Skip to contentSkip to conversation

mfd: mt6397-irq: Fix PM notifier use-after-free

PATCH In review Yuho Choi 2 participants 3 messages Jun 8 – Jun 18

Subsystems (2)
lore
Share & tools
lore

Yuho Choi

author
#0 lore raw

PATCH 1/1 mfd: mt6397-irq: Fix PM notifier use-after-free

mt6397_irq_init() registers chip->pm_nb with the global PM notifier chain. The notifier callback uses container_of() to recover struct mt6397_chip and then dereferences chip fields.

The chip structure is allocated with devm_kzalloc() in mt6397_probe(). If probe fails after mt6397_irq_init() succeeds, for example when devm_mfd_add_devices() fails, devres can release the chip while the PM notifier remains registered. The same lifetime mismatch exists when the driver is unbound.

Check the register_pm_notifier() return value and add a devm cleanup action to unregister the notifier before the devm-managed chip is freed. If adding the cleanup action fails, devm_add_action_or_reset() unregisters the notifier immediately; then remove the IRQ domain in the remaining error path.

Fixes 4e2e7cfec13a ("mfd: mt6397: Modify suspend/resume behavior")
---
 drivers/mfd/mt6397-irq.c | 22 ++++++++++++++++++++--
 1 file changed, 20 insertions(+), 2 deletions(-)
drivers/mfd/mt6397-irq.c +20 -2
index 5d2e5459f744..8947f7e732fa 100644
@@ -169,6 +169,13 @@ static int mt6397_irq_pm_notifier(struct notifier_block *notifier,
 	return NOTIFY_DONE;
 }
 
+static void mt6397_irq_pm_notifier_unregister(void *data)
+{
+	struct mt6397_chip *chip = data;
+
+	unregister_pm_notifier(&chip->pm_nb);
+}
+
 int mt6397_irq_init(struct mt6397_chip *chip)
 {
 	int ret;
@@ -233,6 +240,17 @@ int mt6397_irq_init(struct mt6397_chip *chip)
 		return ret;
 	}
 
-	register_pm_notifier(&chip->pm_nb);
-	return 0;
+	ret = register_pm_notifier(&chip->pm_nb);
+	if (ret) {
+		dev_err(chip->dev, "failed to register PM notifier: %d\n", ret);
+		irq_domain_remove(chip->irq_domain);
+		return ret;
+	}
+
+	ret = devm_add_action_or_reset(chip->dev,
+				       mt6397_irq_pm_notifier_unregister, chip);
+	if (ret)
+		irq_domain_remove(chip->irq_domain);
+
+	return ret;
 }
Signature · 1 line
2.43.0
maintainer
#1 lore raw

/* Sashiko Automation: Issues Found (3 Findings) */

Worth considering.

Quoted 27 lines · Yuho Choi

mt6397_irq_init() registers chip->pm_nb with the global PM notifier chain. The notifier callback uses container_of() to recover struct mt6397_chip and then dereferences chip fields.

The chip structure is allocated with devm_kzalloc() in mt6397_probe(). If probe fails after mt6397_irq_init() succeeds, for example when devm_mfd_add_devices() fails, devres can release the chip while the PM notifier remains registered. The same lifetime mismatch exists when the driver is unbound.

Check the register_pm_notifier() return value and add a devm cleanup action to unregister the notifier before the devm-managed chip is freed. If adding the cleanup action fails, devm_add_action_or_reset() unregisters the notifier immediately; then remove the IRQ domain in the remaining error path.

Fixes 4e2e7cfec13a ("mfd: mt6397: Modify suspend/resume behavior")
---
 drivers/mfd/mt6397-irq.c | 22 ++++++++++++++++++++--
 1 file changed, 20 insertions(+), 2 deletions(-)
drivers/mfd/mt6397-irq.c +0 -0
index 5d2e5459f744..8947f7e732fa 100644

drivers/mfd/mt6397-irq.c static int mt6397_irq_pm_notifier(struct notifier_block *notifier, · line 170

Quoted 2 lines
drivers/mfd/mt6397-irq.c +0 -0
@@ -169,6 +169,13 @@ static int mt6397_irq_pm_notifier(struct notifier_block *notifier,
 	return NOTIFY_DONE;
 }

[Severity: High]
This is a pre-existing issue, but does this PM notifier safely access the hardware registers and software state?

In mt6397_irq_pm_notifier(), the hardware registers and chip->irq_masks_cur are read and written without holding chip->irqlock. Since PM notifiers like PM_SUSPEND_PREPARE run while tasks are still active, if another driver requests or frees an IRQ concurrently, mt6397_irq_sync_unlock() will acquire irqlock and write dynamically updated irq_masks_cur to the hardware registers.

Can this data race allow the PM notifier to silently overwrite the dynamically updated hardware masks with stale data, or corrupt the suspend wake_mask?

Quoted 12 lines
drivers/mfd/mt6397-irq.c +7 -0
+static void mt6397_irq_pm_notifier_unregister(void *data)
+{
+	struct mt6397_chip *chip = data;
+
+	unregister_pm_notifier(&chip->pm_nb);
+}
+
 int mt6397_irq_init(struct mt6397_chip *chip)
 {
 	int ret;

drivers/mfd/mt6397-irq.c int mt6397_irq_init(struct mt6397_chip *chip) · line 246

Quoted 9 lines
drivers/mfd/mt6397-irq.c +4 -2
@@ -233,6 +240,17 @@ int mt6397_irq_init(struct mt6397_chip *chip)
 		return ret;
 	}
 
-	register_pm_notifier(&chip->pm_nb);
-	return 0;
+	ret = register_pm_notifier(&chip->pm_nb);
+	if (ret) {
+		dev_err(chip->dev, "failed to register PM notifier: %d\n", ret);
+		irq_domain_remove(chip->irq_domain);

[Severity: High]
Is it safe to explicitly remove the IRQ domain here while the devm-managed parent threaded IRQ remains active?

The parent IRQ was requested earlier via devm_request_threaded_irq(). Because devres cleanup does not run until the probe function completely fails and returns, the parent IRQ remains active and enabled during this window.

If an interrupt triggers during this error path, the active IRQ handler mt6397_irq_thread() will execute and call mt6397_irq_handle_reg(). This in turn calls irq_find_mapping(), which will dereference the freed irq_domain, resulting in a use-after-free.

[Severity: High]
This is a pre-existing issue, but does the IRQ domain properly unregister during driver unbind?

The irq_domain is created using irq_domain_create_linear() earlier in mt6397_irq_init(), which is not managed by devm. The MT6397 driver lacks a .remove callback and relies entirely on devres for cleanup.

When the device is unbound, devres frees the chip structure. However, the irq_domain is never removed and remains in the global IRQ domain list, leaking memory. Furthermore, its host_data points to the now-freed chip.

If the device is rebound, mt6397_irq_domain_map() will dereference the freed chip pointer, causing a use-after-free.

Quoted 13 lines
drivers/mfd/mt6397-irq.c +9 -0
+		return ret;
+	}
+
+	ret = devm_add_action_or_reset(chip->dev,
+				       mt6397_irq_pm_notifier_unregister, chip);
+	if (ret)
+		irq_domain_remove(chip->irq_domain);
+
+	return ret;
 }
Signature · 1 line
2.43.0
Signature · 1 line
Lee Jones
author
#2 lore raw
Quoted 4 lines · Lee Jones

[Severity: High]
Is it safe to explicitly remove the IRQ domain here while the devm-managed parent threaded IRQ remains active?

Quoted 8 lines

The parent IRQ was requested earlier via devm_request_threaded_irq(). Because devres cleanup does not run until the probe function completely fails and returns, the parent IRQ remains active and enabled during this window.

If an interrupt triggers during this error path, the active IRQ handler mt6397_irq_thread() will execute and call mt6397_irq_handle_reg(). This in turn calls irq_find_mapping(), which will dereference the freed irq_domain, resulting in a use-after-free.

Thanks for pointing this out. Findings 1 and 3 look like pre-existing issues. Finding 2, however, appears to be related to the error-path cleanup I added.

I will rework the patch so the notifier and irq_domain cleanup ordering is handled consistently.