Skip to contentSkip to conversation

dm-multipath: push back requests instead of queueing

PATCH Superseded Hannes Reinecke 3 participants 16 messages Feb 4, 2014 – Feb 12, 2014 Device mapper

v7 of 9 Superseded Latest is v9 →

lore
Share & tools
lore
#0 lore raw

v7 cover dm-multipath: push back requests instead of queueing

Hi Mike,

here's now an updated patchset, which includes the comments/modifications from Jun'ichi.

I've also updated your last patch to remove the last layer of indentation; it's now far clearer to read.

And I've attached the 'Reviewed-By' tags from Jun'ichi where applicable.

Hannes Reinecke (5):
dm mpath: do not call pg_init when it is already running dm mpath: push back requests instead of queueing dm mpath: remove process_queued_ios()
dm mpath: reduce memory pressure when requeuing dm mpath: remove map_io()

Mike Snitzer (2):
dm table: add dm_table_run_md_queue_async
dm mpath: remove extra nesting in map function

 drivers/md/dm-mpath.c         | 208 ++++++++++++++----------------------------
 drivers/md/dm-table.c         |  19 ++++
 drivers/md/dm.c               |   5 +
 drivers/md/dm.h               |   1 +
 include/linux/device-mapper.h |   5 +
 5 files changed, 101 insertions(+), 137 deletions(-)
Signature · 1 line
1.7.12.4
#1 lore raw

PATCH v7 1/7 dm mpath: do not call pg_init when it is already running

This patch moves condition checks as a preparation of following patches and has no effect on behaviour. process_queued_ios() is the only caller of __pg_init_all_paths() and 2 condition checks are moved from outside to inside without side effects.

---
 drivers/md/dm-mpath.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
drivers/md/dm-mpath.c +4 -2
index 6eb9dc9..304ee5c 100644
@@ -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)
@@ -501,8 +504,7 @@ static void process_queued_ios(struct work_struct *work)
 	    (!pgpath && !m->queue_if_no_path))
 		must_queue = 0;
 
-	if (m->pg_init_required && !m->pg_init_in_progress && pgpath &&
-	    !m->pg_init_disabled)
+	if (pgpath && m->pg_init_required)
 		__pg_init_all_paths(m);
 
 	spin_unlock_irqrestore(&m->lock, flags);
Signature · 1 line
1.7.12.4
#2 lore raw

PATCH v7 2/7 dm table: add dm_table_run_md_queue_async

In reply to Hannes Reinecke

From: Mike Snitzer <[email protected]>

Introduce dm_table_run_md_queue_async() to run the request_queue of the mapped_device associated with a request-based DM table.

Also add dm_md_get_queue() wrapper to extract the request_queue from a mapped_device.

---
 drivers/md/dm-table.c         | 19 +++++++++++++++++++
 drivers/md/dm.c               |  5 +++++
 drivers/md/dm.h               |  1 +
 include/linux/device-mapper.h |  5 +++++
 4 files changed, 30 insertions(+)
drivers/md/dm-table.c +19 -0
index 6a7f2b8..8df63ed 100644
@@ -1618,6 +1618,25 @@ 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;
+	struct request_queue *queue;
+	unsigned long flags;
+
+	if (!dm_table_request_based(t))
+		return;
+
+	md = dm_table_get_md(t);
+	queue = dm_get_md_queue(md);
+	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)
 {
drivers/md/dm.c +5 -0
index 8c53b09..341991f 100644
@@ -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;
drivers/md/dm.h +1 -0
index c4569f0..1b2ca8a 100644
@@ -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,
include/linux/device-mapper.h +5 -0
index ed419c6..250225b 100644
@@ -466,6 +466,11 @@ struct mapped_device *dm_table_get_md(struct dm_table *t);
 void dm_table_event(struct dm_table *t);
 
 /*
+ * Run the queue for request-based targets.
+ */
+void dm_table_run_md_queue_async(struct dm_table *t);
+
+/*
  * The device must be suspended before calling this method.
  * Returns the previous table, which the caller must destroy.
  */
Signature · 1 line
1.7.12.4
#3 lore raw

PATCH v7 3/7 dm mpath: push back requests instead of queueing

In reply to 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.

Since mpath no longer does internal queuing of I/O the table info no longer emits the internal queue_size. Instead it displays 1 if queuing is being used or 0 if it is not.

---
 drivers/md/dm-mpath.c | 114 ++++++++++++++++----------------------------------
 1 file changed, 36 insertions(+), 78 deletions(-)
drivers/md/dm-mpath.c +36 -78
index 304ee5c..0c15847 100644
@@ -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 +392,28 @@ 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 (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 +435,7 @@ 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 +443,6 @@ 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 =
@@ -509,7 +467,7 @@ 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_md_queue_async(m->ti->table);
 }
 
 /*
@@ -987,7 +945,7 @@ 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);
 
@@ -1056,7 +1014,7 @@ 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)) {
@@ -1435,7 +1393,7 @@ 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, m->pg_init_count);
 	else {
 		DMEMIT("%u ", m->queue_if_no_path +
 			      (m->pg_init_retries > 0) * 2 +
@@ -1683,7 +1641,7 @@ 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;
 	}
@@ -1736,7 +1694,7 @@ out:
  *---------------------------------------------------------------*/
 static struct target_type multipath_target = {
 	.name = "multipath",
-	.version = {1, 6, 0},
+	.version = {1, 7, 0},
 	.module = THIS_MODULE,
 	.ctr = multipath_ctr,
 	.dtr = multipath_dtr,
Signature · 1 line
1.7.12.4
#4 lore raw

PATCH v7 4/7 dm mpath: remove process_queued_ios()

In reply to Hannes Reinecke

process_queued_ios() has served 3 functions:
1) select pg and pgpath if none is selected
2) start pg_init if requested
3) dispatch queued IOs when pg is ready

Basically, a call to queue_work(process_queued_ios) can be replaced by dm_table_run_md_queue_async(), which runs request queue and ends up calling map_io(), which does 1), 2) and 3).

Exception is when !pg_ready() (which means either pg_init is running or requested), then multipath_busy() prevents map_io() being called from request_fn.

If pg_init is running, it should be ok as long as pg_init_done() does the right thing when pg_init is completed, I.e.: restart pg_init if !pg_ready() or call dm_table_run_md_queue_async() to kick map_io().

If pg_init is requested, we have to make sure the request is detected and pg_init will be started. pg_init is requested in 3 places:
a) __choose_pgpath() in map_io()
b) __choose_pgpath() in multipath_ioctl()
c) pg_init retry in pg_init_done()
a) is ok because map_io() calls __pg_init_all_paths(), which does 2).
b) needs a call to __pg_init_all_paths(), which does 2).
c) needs a call to __pg_init_all_paths(), which does 2).

So this patch removes process_queued_ios() and ensures that __pg_init_all_paths() is called at the appropriate locations.

We only need to take care to add a small delay when calling __pg_init_all_paths() to move processing off to a workqueue; otherwise pg_init_done() might end up calling scsi_dh_activate() directly, which might use non-atomic memory allocations or issue I/O.

---
 drivers/md/dm-mpath.c | 71 +++++++++++++++++++++------------------------------
 1 file changed, 29 insertions(+), 42 deletions(-)
drivers/md/dm-mpath.c +29 -42
index 0c15847..0355adc 100644
@@ -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);
@@ -254,16 +250,21 @@ static void clear_mapinfo(struct multipath *m, union map_info *info)
  * Path selection
  *-----------------------------------------------*/
 
-static void __pg_init_all_paths(struct multipath *m)
+static int __pg_init_all_paths(struct multipath *m, unsigned long min_delay)
 {
 	struct pgpath *pgpath;
-	unsigned long pg_init_delay = 0;
+	unsigned long pg_init_delay = min_delay;
 
 	if (m->pg_init_in_progress || m->pg_init_disabled)
-		return;
+		return 0;
 
 	m->pg_init_count++;
 	m->pg_init_required = 0;
+
+	/* Check here to reset pg_init_required */
+	if (!m->current_pg)
+		return 0;
+
 	if (m->pg_init_delay_retry)
 		pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 						 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
@@ -275,6 +276,7 @@ static void __pg_init_all_paths(struct multipath *m)
 				       pg_init_delay))
 			m->pg_init_in_progress++;
 	}
+	return m->pg_init_in_progress;
 }
 
 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
@@ -404,7 +406,7 @@ static int map_io(struct multipath *m, struct request *clone,
 							      &pgpath->path,
 							      nr_bytes);
 		} else {
-			__pg_init_all_paths(m);
+			__pg_init_all_paths(m, 0);
 			r = DM_MAPIO_REQUEUE;
 		}
 	} else {
@@ -436,40 +438,13 @@ 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_md_queue_async(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 (pgpath && m->pg_init_required)
-		__pg_init_all_paths(m);
-
-	spin_unlock_irqrestore(&m->lock, flags);
-	if (!must_queue)
-		dm_table_run_md_queue_async(m->ti->table);
-}
-
 /*
  * An event is triggered whenever a path is taken out of use.
  * Includes path failure and PG bypass.
@@ -1016,7 +991,7 @@ 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_md_queue_async(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++;
@@ -1214,9 +1189,12 @@ 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 (m->current_pg) {
+		m->pg_init_delay_retry = delay_retry;
+		/* Use a small delay to force the use of workqueue context */
+		if (__pg_init_all_paths(m, HZ/100))
+			goto out;
+	}
 
 	/*
 	 * Wake up any thread waiting to suspend.
@@ -1589,8 +1567,17 @@ 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_pg) {
+			/* Path status changed, redo selection */
+			__choose_pgpath(m, 0);
+		}
+		if (m->pg_init_required)
+			__pg_init_all_paths(m, 0);
+		spin_unlock_irqrestore(&m->lock, flags);
+		dm_table_run_md_queue_async(m->ti->table);
+	}
 
 	return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
 }
Signature · 1 line
1.7.12.4
#5 lore raw

Re: [PATCH 4/7] dm mpath: remove process_queued_ios()

Quoted 4 lines · Hannes Reinecke

We only need to take care to add a small delay when calling __pg_init_all_paths() to move processing off to a workqueue; otherwise pg_init_done() might end up calling scsi_dh_activate() directly, which might use non-atomic memory allocations or issue I/O.

Hi Hannes,

could you tell me how "might end up calling scsi_dh_active()" happens?

queue_delayed_work()

    queue_delayed_work_on()
      __queue_delayed_work()
        if (!delay)
          __queue_work()
            get_work_pool()
            insert_work()
              set_work_pwq()
              get_pwq()
              if (__need_more_worker())
                wake_up_worker()

queue_work() doesn't execute the work itself.

What am I missing?

Signature · 1 line
Jun'ichi Nomura, NEC Corporation
#6 lore raw

Re: [PATCH 4/7] dm mpath: remove process_queued_ios()

Quoted 22 lines · Junichi Nomura
Quoted 4 lines · Hannes Reinecke

We only need to take care to add a small delay when calling __pg_init_all_paths() to move processing off to a workqueue; otherwise pg_init_done() might end up calling scsi_dh_activate() directly, which might use non-atomic memory allocations or issue I/O.

Hi Hannes,

could you tell me how "might end up calling scsi_dh_active()" happens?

queue_delayed_work()

    queue_delayed_work_on()
      __queue_delayed_work()
        if (!delay)
          __queue_work()
            get_work_pool()
            insert_work()
              set_work_pwq()
              get_pwq()
              if (__need_more_worker())
                wake_up_worker()
Quoted 4 lines

queue_work() doesn't execute the work itself.

What am I missing?

As mentioned, I stumbled across the same issue when developing the asynchronous SCSI aborts. I'll see to have it recreated with this patchset.

Cheers,

Hannes

Signature · 9 lines
Dr. Hannes Reinecke		      zSeries & Storage
[email protected]			      +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
[email protected]
https://www.redhat.com/mailman/listinfo/dm-devel
#7 lore raw

Re: [PATCH 4/7] dm mpath: remove process_queued_ios()

Quoted 23 lines · Hannes Reinecke
Quoted 22 lines · Junichi Nomura
Quoted 4 lines · Hannes Reinecke

We only need to take care to add a small delay when calling __pg_init_all_paths() to move processing off to a workqueue; otherwise pg_init_done() might end up calling scsi_dh_activate() directly, which might use non-atomic memory allocations or issue I/O.

Hi Hannes,

could you tell me how "might end up calling scsi_dh_active()" happens?

queue_delayed_work()

    queue_delayed_work_on()
      __queue_delayed_work()
        if (!delay)
          __queue_work()
            get_work_pool()
            insert_work()
              set_work_pwq()
              get_pwq()
              if (__need_more_worker())
                wake_up_worker()
Quoted 7 lines
Quoted 4 lines

queue_work() doesn't execute the work itself.

What am I missing?

As mentioned, I stumbled across the same issue when developing the asynchronous SCSI aborts. I'll see to have it recreated with this patchset.

Just to verify, this seems to be the only outstanding question for this patchset?

What value are you using for HZ? If this portion of the change does turn out to be meaningul: Rather than tieing to HZ should we just use an explicitly non-zero value for __pg_init_all_paths()'s @min_delay?

#8 lore raw

Re: [PATCH 4/7] dm mpath: remove process_queued_ios()

Quoted 34 lines · Mike Snitzer
Quoted 23 lines · Hannes Reinecke
Quoted 22 lines · Junichi Nomura
Quoted 4 lines · Hannes Reinecke

We only need to take care to add a small delay when calling __pg_init_all_paths() to move processing off to a workqueue; otherwise pg_init_done() might end up calling scsi_dh_activate() directly, which might use non-atomic memory allocations or issue I/O.

Hi Hannes,

could you tell me how "might end up calling scsi_dh_active()" happens?

queue_delayed_work()

    queue_delayed_work_on()
      __queue_delayed_work()
        if (!delay)
          __queue_work()
            get_work_pool()
            insert_work()
              set_work_pwq()
              get_pwq()
              if (__need_more_worker())
                wake_up_worker()
Quoted 7 lines
Quoted 4 lines

queue_work() doesn't execute the work itself.

What am I missing?

As mentioned, I stumbled across the same issue when developing the asynchronous SCSI aborts. I'll see to have it recreated with this patchset.

Quoted 7 lines

Just to verify, this seems to be the only outstanding question for this patchset?

What value are you using for HZ? If this portion of the change does turn out to be meaningul: Rather than tieing to HZ should we just use an explicitly non-zero value for __pg_init_all_paths()'s @min_delay?

The actual amount here is irrelevant, as long as it's non-zero. It's just there to force execution of the work item off the current thread.

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke		      zSeries & Storage
[email protected]			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)
#9 lore raw

Re: [PATCH 4/7] dm mpath: remove process_queued_ios()

Quoted 5 lines · Hannes Reinecke
Quoted 4 lines · Mike Snitzer

Just to verify, this seems to be the only outstanding question for this patchset?

Quoted 7 lines
Quoted 4 lines

What value are you using for HZ? If this portion of the change does turn out to be meaningul: Rather than tieing to HZ should we just use an explicitly non-zero value for __pg_init_all_paths()'s @min_delay?

The actual amount here is irrelevant, as long as it's non-zero. It's just there to force execution of the work item off the current thread.

I'm aware we just need a non-zero value. My concern, as originally raised by Junichi in an earlier reply when you had it as HZ/50, is that the value could be 0 if HZ is really small. While unlikely I see no point allowing the variable nature of HZ compromise passing a non-zero value here. Best to just be explicit by passing 1 or something.

All said, the question of why this is actually needed remains. I trust you're working on answering that via reproducer (by not forcing the use of workqueue context)?

#10 lore raw

Re: [PATCH 4/7] dm mpath: remove process_queued_ios()

Quoted 22 lines · Mike Snitzer
Quoted 5 lines · Hannes Reinecke
Quoted 4 lines · Mike Snitzer

Just to verify, this seems to be the only outstanding question for this patchset?

Quoted 7 lines
Quoted 4 lines

What value are you using for HZ? If this portion of the change does turn out to be meaningul: Rather than tieing to HZ should we just use an explicitly non-zero value for __pg_init_all_paths()'s @min_delay?

The actual amount here is irrelevant, as long as it's non-zero. It's just there to force execution of the work item off the current thread.

I'm aware we just need a non-zero value. My concern, as originally raised by Junichi in an earlier reply when you had it as HZ/50, is that the value could be 0 if HZ is really small. While unlikely I see no point allowing the variable nature of HZ compromise passing a non-zero value here. Best to just be explicit by passing 1 or something.

Quoted 4 lines

All said, the question of why this is actually needed remains. I trust you're working on answering that via reproducer (by not forcing the use of workqueue context)?

Precisely.

But as this is a bit hard to trigger it might take some time.
(you'll only be hitting this issue if you have to retry scsi_dh_activate, so you'll need to trigger this somehow).

I hope to get it done this week.
Is there any deadline which I might miss with that?

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke		      zSeries & Storage
[email protected]			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)
#11 lore raw

Re: [PATCH 4/7] dm mpath: remove process_queued_ios()

Quoted 27 lines · Hannes Reinecke
Quoted 22 lines · Mike Snitzer
Quoted 5 lines · Hannes Reinecke
Quoted 4 lines · Mike Snitzer

Just to verify, this seems to be the only outstanding question for this patchset?

Quoted 7 lines
Quoted 4 lines

What value are you using for HZ? If this portion of the change does turn out to be meaningul: Rather than tieing to HZ should we just use an explicitly non-zero value for __pg_init_all_paths()'s @min_delay?

The actual amount here is irrelevant, as long as it's non-zero. It's just there to force execution of the work item off the current thread.

I'm aware we just need a non-zero value. My concern, as originally raised by Junichi in an earlier reply when you had it as HZ/50, is that the value could be 0 if HZ is really small. While unlikely I see no point allowing the variable nature of HZ compromise passing a non-zero value here. Best to just be explicit by passing 1 or something.

Quoted 4 lines

All said, the question of why this is actually needed remains. I trust you're working on answering that via reproducer (by not forcing the use of workqueue context)?

Quoted 5 lines

Precisely.

But as this is a bit hard to trigger it might take some time. (you'll only be hitting this issue if you have to retry scsi_dh_activate, so you'll need to trigger this somehow).

OK.

Quoted 2 lines

I hope to get it done this week.
Is there any deadline which I might miss with that?

No, that'll be great. We have some time until the 3.15 merge window opens.

Thanks,
Mike

#12 lore raw

Re: [PATCH 4/7] dm mpath: remove process_queued_ios()

In reply to Mike Snitzer
Quoted 28 lines · Mike Snitzer
Quoted 25 lines · Hannes Reinecke
Quoted 24 lines · Junichi Nomura
Quoted 4 lines · Hannes Reinecke

We only need to take care to add a small delay when calling __pg_init_all_paths() to move processing off to a workqueue; otherwise pg_init_done() might end up calling scsi_dh_activate() directly, which might use non-atomic memory allocations or issue I/O.

Hi Hannes,

could you tell me how "might end up calling scsi_dh_active()" happens?

queue_delayed_work()

    queue_delayed_work_on()
      __queue_delayed_work()
        if (!delay)
          __queue_work()
            get_work_pool()
            insert_work()
              set_work_pwq()
              get_pwq()
              if (__need_more_worker())
                wake_up_worker()

queue_work() doesn't execute the work itself.

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

What am I missing?

As mentioned, I stumbled across the same issue when developing the asynchronous SCSI aborts. I'll see to have it recreated with this patchset.

Just to verify, this seems to be the only outstanding question for this patchset?

Hi Mike, Hannes,

Yes, this is the only remaining question from me.
For other parts of this patch and patchset, issues I pointed were solved.

Signature · 1 line
Jun'ichi Nomura, NEC Corporation
#13 lore raw

PATCH v7 5/7 dm mpath: reduce memory pressure when requeuing

In reply to 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.

---
 drivers/md/dm-mpath.c | 38 +++++++++++++++-----------------------
 1 file changed, 15 insertions(+), 23 deletions(-)
drivers/md/dm-mpath.c +15 -23
index 0355adc..4f14445 100644
@@ -378,12 +378,12 @@ 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);
 
@@ -396,27 +396,29 @@ static int map_io(struct multipath *m, struct request *clone,
 
 	if (pgpath) {
 		if (pg_ready(m)) {
+			if (set_mapinfo(m, map_context) < 0)
+				/* ENOMEM, requeue */
+				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, 0);
-			r = DM_MAPIO_REQUEUE;
+			r = DM_MAPIO_REMAPPED;
+			goto out_unlock;
 		}
-	} else {
-		/* No path */
-		if (__must_push_back(m))
-			r = DM_MAPIO_REQUEUE;
-		else
-			r = -EIO;	/* Failed */
-	}
+		__pg_init_all_paths(m, 0);
+	} else if (!__must_push_back(m))
+		r = -EIO;	/* Failed */
 
+out_unlock:
 	spin_unlock_irqrestore(&m->lock, flags);
 
 	return r;
@@ -912,19 +914,9 @@ 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);
 }
 
 /*
Signature · 1 line
1.7.12.4
#14 lore raw

PATCH v7 6/7 dm mpath: remove map_io()

In reply to Hannes Reinecke

multipath_map() is now just a wrapper around map_io(), so we can rename map_io() to multipath_map().

---
 drivers/md/dm-mpath.c | 19 ++++++-------------
 1 file changed, 6 insertions(+), 13 deletions(-)
drivers/md/dm-mpath.c +6 -13
index 4f14445..233c6d9 100644
@@ -375,9 +375,13 @@ 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;
@@ -909,17 +913,6 @@ 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)
Signature · 1 line
1.7.12.4
#15 lore raw

PATCH v7 7/7 dm mpath: remove extra nesting in map function

In reply to Hannes Reinecke

From: Mike Snitzer <[email protected]>

Return early for case when no path exists, and when the pathgroup isn't ready. This eliminates the need for extra nesting for the the common case.

---
 drivers/md/dm-mpath.c | 46 ++++++++++++++++++++++++----------------------
 1 file changed, 24 insertions(+), 22 deletions(-)
drivers/md/dm-mpath.c +24 -22
index 233c6d9..0f47038 100644
@@ -398,29 +398,31 @@ static int multipath_map(struct dm_target *ti, struct request *clone,
 
 	pgpath = m->current_pgpath;
 
-	if (pgpath) {
-		if (pg_ready(m)) {
-			if (set_mapinfo(m, map_context) < 0)
-				/* ENOMEM, requeue */
-				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);
-			r = DM_MAPIO_REMAPPED;
-			goto out_unlock;
-		}
+	if (!pgpath) {
+		if (!__must_push_back(m))
+			r = -EIO;	/* Failed */
+		goto out_unlock;
+	}
+	if (!pg_ready(m)) {
 		__pg_init_all_paths(m, 0);
-	} else if (!__must_push_back(m))
-		r = -EIO;	/* Failed */
+		goto out_unlock;
+	}
+	if (set_mapinfo(m, map_context) < 0)
+		/* ENOMEM, requeue */
+		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);
+	r = DM_MAPIO_REMAPPED;
 
 out_unlock:
 	spin_unlock_irqrestore(&m->lock, flags);
Signature · 1 line
1.7.12.4