2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
5 * This file is released under the GPL.
8 #include <linux/device-mapper.h>
10 #include "dm-path-selector.h"
11 #include "dm-uevent.h"
13 #include <linux/ctype.h>
14 #include <linux/init.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/pagemap.h>
18 #include <linux/slab.h>
19 #include <linux/time.h>
20 #include <linux/workqueue.h>
21 #include <scsi/scsi_dh.h>
22 #include <asm/atomic.h>
24 #define DM_MSG_PREFIX "multipath"
25 #define MESG_STR(x) x, sizeof(x)
29 struct list_head list;
31 struct priority_group *pg; /* Owning PG */
32 unsigned is_active; /* Path status */
33 unsigned fail_count; /* Cumulative failure count */
36 struct work_struct deactivate_path;
37 struct work_struct activate_path;
40 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
43 * Paths are grouped into Priority Groups and numbered from 1 upwards.
44 * Each has a path selector which controls which path gets used.
46 struct priority_group {
47 struct list_head list;
49 struct multipath *m; /* Owning multipath instance */
50 struct path_selector ps;
52 unsigned pg_num; /* Reference number */
53 unsigned bypassed; /* Temporarily bypass this PG? */
55 unsigned nr_pgpaths; /* Number of paths in PG */
56 struct list_head pgpaths;
59 /* Multipath context */
61 struct list_head list;
66 const char *hw_handler_name;
67 char *hw_handler_params;
68 unsigned nr_priority_groups;
69 struct list_head priority_groups;
70 unsigned pg_init_required; /* pg_init needs calling? */
71 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
73 unsigned nr_valid_paths; /* Total number of usable paths */
74 struct pgpath *current_pgpath;
75 struct priority_group *current_pg;
76 struct priority_group *next_pg; /* Switch to this PG if set */
77 unsigned repeat_count; /* I/Os left before calling PS again */
79 unsigned queue_io; /* Must we queue all I/O? */
80 unsigned queue_if_no_path; /* Queue I/O if last path fails? */
81 unsigned saved_queue_if_no_path;/* Saved state during suspension */
82 unsigned pg_init_retries; /* Number of times to retry pg_init */
83 unsigned pg_init_count; /* Number of times pg_init called */
85 struct work_struct process_queued_ios;
86 struct list_head queued_ios;
89 struct work_struct trigger_event;
92 * We must use a mempool of dm_mpath_io structs so that we
93 * can resubmit bios on error.
97 struct mutex work_mutex;
101 * Context information attached to each bio we process.
104 struct pgpath *pgpath;
108 typedef int (*action_fn) (struct pgpath *pgpath);
110 #define MIN_IOS 256 /* Mempool size */
112 static struct kmem_cache *_mpio_cache;
114 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
115 static void process_queued_ios(struct work_struct *work);
116 static void trigger_event(struct work_struct *work);
117 static void activate_path(struct work_struct *work);
118 static void deactivate_path(struct work_struct *work);
121 /*-----------------------------------------------
122 * Allocation routines
123 *-----------------------------------------------*/
125 static struct pgpath *alloc_pgpath(void)
127 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
130 pgpath->is_active = 1;
131 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
132 INIT_WORK(&pgpath->activate_path, activate_path);
138 static void free_pgpath(struct pgpath *pgpath)
143 static void deactivate_path(struct work_struct *work)
145 struct pgpath *pgpath =
146 container_of(work, struct pgpath, deactivate_path);
148 blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
151 static struct priority_group *alloc_priority_group(void)
153 struct priority_group *pg;
155 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
158 INIT_LIST_HEAD(&pg->pgpaths);
163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
165 struct pgpath *pgpath, *tmp;
166 struct multipath *m = ti->private;
168 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
169 list_del(&pgpath->list);
170 if (m->hw_handler_name)
171 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
172 dm_put_device(ti, pgpath->path.dev);
177 static void free_priority_group(struct priority_group *pg,
178 struct dm_target *ti)
180 struct path_selector *ps = &pg->ps;
183 ps->type->destroy(ps);
184 dm_put_path_selector(ps->type);
187 free_pgpaths(&pg->pgpaths, ti);
191 static struct multipath *alloc_multipath(struct dm_target *ti)
195 m = kzalloc(sizeof(*m), GFP_KERNEL);
197 INIT_LIST_HEAD(&m->priority_groups);
198 INIT_LIST_HEAD(&m->queued_ios);
199 spin_lock_init(&m->lock);
201 INIT_WORK(&m->process_queued_ios, process_queued_ios);
202 INIT_WORK(&m->trigger_event, trigger_event);
203 mutex_init(&m->work_mutex);
204 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
216 static void free_multipath(struct multipath *m)
218 struct priority_group *pg, *tmp;
220 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
222 free_priority_group(pg, m->ti);
225 kfree(m->hw_handler_name);
226 kfree(m->hw_handler_params);
227 mempool_destroy(m->mpio_pool);
232 /*-----------------------------------------------
234 *-----------------------------------------------*/
236 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
238 m->current_pg = pgpath->pg;
240 /* Must we initialise the PG first, and queue I/O till it's ready? */
241 if (m->hw_handler_name) {
242 m->pg_init_required = 1;
245 m->pg_init_required = 0;
249 m->pg_init_count = 0;
252 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
255 struct dm_path *path;
257 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
261 m->current_pgpath = path_to_pgpath(path);
263 if (m->current_pg != pg)
264 __switch_pg(m, m->current_pgpath);
269 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
271 struct priority_group *pg;
272 unsigned bypassed = 1;
274 if (!m->nr_valid_paths)
277 /* Were we instructed to switch PG? */
281 if (!__choose_path_in_pg(m, pg, nr_bytes))
285 /* Don't change PG until it has no remaining paths */
286 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
290 * Loop through priority groups until we find a valid path.
291 * First time we skip PGs marked 'bypassed'.
292 * Second time we only try the ones we skipped.
295 list_for_each_entry(pg, &m->priority_groups, list) {
296 if (pg->bypassed == bypassed)
298 if (!__choose_path_in_pg(m, pg, nr_bytes))
301 } while (bypassed--);
304 m->current_pgpath = NULL;
305 m->current_pg = NULL;
309 * Check whether bios must be queued in the device-mapper core rather
310 * than here in the target.
312 * m->lock must be held on entry.
314 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
315 * same value then we are not between multipath_presuspend()
316 * and multipath_resume() calls and we have no need to check
317 * for the DMF_NOFLUSH_SUSPENDING flag.
319 static int __must_push_back(struct multipath *m)
321 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
322 dm_noflush_suspending(m->ti));
325 static int map_io(struct multipath *m, struct request *clone,
326 struct dm_mpath_io *mpio, unsigned was_queued)
328 int r = DM_MAPIO_REMAPPED;
329 size_t nr_bytes = blk_rq_bytes(clone);
331 struct pgpath *pgpath;
332 struct block_device *bdev;
334 spin_lock_irqsave(&m->lock, flags);
336 /* Do we need to select a new pgpath? */
337 if (!m->current_pgpath ||
338 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
339 __choose_pgpath(m, nr_bytes);
341 pgpath = m->current_pgpath;
346 if ((pgpath && m->queue_io) ||
347 (!pgpath && m->queue_if_no_path)) {
348 /* Queue for the daemon to resubmit */
349 list_add_tail(&clone->queuelist, &m->queued_ios);
351 if ((m->pg_init_required && !m->pg_init_in_progress) ||
353 queue_work(kmultipathd, &m->process_queued_ios);
355 r = DM_MAPIO_SUBMITTED;
357 bdev = pgpath->path.dev->bdev;
358 clone->q = bdev_get_queue(bdev);
359 clone->rq_disk = bdev->bd_disk;
360 } else if (__must_push_back(m))
361 r = DM_MAPIO_REQUEUE;
363 r = -EIO; /* Failed */
365 mpio->pgpath = pgpath;
366 mpio->nr_bytes = nr_bytes;
368 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
369 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
372 spin_unlock_irqrestore(&m->lock, flags);
378 * If we run out of usable paths, should we queue I/O or error it?
380 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
381 unsigned save_old_value)
385 spin_lock_irqsave(&m->lock, flags);
388 m->saved_queue_if_no_path = m->queue_if_no_path;
390 m->saved_queue_if_no_path = queue_if_no_path;
391 m->queue_if_no_path = queue_if_no_path;
392 if (!m->queue_if_no_path && m->queue_size)
393 queue_work(kmultipathd, &m->process_queued_ios);
395 spin_unlock_irqrestore(&m->lock, flags);
400 /*-----------------------------------------------------------------
401 * The multipath daemon is responsible for resubmitting queued ios.
402 *---------------------------------------------------------------*/
404 static void dispatch_queued_ios(struct multipath *m)
408 struct dm_mpath_io *mpio;
409 union map_info *info;
410 struct request *clone, *n;
413 spin_lock_irqsave(&m->lock, flags);
414 list_splice_init(&m->queued_ios, &cl);
415 spin_unlock_irqrestore(&m->lock, flags);
417 list_for_each_entry_safe(clone, n, &cl, queuelist) {
418 list_del_init(&clone->queuelist);
420 info = dm_get_rq_mapinfo(clone);
423 r = map_io(m, clone, mpio, 1);
425 mempool_free(mpio, m->mpio_pool);
426 dm_kill_unmapped_request(clone, r);
427 } else if (r == DM_MAPIO_REMAPPED)
428 dm_dispatch_request(clone);
429 else if (r == DM_MAPIO_REQUEUE) {
430 mempool_free(mpio, m->mpio_pool);
431 dm_requeue_unmapped_request(clone);
436 static void process_queued_ios(struct work_struct *work)
438 struct multipath *m =
439 container_of(work, struct multipath, process_queued_ios);
440 struct pgpath *pgpath = NULL, *tmp;
441 unsigned must_queue = 1;
444 spin_lock_irqsave(&m->lock, flags);
449 if (!m->current_pgpath)
450 __choose_pgpath(m, 0);
452 pgpath = m->current_pgpath;
454 if ((pgpath && !m->queue_io) ||
455 (!pgpath && !m->queue_if_no_path))
458 if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
460 m->pg_init_required = 0;
461 list_for_each_entry(tmp, &pgpath->pg->pgpaths, list) {
462 /* Skip failed paths */
465 if (queue_work(kmpath_handlerd, &tmp->activate_path))
466 m->pg_init_in_progress++;
470 spin_unlock_irqrestore(&m->lock, flags);
472 dispatch_queued_ios(m);
476 * An event is triggered whenever a path is taken out of use.
477 * Includes path failure and PG bypass.
479 static void trigger_event(struct work_struct *work)
481 struct multipath *m =
482 container_of(work, struct multipath, trigger_event);
484 dm_table_event(m->ti->table);
487 /*-----------------------------------------------------------------
488 * Constructor/argument parsing:
489 * <#multipath feature args> [<arg>]*
490 * <#hw_handler args> [hw_handler [<arg>]*]
492 * <initial priority group>
493 * [<selector> <#selector args> [<arg>]*
494 * <#paths> <#per-path selector args>
495 * [<path> [<arg>]* ]+ ]+
496 *---------------------------------------------------------------*/
503 static int read_param(struct param *param, char *str, unsigned *v, char **error)
506 (sscanf(str, "%u", v) != 1) ||
509 *error = param->error;
521 static char *shift(struct arg_set *as)
535 static void consume(struct arg_set *as, unsigned n)
537 BUG_ON (as->argc < n);
542 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
543 struct dm_target *ti)
546 struct path_selector_type *pst;
549 static struct param _params[] = {
550 {0, 1024, "invalid number of path selector args"},
553 pst = dm_get_path_selector(shift(as));
555 ti->error = "unknown path selector type";
559 r = read_param(_params, shift(as), &ps_argc, &ti->error);
561 dm_put_path_selector(pst);
565 if (ps_argc > as->argc) {
566 dm_put_path_selector(pst);
567 ti->error = "not enough arguments for path selector";
571 r = pst->create(&pg->ps, ps_argc, as->argv);
573 dm_put_path_selector(pst);
574 ti->error = "path selector constructor failed";
579 consume(as, ps_argc);
584 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
585 struct dm_target *ti)
589 struct multipath *m = ti->private;
591 /* we need at least a path arg */
593 ti->error = "no device given";
594 return ERR_PTR(-EINVAL);
599 return ERR_PTR(-ENOMEM);
601 r = dm_get_device(ti, shift(as), ti->begin, ti->len,
602 dm_table_get_mode(ti->table), &p->path.dev);
604 ti->error = "error getting device";
608 if (m->hw_handler_name) {
609 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
611 r = scsi_dh_attach(q, m->hw_handler_name);
614 * Already attached to different hw_handler,
615 * try to reattach with correct one.
618 r = scsi_dh_attach(q, m->hw_handler_name);
622 ti->error = "error attaching hardware handler";
623 dm_put_device(ti, p->path.dev);
627 if (m->hw_handler_params) {
628 r = scsi_dh_set_params(q, m->hw_handler_params);
630 ti->error = "unable to set hardware "
631 "handler parameters";
633 dm_put_device(ti, p->path.dev);
639 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
641 dm_put_device(ti, p->path.dev);
652 static struct priority_group *parse_priority_group(struct arg_set *as,
655 static struct param _params[] = {
656 {1, 1024, "invalid number of paths"},
657 {0, 1024, "invalid number of selector args"}
661 unsigned i, nr_selector_args, nr_params;
662 struct priority_group *pg;
663 struct dm_target *ti = m->ti;
667 ti->error = "not enough priority group arguments";
668 return ERR_PTR(-EINVAL);
671 pg = alloc_priority_group();
673 ti->error = "couldn't allocate priority group";
674 return ERR_PTR(-ENOMEM);
678 r = parse_path_selector(as, pg, ti);
685 r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
689 r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
693 nr_params = 1 + nr_selector_args;
694 for (i = 0; i < pg->nr_pgpaths; i++) {
695 struct pgpath *pgpath;
696 struct arg_set path_args;
698 if (as->argc < nr_params) {
699 ti->error = "not enough path parameters";
703 path_args.argc = nr_params;
704 path_args.argv = as->argv;
706 pgpath = parse_path(&path_args, &pg->ps, ti);
707 if (IS_ERR(pgpath)) {
713 list_add_tail(&pgpath->list, &pg->pgpaths);
714 consume(as, nr_params);
720 free_priority_group(pg, ti);
724 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
728 struct dm_target *ti = m->ti;
730 static struct param _params[] = {
731 {0, 1024, "invalid number of hardware handler args"},
734 if (read_param(_params, shift(as), &hw_argc, &ti->error))
740 if (hw_argc > as->argc) {
741 ti->error = "not enough arguments for hardware handler";
745 m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
746 request_module("scsi_dh_%s", m->hw_handler_name);
747 if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
748 ti->error = "unknown hardware handler type";
757 for (i = 0; i <= hw_argc - 2; i++)
758 len += strlen(as->argv[i]) + 1;
759 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
761 ti->error = "memory allocation failed";
765 j = sprintf(p, "%d", hw_argc - 1);
766 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
767 j = sprintf(p, "%s", as->argv[i]);
769 consume(as, hw_argc - 1);
773 kfree(m->hw_handler_name);
774 m->hw_handler_name = NULL;
778 static int parse_features(struct arg_set *as, struct multipath *m)
782 struct dm_target *ti = m->ti;
783 const char *param_name;
785 static struct param _params[] = {
786 {0, 3, "invalid number of feature args"},
787 {1, 50, "pg_init_retries must be between 1 and 50"},
790 r = read_param(_params, shift(as), &argc, &ti->error);
798 param_name = shift(as);
801 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
802 r = queue_if_no_path(m, 1, 0);
806 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
808 r = read_param(_params + 1, shift(as),
809 &m->pg_init_retries, &ti->error);
814 ti->error = "Unrecognised multipath feature request";
816 } while (argc && !r);
821 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
824 /* target parameters */
825 static struct param _params[] = {
826 {1, 1024, "invalid number of priority groups"},
827 {1, 1024, "invalid initial priority group number"},
833 unsigned pg_count = 0;
834 unsigned next_pg_num;
839 m = alloc_multipath(ti);
841 ti->error = "can't allocate multipath";
845 r = parse_features(&as, m);
849 r = parse_hw_handler(&as, m);
853 r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
857 r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
861 /* parse the priority groups */
863 struct priority_group *pg;
865 pg = parse_priority_group(&as, m);
871 m->nr_valid_paths += pg->nr_pgpaths;
872 list_add_tail(&pg->list, &m->priority_groups);
874 pg->pg_num = pg_count;
879 if (pg_count != m->nr_priority_groups) {
880 ti->error = "priority group count mismatch";
885 ti->num_flush_requests = 1;
894 static void flush_multipath_work(void)
896 flush_workqueue(kmpath_handlerd);
897 flush_workqueue(kmultipathd);
898 flush_scheduled_work();
901 static void multipath_dtr(struct dm_target *ti)
903 struct multipath *m = ti->private;
905 flush_multipath_work();
910 * Map cloned requests
912 static int multipath_map(struct dm_target *ti, struct request *clone,
913 union map_info *map_context)
916 struct dm_mpath_io *mpio;
917 struct multipath *m = (struct multipath *) ti->private;
919 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
921 /* ENOMEM, requeue */
922 return DM_MAPIO_REQUEUE;
923 memset(mpio, 0, sizeof(*mpio));
925 map_context->ptr = mpio;
926 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
927 r = map_io(m, clone, mpio, 0);
928 if (r < 0 || r == DM_MAPIO_REQUEUE)
929 mempool_free(mpio, m->mpio_pool);
935 * Take a path out of use.
937 static int fail_path(struct pgpath *pgpath)
940 struct multipath *m = pgpath->pg->m;
942 spin_lock_irqsave(&m->lock, flags);
944 if (!pgpath->is_active)
947 DMWARN("Failing path %s.", pgpath->path.dev->name);
949 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
950 pgpath->is_active = 0;
951 pgpath->fail_count++;
955 if (pgpath == m->current_pgpath)
956 m->current_pgpath = NULL;
958 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
959 pgpath->path.dev->name, m->nr_valid_paths);
961 schedule_work(&m->trigger_event);
962 queue_work(kmultipathd, &pgpath->deactivate_path);
965 spin_unlock_irqrestore(&m->lock, flags);
971 * Reinstate a previously-failed path
973 static int reinstate_path(struct pgpath *pgpath)
977 struct multipath *m = pgpath->pg->m;
979 spin_lock_irqsave(&m->lock, flags);
981 if (pgpath->is_active)
984 if (!pgpath->pg->ps.type->reinstate_path) {
985 DMWARN("Reinstate path not supported by path selector %s",
986 pgpath->pg->ps.type->name);
991 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
995 pgpath->is_active = 1;
997 if (!m->nr_valid_paths++ && m->queue_size) {
998 m->current_pgpath = NULL;
999 queue_work(kmultipathd, &m->process_queued_ios);
1000 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1001 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
1002 m->pg_init_in_progress++;
1005 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1006 pgpath->path.dev->name, m->nr_valid_paths);
1008 schedule_work(&m->trigger_event);
1011 spin_unlock_irqrestore(&m->lock, flags);
1017 * Fail or reinstate all paths that match the provided struct dm_dev.
1019 static int action_dev(struct multipath *m, struct dm_dev *dev,
1023 struct pgpath *pgpath;
1024 struct priority_group *pg;
1026 list_for_each_entry(pg, &m->priority_groups, list) {
1027 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1028 if (pgpath->path.dev == dev)
1037 * Temporarily try to avoid having to use the specified PG
1039 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1042 unsigned long flags;
1044 spin_lock_irqsave(&m->lock, flags);
1046 pg->bypassed = bypassed;
1047 m->current_pgpath = NULL;
1048 m->current_pg = NULL;
1050 spin_unlock_irqrestore(&m->lock, flags);
1052 schedule_work(&m->trigger_event);
1056 * Switch to using the specified PG from the next I/O that gets mapped
1058 static int switch_pg_num(struct multipath *m, const char *pgstr)
1060 struct priority_group *pg;
1062 unsigned long flags;
1064 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1065 (pgnum > m->nr_priority_groups)) {
1066 DMWARN("invalid PG number supplied to switch_pg_num");
1070 spin_lock_irqsave(&m->lock, flags);
1071 list_for_each_entry(pg, &m->priority_groups, list) {
1076 m->current_pgpath = NULL;
1077 m->current_pg = NULL;
1080 spin_unlock_irqrestore(&m->lock, flags);
1082 schedule_work(&m->trigger_event);
1087 * Set/clear bypassed status of a PG.
1088 * PGs are numbered upwards from 1 in the order they were declared.
1090 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1092 struct priority_group *pg;
1095 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1096 (pgnum > m->nr_priority_groups)) {
1097 DMWARN("invalid PG number supplied to bypass_pg");
1101 list_for_each_entry(pg, &m->priority_groups, list) {
1106 bypass_pg(m, pg, bypassed);
1111 * Should we retry pg_init immediately?
1113 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1115 unsigned long flags;
1116 int limit_reached = 0;
1118 spin_lock_irqsave(&m->lock, flags);
1120 if (m->pg_init_count <= m->pg_init_retries)
1121 m->pg_init_required = 1;
1125 spin_unlock_irqrestore(&m->lock, flags);
1127 return limit_reached;
1130 static void pg_init_done(void *data, int errors)
1132 struct pgpath *pgpath = data;
1133 struct priority_group *pg = pgpath->pg;
1134 struct multipath *m = pg->m;
1135 unsigned long flags;
1137 /* device or driver problems */
1142 if (!m->hw_handler_name) {
1146 DMERR("Could not failover the device: Handler scsi_dh_%s "
1147 "Error %d.", m->hw_handler_name, errors);
1149 * Fail path for now, so we do not ping pong
1153 case SCSI_DH_DEV_TEMP_BUSY:
1155 * Probably doing something like FW upgrade on the
1156 * controller so try the other pg.
1158 bypass_pg(m, pg, 1);
1160 /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1162 case SCSI_DH_IMM_RETRY:
1163 case SCSI_DH_RES_TEMP_UNAVAIL:
1164 if (pg_init_limit_reached(m, pgpath))
1170 * We probably do not want to fail the path for a device
1171 * error, but this is what the old dm did. In future
1172 * patches we can do more advanced handling.
1177 spin_lock_irqsave(&m->lock, flags);
1179 if (pgpath == m->current_pgpath) {
1180 DMERR("Could not failover device. Error %d.", errors);
1181 m->current_pgpath = NULL;
1182 m->current_pg = NULL;
1184 } else if (!m->pg_init_required) {
1189 m->pg_init_in_progress--;
1190 if (!m->pg_init_in_progress)
1191 queue_work(kmultipathd, &m->process_queued_ios);
1192 spin_unlock_irqrestore(&m->lock, flags);
1195 static void activate_path(struct work_struct *work)
1197 struct pgpath *pgpath =
1198 container_of(work, struct pgpath, activate_path);
1200 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1201 pg_init_done, pgpath);
1207 static int do_end_io(struct multipath *m, struct request *clone,
1208 int error, struct dm_mpath_io *mpio)
1211 * We don't queue any clone request inside the multipath target
1212 * during end I/O handling, since those clone requests don't have
1213 * bio clones. If we queue them inside the multipath target,
1214 * we need to make bio clones, that requires memory allocation.
1215 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1216 * don't have bio clones.)
1217 * Instead of queueing the clone request here, we queue the original
1218 * request into dm core, which will remake a clone request and
1219 * clone bios for it and resubmit it later.
1221 int r = DM_ENDIO_REQUEUE;
1222 unsigned long flags;
1224 if (!error && !clone->errors)
1225 return 0; /* I/O complete */
1227 if (error == -EOPNOTSUPP)
1231 fail_path(mpio->pgpath);
1233 spin_lock_irqsave(&m->lock, flags);
1234 if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1236 spin_unlock_irqrestore(&m->lock, flags);
1241 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1242 int error, union map_info *map_context)
1244 struct multipath *m = ti->private;
1245 struct dm_mpath_io *mpio = map_context->ptr;
1246 struct pgpath *pgpath = mpio->pgpath;
1247 struct path_selector *ps;
1250 r = do_end_io(m, clone, error, mpio);
1252 ps = &pgpath->pg->ps;
1253 if (ps->type->end_io)
1254 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1256 mempool_free(mpio, m->mpio_pool);
1262 * Suspend can't complete until all the I/O is processed so if
1263 * the last path fails we must error any remaining I/O.
1264 * Note that if the freeze_bdev fails while suspending, the
1265 * queue_if_no_path state is lost - userspace should reset it.
1267 static void multipath_presuspend(struct dm_target *ti)
1269 struct multipath *m = (struct multipath *) ti->private;
1271 queue_if_no_path(m, 0, 1);
1274 static void multipath_postsuspend(struct dm_target *ti)
1276 struct multipath *m = ti->private;
1278 mutex_lock(&m->work_mutex);
1279 flush_multipath_work();
1280 mutex_unlock(&m->work_mutex);
1284 * Restore the queue_if_no_path setting.
1286 static void multipath_resume(struct dm_target *ti)
1288 struct multipath *m = (struct multipath *) ti->private;
1289 unsigned long flags;
1291 spin_lock_irqsave(&m->lock, flags);
1292 m->queue_if_no_path = m->saved_queue_if_no_path;
1293 spin_unlock_irqrestore(&m->lock, flags);
1297 * Info output has the following format:
1298 * num_multipath_feature_args [multipath_feature_args]*
1299 * num_handler_status_args [handler_status_args]*
1300 * num_groups init_group_number
1301 * [A|D|E num_ps_status_args [ps_status_args]*
1302 * num_paths num_selector_args
1303 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1305 * Table output has the following format (identical to the constructor string):
1306 * num_feature_args [features_args]*
1307 * num_handler_args hw_handler [hw_handler_args]*
1308 * num_groups init_group_number
1309 * [priority selector-name num_ps_args [ps_args]*
1310 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1312 static int multipath_status(struct dm_target *ti, status_type_t type,
1313 char *result, unsigned int maxlen)
1316 unsigned long flags;
1317 struct multipath *m = (struct multipath *) ti->private;
1318 struct priority_group *pg;
1323 spin_lock_irqsave(&m->lock, flags);
1326 if (type == STATUSTYPE_INFO)
1327 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1329 DMEMIT("%u ", m->queue_if_no_path +
1330 (m->pg_init_retries > 0) * 2);
1331 if (m->queue_if_no_path)
1332 DMEMIT("queue_if_no_path ");
1333 if (m->pg_init_retries)
1334 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1337 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1340 DMEMIT("1 %s ", m->hw_handler_name);
1342 DMEMIT("%u ", m->nr_priority_groups);
1345 pg_num = m->next_pg->pg_num;
1346 else if (m->current_pg)
1347 pg_num = m->current_pg->pg_num;
1351 DMEMIT("%u ", pg_num);
1354 case STATUSTYPE_INFO:
1355 list_for_each_entry(pg, &m->priority_groups, list) {
1357 state = 'D'; /* Disabled */
1358 else if (pg == m->current_pg)
1359 state = 'A'; /* Currently Active */
1361 state = 'E'; /* Enabled */
1363 DMEMIT("%c ", state);
1365 if (pg->ps.type->status)
1366 sz += pg->ps.type->status(&pg->ps, NULL, type,
1372 DMEMIT("%u %u ", pg->nr_pgpaths,
1373 pg->ps.type->info_args);
1375 list_for_each_entry(p, &pg->pgpaths, list) {
1376 DMEMIT("%s %s %u ", p->path.dev->name,
1377 p->is_active ? "A" : "F",
1379 if (pg->ps.type->status)
1380 sz += pg->ps.type->status(&pg->ps,
1381 &p->path, type, result + sz,
1387 case STATUSTYPE_TABLE:
1388 list_for_each_entry(pg, &m->priority_groups, list) {
1389 DMEMIT("%s ", pg->ps.type->name);
1391 if (pg->ps.type->status)
1392 sz += pg->ps.type->status(&pg->ps, NULL, type,
1398 DMEMIT("%u %u ", pg->nr_pgpaths,
1399 pg->ps.type->table_args);
1401 list_for_each_entry(p, &pg->pgpaths, list) {
1402 DMEMIT("%s ", p->path.dev->name);
1403 if (pg->ps.type->status)
1404 sz += pg->ps.type->status(&pg->ps,
1405 &p->path, type, result + sz,
1412 spin_unlock_irqrestore(&m->lock, flags);
1417 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1421 struct multipath *m = (struct multipath *) ti->private;
1424 mutex_lock(&m->work_mutex);
1426 if (dm_suspended(ti)) {
1432 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
1433 r = queue_if_no_path(m, 1, 0);
1435 } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
1436 r = queue_if_no_path(m, 0, 0);
1442 DMWARN("Unrecognised multipath message received.");
1446 if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
1447 r = bypass_pg_num(m, argv[1], 1);
1449 } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
1450 r = bypass_pg_num(m, argv[1], 0);
1452 } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
1453 r = switch_pg_num(m, argv[1]);
1455 } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1456 action = reinstate_path;
1457 else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1460 DMWARN("Unrecognised multipath message received.");
1464 r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1465 dm_table_get_mode(ti->table), &dev);
1467 DMWARN("message: error getting device %s",
1472 r = action_dev(m, dev, action);
1474 dm_put_device(ti, dev);
1477 mutex_unlock(&m->work_mutex);
1481 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1484 struct multipath *m = (struct multipath *) ti->private;
1485 struct block_device *bdev = NULL;
1487 unsigned long flags;
1490 spin_lock_irqsave(&m->lock, flags);
1492 if (!m->current_pgpath)
1493 __choose_pgpath(m, 0);
1495 if (m->current_pgpath) {
1496 bdev = m->current_pgpath->path.dev->bdev;
1497 mode = m->current_pgpath->path.dev->mode;
1505 spin_unlock_irqrestore(&m->lock, flags);
1507 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1510 static int multipath_iterate_devices(struct dm_target *ti,
1511 iterate_devices_callout_fn fn, void *data)
1513 struct multipath *m = ti->private;
1514 struct priority_group *pg;
1518 list_for_each_entry(pg, &m->priority_groups, list) {
1519 list_for_each_entry(p, &pg->pgpaths, list) {
1520 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1530 static int __pgpath_busy(struct pgpath *pgpath)
1532 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1534 return dm_underlying_device_busy(q);
1538 * We return "busy", only when we can map I/Os but underlying devices
1539 * are busy (so even if we map I/Os now, the I/Os will wait on
1540 * the underlying queue).
1541 * In other words, if we want to kill I/Os or queue them inside us
1542 * due to map unavailability, we don't return "busy". Otherwise,
1543 * dm core won't give us the I/Os and we can't do what we want.
1545 static int multipath_busy(struct dm_target *ti)
1547 int busy = 0, has_active = 0;
1548 struct multipath *m = ti->private;
1549 struct priority_group *pg;
1550 struct pgpath *pgpath;
1551 unsigned long flags;
1553 spin_lock_irqsave(&m->lock, flags);
1555 /* Guess which priority_group will be used at next mapping time */
1556 if (unlikely(!m->current_pgpath && m->next_pg))
1558 else if (likely(m->current_pg))
1562 * We don't know which pg will be used at next mapping time.
1563 * We don't call __choose_pgpath() here to avoid to trigger
1564 * pg_init just by busy checking.
1565 * So we don't know whether underlying devices we will be using
1566 * at next mapping time are busy or not. Just try mapping.
1571 * If there is one non-busy active path at least, the path selector
1572 * will be able to select it. So we consider such a pg as not busy.
1575 list_for_each_entry(pgpath, &pg->pgpaths, list)
1576 if (pgpath->is_active) {
1579 if (!__pgpath_busy(pgpath)) {
1587 * No active path in this pg, so this pg won't be used and
1588 * the current_pg will be changed at next mapping time.
1589 * We need to try mapping to determine it.
1594 spin_unlock_irqrestore(&m->lock, flags);
1599 /*-----------------------------------------------------------------
1601 *---------------------------------------------------------------*/
1602 static struct target_type multipath_target = {
1603 .name = "multipath",
1604 .version = {1, 1, 1},
1605 .module = THIS_MODULE,
1606 .ctr = multipath_ctr,
1607 .dtr = multipath_dtr,
1608 .map_rq = multipath_map,
1609 .rq_end_io = multipath_end_io,
1610 .presuspend = multipath_presuspend,
1611 .postsuspend = multipath_postsuspend,
1612 .resume = multipath_resume,
1613 .status = multipath_status,
1614 .message = multipath_message,
1615 .ioctl = multipath_ioctl,
1616 .iterate_devices = multipath_iterate_devices,
1617 .busy = multipath_busy,
1620 static int __init dm_multipath_init(void)
1624 /* allocate a slab for the dm_ios */
1625 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1629 r = dm_register_target(&multipath_target);
1631 DMERR("register failed %d", r);
1632 kmem_cache_destroy(_mpio_cache);
1636 kmultipathd = create_workqueue("kmpathd");
1638 DMERR("failed to create workqueue kmpathd");
1639 dm_unregister_target(&multipath_target);
1640 kmem_cache_destroy(_mpio_cache);
1645 * A separate workqueue is used to handle the device handlers
1646 * to avoid overloading existing workqueue. Overloading the
1647 * old workqueue would also create a bottleneck in the
1648 * path of the storage hardware device activation.
1650 kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1651 if (!kmpath_handlerd) {
1652 DMERR("failed to create workqueue kmpath_handlerd");
1653 destroy_workqueue(kmultipathd);
1654 dm_unregister_target(&multipath_target);
1655 kmem_cache_destroy(_mpio_cache);
1659 DMINFO("version %u.%u.%u loaded",
1660 multipath_target.version[0], multipath_target.version[1],
1661 multipath_target.version[2]);
1666 static void __exit dm_multipath_exit(void)
1668 destroy_workqueue(kmpath_handlerd);
1669 destroy_workqueue(kmultipathd);
1671 dm_unregister_target(&multipath_target);
1672 kmem_cache_destroy(_mpio_cache);
1675 module_init(dm_multipath_init);
1676 module_exit(dm_multipath_exit);
1678 MODULE_DESCRIPTION(DM_NAME " multipath target");
1679 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1680 MODULE_LICENSE("GPL");