[PATCHv3 0/5] dm-multipath: push back requests instead of queueing

[PATCHv3 0/5] dm-multipath: push back requests instead of queueing

[Hannes Reinecke]

Hi all,

dm-multipath still carries around it's own queueing framework for
implementing 'queue_if_no_path'.
However, there is no real reason for this; we could as well
push back the requests onto the request_queue.
In doing so we can also reduce the memory pressure during
fail_if_no_path scenarios, as we don't have to allocate a
context for each request when it need to be requeued.

This patchset is the third version, including
the review by Jun'ichi Nomura and suggestions from Mike Snitzer.

Hannes Reinecke (5):
  dm-multipath: Do not call pg_init twice
  dm-multipath: push back requests instead of queueing
  dm-multipath: remove process_queued_ios()
  dm-multipath: reduce memory pressure during requeuing
  dm-multipath: remove map_io()

 drivers/md/dm-mpath.c         | 214 +++++++++++++++---------------------------
 drivers/md/dm.c               |  13 +++
 include/linux/device-mapper.h |   1 +
 3 files changed, 88 insertions(+), 140 deletions(-)

-- 
1.7.12.4

[PATCH 1/5] dm-multipath: Do not call pg_init twice

[Hannes Reinecke]

When pg_init is running we shouldn't be calling the same
routine twice; we need to wait for the first pg_init to
complete.

Signed-off-by: Hannes Reinecke <hare@suse.de>
---
 drivers/md/dm-mpath.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/drivers/md/dm-mpath.c b/drivers/md/dm-mpath.c
index 6eb9dc9..d45290a 100644
--- a/drivers/md/dm-mpath.c
+++ b/drivers/md/dm-mpath.c
@@ -261,6 +261,9 @@ static void __pg_init_all_paths(struct multipath *m)
 	struct pgpath *pgpath;
 	unsigned long pg_init_delay = 0;
 
+	if (m->pg_init_in_progress || m->pg_init_disabled)
+		return;
+
 	m->pg_init_count++;
 	m->pg_init_required = 0;
 	if (m->pg_init_delay_retry)
-- 
1.7.12.4

[PATCH 2/5] dm-multipath: push back requests instead of queueing

[Hannes Reinecke]

There is no reason why multipath needs to queue requests
internally for queue_if_no_path or pg_init; we should
rather push them back onto the request queue.

And while we're at it we can simplify the conditional
statement in map_io() to make it easier to read.

Cc: Mike Snitzer <snitzer@redhat.com>
Cc: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Hannes Reinecke <hare@suse.de>
---
 drivers/md/dm-mpath.c         | 115 ++++++++++++++----------------------------
 drivers/md/dm.c               |  13 +++++
 include/linux/device-mapper.h |   1 +
 3 files changed, 52 insertions(+), 77 deletions(-)

diff --git a/drivers/md/dm-mpath.c b/drivers/md/dm-mpath.c
index d45290a..39fd166 100644
--- a/drivers/md/dm-mpath.c
+++ b/drivers/md/dm-mpath.c
@@ -93,9 +93,7 @@ struct multipath {
 	unsigned pg_init_count;		/* Number of times pg_init called */
 	unsigned pg_init_delay_msecs;	/* Number of msecs before pg_init retry */
 
-	unsigned queue_size;
 	struct work_struct process_queued_ios;
-	struct list_head queued_ios;
 
 	struct work_struct trigger_event;
 
@@ -124,6 +122,7 @@ static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 static void process_queued_ios(struct work_struct *work);
 static void trigger_event(struct work_struct *work);
 static void activate_path(struct work_struct *work);
+static int __pgpath_busy(struct pgpath *pgpath);
 
 
 /*-----------------------------------------------
@@ -195,7 +194,6 @@ static struct multipath *alloc_multipath(struct dm_target *ti)
 	m = kzalloc(sizeof(*m), GFP_KERNEL);
 	if (m) {
 		INIT_LIST_HEAD(&m->priority_groups);
-		INIT_LIST_HEAD(&m->queued_ios);
 		spin_lock_init(&m->lock);
 		m->queue_io = 1;
 		m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
@@ -368,12 +366,15 @@ failed:
  */
 static int __must_push_back(struct multipath *m)
 {
-	return (m->queue_if_no_path != m->saved_queue_if_no_path &&
-		dm_noflush_suspending(m->ti));
+	return (m->queue_if_no_path ||
+		(m->queue_if_no_path != m->saved_queue_if_no_path &&
+		 dm_noflush_suspending(m->ti)));
 }
 
+#define pg_ready(m) (!(m)->queue_io && !(m)->pg_init_required)
+
 static int map_io(struct multipath *m, struct request *clone,
-		  union map_info *map_context, unsigned was_queued)
+		  union map_info *map_context)
 {
 	int r = DM_MAPIO_REMAPPED;
 	size_t nr_bytes = blk_rq_bytes(clone);
@@ -391,37 +392,30 @@ static int map_io(struct multipath *m, struct request *clone,
 
 	pgpath = m->current_pgpath;
 
-	if (was_queued)
-		m->queue_size--;
-
-	if (m->pg_init_required) {
-		if (!m->pg_init_in_progress)
-			queue_work(kmultipathd, &m->process_queued_ios);
-		r = DM_MAPIO_REQUEUE;
-	} else if ((pgpath && m->queue_io) ||
-		   (!pgpath && m->queue_if_no_path)) {
-		/* Queue for the daemon to resubmit */
-		list_add_tail(&clone->queuelist, &m->queued_ios);
-		m->queue_size++;
-		if (!m->queue_io)
-			queue_work(kmultipathd, &m->process_queued_ios);
-		pgpath = NULL;
-		r = DM_MAPIO_SUBMITTED;
-	} else if (pgpath) {
-		bdev = pgpath->path.dev->bdev;
-		clone->q = bdev_get_queue(bdev);
-		clone->rq_disk = bdev->bd_disk;
-	} else if (__must_push_back(m))
-		r = DM_MAPIO_REQUEUE;
-	else
-		r = -EIO;	/* Failed */
-
-	mpio->pgpath = pgpath;
-	mpio->nr_bytes = nr_bytes;
-
-	if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
-		pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
-					      nr_bytes);
+	if (pgpath) {
+		if (__pgpath_busy(pgpath))
+			r = DM_MAPIO_REQUEUE;
+		else if (pg_ready(m)) {
+			bdev = pgpath->path.dev->bdev;
+			clone->q = bdev_get_queue(bdev);
+			clone->rq_disk = bdev->bd_disk;
+			mpio->pgpath = pgpath;
+			mpio->nr_bytes = nr_bytes;
+			if (pgpath->pg->ps.type->start_io)
+				pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
+							      &pgpath->path,
+							      nr_bytes);
+		} else {
+			__pg_init_all_paths(m);
+			r = DM_MAPIO_REQUEUE;
+		}
+	} else {
+		/* No path */
+		if (__must_push_back(m))
+			r = DM_MAPIO_REQUEUE;
+		else
+			r = -EIO;	/* Failed */
+	}
 
 	spin_unlock_irqrestore(&m->lock, flags);
 
@@ -443,7 +437,7 @@ static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 	else
 		m->saved_queue_if_no_path = queue_if_no_path;
 	m->queue_if_no_path = queue_if_no_path;
-	if (!m->queue_if_no_path && m->queue_size)
+	if (!m->queue_if_no_path)
 		queue_work(kmultipathd, &m->process_queued_ios);
 
 	spin_unlock_irqrestore(&m->lock, flags);
@@ -451,40 +445,6 @@ static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 	return 0;
 }
 
-/*-----------------------------------------------------------------
- * The multipath daemon is responsible for resubmitting queued ios.
- *---------------------------------------------------------------*/
-
-static void dispatch_queued_ios(struct multipath *m)
-{
-	int r;
-	unsigned long flags;
-	union map_info *info;
-	struct request *clone, *n;
-	LIST_HEAD(cl);
-
-	spin_lock_irqsave(&m->lock, flags);
-	list_splice_init(&m->queued_ios, &cl);
-	spin_unlock_irqrestore(&m->lock, flags);
-
-	list_for_each_entry_safe(clone, n, &cl, queuelist) {
-		list_del_init(&clone->queuelist);
-
-		info = dm_get_rq_mapinfo(clone);
-
-		r = map_io(m, clone, info, 1);
-		if (r < 0) {
-			clear_mapinfo(m, info);
-			dm_kill_unmapped_request(clone, r);
-		} else if (r == DM_MAPIO_REMAPPED)
-			dm_dispatch_request(clone);
-		else if (r == DM_MAPIO_REQUEUE) {
-			clear_mapinfo(m, info);
-			dm_requeue_unmapped_request(clone);
-		}
-	}
-}
-
 static void process_queued_ios(struct work_struct *work)
 {
 	struct multipath *m =
@@ -510,7 +470,7 @@ static void process_queued_ios(struct work_struct *work)
 
 	spin_unlock_irqrestore(&m->lock, flags);
 	if (!must_queue)
-		dispatch_queued_ios(m);
+		dm_table_run_queue(m->ti->table);
 }
 
 /*
@@ -988,7 +948,7 @@ static int multipath_map(struct dm_target *ti, struct request *clone,
 		return DM_MAPIO_REQUEUE;
 
 	clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
-	r = map_io(m, clone, map_context, 0);
+	r = map_io(m, clone, map_context);
 	if (r < 0 || r == DM_MAPIO_REQUEUE)
 		clear_mapinfo(m, map_context);
 
@@ -1057,7 +1017,7 @@ static int reinstate_path(struct pgpath *pgpath)
 
 	pgpath->is_active = 1;
 
-	if (!m->nr_valid_paths++ && m->queue_size) {
+	if (!m->nr_valid_paths++) {
 		m->current_pgpath = NULL;
 		queue_work(kmultipathd, &m->process_queued_ios);
 	} else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
@@ -1436,7 +1396,8 @@ static void multipath_status(struct dm_target *ti, status_type_t type,
 
 	/* Features */
 	if (type == STATUSTYPE_INFO)
-		DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
+		DMEMIT("2 %u %u ", (m->queue_io << 1) + m->queue_if_no_path,
+		       m->pg_init_count);
 	else {
 		DMEMIT("%u ", m->queue_if_no_path +
 			      (m->pg_init_retries > 0) * 2 +
@@ -1684,7 +1645,7 @@ static int multipath_busy(struct dm_target *ti)
 	spin_lock_irqsave(&m->lock, flags);
 
 	/* pg_init in progress, requeue until done */
-	if (m->pg_init_in_progress) {
+	if (!pg_ready(m)) {
 		busy = 1;
 		goto out;
 	}
diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index 0704c52..291491b 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c
@@ -1912,6 +1912,19 @@ static int dm_any_congested(void *congested_data, int bdi_bits)
 	return r;
 }
 
+void dm_table_run_queue(struct dm_table *t)
+{
+	struct mapped_device *md = dm_table_get_md(t);
+	unsigned long flags;
+
+	if (md->queue) {
+		spin_lock_irqsave(md->queue->queue_lock, flags);
+		blk_run_queue_async(md->queue);
+		spin_unlock_irqrestore(md->queue->queue_lock, flags);
+	}
+}
+EXPORT_SYMBOL_GPL(dm_table_run_queue);
+
 /*-----------------------------------------------------------------
  * An IDR is used to keep track of allocated minor numbers.
  *---------------------------------------------------------------*/
diff --git a/include/linux/device-mapper.h b/include/linux/device-mapper.h
index ed419c6..a33653f 100644
--- a/include/linux/device-mapper.h
+++ b/include/linux/device-mapper.h
@@ -604,5 +604,6 @@ void dm_dispatch_request(struct request *rq);
 void dm_requeue_unmapped_request(struct request *rq);
 void dm_kill_unmapped_request(struct request *rq, int error);
 int dm_underlying_device_busy(struct request_queue *q);
+void dm_table_run_queue(struct dm_table *t);
 
 #endif	/* _LINUX_DEVICE_MAPPER_H */
-- 
1.7.12.4

[PATCH 3/5] dm-multipath: remove process_queued_ios()

[Hannes Reinecke]

Doesn't serve any real purpose anymore; dm_table_run_queue()
will already move things to a workqueue, so we don't need
to do it ourselves.

Signed-off-by: Hannes Reinecke <hare@suse.de>
---
 drivers/md/dm-mpath.c | 67 ++++++++++++++++++++-------------------------------
 1 file changed, 26 insertions(+), 41 deletions(-)

diff --git a/drivers/md/dm-mpath.c b/drivers/md/dm-mpath.c
index 39fd166..7122ac3 100644
--- a/drivers/md/dm-mpath.c
+++ b/drivers/md/dm-mpath.c
@@ -93,8 +93,6 @@ struct multipath {
 	unsigned pg_init_count;		/* Number of times pg_init called */
 	unsigned pg_init_delay_msecs;	/* Number of msecs before pg_init retry */
 
-	struct work_struct process_queued_ios;
-
 	struct work_struct trigger_event;
 
 	/*
@@ -119,7 +117,6 @@ typedef int (*action_fn) (struct pgpath *pgpath);
 static struct kmem_cache *_mpio_cache;
 
 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
-static void process_queued_ios(struct work_struct *work);
 static void trigger_event(struct work_struct *work);
 static void activate_path(struct work_struct *work);
 static int __pgpath_busy(struct pgpath *pgpath);
@@ -197,7 +194,6 @@ static struct multipath *alloc_multipath(struct dm_target *ti)
 		spin_lock_init(&m->lock);
 		m->queue_io = 1;
 		m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
-		INIT_WORK(&m->process_queued_ios, process_queued_ios);
 		INIT_WORK(&m->trigger_event, trigger_event);
 		init_waitqueue_head(&m->pg_init_wait);
 		mutex_init(&m->work_mutex);
@@ -438,41 +434,13 @@ static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 		m->saved_queue_if_no_path = queue_if_no_path;
 	m->queue_if_no_path = queue_if_no_path;
 	if (!m->queue_if_no_path)
-		queue_work(kmultipathd, &m->process_queued_ios);
+		dm_table_run_queue(m->ti->table);
 
 	spin_unlock_irqrestore(&m->lock, flags);
 
 	return 0;
 }
 
-static void process_queued_ios(struct work_struct *work)
-{
-	struct multipath *m =
-		container_of(work, struct multipath, process_queued_ios);
-	struct pgpath *pgpath = NULL;
-	unsigned must_queue = 1;
-	unsigned long flags;
-
-	spin_lock_irqsave(&m->lock, flags);
-
-	if (!m->current_pgpath)
-		__choose_pgpath(m, 0);
-
-	pgpath = m->current_pgpath;
-
-	if ((pgpath && !m->queue_io) ||
-	    (!pgpath && !m->queue_if_no_path))
-		must_queue = 0;
-
-	if (m->pg_init_required && !m->pg_init_in_progress && pgpath &&
-	    !m->pg_init_disabled)
-		__pg_init_all_paths(m);
-
-	spin_unlock_irqrestore(&m->lock, flags);
-	if (!must_queue)
-		dm_table_run_queue(m->ti->table);
-}
-
 /*
  * An event is triggered whenever a path is taken out of use.
  * Includes path failure and PG bypass.
@@ -1019,7 +987,7 @@ static int reinstate_path(struct pgpath *pgpath)
 
 	if (!m->nr_valid_paths++) {
 		m->current_pgpath = NULL;
-		queue_work(kmultipathd, &m->process_queued_ios);
+		dm_table_run_queue(m->ti->table);
 	} else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
 		if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
 			m->pg_init_in_progress++;
@@ -1158,7 +1126,7 @@ static void pg_init_done(void *data, int errors)
 	struct priority_group *pg = pgpath->pg;
 	struct multipath *m = pg->m;
 	unsigned long flags;
-	unsigned delay_retry = 0;
+	unsigned long pg_init_delay = 0;
 
 	/* device or driver problems */
 	switch (errors) {
@@ -1185,7 +1153,7 @@ static void pg_init_done(void *data, int errors)
 		break;
 	case SCSI_DH_RETRY:
 		/* Wait before retrying. */
-		delay_retry = 1;
+		pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs);
 	case SCSI_DH_IMM_RETRY:
 	case SCSI_DH_RES_TEMP_UNAVAIL:
 		if (pg_init_limit_reached(m, pgpath))
@@ -1217,9 +1185,21 @@ static void pg_init_done(void *data, int errors)
 
 	if (!m->pg_init_required)
 		m->queue_io = 0;
-
-	m->pg_init_delay_retry = delay_retry;
-	queue_work(kmultipathd, &m->process_queued_ios);
+	else {
+		if (!pg_init_delay) {
+			m->pg_init_delay_retry = 0;
+			/*
+			 * Add a small delay to move the
+			 * activation to a workqueue
+			 */
+			pg_init_delay = HZ / 50;
+		} else
+			m->pg_init_delay_retry = 1;
+		if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
+				       pg_init_delay))
+			m->pg_init_in_progress++;
+		goto out;
+	}
 
 	/*
 	 * Wake up any thread waiting to suspend.
@@ -1593,8 +1573,13 @@ static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
 	if (!r && ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT)
 		r = scsi_verify_blk_ioctl(NULL, cmd);
 
-	if (r == -ENOTCONN && !fatal_signal_pending(current))
-		queue_work(kmultipathd, &m->process_queued_ios);
+	if (r == -ENOTCONN && !fatal_signal_pending(current)) {
+		spin_lock_irqsave(&m->lock, flags);
+		if (m->current_pgpath && m->pg_init_required)
+			__pg_init_all_paths(m);
+		spin_unlock_irqrestore(&m->lock, flags);
+		dm_table_run_queue(m->ti->table);
+	}
 
 	return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
 }
-- 
1.7.12.4

[PATCH 4/5] dm-multipath: reduce memory pressure during requeuing

[Hannes Reinecke]

When multipath needs to requeue I/O in the block layer
the per-request context shouldn't be allocated, as it will
be freed immediately afterwards anyway.
Avoiding this memory allocation will reduce memory
pressure during requeuing.

Cc: Mike Snitzer <snitzer@redhat.com>
Cc: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Hannes Reinecke <hare@suse.de>
---
 drivers/md/dm-mpath.c | 40 ++++++++++++++++------------------------
 1 file changed, 16 insertions(+), 24 deletions(-)

diff --git a/drivers/md/dm-mpath.c b/drivers/md/dm-mpath.c
index 7122ac3..7146f7b 100644
--- a/drivers/md/dm-mpath.c
+++ b/drivers/md/dm-mpath.c
@@ -372,12 +372,12 @@ static int __must_push_back(struct multipath *m)
 static int map_io(struct multipath *m, struct request *clone,
 		  union map_info *map_context)
 {
-	int r = DM_MAPIO_REMAPPED;
+	int r = DM_MAPIO_REQUEUE;
 	size_t nr_bytes = blk_rq_bytes(clone);
 	unsigned long flags;
 	struct pgpath *pgpath;
 	struct block_device *bdev;
-	struct dm_mpath_io *mpio = map_context->ptr;
+	struct dm_mpath_io *mpio;
 
 	spin_lock_irqsave(&m->lock, flags);
 
@@ -390,29 +390,31 @@ static int map_io(struct multipath *m, struct request *clone,
 
 	if (pgpath) {
 		if (__pgpath_busy(pgpath))
-			r = DM_MAPIO_REQUEUE;
-		else if (pg_ready(m)) {
+			goto out_unlock;
+
+		if (pg_ready(m)) {
+			if (set_mapinfo(m, map_context) < 0)
+				goto out_unlock;
+
 			bdev = pgpath->path.dev->bdev;
 			clone->q = bdev_get_queue(bdev);
 			clone->rq_disk = bdev->bd_disk;
+			clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
+			mpio = map_context->ptr;
 			mpio->pgpath = pgpath;
 			mpio->nr_bytes = nr_bytes;
 			if (pgpath->pg->ps.type->start_io)
 				pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 							      &pgpath->path,
 							      nr_bytes);
-		} else {
-			__pg_init_all_paths(m);
-			r = DM_MAPIO_REQUEUE;
+			r = DM_MAPIO_REMAPPED;
+			goto out_unlock;
 		}
-	} else {
-		/* No path */
-		if (__must_push_back(m))
-			r = DM_MAPIO_REQUEUE;
-		else
+		__pg_init_all_paths(m);
+	} else if (!__must_push_back(m))
 			r = -EIO;	/* Failed */
-	}
 
+out_unlock:
 	spin_unlock_irqrestore(&m->lock, flags);
 
 	return r;
@@ -908,19 +910,9 @@ static void multipath_dtr(struct dm_target *ti)
 static int multipath_map(struct dm_target *ti, struct request *clone,
 			 union map_info *map_context)
 {
-	int r;
 	struct multipath *m = (struct multipath *) ti->private;
 
-	if (set_mapinfo(m, map_context) < 0)
-		/* ENOMEM, requeue */
-		return DM_MAPIO_REQUEUE;
-
-	clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
-	r = map_io(m, clone, map_context);
-	if (r < 0 || r == DM_MAPIO_REQUEUE)
-		clear_mapinfo(m, map_context);
-
-	return r;
+	return map_io(m, clone, map_context);
 }
 
 /*
-- 
1.7.12.4

[PATCH 5/5] dm-multipath: remove map_io()

[Hannes Reinecke]

multipath_map() is now just a wrapper around map_io(), so we
can fold both into multipath_map().

Signed-off-by: Hannes Reinecke <hare@suse.de>
---
 drivers/md/dm-mpath.c | 19 ++++++-------------
 1 file changed, 6 insertions(+), 13 deletions(-)

diff --git a/drivers/md/dm-mpath.c b/drivers/md/dm-mpath.c
index 7146f7b..4974aac 100644
--- a/drivers/md/dm-mpath.c
+++ b/drivers/md/dm-mpath.c
@@ -369,9 +369,13 @@ static int __must_push_back(struct multipath *m)
 
 #define pg_ready(m) (!(m)->queue_io && !(m)->pg_init_required)
 
-static int map_io(struct multipath *m, struct request *clone,
-		  union map_info *map_context)
+/*
+ * Map cloned requests
+ */
+static int multipath_map(struct dm_target *ti, struct request *clone,
+			 union map_info *map_context)
 {
+	struct multipath *m = (struct multipath *) ti->private;
 	int r = DM_MAPIO_REQUEUE;
 	size_t nr_bytes = blk_rq_bytes(clone);
 	unsigned long flags;
@@ -905,17 +909,6 @@ static void multipath_dtr(struct dm_target *ti)
 }
 
 /*
- * Map cloned requests
- */
-static int multipath_map(struct dm_target *ti, struct request *clone,
-			 union map_info *map_context)
-{
-	struct multipath *m = (struct multipath *) ti->private;
-
-	return map_io(m, clone, map_context);
-}
-
-/*
  * Take a path out of use.
  */
 static int fail_path(struct pgpath *pgpath)
-- 
1.7.12.4

Re: [PATCH 3/5] dm-multipath: remove process_queued_ios()

[Junichi Nomura]

On 01/31/14 18:10, Hannes Reinecke wrote:
> @@ -1217,9 +1185,21 @@ static void pg_init_done(void *data, int errors)
>  
>  	if (!m->pg_init_required)
>  		m->queue_io = 0;
> -
> -	m->pg_init_delay_retry = delay_retry;
> -	queue_work(kmultipathd, &m->process_queued_ios);
> +	else {
> +		if (!pg_init_delay) {
> +			m->pg_init_delay_retry = 0;
> +			/*
> +			 * Add a small delay to move the
> +			 * activation to a workqueue
> +			 */
> +			pg_init_delay = HZ / 50;
> +		} else
> +			m->pg_init_delay_retry = 1;
> +		if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
> +				       pg_init_delay))
> +			m->pg_init_in_progress++;
> +		goto out;
> +	}

This code is called when pg_init_progress becomes 0,
that means it is possible that multiple pgpaths have requested retries.
So you have to activate_path for all non-active pgpath in this pg,
like __pg_init_all_paths() does in the current code.

Also, you aren't clearing pg_init_required here.
I think it causes extra pg_init unnecessarily.

What do you think if we just call __pg_init_all_paths() here
instead of open coding it?

-- 
Jun'ichi Nomura, NEC Corporation

Re: [PATCH 3/5] dm-multipath: remove process_queued_ios()

[Hannes Reinecke]

On 01/31/2014 10:43 AM, Junichi Nomura wrote:
> On 01/31/14 18:10, Hannes Reinecke wrote:
>> @@ -1217,9 +1185,21 @@ static void pg_init_done(void *data, int errors)
>>  
>>  	if (!m->pg_init_required)
>>  		m->queue_io = 0;
>> -
>> -	m->pg_init_delay_retry = delay_retry;
>> -	queue_work(kmultipathd, &m->process_queued_ios);
>> +	else {
>> +		if (!pg_init_delay) {
>> +			m->pg_init_delay_retry = 0;
>> +			/*
>> +			 * Add a small delay to move the
>> +			 * activation to a workqueue
>> +			 */
>> +			pg_init_delay = HZ / 50;
>> +		} else
>> +			m->pg_init_delay_retry = 1;
>> +		if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
>> +				       pg_init_delay))
>> +			m->pg_init_in_progress++;
>> +		goto out;
>> +	}
> 
> This code is called when pg_init_progress becomes 0,
> that means it is possible that multiple pgpaths have requested retries.
> So you have to activate_path for all non-active pgpath in this pg,
> like __pg_init_all_paths() does in the current code.
> 
Not necessarily. SCSI_DH_RETRY (initially) is per-path, so it's well
feasible that pg_init on the first path returns SCSI_DH_OK, and
pg_init on the other path returns SCSI_DH_RETRY.

> Also, you aren't clearing pg_init_required here.
> I think it causes extra pg_init unnecessarily.
> 
Indeed. I'll need to fix this up.

> What do you think if we just call __pg_init_all_paths() here
> instead of open coding it?
> 
That depends on the intended policy.

Problem here is that pg_init_done() is per path, so at face value
SCSI_DH_RETRY is per path, too.
So from that we should be retrying this path (and this path only).
Hence it would be correct to call queue_delayed_work directly.

However, typically any pg_init affects _every_ path in the multipath
device (active paths become passive and vice versa).
Which seems to be the intended usage, as we're checking for
pg_init_in_progress prior to invoking queue_delayed_work().

But _if_ we assume that, then we only need to send a _single_
pg_init, as this will switch all paths. So again, a call to
__pg_init_all_paths will not do the correct thing as it'll
send activations to _every_ active path.

(And, in fact, we're trying really hard in scsi_dh_rdac
and scsi_dh_alua to bunch together all the various pg_init
requests precisely for the cited reason).

So my inclination here would be to treat SCSI_DH_RETRY
as _per path_, and retry only this specific path.
IE removing the check to pg_init_in_progress and call
queue_delayed_work() directly.

IMHO this would impose the least restriction on
the internal workings of the various device handlers.

What do you think?

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      zSeries & Storage
hare@suse.de			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel

Re: [PATCH 2/5] dm-multipath: push back requests instead of queueing

[Mike Snitzer]

On Fri, Jan 31 2014 at  4:10am -0500,
Hannes Reinecke <hare@suse.de> wrote:

> There is no reason why multipath needs to queue requests
> internally for queue_if_no_path or pg_init; we should
> rather push them back onto the request queue.
> 
> And while we're at it we can simplify the conditional
> statement in map_io() to make it easier to read.
> 
> Cc: Mike Snitzer <snitzer@redhat.com>
> Cc: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
> Signed-off-by: Hannes Reinecke <hare@suse.de>
> ---
>  drivers/md/dm-mpath.c         | 115 ++++++++++++++----------------------------
>  drivers/md/dm.c               |  13 +++++
>  include/linux/device-mapper.h |   1 +
>  3 files changed, 52 insertions(+), 77 deletions(-)

...

> diff --git a/drivers/md/dm.c b/drivers/md/dm.c
> index 0704c52..291491b 100644
> --- a/drivers/md/dm.c
> +++ b/drivers/md/dm.c
> @@ -1912,6 +1912,19 @@ static int dm_any_congested(void *congested_data, int bdi_bits)
>  	return r;
>  }
>  
> +void dm_table_run_queue(struct dm_table *t)
> +{
> +	struct mapped_device *md = dm_table_get_md(t);
> +	unsigned long flags;
> +
> +	if (md->queue) {
> +		spin_lock_irqsave(md->queue->queue_lock, flags);
> +		blk_run_queue_async(md->queue);
> +		spin_unlock_irqrestore(md->queue->queue_lock, flags);
> +	}
> +}
> +EXPORT_SYMBOL_GPL(dm_table_run_queue);
> +

I think I agree with Junichi's previous point that this should live in
dm-table.c.  You don't need the mapped_device, only the associated
request_queue, so... I know I implemented something like
dm_get_md_queue() before, but obviously it never went in for whatever
reason.

Anyway, I think something like this would be ideal (I renamed to
dm_table_run_md_queue_async).  Please feel free to fold this in and
repost with v4 of your patchset (assuming you'll be fixing up patch 3
after you get Jun'ichi's feedback):

 drivers/md/dm-table.c |   14 ++++++++++++++
 drivers/md/dm.c       |    5 +++++
 drivers/md/dm.h       |    1 +
 3 files changed, 20 insertions(+), 0 deletions(-)

diff --git a/drivers/md/dm-table.c b/drivers/md/dm-table.c
index 6a7f2b8..72362ce 100644
--- a/drivers/md/dm-table.c
+++ b/drivers/md/dm-table.c
@@ -1618,6 +1618,20 @@ struct mapped_device *dm_table_get_md(struct dm_table *t)
 }
 EXPORT_SYMBOL(dm_table_get_md);
 
+void dm_table_run_md_queue_async(struct dm_table *t)
+{
+	struct mapped_device *md = dm_table_get_md(t);
+	struct request_queue *queue = dm_get_md_queue(md);
+	unsigned long flags;
+
+	if (queue) {
+		spin_lock_irqsave(queue->queue_lock, flags);
+		blk_run_queue_async(queue);
+		spin_unlock_irqrestore(queue->queue_lock, flags);
+	}
+}
+EXPORT_SYMBOL(dm_table_run_md_queue_async);
+
 static int device_discard_capable(struct dm_target *ti, struct dm_dev *dev,
 				  sector_t start, sector_t len, void *data)
 {
diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index b49c762..10cae7f 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c
@@ -475,6 +475,11 @@ sector_t dm_get_size(struct mapped_device *md)
 	return get_capacity(md->disk);
 }
 
+struct request_queue *dm_get_md_queue(struct mapped_device *md)
+{
+	return md->queue;
+}
+
 struct dm_stats *dm_get_stats(struct mapped_device *md)
 {
 	return &md->stats;
diff --git a/drivers/md/dm.h b/drivers/md/dm.h
index c4569f0..1b2ca8a 100644
--- a/drivers/md/dm.h
+++ b/drivers/md/dm.h
@@ -189,6 +189,7 @@ int dm_lock_for_deletion(struct mapped_device *md, bool mark_deferred, bool only
 int dm_cancel_deferred_remove(struct mapped_device *md);
 int dm_request_based(struct mapped_device *md);
 sector_t dm_get_size(struct mapped_device *md);
+struct request_queue *dm_get_md_queue(struct mapped_device *md);
 struct dm_stats *dm_get_stats(struct mapped_device *md);
 
 int dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,

Re: [PATCH 2/5] dm-multipath: push back requests instead of queueing

[Hannes Reinecke]

On 01/31/2014 04:40 PM, Mike Snitzer wrote:
> On Fri, Jan 31 2014 at  4:10am -0500,
> Hannes Reinecke <hare@suse.de> wrote:
>
>> There is no reason why multipath needs to queue requests
>> internally for queue_if_no_path or pg_init; we should
>> rather push them back onto the request queue.
>>
>> And while we're at it we can simplify the conditional
>> statement in map_io() to make it easier to read.
>>
>> Cc: Mike Snitzer <snitzer@redhat.com>
>> Cc: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
>> Signed-off-by: Hannes Reinecke <hare@suse.de>
>> ---
>>   drivers/md/dm-mpath.c         | 115 ++++++++++++++----------------------------
>>   drivers/md/dm.c               |  13 +++++
>>   include/linux/device-mapper.h |   1 +
>>   3 files changed, 52 insertions(+), 77 deletions(-)
>
> ...
>
>> diff --git a/drivers/md/dm.c b/drivers/md/dm.c
>> index 0704c52..291491b 100644
>> --- a/drivers/md/dm.c
>> +++ b/drivers/md/dm.c
>> @@ -1912,6 +1912,19 @@ static int dm_any_congested(void *congested_data, int bdi_bits)
>>   	return r;
>>   }
>>
>> +void dm_table_run_queue(struct dm_table *t)
>> +{
>> +	struct mapped_device *md = dm_table_get_md(t);
>> +	unsigned long flags;
>> +
>> +	if (md->queue) {
>> +		spin_lock_irqsave(md->queue->queue_lock, flags);
>> +		blk_run_queue_async(md->queue);
>> +		spin_unlock_irqrestore(md->queue->queue_lock, flags);
>> +	}
>> +}
>> +EXPORT_SYMBOL_GPL(dm_table_run_queue);
>> +
>
> I think I agree with Junichi's previous point that this should live in
> dm-table.c.  You don't need the mapped_device, only the associated
> request_queue, so... I know I implemented something like
> dm_get_md_queue() before, but obviously it never went in for whatever
> reason.
>
> Anyway, I think something like this would be ideal (I renamed to
> dm_table_run_md_queue_async).  Please feel free to fold this in and
> repost with v4 of your patchset (assuming you'll be fixing up patch 3
> after you get Jun'ichi's feedback):
>
As mentioned, I don't mind.
The layering conventions in the device-mapper code still confuses
me occasionally.

I'll be folding it in.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      zSeries & Storage
hare@suse.de			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)

Re: [PATCH 3/5] dm-multipath: remove process_queued_ios()

[Junichi Nomura]

On 01/31/14 23:55, Hannes Reinecke wrote:
> Problem here is that pg_init_done() is per path, so at face value
> SCSI_DH_RETRY is per path, too.
> So from that we should be retrying this path (and this path only).
> Hence it would be correct to call queue_delayed_work directly.
> 
> However, typically any pg_init affects _every_ path in the multipath
> device (active paths become passive and vice versa).
> Which seems to be the intended usage, as we're checking for
> pg_init_in_progress prior to invoking queue_delayed_work().
> 
> But _if_ we assume that, then we only need to send a _single_
> pg_init, as this will switch all paths. So again, a call to
> __pg_init_all_paths will not do the correct thing as it'll
> send activations to _every_ active path.

Sending activation for every paths was introduced by this:
  commit e54f77ddda72781ec1c1696b21aabd6a30cbb7c6
  Author: Chandra Seetharaman <sekharan@us.ibm.com>
  Date:   Mon Jun 22 10:12:12 2009 +0100
So if you make change in the logic, you have to check whether the
change does not break what the above solved.

> (And, in fact, we're trying really hard in scsi_dh_rdac
> and scsi_dh_alua to bunch together all the various pg_init
> requests precisely for the cited reason).
> 
> So my inclination here would be to treat SCSI_DH_RETRY
> as _per path_, and retry only this specific path.
> IE removing the check to pg_init_in_progress and call
> queue_delayed_work() directly.
> 
> IMHO this would impose the least restriction on
> the internal workings of the various device handlers.
> 
> What do you think?

I don't have strong opinion either way. But if we do that, more code
has to be changed, e.g. the management of retry count.

Removing the unnecessary workqueue has already a benefit.
It would be nice to focus on that instead of folding in more changes.

-- 
Jun'ichi Nomura, NEC Corporation

Re: [PATCH 3/5] dm-multipath: remove process_queued_ios()

[Hannes Reinecke]

On 02/03/2014 11:26 AM, Junichi Nomura wrote:
> On 01/31/14 23:55, Hannes Reinecke wrote:
>> Problem here is that pg_init_done() is per path, so at face value
>> SCSI_DH_RETRY is per path, too.
>> So from that we should be retrying this path (and this path only).
>> Hence it would be correct to call queue_delayed_work directly.
>>
>> However, typically any pg_init affects _every_ path in the multipath
>> device (active paths become passive and vice versa).
>> Which seems to be the intended usage, as we're checking for
>> pg_init_in_progress prior to invoking queue_delayed_work().
>>
>> But _if_ we assume that, then we only need to send a _single_
>> pg_init, as this will switch all paths. So again, a call to
>> __pg_init_all_paths will not do the correct thing as it'll
>> send activations to _every_ active path.
> 
> Sending activation for every paths was introduced by this:
>   commit e54f77ddda72781ec1c1696b21aabd6a30cbb7c6
>   Author: Chandra Seetharaman <sekharan@us.ibm.com>
>   Date:   Mon Jun 22 10:12:12 2009 +0100
> So if you make change in the logic, you have to check whether the
> change does not break what the above solved.
> 
>> (And, in fact, we're trying really hard in scsi_dh_rdac
>> and scsi_dh_alua to bunch together all the various pg_init
>> requests precisely for the cited reason).
>>
>> So my inclination here would be to treat SCSI_DH_RETRY
>> as _per path_, and retry only this specific path.
>> IE removing the check to pg_init_in_progress and call
>> queue_delayed_work() directly.
>>
>> IMHO this would impose the least restriction on
>> the internal workings of the various device handlers.
>>
>> What do you think?
> 
> I don't have strong opinion either way. But if we do that, more code
> has to be changed, e.g. the management of retry count.
> 
> Removing the unnecessary workqueue has already a benefit.
> It would be nice to focus on that instead of folding in more changes.
> 
Yes, that's what I've figured, too.
So I've just included the changes you suggested without modifying
the current logic.

I've already sent a new patchset, please check if the changes there
are correct.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      zSeries & Storage
hare@suse.de			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)

--
dm-devel mailing list
dm-devel@redhat.com
https://www.redhat.com/mailman/listinfo/dm-devel