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 <linux/delay.h>
22 #include <scsi/scsi_dh.h>
23 #include <linux/atomic.h>
25 #define DM_MSG_PREFIX "multipath"
26 #define DM_PG_INIT_DELAY_MSECS 2000
27 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
31 struct list_head list;
33 struct priority_group *pg; /* Owning PG */
34 unsigned is_active; /* Path status */
35 unsigned fail_count; /* Cumulative failure count */
38 struct delayed_work activate_path;
41 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
44 * Paths are grouped into Priority Groups and numbered from 1 upwards.
45 * Each has a path selector which controls which path gets used.
47 struct priority_group {
48 struct list_head list;
50 struct multipath *m; /* Owning multipath instance */
51 struct path_selector ps;
53 unsigned pg_num; /* Reference number */
54 unsigned bypassed; /* Temporarily bypass this PG? */
56 unsigned nr_pgpaths; /* Number of paths in PG */
57 struct list_head pgpaths;
60 /* Multipath context */
62 struct list_head list;
65 const char *hw_handler_name;
66 char *hw_handler_params;
70 unsigned nr_priority_groups;
71 struct list_head priority_groups;
73 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
75 unsigned pg_init_required; /* pg_init needs calling? */
76 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
77 unsigned pg_init_delay_retry; /* Delay pg_init retry? */
79 unsigned nr_valid_paths; /* Total number of usable paths */
80 struct pgpath *current_pgpath;
81 struct priority_group *current_pg;
82 struct priority_group *next_pg; /* Switch to this PG if set */
83 unsigned repeat_count; /* I/Os left before calling PS again */
85 unsigned queue_io:1; /* Must we queue all I/O? */
86 unsigned queue_if_no_path:1; /* Queue I/O if last path fails? */
87 unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
88 unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
89 unsigned pg_init_disabled:1; /* pg_init is not currently allowed */
91 unsigned pg_init_retries; /* Number of times to retry pg_init */
92 unsigned pg_init_count; /* Number of times pg_init called */
93 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
96 struct work_struct process_queued_ios;
97 struct list_head queued_ios;
99 struct work_struct trigger_event;
102 * We must use a mempool of dm_mpath_io structs so that we
103 * can resubmit bios on error.
105 mempool_t *mpio_pool;
107 struct mutex work_mutex;
111 * Context information attached to each bio we process.
114 struct pgpath *pgpath;
118 typedef int (*action_fn) (struct pgpath *pgpath);
120 #define MIN_IOS 256 /* Mempool size */
122 static struct kmem_cache *_mpio_cache;
124 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
125 static void process_queued_ios(struct work_struct *work);
126 static void trigger_event(struct work_struct *work);
127 static void activate_path(struct work_struct *work);
130 /*-----------------------------------------------
131 * Allocation routines
132 *-----------------------------------------------*/
134 static struct pgpath *alloc_pgpath(void)
136 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
139 pgpath->is_active = 1;
140 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
146 static void free_pgpath(struct pgpath *pgpath)
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 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
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);
233 static int set_mapinfo(struct multipath *m, union map_info *info)
235 struct dm_mpath_io *mpio;
237 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
241 memset(mpio, 0, sizeof(*mpio));
247 static void clear_mapinfo(struct multipath *m, union map_info *info)
249 struct dm_mpath_io *mpio = info->ptr;
252 mempool_free(mpio, m->mpio_pool);
255 /*-----------------------------------------------
257 *-----------------------------------------------*/
259 static void __pg_init_all_paths(struct multipath *m)
261 struct pgpath *pgpath;
262 unsigned long pg_init_delay = 0;
265 m->pg_init_required = 0;
266 if (m->pg_init_delay_retry)
267 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
268 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
269 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
270 /* Skip failed paths */
271 if (!pgpath->is_active)
273 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
275 m->pg_init_in_progress++;
279 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
281 m->current_pg = pgpath->pg;
283 /* Must we initialise the PG first, and queue I/O till it's ready? */
284 if (m->hw_handler_name) {
285 m->pg_init_required = 1;
288 m->pg_init_required = 0;
292 m->pg_init_count = 0;
295 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
298 struct dm_path *path;
300 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
304 m->current_pgpath = path_to_pgpath(path);
306 if (m->current_pg != pg)
307 __switch_pg(m, m->current_pgpath);
312 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
314 struct priority_group *pg;
315 unsigned bypassed = 1;
317 if (!m->nr_valid_paths)
320 /* Were we instructed to switch PG? */
324 if (!__choose_path_in_pg(m, pg, nr_bytes))
328 /* Don't change PG until it has no remaining paths */
329 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
333 * Loop through priority groups until we find a valid path.
334 * First time we skip PGs marked 'bypassed'.
335 * Second time we only try the ones we skipped, but set
336 * pg_init_delay_retry so we do not hammer controllers.
339 list_for_each_entry(pg, &m->priority_groups, list) {
340 if (pg->bypassed == bypassed)
342 if (!__choose_path_in_pg(m, pg, nr_bytes)) {
344 m->pg_init_delay_retry = 1;
348 } while (bypassed--);
351 m->current_pgpath = NULL;
352 m->current_pg = NULL;
356 * Check whether bios must be queued in the device-mapper core rather
357 * than here in the target.
359 * m->lock must be held on entry.
361 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
362 * same value then we are not between multipath_presuspend()
363 * and multipath_resume() calls and we have no need to check
364 * for the DMF_NOFLUSH_SUSPENDING flag.
366 static int __must_push_back(struct multipath *m)
368 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
369 dm_noflush_suspending(m->ti));
372 static int map_io(struct multipath *m, struct request *clone,
373 union map_info *map_context, unsigned was_queued)
375 int r = DM_MAPIO_REMAPPED;
376 size_t nr_bytes = blk_rq_bytes(clone);
378 struct pgpath *pgpath;
379 struct block_device *bdev;
380 struct dm_mpath_io *mpio = map_context->ptr;
382 spin_lock_irqsave(&m->lock, flags);
384 /* Do we need to select a new pgpath? */
385 if (!m->current_pgpath ||
386 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
387 __choose_pgpath(m, nr_bytes);
389 pgpath = m->current_pgpath;
394 if ((pgpath && m->queue_io) ||
395 (!pgpath && m->queue_if_no_path)) {
396 /* Queue for the daemon to resubmit */
397 list_add_tail(&clone->queuelist, &m->queued_ios);
399 if ((m->pg_init_required && !m->pg_init_in_progress) ||
401 queue_work(kmultipathd, &m->process_queued_ios);
403 r = DM_MAPIO_SUBMITTED;
405 bdev = pgpath->path.dev->bdev;
406 clone->q = bdev_get_queue(bdev);
407 clone->rq_disk = bdev->bd_disk;
408 } else if (__must_push_back(m))
409 r = DM_MAPIO_REQUEUE;
411 r = -EIO; /* Failed */
413 mpio->pgpath = pgpath;
414 mpio->nr_bytes = nr_bytes;
416 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
417 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
420 spin_unlock_irqrestore(&m->lock, flags);
426 * If we run out of usable paths, should we queue I/O or error it?
428 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
429 unsigned save_old_value)
433 spin_lock_irqsave(&m->lock, flags);
436 m->saved_queue_if_no_path = m->queue_if_no_path;
438 m->saved_queue_if_no_path = queue_if_no_path;
439 m->queue_if_no_path = queue_if_no_path;
440 if (!m->queue_if_no_path && m->queue_size)
441 queue_work(kmultipathd, &m->process_queued_ios);
443 spin_unlock_irqrestore(&m->lock, flags);
448 /*-----------------------------------------------------------------
449 * The multipath daemon is responsible for resubmitting queued ios.
450 *---------------------------------------------------------------*/
452 static void dispatch_queued_ios(struct multipath *m)
456 union map_info *info;
457 struct request *clone, *n;
460 spin_lock_irqsave(&m->lock, flags);
461 list_splice_init(&m->queued_ios, &cl);
462 spin_unlock_irqrestore(&m->lock, flags);
464 list_for_each_entry_safe(clone, n, &cl, queuelist) {
465 list_del_init(&clone->queuelist);
467 info = dm_get_rq_mapinfo(clone);
469 r = map_io(m, clone, info, 1);
471 clear_mapinfo(m, info);
472 dm_kill_unmapped_request(clone, r);
473 } else if (r == DM_MAPIO_REMAPPED)
474 dm_dispatch_request(clone);
475 else if (r == DM_MAPIO_REQUEUE) {
476 clear_mapinfo(m, info);
477 dm_requeue_unmapped_request(clone);
482 static void process_queued_ios(struct work_struct *work)
484 struct multipath *m =
485 container_of(work, struct multipath, process_queued_ios);
486 struct pgpath *pgpath = NULL;
487 unsigned must_queue = 1;
490 spin_lock_irqsave(&m->lock, flags);
492 if (!m->current_pgpath)
493 __choose_pgpath(m, 0);
495 pgpath = m->current_pgpath;
497 if ((pgpath && !m->queue_io) ||
498 (!pgpath && !m->queue_if_no_path))
501 if (m->pg_init_required && !m->pg_init_in_progress && pgpath &&
502 !m->pg_init_disabled)
503 __pg_init_all_paths(m);
505 spin_unlock_irqrestore(&m->lock, flags);
507 dispatch_queued_ios(m);
511 * An event is triggered whenever a path is taken out of use.
512 * Includes path failure and PG bypass.
514 static void trigger_event(struct work_struct *work)
516 struct multipath *m =
517 container_of(work, struct multipath, trigger_event);
519 dm_table_event(m->ti->table);
522 /*-----------------------------------------------------------------
523 * Constructor/argument parsing:
524 * <#multipath feature args> [<arg>]*
525 * <#hw_handler args> [hw_handler [<arg>]*]
527 * <initial priority group>
528 * [<selector> <#selector args> [<arg>]*
529 * <#paths> <#per-path selector args>
530 * [<path> [<arg>]* ]+ ]+
531 *---------------------------------------------------------------*/
532 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
533 struct dm_target *ti)
536 struct path_selector_type *pst;
539 static struct dm_arg _args[] = {
540 {0, 1024, "invalid number of path selector args"},
543 pst = dm_get_path_selector(dm_shift_arg(as));
545 ti->error = "unknown path selector type";
549 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
551 dm_put_path_selector(pst);
555 r = pst->create(&pg->ps, ps_argc, as->argv);
557 dm_put_path_selector(pst);
558 ti->error = "path selector constructor failed";
563 dm_consume_args(as, ps_argc);
568 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
569 struct dm_target *ti)
573 struct multipath *m = ti->private;
574 struct request_queue *q = NULL;
575 const char *attached_handler_name;
577 /* we need at least a path arg */
579 ti->error = "no device given";
580 return ERR_PTR(-EINVAL);
585 return ERR_PTR(-ENOMEM);
587 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
590 ti->error = "error getting device";
594 if (m->retain_attached_hw_handler || m->hw_handler_name)
595 q = bdev_get_queue(p->path.dev->bdev);
597 if (m->retain_attached_hw_handler) {
598 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
599 if (attached_handler_name) {
601 * Reset hw_handler_name to match the attached handler
602 * and clear any hw_handler_params associated with the
605 * NB. This modifies the table line to show the actual
606 * handler instead of the original table passed in.
608 kfree(m->hw_handler_name);
609 m->hw_handler_name = attached_handler_name;
611 kfree(m->hw_handler_params);
612 m->hw_handler_params = NULL;
616 if (m->hw_handler_name) {
618 * Increments scsi_dh reference, even when using an
619 * already-attached handler.
621 r = scsi_dh_attach(q, m->hw_handler_name);
624 * Already attached to different hw_handler:
625 * try to reattach with correct one.
628 r = scsi_dh_attach(q, m->hw_handler_name);
632 ti->error = "error attaching hardware handler";
633 dm_put_device(ti, p->path.dev);
637 if (m->hw_handler_params) {
638 r = scsi_dh_set_params(q, m->hw_handler_params);
640 ti->error = "unable to set hardware "
641 "handler parameters";
643 dm_put_device(ti, p->path.dev);
649 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
651 dm_put_device(ti, p->path.dev);
662 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
665 static struct dm_arg _args[] = {
666 {1, 1024, "invalid number of paths"},
667 {0, 1024, "invalid number of selector args"}
671 unsigned i, nr_selector_args, nr_args;
672 struct priority_group *pg;
673 struct dm_target *ti = m->ti;
677 ti->error = "not enough priority group arguments";
678 return ERR_PTR(-EINVAL);
681 pg = alloc_priority_group();
683 ti->error = "couldn't allocate priority group";
684 return ERR_PTR(-ENOMEM);
688 r = parse_path_selector(as, pg, ti);
695 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
699 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
703 nr_args = 1 + nr_selector_args;
704 for (i = 0; i < pg->nr_pgpaths; i++) {
705 struct pgpath *pgpath;
706 struct dm_arg_set path_args;
708 if (as->argc < nr_args) {
709 ti->error = "not enough path parameters";
714 path_args.argc = nr_args;
715 path_args.argv = as->argv;
717 pgpath = parse_path(&path_args, &pg->ps, ti);
718 if (IS_ERR(pgpath)) {
724 list_add_tail(&pgpath->list, &pg->pgpaths);
725 dm_consume_args(as, nr_args);
731 free_priority_group(pg, ti);
735 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
739 struct dm_target *ti = m->ti;
741 static struct dm_arg _args[] = {
742 {0, 1024, "invalid number of hardware handler args"},
745 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
751 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
752 if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
753 "scsi_dh_%s", m->hw_handler_name)) {
754 ti->error = "unknown hardware handler type";
763 for (i = 0; i <= hw_argc - 2; i++)
764 len += strlen(as->argv[i]) + 1;
765 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
767 ti->error = "memory allocation failed";
771 j = sprintf(p, "%d", hw_argc - 1);
772 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
773 j = sprintf(p, "%s", as->argv[i]);
775 dm_consume_args(as, hw_argc - 1);
779 kfree(m->hw_handler_name);
780 m->hw_handler_name = NULL;
784 static int parse_features(struct dm_arg_set *as, struct multipath *m)
788 struct dm_target *ti = m->ti;
789 const char *arg_name;
791 static struct dm_arg _args[] = {
792 {0, 6, "invalid number of feature args"},
793 {1, 50, "pg_init_retries must be between 1 and 50"},
794 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
797 r = dm_read_arg_group(_args, as, &argc, &ti->error);
805 arg_name = dm_shift_arg(as);
808 if (!strcasecmp(arg_name, "queue_if_no_path")) {
809 r = queue_if_no_path(m, 1, 0);
813 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
814 m->retain_attached_hw_handler = 1;
818 if (!strcasecmp(arg_name, "pg_init_retries") &&
820 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
825 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
827 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
832 ti->error = "Unrecognised multipath feature request";
834 } while (argc && !r);
839 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
842 /* target arguments */
843 static struct dm_arg _args[] = {
844 {0, 1024, "invalid number of priority groups"},
845 {0, 1024, "invalid initial priority group number"},
850 struct dm_arg_set as;
851 unsigned pg_count = 0;
852 unsigned next_pg_num;
857 m = alloc_multipath(ti);
859 ti->error = "can't allocate multipath";
863 r = parse_features(&as, m);
867 r = parse_hw_handler(&as, m);
871 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
875 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
879 if ((!m->nr_priority_groups && next_pg_num) ||
880 (m->nr_priority_groups && !next_pg_num)) {
881 ti->error = "invalid initial priority group";
886 /* parse the priority groups */
888 struct priority_group *pg;
890 pg = parse_priority_group(&as, m);
896 m->nr_valid_paths += pg->nr_pgpaths;
897 list_add_tail(&pg->list, &m->priority_groups);
899 pg->pg_num = pg_count;
904 if (pg_count != m->nr_priority_groups) {
905 ti->error = "priority group count mismatch";
910 ti->num_flush_bios = 1;
911 ti->num_discard_bios = 1;
912 ti->num_write_same_bios = 1;
921 static void multipath_wait_for_pg_init_completion(struct multipath *m)
923 DECLARE_WAITQUEUE(wait, current);
926 add_wait_queue(&m->pg_init_wait, &wait);
929 set_current_state(TASK_UNINTERRUPTIBLE);
931 spin_lock_irqsave(&m->lock, flags);
932 if (!m->pg_init_in_progress) {
933 spin_unlock_irqrestore(&m->lock, flags);
936 spin_unlock_irqrestore(&m->lock, flags);
940 set_current_state(TASK_RUNNING);
942 remove_wait_queue(&m->pg_init_wait, &wait);
945 static void flush_multipath_work(struct multipath *m)
949 spin_lock_irqsave(&m->lock, flags);
950 m->pg_init_disabled = 1;
951 spin_unlock_irqrestore(&m->lock, flags);
953 flush_workqueue(kmpath_handlerd);
954 multipath_wait_for_pg_init_completion(m);
955 flush_workqueue(kmultipathd);
956 flush_work(&m->trigger_event);
958 spin_lock_irqsave(&m->lock, flags);
959 m->pg_init_disabled = 0;
960 spin_unlock_irqrestore(&m->lock, flags);
963 static void multipath_dtr(struct dm_target *ti)
965 struct multipath *m = ti->private;
967 flush_multipath_work(m);
972 * Map cloned requests
974 static int multipath_map(struct dm_target *ti, struct request *clone,
975 union map_info *map_context)
978 struct multipath *m = (struct multipath *) ti->private;
980 if (set_mapinfo(m, map_context) < 0)
981 /* ENOMEM, requeue */
982 return DM_MAPIO_REQUEUE;
984 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
985 r = map_io(m, clone, map_context, 0);
986 if (r < 0 || r == DM_MAPIO_REQUEUE)
987 clear_mapinfo(m, map_context);
993 * Take a path out of use.
995 static int fail_path(struct pgpath *pgpath)
998 struct multipath *m = pgpath->pg->m;
1000 spin_lock_irqsave(&m->lock, flags);
1002 if (!pgpath->is_active)
1005 DMWARN("Failing path %s.", pgpath->path.dev->name);
1007 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1008 pgpath->is_active = 0;
1009 pgpath->fail_count++;
1011 m->nr_valid_paths--;
1013 if (pgpath == m->current_pgpath)
1014 m->current_pgpath = NULL;
1016 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1017 pgpath->path.dev->name, m->nr_valid_paths);
1019 schedule_work(&m->trigger_event);
1022 spin_unlock_irqrestore(&m->lock, flags);
1028 * Reinstate a previously-failed path
1030 static int reinstate_path(struct pgpath *pgpath)
1033 unsigned long flags;
1034 struct multipath *m = pgpath->pg->m;
1036 spin_lock_irqsave(&m->lock, flags);
1038 if (pgpath->is_active)
1041 if (!pgpath->pg->ps.type->reinstate_path) {
1042 DMWARN("Reinstate path not supported by path selector %s",
1043 pgpath->pg->ps.type->name);
1048 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1052 pgpath->is_active = 1;
1054 if (!m->nr_valid_paths++ && m->queue_size) {
1055 m->current_pgpath = NULL;
1056 queue_work(kmultipathd, &m->process_queued_ios);
1057 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1058 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1059 m->pg_init_in_progress++;
1062 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1063 pgpath->path.dev->name, m->nr_valid_paths);
1065 schedule_work(&m->trigger_event);
1068 spin_unlock_irqrestore(&m->lock, flags);
1074 * Fail or reinstate all paths that match the provided struct dm_dev.
1076 static int action_dev(struct multipath *m, struct dm_dev *dev,
1080 struct pgpath *pgpath;
1081 struct priority_group *pg;
1083 list_for_each_entry(pg, &m->priority_groups, list) {
1084 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1085 if (pgpath->path.dev == dev)
1094 * Temporarily try to avoid having to use the specified PG
1096 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1099 unsigned long flags;
1101 spin_lock_irqsave(&m->lock, flags);
1103 pg->bypassed = bypassed;
1104 m->current_pgpath = NULL;
1105 m->current_pg = NULL;
1107 spin_unlock_irqrestore(&m->lock, flags);
1109 schedule_work(&m->trigger_event);
1113 * Switch to using the specified PG from the next I/O that gets mapped
1115 static int switch_pg_num(struct multipath *m, const char *pgstr)
1117 struct priority_group *pg;
1119 unsigned long flags;
1122 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1123 (pgnum > m->nr_priority_groups)) {
1124 DMWARN("invalid PG number supplied to switch_pg_num");
1128 spin_lock_irqsave(&m->lock, flags);
1129 list_for_each_entry(pg, &m->priority_groups, list) {
1134 m->current_pgpath = NULL;
1135 m->current_pg = NULL;
1138 spin_unlock_irqrestore(&m->lock, flags);
1140 schedule_work(&m->trigger_event);
1145 * Set/clear bypassed status of a PG.
1146 * PGs are numbered upwards from 1 in the order they were declared.
1148 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1150 struct priority_group *pg;
1154 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1155 (pgnum > m->nr_priority_groups)) {
1156 DMWARN("invalid PG number supplied to bypass_pg");
1160 list_for_each_entry(pg, &m->priority_groups, list) {
1165 bypass_pg(m, pg, bypassed);
1170 * Should we retry pg_init immediately?
1172 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1174 unsigned long flags;
1175 int limit_reached = 0;
1177 spin_lock_irqsave(&m->lock, flags);
1179 if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1180 m->pg_init_required = 1;
1184 spin_unlock_irqrestore(&m->lock, flags);
1186 return limit_reached;
1189 static void pg_init_done(void *data, int errors)
1191 struct pgpath *pgpath = data;
1192 struct priority_group *pg = pgpath->pg;
1193 struct multipath *m = pg->m;
1194 unsigned long flags;
1195 unsigned delay_retry = 0;
1197 /* device or driver problems */
1202 if (!m->hw_handler_name) {
1206 DMERR("Could not failover the device: Handler scsi_dh_%s "
1207 "Error %d.", m->hw_handler_name, errors);
1209 * Fail path for now, so we do not ping pong
1213 case SCSI_DH_DEV_TEMP_BUSY:
1215 * Probably doing something like FW upgrade on the
1216 * controller so try the other pg.
1218 bypass_pg(m, pg, 1);
1221 /* Wait before retrying. */
1223 case SCSI_DH_IMM_RETRY:
1224 case SCSI_DH_RES_TEMP_UNAVAIL:
1225 if (pg_init_limit_reached(m, pgpath))
1231 * We probably do not want to fail the path for a device
1232 * error, but this is what the old dm did. In future
1233 * patches we can do more advanced handling.
1238 spin_lock_irqsave(&m->lock, flags);
1240 if (pgpath == m->current_pgpath) {
1241 DMERR("Could not failover device. Error %d.", errors);
1242 m->current_pgpath = NULL;
1243 m->current_pg = NULL;
1245 } else if (!m->pg_init_required)
1248 if (--m->pg_init_in_progress)
1249 /* Activations of other paths are still on going */
1252 if (!m->pg_init_required)
1255 m->pg_init_delay_retry = delay_retry;
1256 queue_work(kmultipathd, &m->process_queued_ios);
1259 * Wake up any thread waiting to suspend.
1261 wake_up(&m->pg_init_wait);
1264 spin_unlock_irqrestore(&m->lock, flags);
1267 static void activate_path(struct work_struct *work)
1269 struct pgpath *pgpath =
1270 container_of(work, struct pgpath, activate_path.work);
1272 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1273 pg_init_done, pgpath);
1279 static int do_end_io(struct multipath *m, struct request *clone,
1280 int error, struct dm_mpath_io *mpio)
1283 * We don't queue any clone request inside the multipath target
1284 * during end I/O handling, since those clone requests don't have
1285 * bio clones. If we queue them inside the multipath target,
1286 * we need to make bio clones, that requires memory allocation.
1287 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1288 * don't have bio clones.)
1289 * Instead of queueing the clone request here, we queue the original
1290 * request into dm core, which will remake a clone request and
1291 * clone bios for it and resubmit it later.
1293 int r = DM_ENDIO_REQUEUE;
1294 unsigned long flags;
1296 if (!error && !clone->errors)
1297 return 0; /* I/O complete */
1299 if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ) {
1300 if ((clone->cmd_flags & REQ_WRITE_SAME) &&
1301 !clone->q->limits.max_write_same_sectors) {
1302 struct queue_limits *limits;
1304 /* device doesn't really support WRITE SAME, disable it */
1305 limits = dm_get_queue_limits(dm_table_get_md(m->ti->table));
1306 limits->max_write_same_sectors = 0;
1312 fail_path(mpio->pgpath);
1314 spin_lock_irqsave(&m->lock, flags);
1315 if (!m->nr_valid_paths) {
1316 if (!m->queue_if_no_path) {
1317 if (!__must_push_back(m))
1320 if (error == -EBADE)
1324 spin_unlock_irqrestore(&m->lock, flags);
1329 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1330 int error, union map_info *map_context)
1332 struct multipath *m = ti->private;
1333 struct dm_mpath_io *mpio = map_context->ptr;
1334 struct pgpath *pgpath;
1335 struct path_selector *ps;
1340 r = do_end_io(m, clone, error, mpio);
1341 pgpath = mpio->pgpath;
1343 ps = &pgpath->pg->ps;
1344 if (ps->type->end_io)
1345 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1347 clear_mapinfo(m, map_context);
1353 * Suspend can't complete until all the I/O is processed so if
1354 * the last path fails we must error any remaining I/O.
1355 * Note that if the freeze_bdev fails while suspending, the
1356 * queue_if_no_path state is lost - userspace should reset it.
1358 static void multipath_presuspend(struct dm_target *ti)
1360 struct multipath *m = (struct multipath *) ti->private;
1362 queue_if_no_path(m, 0, 1);
1365 static void multipath_postsuspend(struct dm_target *ti)
1367 struct multipath *m = ti->private;
1369 mutex_lock(&m->work_mutex);
1370 flush_multipath_work(m);
1371 mutex_unlock(&m->work_mutex);
1375 * Restore the queue_if_no_path setting.
1377 static void multipath_resume(struct dm_target *ti)
1379 struct multipath *m = (struct multipath *) ti->private;
1380 unsigned long flags;
1382 spin_lock_irqsave(&m->lock, flags);
1383 m->queue_if_no_path = m->saved_queue_if_no_path;
1384 spin_unlock_irqrestore(&m->lock, flags);
1388 * Info output has the following format:
1389 * num_multipath_feature_args [multipath_feature_args]*
1390 * num_handler_status_args [handler_status_args]*
1391 * num_groups init_group_number
1392 * [A|D|E num_ps_status_args [ps_status_args]*
1393 * num_paths num_selector_args
1394 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1396 * Table output has the following format (identical to the constructor string):
1397 * num_feature_args [features_args]*
1398 * num_handler_args hw_handler [hw_handler_args]*
1399 * num_groups init_group_number
1400 * [priority selector-name num_ps_args [ps_args]*
1401 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1403 static void multipath_status(struct dm_target *ti, status_type_t type,
1404 unsigned status_flags, char *result, unsigned maxlen)
1407 unsigned long flags;
1408 struct multipath *m = (struct multipath *) ti->private;
1409 struct priority_group *pg;
1414 spin_lock_irqsave(&m->lock, flags);
1417 if (type == STATUSTYPE_INFO)
1418 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1420 DMEMIT("%u ", m->queue_if_no_path +
1421 (m->pg_init_retries > 0) * 2 +
1422 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1423 m->retain_attached_hw_handler);
1424 if (m->queue_if_no_path)
1425 DMEMIT("queue_if_no_path ");
1426 if (m->pg_init_retries)
1427 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1428 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1429 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1430 if (m->retain_attached_hw_handler)
1431 DMEMIT("retain_attached_hw_handler ");
1434 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1437 DMEMIT("1 %s ", m->hw_handler_name);
1439 DMEMIT("%u ", m->nr_priority_groups);
1442 pg_num = m->next_pg->pg_num;
1443 else if (m->current_pg)
1444 pg_num = m->current_pg->pg_num;
1446 pg_num = (m->nr_priority_groups ? 1 : 0);
1448 DMEMIT("%u ", pg_num);
1451 case STATUSTYPE_INFO:
1452 list_for_each_entry(pg, &m->priority_groups, list) {
1454 state = 'D'; /* Disabled */
1455 else if (pg == m->current_pg)
1456 state = 'A'; /* Currently Active */
1458 state = 'E'; /* Enabled */
1460 DMEMIT("%c ", state);
1462 if (pg->ps.type->status)
1463 sz += pg->ps.type->status(&pg->ps, NULL, type,
1469 DMEMIT("%u %u ", pg->nr_pgpaths,
1470 pg->ps.type->info_args);
1472 list_for_each_entry(p, &pg->pgpaths, list) {
1473 DMEMIT("%s %s %u ", p->path.dev->name,
1474 p->is_active ? "A" : "F",
1476 if (pg->ps.type->status)
1477 sz += pg->ps.type->status(&pg->ps,
1478 &p->path, type, result + sz,
1484 case STATUSTYPE_TABLE:
1485 list_for_each_entry(pg, &m->priority_groups, list) {
1486 DMEMIT("%s ", pg->ps.type->name);
1488 if (pg->ps.type->status)
1489 sz += pg->ps.type->status(&pg->ps, NULL, type,
1495 DMEMIT("%u %u ", pg->nr_pgpaths,
1496 pg->ps.type->table_args);
1498 list_for_each_entry(p, &pg->pgpaths, list) {
1499 DMEMIT("%s ", p->path.dev->name);
1500 if (pg->ps.type->status)
1501 sz += pg->ps.type->status(&pg->ps,
1502 &p->path, type, result + sz,
1509 spin_unlock_irqrestore(&m->lock, flags);
1512 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1516 struct multipath *m = (struct multipath *) ti->private;
1519 mutex_lock(&m->work_mutex);
1521 if (dm_suspended(ti)) {
1527 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1528 r = queue_if_no_path(m, 1, 0);
1530 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1531 r = queue_if_no_path(m, 0, 0);
1537 DMWARN("Unrecognised multipath message received.");
1541 if (!strcasecmp(argv[0], "disable_group")) {
1542 r = bypass_pg_num(m, argv[1], 1);
1544 } else if (!strcasecmp(argv[0], "enable_group")) {
1545 r = bypass_pg_num(m, argv[1], 0);
1547 } else if (!strcasecmp(argv[0], "switch_group")) {
1548 r = switch_pg_num(m, argv[1]);
1550 } else if (!strcasecmp(argv[0], "reinstate_path"))
1551 action = reinstate_path;
1552 else if (!strcasecmp(argv[0], "fail_path"))
1555 DMWARN("Unrecognised multipath message received.");
1559 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1561 DMWARN("message: error getting device %s",
1566 r = action_dev(m, dev, action);
1568 dm_put_device(ti, dev);
1571 mutex_unlock(&m->work_mutex);
1575 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1578 struct multipath *m = ti->private;
1579 struct pgpath *pgpath;
1580 struct block_device *bdev;
1582 unsigned long flags;
1589 spin_lock_irqsave(&m->lock, flags);
1591 if (!m->current_pgpath)
1592 __choose_pgpath(m, 0);
1594 pgpath = m->current_pgpath;
1597 bdev = pgpath->path.dev->bdev;
1598 mode = pgpath->path.dev->mode;
1601 if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1606 spin_unlock_irqrestore(&m->lock, flags);
1609 * Only pass ioctls through if the device sizes match exactly.
1611 if (!bdev || ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) {
1612 int err = scsi_verify_blk_ioctl(NULL, cmd);
1617 if (r == -ENOTCONN && !fatal_signal_pending(current))
1618 queue_work(kmultipathd, &m->process_queued_ios);
1620 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1623 static int multipath_iterate_devices(struct dm_target *ti,
1624 iterate_devices_callout_fn fn, void *data)
1626 struct multipath *m = ti->private;
1627 struct priority_group *pg;
1631 list_for_each_entry(pg, &m->priority_groups, list) {
1632 list_for_each_entry(p, &pg->pgpaths, list) {
1633 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1643 static int __pgpath_busy(struct pgpath *pgpath)
1645 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1647 return dm_underlying_device_busy(q);
1651 * We return "busy", only when we can map I/Os but underlying devices
1652 * are busy (so even if we map I/Os now, the I/Os will wait on
1653 * the underlying queue).
1654 * In other words, if we want to kill I/Os or queue them inside us
1655 * due to map unavailability, we don't return "busy". Otherwise,
1656 * dm core won't give us the I/Os and we can't do what we want.
1658 static int multipath_busy(struct dm_target *ti)
1660 int busy = 0, has_active = 0;
1661 struct multipath *m = ti->private;
1662 struct priority_group *pg;
1663 struct pgpath *pgpath;
1664 unsigned long flags;
1666 spin_lock_irqsave(&m->lock, flags);
1668 /* Guess which priority_group will be used at next mapping time */
1669 if (unlikely(!m->current_pgpath && m->next_pg))
1671 else if (likely(m->current_pg))
1675 * We don't know which pg will be used at next mapping time.
1676 * We don't call __choose_pgpath() here to avoid to trigger
1677 * pg_init just by busy checking.
1678 * So we don't know whether underlying devices we will be using
1679 * at next mapping time are busy or not. Just try mapping.
1684 * If there is one non-busy active path at least, the path selector
1685 * will be able to select it. So we consider such a pg as not busy.
1688 list_for_each_entry(pgpath, &pg->pgpaths, list)
1689 if (pgpath->is_active) {
1692 if (!__pgpath_busy(pgpath)) {
1700 * No active path in this pg, so this pg won't be used and
1701 * the current_pg will be changed at next mapping time.
1702 * We need to try mapping to determine it.
1707 spin_unlock_irqrestore(&m->lock, flags);
1712 /*-----------------------------------------------------------------
1714 *---------------------------------------------------------------*/
1715 static struct target_type multipath_target = {
1716 .name = "multipath",
1717 .version = {1, 6, 0},
1718 .module = THIS_MODULE,
1719 .ctr = multipath_ctr,
1720 .dtr = multipath_dtr,
1721 .map_rq = multipath_map,
1722 .rq_end_io = multipath_end_io,
1723 .presuspend = multipath_presuspend,
1724 .postsuspend = multipath_postsuspend,
1725 .resume = multipath_resume,
1726 .status = multipath_status,
1727 .message = multipath_message,
1728 .ioctl = multipath_ioctl,
1729 .iterate_devices = multipath_iterate_devices,
1730 .busy = multipath_busy,
1733 static int __init dm_multipath_init(void)
1737 /* allocate a slab for the dm_ios */
1738 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1742 r = dm_register_target(&multipath_target);
1744 DMERR("register failed %d", r);
1745 kmem_cache_destroy(_mpio_cache);
1749 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1751 DMERR("failed to create workqueue kmpathd");
1752 dm_unregister_target(&multipath_target);
1753 kmem_cache_destroy(_mpio_cache);
1758 * A separate workqueue is used to handle the device handlers
1759 * to avoid overloading existing workqueue. Overloading the
1760 * old workqueue would also create a bottleneck in the
1761 * path of the storage hardware device activation.
1763 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1765 if (!kmpath_handlerd) {
1766 DMERR("failed to create workqueue kmpath_handlerd");
1767 destroy_workqueue(kmultipathd);
1768 dm_unregister_target(&multipath_target);
1769 kmem_cache_destroy(_mpio_cache);
1773 DMINFO("version %u.%u.%u loaded",
1774 multipath_target.version[0], multipath_target.version[1],
1775 multipath_target.version[2]);
1780 static void __exit dm_multipath_exit(void)
1782 destroy_workqueue(kmpath_handlerd);
1783 destroy_workqueue(kmultipathd);
1785 dm_unregister_target(&multipath_target);
1786 kmem_cache_destroy(_mpio_cache);
1789 module_init(dm_multipath_init);
1790 module_exit(dm_multipath_exit);
1792 MODULE_DESCRIPTION(DM_NAME " multipath target");
1793 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1794 MODULE_LICENSE("GPL");