Skip to contentSkip to conversation

thermal/drivers/ti-soc-thermal: Cancel pending alert work on remove

PATCH Applied Weigang He 4 participants 4 messages Sep 26 – Oct 14

Subsystems (2)
lore
Share & tools
lore

Weigang He

author
#0 lore raw

PATCH 1/1 thermal/drivers/ti-soc-thermal: Cancel pending alert work on remove

On a threshold crossing, the talert IRQ handler calls ti_thermal_report_sensor_temperature(), which queues the work embedded in the sensor's ti_thermal_data on the system workqueue. Nothing ever cancels or flushes that work.

ti_bandgap_remove() frees the talert IRQ last. free_irq() waits for a running handler, but not for the work the handler has queued. When remove returns, devres unregisters and frees the thermal zones and then frees the devm-allocated ti_thermal_data, so a work item that is still pending runs ti_thermal_work() on freed memory:

	ti_thermal_work()
	  data = container_of(work, struct ti_thermal_data, thermal_wq);
	  thermal_zone_device_update(data->ti_thermal, ...);

Free the talert IRQ before removing the sensors, so that no new work can be queued, and cancel the work in ti_thermal_remove_sensor().

This needs an OMAP4460/4470 or OMAP5 SoC (DRA7 has TALERT but no ->report_temperature, so it never queues the work), a threshold crossing just before the driver is unbound or unloaded, and the work still pending when devres frees the data.

Found by static analysis tool CodeQL.

Fixes 445eaf871bf9 ("staging: omap-thermal: common code to expose driver to thermal framework")
Assisted-by LLM codeql

---

Notes:
Compile-tested only (ARCH=arm64 and ARCH=x86_64 allmodconfig, and ARCH=arm multi_v7_defconfig, W=1). Not tested on hardware: I have no OMAP4460/4470 or OMAP5 board, and there is no reproducer. Unbinding or unloading the driver needs root, so this is sent as a regular bug; no stable Cc, but please add one if you think it is warranted.

Found while reviewing a CodeQL report for this driver.

 drivers/thermal/ti-soc-thermal/ti-bandgap.c        | 7 ++++---
 drivers/thermal/ti-soc-thermal/ti-thermal-common.c | 4 ++++
 2 files changed, 8 insertions(+), 3 deletions(-)
drivers/thermal/ti-soc-thermal/ti-bandgap.c +4 -3
index ba43399d0b384..507fe67eaffae 100644
@@ -1077,6 +1077,10 @@ void ti_bandgap_remove(struct platform_device *pdev)
 	if (!soc_device_match(soc_no_cpu_notifier))
 		cpu_pm_unregister_notifier(&bgp->nb);
 
+	/* Stop the alerts first: they queue work on the sensors' data */
+	if (TI_BANDGAP_HAS(bgp, TALERT))
+		free_irq(bgp->irq, bgp);
+
 	/* Remove sensor interfaces */
 	for (i = 0; i < bgp->conf->sensor_count; i++) {
 		if (bgp->conf->sensors[i].unregister_cooling)
@@ -1093,9 +1097,6 @@ void ti_bandgap_remove(struct platform_device *pdev)
 	clk_put(bgp->fclock);
 	clk_put(bgp->div_clk);
 
-	if (TI_BANDGAP_HAS(bgp, TALERT))
-		free_irq(bgp->irq, bgp);
-
 	if (TI_BANDGAP_HAS(bgp, TSHUT))
 		free_irq(gpiod_to_irq(bgp->tshut_gpiod), NULL);
 }
drivers/thermal/ti-soc-thermal/ti-thermal-common.c +4 -0
index 6e1bbdee53637..3073d4d9e4f36 100644
@@ -195,6 +195,10 @@ int ti_thermal_remove_sensor(struct ti_bandgap *bgp, int id)
 
 	data = ti_bandgap_get_sensor_data(bgp, id);
 
+	/* Work queued by the talert IRQ must not outlive the data */
+	if (!IS_ERR_OR_NULL(data))
+		cancel_work_sync(&data->thermal_wq);
+
 	if (!IS_ERR_OR_NULL(data) && data->ti_thermal) {
 		if (data->our_zone)
 			thermal_zone_device_unregister(data->ti_thermal);

base-commit: 165768bb70265b5c38cf0b73fafd75be235f8b14

Signature · 1 line
2.43.0
#1 lore raw
Quoted 23 lines · Weigang He

On a threshold crossing, the talert IRQ handler calls ti_thermal_report_sensor_temperature(), which queues the work embedded in the sensor's ti_thermal_data on the system workqueue. Nothing ever cancels or flushes that work.

ti_bandgap_remove() frees the talert IRQ last. free_irq() waits for a running handler, but not for the work the handler has queued. When remove returns, devres unregisters and frees the thermal zones and then frees the devm-allocated ti_thermal_data, so a work item that is still pending runs ti_thermal_work() on freed memory:

	ti_thermal_work()
	  data = container_of(work, struct ti_thermal_data, thermal_wq);
	  thermal_zone_device_update(data->ti_thermal, ...);

Free the talert IRQ before removing the sensors, so that no new work can be queued, and cancel the work in ti_thermal_remove_sensor().

This needs an OMAP4460/4470 or OMAP5 SoC (DRA7 has TALERT but no ->report_temperature, so it never queues the work), a threshold crossing just before the driver is unbound or unloaded, and the work still pending when devres frees the data.

Quoted 6 lines

Found by static analysis tool CodeQL.

Fixes 445eaf871bf9 ("staging: omap-thermal: common code to expose driver to thermal framework")
Assisted-by LLM codeql

---

Sounds correct to me, so if nobody is against this change I'll apply it

Thanks

maintainer
#2 lore raw

-----Original Message-----
From: Daniel Lezcano <[email protected]>
Sent: Tuesday, September 29, 2026 12:45 PM
To: Weigang He <[email protected]>; Daniel Lezcano <[email protected]>; Rafael J . Wysocki <[email protected]>; Eduardo Valentin <[email protected]>; J, KEERTHY <[email protected]>
Cc: Zhang Rui <[email protected]>; Lukasz Luba <[email protected]>; [email protected]; [email protected]; [email protected]
Subject: Re: [PATCH] thermal/drivers/ti-soc-thermal: Cancel pending alert work on remove

Quoted 24 lines · Weigang He

On a threshold crossing, the talert IRQ handler calls ti_thermal_report_sensor_temperature(), which queues the work embedded in the sensor's ti_thermal_data on the system workqueue. Nothing ever cancels or flushes that work.

ti_bandgap_remove() frees the talert IRQ last. free_irq() waits for a running handler, but not for the work the handler has queued. When remove returns, devres unregisters and frees the thermal zones and then frees the devm-allocated ti_thermal_data, so a work item that is still pending runs ti_thermal_work() on freed memory:

	ti_thermal_work()
	  data = container_of(work, struct ti_thermal_data, thermal_wq);
	  thermal_zone_device_update(data->ti_thermal, ...);

Free the talert IRQ before removing the sensors, so that no new work can be queued, and cancel the work in ti_thermal_remove_sensor().

This needs an OMAP4460/4470 or OMAP5 SoC (DRA7 has TALERT but no ->report_temperature, so it never queues the work), a threshold ->crossing
just before the driver is unbound or unloaded, and the work still pending when devres frees the data.

Quoted 7 lines

Found by static analysis tool CodeQL.

Fixes 445eaf871bf9 ("staging: omap-thermal: common code to expose

driver to thermal framework")

Assisted-by LLM codeql

---

Sounds correct to me, so if nobody is against this change I'll apply it

Quoted 1 line

FWIW: Acked-by: Keerthy <[email protected]>

Thanks

merge bot Applied
#3 lore raw
In reply to Weigang He
The following commit has been merged into the thermal/linux-next branch of thermal:

The following commit has been merged into the thermal/linux-next branch of thermal:

Commit-ID:     aaa6c8dec49339710f9fd2a2af5646cf4111dbd4
Gitweb:        https://git.kernel.org/pub/scm/linux/kernel/git/thermal/linux.git//aaa6c8dec49339710f9fd2a2af5646cf4111dbd4
Author:        Weigang He <[email protected]>
AuthorDate:    Sat, 26 Sep 2026 22:49:31 +10:00
Committer:     Daniel Lezcano <[email protected]>
CommitterDate: Tue, 13 Oct 2026 16:26:37 +02:00

thermal/drivers/ti-soc-thermal: Cancel pending alert work on remove

On a threshold crossing, the talert IRQ handler calls ti_thermal_report_sensor_temperature(), which queues the work embedded in the sensor's ti_thermal_data on the system workqueue. Nothing ever cancels or flushes that work.

ti_bandgap_remove() frees the talert IRQ last. free_irq() waits for a running handler, but not for the work the handler has queued. When remove returns, devres unregisters and frees the thermal zones and then frees the devm-allocated ti_thermal_data, so a work item that is still pending runs ti_thermal_work() on freed memory:

	ti_thermal_work()
	  data = container_of(work, struct ti_thermal_data, thermal_wq);
	  thermal_zone_device_update(data->ti_thermal, ...);

Free the talert IRQ before removing the sensors, so that no new work can be queued, and cancel the work in ti_thermal_remove_sensor().

This needs an OMAP4460/4470 or OMAP5 SoC (DRA7 has TALERT but no ->report_temperature, so it never queues the work), a threshold crossing just before the driver is unbound or unloaded, and the work still pending when devres frees the data.

Found by static analysis tool CodeQL.

Fixes 445eaf871bf9 ("staging: omap-thermal: common code to expose driver to thermal framework")
Assisted-by LLM codeql
---
 drivers/thermal/ti-soc-thermal/ti-bandgap.c        | 7 ++++---
 drivers/thermal/ti-soc-thermal/ti-thermal-common.c | 4 ++++
 2 files changed, 8 insertions(+), 3 deletions(-)
drivers/thermal/ti-soc-thermal/ti-bandgap.c +4 -3
index ba43399..507fe67 100644
@@ -1077,6 +1077,10 @@ void ti_bandgap_remove(struct platform_device *pdev)
 	if (!soc_device_match(soc_no_cpu_notifier))
 		cpu_pm_unregister_notifier(&bgp->nb);
 
+	/* Stop the alerts first: they queue work on the sensors' data */
+	if (TI_BANDGAP_HAS(bgp, TALERT))
+		free_irq(bgp->irq, bgp);
+
 	/* Remove sensor interfaces */
 	for (i = 0; i < bgp->conf->sensor_count; i++) {
 		if (bgp->conf->sensors[i].unregister_cooling)
@@ -1093,9 +1097,6 @@ void ti_bandgap_remove(struct platform_device *pdev)
 	clk_put(bgp->fclock);
 	clk_put(bgp->div_clk);
 
-	if (TI_BANDGAP_HAS(bgp, TALERT))
-		free_irq(bgp->irq, bgp);
-
 	if (TI_BANDGAP_HAS(bgp, TSHUT))
 		free_irq(gpiod_to_irq(bgp->tshut_gpiod), NULL);
 }
drivers/thermal/ti-soc-thermal/ti-thermal-common.c +4 -0
index 0cf0826..4bf3b2e 100644
@@ -195,6 +195,10 @@ int ti_thermal_remove_sensor(struct ti_bandgap *bgp, int id)
 
 	data = ti_bandgap_get_sensor_data(bgp, id);
 
+	/* Work queued by the talert IRQ must not outlive the data */
+	if (!IS_ERR_OR_NULL(data))
+		cancel_work_sync(&data->thermal_wq);
+
 	if (!IS_ERR_OR_NULL(data) && data->ti_thermal) {
 		if (data->our_zone)
 			thermal_zone_device_unregister(data->ti_thermal);