Date: Mon, 2 Oct 2006 22:17:32 +0100
From: Alasdair G Kergon <agk@redhat.com>
Subject: [RHEL5 PATCH] dm mpath: fix adding new path while I/O is pending

https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=169302                                                                          

If you lose all the paths to a dm multipath device you can set
queue_if_no_path so that I/O will get queued.  If one of your paths comes
back, userspace will re-enabled it and the queued I/O will be released.

In the current implementation if paths need to be reconfigured while I/O
is queued like this any pending I/O must be cleared (with EIO) first:
the application receives I/O errors.

This patch from NEC allows the multipath tools to set DM_NOFLUSH_FLAG during a
suspend operation to request that I/O be queued during such a path
reconfiguration.  Targets such as multipath then transfer ('push back') their
queue of waiting I/O onto the existing queue held in the device-mapper core
during a 'suspend' operation.

I shall be submitting this upstream via -mm shortly.

Index: linux-2.6.18.noarch/drivers/md/dm.c
===================================================================
--- linux-2.6.18.noarch.orig/drivers/md/dm.c	2006-10-02 21:18:07.000000000 +0100
+++ linux-2.6.18.noarch/drivers/md/dm.c	2006-10-02 21:23:51.000000000 +0100
@@ -68,10 +68,12 @@ union map_info *dm_get_mapinfo(struct bi
 #define DMF_FROZEN 2
 #define DMF_FREEING 3
 #define DMF_DELETING 4
+#define DMF_NOFLUSH_SUSPENDING 5
 
 struct mapped_device {
 	struct rw_semaphore io_lock;
 	struct semaphore suspend_lock;
+	spinlock_t pushback_lock;
 	rwlock_t map_lock;
 	atomic_t holders;
 	atomic_t open_count;
@@ -90,6 +92,7 @@ struct mapped_device {
 	atomic_t pending;
 	wait_queue_head_t wait;
  	struct bio_list deferred;
+	struct bio_list pushback;
 
 	/*
 	 * The current mapping.
@@ -444,23 +447,50 @@ int dm_set_geometry(struct mapped_device
  *   you this clearly demarcated crap.
  *---------------------------------------------------------------*/
 
+static int __noflush_suspending(struct mapped_device *md)
+{
+	return test_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
+}
+
 /*
  * Decrements the number of outstanding ios that a bio has been
  * cloned into, completing the original io if necc.
  */
 static void dec_pending(struct dm_io *io, int error)
 {
-	if (error)
+	unsigned long flags;
+
+	/* Push-back supersedes any I/O errors */
+	if (error && !(io->error > 0 && __noflush_suspending(io->md)))
 		io->error = error;
 
 	if (atomic_dec_and_test(&io->io_count)) {
+		if (io->error == DM_ENDIO_REQUEUE) {
+			/*
+			 * Target requested pushing back the I/O.
+			 * This must be handled before the sleeper on
+			 * suspend queue merges the pushback list.
+			 */
+			spin_lock_irqsave(&io->md->pushback_lock, flags);
+			if (__noflush_suspending(io->md))
+				bio_list_add(&io->md->pushback, io->bio);
+			else
+				/* noflush suspend was interrupted. */
+				io->error = -EIO;
+			spin_unlock_irqrestore(&io->md->pushback_lock, flags);
+		}
+
 		if (end_io_acct(io))
 			/* nudge anyone waiting on suspend queue */
 			wake_up(&io->md->wait);
 
-		blk_add_trace_bio(io->md->queue, io->bio, BLK_TA_COMPLETE);
+		if (io->error != DM_ENDIO_REQUEUE) {
+			blk_add_trace_bio(io->md->queue, io->bio,
+					  BLK_TA_COMPLETE);
+
+			bio_endio(io->bio, io->bio->bi_size, io->error);
+		}
 
-		bio_endio(io->bio, io->bio->bi_size, io->error);
 		free_io(io->md, io);
 	}
 }
@@ -480,12 +510,19 @@ static int clone_endio(struct bio *bio, 
 
 	if (endio) {
 		r = endio(tio->ti, bio, error, &tio->info);
-		if (r < 0)
+		if (r < 0 || r == DM_ENDIO_REQUEUE)
+			/*
+			 * error and requeue request are handled
+			 * in dec_pending().
+			 */
 			error = r;
-
-		else if (r > 0)
-			/* the target wants another shot at the io */
+		else if (r == DM_ENDIO_INCOMPLETE)
+			/* The target will handle the io */
 			return 1;
+		else if (r) {
+			DMWARN("unimplemented target endio return value: %d", r);
+			BUG();
+		}
 	}
 
 	dec_pending(tio->io, error);
@@ -543,7 +580,7 @@ static void __map_bio(struct dm_target *
 	atomic_inc(&tio->io->io_count);
 	sector = clone->bi_sector;
 	r = ti->type->map(ti, clone, &tio->info);
-	if (r > 0) {
+	if (r == DM_MAPIO_REMAPPED) {
 		/* the bio has been remapped so dispatch it */
 
 		blk_add_trace_remap(bdev_get_queue(clone->bi_bdev), clone,
@@ -551,10 +588,8 @@ static void __map_bio(struct dm_target *
 				    clone->bi_sector);
 
 		generic_make_request(clone);
-	}
-
-	else if (r < 0) {
-		/* error the io and bail out */
+	} else if (r < 0 || r == DM_MAPIO_REQUEUE) {
+		/* error the io and bail out, or requeue it if needed */
 		md = tio->io->md;
 		dec_pending(tio->io, r);
 		/*
@@ -563,6 +598,9 @@ static void __map_bio(struct dm_target *
 		clone->bi_private = md->bs;
 		bio_put(clone);
 		free_tio(md, tio);
+	} else if (r) {
+		DMWARN("unimplemented target map return value: %d", r);
+		BUG();
 	}
 }
 
@@ -948,6 +986,7 @@ static struct mapped_device *alloc_dev(i
 	memset(md, 0, sizeof(*md));
 	init_rwsem(&md->io_lock);
 	init_MUTEX(&md->suspend_lock);
+	spin_lock_init(&md->pushback_lock);
 	rwlock_init(&md->map_lock);
 	atomic_set(&md->holders, 1);
 	atomic_set(&md->open_count, 0);
@@ -1275,12 +1314,15 @@ static void unlock_fs(struct mapped_devi
  * dm_bind_table, dm_suspend must be called to flush any in
  * flight bios and ensure that any further io gets deferred.
  */
-int dm_suspend(struct mapped_device *md, int do_lockfs)
+int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
 {
 	struct dm_table *map = NULL;
+	unsigned long flags;
 	DECLARE_WAITQUEUE(wait, current);
 	struct bio *def;
 	int r = -EINVAL;
+	int do_lockfs = suspend_flags & DM_SUSPEND_LOCKFS_FLAG ? 1 : 0;
+	int noflush = suspend_flags & DM_SUSPEND_NOFLUSH_FLAG ? 1 : 0;
 
 	down(&md->suspend_lock);
 
@@ -1289,6 +1331,13 @@ int dm_suspend(struct mapped_device *md,
 
 	map = dm_get_table(md);
 
+	/*
+	 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
+	 * This flag is cleared before dm_suspend returns.
+	 */
+	if (noflush)
+		set_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
+
 	/* This does not get reverted if there's an error later. */
 	dm_table_presuspend_targets(map);
 
@@ -1296,11 +1345,14 @@ int dm_suspend(struct mapped_device *md,
 	if (!md->suspended_bdev) {
 		DMWARN("bdget failed in dm_suspend");
 		r = -ENOMEM;
-		goto out;
+		goto flush_and_out;
 	}
 
-	/* Flush I/O to the device. */
-	if (do_lockfs) {
+	/*
+	 * Flush I/O to the device.
+	 * noflush supersedes do_lockfs, because lock_fs() needs to flush I/Os.
+	 */
+	if (do_lockfs && !noflush) {
 		r = lock_fs(md);
 		if (r)
 			goto out;
@@ -1336,6 +1388,14 @@ int dm_suspend(struct mapped_device *md,
 	down_write(&md->io_lock);
 	remove_wait_queue(&md->wait, &wait);
 
+	if (noflush) {
+		spin_lock_irqsave(&md->pushback_lock, flags);
+		clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
+		bio_list_merge_head(&md->deferred, &md->pushback);
+		bio_list_init(&md->pushback);
+		spin_unlock_irqrestore(&md->pushback_lock, flags);
+	}
+
 	/* were we interrupted ? */
 	r = -EINTR;
 	if (atomic_read(&md->pending)) {
@@ -1344,7 +1404,7 @@ int dm_suspend(struct mapped_device *md,
 		__flush_deferred_io(md, def);
 		up_write(&md->io_lock);
 		unlock_fs(md);
-		goto out;
+		goto out; /* pushback list is already flushed, so skip flush */
 	}
 	up_write(&md->io_lock);
 
@@ -1354,6 +1414,25 @@ int dm_suspend(struct mapped_device *md,
 
 	r = 0;
 
+flush_and_out:
+	if (r && noflush) {
+		/*
+		 * Because there may be already I/Os in the pushback list,
+		 * flush them before return.
+		 */
+		down_write(&md->io_lock);
+
+		spin_lock_irqsave(&md->pushback_lock, flags);
+		clear_bit(DMF_NOFLUSH_SUSPENDING, &md->flags);
+		bio_list_merge_head(&md->deferred, &md->pushback);
+		bio_list_init(&md->pushback);
+		spin_unlock_irqrestore(&md->pushback_lock, flags);
+
+		def = bio_list_get(&md->deferred);
+		__flush_deferred_io(md, def);
+		up_write(&md->io_lock);
+	}
+
 out:
 	if (r && md->suspended_bdev) {
 		bdput(md->suspended_bdev);
@@ -1438,6 +1517,17 @@ int dm_suspended(struct mapped_device *m
 	return test_bit(DMF_SUSPENDED, &md->flags);
 }
 
+int dm_noflush_suspending(struct dm_target *ti)
+{
+	struct mapped_device *md = dm_table_get_md(ti->table);
+	int r = __noflush_suspending(md);
+
+	dm_put(md);
+
+	return r;
+}
+EXPORT_SYMBOL_GPL(dm_noflush_suspending);
+
 static struct block_device_operations dm_blk_dops = {
 	.open = dm_blk_open,
 	.release = dm_blk_close,
Index: linux-2.6.18.noarch/drivers/md/dm-ioctl.c
===================================================================
--- linux-2.6.18.noarch.orig/drivers/md/dm-ioctl.c	2006-10-02 21:18:07.000000000 +0100
+++ linux-2.6.18.noarch/drivers/md/dm-ioctl.c	2006-10-02 21:20:03.000000000 +0100
@@ -760,7 +760,7 @@ out:
 static int do_suspend(struct dm_ioctl *param)
 {
 	int r = 0;
-	int do_lockfs = 1;
+	unsigned suspend_flags = DM_SUSPEND_LOCKFS_FLAG;
 	struct mapped_device *md;
 
 	md = find_device(param);
@@ -768,10 +768,12 @@ static int do_suspend(struct dm_ioctl *p
 		return -ENXIO;
 
 	if (param->flags & DM_SKIP_LOCKFS_FLAG)
-		do_lockfs = 0;
+		suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
+	if (param->flags & DM_NOFLUSH_FLAG)
+		suspend_flags |= DM_SUSPEND_NOFLUSH_FLAG;
 
 	if (!dm_suspended(md))
-		r = dm_suspend(md, do_lockfs);
+		r = dm_suspend(md, suspend_flags);
 
 	if (!r)
 		r = __dev_status(md, param);
@@ -783,7 +785,7 @@ static int do_suspend(struct dm_ioctl *p
 static int do_resume(struct dm_ioctl *param)
 {
 	int r = 0;
-	int do_lockfs = 1;
+	unsigned suspend_flags = DM_SUSPEND_LOCKFS_FLAG;
 	struct hash_cell *hc;
 	struct mapped_device *md;
 	struct dm_table *new_map;
@@ -809,9 +811,11 @@ static int do_resume(struct dm_ioctl *pa
 	if (new_map) {
 		/* Suspend if it isn't already suspended */
 		if (param->flags & DM_SKIP_LOCKFS_FLAG)
-			do_lockfs = 0;
+			suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
+		if (param->flags & DM_NOFLUSH_FLAG)
+			suspend_flags |= DM_SUSPEND_NOFLUSH_FLAG;
 		if (!dm_suspended(md))
-			dm_suspend(md, do_lockfs);
+			dm_suspend(md, suspend_flags);
 
 		r = dm_swap_table(md, new_map);
 		if (r) {
Index: linux-2.6.18.noarch/include/linux/device-mapper.h
===================================================================
--- linux-2.6.18.noarch.orig/include/linux/device-mapper.h	2006-10-02 21:20:01.000000000 +0100
+++ linux-2.6.18.noarch/include/linux/device-mapper.h	2006-10-02 21:20:03.000000000 +0100
@@ -39,7 +39,8 @@ typedef void (*dm_dtr_fn) (struct dm_tar
  * The map function must return:
  * < 0: error
  * = 0: The target will handle the io by resubmitting it later
- * > 0: simple remap complete
+ * = 1: simple remap complete
+ * = 2: The target wants to push back the io
  */
 typedef int (*dm_map_fn) (struct dm_target *ti, struct bio *bio,
 			  union map_info *map_context);
@@ -50,6 +51,7 @@ typedef int (*dm_map_fn) (struct dm_targ
  * 0   : ended successfully
  * 1   : for some reason the io has still not completed (eg,
  *       multipath target might want to requeue a failed io).
+ * 2   : The target wants to push back the io
  */
 typedef int (*dm_endio_fn) (struct dm_target *ti,
 			    struct bio *bio, int error,
@@ -174,7 +176,7 @@ void *dm_get_mdptr(struct mapped_device 
 /*
  * A device can still be used while suspended, but I/O is deferred.
  */
-int dm_suspend(struct mapped_device *md, int with_lockfs);
+int dm_suspend(struct mapped_device *md, unsigned suspend_flags);
 int dm_resume(struct mapped_device *md);
 
 /*
@@ -189,6 +191,7 @@ int dm_wait_event(struct mapped_device *
 const char *dm_device_name(struct mapped_device *md);
 struct gendisk *dm_disk(struct mapped_device *md);
 int dm_suspended(struct mapped_device *md);
+int dm_noflush_suspending(struct dm_target *ti);
 
 /*
  * Geometry functions.
Index: linux-2.6.18.noarch/drivers/md/dm.h
===================================================================
--- linux-2.6.18.noarch.orig/drivers/md/dm.h	2006-10-02 21:18:07.000000000 +0100
+++ linux-2.6.18.noarch/drivers/md/dm.h	2006-10-02 21:20:03.000000000 +0100
@@ -33,6 +33,25 @@
 #define SECTOR_SHIFT 9
 
 /*
+ * Definitions of return values from target end_io function.
+ */
+#define DM_ENDIO_INCOMPLETE	1
+#define DM_ENDIO_REQUEUE	2
+
+/*
+ * Definitions of return values from target map function.
+ */
+#define DM_MAPIO_SUBMITTED		0
+#define DM_MAPIO_REMAPPED		1
+#define DM_MAPIO_REQUEUE		DM_ENDIO_REQUEUE
+
+/*
+ * Suspend feature flags
+ */
+#define DM_SUSPEND_LOCKFS_FLAG		(1 << 0)
+#define DM_SUSPEND_NOFLUSH_FLAG		(1 << 1)
+
+/*
  * List of devices that a metadevice uses and should open/close.
  */
 struct dm_dev {
Index: linux-2.6.18.noarch/drivers/md/dm-bio-list.h
===================================================================
--- linux-2.6.18.noarch.orig/drivers/md/dm-bio-list.h	2006-10-02 21:18:07.000000000 +0100
+++ linux-2.6.18.noarch/drivers/md/dm-bio-list.h	2006-10-02 21:20:03.000000000 +0100
@@ -44,6 +44,20 @@ static inline void bio_list_merge(struct
 	bl->tail = bl2->tail;
 }
 
+static inline void bio_list_merge_head(struct bio_list *bl,
+				       struct bio_list *bl2)
+{
+	if (!bl2->head)
+		return;
+
+	if (bl->head)
+		bl2->tail->bi_next = bl->head;
+	else
+		bl->tail = bl2->tail;
+
+	bl->head = bl2->head;
+}
+
 static inline struct bio *bio_list_pop(struct bio_list *bl)
 {
 	struct bio *bio = bl->head;
Index: linux-2.6.18.noarch/include/linux/dm-ioctl.h
===================================================================
--- linux-2.6.18.noarch.orig/include/linux/dm-ioctl.h	2006-10-02 21:18:07.000000000 +0100
+++ linux-2.6.18.noarch/include/linux/dm-ioctl.h	2006-10-02 21:20:03.000000000 +0100
@@ -285,7 +285,7 @@ typedef char ioctl_struct[308];
 #define DM_DEV_SET_GEOMETRY	_IOWR(DM_IOCTL, DM_DEV_SET_GEOMETRY_CMD, struct dm_ioctl)
 
 #define DM_VERSION_MAJOR	4
-#define DM_VERSION_MINOR	9
+#define DM_VERSION_MINOR	11
 #define DM_VERSION_PATCHLEVEL	0
 #define DM_VERSION_EXTRA	"-ioctl (2006-09-14)"
 
@@ -323,4 +323,9 @@ typedef char ioctl_struct[308];
  */
 #define DM_SKIP_LOCKFS_FLAG	(1 << 10) /* In */
 
+/*
+ * Set this to suspend without flushing queued ios.
+ */
+#define DM_NOFLUSH_FLAG                (1 << 11) /* In */
+
 #endif				/* _LINUX_DM_IOCTL_H */
Index: linux-2.6.18.noarch/drivers/md/dm-mpath.c
===================================================================
--- linux-2.6.18.noarch.orig/drivers/md/dm-mpath.c	2006-10-02 21:18:07.000000000 +0100
+++ linux-2.6.18.noarch/drivers/md/dm-mpath.c	2006-10-02 21:27:55.000000000 +0100
@@ -282,10 +282,13 @@ failed:
 	m->current_pg = NULL;
 }
 
+#define __PUSHBACK(m) (!(m)->queue_if_no_path && (m)->saved_queue_if_no_path && \
+		       dm_noflush_suspending((m)->ti))
+
 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
 		  unsigned was_queued)
 {
-	int r = 1;
+	int r = DM_MAPIO_REMAPPED;
 	unsigned long flags;
 	struct pgpath *pgpath;
 
@@ -310,10 +313,13 @@ static int map_io(struct multipath *m, s
 		    !m->queue_io)
 			queue_work(kmultipathd, &m->process_queued_ios);
 		pgpath = NULL;
-		r = 0;
-	} else if (!pgpath)
+		r = DM_MAPIO_SUBMITTED;
+	} else if (!pgpath) {
 		r = -EIO;		/* Failed */
-	else
+
+		if (__PUSHBACK(m))
+			r = DM_MAPIO_REQUEUE;
+	} else
 		bio->bi_bdev = pgpath->path.dev->bdev;
 
 	mpio->pgpath = pgpath;
@@ -372,8 +378,16 @@ static void dispatch_queued_ios(struct m
 		r = map_io(m, bio, mpio, 1);
 		if (r < 0)
 			bio_endio(bio, bio->bi_size, r);
-		else if (r == 1)
+		else if (r == DM_MAPIO_REMAPPED)
 			generic_make_request(bio);
+		else if (r == DM_MAPIO_REQUEUE)
+			/*
+			 * end_io handles the requeue request by
+			 * returning the bio with error status.
+			 * We don't return the r value to end_io,
+			 * since it is probably not needed.
+			 */
+			bio_endio(bio, bio->bi_size, -EIO);
 
 		bio = next;
 	}
@@ -781,7 +795,7 @@ static int multipath_map(struct dm_targe
 	map_context->ptr = mpio;
 	bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 	r = map_io(m, bio, mpio, 0);
-	if (r < 0)
+	if (r < 0 || r == DM_MAPIO_REQUEUE)
 		mempool_free(mpio, m->mpio_pool);
 
 	return r;
@@ -1005,7 +1019,10 @@ static int do_end_io(struct multipath *m
 
 	spin_lock_irqsave(&m->lock, flags);
 	if (!m->nr_valid_paths) {
-		if (!m->queue_if_no_path) {
+		if (__PUSHBACK(m)) {
+			spin_unlock_irqrestore(&m->lock, flags);
+			return DM_ENDIO_REQUEUE;
+		} else if (!m->queue_if_no_path) {
 			spin_unlock_irqrestore(&m->lock, flags);
 			return -EIO;
 		} else {
@@ -1040,7 +1057,7 @@ static int do_end_io(struct multipath *m
 		queue_work(kmultipathd, &m->process_queued_ios);
 	spin_unlock_irqrestore(&m->lock, flags);
 
-	return 1;	/* io not complete */
+	return DM_ENDIO_INCOMPLETE;	/* io not complete */
 }
 
 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
@@ -1058,7 +1075,7 @@ static int multipath_end_io(struct dm_ta
 		if (ps->type->end_io)
 			ps->type->end_io(ps, &pgpath->path);
 	}
-	if (r <= 0)
+	if (r != DM_ENDIO_INCOMPLETE)
 		mempool_free(mpio, m->mpio_pool);
 
 	return r;