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 */
72 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
74 unsigned nr_valid_paths; /* Total number of usable paths */
75 struct pgpath *current_pgpath;
76 struct priority_group *current_pg;
77 struct priority_group *next_pg; /* Switch to this PG if set */
78 unsigned repeat_count; /* I/Os left before calling PS again */
80 unsigned queue_io; /* Must we queue all I/O? */
81 unsigned queue_if_no_path; /* Queue I/O if last path fails? */
82 unsigned saved_queue_if_no_path;/* Saved state during suspension */
83 unsigned pg_init_retries; /* Number of times to retry pg_init */
84 unsigned pg_init_count; /* Number of times pg_init called */
86 struct work_struct process_queued_ios;
87 struct list_head queued_ios;
90 struct work_struct trigger_event;
93 * We must use a mempool of dm_mpath_io structs so that we
94 * can resubmit bios on error.
98 struct mutex work_mutex;
102 * Context information attached to each bio we process.
105 struct pgpath *pgpath;
109 typedef int (*action_fn) (struct pgpath *pgpath);
111 #define MIN_IOS 256 /* Mempool size */
113 static struct kmem_cache *_mpio_cache;
115 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
116 static void process_queued_ios(struct work_struct *work);
117 static void trigger_event(struct work_struct *work);
118 static void activate_path(struct work_struct *work);
119 static void deactivate_path(struct work_struct *work);
122 /*-----------------------------------------------
123 * Allocation routines
124 *-----------------------------------------------*/
126 static struct pgpath *alloc_pgpath(void)
128 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
131 pgpath->is_active = 1;
132 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
133 INIT_WORK(&pgpath->activate_path, activate_path);
139 static void free_pgpath(struct pgpath *pgpath)
144 static void deactivate_path(struct work_struct *work)
146 struct pgpath *pgpath =
147 container_of(work, struct pgpath, deactivate_path);
149 blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
152 static struct priority_group *alloc_priority_group(void)
154 struct priority_group *pg;
156 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
159 INIT_LIST_HEAD(&pg->pgpaths);
164 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
166 struct pgpath *pgpath, *tmp;
167 struct multipath *m = ti->private;
169 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
170 list_del(&pgpath->list);
171 if (m->hw_handler_name)
172 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
173 dm_put_device(ti, pgpath->path.dev);
178 static void free_priority_group(struct priority_group *pg,
179 struct dm_target *ti)
181 struct path_selector *ps = &pg->ps;
184 ps->type->destroy(ps);
185 dm_put_path_selector(ps->type);
188 free_pgpaths(&pg->pgpaths, ti);
192 static struct multipath *alloc_multipath(struct dm_target *ti)
196 m = kzalloc(sizeof(*m), GFP_KERNEL);
198 INIT_LIST_HEAD(&m->priority_groups);
199 INIT_LIST_HEAD(&m->queued_ios);
200 spin_lock_init(&m->lock);
202 INIT_WORK(&m->process_queued_ios, process_queued_ios);
203 INIT_WORK(&m->trigger_event, trigger_event);
204 init_waitqueue_head(&m->pg_init_wait);
205 mutex_init(&m->work_mutex);
206 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
218 static void free_multipath(struct multipath *m)
220 struct priority_group *pg, *tmp;
222 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
224 free_priority_group(pg, m->ti);
227 kfree(m->hw_handler_name);
228 kfree(m->hw_handler_params);
229 mempool_destroy(m->mpio_pool);
234 /*-----------------------------------------------
236 *-----------------------------------------------*/
238 static void __pg_init_all_paths(struct multipath *m)
240 struct pgpath *pgpath;
243 m->pg_init_required = 0;
244 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
245 /* Skip failed paths */
246 if (!pgpath->is_active)
248 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
249 m->pg_init_in_progress++;
253 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
255 m->current_pg = pgpath->pg;
257 /* Must we initialise the PG first, and queue I/O till it's ready? */
258 if (m->hw_handler_name) {
259 m->pg_init_required = 1;
262 m->pg_init_required = 0;
266 m->pg_init_count = 0;
269 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
272 struct dm_path *path;
274 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
278 m->current_pgpath = path_to_pgpath(path);
280 if (m->current_pg != pg)
281 __switch_pg(m, m->current_pgpath);
286 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
288 struct priority_group *pg;
289 unsigned bypassed = 1;
291 if (!m->nr_valid_paths)
294 /* Were we instructed to switch PG? */
298 if (!__choose_path_in_pg(m, pg, nr_bytes))
302 /* Don't change PG until it has no remaining paths */
303 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
307 * Loop through priority groups until we find a valid path.
308 * First time we skip PGs marked 'bypassed'.
309 * Second time we only try the ones we skipped.
312 list_for_each_entry(pg, &m->priority_groups, list) {
313 if (pg->bypassed == bypassed)
315 if (!__choose_path_in_pg(m, pg, nr_bytes))
318 } while (bypassed--);
321 m->current_pgpath = NULL;
322 m->current_pg = NULL;
326 * Check whether bios must be queued in the device-mapper core rather
327 * than here in the target.
329 * m->lock must be held on entry.
331 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
332 * same value then we are not between multipath_presuspend()
333 * and multipath_resume() calls and we have no need to check
334 * for the DMF_NOFLUSH_SUSPENDING flag.
336 static int __must_push_back(struct multipath *m)
338 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
339 dm_noflush_suspending(m->ti));
342 static int map_io(struct multipath *m, struct request *clone,
343 struct dm_mpath_io *mpio, unsigned was_queued)
345 int r = DM_MAPIO_REMAPPED;
346 size_t nr_bytes = blk_rq_bytes(clone);
348 struct pgpath *pgpath;
349 struct block_device *bdev;
351 spin_lock_irqsave(&m->lock, flags);
353 /* Do we need to select a new pgpath? */
354 if (!m->current_pgpath ||
355 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
356 __choose_pgpath(m, nr_bytes);
358 pgpath = m->current_pgpath;
363 if ((pgpath && m->queue_io) ||
364 (!pgpath && m->queue_if_no_path)) {
365 /* Queue for the daemon to resubmit */
366 list_add_tail(&clone->queuelist, &m->queued_ios);
368 if ((m->pg_init_required && !m->pg_init_in_progress) ||
370 queue_work(kmultipathd, &m->process_queued_ios);
372 r = DM_MAPIO_SUBMITTED;
374 bdev = pgpath->path.dev->bdev;
375 clone->q = bdev_get_queue(bdev);
376 clone->rq_disk = bdev->bd_disk;
377 } else if (__must_push_back(m))
378 r = DM_MAPIO_REQUEUE;
380 r = -EIO; /* Failed */
382 mpio->pgpath = pgpath;
383 mpio->nr_bytes = nr_bytes;
385 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
386 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
389 spin_unlock_irqrestore(&m->lock, flags);
395 * If we run out of usable paths, should we queue I/O or error it?
397 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
398 unsigned save_old_value)
402 spin_lock_irqsave(&m->lock, flags);
405 m->saved_queue_if_no_path = m->queue_if_no_path;
407 m->saved_queue_if_no_path = queue_if_no_path;
408 m->queue_if_no_path = queue_if_no_path;
409 if (!m->queue_if_no_path && m->queue_size)
410 queue_work(kmultipathd, &m->process_queued_ios);
412 spin_unlock_irqrestore(&m->lock, flags);
417 /*-----------------------------------------------------------------
418 * The multipath daemon is responsible for resubmitting queued ios.
419 *---------------------------------------------------------------*/
421 static void dispatch_queued_ios(struct multipath *m)
425 struct dm_mpath_io *mpio;
426 union map_info *info;
427 struct request *clone, *n;
430 spin_lock_irqsave(&m->lock, flags);
431 list_splice_init(&m->queued_ios, &cl);
432 spin_unlock_irqrestore(&m->lock, flags);
434 list_for_each_entry_safe(clone, n, &cl, queuelist) {
435 list_del_init(&clone->queuelist);
437 info = dm_get_rq_mapinfo(clone);
440 r = map_io(m, clone, mpio, 1);
442 mempool_free(mpio, m->mpio_pool);
443 dm_kill_unmapped_request(clone, r);
444 } else if (r == DM_MAPIO_REMAPPED)
445 dm_dispatch_request(clone);
446 else if (r == DM_MAPIO_REQUEUE) {
447 mempool_free(mpio, m->mpio_pool);
448 dm_requeue_unmapped_request(clone);
453 static void process_queued_ios(struct work_struct *work)
455 struct multipath *m =
456 container_of(work, struct multipath, process_queued_ios);
457 struct pgpath *pgpath = NULL;
458 unsigned must_queue = 1;
461 spin_lock_irqsave(&m->lock, flags);
466 if (!m->current_pgpath)
467 __choose_pgpath(m, 0);
469 pgpath = m->current_pgpath;
471 if ((pgpath && !m->queue_io) ||
472 (!pgpath && !m->queue_if_no_path))
475 if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
476 __pg_init_all_paths(m);
479 spin_unlock_irqrestore(&m->lock, flags);
481 dispatch_queued_ios(m);
485 * An event is triggered whenever a path is taken out of use.
486 * Includes path failure and PG bypass.
488 static void trigger_event(struct work_struct *work)
490 struct multipath *m =
491 container_of(work, struct multipath, trigger_event);
493 dm_table_event(m->ti->table);
496 /*-----------------------------------------------------------------
497 * Constructor/argument parsing:
498 * <#multipath feature args> [<arg>]*
499 * <#hw_handler args> [hw_handler [<arg>]*]
501 * <initial priority group>
502 * [<selector> <#selector args> [<arg>]*
503 * <#paths> <#per-path selector args>
504 * [<path> [<arg>]* ]+ ]+
505 *---------------------------------------------------------------*/
512 static int read_param(struct param *param, char *str, unsigned *v, char **error)
515 (sscanf(str, "%u", v) != 1) ||
518 *error = param->error;
530 static char *shift(struct arg_set *as)
544 static void consume(struct arg_set *as, unsigned n)
546 BUG_ON (as->argc < n);
551 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
552 struct dm_target *ti)
555 struct path_selector_type *pst;
558 static struct param _params[] = {
559 {0, 1024, "invalid number of path selector args"},
562 pst = dm_get_path_selector(shift(as));
564 ti->error = "unknown path selector type";
568 r = read_param(_params, shift(as), &ps_argc, &ti->error);
570 dm_put_path_selector(pst);
574 if (ps_argc > as->argc) {
575 dm_put_path_selector(pst);
576 ti->error = "not enough arguments for path selector";
580 r = pst->create(&pg->ps, ps_argc, as->argv);
582 dm_put_path_selector(pst);
583 ti->error = "path selector constructor failed";
588 consume(as, ps_argc);
593 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
594 struct dm_target *ti)
598 struct multipath *m = ti->private;
600 /* we need at least a path arg */
602 ti->error = "no device given";
603 return ERR_PTR(-EINVAL);
608 return ERR_PTR(-ENOMEM);
610 r = dm_get_device(ti, shift(as), dm_table_get_mode(ti->table),
613 ti->error = "error getting device";
617 if (m->hw_handler_name) {
618 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
620 r = scsi_dh_attach(q, m->hw_handler_name);
623 * Already attached to different hw_handler,
624 * try to reattach with correct one.
627 r = scsi_dh_attach(q, m->hw_handler_name);
631 ti->error = "error attaching hardware handler";
632 dm_put_device(ti, p->path.dev);
636 if (m->hw_handler_params) {
637 r = scsi_dh_set_params(q, m->hw_handler_params);
639 ti->error = "unable to set hardware "
640 "handler parameters";
642 dm_put_device(ti, p->path.dev);
648 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
650 dm_put_device(ti, p->path.dev);
661 static struct priority_group *parse_priority_group(struct arg_set *as,
664 static struct param _params[] = {
665 {1, 1024, "invalid number of paths"},
666 {0, 1024, "invalid number of selector args"}
670 unsigned i, nr_selector_args, nr_params;
671 struct priority_group *pg;
672 struct dm_target *ti = m->ti;
676 ti->error = "not enough priority group arguments";
677 return ERR_PTR(-EINVAL);
680 pg = alloc_priority_group();
682 ti->error = "couldn't allocate priority group";
683 return ERR_PTR(-ENOMEM);
687 r = parse_path_selector(as, pg, ti);
694 r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
698 r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
702 nr_params = 1 + nr_selector_args;
703 for (i = 0; i < pg->nr_pgpaths; i++) {
704 struct pgpath *pgpath;
705 struct arg_set path_args;
707 if (as->argc < nr_params) {
708 ti->error = "not enough path parameters";
712 path_args.argc = nr_params;
713 path_args.argv = as->argv;
715 pgpath = parse_path(&path_args, &pg->ps, ti);
716 if (IS_ERR(pgpath)) {
722 list_add_tail(&pgpath->list, &pg->pgpaths);
723 consume(as, nr_params);
729 free_priority_group(pg, ti);
733 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
737 struct dm_target *ti = m->ti;
739 static struct param _params[] = {
740 {0, 1024, "invalid number of hardware handler args"},
743 if (read_param(_params, shift(as), &hw_argc, &ti->error))
749 if (hw_argc > as->argc) {
750 ti->error = "not enough arguments for hardware handler";
754 m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
755 request_module("scsi_dh_%s", m->hw_handler_name);
756 if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
757 ti->error = "unknown hardware handler type";
766 for (i = 0; i <= hw_argc - 2; i++)
767 len += strlen(as->argv[i]) + 1;
768 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
770 ti->error = "memory allocation failed";
774 j = sprintf(p, "%d", hw_argc - 1);
775 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
776 j = sprintf(p, "%s", as->argv[i]);
778 consume(as, hw_argc - 1);
782 kfree(m->hw_handler_name);
783 m->hw_handler_name = NULL;
787 static int parse_features(struct arg_set *as, struct multipath *m)
791 struct dm_target *ti = m->ti;
792 const char *param_name;
794 static struct param _params[] = {
795 {0, 3, "invalid number of feature args"},
796 {1, 50, "pg_init_retries must be between 1 and 50"},
799 r = read_param(_params, shift(as), &argc, &ti->error);
807 param_name = shift(as);
810 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
811 r = queue_if_no_path(m, 1, 0);
815 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
817 r = read_param(_params + 1, shift(as),
818 &m->pg_init_retries, &ti->error);
823 ti->error = "Unrecognised multipath feature request";
825 } while (argc && !r);
830 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
833 /* target parameters */
834 static struct param _params[] = {
835 {1, 1024, "invalid number of priority groups"},
836 {1, 1024, "invalid initial priority group number"},
842 unsigned pg_count = 0;
843 unsigned next_pg_num;
848 m = alloc_multipath(ti);
850 ti->error = "can't allocate multipath";
854 r = parse_features(&as, m);
858 r = parse_hw_handler(&as, m);
862 r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
866 r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
870 /* parse the priority groups */
872 struct priority_group *pg;
874 pg = parse_priority_group(&as, m);
880 m->nr_valid_paths += pg->nr_pgpaths;
881 list_add_tail(&pg->list, &m->priority_groups);
883 pg->pg_num = pg_count;
888 if (pg_count != m->nr_priority_groups) {
889 ti->error = "priority group count mismatch";
894 ti->num_flush_requests = 1;
903 static void multipath_wait_for_pg_init_completion(struct multipath *m)
905 DECLARE_WAITQUEUE(wait, current);
908 add_wait_queue(&m->pg_init_wait, &wait);
911 set_current_state(TASK_UNINTERRUPTIBLE);
913 spin_lock_irqsave(&m->lock, flags);
914 if (!m->pg_init_in_progress) {
915 spin_unlock_irqrestore(&m->lock, flags);
918 spin_unlock_irqrestore(&m->lock, flags);
922 set_current_state(TASK_RUNNING);
924 remove_wait_queue(&m->pg_init_wait, &wait);
927 static void flush_multipath_work(struct multipath *m)
929 flush_workqueue(kmpath_handlerd);
930 multipath_wait_for_pg_init_completion(m);
931 flush_workqueue(kmultipathd);
932 flush_scheduled_work();
935 static void multipath_dtr(struct dm_target *ti)
937 struct multipath *m = ti->private;
939 flush_multipath_work(m);
944 * Map cloned requests
946 static int multipath_map(struct dm_target *ti, struct request *clone,
947 union map_info *map_context)
950 struct dm_mpath_io *mpio;
951 struct multipath *m = (struct multipath *) ti->private;
953 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
955 /* ENOMEM, requeue */
956 return DM_MAPIO_REQUEUE;
957 memset(mpio, 0, sizeof(*mpio));
959 map_context->ptr = mpio;
960 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
961 r = map_io(m, clone, mpio, 0);
962 if (r < 0 || r == DM_MAPIO_REQUEUE)
963 mempool_free(mpio, m->mpio_pool);
969 * Take a path out of use.
971 static int fail_path(struct pgpath *pgpath)
974 struct multipath *m = pgpath->pg->m;
976 spin_lock_irqsave(&m->lock, flags);
978 if (!pgpath->is_active)
981 DMWARN("Failing path %s.", pgpath->path.dev->name);
983 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
984 pgpath->is_active = 0;
985 pgpath->fail_count++;
989 if (pgpath == m->current_pgpath)
990 m->current_pgpath = NULL;
992 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
993 pgpath->path.dev->name, m->nr_valid_paths);
995 schedule_work(&m->trigger_event);
996 queue_work(kmultipathd, &pgpath->deactivate_path);
999 spin_unlock_irqrestore(&m->lock, flags);
1005 * Reinstate a previously-failed path
1007 static int reinstate_path(struct pgpath *pgpath)
1010 unsigned long flags;
1011 struct multipath *m = pgpath->pg->m;
1013 spin_lock_irqsave(&m->lock, flags);
1015 if (pgpath->is_active)
1018 if (!pgpath->pg->ps.type->reinstate_path) {
1019 DMWARN("Reinstate path not supported by path selector %s",
1020 pgpath->pg->ps.type->name);
1025 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1029 pgpath->is_active = 1;
1031 if (!m->nr_valid_paths++ && m->queue_size) {
1032 m->current_pgpath = NULL;
1033 queue_work(kmultipathd, &m->process_queued_ios);
1034 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1035 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
1036 m->pg_init_in_progress++;
1039 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1040 pgpath->path.dev->name, m->nr_valid_paths);
1042 schedule_work(&m->trigger_event);
1045 spin_unlock_irqrestore(&m->lock, flags);
1051 * Fail or reinstate all paths that match the provided struct dm_dev.
1053 static int action_dev(struct multipath *m, struct dm_dev *dev,
1057 struct pgpath *pgpath;
1058 struct priority_group *pg;
1060 list_for_each_entry(pg, &m->priority_groups, list) {
1061 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1062 if (pgpath->path.dev == dev)
1071 * Temporarily try to avoid having to use the specified PG
1073 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1076 unsigned long flags;
1078 spin_lock_irqsave(&m->lock, flags);
1080 pg->bypassed = bypassed;
1081 m->current_pgpath = NULL;
1082 m->current_pg = NULL;
1084 spin_unlock_irqrestore(&m->lock, flags);
1086 schedule_work(&m->trigger_event);
1090 * Switch to using the specified PG from the next I/O that gets mapped
1092 static int switch_pg_num(struct multipath *m, const char *pgstr)
1094 struct priority_group *pg;
1096 unsigned long flags;
1098 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1099 (pgnum > m->nr_priority_groups)) {
1100 DMWARN("invalid PG number supplied to switch_pg_num");
1104 spin_lock_irqsave(&m->lock, flags);
1105 list_for_each_entry(pg, &m->priority_groups, list) {
1110 m->current_pgpath = NULL;
1111 m->current_pg = NULL;
1114 spin_unlock_irqrestore(&m->lock, flags);
1116 schedule_work(&m->trigger_event);
1121 * Set/clear bypassed status of a PG.
1122 * PGs are numbered upwards from 1 in the order they were declared.
1124 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1126 struct priority_group *pg;
1129 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1130 (pgnum > m->nr_priority_groups)) {
1131 DMWARN("invalid PG number supplied to bypass_pg");
1135 list_for_each_entry(pg, &m->priority_groups, list) {
1140 bypass_pg(m, pg, bypassed);
1145 * Should we retry pg_init immediately?
1147 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1149 unsigned long flags;
1150 int limit_reached = 0;
1152 spin_lock_irqsave(&m->lock, flags);
1154 if (m->pg_init_count <= m->pg_init_retries)
1155 m->pg_init_required = 1;
1159 spin_unlock_irqrestore(&m->lock, flags);
1161 return limit_reached;
1164 static void pg_init_done(void *data, int errors)
1166 struct pgpath *pgpath = data;
1167 struct priority_group *pg = pgpath->pg;
1168 struct multipath *m = pg->m;
1169 unsigned long flags;
1171 /* device or driver problems */
1176 if (!m->hw_handler_name) {
1180 DMERR("Could not failover the device: Handler scsi_dh_%s "
1181 "Error %d.", m->hw_handler_name, errors);
1183 * Fail path for now, so we do not ping pong
1187 case SCSI_DH_DEV_TEMP_BUSY:
1189 * Probably doing something like FW upgrade on the
1190 * controller so try the other pg.
1192 bypass_pg(m, pg, 1);
1194 /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1196 case SCSI_DH_IMM_RETRY:
1197 case SCSI_DH_RES_TEMP_UNAVAIL:
1198 if (pg_init_limit_reached(m, pgpath))
1204 * We probably do not want to fail the path for a device
1205 * error, but this is what the old dm did. In future
1206 * patches we can do more advanced handling.
1211 spin_lock_irqsave(&m->lock, flags);
1213 if (pgpath == m->current_pgpath) {
1214 DMERR("Could not failover device. Error %d.", errors);
1215 m->current_pgpath = NULL;
1216 m->current_pg = NULL;
1218 } else if (!m->pg_init_required)
1221 if (--m->pg_init_in_progress)
1222 /* Activations of other paths are still on going */
1225 if (!m->pg_init_required)
1228 queue_work(kmultipathd, &m->process_queued_ios);
1231 * Wake up any thread waiting to suspend.
1233 wake_up(&m->pg_init_wait);
1236 spin_unlock_irqrestore(&m->lock, flags);
1239 static void activate_path(struct work_struct *work)
1241 struct pgpath *pgpath =
1242 container_of(work, struct pgpath, activate_path);
1244 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1245 pg_init_done, pgpath);
1251 static int do_end_io(struct multipath *m, struct request *clone,
1252 int error, struct dm_mpath_io *mpio)
1255 * We don't queue any clone request inside the multipath target
1256 * during end I/O handling, since those clone requests don't have
1257 * bio clones. If we queue them inside the multipath target,
1258 * we need to make bio clones, that requires memory allocation.
1259 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1260 * don't have bio clones.)
1261 * Instead of queueing the clone request here, we queue the original
1262 * request into dm core, which will remake a clone request and
1263 * clone bios for it and resubmit it later.
1265 int r = DM_ENDIO_REQUEUE;
1266 unsigned long flags;
1268 if (!error && !clone->errors)
1269 return 0; /* I/O complete */
1271 if (error == -EOPNOTSUPP)
1275 fail_path(mpio->pgpath);
1277 spin_lock_irqsave(&m->lock, flags);
1278 if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1280 spin_unlock_irqrestore(&m->lock, flags);
1285 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1286 int error, union map_info *map_context)
1288 struct multipath *m = ti->private;
1289 struct dm_mpath_io *mpio = map_context->ptr;
1290 struct pgpath *pgpath = mpio->pgpath;
1291 struct path_selector *ps;
1294 r = do_end_io(m, clone, error, mpio);
1296 ps = &pgpath->pg->ps;
1297 if (ps->type->end_io)
1298 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1300 mempool_free(mpio, m->mpio_pool);
1306 * Suspend can't complete until all the I/O is processed so if
1307 * the last path fails we must error any remaining I/O.
1308 * Note that if the freeze_bdev fails while suspending, the
1309 * queue_if_no_path state is lost - userspace should reset it.
1311 static void multipath_presuspend(struct dm_target *ti)
1313 struct multipath *m = (struct multipath *) ti->private;
1315 queue_if_no_path(m, 0, 1);
1318 static void multipath_postsuspend(struct dm_target *ti)
1320 struct multipath *m = ti->private;
1322 mutex_lock(&m->work_mutex);
1323 flush_multipath_work(m);
1324 mutex_unlock(&m->work_mutex);
1328 * Restore the queue_if_no_path setting.
1330 static void multipath_resume(struct dm_target *ti)
1332 struct multipath *m = (struct multipath *) ti->private;
1333 unsigned long flags;
1335 spin_lock_irqsave(&m->lock, flags);
1336 m->queue_if_no_path = m->saved_queue_if_no_path;
1337 spin_unlock_irqrestore(&m->lock, flags);
1341 * Info output has the following format:
1342 * num_multipath_feature_args [multipath_feature_args]*
1343 * num_handler_status_args [handler_status_args]*
1344 * num_groups init_group_number
1345 * [A|D|E num_ps_status_args [ps_status_args]*
1346 * num_paths num_selector_args
1347 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1349 * Table output has the following format (identical to the constructor string):
1350 * num_feature_args [features_args]*
1351 * num_handler_args hw_handler [hw_handler_args]*
1352 * num_groups init_group_number
1353 * [priority selector-name num_ps_args [ps_args]*
1354 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1356 static int multipath_status(struct dm_target *ti, status_type_t type,
1357 char *result, unsigned int maxlen)
1360 unsigned long flags;
1361 struct multipath *m = (struct multipath *) ti->private;
1362 struct priority_group *pg;
1367 spin_lock_irqsave(&m->lock, flags);
1370 if (type == STATUSTYPE_INFO)
1371 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1373 DMEMIT("%u ", m->queue_if_no_path +
1374 (m->pg_init_retries > 0) * 2);
1375 if (m->queue_if_no_path)
1376 DMEMIT("queue_if_no_path ");
1377 if (m->pg_init_retries)
1378 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1381 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1384 DMEMIT("1 %s ", m->hw_handler_name);
1386 DMEMIT("%u ", m->nr_priority_groups);
1389 pg_num = m->next_pg->pg_num;
1390 else if (m->current_pg)
1391 pg_num = m->current_pg->pg_num;
1395 DMEMIT("%u ", pg_num);
1398 case STATUSTYPE_INFO:
1399 list_for_each_entry(pg, &m->priority_groups, list) {
1401 state = 'D'; /* Disabled */
1402 else if (pg == m->current_pg)
1403 state = 'A'; /* Currently Active */
1405 state = 'E'; /* Enabled */
1407 DMEMIT("%c ", state);
1409 if (pg->ps.type->status)
1410 sz += pg->ps.type->status(&pg->ps, NULL, type,
1416 DMEMIT("%u %u ", pg->nr_pgpaths,
1417 pg->ps.type->info_args);
1419 list_for_each_entry(p, &pg->pgpaths, list) {
1420 DMEMIT("%s %s %u ", p->path.dev->name,
1421 p->is_active ? "A" : "F",
1423 if (pg->ps.type->status)
1424 sz += pg->ps.type->status(&pg->ps,
1425 &p->path, type, result + sz,
1431 case STATUSTYPE_TABLE:
1432 list_for_each_entry(pg, &m->priority_groups, list) {
1433 DMEMIT("%s ", pg->ps.type->name);
1435 if (pg->ps.type->status)
1436 sz += pg->ps.type->status(&pg->ps, NULL, type,
1442 DMEMIT("%u %u ", pg->nr_pgpaths,
1443 pg->ps.type->table_args);
1445 list_for_each_entry(p, &pg->pgpaths, list) {
1446 DMEMIT("%s ", p->path.dev->name);
1447 if (pg->ps.type->status)
1448 sz += pg->ps.type->status(&pg->ps,
1449 &p->path, type, result + sz,
1456 spin_unlock_irqrestore(&m->lock, flags);
1461 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1465 struct multipath *m = (struct multipath *) ti->private;
1468 mutex_lock(&m->work_mutex);
1470 if (dm_suspended(ti)) {
1476 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
1477 r = queue_if_no_path(m, 1, 0);
1479 } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
1480 r = queue_if_no_path(m, 0, 0);
1486 DMWARN("Unrecognised multipath message received.");
1490 if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
1491 r = bypass_pg_num(m, argv[1], 1);
1493 } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
1494 r = bypass_pg_num(m, argv[1], 0);
1496 } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
1497 r = switch_pg_num(m, argv[1]);
1499 } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1500 action = reinstate_path;
1501 else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1504 DMWARN("Unrecognised multipath message received.");
1508 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1510 DMWARN("message: error getting device %s",
1515 r = action_dev(m, dev, action);
1517 dm_put_device(ti, dev);
1520 mutex_unlock(&m->work_mutex);
1524 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1527 struct multipath *m = (struct multipath *) ti->private;
1528 struct block_device *bdev = NULL;
1530 unsigned long flags;
1533 spin_lock_irqsave(&m->lock, flags);
1535 if (!m->current_pgpath)
1536 __choose_pgpath(m, 0);
1538 if (m->current_pgpath) {
1539 bdev = m->current_pgpath->path.dev->bdev;
1540 mode = m->current_pgpath->path.dev->mode;
1548 spin_unlock_irqrestore(&m->lock, flags);
1550 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1553 static int multipath_iterate_devices(struct dm_target *ti,
1554 iterate_devices_callout_fn fn, void *data)
1556 struct multipath *m = ti->private;
1557 struct priority_group *pg;
1561 list_for_each_entry(pg, &m->priority_groups, list) {
1562 list_for_each_entry(p, &pg->pgpaths, list) {
1563 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1573 static int __pgpath_busy(struct pgpath *pgpath)
1575 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1577 return dm_underlying_device_busy(q);
1581 * We return "busy", only when we can map I/Os but underlying devices
1582 * are busy (so even if we map I/Os now, the I/Os will wait on
1583 * the underlying queue).
1584 * In other words, if we want to kill I/Os or queue them inside us
1585 * due to map unavailability, we don't return "busy". Otherwise,
1586 * dm core won't give us the I/Os and we can't do what we want.
1588 static int multipath_busy(struct dm_target *ti)
1590 int busy = 0, has_active = 0;
1591 struct multipath *m = ti->private;
1592 struct priority_group *pg;
1593 struct pgpath *pgpath;
1594 unsigned long flags;
1596 spin_lock_irqsave(&m->lock, flags);
1598 /* Guess which priority_group will be used at next mapping time */
1599 if (unlikely(!m->current_pgpath && m->next_pg))
1601 else if (likely(m->current_pg))
1605 * We don't know which pg will be used at next mapping time.
1606 * We don't call __choose_pgpath() here to avoid to trigger
1607 * pg_init just by busy checking.
1608 * So we don't know whether underlying devices we will be using
1609 * at next mapping time are busy or not. Just try mapping.
1614 * If there is one non-busy active path at least, the path selector
1615 * will be able to select it. So we consider such a pg as not busy.
1618 list_for_each_entry(pgpath, &pg->pgpaths, list)
1619 if (pgpath->is_active) {
1622 if (!__pgpath_busy(pgpath)) {
1630 * No active path in this pg, so this pg won't be used and
1631 * the current_pg will be changed at next mapping time.
1632 * We need to try mapping to determine it.
1637 spin_unlock_irqrestore(&m->lock, flags);
1642 /*-----------------------------------------------------------------
1644 *---------------------------------------------------------------*/
1645 static struct target_type multipath_target = {
1646 .name = "multipath",
1647 .version = {1, 1, 1},
1648 .module = THIS_MODULE,
1649 .ctr = multipath_ctr,
1650 .dtr = multipath_dtr,
1651 .map_rq = multipath_map,
1652 .rq_end_io = multipath_end_io,
1653 .presuspend = multipath_presuspend,
1654 .postsuspend = multipath_postsuspend,
1655 .resume = multipath_resume,
1656 .status = multipath_status,
1657 .message = multipath_message,
1658 .ioctl = multipath_ioctl,
1659 .iterate_devices = multipath_iterate_devices,
1660 .busy = multipath_busy,
1663 static int __init dm_multipath_init(void)
1667 /* allocate a slab for the dm_ios */
1668 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1672 r = dm_register_target(&multipath_target);
1674 DMERR("register failed %d", r);
1675 kmem_cache_destroy(_mpio_cache);
1679 kmultipathd = create_workqueue("kmpathd");
1681 DMERR("failed to create workqueue kmpathd");
1682 dm_unregister_target(&multipath_target);
1683 kmem_cache_destroy(_mpio_cache);
1688 * A separate workqueue is used to handle the device handlers
1689 * to avoid overloading existing workqueue. Overloading the
1690 * old workqueue would also create a bottleneck in the
1691 * path of the storage hardware device activation.
1693 kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1694 if (!kmpath_handlerd) {
1695 DMERR("failed to create workqueue kmpath_handlerd");
1696 destroy_workqueue(kmultipathd);
1697 dm_unregister_target(&multipath_target);
1698 kmem_cache_destroy(_mpio_cache);
1702 DMINFO("version %u.%u.%u loaded",
1703 multipath_target.version[0], multipath_target.version[1],
1704 multipath_target.version[2]);
1709 static void __exit dm_multipath_exit(void)
1711 destroy_workqueue(kmpath_handlerd);
1712 destroy_workqueue(kmultipathd);
1714 dm_unregister_target(&multipath_target);
1715 kmem_cache_destroy(_mpio_cache);
1718 module_init(dm_multipath_init);
1719 module_exit(dm_multipath_exit);
1721 MODULE_DESCRIPTION(DM_NAME " multipath target");
1722 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1723 MODULE_LICENSE("GPL");