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>
11 #include "dm-path-selector.h"
12 #include "dm-uevent.h"
14 #include <linux/ctype.h>
15 #include <linux/init.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/pagemap.h>
19 #include <linux/slab.h>
20 #include <linux/time.h>
21 #include <linux/workqueue.h>
22 #include <linux/delay.h>
23 #include <scsi/scsi_dh.h>
24 #include <linux/atomic.h>
26 #define DM_MSG_PREFIX "multipath"
27 #define DM_PG_INIT_DELAY_MSECS 2000
28 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
32 struct list_head list;
34 struct priority_group *pg; /* Owning PG */
35 unsigned is_active; /* Path status */
36 unsigned fail_count; /* Cumulative failure count */
39 struct delayed_work activate_path;
42 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
45 * Paths are grouped into Priority Groups and numbered from 1 upwards.
46 * Each has a path selector which controls which path gets used.
48 struct priority_group {
49 struct list_head list;
51 struct multipath *m; /* Owning multipath instance */
52 struct path_selector ps;
54 unsigned pg_num; /* Reference number */
55 unsigned bypassed; /* Temporarily bypass this PG? */
57 unsigned nr_pgpaths; /* Number of paths in PG */
58 struct list_head pgpaths;
61 /* Multipath context */
63 struct list_head list;
66 const char *hw_handler_name;
67 char *hw_handler_params;
71 unsigned nr_priority_groups;
72 struct list_head priority_groups;
74 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
76 unsigned pg_init_required; /* pg_init needs calling? */
77 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
78 unsigned pg_init_delay_retry; /* Delay pg_init retry? */
80 unsigned nr_valid_paths; /* Total number of usable paths */
81 struct pgpath *current_pgpath;
82 struct priority_group *current_pg;
83 struct priority_group *next_pg; /* Switch to this PG if set */
84 unsigned repeat_count; /* I/Os left before calling PS again */
86 unsigned queue_io:1; /* Must we queue all I/O? */
87 unsigned queue_if_no_path:1; /* Queue I/O if last path fails? */
88 unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
89 unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
90 unsigned pg_init_disabled:1; /* pg_init is not currently allowed */
92 unsigned pg_init_retries; /* Number of times to retry pg_init */
93 unsigned pg_init_count; /* Number of times pg_init called */
94 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
96 struct work_struct trigger_event;
99 * We must use a mempool of dm_mpath_io structs so that we
100 * can resubmit bios on error.
102 mempool_t *mpio_pool;
104 struct mutex work_mutex;
108 * Context information attached to each bio we process.
111 struct pgpath *pgpath;
115 typedef int (*action_fn) (struct pgpath *pgpath);
117 static struct kmem_cache *_mpio_cache;
119 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
120 static void trigger_event(struct work_struct *work);
121 static void activate_path(struct work_struct *work);
122 static int __pgpath_busy(struct pgpath *pgpath);
125 /*-----------------------------------------------
126 * Allocation routines
127 *-----------------------------------------------*/
129 static struct pgpath *alloc_pgpath(void)
131 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
134 pgpath->is_active = 1;
135 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
141 static void free_pgpath(struct pgpath *pgpath)
146 static struct priority_group *alloc_priority_group(void)
148 struct priority_group *pg;
150 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
153 INIT_LIST_HEAD(&pg->pgpaths);
158 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
160 struct pgpath *pgpath, *tmp;
161 struct multipath *m = ti->private;
163 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
164 list_del(&pgpath->list);
165 if (m->hw_handler_name)
166 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
167 dm_put_device(ti, pgpath->path.dev);
172 static void free_priority_group(struct priority_group *pg,
173 struct dm_target *ti)
175 struct path_selector *ps = &pg->ps;
178 ps->type->destroy(ps);
179 dm_put_path_selector(ps->type);
182 free_pgpaths(&pg->pgpaths, ti);
186 static struct multipath *alloc_multipath(struct dm_target *ti)
189 unsigned min_ios = dm_get_reserved_rq_based_ios();
191 m = kzalloc(sizeof(*m), GFP_KERNEL);
193 INIT_LIST_HEAD(&m->priority_groups);
194 spin_lock_init(&m->lock);
196 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
197 INIT_WORK(&m->trigger_event, trigger_event);
198 init_waitqueue_head(&m->pg_init_wait);
199 mutex_init(&m->work_mutex);
200 m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
212 static void free_multipath(struct multipath *m)
214 struct priority_group *pg, *tmp;
216 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
218 free_priority_group(pg, m->ti);
221 kfree(m->hw_handler_name);
222 kfree(m->hw_handler_params);
223 mempool_destroy(m->mpio_pool);
227 static int set_mapinfo(struct multipath *m, union map_info *info)
229 struct dm_mpath_io *mpio;
231 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
235 memset(mpio, 0, sizeof(*mpio));
241 static void clear_mapinfo(struct multipath *m, union map_info *info)
243 struct dm_mpath_io *mpio = info->ptr;
246 mempool_free(mpio, m->mpio_pool);
249 /*-----------------------------------------------
251 *-----------------------------------------------*/
253 static int __pg_init_all_paths(struct multipath *m)
255 struct pgpath *pgpath;
256 unsigned long pg_init_delay = 0;
258 if (m->pg_init_in_progress || m->pg_init_disabled)
262 m->pg_init_required = 0;
264 /* Check here to reset pg_init_required */
268 if (m->pg_init_delay_retry)
269 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
270 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
271 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
272 /* Skip failed paths */
273 if (!pgpath->is_active)
275 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
277 m->pg_init_in_progress++;
279 return m->pg_init_in_progress;
282 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
284 m->current_pg = pgpath->pg;
286 /* Must we initialise the PG first, and queue I/O till it's ready? */
287 if (m->hw_handler_name) {
288 m->pg_init_required = 1;
291 m->pg_init_required = 0;
295 m->pg_init_count = 0;
298 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
301 struct dm_path *path;
303 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
307 m->current_pgpath = path_to_pgpath(path);
309 if (m->current_pg != pg)
310 __switch_pg(m, m->current_pgpath);
315 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
317 struct priority_group *pg;
318 unsigned bypassed = 1;
320 if (!m->nr_valid_paths)
323 /* Were we instructed to switch PG? */
327 if (!__choose_path_in_pg(m, pg, nr_bytes))
331 /* Don't change PG until it has no remaining paths */
332 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
336 * Loop through priority groups until we find a valid path.
337 * First time we skip PGs marked 'bypassed'.
338 * Second time we only try the ones we skipped, but set
339 * pg_init_delay_retry so we do not hammer controllers.
342 list_for_each_entry(pg, &m->priority_groups, list) {
343 if (pg->bypassed == bypassed)
345 if (!__choose_path_in_pg(m, pg, nr_bytes)) {
347 m->pg_init_delay_retry = 1;
351 } while (bypassed--);
354 m->current_pgpath = NULL;
355 m->current_pg = NULL;
359 * Check whether bios must be queued in the device-mapper core rather
360 * than here in the target.
362 * m->lock must be held on entry.
364 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
365 * same value then we are not between multipath_presuspend()
366 * and multipath_resume() calls and we have no need to check
367 * for the DMF_NOFLUSH_SUSPENDING flag.
369 static int __must_push_back(struct multipath *m)
371 return (m->queue_if_no_path ||
372 (m->queue_if_no_path != m->saved_queue_if_no_path &&
373 dm_noflush_suspending(m->ti)));
377 * Map cloned requests
379 static int multipath_map(struct dm_target *ti, struct request *clone,
380 union map_info *map_context)
382 struct multipath *m = (struct multipath *) ti->private;
383 int r = DM_MAPIO_REQUEUE;
384 size_t nr_bytes = blk_rq_bytes(clone);
386 struct pgpath *pgpath;
387 struct block_device *bdev;
388 struct dm_mpath_io *mpio;
390 spin_lock_irqsave(&m->lock, flags);
392 /* Do we need to select a new pgpath? */
393 if (!m->current_pgpath ||
394 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
395 __choose_pgpath(m, nr_bytes);
397 pgpath = m->current_pgpath;
400 if (!__must_push_back(m))
401 r = -EIO; /* Failed */
403 } else if (m->queue_io || m->pg_init_required) {
404 __pg_init_all_paths(m);
408 if (set_mapinfo(m, map_context) < 0)
409 /* ENOMEM, requeue */
412 bdev = pgpath->path.dev->bdev;
413 clone->q = bdev_get_queue(bdev);
414 clone->rq_disk = bdev->bd_disk;
415 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
416 mpio = map_context->ptr;
417 mpio->pgpath = pgpath;
418 mpio->nr_bytes = nr_bytes;
419 if (pgpath->pg->ps.type->start_io)
420 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
423 r = DM_MAPIO_REMAPPED;
426 spin_unlock_irqrestore(&m->lock, flags);
432 * If we run out of usable paths, should we queue I/O or error it?
434 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
435 unsigned save_old_value)
439 spin_lock_irqsave(&m->lock, flags);
442 m->saved_queue_if_no_path = m->queue_if_no_path;
444 m->saved_queue_if_no_path = queue_if_no_path;
445 m->queue_if_no_path = queue_if_no_path;
446 spin_unlock_irqrestore(&m->lock, flags);
448 if (!queue_if_no_path)
449 dm_table_run_md_queue_async(m->ti->table);
455 * An event is triggered whenever a path is taken out of use.
456 * Includes path failure and PG bypass.
458 static void trigger_event(struct work_struct *work)
460 struct multipath *m =
461 container_of(work, struct multipath, trigger_event);
463 dm_table_event(m->ti->table);
466 /*-----------------------------------------------------------------
467 * Constructor/argument parsing:
468 * <#multipath feature args> [<arg>]*
469 * <#hw_handler args> [hw_handler [<arg>]*]
471 * <initial priority group>
472 * [<selector> <#selector args> [<arg>]*
473 * <#paths> <#per-path selector args>
474 * [<path> [<arg>]* ]+ ]+
475 *---------------------------------------------------------------*/
476 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
477 struct dm_target *ti)
480 struct path_selector_type *pst;
483 static struct dm_arg _args[] = {
484 {0, 1024, "invalid number of path selector args"},
487 pst = dm_get_path_selector(dm_shift_arg(as));
489 ti->error = "unknown path selector type";
493 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
495 dm_put_path_selector(pst);
499 r = pst->create(&pg->ps, ps_argc, as->argv);
501 dm_put_path_selector(pst);
502 ti->error = "path selector constructor failed";
507 dm_consume_args(as, ps_argc);
512 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
513 struct dm_target *ti)
517 struct multipath *m = ti->private;
518 struct request_queue *q = NULL;
519 const char *attached_handler_name;
521 /* we need at least a path arg */
523 ti->error = "no device given";
524 return ERR_PTR(-EINVAL);
529 return ERR_PTR(-ENOMEM);
531 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
534 ti->error = "error getting device";
538 if (m->retain_attached_hw_handler || m->hw_handler_name)
539 q = bdev_get_queue(p->path.dev->bdev);
541 if (m->retain_attached_hw_handler) {
542 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
543 if (attached_handler_name) {
545 * Reset hw_handler_name to match the attached handler
546 * and clear any hw_handler_params associated with the
549 * NB. This modifies the table line to show the actual
550 * handler instead of the original table passed in.
552 kfree(m->hw_handler_name);
553 m->hw_handler_name = attached_handler_name;
555 kfree(m->hw_handler_params);
556 m->hw_handler_params = NULL;
560 if (m->hw_handler_name) {
562 * Increments scsi_dh reference, even when using an
563 * already-attached handler.
565 r = scsi_dh_attach(q, m->hw_handler_name);
568 * Already attached to different hw_handler:
569 * try to reattach with correct one.
572 r = scsi_dh_attach(q, m->hw_handler_name);
576 ti->error = "error attaching hardware handler";
577 dm_put_device(ti, p->path.dev);
581 if (m->hw_handler_params) {
582 r = scsi_dh_set_params(q, m->hw_handler_params);
584 ti->error = "unable to set hardware "
585 "handler parameters";
587 dm_put_device(ti, p->path.dev);
593 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
595 dm_put_device(ti, p->path.dev);
606 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
609 static struct dm_arg _args[] = {
610 {1, 1024, "invalid number of paths"},
611 {0, 1024, "invalid number of selector args"}
615 unsigned i, nr_selector_args, nr_args;
616 struct priority_group *pg;
617 struct dm_target *ti = m->ti;
621 ti->error = "not enough priority group arguments";
622 return ERR_PTR(-EINVAL);
625 pg = alloc_priority_group();
627 ti->error = "couldn't allocate priority group";
628 return ERR_PTR(-ENOMEM);
632 r = parse_path_selector(as, pg, ti);
639 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
643 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
647 nr_args = 1 + nr_selector_args;
648 for (i = 0; i < pg->nr_pgpaths; i++) {
649 struct pgpath *pgpath;
650 struct dm_arg_set path_args;
652 if (as->argc < nr_args) {
653 ti->error = "not enough path parameters";
658 path_args.argc = nr_args;
659 path_args.argv = as->argv;
661 pgpath = parse_path(&path_args, &pg->ps, ti);
662 if (IS_ERR(pgpath)) {
668 list_add_tail(&pgpath->list, &pg->pgpaths);
669 dm_consume_args(as, nr_args);
675 free_priority_group(pg, ti);
679 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
683 struct dm_target *ti = m->ti;
685 static struct dm_arg _args[] = {
686 {0, 1024, "invalid number of hardware handler args"},
689 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
695 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
696 if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
697 "scsi_dh_%s", m->hw_handler_name)) {
698 ti->error = "unknown hardware handler type";
707 for (i = 0; i <= hw_argc - 2; i++)
708 len += strlen(as->argv[i]) + 1;
709 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
711 ti->error = "memory allocation failed";
715 j = sprintf(p, "%d", hw_argc - 1);
716 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
717 j = sprintf(p, "%s", as->argv[i]);
719 dm_consume_args(as, hw_argc - 1);
723 kfree(m->hw_handler_name);
724 m->hw_handler_name = NULL;
728 static int parse_features(struct dm_arg_set *as, struct multipath *m)
732 struct dm_target *ti = m->ti;
733 const char *arg_name;
735 static struct dm_arg _args[] = {
736 {0, 6, "invalid number of feature args"},
737 {1, 50, "pg_init_retries must be between 1 and 50"},
738 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
741 r = dm_read_arg_group(_args, as, &argc, &ti->error);
749 arg_name = dm_shift_arg(as);
752 if (!strcasecmp(arg_name, "queue_if_no_path")) {
753 r = queue_if_no_path(m, 1, 0);
757 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
758 m->retain_attached_hw_handler = 1;
762 if (!strcasecmp(arg_name, "pg_init_retries") &&
764 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
769 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
771 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
776 ti->error = "Unrecognised multipath feature request";
778 } while (argc && !r);
783 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
786 /* target arguments */
787 static struct dm_arg _args[] = {
788 {0, 1024, "invalid number of priority groups"},
789 {0, 1024, "invalid initial priority group number"},
794 struct dm_arg_set as;
795 unsigned pg_count = 0;
796 unsigned next_pg_num;
801 m = alloc_multipath(ti);
803 ti->error = "can't allocate multipath";
807 r = parse_features(&as, m);
811 r = parse_hw_handler(&as, m);
815 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
819 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
823 if ((!m->nr_priority_groups && next_pg_num) ||
824 (m->nr_priority_groups && !next_pg_num)) {
825 ti->error = "invalid initial priority group";
830 /* parse the priority groups */
832 struct priority_group *pg;
834 pg = parse_priority_group(&as, m);
840 m->nr_valid_paths += pg->nr_pgpaths;
841 list_add_tail(&pg->list, &m->priority_groups);
843 pg->pg_num = pg_count;
848 if (pg_count != m->nr_priority_groups) {
849 ti->error = "priority group count mismatch";
854 ti->num_flush_bios = 1;
855 ti->num_discard_bios = 1;
856 ti->num_write_same_bios = 1;
865 static void multipath_wait_for_pg_init_completion(struct multipath *m)
867 DECLARE_WAITQUEUE(wait, current);
870 add_wait_queue(&m->pg_init_wait, &wait);
873 set_current_state(TASK_UNINTERRUPTIBLE);
875 spin_lock_irqsave(&m->lock, flags);
876 if (!m->pg_init_in_progress) {
877 spin_unlock_irqrestore(&m->lock, flags);
880 spin_unlock_irqrestore(&m->lock, flags);
884 set_current_state(TASK_RUNNING);
886 remove_wait_queue(&m->pg_init_wait, &wait);
889 static void flush_multipath_work(struct multipath *m)
893 spin_lock_irqsave(&m->lock, flags);
894 m->pg_init_disabled = 1;
895 spin_unlock_irqrestore(&m->lock, flags);
897 flush_workqueue(kmpath_handlerd);
898 multipath_wait_for_pg_init_completion(m);
899 flush_workqueue(kmultipathd);
900 flush_work(&m->trigger_event);
902 spin_lock_irqsave(&m->lock, flags);
903 m->pg_init_disabled = 0;
904 spin_unlock_irqrestore(&m->lock, flags);
907 static void multipath_dtr(struct dm_target *ti)
909 struct multipath *m = ti->private;
911 flush_multipath_work(m);
916 * Take a path out of use.
918 static int fail_path(struct pgpath *pgpath)
921 struct multipath *m = pgpath->pg->m;
923 spin_lock_irqsave(&m->lock, flags);
925 if (!pgpath->is_active)
928 DMWARN("Failing path %s.", pgpath->path.dev->name);
930 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
931 pgpath->is_active = 0;
932 pgpath->fail_count++;
936 if (pgpath == m->current_pgpath)
937 m->current_pgpath = NULL;
939 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
940 pgpath->path.dev->name, m->nr_valid_paths);
942 schedule_work(&m->trigger_event);
945 spin_unlock_irqrestore(&m->lock, flags);
951 * Reinstate a previously-failed path
953 static int reinstate_path(struct pgpath *pgpath)
955 int r = 0, run_queue = 0;
957 struct multipath *m = pgpath->pg->m;
959 spin_lock_irqsave(&m->lock, flags);
961 if (pgpath->is_active)
964 if (!pgpath->pg->ps.type->reinstate_path) {
965 DMWARN("Reinstate path not supported by path selector %s",
966 pgpath->pg->ps.type->name);
971 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
975 pgpath->is_active = 1;
977 if (!m->nr_valid_paths++) {
978 m->current_pgpath = NULL;
980 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
981 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
982 m->pg_init_in_progress++;
985 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
986 pgpath->path.dev->name, m->nr_valid_paths);
988 schedule_work(&m->trigger_event);
991 spin_unlock_irqrestore(&m->lock, flags);
993 dm_table_run_md_queue_async(m->ti->table);
999 * Fail or reinstate all paths that match the provided struct dm_dev.
1001 static int action_dev(struct multipath *m, struct dm_dev *dev,
1005 struct pgpath *pgpath;
1006 struct priority_group *pg;
1008 list_for_each_entry(pg, &m->priority_groups, list) {
1009 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1010 if (pgpath->path.dev == dev)
1019 * Temporarily try to avoid having to use the specified PG
1021 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1024 unsigned long flags;
1026 spin_lock_irqsave(&m->lock, flags);
1028 pg->bypassed = bypassed;
1029 m->current_pgpath = NULL;
1030 m->current_pg = NULL;
1032 spin_unlock_irqrestore(&m->lock, flags);
1034 schedule_work(&m->trigger_event);
1038 * Switch to using the specified PG from the next I/O that gets mapped
1040 static int switch_pg_num(struct multipath *m, const char *pgstr)
1042 struct priority_group *pg;
1044 unsigned long flags;
1047 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1048 (pgnum > m->nr_priority_groups)) {
1049 DMWARN("invalid PG number supplied to switch_pg_num");
1053 spin_lock_irqsave(&m->lock, flags);
1054 list_for_each_entry(pg, &m->priority_groups, list) {
1059 m->current_pgpath = NULL;
1060 m->current_pg = NULL;
1063 spin_unlock_irqrestore(&m->lock, flags);
1065 schedule_work(&m->trigger_event);
1070 * Set/clear bypassed status of a PG.
1071 * PGs are numbered upwards from 1 in the order they were declared.
1073 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1075 struct priority_group *pg;
1079 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1080 (pgnum > m->nr_priority_groups)) {
1081 DMWARN("invalid PG number supplied to bypass_pg");
1085 list_for_each_entry(pg, &m->priority_groups, list) {
1090 bypass_pg(m, pg, bypassed);
1095 * Should we retry pg_init immediately?
1097 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1099 unsigned long flags;
1100 int limit_reached = 0;
1102 spin_lock_irqsave(&m->lock, flags);
1104 if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1105 m->pg_init_required = 1;
1109 spin_unlock_irqrestore(&m->lock, flags);
1111 return limit_reached;
1114 static void pg_init_done(void *data, int errors)
1116 struct pgpath *pgpath = data;
1117 struct priority_group *pg = pgpath->pg;
1118 struct multipath *m = pg->m;
1119 unsigned long flags;
1120 unsigned delay_retry = 0;
1122 /* device or driver problems */
1127 if (!m->hw_handler_name) {
1131 DMERR("Could not failover the device: Handler scsi_dh_%s "
1132 "Error %d.", m->hw_handler_name, errors);
1134 * Fail path for now, so we do not ping pong
1138 case SCSI_DH_DEV_TEMP_BUSY:
1140 * Probably doing something like FW upgrade on the
1141 * controller so try the other pg.
1143 bypass_pg(m, pg, 1);
1146 /* Wait before retrying. */
1148 case SCSI_DH_IMM_RETRY:
1149 case SCSI_DH_RES_TEMP_UNAVAIL:
1150 if (pg_init_limit_reached(m, pgpath))
1156 * We probably do not want to fail the path for a device
1157 * error, but this is what the old dm did. In future
1158 * patches we can do more advanced handling.
1163 spin_lock_irqsave(&m->lock, flags);
1165 if (pgpath == m->current_pgpath) {
1166 DMERR("Could not failover device. Error %d.", errors);
1167 m->current_pgpath = NULL;
1168 m->current_pg = NULL;
1170 } else if (!m->pg_init_required)
1173 if (--m->pg_init_in_progress)
1174 /* Activations of other paths are still on going */
1177 if (m->pg_init_required) {
1178 m->pg_init_delay_retry = delay_retry;
1179 if (__pg_init_all_paths(m))
1185 * Wake up any thread waiting to suspend.
1187 wake_up(&m->pg_init_wait);
1190 spin_unlock_irqrestore(&m->lock, flags);
1193 static void activate_path(struct work_struct *work)
1195 struct pgpath *pgpath =
1196 container_of(work, struct pgpath, activate_path.work);
1198 if (pgpath->is_active)
1199 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1200 pg_init_done, pgpath);
1202 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1205 static int noretry_error(int error)
1216 /* Anything else could be a path failure, so should be retried */
1223 static int do_end_io(struct multipath *m, struct request *clone,
1224 int error, struct dm_mpath_io *mpio)
1227 * We don't queue any clone request inside the multipath target
1228 * during end I/O handling, since those clone requests don't have
1229 * bio clones. If we queue them inside the multipath target,
1230 * we need to make bio clones, that requires memory allocation.
1231 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1232 * don't have bio clones.)
1233 * Instead of queueing the clone request here, we queue the original
1234 * request into dm core, which will remake a clone request and
1235 * clone bios for it and resubmit it later.
1237 int r = DM_ENDIO_REQUEUE;
1238 unsigned long flags;
1240 if (!error && !clone->errors)
1241 return 0; /* I/O complete */
1243 if (noretry_error(error))
1247 fail_path(mpio->pgpath);
1249 spin_lock_irqsave(&m->lock, flags);
1250 if (!m->nr_valid_paths) {
1251 if (!m->queue_if_no_path) {
1252 if (!__must_push_back(m))
1255 if (error == -EBADE)
1259 spin_unlock_irqrestore(&m->lock, flags);
1264 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1265 int error, union map_info *map_context)
1267 struct multipath *m = ti->private;
1268 struct dm_mpath_io *mpio = map_context->ptr;
1269 struct pgpath *pgpath;
1270 struct path_selector *ps;
1275 r = do_end_io(m, clone, error, mpio);
1276 pgpath = mpio->pgpath;
1278 ps = &pgpath->pg->ps;
1279 if (ps->type->end_io)
1280 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1282 clear_mapinfo(m, map_context);
1288 * Suspend can't complete until all the I/O is processed so if
1289 * the last path fails we must error any remaining I/O.
1290 * Note that if the freeze_bdev fails while suspending, the
1291 * queue_if_no_path state is lost - userspace should reset it.
1293 static void multipath_presuspend(struct dm_target *ti)
1295 struct multipath *m = (struct multipath *) ti->private;
1297 queue_if_no_path(m, 0, 1);
1300 static void multipath_postsuspend(struct dm_target *ti)
1302 struct multipath *m = ti->private;
1304 mutex_lock(&m->work_mutex);
1305 flush_multipath_work(m);
1306 mutex_unlock(&m->work_mutex);
1310 * Restore the queue_if_no_path setting.
1312 static void multipath_resume(struct dm_target *ti)
1314 struct multipath *m = (struct multipath *) ti->private;
1315 unsigned long flags;
1317 spin_lock_irqsave(&m->lock, flags);
1318 m->queue_if_no_path = m->saved_queue_if_no_path;
1319 spin_unlock_irqrestore(&m->lock, flags);
1323 * Info output has the following format:
1324 * num_multipath_feature_args [multipath_feature_args]*
1325 * num_handler_status_args [handler_status_args]*
1326 * num_groups init_group_number
1327 * [A|D|E num_ps_status_args [ps_status_args]*
1328 * num_paths num_selector_args
1329 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1331 * Table output has the following format (identical to the constructor string):
1332 * num_feature_args [features_args]*
1333 * num_handler_args hw_handler [hw_handler_args]*
1334 * num_groups init_group_number
1335 * [priority selector-name num_ps_args [ps_args]*
1336 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1338 static void multipath_status(struct dm_target *ti, status_type_t type,
1339 unsigned status_flags, char *result, unsigned maxlen)
1342 unsigned long flags;
1343 struct multipath *m = (struct multipath *) ti->private;
1344 struct priority_group *pg;
1349 spin_lock_irqsave(&m->lock, flags);
1352 if (type == STATUSTYPE_INFO)
1353 DMEMIT("2 %u %u ", m->queue_io, m->pg_init_count);
1355 DMEMIT("%u ", m->queue_if_no_path +
1356 (m->pg_init_retries > 0) * 2 +
1357 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1358 m->retain_attached_hw_handler);
1359 if (m->queue_if_no_path)
1360 DMEMIT("queue_if_no_path ");
1361 if (m->pg_init_retries)
1362 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1363 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1364 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1365 if (m->retain_attached_hw_handler)
1366 DMEMIT("retain_attached_hw_handler ");
1369 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1372 DMEMIT("1 %s ", m->hw_handler_name);
1374 DMEMIT("%u ", m->nr_priority_groups);
1377 pg_num = m->next_pg->pg_num;
1378 else if (m->current_pg)
1379 pg_num = m->current_pg->pg_num;
1381 pg_num = (m->nr_priority_groups ? 1 : 0);
1383 DMEMIT("%u ", pg_num);
1386 case STATUSTYPE_INFO:
1387 list_for_each_entry(pg, &m->priority_groups, list) {
1389 state = 'D'; /* Disabled */
1390 else if (pg == m->current_pg)
1391 state = 'A'; /* Currently Active */
1393 state = 'E'; /* Enabled */
1395 DMEMIT("%c ", state);
1397 if (pg->ps.type->status)
1398 sz += pg->ps.type->status(&pg->ps, NULL, type,
1404 DMEMIT("%u %u ", pg->nr_pgpaths,
1405 pg->ps.type->info_args);
1407 list_for_each_entry(p, &pg->pgpaths, list) {
1408 DMEMIT("%s %s %u ", p->path.dev->name,
1409 p->is_active ? "A" : "F",
1411 if (pg->ps.type->status)
1412 sz += pg->ps.type->status(&pg->ps,
1413 &p->path, type, result + sz,
1419 case STATUSTYPE_TABLE:
1420 list_for_each_entry(pg, &m->priority_groups, list) {
1421 DMEMIT("%s ", pg->ps.type->name);
1423 if (pg->ps.type->status)
1424 sz += pg->ps.type->status(&pg->ps, NULL, type,
1430 DMEMIT("%u %u ", pg->nr_pgpaths,
1431 pg->ps.type->table_args);
1433 list_for_each_entry(p, &pg->pgpaths, list) {
1434 DMEMIT("%s ", p->path.dev->name);
1435 if (pg->ps.type->status)
1436 sz += pg->ps.type->status(&pg->ps,
1437 &p->path, type, result + sz,
1444 spin_unlock_irqrestore(&m->lock, flags);
1447 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1451 struct multipath *m = (struct multipath *) ti->private;
1454 mutex_lock(&m->work_mutex);
1456 if (dm_suspended(ti)) {
1462 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1463 r = queue_if_no_path(m, 1, 0);
1465 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1466 r = queue_if_no_path(m, 0, 0);
1472 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1476 if (!strcasecmp(argv[0], "disable_group")) {
1477 r = bypass_pg_num(m, argv[1], 1);
1479 } else if (!strcasecmp(argv[0], "enable_group")) {
1480 r = bypass_pg_num(m, argv[1], 0);
1482 } else if (!strcasecmp(argv[0], "switch_group")) {
1483 r = switch_pg_num(m, argv[1]);
1485 } else if (!strcasecmp(argv[0], "reinstate_path"))
1486 action = reinstate_path;
1487 else if (!strcasecmp(argv[0], "fail_path"))
1490 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1494 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1496 DMWARN("message: error getting device %s",
1501 r = action_dev(m, dev, action);
1503 dm_put_device(ti, dev);
1506 mutex_unlock(&m->work_mutex);
1510 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1513 struct multipath *m = ti->private;
1514 struct pgpath *pgpath;
1515 struct block_device *bdev;
1517 unsigned long flags;
1524 spin_lock_irqsave(&m->lock, flags);
1526 if (!m->current_pgpath)
1527 __choose_pgpath(m, 0);
1529 pgpath = m->current_pgpath;
1532 bdev = pgpath->path.dev->bdev;
1533 mode = pgpath->path.dev->mode;
1536 if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1541 spin_unlock_irqrestore(&m->lock, flags);
1544 * Only pass ioctls through if the device sizes match exactly.
1546 if (!bdev || ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) {
1547 int err = scsi_verify_blk_ioctl(NULL, cmd);
1552 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
1553 spin_lock_irqsave(&m->lock, flags);
1554 if (!m->current_pg) {
1555 /* Path status changed, redo selection */
1556 __choose_pgpath(m, 0);
1558 if (m->pg_init_required)
1559 __pg_init_all_paths(m);
1560 spin_unlock_irqrestore(&m->lock, flags);
1561 dm_table_run_md_queue_async(m->ti->table);
1564 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1567 static int multipath_iterate_devices(struct dm_target *ti,
1568 iterate_devices_callout_fn fn, void *data)
1570 struct multipath *m = ti->private;
1571 struct priority_group *pg;
1575 list_for_each_entry(pg, &m->priority_groups, list) {
1576 list_for_each_entry(p, &pg->pgpaths, list) {
1577 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1587 static int __pgpath_busy(struct pgpath *pgpath)
1589 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1591 return dm_underlying_device_busy(q);
1595 * We return "busy", only when we can map I/Os but underlying devices
1596 * are busy (so even if we map I/Os now, the I/Os will wait on
1597 * the underlying queue).
1598 * In other words, if we want to kill I/Os or queue them inside us
1599 * due to map unavailability, we don't return "busy". Otherwise,
1600 * dm core won't give us the I/Os and we can't do what we want.
1602 static int multipath_busy(struct dm_target *ti)
1604 int busy = 0, has_active = 0;
1605 struct multipath *m = ti->private;
1606 struct priority_group *pg;
1607 struct pgpath *pgpath;
1608 unsigned long flags;
1610 spin_lock_irqsave(&m->lock, flags);
1612 /* pg_init in progress or no paths available */
1613 if (m->pg_init_in_progress ||
1614 (!m->nr_valid_paths && m->queue_if_no_path)) {
1618 /* Guess which priority_group will be used at next mapping time */
1619 if (unlikely(!m->current_pgpath && m->next_pg))
1621 else if (likely(m->current_pg))
1625 * We don't know which pg will be used at next mapping time.
1626 * We don't call __choose_pgpath() here to avoid to trigger
1627 * pg_init just by busy checking.
1628 * So we don't know whether underlying devices we will be using
1629 * at next mapping time are busy or not. Just try mapping.
1634 * If there is one non-busy active path at least, the path selector
1635 * will be able to select it. So we consider such a pg as not busy.
1638 list_for_each_entry(pgpath, &pg->pgpaths, list)
1639 if (pgpath->is_active) {
1642 if (!__pgpath_busy(pgpath)) {
1650 * No active path in this pg, so this pg won't be used and
1651 * the current_pg will be changed at next mapping time.
1652 * We need to try mapping to determine it.
1657 spin_unlock_irqrestore(&m->lock, flags);
1662 /*-----------------------------------------------------------------
1664 *---------------------------------------------------------------*/
1665 static struct target_type multipath_target = {
1666 .name = "multipath",
1667 .version = {1, 7, 0},
1668 .module = THIS_MODULE,
1669 .ctr = multipath_ctr,
1670 .dtr = multipath_dtr,
1671 .map_rq = multipath_map,
1672 .rq_end_io = multipath_end_io,
1673 .presuspend = multipath_presuspend,
1674 .postsuspend = multipath_postsuspend,
1675 .resume = multipath_resume,
1676 .status = multipath_status,
1677 .message = multipath_message,
1678 .ioctl = multipath_ioctl,
1679 .iterate_devices = multipath_iterate_devices,
1680 .busy = multipath_busy,
1683 static int __init dm_multipath_init(void)
1687 /* allocate a slab for the dm_ios */
1688 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1692 r = dm_register_target(&multipath_target);
1694 DMERR("register failed %d", r);
1695 kmem_cache_destroy(_mpio_cache);
1699 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1701 DMERR("failed to create workqueue kmpathd");
1702 dm_unregister_target(&multipath_target);
1703 kmem_cache_destroy(_mpio_cache);
1708 * A separate workqueue is used to handle the device handlers
1709 * to avoid overloading existing workqueue. Overloading the
1710 * old workqueue would also create a bottleneck in the
1711 * path of the storage hardware device activation.
1713 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1715 if (!kmpath_handlerd) {
1716 DMERR("failed to create workqueue kmpath_handlerd");
1717 destroy_workqueue(kmultipathd);
1718 dm_unregister_target(&multipath_target);
1719 kmem_cache_destroy(_mpio_cache);
1723 DMINFO("version %u.%u.%u loaded",
1724 multipath_target.version[0], multipath_target.version[1],
1725 multipath_target.version[2]);
1730 static void __exit dm_multipath_exit(void)
1732 destroy_workqueue(kmpath_handlerd);
1733 destroy_workqueue(kmultipathd);
1735 dm_unregister_target(&multipath_target);
1736 kmem_cache_destroy(_mpio_cache);
1739 module_init(dm_multipath_init);
1740 module_exit(dm_multipath_exit);
1742 MODULE_DESCRIPTION(DM_NAME " multipath target");
1743 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1744 MODULE_LICENSE("GPL");