[v3,1/1] scsi: Synchronize request queue PM status only on successful resume

Commit Message

Stanley Chu Jan. 3, 2019, 2:08 p.m. UTC

From: Stanley Chu <[e-mail]>

The commit 356fd2663cff ("scsi: Set request queue runtime PM status
back to active on resume") fixed up the inconsistent RPM status between
request queue and device. However changing request queue RPM status
shall be done only on successful resume, otherwise status may be still
inconsistent as below,

Request queue: RPM_ACTIVE
Device: RPM_SUSPENDED

This ends up soft lockup because requests can be submitted to
underlying devices but those devices and their required resource
are not resumed.

For example,

After above inconsistent status happens, IO request can be submitted
to UFS device driver but required resource (like clock) is not resumed
yet thus lead to warning as below call stack,

WARN_ON(hba->clk_gating.state != CLKS_ON);
ufshcd_queuecommand
scsi_dispatch_cmd
scsi_request_fn
__blk_run_queue
cfq_insert_request
__elv_add_request
blk_flush_plug_list
blk_finish_plug
jbd2_journal_commit_transaction
kjournald2

We may see all behind IO requests hang because of no response from
storage host or device and then soft lockup happens in system. In the
end, system may crash in many ways.

Fixes: 356fd2663cff (scsi: Set request queue runtime PM status
back to active on resume)
Cc: [e-mail]
Signed-off-by: Stanley Chu <[e-mail]>
---
 drivers/scsi/scsi_pm.c | 26 +++++++++++++++-----------
 1 file changed, 15 insertions(+), 11 deletions(-)

Comments

Bart Van Assche Jan. 3, 2019, 7:03 p.m. UTC | #1

On Thu, 2019-01-03 at 22:08 +0800, [e-mail] wrote:
> Fixes: 356fd2663cff (scsi: Set request queue runtime PM status
> back to active on resume)

Please do not split "Fixes" lines. Otherwise this patch looks good to me.
Hence:

Reviewed-by: Bart Van Assche <[e-mail]>

Martin K. Petersen Jan. 4, 2019, 6:02 a.m. UTC | #2

Stanley,

> The commit 356fd2663cff ("scsi: Set request queue runtime PM status
> back to active on resume") fixed up the inconsistent RPM status
> between request queue and device. However changing request queue RPM
> status shall be done only on successful resume, otherwise status may
> be still inconsistent as below,

Applied to 4.21/scsi-fixes, thanks!

Patch

diff --git a/drivers/scsi/scsi_pm.c b/drivers/scsi/scsi_pm.c
index a2b4179bfdf7..7639df91b110 100644
--- a/drivers/scsi/scsi_pm.c
+++ b/drivers/scsi/scsi_pm.c
@@ -80,8 +80,22 @@  static int scsi_dev_type_resume(struct device *dev,
 
 	if (err == 0) {
 		pm_runtime_disable(dev);
-		pm_runtime_set_active(dev);
+		err = pm_runtime_set_active(dev);
 		pm_runtime_enable(dev);
+
+		/*
+		 * Forcibly set runtime PM status of request queue to "active"
+		 * to make sure we can again get requests from the queue
+		 * (see also blk_pm_peek_request()).
+		 *
+		 * The resume hook will correct runtime PM status of the disk.
+		 */
+		if (!err && scsi_is_sdev_device(dev)) {
+			struct scsi_device *sdev = to_scsi_device(dev);
+
+			if (sdev->request_queue->dev)
+				blk_set_runtime_active(sdev->request_queue);
+		}
 	}
 
 	return err;
@@ -140,16 +154,6 @@  static int scsi_bus_resume_common(struct device *dev,
 	else
 		fn = NULL;
 
-	/*
-	 * Forcibly set runtime PM status of request queue to "active" to
-	 * make sure we can again get requests from the queue (see also
-	 * blk_pm_peek_request()).
-	 *
-	 * The resume hook will correct runtime PM status of the disk.
-	 */
-	if (scsi_is_sdev_device(dev) && pm_runtime_suspended(dev))
-		blk_set_runtime_active(to_scsi_device(dev)->request_queue);
-
 	if (fn) {
 		async_schedule_domain(fn, dev, &scsi_sd_pm_domain);