rcar_thermal: wait for IRQs before cancelling work Yaohan Li
2026-11-18  9:00 ` Niklas Söderlund

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®