Skip to contentSkip to conversation

media: qcom: iris: guard IRQ handler with runtime PM check

PATCH In review Hungyu Lin 3 participants 5 messages Jun 8 – Jun 10

Subsystems (4)
lore
Share & tools
lore

Hungyu Lin

author
#0 lore raw

PATCH v4 1/1 media: qcom: iris: guard IRQ handler with runtime PM check

Guard hardware register access in the threaded IRQ handler with pm_runtime_get_if_active().

A possible ordering exists where the top-half IRQ handler returns IRQ_WAKE_THREAD, runtime PM suspend powers down the VPU, and the threaded IRQ handler subsequently runs and accesses hardware registers through iris_vpu_clear_interrupt().

Avoid touching registers when the device is no longer active by skipping interrupt processing when runtime PM indicates that the device is suspended.

---
v4:
- Clarify the possible ordering between IRQ_WAKE_THREAD, runtime PM suspend, and the threaded IRQ handler.
- Describe the race condition motivating the runtime PM check.

v3:
- Remove the early enable_irq() from the PM-inactive early-return path.
- IRQ re-enablement is already handled by iris_vpu_power_on() after power-on.

v2:
- Use pm_runtime_get_if_active() instead of pm_runtime_get_if_in_use().
- Handle negative runtime PM return values correctly.
- Return IRQ_NONE when interrupt processing is skipped.

 drivers/media/platform/qcom/iris/iris_hfi_common.c | 9 +++++++++
 1 file changed, 9 insertions(+)
drivers/media/platform/qcom/iris/iris_hfi_common.c +9 -0
index 621c66593d88..ab2ec1e75309 100644
@@ -100,10 +100,17 @@ irqreturn_t iris_hfi_isr(int irq, void *data)
 irqreturn_t iris_hfi_isr_handler(int irq, void *data)
 {
 	struct iris_core *core = data;
+	int ret;
 
 	if (!core)
 		return IRQ_NONE;
 
+	if (IS_ENABLED(CONFIG_PM)) {
+		ret = pm_runtime_get_if_active(core->dev);
+		if (ret <= 0)
+			return IRQ_NONE;
+	}
+
 	mutex_lock(&core->lock);
 	pm_runtime_mark_last_busy(core->dev);
 	iris_vpu_clear_interrupt(core);
@@ -111,6 +118,8 @@ irqreturn_t iris_hfi_isr_handler(int irq, void *data)
 
 	core->hfi_response_ops->hfi_response_handler(core);
 
+	pm_runtime_put_autosuspend(core->dev);
+
 	if (!iris_vpu_watchdog(core, core->intr_status))
 		enable_irq(irq);
 
Signature · 1 line
2.34.1
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider: - [High] The threaded IRQ handler returns early without calling `enable_irq()` when the device is runtime…

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] The threaded IRQ handler returns early without calling enable_irq() when the device is runtime suspended or runtime PM is disabled, permanently leaking the IRQ disable depth and disabling the interrupt. --

commit 300804196935f196b366554c39d9f042b16bde60
Author: Hungyu Lin <[email protected]>

media: qcom: iris: guard IRQ handler with runtime PM check

Guard hardware register access in the threaded IRQ handler with pm_runtime_get_if_active() to avoid touching registers when suspended.

Quoted 7 lines
drivers/media/platform/qcom/iris/iris_hfi_common.c +0 -0
index 621c66593d88d..ab2ec1e753097 100644
@@ -100,10 +100,17 @@ irqreturn_t iris_hfi_isr(int irq, void *data)
 irqreturn_t iris_hfi_isr_handler(int irq, void *data)
 {

drivers/media/platform/qcom/iris/iris_hfi_common.c irqreturn_t iris_hfi_isr(int irq, void *data) · line 111

Quoted 10 lines
drivers/media/platform/qcom/iris/iris_hfi_common.c +5 -0
 	struct iris_core *core = data;
+	int ret;
 
 	if (!core)
 		return IRQ_NONE;
 
+	if (IS_ENABLED(CONFIG_PM)) {
+		ret = pm_runtime_get_if_active(core->dev);
+		if (ret <= 0)
+			return IRQ_NONE;

[Severity: High]
Does this early return leave the interrupt permanently disabled?

Looking at the top-half handler iris_hfi_isr() in this same file, it unconditionally calls disable_irq_nosync(irq) before returning IRQ_WAKE_THREAD.

If a suspend sequence begins between the execution of the top-half and the threaded handler, pm_runtime_get_if_active() will return 0. The threaded handler will then exit early here without calling enable_irq(irq).

This leaves the IRQ disable depth permanently incremented. During resume, even if the power-on sequence calls enable_irq(), the depth never returns to 0, leaving the hardware interrupt permanently masked and potentially causing all future device operations to timeout.

Should this early return path also call enable_irq(irq) to balance the disable depth?

Quoted 5 lines
drivers/media/platform/qcom/iris/iris_hfi_common.c +2 -0
+	}
+
 	mutex_lock(&core->lock);
 	pm_runtime_mark_last_busy(core->dev);
 	iris_vpu_clear_interrupt(core);
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Hungyu Lin
Quoted 14 lines · Hungyu Lin

Guard hardware register access in the threaded IRQ handler with pm_runtime_get_if_active().

A possible ordering exists where the top-half IRQ handler returns IRQ_WAKE_THREAD, runtime PM suspend powers down the VPU, and the threaded IRQ handler subsequently runs and accesses hardware registers through iris_vpu_clear_interrupt().

Avoid touching registers when the device is no longer active by skipping interrupt processing when runtime PM indicates that the device is suspended.

---

[...]

drivers/media/platform/qcom/iris/iris_hfi_common.c irqreturn_t iris_hfi_isr(int irq, void *data) · line 108

Quoted 12 lines
drivers/media/platform/qcom/iris/iris_hfi_common.c +2 -0
@@ -100,10 +100,17 @@ irqreturn_t iris_hfi_isr(int irq, void *data)
 irqreturn_t iris_hfi_isr_handler(int irq, void *data)
 {
 	struct iris_core *core = data;
+	int ret;
 
 	if (!core)
 		return IRQ_NONE;
 
+	if (IS_ENABLED(CONFIG_PM)) {

I'm not sure this driver is much useful without CONFIG_PM

drivers/media/platform/qcom/iris/iris_hfi_common.c irqreturn_t iris_hfi_isr(int irq, void *data) · line 112

Quoted 4 lines
drivers/media/platform/qcom/iris/iris_hfi_common.c +4 -0
+		ret = pm_runtime_get_if_active(core->dev);
+		if (ret <= 0)
+			return IRQ_NONE;
+	}

Have you actually hit this issue, or is it purely theoretical? We shouldn't be receiving interrupts at the tail end of suspend callbacks (and there's a disable_irq_nosync() right after the HW is disabled)

Konrad

author
#3 lore raw
Quoted 3 lines

Have you actually hit this issue, or is it purely theoretical? We shouldn't be receiving interrupts at the tail end of suspend callbacks (and there's a disable_irq_nosync() right after the HW is disabled)

I have not observed this on hardware.

My concern comes from disable_irq_nosync().

Its documentation states that it does not ensure existing instances of the IRQ handler have completed before returning.

Given that, could a queued threaded handler still
execute after the VPU has been powered down?

Thanks,
Hungyu

#4 lore raw
Quoted 11 lines · Hungyu Lin
Quoted 3 lines

Have you actually hit this issue, or is it purely theoretical? We shouldn't be receiving interrupts at the tail end of suspend callbacks (and there's a disable_irq_nosync() right after the HW is disabled)

I have not observed this on hardware.

My concern comes from disable_irq_nosync().

Its documentation states that it does not ensure existing instances of the IRQ handler have completed before returning.

Then let's just replace disable_irq_nosync with disable_irq, I think this will solve this issue

Konrad