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 <linux/atomic.h>
24 #define DM_MSG_PREFIX "multipath"
25 #define DM_PG_INIT_DELAY_MSECS 2000
26 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
30 struct list_head list;
32 struct priority_group *pg; /* Owning PG */
33 unsigned is_active; /* Path status */
34 unsigned fail_count; /* Cumulative failure count */
37 struct delayed_work 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;
69 unsigned nr_priority_groups;
70 struct list_head priority_groups;
72 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
74 unsigned pg_init_required; /* pg_init needs calling? */
75 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
76 unsigned pg_init_delay_retry; /* Delay pg_init retry? */
78 unsigned nr_valid_paths; /* Total number of usable paths */
79 struct pgpath *current_pgpath;
80 struct priority_group *current_pg;
81 struct priority_group *next_pg; /* Switch to this PG if set */
82 unsigned repeat_count; /* I/Os left before calling PS again */
84 unsigned queue_io; /* Must we queue all I/O? */
85 unsigned queue_if_no_path; /* Queue I/O if last path fails? */
86 unsigned saved_queue_if_no_path;/* Saved state during suspension */
87 unsigned pg_init_disabled:1; /* pg_init is not currently allowed */
88 unsigned pg_init_retries; /* Number of times to retry pg_init */
89 unsigned pg_init_count; /* Number of times pg_init called */
90 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
92 struct work_struct process_queued_ios;
93 struct list_head queued_ios;
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 #define MIN_IOS 256 /* Mempool size */
119 static struct kmem_cache *_mpio_cache;
121 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
122 static void process_queued_ios(struct work_struct *work);
123 static void trigger_event(struct work_struct *work);
124 static void activate_path(struct work_struct *work);
127 /*-----------------------------------------------
128 * Allocation routines
129 *-----------------------------------------------*/
131 static struct pgpath *alloc_pgpath(void)
133 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
136 pgpath->is_active = 1;
137 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
143 static void free_pgpath(struct pgpath *pgpath)
148 static struct priority_group *alloc_priority_group(void)
150 struct priority_group *pg;
152 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
155 INIT_LIST_HEAD(&pg->pgpaths);
160 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
162 struct pgpath *pgpath, *tmp;
163 struct multipath *m = ti->private;
165 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
166 list_del(&pgpath->list);
167 if (m->hw_handler_name)
168 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
169 dm_put_device(ti, pgpath->path.dev);
174 static void free_priority_group(struct priority_group *pg,
175 struct dm_target *ti)
177 struct path_selector *ps = &pg->ps;
180 ps->type->destroy(ps);
181 dm_put_path_selector(ps->type);
184 free_pgpaths(&pg->pgpaths, ti);
188 static struct multipath *alloc_multipath(struct dm_target *ti)
192 m = kzalloc(sizeof(*m), GFP_KERNEL);
194 INIT_LIST_HEAD(&m->priority_groups);
195 INIT_LIST_HEAD(&m->queued_ios);
196 spin_lock_init(&m->lock);
198 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
199 INIT_WORK(&m->process_queued_ios, process_queued_ios);
200 INIT_WORK(&m->trigger_event, trigger_event);
201 init_waitqueue_head(&m->pg_init_wait);
202 mutex_init(&m->work_mutex);
203 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
215 static void free_multipath(struct multipath *m)
217 struct priority_group *pg, *tmp;
219 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
221 free_priority_group(pg, m->ti);
224 kfree(m->hw_handler_name);
225 kfree(m->hw_handler_params);
226 mempool_destroy(m->mpio_pool);
231 /*-----------------------------------------------
233 *-----------------------------------------------*/
235 static void __pg_init_all_paths(struct multipath *m)
237 struct pgpath *pgpath;
238 unsigned long pg_init_delay = 0;
241 m->pg_init_required = 0;
242 if (m->pg_init_delay_retry)
243 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
244 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
245 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
246 /* Skip failed paths */
247 if (!pgpath->is_active)
249 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
251 m->pg_init_in_progress++;
255 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
257 m->current_pg = pgpath->pg;
259 /* Must we initialise the PG first, and queue I/O till it's ready? */
260 if (m->hw_handler_name) {
261 m->pg_init_required = 1;
264 m->pg_init_required = 0;
268 m->pg_init_count = 0;
271 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
274 struct dm_path *path;
276 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
280 m->current_pgpath = path_to_pgpath(path);
282 if (m->current_pg != pg)
283 __switch_pg(m, m->current_pgpath);
288 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
290 struct priority_group *pg;
291 unsigned bypassed = 1;
293 if (!m->nr_valid_paths)
296 /* Were we instructed to switch PG? */
300 if (!__choose_path_in_pg(m, pg, nr_bytes))
304 /* Don't change PG until it has no remaining paths */
305 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
309 * Loop through priority groups until we find a valid path.
310 * First time we skip PGs marked 'bypassed'.
311 * Second time we only try the ones we skipped.
314 list_for_each_entry(pg, &m->priority_groups, list) {
315 if (pg->bypassed == bypassed)
317 if (!__choose_path_in_pg(m, pg, nr_bytes))
320 } while (bypassed--);
323 m->current_pgpath = NULL;
324 m->current_pg = NULL;
328 * Check whether bios must be queued in the device-mapper core rather
329 * than here in the target.
331 * m->lock must be held on entry.
333 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
334 * same value then we are not between multipath_presuspend()
335 * and multipath_resume() calls and we have no need to check
336 * for the DMF_NOFLUSH_SUSPENDING flag.
338 static int __must_push_back(struct multipath *m)
340 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
341 dm_noflush_suspending(m->ti));
344 static int map_io(struct multipath *m, struct request *clone,
345 struct dm_mpath_io *mpio, unsigned was_queued)
347 int r = DM_MAPIO_REMAPPED;
348 size_t nr_bytes = blk_rq_bytes(clone);
350 struct pgpath *pgpath;
351 struct block_device *bdev;
353 spin_lock_irqsave(&m->lock, flags);
355 /* Do we need to select a new pgpath? */
356 if (!m->current_pgpath ||
357 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
358 __choose_pgpath(m, nr_bytes);
360 pgpath = m->current_pgpath;
365 if ((pgpath && m->queue_io) ||
366 (!pgpath && m->queue_if_no_path)) {
367 /* Queue for the daemon to resubmit */
368 list_add_tail(&clone->queuelist, &m->queued_ios);
370 if ((m->pg_init_required && !m->pg_init_in_progress) ||
372 queue_work(kmultipathd, &m->process_queued_ios);
374 r = DM_MAPIO_SUBMITTED;
376 bdev = pgpath->path.dev->bdev;
377 clone->q = bdev_get_queue(bdev);
378 clone->rq_disk = bdev->bd_disk;
379 } else if (__must_push_back(m))
380 r = DM_MAPIO_REQUEUE;
382 r = -EIO; /* Failed */
384 mpio->pgpath = pgpath;
385 mpio->nr_bytes = nr_bytes;
387 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
388 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
391 spin_unlock_irqrestore(&m->lock, flags);
397 * If we run out of usable paths, should we queue I/O or error it?
399 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
400 unsigned save_old_value)
404 spin_lock_irqsave(&m->lock, flags);
407 m->saved_queue_if_no_path = m->queue_if_no_path;
409 m->saved_queue_if_no_path = queue_if_no_path;
410 m->queue_if_no_path = queue_if_no_path;
411 if (!m->queue_if_no_path && m->queue_size)
412 queue_work(kmultipathd, &m->process_queued_ios);
414 spin_unlock_irqrestore(&m->lock, flags);
419 /*-----------------------------------------------------------------
420 * The multipath daemon is responsible for resubmitting queued ios.
421 *---------------------------------------------------------------*/
423 static void dispatch_queued_ios(struct multipath *m)
427 struct dm_mpath_io *mpio;
428 union map_info *info;
429 struct request *clone, *n;
432 spin_lock_irqsave(&m->lock, flags);
433 list_splice_init(&m->queued_ios, &cl);
434 spin_unlock_irqrestore(&m->lock, flags);
436 list_for_each_entry_safe(clone, n, &cl, queuelist) {
437 list_del_init(&clone->queuelist);
439 info = dm_get_rq_mapinfo(clone);
442 r = map_io(m, clone, mpio, 1);
444 mempool_free(mpio, m->mpio_pool);
445 dm_kill_unmapped_request(clone, r);
446 } else if (r == DM_MAPIO_REMAPPED)
447 dm_dispatch_request(clone);
448 else if (r == DM_MAPIO_REQUEUE) {
449 mempool_free(mpio, m->mpio_pool);
450 dm_requeue_unmapped_request(clone);
455 static void process_queued_ios(struct work_struct *work)
457 struct multipath *m =
458 container_of(work, struct multipath, process_queued_ios);
459 struct pgpath *pgpath = NULL;
460 unsigned must_queue = 1;
463 spin_lock_irqsave(&m->lock, flags);
468 if (!m->current_pgpath)
469 __choose_pgpath(m, 0);
471 pgpath = m->current_pgpath;
473 if ((pgpath && !m->queue_io) ||
474 (!pgpath && !m->queue_if_no_path))
477 if (m->pg_init_required && !m->pg_init_in_progress && pgpath &&
478 !m->pg_init_disabled)
479 __pg_init_all_paths(m);
482 spin_unlock_irqrestore(&m->lock, flags);
484 dispatch_queued_ios(m);
488 * An event is triggered whenever a path is taken out of use.
489 * Includes path failure and PG bypass.
491 static void trigger_event(struct work_struct *work)
493 struct multipath *m =
494 container_of(work, struct multipath, trigger_event);
496 dm_table_event(m->ti->table);
499 /*-----------------------------------------------------------------
500 * Constructor/argument parsing:
501 * <#multipath feature args> [<arg>]*
502 * <#hw_handler args> [hw_handler [<arg>]*]
504 * <initial priority group>
505 * [<selector> <#selector args> [<arg>]*
506 * <#paths> <#per-path selector args>
507 * [<path> [<arg>]* ]+ ]+
508 *---------------------------------------------------------------*/
509 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
510 struct dm_target *ti)
513 struct path_selector_type *pst;
516 static struct dm_arg _args[] = {
517 {0, 1024, "invalid number of path selector args"},
520 pst = dm_get_path_selector(dm_shift_arg(as));
522 ti->error = "unknown path selector type";
526 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
528 dm_put_path_selector(pst);
532 r = pst->create(&pg->ps, ps_argc, as->argv);
534 dm_put_path_selector(pst);
535 ti->error = "path selector constructor failed";
540 dm_consume_args(as, ps_argc);
545 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
546 struct dm_target *ti)
550 struct multipath *m = ti->private;
552 /* we need at least a path arg */
554 ti->error = "no device given";
555 return ERR_PTR(-EINVAL);
560 return ERR_PTR(-ENOMEM);
562 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
565 ti->error = "error getting device";
569 if (m->hw_handler_name) {
570 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
572 r = scsi_dh_attach(q, m->hw_handler_name);
575 * Already attached to different hw_handler,
576 * try to reattach with correct one.
579 r = scsi_dh_attach(q, m->hw_handler_name);
583 ti->error = "error attaching hardware handler";
584 dm_put_device(ti, p->path.dev);
588 if (m->hw_handler_params) {
589 r = scsi_dh_set_params(q, m->hw_handler_params);
591 ti->error = "unable to set hardware "
592 "handler parameters";
594 dm_put_device(ti, p->path.dev);
600 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
602 dm_put_device(ti, p->path.dev);
613 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
616 static struct dm_arg _args[] = {
617 {1, 1024, "invalid number of paths"},
618 {0, 1024, "invalid number of selector args"}
622 unsigned i, nr_selector_args, nr_args;
623 struct priority_group *pg;
624 struct dm_target *ti = m->ti;
628 ti->error = "not enough priority group arguments";
629 return ERR_PTR(-EINVAL);
632 pg = alloc_priority_group();
634 ti->error = "couldn't allocate priority group";
635 return ERR_PTR(-ENOMEM);
639 r = parse_path_selector(as, pg, ti);
646 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
650 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
654 nr_args = 1 + nr_selector_args;
655 for (i = 0; i < pg->nr_pgpaths; i++) {
656 struct pgpath *pgpath;
657 struct dm_arg_set path_args;
659 if (as->argc < nr_args) {
660 ti->error = "not enough path parameters";
665 path_args.argc = nr_args;
666 path_args.argv = as->argv;
668 pgpath = parse_path(&path_args, &pg->ps, ti);
669 if (IS_ERR(pgpath)) {
675 list_add_tail(&pgpath->list, &pg->pgpaths);
676 dm_consume_args(as, nr_args);
682 free_priority_group(pg, ti);
686 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
690 struct dm_target *ti = m->ti;
692 static struct dm_arg _args[] = {
693 {0, 1024, "invalid number of hardware handler args"},
696 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
702 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
703 if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
704 "scsi_dh_%s", m->hw_handler_name)) {
705 ti->error = "unknown hardware handler type";
714 for (i = 0; i <= hw_argc - 2; i++)
715 len += strlen(as->argv[i]) + 1;
716 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
718 ti->error = "memory allocation failed";
722 j = sprintf(p, "%d", hw_argc - 1);
723 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
724 j = sprintf(p, "%s", as->argv[i]);
726 dm_consume_args(as, hw_argc - 1);
730 kfree(m->hw_handler_name);
731 m->hw_handler_name = NULL;
735 static int parse_features(struct dm_arg_set *as, struct multipath *m)
739 struct dm_target *ti = m->ti;
740 const char *arg_name;
742 static struct dm_arg _args[] = {
743 {0, 5, "invalid number of feature args"},
744 {1, 50, "pg_init_retries must be between 1 and 50"},
745 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
748 r = dm_read_arg_group(_args, as, &argc, &ti->error);
756 arg_name = dm_shift_arg(as);
759 if (!strcasecmp(arg_name, "queue_if_no_path")) {
760 r = queue_if_no_path(m, 1, 0);
764 if (!strcasecmp(arg_name, "pg_init_retries") &&
766 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
771 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
773 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
778 ti->error = "Unrecognised multipath feature request";
780 } while (argc && !r);
785 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
788 /* target arguments */
789 static struct dm_arg _args[] = {
790 {0, 1024, "invalid number of priority groups"},
791 {0, 1024, "invalid initial priority group number"},
796 struct dm_arg_set as;
797 unsigned pg_count = 0;
798 unsigned next_pg_num;
803 m = alloc_multipath(ti);
805 ti->error = "can't allocate multipath";
809 r = parse_features(&as, m);
813 r = parse_hw_handler(&as, m);
817 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
821 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
825 if ((!m->nr_priority_groups && next_pg_num) ||
826 (m->nr_priority_groups && !next_pg_num)) {
827 ti->error = "invalid initial priority group";
832 /* parse the priority groups */
834 struct priority_group *pg;
836 pg = parse_priority_group(&as, m);
842 m->nr_valid_paths += pg->nr_pgpaths;
843 list_add_tail(&pg->list, &m->priority_groups);
845 pg->pg_num = pg_count;
850 if (pg_count != m->nr_priority_groups) {
851 ti->error = "priority group count mismatch";
856 ti->num_flush_requests = 1;
857 ti->num_discard_requests = 1;
866 static void multipath_wait_for_pg_init_completion(struct multipath *m)
868 DECLARE_WAITQUEUE(wait, current);
871 add_wait_queue(&m->pg_init_wait, &wait);
874 set_current_state(TASK_UNINTERRUPTIBLE);
876 spin_lock_irqsave(&m->lock, flags);
877 if (!m->pg_init_in_progress) {
878 spin_unlock_irqrestore(&m->lock, flags);
881 spin_unlock_irqrestore(&m->lock, flags);
885 set_current_state(TASK_RUNNING);
887 remove_wait_queue(&m->pg_init_wait, &wait);
890 static void flush_multipath_work(struct multipath *m)
894 spin_lock_irqsave(&m->lock, flags);
895 m->pg_init_disabled = 1;
896 spin_unlock_irqrestore(&m->lock, flags);
898 flush_workqueue(kmpath_handlerd);
899 multipath_wait_for_pg_init_completion(m);
900 flush_workqueue(kmultipathd);
901 flush_work_sync(&m->trigger_event);
903 spin_lock_irqsave(&m->lock, flags);
904 m->pg_init_disabled = 0;
905 spin_unlock_irqrestore(&m->lock, flags);
908 static void multipath_dtr(struct dm_target *ti)
910 struct multipath *m = ti->private;
912 flush_multipath_work(m);
917 * Map cloned requests
919 static int multipath_map(struct dm_target *ti, struct request *clone,
920 union map_info *map_context)
923 struct dm_mpath_io *mpio;
924 struct multipath *m = (struct multipath *) ti->private;
926 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
928 /* ENOMEM, requeue */
929 return DM_MAPIO_REQUEUE;
930 memset(mpio, 0, sizeof(*mpio));
932 map_context->ptr = mpio;
933 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
934 r = map_io(m, clone, mpio, 0);
935 if (r < 0 || r == DM_MAPIO_REQUEUE)
936 mempool_free(mpio, m->mpio_pool);
942 * Take a path out of use.
944 static int fail_path(struct pgpath *pgpath)
947 struct multipath *m = pgpath->pg->m;
949 spin_lock_irqsave(&m->lock, flags);
951 if (!pgpath->is_active)
954 DMWARN("Failing path %s.", pgpath->path.dev->name);
956 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
957 pgpath->is_active = 0;
958 pgpath->fail_count++;
962 if (pgpath == m->current_pgpath)
963 m->current_pgpath = NULL;
965 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
966 pgpath->path.dev->name, m->nr_valid_paths);
968 schedule_work(&m->trigger_event);
971 spin_unlock_irqrestore(&m->lock, flags);
977 * Reinstate a previously-failed path
979 static int reinstate_path(struct pgpath *pgpath)
983 struct multipath *m = pgpath->pg->m;
985 spin_lock_irqsave(&m->lock, flags);
987 if (pgpath->is_active)
990 if (!pgpath->pg->ps.type->reinstate_path) {
991 DMWARN("Reinstate path not supported by path selector %s",
992 pgpath->pg->ps.type->name);
997 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1001 pgpath->is_active = 1;
1003 if (!m->nr_valid_paths++ && m->queue_size) {
1004 m->current_pgpath = NULL;
1005 queue_work(kmultipathd, &m->process_queued_ios);
1006 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1007 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1008 m->pg_init_in_progress++;
1011 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1012 pgpath->path.dev->name, m->nr_valid_paths);
1014 schedule_work(&m->trigger_event);
1017 spin_unlock_irqrestore(&m->lock, flags);
1023 * Fail or reinstate all paths that match the provided struct dm_dev.
1025 static int action_dev(struct multipath *m, struct dm_dev *dev,
1029 struct pgpath *pgpath;
1030 struct priority_group *pg;
1032 list_for_each_entry(pg, &m->priority_groups, list) {
1033 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1034 if (pgpath->path.dev == dev)
1043 * Temporarily try to avoid having to use the specified PG
1045 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1048 unsigned long flags;
1050 spin_lock_irqsave(&m->lock, flags);
1052 pg->bypassed = bypassed;
1053 m->current_pgpath = NULL;
1054 m->current_pg = NULL;
1056 spin_unlock_irqrestore(&m->lock, flags);
1058 schedule_work(&m->trigger_event);
1062 * Switch to using the specified PG from the next I/O that gets mapped
1064 static int switch_pg_num(struct multipath *m, const char *pgstr)
1066 struct priority_group *pg;
1068 unsigned long flags;
1070 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1071 (pgnum > m->nr_priority_groups)) {
1072 DMWARN("invalid PG number supplied to switch_pg_num");
1076 spin_lock_irqsave(&m->lock, flags);
1077 list_for_each_entry(pg, &m->priority_groups, list) {
1082 m->current_pgpath = NULL;
1083 m->current_pg = NULL;
1086 spin_unlock_irqrestore(&m->lock, flags);
1088 schedule_work(&m->trigger_event);
1093 * Set/clear bypassed status of a PG.
1094 * PGs are numbered upwards from 1 in the order they were declared.
1096 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1098 struct priority_group *pg;
1101 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1102 (pgnum > m->nr_priority_groups)) {
1103 DMWARN("invalid PG number supplied to bypass_pg");
1107 list_for_each_entry(pg, &m->priority_groups, list) {
1112 bypass_pg(m, pg, bypassed);
1117 * Should we retry pg_init immediately?
1119 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1121 unsigned long flags;
1122 int limit_reached = 0;
1124 spin_lock_irqsave(&m->lock, flags);
1126 if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1127 m->pg_init_required = 1;
1131 spin_unlock_irqrestore(&m->lock, flags);
1133 return limit_reached;
1136 static void pg_init_done(void *data, int errors)
1138 struct pgpath *pgpath = data;
1139 struct priority_group *pg = pgpath->pg;
1140 struct multipath *m = pg->m;
1141 unsigned long flags;
1142 unsigned delay_retry = 0;
1144 /* device or driver problems */
1149 if (!m->hw_handler_name) {
1153 DMERR("Could not failover the device: Handler scsi_dh_%s "
1154 "Error %d.", m->hw_handler_name, errors);
1156 * Fail path for now, so we do not ping pong
1160 case SCSI_DH_DEV_TEMP_BUSY:
1162 * Probably doing something like FW upgrade on the
1163 * controller so try the other pg.
1165 bypass_pg(m, pg, 1);
1168 /* Wait before retrying. */
1170 case SCSI_DH_IMM_RETRY:
1171 case SCSI_DH_RES_TEMP_UNAVAIL:
1172 if (pg_init_limit_reached(m, pgpath))
1178 * We probably do not want to fail the path for a device
1179 * error, but this is what the old dm did. In future
1180 * patches we can do more advanced handling.
1185 spin_lock_irqsave(&m->lock, flags);
1187 if (pgpath == m->current_pgpath) {
1188 DMERR("Could not failover device. Error %d.", errors);
1189 m->current_pgpath = NULL;
1190 m->current_pg = NULL;
1192 } else if (!m->pg_init_required)
1195 if (--m->pg_init_in_progress)
1196 /* Activations of other paths are still on going */
1199 if (!m->pg_init_required)
1202 m->pg_init_delay_retry = delay_retry;
1203 queue_work(kmultipathd, &m->process_queued_ios);
1206 * Wake up any thread waiting to suspend.
1208 wake_up(&m->pg_init_wait);
1211 spin_unlock_irqrestore(&m->lock, flags);
1214 static void activate_path(struct work_struct *work)
1216 struct pgpath *pgpath =
1217 container_of(work, struct pgpath, activate_path.work);
1219 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1220 pg_init_done, pgpath);
1226 static int do_end_io(struct multipath *m, struct request *clone,
1227 int error, struct dm_mpath_io *mpio)
1230 * We don't queue any clone request inside the multipath target
1231 * during end I/O handling, since those clone requests don't have
1232 * bio clones. If we queue them inside the multipath target,
1233 * we need to make bio clones, that requires memory allocation.
1234 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1235 * don't have bio clones.)
1236 * Instead of queueing the clone request here, we queue the original
1237 * request into dm core, which will remake a clone request and
1238 * clone bios for it and resubmit it later.
1240 int r = DM_ENDIO_REQUEUE;
1241 unsigned long flags;
1243 if (!error && !clone->errors)
1244 return 0; /* I/O complete */
1246 if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1250 fail_path(mpio->pgpath);
1252 spin_lock_irqsave(&m->lock, flags);
1253 if (!m->nr_valid_paths) {
1254 if (!m->queue_if_no_path) {
1255 if (!__must_push_back(m))
1258 if (error == -EBADE)
1262 spin_unlock_irqrestore(&m->lock, flags);
1267 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1268 int error, union map_info *map_context)
1270 struct multipath *m = ti->private;
1271 struct dm_mpath_io *mpio = map_context->ptr;
1272 struct pgpath *pgpath = mpio->pgpath;
1273 struct path_selector *ps;
1276 r = do_end_io(m, clone, error, mpio);
1278 ps = &pgpath->pg->ps;
1279 if (ps->type->end_io)
1280 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1282 mempool_free(mpio, m->mpio_pool);
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 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_size, 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 if (m->queue_if_no_path)
1359 DMEMIT("queue_if_no_path ");
1360 if (m->pg_init_retries)
1361 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1362 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1363 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1366 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1369 DMEMIT("1 %s ", m->hw_handler_name);
1371 DMEMIT("%u ", m->nr_priority_groups);
1374 pg_num = m->next_pg->pg_num;
1375 else if (m->current_pg)
1376 pg_num = m->current_pg->pg_num;
1378 pg_num = (m->nr_priority_groups ? 1 : 0);
1380 DMEMIT("%u ", pg_num);
1383 case STATUSTYPE_INFO:
1384 list_for_each_entry(pg, &m->priority_groups, list) {
1386 state = 'D'; /* Disabled */
1387 else if (pg == m->current_pg)
1388 state = 'A'; /* Currently Active */
1390 state = 'E'; /* Enabled */
1392 DMEMIT("%c ", state);
1394 if (pg->ps.type->status)
1395 sz += pg->ps.type->status(&pg->ps, NULL, type,
1401 DMEMIT("%u %u ", pg->nr_pgpaths,
1402 pg->ps.type->info_args);
1404 list_for_each_entry(p, &pg->pgpaths, list) {
1405 DMEMIT("%s %s %u ", p->path.dev->name,
1406 p->is_active ? "A" : "F",
1408 if (pg->ps.type->status)
1409 sz += pg->ps.type->status(&pg->ps,
1410 &p->path, type, result + sz,
1416 case STATUSTYPE_TABLE:
1417 list_for_each_entry(pg, &m->priority_groups, list) {
1418 DMEMIT("%s ", pg->ps.type->name);
1420 if (pg->ps.type->status)
1421 sz += pg->ps.type->status(&pg->ps, NULL, type,
1427 DMEMIT("%u %u ", pg->nr_pgpaths,
1428 pg->ps.type->table_args);
1430 list_for_each_entry(p, &pg->pgpaths, list) {
1431 DMEMIT("%s ", p->path.dev->name);
1432 if (pg->ps.type->status)
1433 sz += pg->ps.type->status(&pg->ps,
1434 &p->path, type, result + sz,
1441 spin_unlock_irqrestore(&m->lock, flags);
1444 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1448 struct multipath *m = (struct multipath *) ti->private;
1451 mutex_lock(&m->work_mutex);
1453 if (dm_suspended(ti)) {
1459 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1460 r = queue_if_no_path(m, 1, 0);
1462 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1463 r = queue_if_no_path(m, 0, 0);
1469 DMWARN("Unrecognised multipath message received.");
1473 if (!strcasecmp(argv[0], "disable_group")) {
1474 r = bypass_pg_num(m, argv[1], 1);
1476 } else if (!strcasecmp(argv[0], "enable_group")) {
1477 r = bypass_pg_num(m, argv[1], 0);
1479 } else if (!strcasecmp(argv[0], "switch_group")) {
1480 r = switch_pg_num(m, argv[1]);
1482 } else if (!strcasecmp(argv[0], "reinstate_path"))
1483 action = reinstate_path;
1484 else if (!strcasecmp(argv[0], "fail_path"))
1487 DMWARN("Unrecognised multipath message received.");
1491 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1493 DMWARN("message: error getting device %s",
1498 r = action_dev(m, dev, action);
1500 dm_put_device(ti, dev);
1503 mutex_unlock(&m->work_mutex);
1507 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1510 struct multipath *m = (struct multipath *) ti->private;
1511 struct block_device *bdev = NULL;
1513 unsigned long flags;
1516 spin_lock_irqsave(&m->lock, flags);
1518 if (!m->current_pgpath)
1519 __choose_pgpath(m, 0);
1521 if (m->current_pgpath) {
1522 bdev = m->current_pgpath->path.dev->bdev;
1523 mode = m->current_pgpath->path.dev->mode;
1531 spin_unlock_irqrestore(&m->lock, flags);
1534 * Only pass ioctls through if the device sizes match exactly.
1536 if (!r && ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT)
1537 r = scsi_verify_blk_ioctl(NULL, cmd);
1539 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1542 static int multipath_iterate_devices(struct dm_target *ti,
1543 iterate_devices_callout_fn fn, void *data)
1545 struct multipath *m = ti->private;
1546 struct priority_group *pg;
1550 list_for_each_entry(pg, &m->priority_groups, list) {
1551 list_for_each_entry(p, &pg->pgpaths, list) {
1552 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1562 static int __pgpath_busy(struct pgpath *pgpath)
1564 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1566 return dm_underlying_device_busy(q);
1570 * We return "busy", only when we can map I/Os but underlying devices
1571 * are busy (so even if we map I/Os now, the I/Os will wait on
1572 * the underlying queue).
1573 * In other words, if we want to kill I/Os or queue them inside us
1574 * due to map unavailability, we don't return "busy". Otherwise,
1575 * dm core won't give us the I/Os and we can't do what we want.
1577 static int multipath_busy(struct dm_target *ti)
1579 int busy = 0, has_active = 0;
1580 struct multipath *m = ti->private;
1581 struct priority_group *pg;
1582 struct pgpath *pgpath;
1583 unsigned long flags;
1585 spin_lock_irqsave(&m->lock, flags);
1587 /* Guess which priority_group will be used at next mapping time */
1588 if (unlikely(!m->current_pgpath && m->next_pg))
1590 else if (likely(m->current_pg))
1594 * We don't know which pg will be used at next mapping time.
1595 * We don't call __choose_pgpath() here to avoid to trigger
1596 * pg_init just by busy checking.
1597 * So we don't know whether underlying devices we will be using
1598 * at next mapping time are busy or not. Just try mapping.
1603 * If there is one non-busy active path at least, the path selector
1604 * will be able to select it. So we consider such a pg as not busy.
1607 list_for_each_entry(pgpath, &pg->pgpaths, list)
1608 if (pgpath->is_active) {
1611 if (!__pgpath_busy(pgpath)) {
1619 * No active path in this pg, so this pg won't be used and
1620 * the current_pg will be changed at next mapping time.
1621 * We need to try mapping to determine it.
1626 spin_unlock_irqrestore(&m->lock, flags);
1631 /*-----------------------------------------------------------------
1633 *---------------------------------------------------------------*/
1634 static struct target_type multipath_target = {
1635 .name = "multipath",
1636 .version = {1, 3, 2},
1637 .module = THIS_MODULE,
1638 .ctr = multipath_ctr,
1639 .dtr = multipath_dtr,
1640 .map_rq = multipath_map,
1641 .rq_end_io = multipath_end_io,
1642 .presuspend = multipath_presuspend,
1643 .postsuspend = multipath_postsuspend,
1644 .resume = multipath_resume,
1645 .status = multipath_status,
1646 .message = multipath_message,
1647 .ioctl = multipath_ioctl,
1648 .iterate_devices = multipath_iterate_devices,
1649 .busy = multipath_busy,
1652 static int __init dm_multipath_init(void)
1656 /* allocate a slab for the dm_ios */
1657 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1661 r = dm_register_target(&multipath_target);
1663 DMERR("register failed %d", r);
1664 kmem_cache_destroy(_mpio_cache);
1668 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1670 DMERR("failed to create workqueue kmpathd");
1671 dm_unregister_target(&multipath_target);
1672 kmem_cache_destroy(_mpio_cache);
1677 * A separate workqueue is used to handle the device handlers
1678 * to avoid overloading existing workqueue. Overloading the
1679 * old workqueue would also create a bottleneck in the
1680 * path of the storage hardware device activation.
1682 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1684 if (!kmpath_handlerd) {
1685 DMERR("failed to create workqueue kmpath_handlerd");
1686 destroy_workqueue(kmultipathd);
1687 dm_unregister_target(&multipath_target);
1688 kmem_cache_destroy(_mpio_cache);
1692 DMINFO("version %u.%u.%u loaded",
1693 multipath_target.version[0], multipath_target.version[1],
1694 multipath_target.version[2]);
1699 static void __exit dm_multipath_exit(void)
1701 destroy_workqueue(kmpath_handlerd);
1702 destroy_workqueue(kmultipathd);
1704 dm_unregister_target(&multipath_target);
1705 kmem_cache_destroy(_mpio_cache);
1708 module_init(dm_multipath_init);
1709 module_exit(dm_multipath_exit);
1711 MODULE_DESCRIPTION(DM_NAME " multipath target");
1712 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1713 MODULE_LICENSE("GPL");