[PATCH v4 1/1] scsi: pm: Balance pm_only counter of request queue during system resume

[PATCH v4 1/1] scsi: pm: Balance pm_only counter of request queue during system resume

[Can Guo]

During system resume, scsi_resume_device() decreases a request queue's
pm_only counter if the scsi device was quiesced before. But after that,
if the scsi device's RPM status is RPM_SUSPENDED, the pm_only counter is
still held (non-zero). Current scsi resume hook only sets the RPM status
of the scsi device and its request queue to RPM_ACTIVE, but leaves the
pm_only counter unchanged. This may make the request queue's pm_only
counter remain non-zero after resume hook returns, hence those who are
waiting on the mq_freeze_wq would never be woken up. Fix this by calling
blk_post_runtime_resume() if a sdev's RPM status was RPM_SUSPENDED.

(struct request_queue)0xFFFFFF815B69E938
	pm_only = (counter = 2),
	rpm_status = 0,
	dev = 0xFFFFFF815B0511A0,

((struct device)0xFFFFFF815B0511A0)).power
	is_suspended = FALSE,
	runtime_status = RPM_ACTIVE,

(struct scsi_device)0xffffff815b051000
	request_queue = 0xFFFFFF815B69E938,
	sdev_state = SDEV_RUNNING,
	quiesced_by = 0x0,

B::v.f_/task_0xFFFFFF810C246940
-000|__switch_to(prev = 0xFFFFFF810C246940, next = 0xFFFFFF80A49357C0)
-001|context_switch(inline)
-001|__schedule(?)
-002|schedule()
-003|blk_queue_enter(q = 0xFFFFFF815B69E938, flags = 0)
-004|generic_make_request(?)
-005|submit_bio(bio = 0xFFFFFF80A8195B80)

Signed-off-by: Can Guo <cang@codeaurora.org>
---

Change since v3:
- Instead of relying on pm_only counter, rely on the runtime status of a sdev. (Based on the discussion with Bart)

Change since v2:
- Rebased on 5.8-scsi-queue

Change since v1:
- Added more debugging context info

 drivers/scsi/scsi_pm.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/drivers/scsi/scsi_pm.c b/drivers/scsi/scsi_pm.c
index 3717eea..5f0ad8b 100644
--- a/drivers/scsi/scsi_pm.c
+++ b/drivers/scsi/scsi_pm.c
@@ -80,6 +80,10 @@ static int scsi_dev_type_resume(struct device *dev,
 	dev_dbg(dev, "scsi resume: %d\n", err);
 
 	if (err == 0) {
+		bool was_runtime_suspended;
+
+		was_runtime_suspended = pm_runtime_suspended(dev);
+
 		pm_runtime_disable(dev);
 		err = pm_runtime_set_active(dev);
 		pm_runtime_enable(dev);
@@ -93,8 +97,10 @@ static int scsi_dev_type_resume(struct device *dev,
 		 */
 		if (!err && scsi_is_sdev_device(dev)) {
 			struct scsi_device *sdev = to_scsi_device(dev);
-
-			blk_set_runtime_active(sdev->request_queue);
+			if (was_runtime_suspended)
+				blk_post_runtime_resume(sdev->request_queue, 0);
+			else
+				blk_set_runtime_active(sdev->request_queue);
 		}
 	}
 
-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project.

Re: [PATCH v4 1/1] scsi: pm: Balance pm_only counter of request queue during system resume

[Bart Van Assche]

On 2020-05-05 21:55, Can Guo wrote:
> During system resume, scsi_resume_device() decreases a request queue's
> pm_only counter if the scsi device was quiesced before. But after that,
> if the scsi device's RPM status is RPM_SUSPENDED, the pm_only counter is
> still held (non-zero). Current scsi resume hook only sets the RPM status
> of the scsi device and its request queue to RPM_ACTIVE, but leaves the
> pm_only counter unchanged. This may make the request queue's pm_only
> counter remain non-zero after resume hook returns, hence those who are
> waiting on the mq_freeze_wq would never be woken up. Fix this by calling
> blk_post_runtime_resume() if a sdev's RPM status was RPM_SUSPENDED.

Reviewed-by: Bart Van Assche <bvanassche@acm.org>

Thank you for having root-caused and fixed this!

Bart.

Re: [PATCH v4 1/1] scsi: pm: Balance pm_only counter of request queue during system resume

[Martin K. Petersen]

On Tue, 5 May 2020 21:55:35 -0700, Can Guo wrote:

> During system resume, scsi_resume_device() decreases a request queue's
> pm_only counter if the scsi device was quiesced before. But after that,
> if the scsi device's RPM status is RPM_SUSPENDED, the pm_only counter is
> still held (non-zero). Current scsi resume hook only sets the RPM status
> of the scsi device and its request queue to RPM_ACTIVE, but leaves the
> pm_only counter unchanged. This may make the request queue's pm_only
> counter remain non-zero after resume hook returns, hence those who are
> waiting on the mq_freeze_wq would never be woken up. Fix this by calling
> blk_post_runtime_resume() if a sdev's RPM status was RPM_SUSPENDED.
> 
> [...]

Applied to 5.7/scsi-queue, thanks!

[1/1] scsi: pm: Balance pm_only counter of request queue during system resume
      https://git.kernel.org/mkp/scsi/c/a3b923842626

-- 
Martin K. Petersen	Oracle Linux Engineering