atomic: use <linux/atomic.h>
[linux-2.6.git] / drivers / md / dm-mpath.c
index cf8bf05..c354701 100644 (file)
@@ -5,11 +5,10 @@
  * This file is released under the GPL.
  */
 
-#include "dm.h"
+#include <linux/device-mapper.h>
+
 #include "dm-path-selector.h"
-#include "dm-hw-handler.h"
-#include "dm-bio-list.h"
-#include "dm-bio-record.h"
+#include "dm-uevent.h"
 
 #include <linux/ctype.h>
 #include <linux/init.h>
 #include <linux/slab.h>
 #include <linux/time.h>
 #include <linux/workqueue.h>
-#include <asm/atomic.h>
+#include <scsi/scsi_dh.h>
+#include <linux/atomic.h>
 
 #define DM_MSG_PREFIX "multipath"
 #define MESG_STR(x) x, sizeof(x)
+#define DM_PG_INIT_DELAY_MSECS 2000
+#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
 
 /* Path properties */
 struct pgpath {
        struct list_head list;
 
        struct priority_group *pg;      /* Owning PG */
+       unsigned is_active;             /* Path status */
        unsigned fail_count;            /* Cumulative failure count */
 
-       struct path path;
+       struct dm_path path;
+       struct delayed_work activate_path;
 };
 
 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
@@ -60,11 +64,17 @@ struct multipath {
 
        spinlock_t lock;
 
-       struct hw_handler hw_handler;
+       const char *hw_handler_name;
+       char *hw_handler_params;
+
        unsigned nr_priority_groups;
        struct list_head priority_groups;
+
+       wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
+
        unsigned pg_init_required;      /* pg_init needs calling? */
        unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
+       unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
 
        unsigned nr_valid_paths;        /* Total number of usable paths */
        struct pgpath *current_pgpath;
@@ -75,26 +85,31 @@ struct multipath {
        unsigned queue_io;              /* Must we queue all I/O? */
        unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
        unsigned saved_queue_if_no_path;/* Saved state during suspension */
+       unsigned pg_init_retries;       /* Number of times to retry pg_init */
+       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 bio_list queued_ios;
+       struct list_head queued_ios;
        unsigned queue_size;
 
        struct work_struct trigger_event;
 
        /*
-        * We must use a mempool of mpath_io structs so that we
+        * We must use a mempool of dm_mpath_io structs so that we
         * can resubmit bios on error.
         */
        mempool_t *mpio_pool;
+
+       struct mutex work_mutex;
 };
 
 /*
  * Context information attached to each bio we process.
  */
-struct mpath_io {
+struct dm_mpath_io {
        struct pgpath *pgpath;
-       struct dm_bio_details details;
+       size_t nr_bytes;
 };
 
 typedef int (*action_fn) (struct pgpath *pgpath);
@@ -103,9 +118,10 @@ typedef int (*action_fn) (struct pgpath *pgpath);
 
 static struct kmem_cache *_mpio_cache;
 
-struct workqueue_struct *kmultipathd;
+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);
 
 
 /*-----------------------------------------------
@@ -116,13 +132,15 @@ static struct pgpath *alloc_pgpath(void)
 {
        struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 
-       if (pgpath)
-               pgpath->path.is_active = 1;
+       if (pgpath) {
+               pgpath->is_active = 1;
+               INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
+       }
 
        return pgpath;
 }
 
-static inline void free_pgpath(struct pgpath *pgpath)
+static void free_pgpath(struct pgpath *pgpath)
 {
        kfree(pgpath);
 }
@@ -142,9 +160,12 @@ static struct priority_group *alloc_priority_group(void)
 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 {
        struct pgpath *pgpath, *tmp;
+       struct multipath *m = ti->private;
 
        list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
                list_del(&pgpath->list);
+               if (m->hw_handler_name)
+                       scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
                dm_put_device(ti, pgpath->path.dev);
                free_pgpath(pgpath);
        }
@@ -171,10 +192,14 @@ 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;
                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);
                m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
                if (!m->mpio_pool) {
                        kfree(m);
@@ -190,18 +215,14 @@ static struct multipath *alloc_multipath(struct dm_target *ti)
 static void free_multipath(struct multipath *m)
 {
        struct priority_group *pg, *tmp;
-       struct hw_handler *hwh = &m->hw_handler;
 
        list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
                list_del(&pg->list);
                free_priority_group(pg, m->ti);
        }
 
-       if (hwh->type) {
-               hwh->type->destroy(hwh);
-               dm_put_hw_handler(hwh->type);
-       }
-
+       kfree(m->hw_handler_name);
+       kfree(m->hw_handler_params);
        mempool_destroy(m->mpio_pool);
        kfree(m);
 }
@@ -211,27 +232,48 @@ static void free_multipath(struct multipath *m)
  * Path selection
  *-----------------------------------------------*/
 
-static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
+static void __pg_init_all_paths(struct multipath *m)
 {
-       struct hw_handler *hwh = &m->hw_handler;
+       struct pgpath *pgpath;
+       unsigned long pg_init_delay = 0;
+
+       m->pg_init_count++;
+       m->pg_init_required = 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);
+       list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
+               /* Skip failed paths */
+               if (!pgpath->is_active)
+                       continue;
+               if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
+                                      pg_init_delay))
+                       m->pg_init_in_progress++;
+       }
+}
 
+static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
+{
        m->current_pg = pgpath->pg;
 
        /* Must we initialise the PG first, and queue I/O till it's ready? */
-       if (hwh->type && hwh->type->pg_init) {
+       if (m->hw_handler_name) {
                m->pg_init_required = 1;
                m->queue_io = 1;
        } else {
                m->pg_init_required = 0;
                m->queue_io = 0;
        }
+
+       m->pg_init_count = 0;
 }
 
-static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
+static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
+                              size_t nr_bytes)
 {
-       struct path *path;
+       struct dm_path *path;
 
-       path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
+       path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
        if (!path)
                return -ENXIO;
 
@@ -243,7 +285,7 @@ static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
        return 0;
 }
 
-static void __choose_pgpath(struct multipath *m)
+static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 {
        struct priority_group *pg;
        unsigned bypassed = 1;
@@ -255,12 +297,12 @@ static void __choose_pgpath(struct multipath *m)
        if (m->next_pg) {
                pg = m->next_pg;
                m->next_pg = NULL;
-               if (!__choose_path_in_pg(m, pg))
+               if (!__choose_path_in_pg(m, pg, nr_bytes))
                        return;
        }
 
        /* Don't change PG until it has no remaining paths */
-       if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
+       if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
                return;
 
        /*
@@ -272,7 +314,7 @@ static void __choose_pgpath(struct multipath *m)
                list_for_each_entry(pg, &m->priority_groups, list) {
                        if (pg->bypassed == bypassed)
                                continue;
-                       if (!__choose_path_in_pg(m, pg))
+                       if (!__choose_path_in_pg(m, pg, nr_bytes))
                                return;
                }
        } while (bypassed--);
@@ -282,19 +324,38 @@ failed:
        m->current_pg = NULL;
 }
 
-static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
-                 unsigned was_queued)
+/*
+ * Check whether bios must be queued in the device-mapper core rather
+ * than here in the target.
+ *
+ * m->lock must be held on entry.
+ *
+ * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
+ * same value then we are not between multipath_presuspend()
+ * and multipath_resume() calls and we have no need to check
+ * for the DMF_NOFLUSH_SUSPENDING flag.
+ */
+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));
+}
+
+static int map_io(struct multipath *m, struct request *clone,
+                 struct dm_mpath_io *mpio, unsigned was_queued)
 {
-       int r = 1;
+       int r = DM_MAPIO_REMAPPED;
+       size_t nr_bytes = blk_rq_bytes(clone);
        unsigned long flags;
        struct pgpath *pgpath;
+       struct block_device *bdev;
 
        spin_lock_irqsave(&m->lock, flags);
 
        /* Do we need to select a new pgpath? */
        if (!m->current_pgpath ||
            (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
-               __choose_pgpath(m);
+               __choose_pgpath(m, nr_bytes);
 
        pgpath = m->current_pgpath;
 
@@ -304,19 +365,28 @@ static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
        if ((pgpath && m->queue_io) ||
            (!pgpath && m->queue_if_no_path)) {
                /* Queue for the daemon to resubmit */
-               bio_list_add(&m->queued_ios, bio);
+               list_add_tail(&clone->queuelist, &m->queued_ios);
                m->queue_size++;
                if ((m->pg_init_required && !m->pg_init_in_progress) ||
                    !m->queue_io)
                        queue_work(kmultipathd, &m->process_queued_ios);
                pgpath = NULL;
-               r = 0;
-       } else if (!pgpath)
-               r = -EIO;               /* Failed */
+               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
-               bio->bi_bdev = pgpath->path.dev->bdev;
+               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);
 
        spin_unlock_irqrestore(&m->lock, flags);
 
@@ -354,28 +424,31 @@ static void dispatch_queued_ios(struct multipath *m)
 {
        int r;
        unsigned long flags;
-       struct bio *bio = NULL, *next;
-       struct mpath_io *mpio;
+       struct dm_mpath_io *mpio;
        union map_info *info;
+       struct request *clone, *n;
+       LIST_HEAD(cl);
 
        spin_lock_irqsave(&m->lock, flags);
-       bio = bio_list_get(&m->queued_ios);
+       list_splice_init(&m->queued_ios, &cl);
        spin_unlock_irqrestore(&m->lock, flags);
 
-       while (bio) {
-               next = bio->bi_next;
-               bio->bi_next = NULL;
+       list_for_each_entry_safe(clone, n, &cl, queuelist) {
+               list_del_init(&clone->queuelist);
 
-               info = dm_get_mapinfo(bio);
+               info = dm_get_rq_mapinfo(clone);
                mpio = info->ptr;
 
-               r = map_io(m, bio, mpio, 1);
-               if (r < 0)
-                       bio_endio(bio, bio->bi_size, r);
-               else if (r == 1)
-                       generic_make_request(bio);
-
-               bio = next;
+               r = map_io(m, clone, mpio, 1);
+               if (r < 0) {
+                       mempool_free(mpio, m->mpio_pool);
+                       dm_kill_unmapped_request(clone, r);
+               } else if (r == DM_MAPIO_REMAPPED)
+                       dm_dispatch_request(clone);
+               else if (r == DM_MAPIO_REQUEUE) {
+                       mempool_free(mpio, m->mpio_pool);
+                       dm_requeue_unmapped_request(clone);
+               }
        }
 }
 
@@ -383,9 +456,8 @@ static void process_queued_ios(struct work_struct *work)
 {
        struct multipath *m =
                container_of(work, struct multipath, process_queued_ios);
-       struct hw_handler *hwh = &m->hw_handler;
        struct pgpath *pgpath = NULL;
-       unsigned init_required = 0, must_queue = 1;
+       unsigned must_queue = 1;
        unsigned long flags;
 
        spin_lock_irqsave(&m->lock, flags);
@@ -394,7 +466,7 @@ static void process_queued_ios(struct work_struct *work)
                goto out;
 
        if (!m->current_pgpath)
-               __choose_pgpath(m);
+               __choose_pgpath(m, 0);
 
        pgpath = m->current_pgpath;
 
@@ -402,18 +474,11 @@ 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) {
-               m->pg_init_required = 0;
-               m->pg_init_in_progress = 1;
-               init_required = 1;
-       }
+       if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
+               __pg_init_all_paths(m);
 
 out:
        spin_unlock_irqrestore(&m->lock, flags);
-
-       if (init_required)
-               hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
-
        if (!must_queue)
                dispatch_queued_ios(m);
 }
@@ -503,8 +568,16 @@ static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
        }
 
        r = read_param(_params, shift(as), &ps_argc, &ti->error);
-       if (r)
+       if (r) {
+               dm_put_path_selector(pst);
+               return -EINVAL;
+       }
+
+       if (ps_argc > as->argc) {
+               dm_put_path_selector(pst);
+               ti->error = "not enough arguments for path selector";
                return -EINVAL;
+       }
 
        r = pst->create(&pg->ps, ps_argc, as->argv);
        if (r) {
@@ -524,24 +597,56 @@ static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 {
        int r;
        struct pgpath *p;
+       struct multipath *m = ti->private;
 
        /* we need at least a path arg */
        if (as->argc < 1) {
                ti->error = "no device given";
-               return NULL;
+               return ERR_PTR(-EINVAL);
        }
 
        p = alloc_pgpath();
        if (!p)
-               return NULL;
+               return ERR_PTR(-ENOMEM);
 
-       r = dm_get_device(ti, shift(as), ti->begin, ti->len,
-                         dm_table_get_mode(ti->table), &p->path.dev);
+       r = dm_get_device(ti, shift(as), dm_table_get_mode(ti->table),
+                         &p->path.dev);
        if (r) {
                ti->error = "error getting device";
                goto bad;
        }
 
+       if (m->hw_handler_name) {
+               struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
+
+               r = scsi_dh_attach(q, m->hw_handler_name);
+               if (r == -EBUSY) {
+                       /*
+                        * Already attached to different hw_handler,
+                        * try to reattach with correct one.
+                        */
+                       scsi_dh_detach(q);
+                       r = scsi_dh_attach(q, m->hw_handler_name);
+               }
+
+               if (r < 0) {
+                       ti->error = "error attaching hardware handler";
+                       dm_put_device(ti, p->path.dev);
+                       goto bad;
+               }
+
+               if (m->hw_handler_params) {
+                       r = scsi_dh_set_params(q, m->hw_handler_params);
+                       if (r < 0) {
+                               ti->error = "unable to set hardware "
+                                                       "handler parameters";
+                               scsi_dh_detach(q);
+                               dm_put_device(ti, p->path.dev);
+                               goto bad;
+                       }
+               }
+       }
+
        r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
        if (r) {
                dm_put_device(ti, p->path.dev);
@@ -552,7 +657,7 @@ static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 
  bad:
        free_pgpath(p);
-       return NULL;
+       return ERR_PTR(r);
 }
 
 static struct priority_group *parse_priority_group(struct arg_set *as,
@@ -570,14 +675,14 @@ static struct priority_group *parse_priority_group(struct arg_set *as,
 
        if (as->argc < 2) {
                as->argc = 0;
-               ti->error = "not enough priority group aruments";
-               return NULL;
+               ti->error = "not enough priority group arguments";
+               return ERR_PTR(-EINVAL);
        }
 
        pg = alloc_priority_group();
        if (!pg) {
                ti->error = "couldn't allocate priority group";
-               return NULL;
+               return ERR_PTR(-ENOMEM);
        }
        pg->m = m;
 
@@ -601,15 +706,20 @@ static struct priority_group *parse_priority_group(struct arg_set *as,
                struct pgpath *pgpath;
                struct arg_set path_args;
 
-               if (as->argc < nr_params)
+               if (as->argc < nr_params) {
+                       ti->error = "not enough path parameters";
+                       r = -EINVAL;
                        goto bad;
+               }
 
                path_args.argc = nr_params;
                path_args.argv = as->argv;
 
                pgpath = parse_path(&path_args, &pg->ps, ti);
-               if (!pgpath)
+               if (IS_ERR(pgpath)) {
+                       r = PTR_ERR(pgpath);
                        goto bad;
+               }
 
                pgpath->pg = pg;
                list_add_tail(&pgpath->list, &pg->pgpaths);
@@ -620,44 +730,61 @@ static struct priority_group *parse_priority_group(struct arg_set *as,
 
  bad:
        free_priority_group(pg, ti);
-       return NULL;
+       return ERR_PTR(r);
 }
 
 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 {
-       int r;
-       struct hw_handler_type *hwht;
        unsigned hw_argc;
+       int ret;
        struct dm_target *ti = m->ti;
 
        static struct param _params[] = {
                {0, 1024, "invalid number of hardware handler args"},
        };
 
-       r = read_param(_params, shift(as), &hw_argc, &ti->error);
-       if (r)
+       if (read_param(_params, shift(as), &hw_argc, &ti->error))
                return -EINVAL;
 
        if (!hw_argc)
                return 0;
 
-       hwht = dm_get_hw_handler(shift(as));
-       if (!hwht) {
-               ti->error = "unknown hardware handler type";
+       if (hw_argc > as->argc) {
+               ti->error = "not enough arguments for hardware handler";
                return -EINVAL;
        }
 
-       r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
-       if (r) {
-               dm_put_hw_handler(hwht);
-               ti->error = "hardware handler constructor failed";
-               return r;
+       m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
+       request_module("scsi_dh_%s", m->hw_handler_name);
+       if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
+               ti->error = "unknown hardware handler type";
+               ret = -EINVAL;
+               goto fail;
        }
 
-       m->hw_handler.type = hwht;
+       if (hw_argc > 1) {
+               char *p;
+               int i, j, len = 4;
+
+               for (i = 0; i <= hw_argc - 2; i++)
+                       len += strlen(as->argv[i]) + 1;
+               p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
+               if (!p) {
+                       ti->error = "memory allocation failed";
+                       ret = -ENOMEM;
+                       goto fail;
+               }
+               j = sprintf(p, "%d", hw_argc - 1);
+               for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
+                       j = sprintf(p, "%s", as->argv[i]);
+       }
        consume(as, hw_argc - 1);
 
        return 0;
+fail:
+       kfree(m->hw_handler_name);
+       m->hw_handler_name = NULL;
+       return ret;
 }
 
 static int parse_features(struct arg_set *as, struct multipath *m)
@@ -665,9 +792,12 @@ static int parse_features(struct arg_set *as, struct multipath *m)
        int r;
        unsigned argc;
        struct dm_target *ti = m->ti;
+       const char *param_name;
 
        static struct param _params[] = {
-               {0, 1, "invalid number of feature args"},
+               {0, 5, "invalid number of feature args"},
+               {1, 50, "pg_init_retries must be between 1 and 50"},
+               {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
        };
 
        r = read_param(_params, shift(as), &argc, &ti->error);
@@ -677,12 +807,36 @@ static int parse_features(struct arg_set *as, struct multipath *m)
        if (!argc)
                return 0;
 
-       if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
-               return queue_if_no_path(m, 1, 0);
-       else {
+       do {
+               param_name = shift(as);
+               argc--;
+
+               if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
+                       r = queue_if_no_path(m, 1, 0);
+                       continue;
+               }
+
+               if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
+                   (argc >= 1)) {
+                       r = read_param(_params + 1, shift(as),
+                                      &m->pg_init_retries, &ti->error);
+                       argc--;
+                       continue;
+               }
+
+               if (!strnicmp(param_name, MESG_STR("pg_init_delay_msecs")) &&
+                   (argc >= 1)) {
+                       r = read_param(_params + 2, shift(as),
+                                      &m->pg_init_delay_msecs, &ti->error);
+                       argc--;
+                       continue;
+               }
+
                ti->error = "Unrecognised multipath feature request";
-               return -EINVAL;
-       }
+               r = -EINVAL;
+       } while (argc && !r);
+
+       return r;
 }
 
 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
@@ -690,8 +844,8 @@ static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 {
        /* target parameters */
        static struct param _params[] = {
-               {1, 1024, "invalid number of priority groups"},
-               {1, 1024, "invalid initial priority group number"},
+               {0, 1024, "invalid number of priority groups"},
+               {0, 1024, "invalid initial priority group number"},
        };
 
        int r;
@@ -725,13 +879,20 @@ static int multipath_ctr(struct dm_target *ti, unsigned int argc,
        if (r)
                goto bad;
 
+       if ((!m->nr_priority_groups && next_pg_num) ||
+           (m->nr_priority_groups && !next_pg_num)) {
+               ti->error = "invalid initial priority group";
+               r = -EINVAL;
+               goto bad;
+       }
+
        /* parse the priority groups */
        while (as.argc) {
                struct priority_group *pg;
 
                pg = parse_priority_group(&as, m);
-               if (!pg) {
-                       r = -EINVAL;
+               if (IS_ERR(pg)) {
+                       r = PTR_ERR(pg);
                        goto bad;
                }
 
@@ -749,6 +910,9 @@ static int multipath_ctr(struct dm_target *ti, unsigned int argc,
                goto bad;
        }
 
+       ti->num_flush_requests = 1;
+       ti->num_discard_requests = 1;
+
        return 0;
 
  bad:
@@ -756,34 +920,66 @@ static int multipath_ctr(struct dm_target *ti, unsigned int argc,
        return r;
 }
 
-static void multipath_dtr(struct dm_target *ti)
+static void multipath_wait_for_pg_init_completion(struct multipath *m)
 {
-       struct multipath *m = (struct multipath *) ti->private;
+       DECLARE_WAITQUEUE(wait, current);
+       unsigned long flags;
+
+       add_wait_queue(&m->pg_init_wait, &wait);
+
+       while (1) {
+               set_current_state(TASK_UNINTERRUPTIBLE);
 
+               spin_lock_irqsave(&m->lock, flags);
+               if (!m->pg_init_in_progress) {
+                       spin_unlock_irqrestore(&m->lock, flags);
+                       break;
+               }
+               spin_unlock_irqrestore(&m->lock, flags);
+
+               io_schedule();
+       }
+       set_current_state(TASK_RUNNING);
+
+       remove_wait_queue(&m->pg_init_wait, &wait);
+}
+
+static void flush_multipath_work(struct multipath *m)
+{
+       flush_workqueue(kmpath_handlerd);
+       multipath_wait_for_pg_init_completion(m);
        flush_workqueue(kmultipathd);
+       flush_work_sync(&m->trigger_event);
+}
+
+static void multipath_dtr(struct dm_target *ti)
+{
+       struct multipath *m = ti->private;
+
+       flush_multipath_work(m);
        free_multipath(m);
 }
 
 /*
- * Map bios, recording original fields for later in case we have to resubmit
+ * Map cloned requests
  */
-static int multipath_map(struct dm_target *ti, struct bio *bio,
+static int multipath_map(struct dm_target *ti, struct request *clone,
                         union map_info *map_context)
 {
        int r;
-       struct mpath_io *mpio;
+       struct dm_mpath_io *mpio;
        struct multipath *m = (struct multipath *) ti->private;
 
-       if (bio_barrier(bio))
-               return -EOPNOTSUPP;
-
-       mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
-       dm_bio_record(&mpio->details, bio);
+       mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
+       if (!mpio)
+               /* ENOMEM, requeue */
+               return DM_MAPIO_REQUEUE;
+       memset(mpio, 0, sizeof(*mpio));
 
        map_context->ptr = mpio;
-       bio->bi_rw |= (1 << BIO_RW_FAILFAST);
-       r = map_io(m, bio, mpio, 0);
-       if (r < 0)
+       clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
+       r = map_io(m, clone, mpio, 0);
+       if (r < 0 || r == DM_MAPIO_REQUEUE)
                mempool_free(mpio, m->mpio_pool);
 
        return r;
@@ -799,13 +995,13 @@ static int fail_path(struct pgpath *pgpath)
 
        spin_lock_irqsave(&m->lock, flags);
 
-       if (!pgpath->path.is_active)
+       if (!pgpath->is_active)
                goto out;
 
        DMWARN("Failing path %s.", pgpath->path.dev->name);
 
        pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
-       pgpath->path.is_active = 0;
+       pgpath->is_active = 0;
        pgpath->fail_count++;
 
        m->nr_valid_paths--;
@@ -813,7 +1009,10 @@ static int fail_path(struct pgpath *pgpath)
        if (pgpath == m->current_pgpath)
                m->current_pgpath = NULL;
 
-       queue_work(kmultipathd, &m->trigger_event);
+       dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
+                     pgpath->path.dev->name, m->nr_valid_paths);
+
+       schedule_work(&m->trigger_event);
 
 out:
        spin_unlock_irqrestore(&m->lock, flags);
@@ -832,10 +1031,10 @@ static int reinstate_path(struct pgpath *pgpath)
 
        spin_lock_irqsave(&m->lock, flags);
 
-       if (pgpath->path.is_active)
+       if (pgpath->is_active)
                goto out;
 
-       if (!pgpath->pg->ps.type) {
+       if (!pgpath->pg->ps.type->reinstate_path) {
                DMWARN("Reinstate path not supported by path selector %s",
                       pgpath->pg->ps.type->name);
                r = -EINVAL;
@@ -846,13 +1045,20 @@ static int reinstate_path(struct pgpath *pgpath)
        if (r)
                goto out;
 
-       pgpath->path.is_active = 1;
+       pgpath->is_active = 1;
 
-       m->current_pgpath = NULL;
-       if (!m->nr_valid_paths++ && m->queue_size)
+       if (!m->nr_valid_paths++ && m->queue_size) {
+               m->current_pgpath = NULL;
                queue_work(kmultipathd, &m->process_queued_ios);
+       } 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++;
+       }
+
+       dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
+                     pgpath->path.dev->name, m->nr_valid_paths);
 
-       queue_work(kmultipathd, &m->trigger_event);
+       schedule_work(&m->trigger_event);
 
 out:
        spin_unlock_irqrestore(&m->lock, flags);
@@ -866,7 +1072,7 @@ out:
 static int action_dev(struct multipath *m, struct dm_dev *dev,
                      action_fn action)
 {
-       int r = 0;
+       int r = -EINVAL;
        struct pgpath *pgpath;
        struct priority_group *pg;
 
@@ -896,7 +1102,7 @@ static void bypass_pg(struct multipath *m, struct priority_group *pg,
 
        spin_unlock_irqrestore(&m->lock, flags);
 
-       queue_work(kmultipathd, &m->trigger_event);
+       schedule_work(&m->trigger_event);
 }
 
 /*
@@ -926,7 +1132,7 @@ static int switch_pg_num(struct multipath *m, const char *pgstr)
        }
        spin_unlock_irqrestore(&m->lock, flags);
 
-       queue_work(kmultipathd, &m->trigger_event);
+       schedule_work(&m->trigger_event);
        return 0;
 }
 
@@ -955,113 +1161,172 @@ static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 }
 
 /*
- * pg_init must call this when it has completed its initialisation
+ * Should we retry pg_init immediately?
  */
-void dm_pg_init_complete(struct path *path, unsigned err_flags)
+static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
+{
+       unsigned long flags;
+       int limit_reached = 0;
+
+       spin_lock_irqsave(&m->lock, flags);
+
+       if (m->pg_init_count <= m->pg_init_retries)
+               m->pg_init_required = 1;
+       else
+               limit_reached = 1;
+
+       spin_unlock_irqrestore(&m->lock, flags);
+
+       return limit_reached;
+}
+
+static void pg_init_done(void *data, int errors)
 {
-       struct pgpath *pgpath = path_to_pgpath(path);
+       struct pgpath *pgpath = data;
        struct priority_group *pg = pgpath->pg;
        struct multipath *m = pg->m;
        unsigned long flags;
+       unsigned delay_retry = 0;
 
-       /* We insist on failing the path if the PG is already bypassed. */
-       if (err_flags && pg->bypassed)
-               err_flags |= MP_FAIL_PATH;
-
-       if (err_flags & MP_FAIL_PATH)
+       /* device or driver problems */
+       switch (errors) {
+       case SCSI_DH_OK:
+               break;
+       case SCSI_DH_NOSYS:
+               if (!m->hw_handler_name) {
+                       errors = 0;
+                       break;
+               }
+               DMERR("Could not failover the device: Handler scsi_dh_%s "
+                     "Error %d.", m->hw_handler_name, errors);
+               /*
+                * Fail path for now, so we do not ping pong
+                */
                fail_path(pgpath);
-
-       if (err_flags & MP_BYPASS_PG)
+               break;
+       case SCSI_DH_DEV_TEMP_BUSY:
+               /*
+                * Probably doing something like FW upgrade on the
+                * controller so try the other pg.
+                */
                bypass_pg(m, pg, 1);
+               break;
+       case SCSI_DH_RETRY:
+               /* Wait before retrying. */
+               delay_retry = 1;
+       case SCSI_DH_IMM_RETRY:
+       case SCSI_DH_RES_TEMP_UNAVAIL:
+               if (pg_init_limit_reached(m, pgpath))
+                       fail_path(pgpath);
+               errors = 0;
+               break;
+       default:
+               /*
+                * We probably do not want to fail the path for a device
+                * error, but this is what the old dm did. In future
+                * patches we can do more advanced handling.
+                */
+               fail_path(pgpath);
+       }
 
        spin_lock_irqsave(&m->lock, flags);
-       if (err_flags) {
-               m->current_pgpath = NULL;
-               m->current_pg = NULL;
+       if (errors) {
+               if (pgpath == m->current_pgpath) {
+                       DMERR("Could not failover device. Error %d.", errors);
+                       m->current_pgpath = NULL;
+                       m->current_pg = NULL;
+               }
        } else if (!m->pg_init_required)
+               pg->bypassed = 0;
+
+       if (--m->pg_init_in_progress)
+               /* Activations of other paths are still on going */
+               goto out;
+
+       if (!m->pg_init_required)
                m->queue_io = 0;
 
-       m->pg_init_in_progress = 0;
+       m->pg_init_delay_retry = delay_retry;
        queue_work(kmultipathd, &m->process_queued_ios);
+
+       /*
+        * Wake up any thread waiting to suspend.
+        */
+       wake_up(&m->pg_init_wait);
+
+out:
        spin_unlock_irqrestore(&m->lock, flags);
 }
 
+static void activate_path(struct work_struct *work)
+{
+       struct pgpath *pgpath =
+               container_of(work, struct pgpath, activate_path.work);
+
+       scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
+                               pg_init_done, pgpath);
+}
+
 /*
  * end_io handling
  */
-static int do_end_io(struct multipath *m, struct bio *bio,
-                    int error, struct mpath_io *mpio)
+static int do_end_io(struct multipath *m, struct request *clone,
+                    int error, struct dm_mpath_io *mpio)
 {
-       struct hw_handler *hwh = &m->hw_handler;
-       unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
+       /*
+        * We don't queue any clone request inside the multipath target
+        * during end I/O handling, since those clone requests don't have
+        * bio clones.  If we queue them inside the multipath target,
+        * we need to make bio clones, that requires memory allocation.
+        * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
+        *  don't have bio clones.)
+        * Instead of queueing the clone request here, we queue the original
+        * request into dm core, which will remake a clone request and
+        * clone bios for it and resubmit it later.
+        */
+       int r = DM_ENDIO_REQUEUE;
        unsigned long flags;
 
-       if (!error)
+       if (!error && !clone->errors)
                return 0;       /* I/O complete */
 
-       if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
+       if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
                return error;
 
-       if (error == -EOPNOTSUPP)
-               return error;
+       if (mpio->pgpath)
+               fail_path(mpio->pgpath);
 
        spin_lock_irqsave(&m->lock, flags);
        if (!m->nr_valid_paths) {
                if (!m->queue_if_no_path) {
-                       spin_unlock_irqrestore(&m->lock, flags);
-                       return -EIO;
+                       if (!__must_push_back(m))
+                               r = -EIO;
                } else {
-                       spin_unlock_irqrestore(&m->lock, flags);
-                       goto requeue;
+                       if (error == -EBADE)
+                               r = error;
                }
        }
        spin_unlock_irqrestore(&m->lock, flags);
 
-       if (hwh->type && hwh->type->error)
-               err_flags = hwh->type->error(hwh, bio);
-
-       if (mpio->pgpath) {
-               if (err_flags & MP_FAIL_PATH)
-                       fail_path(mpio->pgpath);
-
-               if (err_flags & MP_BYPASS_PG)
-                       bypass_pg(m, mpio->pgpath->pg, 1);
-       }
-
-       if (err_flags & MP_ERROR_IO)
-               return -EIO;
-
-      requeue:
-       dm_bio_restore(&mpio->details, bio);
-
-       /* queue for the daemon to resubmit or fail */
-       spin_lock_irqsave(&m->lock, flags);
-       bio_list_add(&m->queued_ios, bio);
-       m->queue_size++;
-       if (!m->queue_io)
-               queue_work(kmultipathd, &m->process_queued_ios);
-       spin_unlock_irqrestore(&m->lock, flags);
-
-       return 1;       /* io not complete */
+       return r;
 }
 
-static int multipath_end_io(struct dm_target *ti, struct bio *bio,
+static int multipath_end_io(struct dm_target *ti, struct request *clone,
                            int error, union map_info *map_context)
 {
-       struct multipath *m = (struct multipath *) ti->private;
-       struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
+       struct multipath *m = ti->private;
+       struct dm_mpath_io *mpio = map_context->ptr;
        struct pgpath *pgpath = mpio->pgpath;
        struct path_selector *ps;
        int r;
 
-       r  = do_end_io(m, bio, error, mpio);
+       r  = do_end_io(m, clone, error, mpio);
        if (pgpath) {
                ps = &pgpath->pg->ps;
                if (ps->type->end_io)
-                       ps->type->end_io(ps, &pgpath->path);
+                       ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
        }
-       if (r <= 0)
-               mempool_free(mpio, m->mpio_pool);
+       mempool_free(mpio, m->mpio_pool);
 
        return r;
 }
@@ -1079,6 +1344,15 @@ static void multipath_presuspend(struct dm_target *ti)
        queue_if_no_path(m, 0, 1);
 }
 
+static void multipath_postsuspend(struct dm_target *ti)
+{
+       struct multipath *m = ti->private;
+
+       mutex_lock(&m->work_mutex);
+       flush_multipath_work(m);
+       mutex_unlock(&m->work_mutex);
+}
+
 /*
  * Restore the queue_if_no_path setting.
  */
@@ -1114,7 +1388,6 @@ static int multipath_status(struct dm_target *ti, status_type_t type,
        int sz = 0;
        unsigned long flags;
        struct multipath *m = (struct multipath *) ti->private;
-       struct hw_handler *hwh = &m->hw_handler;
        struct priority_group *pg;
        struct pgpath *p;
        unsigned pg_num;
@@ -1124,18 +1397,23 @@ static int multipath_status(struct dm_target *ti, status_type_t type,
 
        /* Features */
        if (type == STATUSTYPE_INFO)
-               DMEMIT("1 %u ", m->queue_size);
-       else if (m->queue_if_no_path)
-               DMEMIT("1 queue_if_no_path ");
-       else
-               DMEMIT("0 ");
+               DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
+       else {
+               DMEMIT("%u ", m->queue_if_no_path +
+                             (m->pg_init_retries > 0) * 2 +
+                             (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2);
+               if (m->queue_if_no_path)
+                       DMEMIT("queue_if_no_path ");
+               if (m->pg_init_retries)
+                       DMEMIT("pg_init_retries %u ", m->pg_init_retries);
+               if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
+                       DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
+       }
 
-       if (hwh->type && hwh->type->status)
-               sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
-       else if (!hwh->type || type == STATUSTYPE_INFO)
+       if (!m->hw_handler_name || type == STATUSTYPE_INFO)
                DMEMIT("0 ");
        else
-               DMEMIT("1 %s ", hwh->type->name);
+               DMEMIT("1 %s ", m->hw_handler_name);
 
        DMEMIT("%u ", m->nr_priority_groups);
 
@@ -1144,7 +1422,7 @@ static int multipath_status(struct dm_target *ti, status_type_t type,
        else if (m->current_pg)
                pg_num = m->current_pg->pg_num;
        else
-                       pg_num = 1;
+               pg_num = (m->nr_priority_groups ? 1 : 0);
 
        DMEMIT("%u ", pg_num);
 
@@ -1172,7 +1450,7 @@ static int multipath_status(struct dm_target *ti, status_type_t type,
 
                        list_for_each_entry(p, &pg->pgpaths, list) {
                                DMEMIT("%s %s %u ", p->path.dev->name,
-                                      p->path.is_active ? "A" : "F",
+                                      p->is_active ? "A" : "F",
                                       p->fail_count);
                                if (pg->ps.type->status)
                                        sz += pg->ps.type->status(&pg->ps,
@@ -1214,75 +1492,84 @@ static int multipath_status(struct dm_target *ti, status_type_t type,
 
 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
 {
-       int r;
+       int r = -EINVAL;
        struct dm_dev *dev;
        struct multipath *m = (struct multipath *) ti->private;
        action_fn action;
 
+       mutex_lock(&m->work_mutex);
+
+       if (dm_suspended(ti)) {
+               r = -EBUSY;
+               goto out;
+       }
+
        if (argc == 1) {
-               if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
-                       return queue_if_no_path(m, 1, 0);
-               else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
-                       return queue_if_no_path(m, 0, 0);
+               if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
+                       r = queue_if_no_path(m, 1, 0);
+                       goto out;
+               } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
+                       r = queue_if_no_path(m, 0, 0);
+                       goto out;
+               }
        }
 
-       if (argc != 2)
-               goto error;
+       if (argc != 2) {
+               DMWARN("Unrecognised multipath message received.");
+               goto out;
+       }
 
-       if (!strnicmp(argv[0], MESG_STR("disable_group")))
-               return bypass_pg_num(m, argv[1], 1);
-       else if (!strnicmp(argv[0], MESG_STR("enable_group")))
-               return bypass_pg_num(m, argv[1], 0);
-       else if (!strnicmp(argv[0], MESG_STR("switch_group")))
-               return switch_pg_num(m, argv[1]);
-       else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
+       if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
+               r = bypass_pg_num(m, argv[1], 1);
+               goto out;
+       } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
+               r = bypass_pg_num(m, argv[1], 0);
+               goto out;
+       } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
+               r = switch_pg_num(m, argv[1]);
+               goto out;
+       } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
                action = reinstate_path;
        else if (!strnicmp(argv[0], MESG_STR("fail_path")))
                action = fail_path;
-       else
-               goto error;
+       else {
+               DMWARN("Unrecognised multipath message received.");
+               goto out;
+       }
 
-       r = dm_get_device(ti, argv[1], ti->begin, ti->len,
-                         dm_table_get_mode(ti->table), &dev);
+       r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
        if (r) {
                DMWARN("message: error getting device %s",
                       argv[1]);
-               return -EINVAL;
+               goto out;
        }
 
        r = action_dev(m, dev, action);
 
        dm_put_device(ti, dev);
 
+out:
+       mutex_unlock(&m->work_mutex);
        return r;
-
-error:
-       DMWARN("Unrecognised multipath message received.");
-       return -EINVAL;
 }
 
-static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
-                          struct file *filp, unsigned int cmd,
+static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
                           unsigned long arg)
 {
        struct multipath *m = (struct multipath *) ti->private;
        struct block_device *bdev = NULL;
+       fmode_t mode = 0;
        unsigned long flags;
-       struct file fake_file = {};
-       struct dentry fake_dentry = {};
        int r = 0;
 
-       fake_file.f_dentry = &fake_dentry;
-
        spin_lock_irqsave(&m->lock, flags);
 
        if (!m->current_pgpath)
-               __choose_pgpath(m);
+               __choose_pgpath(m, 0);
 
        if (m->current_pgpath) {
                bdev = m->current_pgpath->path.dev->bdev;
-               fake_dentry.d_inode = bdev->bd_inode;
-               fake_file.f_mode = m->current_pgpath->path.dev->mode;
+               mode = m->current_pgpath->path.dev->mode;
        }
 
        if (m->queue_io)
@@ -1292,8 +1579,96 @@ static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
 
        spin_unlock_irqrestore(&m->lock, flags);
 
-       return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
-                                        bdev->bd_disk, cmd, arg);
+       return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
+}
+
+static int multipath_iterate_devices(struct dm_target *ti,
+                                    iterate_devices_callout_fn fn, void *data)
+{
+       struct multipath *m = ti->private;
+       struct priority_group *pg;
+       struct pgpath *p;
+       int ret = 0;
+
+       list_for_each_entry(pg, &m->priority_groups, list) {
+               list_for_each_entry(p, &pg->pgpaths, list) {
+                       ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
+                       if (ret)
+                               goto out;
+               }
+       }
+
+out:
+       return ret;
+}
+
+static int __pgpath_busy(struct pgpath *pgpath)
+{
+       struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
+
+       return dm_underlying_device_busy(q);
+}
+
+/*
+ * We return "busy", only when we can map I/Os but underlying devices
+ * are busy (so even if we map I/Os now, the I/Os will wait on
+ * the underlying queue).
+ * In other words, if we want to kill I/Os or queue them inside us
+ * due to map unavailability, we don't return "busy".  Otherwise,
+ * dm core won't give us the I/Os and we can't do what we want.
+ */
+static int multipath_busy(struct dm_target *ti)
+{
+       int busy = 0, has_active = 0;
+       struct multipath *m = ti->private;
+       struct priority_group *pg;
+       struct pgpath *pgpath;
+       unsigned long flags;
+
+       spin_lock_irqsave(&m->lock, flags);
+
+       /* Guess which priority_group will be used at next mapping time */
+       if (unlikely(!m->current_pgpath && m->next_pg))
+               pg = m->next_pg;
+       else if (likely(m->current_pg))
+               pg = m->current_pg;
+       else
+               /*
+                * We don't know which pg will be used at next mapping time.
+                * We don't call __choose_pgpath() here to avoid to trigger
+                * pg_init just by busy checking.
+                * So we don't know whether underlying devices we will be using
+                * at next mapping time are busy or not. Just try mapping.
+                */
+               goto out;
+
+       /*
+        * If there is one non-busy active path at least, the path selector
+        * will be able to select it. So we consider such a pg as not busy.
+        */
+       busy = 1;
+       list_for_each_entry(pgpath, &pg->pgpaths, list)
+               if (pgpath->is_active) {
+                       has_active = 1;
+
+                       if (!__pgpath_busy(pgpath)) {
+                               busy = 0;
+                               break;
+                       }
+               }
+
+       if (!has_active)
+               /*
+                * No active path in this pg, so this pg won't be used and
+                * the current_pg will be changed at next mapping time.
+                * We need to try mapping to determine it.
+                */
+               busy = 0;
+
+out:
+       spin_unlock_irqrestore(&m->lock, flags);
+
+       return busy;
 }
 
 /*-----------------------------------------------------------------
@@ -1301,17 +1676,20 @@ static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
  *---------------------------------------------------------------*/
 static struct target_type multipath_target = {
        .name = "multipath",
-       .version = {1, 0, 5},
+       .version = {1, 3, 0},
        .module = THIS_MODULE,
        .ctr = multipath_ctr,
        .dtr = multipath_dtr,
-       .map = multipath_map,
-       .end_io = multipath_end_io,
+       .map_rq = multipath_map,
+       .rq_end_io = multipath_end_io,
        .presuspend = multipath_presuspend,
+       .postsuspend = multipath_postsuspend,
        .resume = multipath_resume,
        .status = multipath_status,
        .message = multipath_message,
        .ioctl  = multipath_ioctl,
+       .iterate_devices = multipath_iterate_devices,
+       .busy = multipath_busy,
 };
 
 static int __init dm_multipath_init(void)
@@ -1319,22 +1697,36 @@ static int __init dm_multipath_init(void)
        int r;
 
        /* allocate a slab for the dm_ios */
-       _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
-                                       0, 0, NULL, NULL);
+       _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
        if (!_mpio_cache)
                return -ENOMEM;
 
        r = dm_register_target(&multipath_target);
        if (r < 0) {
-               DMERR("%s: register failed %d", multipath_target.name, r);
+               DMERR("register failed %d", r);
                kmem_cache_destroy(_mpio_cache);
                return -EINVAL;
        }
 
-       kmultipathd = create_workqueue("kmpathd");
+       kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
        if (!kmultipathd) {
-               DMERR("%s: failed to create workqueue kmpathd",
-                               multipath_target.name);
+               DMERR("failed to create workqueue kmpathd");
+               dm_unregister_target(&multipath_target);
+               kmem_cache_destroy(_mpio_cache);
+               return -ENOMEM;
+       }
+
+       /*
+        * A separate workqueue is used to handle the device handlers
+        * to avoid overloading existing workqueue. Overloading the
+        * old workqueue would also create a bottleneck in the
+        * path of the storage hardware device activation.
+        */
+       kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
+                                                 WQ_MEM_RECLAIM);
+       if (!kmpath_handlerd) {
+               DMERR("failed to create workqueue kmpath_handlerd");
+               destroy_workqueue(kmultipathd);
                dm_unregister_target(&multipath_target);
                kmem_cache_destroy(_mpio_cache);
                return -ENOMEM;
@@ -1349,19 +1741,13 @@ static int __init dm_multipath_init(void)
 
 static void __exit dm_multipath_exit(void)
 {
-       int r;
-
+       destroy_workqueue(kmpath_handlerd);
        destroy_workqueue(kmultipathd);
 
-       r = dm_unregister_target(&multipath_target);
-       if (r < 0)
-               DMERR("%s: target unregister failed %d",
-                     multipath_target.name, r);
+       dm_unregister_target(&multipath_target);
        kmem_cache_destroy(_mpio_cache);
 }
 
-EXPORT_SYMBOL_GPL(dm_pg_init_complete);
-
 module_init(dm_multipath_init);
 module_exit(dm_multipath_exit);