Skip to contentSkip to conversation

ufs: core: Fix runtime suspend error deadlock

PATCH Superseded [email protected] 2 participants 5 messages Sep 23, 2025 – Sep 25, 2025

Subsystems (3)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ufs: core: Fix runtime suspend error deadlock

From: Peter Wang <[email protected]>

Resolve the deadlock issue during runtime suspend when an error triggers the error handler. Prevent the deadlock by checking pm_op_in_progress and performing a quick recovery. This approach ensures that the error handler does not wait indefinitely for runtime PM to resume, allowing runtime suspend to proceed smoothly.

---
 drivers/ufs/core/ufshcd.c | 5 +++++
 1 file changed, 5 insertions(+)
drivers/ufs/core/ufshcd.c +5 -0
index 88a0e9289ca6..0735bd5df1cc 100644
@@ -6681,6 +6681,11 @@ static void ufshcd_err_handler(struct work_struct *work)
 	}
 	spin_unlock_irqrestore(hba->host->host_lock, flags);
 
+	if (hba->pm_op_in_progress) {
+		ufshcd_link_recovery(hba);
+		return;
+	}
+
 	ufshcd_err_handling_prepare(hba);
 
 	spin_lock_irqsave(hba->host->host_lock, flags);
Signature · 1 line
2.45.2
#1 lore raw
Quoted 22 lines · On 9/23/25 1:20 AM, [email protected] wrote:

From: Peter Wang <[email protected]>

Resolve the deadlock issue during runtime suspend when an error triggers the error handler. Prevent the deadlock by checking pm_op_in_progress and performing a quick recovery. This approach ensures that the error handler does not wait indefinitely for runtime PM to resume, allowing runtime suspend to proceed smoothly.

---
  drivers/ufs/core/ufshcd.c | 5 +++++
  1 file changed, 5 insertions(+)
drivers/ufs/core/ufshcd.c +0 -0
index 88a0e9289ca6..0735bd5df1cc 100644
@@ -6681,6 +6681,11 @@ static void ufshcd_err_handler(struct work_struct *work)
  	}

drivers/ufs/core/ufshcd.c static void ufshcd_err_handler(struct work_struct *work) · line 6691

Quoted 10 lines
drivers/ufs/core/ufshcd.c +5 -0
  	spin_unlock_irqrestore(hba->host->host_lock, flags);
 
+	if (hba->pm_op_in_progress) {
+		ufshcd_link_recovery(hba);
+		return;
+	}
+
  	ufshcd_err_handling_prepare(hba);
 
  	spin_lock_irqsave(hba->host->host_lock, flags);

Does this patch introduce a race condition? Can a power management operation start after hba->pm_op_in_progress has been tested and before ufshcd_err_handling_prepare() is called? Should the added code perhaps be surrounded with ufshcd_rpm_get_noresume() and ufshcd_rpm_put()?

Thanks,

Bart.

#2 lore raw
Quoted 9 lines · Bart Van Assche

Does this patch introduce a race condition? Can a power management operation start after hba->pm_op_in_progress has been tested and before
ufshcd_err_handling_prepare() is called? Should the added code perhaps
be surrounded with ufshcd_rpm_get_noresume() and ufshcd_rpm_put()?

Quoted 3 lines

Thanks,

Bart.

Hi Bart,

Are you saying that after the runtime suspend timer expires, the error handler might be triggered just before runtime suspend? I don't think this can happen, since there shouldn't be anything triggering the error handler when the bus has been idle for a long time.

Thanks
Peter

#3 lore raw
Quoted 5 lines · Peter Wang (王信友)

Are you saying that after the runtime suspend timer expires, the error handler might be triggered just before runtime suspend? I don't think this can happen, since there shouldn't be anything triggering the error handler when the bus has been idle for a long time.

Hi Peter,

The UFS error handler is not only triggered by bus errors. It can also be activated via debugfs. Do you agree that activation via debugfs can happen concurrently with runtime suspend?

Thanks,

Bart.

#4 lore raw
Quoted 9 lines · Bart Van Assche

Hi Peter,

The UFS error handler is not only triggered by bus errors. It can also
be activated via debugfs. Do you agree that activation via debugfs can
happen concurrently with runtime suspend?

Quoted 3 lines

Thanks,

Bart.

Hi Bart,

Do you mean that debugfs could schedule EH work just before suspend? If so, there is indeed a chance that this patch might not work as intended. Even if we surround it with ufshcd_rpm_get_noresume() and ufshcd_rpm_put(), there is still a possibility that runtime PM is ongoing and the suspend callback hasn't been called yet, so the result would be the same, just with a lower chance of occurrence.

Because this patch mainly targets errors that occur during the suspend process. If the error happens before suspend, it’s outside the scope of what this patch is intended to fix. But I will add ufshcd_rpm_get_noresume() and ufshcd_rpm_put() next version.

Thanks.
Peter