Skip to contentSkip to conversation

Rework runtime suspend and SPI domain validation

PATCH Applied Bart Van Assche 6 participants 30 messages Nov 30, 2020 – Dec 8, 2020

Subsystems (3)

v4 of 5 Applied Latest is v5 →

lore
Share & tools
lore
#0 lore raw

v4 cover Rework runtime suspend and SPI domain validation

Hi Martin,

The SCSI runtime suspend and SPI domain validation mechanisms both use scsi_device_quiesce(). scsi_device_quiesce() restricts blk_queue_enter() to BLK_MQ_REQ_PREEMPT requests. There is a conflict between the requirements of runtime suspend and SCSI domain validation: no requests must be sent to runtime suspended devices that are in the state RPM_SUSPENDED while BLK_MQ_REQ_PREEMPT requests must be processed during SCSI domain validation. This conflict is resolved by reworking the SCSI domain validation implementation.

Hybernation, runtime suspend and SCSI domain validation have been retested.

Please consider this patch series for kernel v5.11.

Thanks,

Bart.

Changes between v3 and v4:
- Instead of creating a second request queue for SPI DV, set RQF_PM.

Changes between v2 and v3:
- Inlined scsi_mq_alloc_queue() into scsi_alloc_sdev() as requested by Christoph.

Changes between v1 and v2:
- Rebased this patch series on top of kernel v5.10-rc1.

Alan Stern (1):
block: Do not accept any requests while suspended

Bart Van Assche (8):
block: Fix a race in the runtime power management code
block: Introduce BLK_MQ_REQ_PM
ide: Do not set the RQF_PREEMPT flag for sense requests
ide: Mark power management requests with RQF_PM instead of RQF_PREEMPT
scsi: Do not wait for a request in scsi_eh_lock_door()
scsi_transport_spi: Set RQF_PM for domain validation commands
scsi: Only process PM requests if rpm_status != RPM_ACTIVE
block: Remove RQF_PREEMPT and BLK_MQ_REQ_PREEMPT

 block/blk-core.c                  | 12 ++++++------
 block/blk-mq-debugfs.c            |  1 -
 block/blk-mq.c                    |  4 ++--
 block/blk-pm.c                    | 15 +++++++++------
 block/blk-pm.h                    | 14 +++++++++-----
 drivers/ide/ide-atapi.c           |  1 -
 drivers/ide/ide-io.c              |  7 +------
 drivers/ide/ide-pm.c              |  2 +-
 drivers/scsi/scsi_error.c         |  7 ++++++-
 drivers/scsi/scsi_lib.c           | 27 ++++++++++++++-------------
 drivers/scsi/scsi_transport_spi.c | 27 +++++++++++++++++++--------
 include/linux/blk-mq.h            |  4 ++--
 include/linux/blkdev.h            |  6 +-----
 13 files changed, 70 insertions(+), 57 deletions(-)
#1 lore raw

PATCH v4 1/9 block: Fix a race in the runtime power management code

With the current implementation the following race can happen:
* blk_pre_runtime_suspend() calls blk_freeze_queue_start() and blk_mq_unfreeze_queue().
* blk_queue_enter() calls blk_queue_pm_only() and that function returns true.
* blk_queue_enter() calls blk_pm_request_resume() and that function does not call pm_request_resume() because the queue runtime status is RPM_ACTIVE.
* blk_pre_runtime_suspend() changes the queue status into RPM_SUSPENDING.

Fix this race by changing the queue runtime status into RPM_SUSPENDING before switching q_usage_counter to atomic mode.

Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
Cc stable <[email protected]>
Fixes 986d413b7c15 ("blk-mq: Enable support for runtime power management")
---
 block/blk-pm.c | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)
block/blk-pm.c +9 -6
index b85234d758f7..17bd020268d4 100644
@@ -67,6 +67,10 @@ int blk_pre_runtime_suspend(struct request_queue *q)
 
 	WARN_ON_ONCE(q->rpm_status != RPM_ACTIVE);
 
+	spin_lock_irq(&q->queue_lock);
+	q->rpm_status = RPM_SUSPENDING;
+	spin_unlock_irq(&q->queue_lock);
+
 	/*
 	 * Increase the pm_only counter before checking whether any
 	 * non-PM blk_queue_enter() calls are in progress to avoid that any
@@ -89,15 +93,14 @@ int blk_pre_runtime_suspend(struct request_queue *q)
 	/* Switch q_usage_counter back to per-cpu mode. */
 	blk_mq_unfreeze_queue(q);
 
-	spin_lock_irq(&q->queue_lock);
-	if (ret < 0)
+	if (ret < 0) {
+		spin_lock_irq(&q->queue_lock);
+		q->rpm_status = RPM_ACTIVE;
 		pm_runtime_mark_last_busy(q->dev);
-	else
-		q->rpm_status = RPM_SUSPENDING;
-	spin_unlock_irq(&q->queue_lock);
+		spin_unlock_irq(&q->queue_lock);
 
-	if (ret)
 		blk_clear_pm_only(q);
+	}
 
 	return ret;
 }
#2 lore raw

PATCH v4 2/9 block: Introduce BLK_MQ_REQ_PM

In reply to Bart Van Assche

Introduce the BLK_MQ_REQ_PM flag. This flag makes the request allocation functions set RQF_PM. This is the first step towards removing BLK_MQ_REQ_PREEMPT.

Cc Alan Stern <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
Cc Can Guo <[email protected]>
---
 block/blk-core.c       | 7 ++++---
 block/blk-mq.c         | 2 ++
 include/linux/blk-mq.h | 2 ++
 3 files changed, 8 insertions(+), 3 deletions(-)
block/blk-core.c +4 -3
index 2db8bda43b6e..10696f9fb6ac 100644
@@ -424,11 +424,11 @@ EXPORT_SYMBOL(blk_cleanup_queue);
 /**
  * blk_queue_enter() - try to increase q->q_usage_counter
  * @q: request queue pointer
- * @flags: BLK_MQ_REQ_NOWAIT and/or BLK_MQ_REQ_PREEMPT
+ * @flags: BLK_MQ_REQ_NOWAIT, BLK_MQ_REQ_PM and/or BLK_MQ_REQ_PREEMPT
  */
 int blk_queue_enter(struct request_queue *q, blk_mq_req_flags_t flags)
 {
-	const bool pm = flags & BLK_MQ_REQ_PREEMPT;
+	const bool pm = flags & (BLK_MQ_REQ_PM | BLK_MQ_REQ_PREEMPT);
 
 	while (true) {
 		bool success = false;
@@ -630,7 +630,8 @@ struct request *blk_get_request(struct request_queue *q, unsigned int op,
 	struct request *req;
 
 	WARN_ON_ONCE(op & REQ_NOWAIT);
-	WARN_ON_ONCE(flags & ~(BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_PREEMPT));
+	WARN_ON_ONCE(flags & ~(BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_PM |
+			       BLK_MQ_REQ_PREEMPT));
 
 	req = blk_mq_alloc_request(q, op, flags);
 	if (!IS_ERR(req) && q->mq_ops->initialize_rq_fn)
block/blk-mq.c +2 -0
index 1b25ec2fe9be..b5880a1fb38d 100644
@@ -292,6 +292,8 @@ static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 	rq->mq_hctx = data->hctx;
 	rq->rq_flags = 0;
 	rq->cmd_flags = data->cmd_flags;
+	if (data->flags & BLK_MQ_REQ_PM)
+		rq->rq_flags |= RQF_PM;
 	if (data->flags & BLK_MQ_REQ_PREEMPT)
 		rq->rq_flags |= RQF_PREEMPT;
 	if (blk_queue_io_stat(data->q))
include/linux/blk-mq.h +2 -0
index b23eeca4d677..c00e856c6fb1 100644
@@ -444,6 +444,8 @@ enum {
 	BLK_MQ_REQ_NOWAIT	= (__force blk_mq_req_flags_t)(1 << 0),
 	/* allocate from reserved pool */
 	BLK_MQ_REQ_RESERVED	= (__force blk_mq_req_flags_t)(1 << 1),
+	/* set RQF_PM */
+	BLK_MQ_REQ_PM		= (__force blk_mq_req_flags_t)(1 << 2),
 	/* set RQF_PREEMPT */
 	BLK_MQ_REQ_PREEMPT	= (__force blk_mq_req_flags_t)(1 << 3),
 };
Reviewed-by on v4 2/9
#4 lore raw

Re: [PATCH v4 2/9] block: Introduce BLK_MQ_REQ_PM

In reply to Bart Van Assche
Quoted 12 lines · Bart Van Assche

Introduce the BLK_MQ_REQ_PM flag. This flag makes the request allocation functions set RQF_PM. This is the first step towards removing BLK_MQ_REQ_PREEMPT.

Cc Alan Stern <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
Cc Can Guo <[email protected]>
Quoted 6 lines
---
  block/blk-core.c       | 7 ++++---
  block/blk-mq.c         | 2 ++
  include/linux/blk-mq.h | 2 ++
  3 files changed, 8 insertions(+), 3 deletions(-)

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer
#5 lore raw

PATCH v4 3/9 ide: Do not set the RQF_PREEMPT flag for sense requests

In reply to Bart Van Assche

RQF_PREEMPT is used for two different purposes in the legacy IDE code:
1. To mark power management requests.
2. To mark requests that should preempt another request. An (old) explanation of that feature is as follows:
"The IDE driver in the Linux kernel normally uses a series of busywait delays during its initialization. When the driver executes these busywaits, the kernel does nothing for the duration of the wait. The time spent in these waits could be used for other initialization activities, if they could be run concurrently with these waits.

More specifically, busywait-style delays such as udelay() in module init functions inhibit kernel preemption because the Big Kernel Lock is held, while yielding APIs such as schedule_timeout() allow preemption. This is true because the kernel handles the BKL specially and releases and reacquires it across reschedules allowed by the current thread.

This IDE-preempt specification requires that the driver eliminate these busywaits and replace them with a mechanism that allows other work to proceed while the IDE driver is initializing."

Since I haven't found an implementation of (2), do not set the PREEMPT flag for sense requests. This patch causes sense requests to be postponed while a drive is suspended instead of being submitted to ide_queue_rq().

If it would ever be necessary to restore the IDE PREEMPT functionality, that can be done by introducing a new flag in struct ide_request.

Cc Christoph Hellwig <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc David S. Miller <[email protected]>
Cc Alan Stern <[email protected]>
Cc Can Guo <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
---
 drivers/ide/ide-atapi.c | 1 -
 drivers/ide/ide-io.c    | 5 -----
 2 files changed, 6 deletions(-)
drivers/ide/ide-atapi.c +0 -1
index 2162bc80f09e..013ad33fbbc8 100644
@@ -223,7 +223,6 @@ void ide_prep_sense(ide_drive_t *drive, struct request *rq)
 	sense_rq->rq_disk = rq->rq_disk;
 	sense_rq->cmd_flags = REQ_OP_DRV_IN;
 	ide_req(sense_rq)->type = ATA_PRIV_SENSE;
-	sense_rq->rq_flags |= RQF_PREEMPT;
 
 	req->cmd[0] = GPCMD_REQUEST_SENSE;
 	req->cmd[4] = cmd_len;
drivers/ide/ide-io.c +0 -5
index 1a53c7a75224..c210ea3bd02f 100644
@@ -515,11 +515,6 @@ blk_status_t ide_issue_rq(ide_drive_t *drive, struct request *rq,
 		 * above to return us whatever is in the queue. Since we call
 		 * ide_do_request() ourselves, we end up taking requests while
 		 * the queue is blocked...
-		 * 
-		 * We let requests forced at head of queue with ide-preempt
-		 * though. I hope that doesn't happen too much, hopefully not
-		 * unless the subdriver triggers such a thing in its own PM
-		 * state machine.
 		 */
 		if ((drive->dev_flags & IDE_DFLAG_BLOCKED) &&
 		    ata_pm_request(rq) == 0 &&
Reviewed-by on v4 3/9
#7 lore raw

Re: [PATCH v4 3/9] ide: Do not set the RQF_PREEMPT flag for sense requests

In reply to Bart Van Assche
Quoted 38 lines · Bart Van Assche

RQF_PREEMPT is used for two different purposes in the legacy IDE code:
1. To mark power management requests.
2. To mark requests that should preempt another request. An (old) explanation of that feature is as follows:
"The IDE driver in the Linux kernel normally uses a series of busywait delays during its initialization. When the driver executes these busywaits, the kernel does nothing for the duration of the wait. The time spent in these waits could be used for other initialization activities, if they could be run concurrently with these waits.

More specifically, busywait-style delays such as udelay() in module init functions inhibit kernel preemption because the Big Kernel Lock is held, while yielding APIs such as schedule_timeout() allow preemption. This is true because the kernel handles the BKL specially and releases and reacquires it across reschedules allowed by the current thread.

This IDE-preempt specification requires that the driver eliminate these busywaits and replace them with a mechanism that allows other work to proceed while the IDE driver is initializing."

Since I haven't found an implementation of (2), do not set the PREEMPT flag for sense requests. This patch causes sense requests to be postponed while a drive is suspended instead of being submitted to ide_queue_rq().

If it would ever be necessary to restore the IDE PREEMPT functionality, that can be done by introducing a new flag in struct ide_request.

Cc Christoph Hellwig <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc David S. Miller <[email protected]>
Cc Alan Stern <[email protected]>
Cc Can Guo <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
Quoted 5 lines
---
  drivers/ide/ide-atapi.c | 1 -
  drivers/ide/ide-io.c    | 5 -----
  2 files changed, 6 deletions(-)

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer
#8 lore raw

PATCH v4 4/9 ide: Mark power management requests with RQF_PM instead of RQF_PREEMPT

In reply to Bart Van Assche

This is another step that prepares for the removal of RQF_PREEMPT.

Cc Christoph Hellwig <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc David S. Miller <[email protected]>
Cc Alan Stern <[email protected]>
Cc Can Guo <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
---
 drivers/ide/ide-io.c | 2 +-
 drivers/ide/ide-pm.c | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
drivers/ide/ide-io.c +1 -1
index c210ea3bd02f..4867b67b60d6 100644
@@ -518,7 +518,7 @@ blk_status_t ide_issue_rq(ide_drive_t *drive, struct request *rq,
 		 */
 		if ((drive->dev_flags & IDE_DFLAG_BLOCKED) &&
 		    ata_pm_request(rq) == 0 &&
-		    (rq->rq_flags & RQF_PREEMPT) == 0) {
+		    (rq->rq_flags & RQF_PM) == 0) {
 			/* there should be no pending command at this point */
 			ide_unlock_port(hwif);
 			goto plug_device;
drivers/ide/ide-pm.c +1 -1
index 192e6c65d34e..82ab308f1aaf 100644
@@ -77,7 +77,7 @@ int generic_ide_resume(struct device *dev)
 	}
 
 	memset(&rqpm, 0, sizeof(rqpm));
-	rq = blk_get_request(drive->queue, REQ_OP_DRV_IN, BLK_MQ_REQ_PREEMPT);
+	rq = blk_get_request(drive->queue, REQ_OP_DRV_IN, BLK_MQ_REQ_PM);
 	ide_req(rq)->type = ATA_PRIV_PM_RESUME;
 	ide_req(rq)->special = &rqpm;
 	rqpm.pm_step = IDE_PM_START_RESUME;
Reviewed-by on v4 4/9
#10 lore raw

Re: [PATCH v4 4/9] ide: Mark power management requests with RQF_PM instead of RQF_PREEMPT

In reply to Bart Van Assche
Quoted 11 lines · Bart Van Assche

This is another step that prepares for the removal of RQF_PREEMPT.

Cc Christoph Hellwig <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc David S. Miller <[email protected]>
Cc Alan Stern <[email protected]>
Cc Can Guo <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
Quoted 5 lines
---
  drivers/ide/ide-io.c | 2 +-
  drivers/ide/ide-pm.c | 2 +-
  2 files changed, 2 insertions(+), 2 deletions(-)

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer
#11 lore raw

PATCH v4 5/9 scsi: Do not wait for a request in scsi_eh_lock_door()

In reply to Bart Van Assche

scsi_eh_lock_door() is the only function in the SCSI error handler that calls blk_get_request(). It is not guaranteed that a request is available when scsi_eh_lock_door() is called. Hence pass the BLK_MQ_REQ_NOWAIT flag to blk_get_request().

---
 drivers/scsi/scsi_error.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/scsi/scsi_error.c +6 -1
index d94449188270..6de6e1bf3dcb 100644
@@ -1993,7 +1993,12 @@ static void scsi_eh_lock_door(struct scsi_device *sdev)
 	struct request *req;
 	struct scsi_request *rq;
 
-	req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN, 0);
+	/*
+	 * It is not guaranteed that a request is available nor that
+	 * sdev->request_queue is unfrozen. Hence the BLK_MQ_REQ_NOWAIT below.
+	 */
+	req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN,
+			      BLK_MQ_REQ_NOWAIT);
 	if (IS_ERR(req))
 		return;
 	rq = scsi_req(req);
#12 lore raw

Re: [PATCH v4 5/9] scsi: Do not wait for a request in scsi_eh_lock_door()

Quoted 25 lines · Bart Van Assche

scsi_eh_lock_door() is the only function in the SCSI error handler that calls blk_get_request(). It is not guaranteed that a request is available when scsi_eh_lock_door() is called. Hence pass the BLK_MQ_REQ_NOWAIT flag to blk_get_request().

---
  drivers/scsi/scsi_error.c | 7 ++++++-
  1 file changed, 6 insertions(+), 1 deletion(-)
drivers/scsi/scsi_error.c +0 -0
index d94449188270..6de6e1bf3dcb 100644
@@ -1993,7 +1993,12 @@ static void scsi_eh_lock_door(struct scsi_device *sdev)
  	struct request *req;
  	struct scsi_request *rq;
 

drivers/scsi/scsi_error.c static void scsi_eh_lock_door(struct scsi_device *sdev) · line 2004

Quoted 10 lines
drivers/scsi/scsi_error.c +6 -1
-	req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN, 0);
+	/*
+	 * It is not guaranteed that a request is available nor that
+	 * sdev->request_queue is unfrozen. Hence the BLK_MQ_REQ_NOWAIT below.
+	 */
+	req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN,
+			      BLK_MQ_REQ_NOWAIT);
  	if (IS_ERR(req))
  		return;
  	rq = scsi_req(req);

Well ... had been thinking about that one, too.
The idea of this function is that prior to SCSI EH the device was locked
via scsi_set_medium_removal(). And during SCSI EH the device might have become unlocked, so we need to lock it again. However, scsi_set_medium_removal() not only issues the PREVENT_ALLOW_MEDIUM_REMOVAL command, but also sets the 'locked' flag based on the result. So if we fail to get a request here, shouldn't we unset the 'locked' flag, too?
And what does happen if we fail here? There is no return value, hence SCSI EH might run to completion, and the system will continue with an unlocked door ...
Not sure if that's a good idea.

But anyway, at the very least unset the 'locked' flag upon failure such that the internal state is correctly updated.

_Actually_, the flag should be unset after each successful SCSI EH step, to mirror the actual state. But this is probably out of scope for this patch.

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer
#13 lore raw

Re: [PATCH v4 5/9] scsi: Do not wait for a request in scsi_eh_lock_door()

Quoted 23 lines · Hannes Reinecke
Quoted 12 lines · Bart Van Assche
drivers/scsi/scsi_error.c +3 -0
index d94449188270..6de6e1bf3dcb 100644
@@ -1993,7 +1993,12 @@ static void scsi_eh_lock_door(struct
 scsi_device *sdev)
       struct request *req;
       struct scsi_request *rq;
   -    req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN, 0);
+    /*
+     * It is not guaranteed that a request is available nor that
+     * sdev->request_queue is unfrozen. Hence the BLK_MQ_REQ_NOWAIT
Quoted 8 lines

below.
+     */
+    req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN, +                  BLK_MQ_REQ_NOWAIT);
      if (IS_ERR(req))
          return;
      rq = scsi_req(req);

Well ... had been thinking about that one, too.

Quoted 8 lines

The idea of this function is that prior to SCSI EH the device was locked via scsi_set_medium_removal(). And during SCSI EH the device might have become unlocked, so we need to lock it again.
However, scsi_set_medium_removal() not only issues the PREVENT_ALLOW_MEDIUM_REMOVAL command, but also sets the 'locked' flag based on the result.
So if we fail to get a request here, shouldn't we unset the 'locked' flag, too?

Probably not. My interpretation of the 'locked' flag is that it represents the door state before error handling began. The following code in the SCSI error handler restores the door state after a bus reset:

	if (scsi_device_online(sdev) && sdev->was_reset && sdev->locked) {
		scsi_eh_lock_door(sdev);
		sdev->was_reset = 0;
	}
Quoted 4 lines

And what does happen if we fail here? There is no return value, hence SCSI EH might run to completion, and the system will continue with an unlocked door ...
Not sure if that's a good idea.

How about applying the following patch on top of patch 5/9?

drivers/scsi/scsi_error.c +5 -4
index 6de6e1bf3dcb..feac7262e40e 100644
@@ -1988,7 +1988,7 @@ static void eh_lock_door_done(struct request *req, blk_status_t status)
  * 	We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
  * 	head of the devices request queue, and continue.
  */
-static void scsi_eh_lock_door(struct scsi_device *sdev)
+static int scsi_eh_lock_door(struct scsi_device *sdev)
 {
 	struct request *req;
 	struct scsi_request *rq;
@@ -2000,7 +2000,7 @@ static void scsi_eh_lock_door(struct scsi_device *sdev)
 	req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN,
 			      BLK_MQ_REQ_NOWAIT);
 	if (IS_ERR(req))
-		return;
+		return PTR_ERR(req);
 	rq = scsi_req(req);
 
 	rq->cmd[0] = ALLOW_MEDIUM_REMOVAL;
@@ -2016,6 +2016,7 @@ static void scsi_eh_lock_door(struct scsi_device *sdev)
 	rq->retries = 5;
 
 	blk_execute_rq_nowait(req->q, NULL, req, 1, eh_lock_door_done);
+	return 0;
 }
 
 /**
@@ -2037,8 +2038,8 @@ static void scsi_restart_operations(struct Scsi_Host *shost)
 	 * is no point trying to lock the door of an off-line device.
 	 */
 	shost_for_each_device(sdev, shost) {
-		if (scsi_device_online(sdev) && sdev->was_reset && sdev->locked) {
-			scsi_eh_lock_door(sdev);
+		if (scsi_device_online(sdev) && sdev->was_reset &&
+		    sdev->locked && scsi_eh_lock_door(sdev) == 0) {
 			sdev->was_reset = 0;
 		}
 	}

Thanks,

Bart.

#14 lore raw

Re: [PATCH v4 5/9] scsi: Do not wait for a request in scsi_eh_lock_door()

Quoted 81 lines · Bart Van Assche
Quoted 23 lines · Hannes Reinecke
Quoted 12 lines · Bart Van Assche
drivers/scsi/scsi_error.c +3 -0
index d94449188270..6de6e1bf3dcb 100644
@@ -1993,7 +1993,12 @@ static void scsi_eh_lock_door(struct
 scsi_device *sdev)
       struct request *req;
       struct scsi_request *rq;
   -    req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN, 0);
+    /*
+     * It is not guaranteed that a request is available nor that
+     * sdev->request_queue is unfrozen. Hence the BLK_MQ_REQ_NOWAIT
Quoted 8 lines

below.
+     */
+    req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN, +                  BLK_MQ_REQ_NOWAIT);
      if (IS_ERR(req))
          return;
      rq = scsi_req(req);

Well ... had been thinking about that one, too.

Quoted 8 lines

The idea of this function is that prior to SCSI EH the device was locked via scsi_set_medium_removal(). And during SCSI EH the device might have become unlocked, so we need to lock it again.
However, scsi_set_medium_removal() not only issues the PREVENT_ALLOW_MEDIUM_REMOVAL command, but also sets the 'locked' flag based on the result.
So if we fail to get a request here, shouldn't we unset the 'locked' flag, too?

Probably not. My interpretation of the 'locked' flag is that it represents the door state before error handling began. The following code in the SCSI error handler restores the door state after a bus reset:

	if (scsi_device_online(sdev) && sdev->was_reset && sdev->locked) {
		scsi_eh_lock_door(sdev);
		sdev->was_reset = 0;
	}
Quoted 4 lines

And what does happen if we fail here? There is no return value, hence SCSI EH might run to completion, and the system will continue with an unlocked door ...
Not sure if that's a good idea.

How about applying the following patch on top of patch 5/9?

drivers/scsi/scsi_error.c +3 -2
index 6de6e1bf3dcb..feac7262e40e 100644
@@ -1988,7 +1988,7 @@ static void eh_lock_door_done(struct request *req, blk_status_t status)
   * 	We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
   * 	head of the devices request queue, and continue.
   */
-static void scsi_eh_lock_door(struct scsi_device *sdev)
+static int scsi_eh_lock_door(struct scsi_device *sdev)
  {
  	struct request *req;
  	struct scsi_request *rq;
@@ -2000,7 +2000,7 @@ static void scsi_eh_lock_door(struct scsi_device *sdev)
  	req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN,
  			      BLK_MQ_REQ_NOWAIT);
  	if (IS_ERR(req))
-		return;
+		return PTR_ERR(req);
  	rq = scsi_req(req);
 
  	rq->cmd[0] = ALLOW_MEDIUM_REMOVAL;
@@ -2016,6 +2016,7 @@ static void scsi_eh_lock_door(struct scsi_device *sdev)
  	rq->retries = 5;
 
  	blk_execute_rq_nowait(req->q, NULL, req, 1, eh_lock_door_done);
+	return 0;
  }
 
  /**

drivers/scsi/scsi_error.c static void scsi_restart_operations(struct Scsi_Host *shost) · line 2045

Quoted 10 lines
drivers/scsi/scsi_error.c +2 -2
@@ -2037,8 +2038,8 @@ static void scsi_restart_operations(struct Scsi_Host *shost)
  	 * is no point trying to lock the door of an off-line device.
  	 */
  	shost_for_each_device(sdev, shost) {
-		if (scsi_device_online(sdev) && sdev->was_reset && sdev->locked) {
-			scsi_eh_lock_door(sdev);
+		if (scsi_device_online(sdev) && sdev->was_reset &&
+		    sdev->locked && scsi_eh_lock_door(sdev) == 0) {
  			sdev->was_reset = 0;
  		}
  	}

I probably didn't make myself clear.
As per SBC (in this case, sbc3r36) the effects of PREVENT_ALLOW_MEDIUM_REMOVAL are being reset by a successfull LUN Reset, Hard Reset, Power/On Reset, or an I_T Nexus loss. Which incidentally maps nicely onto SCSI EH, so after a successful SCSI EH the door will be unlocked (which is why we need to call scsi_eh_lock_door()). In the SCSI midlayer this state is being reflected by the 'locked' flag.
Now, if scsi_eh_lock_door() is _not_ being executed due to a blk_get_request() failure, the device remains unlocked, and as such the 'locked' flag would need to be _unset_.

So I was thinking more along these lines:

@@ -2030,7 +2037,8 @@ static void scsi_restart_operations(struct Scsi_Host *shost)

         */
        shost_for_each_device(sdev, shost) {
                if (scsi_device_online(sdev) && sdev->was_reset && 
sdev->locked) {
-                       scsi_eh_lock_door(sdev);
+                       if (scsi_eh_lock_door(sdev) < 0)
+                               sdev->locked = 0;
                        sdev->was_reset = 0;
                }
        }

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer
#15 lore raw

Re: [PATCH v4 5/9] scsi: Do not wait for a request in scsi_eh_lock_door()

Quoted 106 lines · Hannes Reinecke
Quoted 81 lines · Bart Van Assche
Quoted 23 lines · Hannes Reinecke
Quoted 12 lines · Bart Van Assche
drivers/scsi/scsi_error.c +3 -0
index d94449188270..6de6e1bf3dcb 100644
@@ -1993,7 +1993,12 @@ static void scsi_eh_lock_door(struct
 scsi_device *sdev)
       struct request *req;
       struct scsi_request *rq;
   -    req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN, 0);
+    /*
+     * It is not guaranteed that a request is available nor that
+     * sdev->request_queue is unfrozen. Hence the BLK_MQ_REQ_NOWAIT
Quoted 8 lines

below.
+     */
+    req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN, +                  BLK_MQ_REQ_NOWAIT);
      if (IS_ERR(req))
          return;
      rq = scsi_req(req);

Well ... had been thinking about that one, too.

Quoted 8 lines

The idea of this function is that prior to SCSI EH the device was locked via scsi_set_medium_removal(). And during SCSI EH the device might have become unlocked, so we need to lock it again.
However, scsi_set_medium_removal() not only issues the PREVENT_ALLOW_MEDIUM_REMOVAL command, but also sets the 'locked' flag based on the result.
So if we fail to get a request here, shouldn't we unset the 'locked' flag, too?

Probably not. My interpretation of the 'locked' flag is that it represents the door state before error handling began. The following code in the SCSI error handler restores the door state after a bus reset:

	if (scsi_device_online(sdev) && sdev->was_reset && sdev->locked) {
		scsi_eh_lock_door(sdev);
		sdev->was_reset = 0;
	}
Quoted 4 lines

And what does happen if we fail here? There is no return value, hence SCSI EH might run to completion, and the system will continue with an unlocked door ...
Not sure if that's a good idea.

How about applying the following patch on top of patch 5/9?

drivers/scsi/scsi_error.c +3 -2
index 6de6e1bf3dcb..feac7262e40e 100644
@@ -1988,7 +1988,7 @@ static void eh_lock_door_done(struct request *req, blk_status_t status)
   * 	We queue up an asynchronous "ALLOW MEDIUM REMOVAL" request on the
   * 	head of the devices request queue, and continue.
   */
-static void scsi_eh_lock_door(struct scsi_device *sdev)
+static int scsi_eh_lock_door(struct scsi_device *sdev)
  {
  	struct request *req;
  	struct scsi_request *rq;
@@ -2000,7 +2000,7 @@ static void scsi_eh_lock_door(struct scsi_device *sdev)
  	req = blk_get_request(sdev->request_queue, REQ_OP_SCSI_IN,
  			      BLK_MQ_REQ_NOWAIT);
  	if (IS_ERR(req))
-		return;
+		return PTR_ERR(req);
  	rq = scsi_req(req);
 
  	rq->cmd[0] = ALLOW_MEDIUM_REMOVAL;
@@ -2016,6 +2016,7 @@ static void scsi_eh_lock_door(struct scsi_device *sdev)
  	rq->retries = 5;
 
  	blk_execute_rq_nowait(req->q, NULL, req, 1, eh_lock_door_done);
+	return 0;
  }
 
  /**

drivers/scsi/scsi_error.c static void scsi_restart_operations(struct Scsi_Host *shost) · line 2045

Quoted 10 lines
drivers/scsi/scsi_error.c +2 -2
@@ -2037,8 +2038,8 @@ static void scsi_restart_operations(struct Scsi_Host *shost)
  	 * is no point trying to lock the door of an off-line device.
  	 */
  	shost_for_each_device(sdev, shost) {
-		if (scsi_device_online(sdev) && sdev->was_reset && sdev->locked) {
-			scsi_eh_lock_door(sdev);
+		if (scsi_device_online(sdev) && sdev->was_reset &&
+		    sdev->locked && scsi_eh_lock_door(sdev) == 0) {
  			sdev->was_reset = 0;
  		}
  	}

I probably didn't make myself clear.
As per SBC (in this case, sbc3r36) the effects of PREVENT_ALLOW_MEDIUM_REMOVAL are being reset by a successfull LUN Reset, Hard Reset, Power/On Reset, or an I_T Nexus loss. Which incidentally maps nicely onto SCSI EH, so after a successful SCSI EH the door will be unlocked (which is why we need to call scsi_eh_lock_door()).
In the SCSI midlayer this state is being reflected by the 'locked' flag. Now, if scsi_eh_lock_door() is _not_ being executed due to a blk_get_request() failure, the device remains unlocked, and as such the 'locked' flag would need to be _unset_.

So I was thinking more along these lines:

Review static void scsi_restart_operations(struct Scsi_Host · line 2043

Quoted 8 lines
(quoted diff) +2 -1
@@ -2030,7 +2037,8 @@ static void scsi_restart_operations(struct Scsi_Host
 *shost)
         */
        shost_for_each_device(sdev, shost) {
                if (scsi_device_online(sdev) && sdev->was_reset &&
 sdev->locked) {
-                       scsi_eh_lock_door(sdev);
+                       if (scsi_eh_lock_door(sdev) < 0)
+                               sdev->locked = 0;

BTW, scsi_eh_lock_door() returns void, and it can't be sync because there may not be any driver tag available. Even though it is available, the host state isn't running yet, so the command can't be queued to LLD yet.

Maybe the above lines should be put after host state is updated to RUNNING.

Also changing to NOWAIT can't avoid the issue completely, what if 'none' is used?

Thanks,
Ming

#16 lore raw

Re: [PATCH v4 5/9] scsi: Do not wait for a request in scsi_eh_lock_door()

Quoted 10 lines · Ming Lei

BTW, scsi_eh_lock_door() returns void, and it can't be sync because there may not be any driver tag available. Even though it is available, the host state isn't running yet, so the command can't be queued to LLD yet.

Maybe the above lines should be put after host state is updated to RUNNING.

Also changing to NOWAIT can't avoid the issue completely, what if 'none' is used?

Hi Ming,

I am considering to drop this patch since the latest version of the SPI DV patch no longer introduces a new blk_mq_freeze_queue() call in the SPI DV code. In other words, any potential issues with scsi_eh_lock_door() are existing issues and are not made worse by my patch series.

Thanks,

Bart.

#17 lore raw

PATCH v4 6/9 scsi_transport_spi: Set RQF_PM for domain validation commands

In reply to Bart Van Assche

Disable runtime power management during domain validation. Since a later patch removes RQF_PREEMPT, set RQF_PM for domain validation commands such that these are executed in the quiesced SCSI device state.

Cc Alan Stern <[email protected]>
Cc Christoph Hellwig <[email protected]>
Cc James Bottomley <[email protected]>
Cc Woody Suwalski <[email protected]>
Cc Can Guo <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
Cc Stan Johnson <[email protected]>
---
 drivers/scsi/scsi_transport_spi.c | 27 +++++++++++++++++++--------
 1 file changed, 19 insertions(+), 8 deletions(-)
drivers/scsi/scsi_transport_spi.c +19 -8
index f3d5b1bbd5aa..c37dd15d16d2 100644
@@ -117,12 +117,16 @@ static int spi_execute(struct scsi_device *sdev, const void *cmd,
 		sshdr = &sshdr_tmp;
 
 	for(i = 0; i < DV_RETRIES; i++) {
+		/*
+		 * The purpose of the RQF_PM flag below is to bypass the
+		 * SDEV_QUIESCE state.
+		 */
 		result = scsi_execute(sdev, cmd, dir, buffer, bufflen, sense,
 				      sshdr, DV_TIMEOUT, /* retries */ 1,
 				      REQ_FAILFAST_DEV |
 				      REQ_FAILFAST_TRANSPORT |
 				      REQ_FAILFAST_DRIVER,
-				      0, NULL);
+				      RQF_PM, NULL);
 		if (driver_byte(result) != DRIVER_SENSE ||
 		    sshdr->sense_key != UNIT_ATTENTION)
 			break;
@@ -1005,23 +1009,26 @@ spi_dv_device(struct scsi_device *sdev)
 	 */
 	lock_system_sleep();
 
+	if (scsi_autopm_get_device(sdev))
+		goto unlock_system_sleep;
+
 	if (unlikely(spi_dv_in_progress(starget)))
-		goto unlock;
+		goto put_autopm;
 
 	if (unlikely(scsi_device_get(sdev)))
-		goto unlock;
+		goto put_autopm;
 
 	spi_dv_in_progress(starget) = 1;
 
 	buffer = kzalloc(len, GFP_KERNEL);
 
 	if (unlikely(!buffer))
-		goto out_put;
+		goto put_sdev;
 
 	/* We need to verify that the actual device will quiesce; the
 	 * later target quiesce is just a nice to have */
 	if (unlikely(scsi_device_quiesce(sdev)))
-		goto out_free;
+		goto free_buffer;
 
 	scsi_target_quiesce(starget);
 
@@ -1041,12 +1048,16 @@ spi_dv_device(struct scsi_device *sdev)
 
 	spi_initial_dv(starget) = 1;
 
- out_free:
+free_buffer:
 	kfree(buffer);
- out_put:
+
+put_sdev:
 	spi_dv_in_progress(starget) = 0;
 	scsi_device_put(sdev);
-unlock:
+put_autopm:
+	scsi_autopm_put_device(sdev);
+
+unlock_system_sleep:
 	unlock_system_sleep();
 }
 EXPORT_SYMBOL(spi_dv_device);
#19 lore raw

PATCH v4 7/9 scsi: Only process PM requests if rpm_status != RPM_ACTIVE

In reply to Bart Van Assche

Instead of submitting all SCSI commands submitted with scsi_execute() to a SCSI device if rpm_status != RPM_ACTIVE, only submit RQF_PM (power management requests) if rpm_status != RPM_ACTIVE. This patch makes the SCSI core handle the runtime power management status (rpm_status) as it should be handled.

Cc Christoph Hellwig <[email protected]>
Cc Can Guo <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Alan Stern <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
Cc Martin Kepplinger <[email protected]>
---
 drivers/scsi/scsi_lib.c | 27 ++++++++++++++-------------
 1 file changed, 14 insertions(+), 13 deletions(-)
drivers/scsi/scsi_lib.c +14 -13
index b7ac14571415..91bc39a4c3c3 100644
@@ -249,7 +249,8 @@ int __scsi_execute(struct scsi_device *sdev, const unsigned char *cmd,
 
 	req = blk_get_request(sdev->request_queue,
 			data_direction == DMA_TO_DEVICE ?
-			REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN, BLK_MQ_REQ_PREEMPT);
+			REQ_OP_SCSI_OUT : REQ_OP_SCSI_IN,
+			rq_flags & RQF_PM ? BLK_MQ_REQ_PM : 0);
 	if (IS_ERR(req))
 		return ret;
 	rq = scsi_req(req);
@@ -1206,6 +1207,8 @@ static blk_status_t
 scsi_device_state_check(struct scsi_device *sdev, struct request *req)
 {
 	switch (sdev->sdev_state) {
+	case SDEV_CREATED:
+		return BLK_STS_OK;
 	case SDEV_OFFLINE:
 	case SDEV_TRANSPORT_OFFLINE:
 		/*
@@ -1232,18 +1235,18 @@ scsi_device_state_check(struct scsi_device *sdev, struct request *req)
 		return BLK_STS_RESOURCE;
 	case SDEV_QUIESCE:
 		/*
-		 * If the devices is blocked we defer normal commands.
+		 * If the device is blocked we only accept power management
+		 * commands.
 		 */
-		if (req && !(req->rq_flags & RQF_PREEMPT))
+		if (req && WARN_ON_ONCE(!(req->rq_flags & RQF_PM)))
 			return BLK_STS_RESOURCE;
 		return BLK_STS_OK;
 	default:
 		/*
 		 * For any other not fully online state we only allow
-		 * special commands.  In particular any user initiated
-		 * command is not allowed.
+		 * power management commands.
 		 */
-		if (req && !(req->rq_flags & RQF_PREEMPT))
+		if (req && !(req->rq_flags & RQF_PM))
 			return BLK_STS_IOERR;
 		return BLK_STS_OK;
 	}
@@ -2517,15 +2520,13 @@ void sdev_evt_send_simple(struct scsi_device *sdev,
 EXPORT_SYMBOL_GPL(sdev_evt_send_simple);
 
 /**
- *	scsi_device_quiesce - Block user issued commands.
+ *	scsi_device_quiesce - Block all commands except power management.
  *	@sdev:	scsi device to quiesce.
  *
  *	This works by trying to transition to the SDEV_QUIESCE state
  *	(which must be a legal transition).  When the device is in this
- *	state, only special requests will be accepted, all others will
- *	be deferred.  Since special requests may also be requeued requests,
- *	a successful return doesn't guarantee the device will be
- *	totally quiescent.
+ *	state, only power management requests will be accepted, all others will
+ *	be deferred.
  *
  *	Must be called with user context, may sleep.
  *
@@ -2587,12 +2588,12 @@ void scsi_device_resume(struct scsi_device *sdev)
 	 * device deleted during suspend)
 	 */
 	mutex_lock(&sdev->state_mutex);
+	if (sdev->sdev_state == SDEV_QUIESCE)
+		scsi_device_set_state(sdev, SDEV_RUNNING);
 	if (sdev->quiesced_by) {
 		sdev->quiesced_by = NULL;
 		blk_clear_pm_only(sdev->request_queue);
 	}
-	if (sdev->sdev_state == SDEV_QUIESCE)
-		scsi_device_set_state(sdev, SDEV_RUNNING);
 	mutex_unlock(&sdev->state_mutex);
 }
 EXPORT_SYMBOL(scsi_device_resume);
#21 lore raw

Re: [PATCH v4 7/9] scsi: Only process PM requests if rpm_status != RPM_ACTIVE

In reply to Bart Van Assche
Quoted 14 lines · Bart Van Assche

Instead of submitting all SCSI commands submitted with scsi_execute() to a SCSI device if rpm_status != RPM_ACTIVE, only submit RQF_PM (power management requests) if rpm_status != RPM_ACTIVE. This patch makes the SCSI core handle the runtime power management status (rpm_status) as it should be handled.

Cc Christoph Hellwig <[email protected]>
Cc Can Guo <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Alan Stern <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
Cc Martin Kepplinger <[email protected]>
Quoted 4 lines
---
  drivers/scsi/scsi_lib.c | 27 ++++++++++++++-------------
  1 file changed, 14 insertions(+), 13 deletions(-)

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer
#22 lore raw

PATCH v4 8/9 block: Remove RQF_PREEMPT and BLK_MQ_REQ_PREEMPT

In reply to Bart Van Assche

Remove flag RQF_PREEMPT and BLK_MQ_REQ_PREEMPT since these are no longer used by any kernel code.

Cc Christoph Hellwig <[email protected]>
Cc Can Guo <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Alan Stern <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
Cc Martin Kepplinger <[email protected]>
---
 block/blk-core.c       | 7 +++----
 block/blk-mq-debugfs.c | 1 -
 block/blk-mq.c         | 2 --
 include/linux/blk-mq.h | 2 --
 include/linux/blkdev.h | 6 +-----
 5 files changed, 4 insertions(+), 14 deletions(-)
block/blk-core.c +3 -4
index 10696f9fb6ac..a00bce9f46d8 100644
@@ -424,11 +424,11 @@ EXPORT_SYMBOL(blk_cleanup_queue);
 /**
  * blk_queue_enter() - try to increase q->q_usage_counter
  * @q: request queue pointer
- * @flags: BLK_MQ_REQ_NOWAIT, BLK_MQ_REQ_PM and/or BLK_MQ_REQ_PREEMPT
+ * @flags: BLK_MQ_REQ_NOWAIT and/or BLK_MQ_REQ_PM
  */
 int blk_queue_enter(struct request_queue *q, blk_mq_req_flags_t flags)
 {
-	const bool pm = flags & (BLK_MQ_REQ_PM | BLK_MQ_REQ_PREEMPT);
+	const bool pm = flags & BLK_MQ_REQ_PM;
 
 	while (true) {
 		bool success = false;
@@ -630,8 +630,7 @@ struct request *blk_get_request(struct request_queue *q, unsigned int op,
 	struct request *req;
 
 	WARN_ON_ONCE(op & REQ_NOWAIT);
-	WARN_ON_ONCE(flags & ~(BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_PM |
-			       BLK_MQ_REQ_PREEMPT));
+	WARN_ON_ONCE(flags & ~(BLK_MQ_REQ_NOWAIT | BLK_MQ_REQ_PM));
 
 	req = blk_mq_alloc_request(q, op, flags);
 	if (!IS_ERR(req) && q->mq_ops->initialize_rq_fn)
block/blk-mq-debugfs.c +0 -1
index 3094542e12ae..9336a6f8d6ef 100644
@@ -297,7 +297,6 @@ static const char *const rqf_name[] = {
 	RQF_NAME(MIXED_MERGE),
 	RQF_NAME(MQ_INFLIGHT),
 	RQF_NAME(DONTPREP),
-	RQF_NAME(PREEMPT),
 	RQF_NAME(FAILED),
 	RQF_NAME(QUIET),
 	RQF_NAME(ELVPRIV),
block/blk-mq.c +0 -2
index b5880a1fb38d..d50504888b68 100644
@@ -294,8 +294,6 @@ static struct request *blk_mq_rq_ctx_init(struct blk_mq_alloc_data *data,
 	rq->cmd_flags = data->cmd_flags;
 	if (data->flags & BLK_MQ_REQ_PM)
 		rq->rq_flags |= RQF_PM;
-	if (data->flags & BLK_MQ_REQ_PREEMPT)
-		rq->rq_flags |= RQF_PREEMPT;
 	if (blk_queue_io_stat(data->q))
 		rq->rq_flags |= RQF_IO_STAT;
 	INIT_LIST_HEAD(&rq->queuelist);
include/linux/blk-mq.h +0 -2
index c00e856c6fb1..88af1df94308 100644
@@ -446,8 +446,6 @@ enum {
 	BLK_MQ_REQ_RESERVED	= (__force blk_mq_req_flags_t)(1 << 1),
 	/* set RQF_PM */
 	BLK_MQ_REQ_PM		= (__force blk_mq_req_flags_t)(1 << 2),
-	/* set RQF_PREEMPT */
-	BLK_MQ_REQ_PREEMPT	= (__force blk_mq_req_flags_t)(1 << 3),
 };
 
 struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op,
include/linux/blkdev.h +1 -5
index 639cae2c158b..7d4b746f7e6a 100644
@@ -79,9 +79,6 @@ typedef __u32 __bitwise req_flags_t;
 #define RQF_MQ_INFLIGHT		((__force req_flags_t)(1 << 6))
 /* don't call prep for this one */
 #define RQF_DONTPREP		((__force req_flags_t)(1 << 7))
-/* set for "ide_preempt" requests and also for requests for which the SCSI
-   "quiesce" state must be ignored. */
-#define RQF_PREEMPT		((__force req_flags_t)(1 << 8))
 /* vaguely specified driver internal error.  Ignored by the block layer */
 #define RQF_FAILED		((__force req_flags_t)(1 << 10))
 /* don't warn about errors */
@@ -430,8 +427,7 @@ struct request_queue {
 	unsigned long		queue_flags;
 	/*
 	 * Number of contexts that have called blk_set_pm_only(). If this
-	 * counter is above zero then only RQF_PM and RQF_PREEMPT requests are
-	 * processed.
+	 * counter is above zero then only RQF_PM requests are processed.
 	 */
 	atomic_t		pm_only;
 
Reviewed-by on v4 8/9
#24 lore raw

Re: [PATCH v4 8/9] block: Remove RQF_PREEMPT and BLK_MQ_REQ_PREEMPT

In reply to Bart Van Assche
Quoted 11 lines · Bart Van Assche

Remove flag RQF_PREEMPT and BLK_MQ_REQ_PREEMPT since these are no longer used by any kernel code.

Cc Christoph Hellwig <[email protected]>
Cc Can Guo <[email protected]>
Cc Stanley Chu <[email protected]>
Cc Alan Stern <[email protected]>
Cc Ming Lei <[email protected]>
Cc Rafael J. Wysocki <[email protected]>
Cc Martin Kepplinger <[email protected]>
Quoted 8 lines
---
  block/blk-core.c       | 7 +++----
  block/blk-mq-debugfs.c | 1 -
  block/blk-mq.c         | 2 --
  include/linux/blk-mq.h | 2 --
  include/linux/blkdev.h | 6 +-----
  5 files changed, 4 insertions(+), 14 deletions(-)

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer
#25 lore raw

PATCH v4 9/9 block: Do not accept any requests while suspended

In reply to Bart Van Assche

From: Alan Stern <[email protected]>

blk_queue_enter() accepts BLK_MQ_REQ_PM requests independent of the runtime power management state. Now that SCSI domain validation no longer depends on this behavior, modify the behavior of blk_queue_enter() as follows:
- Do not accept any requests while suspended.
- Only process power management requests while suspending or resuming.

Submitting BLK_MQ_REQ_PM requests to a device that is runtime suspended causes runtime-suspended devices not to resume as they should. The request which should cause a runtime resume instead gets issued directly, without resuming the device first. Of course the device can't handle it properly, the I/O fails, and the device remains suspended.

The problem is fixed by checking that the queue's runtime-PM status isn't RPM_SUSPENDED before allowing a request to be issued, and queuing a runtime-resume request if it is. In particular, the inline blk_pm_request_resume() routine is renamed blk_pm_resume_queue() and the code is unified by merging the surrounding checks into the routine. If the queue isn't set up for runtime PM, or there currently is no restriction on allowed requests, the request is allowed. Likewise if the BLK_MQ_REQ_PM flag is set and the status isn't RPM_SUSPENDED. Otherwise a runtime resume is queued and the request is blocked until conditions are more suitable.

[ bvanassche: modified commit message and removed Cc: stable because without
the previous patches from this series this patch would break parallel SCSI domain validation ]

---
 block/blk-core.c |  6 +++---
 block/blk-pm.h   | 14 +++++++++-----
 2 files changed, 12 insertions(+), 8 deletions(-)
block/blk-core.c +3 -3
index a00bce9f46d8..230880cbf8c8 100644
@@ -440,7 +440,8 @@ int blk_queue_enter(struct request_queue *q, blk_mq_req_flags_t flags)
 			 * responsible for ensuring that that counter is
 			 * globally visible before the queue is unfrozen.
 			 */
-			if (pm || !blk_queue_pm_only(q)) {
+			if ((pm && q->rpm_status != RPM_SUSPENDED) ||
+			    !blk_queue_pm_only(q)) {
 				success = true;
 			} else {
 				percpu_ref_put(&q->q_usage_counter);
@@ -465,8 +466,7 @@ int blk_queue_enter(struct request_queue *q, blk_mq_req_flags_t flags)
 
 		wait_event(q->mq_freeze_wq,
 			   (!q->mq_freeze_depth &&
-			    (pm || (blk_pm_request_resume(q),
-				    !blk_queue_pm_only(q)))) ||
+			    blk_pm_resume_queue(pm, q)) ||
 			   blk_queue_dying(q));
 		if (blk_queue_dying(q))
 			return -ENODEV;
block/blk-pm.h +9 -5
index ea5507d23e75..a2283cc9f716 100644
@@ -6,11 +6,14 @@
 #include <linux/pm_runtime.h>
 
 #ifdef CONFIG_PM
-static inline void blk_pm_request_resume(struct request_queue *q)
+static inline int blk_pm_resume_queue(const bool pm, struct request_queue *q)
 {
-	if (q->dev && (q->rpm_status == RPM_SUSPENDED ||
-		       q->rpm_status == RPM_SUSPENDING))
-		pm_request_resume(q->dev);
+	if (!q->dev || !blk_queue_pm_only(q))
+		return 1;	/* Nothing to do */
+	if (pm && q->rpm_status != RPM_SUSPENDED)
+		return 1;	/* Request allowed */
+	pm_request_resume(q->dev);
+	return 0;
 }
 
 static inline void blk_pm_mark_last_busy(struct request *rq)
@@ -44,8 +47,9 @@ static inline void blk_pm_put_request(struct request *rq)
 		--rq->q->nr_pending;
 }
 #else
-static inline void blk_pm_request_resume(struct request_queue *q)
+static inline int blk_pm_resume_queue(const bool pm, struct request_queue *q)
 {
+	return 1;
 }
 
 static inline void blk_pm_mark_last_busy(struct request *rq)
Reviewed-by on v4 9/9
#26 lore raw

Re: [PATCH v4 9/9] block: Do not accept any requests while suspended

Quoted 33 lines · Bart Van Assche

From: Alan Stern <[email protected]>

blk_queue_enter() accepts BLK_MQ_REQ_PM requests independent of the runtime power management state. Now that SCSI domain validation no longer depends on this behavior, modify the behavior of blk_queue_enter() as follows:
- Do not accept any requests while suspended.
- Only process power management requests while suspending or resuming.

Submitting BLK_MQ_REQ_PM requests to a device that is runtime suspended causes runtime-suspended devices not to resume as they should. The request which should cause a runtime resume instead gets issued directly, without resuming the device first. Of course the device can't handle it properly, the I/O fails, and the device remains suspended.

The problem is fixed by checking that the queue's runtime-PM status isn't RPM_SUSPENDED before allowing a request to be issued, and queuing a runtime-resume request if it is. In particular, the inline blk_pm_request_resume() routine is renamed blk_pm_resume_queue() and the code is unified by merging the surrounding checks into the routine. If the queue isn't set up for runtime PM, or there currently is no restriction on allowed requests, the request is allowed. Likewise if the BLK_MQ_REQ_PM flag is set and the status isn't RPM_SUSPENDED. Otherwise a runtime resume is queued and the request is blocked until conditions are more suitable.

Quoted 8 lines

[ bvanassche: modified commit message and removed Cc: stable because without
the previous patches from this series this patch would break parallel SCSI domain validation ]

---
  block/blk-core.c |  6 +++---
  block/blk-pm.h   | 14 +++++++++-----
  2 files changed, 12 insertions(+), 8 deletions(-)

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer
#27 lore raw
In reply to Bart Van Assche

Jens,

Any objections to me picking this up?

Quoted 11 lines

The SCSI runtime suspend and SPI domain validation mechanisms both use scsi_device_quiesce(). scsi_device_quiesce() restricts blk_queue_enter() to BLK_MQ_REQ_PREEMPT requests. There is a conflict between the requirements of runtime suspend and SCSI domain
validation: no requests must be sent to runtime suspended devices that are in the state RPM_SUSPENDED while BLK_MQ_REQ_PREEMPT requests must be processed during SCSI domain validation. This conflict is resolved by reworking the SCSI domain validation implementation.

Hybernation, runtime suspend and SCSI domain validation have been retested.

Signature · 1 line
Martin K. Petersen	Oracle Linux Engineering
maintainer Reviewed-by
#28 lore raw
Quoted 4 lines · Martin K. Petersen

Jens,

Any objections to me picking this up?

No, go ahead. Looks good to me, you can add:

to the series.

Signature · 1 line
Jens Axboe
#29 lore raw
In reply to Bart Van Assche

Bart,

Quoted 8 lines

The SCSI runtime suspend and SPI domain validation mechanisms both use scsi_device_quiesce(). scsi_device_quiesce() restricts blk_queue_enter() to BLK_MQ_REQ_PREEMPT requests. There is a conflict between the requirements of runtime suspend and SCSI domain
validation: no requests must be sent to runtime suspended devices that are in the state RPM_SUSPENDED while BLK_MQ_REQ_PREEMPT requests must be processed during SCSI domain validation. This conflict is resolved by reworking the SCSI domain validation implementation.

Applied to 5.11/scsi-staging except for patch #5. Appreciate all that work that went into this series!

Signature · 1 line
Martin K. Petersen	Oracle Linux Engineering