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/blkdev.h>
15 #include <linux/ctype.h>
16 #include <linux/init.h>
17 #include <linux/mempool.h>
18 #include <linux/module.h>
19 #include <linux/pagemap.h>
20 #include <linux/slab.h>
21 #include <linux/time.h>
22 #include <linux/workqueue.h>
23 #include <linux/delay.h>
24 #include <scsi/scsi_dh.h>
25 #include <linux/atomic.h>
27 #define DM_MSG_PREFIX "multipath"
28 #define DM_PG_INIT_DELAY_MSECS 2000
29 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
33 struct list_head list;
35 struct priority_group *pg; /* Owning PG */
36 unsigned is_active; /* Path status */
37 unsigned fail_count; /* Cumulative failure count */
40 struct delayed_work activate_path;
43 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
46 * Paths are grouped into Priority Groups and numbered from 1 upwards.
47 * Each has a path selector which controls which path gets used.
49 struct priority_group {
50 struct list_head list;
52 struct multipath *m; /* Owning multipath instance */
53 struct path_selector ps;
55 unsigned pg_num; /* Reference number */
56 unsigned bypassed; /* Temporarily bypass this PG? */
58 unsigned nr_pgpaths; /* Number of paths in PG */
59 struct list_head pgpaths;
62 /* Multipath context */
64 struct list_head list;
67 const char *hw_handler_name;
68 char *hw_handler_params;
72 unsigned nr_priority_groups;
73 struct list_head priority_groups;
75 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
77 unsigned pg_init_required; /* pg_init needs calling? */
78 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
79 unsigned pg_init_delay_retry; /* Delay pg_init retry? */
81 unsigned nr_valid_paths; /* Total number of usable paths */
82 struct pgpath *current_pgpath;
83 struct priority_group *current_pg;
84 struct priority_group *next_pg; /* Switch to this PG if set */
85 unsigned repeat_count; /* I/Os left before calling PS again */
87 unsigned queue_io:1; /* Must we queue all I/O? */
88 unsigned queue_if_no_path:1; /* Queue I/O if last path fails? */
89 unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
90 unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
91 unsigned pg_init_disabled:1; /* pg_init is not currently allowed */
93 unsigned pg_init_retries; /* Number of times to retry pg_init */
94 unsigned pg_init_count; /* Number of times pg_init called */
95 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
97 struct work_struct trigger_event;
100 * We must use a mempool of dm_mpath_io structs so that we
101 * can resubmit bios on error.
103 mempool_t *mpio_pool;
105 struct mutex work_mutex;
109 * Context information attached to each bio we process.
112 struct pgpath *pgpath;
116 typedef int (*action_fn) (struct pgpath *pgpath);
118 static struct kmem_cache *_mpio_cache;
120 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
121 static void trigger_event(struct work_struct *work);
122 static void activate_path(struct work_struct *work);
123 static int __pgpath_busy(struct pgpath *pgpath);
126 /*-----------------------------------------------
127 * Allocation routines
128 *-----------------------------------------------*/
130 static struct pgpath *alloc_pgpath(void)
132 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
135 pgpath->is_active = 1;
136 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
142 static void free_pgpath(struct pgpath *pgpath)
147 static struct priority_group *alloc_priority_group(void)
149 struct priority_group *pg;
151 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
154 INIT_LIST_HEAD(&pg->pgpaths);
159 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
161 struct pgpath *pgpath, *tmp;
163 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
164 list_del(&pgpath->list);
165 dm_put_device(ti, pgpath->path.dev);
170 static void free_priority_group(struct priority_group *pg,
171 struct dm_target *ti)
173 struct path_selector *ps = &pg->ps;
176 ps->type->destroy(ps);
177 dm_put_path_selector(ps->type);
180 free_pgpaths(&pg->pgpaths, ti);
184 static struct multipath *alloc_multipath(struct dm_target *ti)
187 unsigned min_ios = dm_get_reserved_rq_based_ios();
189 m = kzalloc(sizeof(*m), GFP_KERNEL);
191 INIT_LIST_HEAD(&m->priority_groups);
192 spin_lock_init(&m->lock);
194 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
195 INIT_WORK(&m->trigger_event, trigger_event);
196 init_waitqueue_head(&m->pg_init_wait);
197 mutex_init(&m->work_mutex);
198 m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
210 static void free_multipath(struct multipath *m)
212 struct priority_group *pg, *tmp;
214 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
216 free_priority_group(pg, m->ti);
219 kfree(m->hw_handler_name);
220 kfree(m->hw_handler_params);
221 mempool_destroy(m->mpio_pool);
225 static int set_mapinfo(struct multipath *m, union map_info *info)
227 struct dm_mpath_io *mpio;
229 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
233 memset(mpio, 0, sizeof(*mpio));
239 static void clear_mapinfo(struct multipath *m, union map_info *info)
241 struct dm_mpath_io *mpio = info->ptr;
244 mempool_free(mpio, m->mpio_pool);
247 /*-----------------------------------------------
249 *-----------------------------------------------*/
251 static int __pg_init_all_paths(struct multipath *m)
253 struct pgpath *pgpath;
254 unsigned long pg_init_delay = 0;
256 if (m->pg_init_in_progress || m->pg_init_disabled)
260 m->pg_init_required = 0;
262 /* Check here to reset pg_init_required */
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++;
277 return m->pg_init_in_progress;
280 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
282 m->current_pg = pgpath->pg;
284 /* Must we initialise the PG first, and queue I/O till it's ready? */
285 if (m->hw_handler_name) {
286 m->pg_init_required = 1;
289 m->pg_init_required = 0;
293 m->pg_init_count = 0;
296 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
299 struct dm_path *path;
301 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
305 m->current_pgpath = path_to_pgpath(path);
307 if (m->current_pg != pg)
308 __switch_pg(m, m->current_pgpath);
313 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
315 struct priority_group *pg;
316 unsigned bypassed = 1;
318 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,
381 struct request *rq, struct request **__clone)
383 struct multipath *m = (struct multipath *) ti->private;
384 int r = DM_MAPIO_REQUEUE;
385 size_t nr_bytes = clone ? blk_rq_bytes(clone) : blk_rq_bytes(rq);
386 struct pgpath *pgpath;
387 struct block_device *bdev;
388 struct dm_mpath_io *mpio;
390 spin_lock_irq(&m->lock);
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 mpio = map_context->ptr;
413 mpio->pgpath = pgpath;
414 mpio->nr_bytes = nr_bytes;
416 bdev = pgpath->path.dev->bdev;
418 spin_unlock_irq(&m->lock);
421 /* Old request-based interface: allocated clone is passed in */
422 clone->q = bdev_get_queue(bdev);
423 clone->rq_disk = bdev->bd_disk;
424 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
426 /* blk-mq request-based interface */
427 *__clone = blk_get_request(bdev_get_queue(bdev),
428 rq_data_dir(rq), GFP_ATOMIC);
429 if (IS_ERR(*__clone)) {
430 /* ENOMEM, requeue */
431 clear_mapinfo(m, map_context);
434 (*__clone)->bio = (*__clone)->biotail = NULL;
435 (*__clone)->rq_disk = bdev->bd_disk;
436 (*__clone)->cmd_flags |= REQ_FAILFAST_TRANSPORT;
439 if (pgpath->pg->ps.type->start_io)
440 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
443 return DM_MAPIO_REMAPPED;
446 spin_unlock_irq(&m->lock);
451 static int multipath_map(struct dm_target *ti, struct request *clone,
452 union map_info *map_context)
454 return __multipath_map(ti, clone, map_context, NULL, NULL);
457 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
458 union map_info *map_context,
459 struct request **clone)
461 return __multipath_map(ti, NULL, map_context, rq, clone);
464 static void multipath_release_clone(struct request *clone)
466 blk_put_request(clone);
470 * If we run out of usable paths, should we queue I/O or error it?
472 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
473 unsigned save_old_value)
477 spin_lock_irqsave(&m->lock, flags);
480 m->saved_queue_if_no_path = m->queue_if_no_path;
482 m->saved_queue_if_no_path = queue_if_no_path;
483 m->queue_if_no_path = queue_if_no_path;
484 spin_unlock_irqrestore(&m->lock, flags);
486 if (!queue_if_no_path)
487 dm_table_run_md_queue_async(m->ti->table);
493 * An event is triggered whenever a path is taken out of use.
494 * Includes path failure and PG bypass.
496 static void trigger_event(struct work_struct *work)
498 struct multipath *m =
499 container_of(work, struct multipath, trigger_event);
501 dm_table_event(m->ti->table);
504 /*-----------------------------------------------------------------
505 * Constructor/argument parsing:
506 * <#multipath feature args> [<arg>]*
507 * <#hw_handler args> [hw_handler [<arg>]*]
509 * <initial priority group>
510 * [<selector> <#selector args> [<arg>]*
511 * <#paths> <#per-path selector args>
512 * [<path> [<arg>]* ]+ ]+
513 *---------------------------------------------------------------*/
514 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
515 struct dm_target *ti)
518 struct path_selector_type *pst;
521 static struct dm_arg _args[] = {
522 {0, 1024, "invalid number of path selector args"},
525 pst = dm_get_path_selector(dm_shift_arg(as));
527 ti->error = "unknown path selector type";
531 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
533 dm_put_path_selector(pst);
537 r = pst->create(&pg->ps, ps_argc, as->argv);
539 dm_put_path_selector(pst);
540 ti->error = "path selector constructor failed";
545 dm_consume_args(as, ps_argc);
550 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
551 struct dm_target *ti)
555 struct multipath *m = ti->private;
556 struct request_queue *q = NULL;
557 const char *attached_handler_name;
559 /* we need at least a path arg */
561 ti->error = "no device given";
562 return ERR_PTR(-EINVAL);
567 return ERR_PTR(-ENOMEM);
569 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
572 ti->error = "error getting device";
576 if (m->retain_attached_hw_handler || m->hw_handler_name)
577 q = bdev_get_queue(p->path.dev->bdev);
579 if (m->retain_attached_hw_handler) {
581 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
582 if (attached_handler_name) {
584 * Reset hw_handler_name to match the attached handler
585 * and clear any hw_handler_params associated with the
588 * NB. This modifies the table line to show the actual
589 * handler instead of the original table passed in.
591 kfree(m->hw_handler_name);
592 m->hw_handler_name = attached_handler_name;
594 kfree(m->hw_handler_params);
595 m->hw_handler_params = NULL;
599 if (m->hw_handler_name) {
600 r = scsi_dh_attach(q, m->hw_handler_name);
602 char b[BDEVNAME_SIZE];
604 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
605 bdevname(p->path.dev->bdev, b));
609 ti->error = "error attaching hardware handler";
610 dm_put_device(ti, p->path.dev);
614 if (m->hw_handler_params) {
615 r = scsi_dh_set_params(q, m->hw_handler_params);
617 ti->error = "unable to set hardware "
618 "handler parameters";
619 dm_put_device(ti, p->path.dev);
625 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
627 dm_put_device(ti, p->path.dev);
638 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
641 static struct dm_arg _args[] = {
642 {1, 1024, "invalid number of paths"},
643 {0, 1024, "invalid number of selector args"}
647 unsigned i, nr_selector_args, nr_args;
648 struct priority_group *pg;
649 struct dm_target *ti = m->ti;
653 ti->error = "not enough priority group arguments";
654 return ERR_PTR(-EINVAL);
657 pg = alloc_priority_group();
659 ti->error = "couldn't allocate priority group";
660 return ERR_PTR(-ENOMEM);
664 r = parse_path_selector(as, pg, ti);
671 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
675 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
679 nr_args = 1 + nr_selector_args;
680 for (i = 0; i < pg->nr_pgpaths; i++) {
681 struct pgpath *pgpath;
682 struct dm_arg_set path_args;
684 if (as->argc < nr_args) {
685 ti->error = "not enough path parameters";
690 path_args.argc = nr_args;
691 path_args.argv = as->argv;
693 pgpath = parse_path(&path_args, &pg->ps, ti);
694 if (IS_ERR(pgpath)) {
700 list_add_tail(&pgpath->list, &pg->pgpaths);
701 dm_consume_args(as, nr_args);
707 free_priority_group(pg, ti);
711 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
715 struct dm_target *ti = m->ti;
717 static struct dm_arg _args[] = {
718 {0, 1024, "invalid number of hardware handler args"},
721 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
727 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
733 for (i = 0; i <= hw_argc - 2; i++)
734 len += strlen(as->argv[i]) + 1;
735 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
737 ti->error = "memory allocation failed";
741 j = sprintf(p, "%d", hw_argc - 1);
742 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
743 j = sprintf(p, "%s", as->argv[i]);
745 dm_consume_args(as, hw_argc - 1);
749 kfree(m->hw_handler_name);
750 m->hw_handler_name = NULL;
754 static int parse_features(struct dm_arg_set *as, struct multipath *m)
758 struct dm_target *ti = m->ti;
759 const char *arg_name;
761 static struct dm_arg _args[] = {
762 {0, 6, "invalid number of feature args"},
763 {1, 50, "pg_init_retries must be between 1 and 50"},
764 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
767 r = dm_read_arg_group(_args, as, &argc, &ti->error);
775 arg_name = dm_shift_arg(as);
778 if (!strcasecmp(arg_name, "queue_if_no_path")) {
779 r = queue_if_no_path(m, 1, 0);
783 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
784 m->retain_attached_hw_handler = 1;
788 if (!strcasecmp(arg_name, "pg_init_retries") &&
790 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
795 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
797 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
802 ti->error = "Unrecognised multipath feature request";
804 } while (argc && !r);
809 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
812 /* target arguments */
813 static struct dm_arg _args[] = {
814 {0, 1024, "invalid number of priority groups"},
815 {0, 1024, "invalid initial priority group number"},
820 struct dm_arg_set as;
821 unsigned pg_count = 0;
822 unsigned next_pg_num;
827 m = alloc_multipath(ti);
829 ti->error = "can't allocate multipath";
833 r = parse_features(&as, m);
837 r = parse_hw_handler(&as, m);
841 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
845 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
849 if ((!m->nr_priority_groups && next_pg_num) ||
850 (m->nr_priority_groups && !next_pg_num)) {
851 ti->error = "invalid initial priority group";
856 /* parse the priority groups */
858 struct priority_group *pg;
860 pg = parse_priority_group(&as, m);
866 m->nr_valid_paths += pg->nr_pgpaths;
867 list_add_tail(&pg->list, &m->priority_groups);
869 pg->pg_num = pg_count;
874 if (pg_count != m->nr_priority_groups) {
875 ti->error = "priority group count mismatch";
880 ti->num_flush_bios = 1;
881 ti->num_discard_bios = 1;
882 ti->num_write_same_bios = 1;
891 static void multipath_wait_for_pg_init_completion(struct multipath *m)
893 DECLARE_WAITQUEUE(wait, current);
896 add_wait_queue(&m->pg_init_wait, &wait);
899 set_current_state(TASK_UNINTERRUPTIBLE);
901 spin_lock_irqsave(&m->lock, flags);
902 if (!m->pg_init_in_progress) {
903 spin_unlock_irqrestore(&m->lock, flags);
906 spin_unlock_irqrestore(&m->lock, flags);
910 set_current_state(TASK_RUNNING);
912 remove_wait_queue(&m->pg_init_wait, &wait);
915 static void flush_multipath_work(struct multipath *m)
919 spin_lock_irqsave(&m->lock, flags);
920 m->pg_init_disabled = 1;
921 spin_unlock_irqrestore(&m->lock, flags);
923 flush_workqueue(kmpath_handlerd);
924 multipath_wait_for_pg_init_completion(m);
925 flush_workqueue(kmultipathd);
926 flush_work(&m->trigger_event);
928 spin_lock_irqsave(&m->lock, flags);
929 m->pg_init_disabled = 0;
930 spin_unlock_irqrestore(&m->lock, flags);
933 static void multipath_dtr(struct dm_target *ti)
935 struct multipath *m = ti->private;
937 flush_multipath_work(m);
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)
981 int r = 0, run_queue = 0;
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++) {
1004 m->current_pgpath = NULL;
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);
1019 dm_table_run_md_queue_async(m->ti->table);
1025 * Fail or reinstate all paths that match the provided struct dm_dev.
1027 static int action_dev(struct multipath *m, struct dm_dev *dev,
1031 struct pgpath *pgpath;
1032 struct priority_group *pg;
1034 list_for_each_entry(pg, &m->priority_groups, list) {
1035 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1036 if (pgpath->path.dev == dev)
1045 * Temporarily try to avoid having to use the specified PG
1047 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1050 unsigned long flags;
1052 spin_lock_irqsave(&m->lock, flags);
1054 pg->bypassed = bypassed;
1055 m->current_pgpath = NULL;
1056 m->current_pg = NULL;
1058 spin_unlock_irqrestore(&m->lock, flags);
1060 schedule_work(&m->trigger_event);
1064 * Switch to using the specified PG from the next I/O that gets mapped
1066 static int switch_pg_num(struct multipath *m, const char *pgstr)
1068 struct priority_group *pg;
1070 unsigned long flags;
1073 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1074 (pgnum > m->nr_priority_groups)) {
1075 DMWARN("invalid PG number supplied to switch_pg_num");
1079 spin_lock_irqsave(&m->lock, flags);
1080 list_for_each_entry(pg, &m->priority_groups, list) {
1085 m->current_pgpath = NULL;
1086 m->current_pg = NULL;
1089 spin_unlock_irqrestore(&m->lock, flags);
1091 schedule_work(&m->trigger_event);
1096 * Set/clear bypassed status of a PG.
1097 * PGs are numbered upwards from 1 in the order they were declared.
1099 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1101 struct priority_group *pg;
1105 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1106 (pgnum > m->nr_priority_groups)) {
1107 DMWARN("invalid PG number supplied to bypass_pg");
1111 list_for_each_entry(pg, &m->priority_groups, list) {
1116 bypass_pg(m, pg, bypassed);
1121 * Should we retry pg_init immediately?
1123 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1125 unsigned long flags;
1126 int limit_reached = 0;
1128 spin_lock_irqsave(&m->lock, flags);
1130 if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1131 m->pg_init_required = 1;
1135 spin_unlock_irqrestore(&m->lock, flags);
1137 return limit_reached;
1140 static void pg_init_done(void *data, int errors)
1142 struct pgpath *pgpath = data;
1143 struct priority_group *pg = pgpath->pg;
1144 struct multipath *m = pg->m;
1145 unsigned long flags;
1146 unsigned delay_retry = 0;
1148 /* device or driver problems */
1153 if (!m->hw_handler_name) {
1157 DMERR("Could not failover the device: Handler scsi_dh_%s "
1158 "Error %d.", m->hw_handler_name, errors);
1160 * Fail path for now, so we do not ping pong
1164 case SCSI_DH_DEV_TEMP_BUSY:
1166 * Probably doing something like FW upgrade on the
1167 * controller so try the other pg.
1169 bypass_pg(m, pg, 1);
1172 /* Wait before retrying. */
1174 case SCSI_DH_IMM_RETRY:
1175 case SCSI_DH_RES_TEMP_UNAVAIL:
1176 if (pg_init_limit_reached(m, pgpath))
1182 * We probably do not want to fail the path for a device
1183 * error, but this is what the old dm did. In future
1184 * patches we can do more advanced handling.
1189 spin_lock_irqsave(&m->lock, flags);
1191 if (pgpath == m->current_pgpath) {
1192 DMERR("Could not failover device. Error %d.", errors);
1193 m->current_pgpath = NULL;
1194 m->current_pg = NULL;
1196 } else if (!m->pg_init_required)
1199 if (--m->pg_init_in_progress)
1200 /* Activations of other paths are still on going */
1203 if (m->pg_init_required) {
1204 m->pg_init_delay_retry = delay_retry;
1205 if (__pg_init_all_paths(m))
1211 * Wake up any thread waiting to suspend.
1213 wake_up(&m->pg_init_wait);
1216 spin_unlock_irqrestore(&m->lock, flags);
1219 static void activate_path(struct work_struct *work)
1221 struct pgpath *pgpath =
1222 container_of(work, struct pgpath, activate_path.work);
1224 if (pgpath->is_active)
1225 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1226 pg_init_done, pgpath);
1228 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1231 static int noretry_error(int error)
1242 /* Anything else could be a path failure, so should be retried */
1249 static int do_end_io(struct multipath *m, struct request *clone,
1250 int error, struct dm_mpath_io *mpio)
1253 * We don't queue any clone request inside the multipath target
1254 * during end I/O handling, since those clone requests don't have
1255 * bio clones. If we queue them inside the multipath target,
1256 * we need to make bio clones, that requires memory allocation.
1257 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1258 * don't have bio clones.)
1259 * Instead of queueing the clone request here, we queue the original
1260 * request into dm core, which will remake a clone request and
1261 * clone bios for it and resubmit it later.
1263 int r = DM_ENDIO_REQUEUE;
1264 unsigned long flags;
1266 if (!error && !clone->errors)
1267 return 0; /* I/O complete */
1269 if (noretry_error(error))
1273 fail_path(mpio->pgpath);
1275 spin_lock_irqsave(&m->lock, flags);
1276 if (!m->nr_valid_paths) {
1277 if (!m->queue_if_no_path) {
1278 if (!__must_push_back(m))
1281 if (error == -EBADE)
1285 spin_unlock_irqrestore(&m->lock, flags);
1290 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1291 int error, union map_info *map_context)
1293 struct multipath *m = ti->private;
1294 struct dm_mpath_io *mpio = map_context->ptr;
1295 struct pgpath *pgpath;
1296 struct path_selector *ps;
1301 r = do_end_io(m, clone, error, mpio);
1302 pgpath = mpio->pgpath;
1304 ps = &pgpath->pg->ps;
1305 if (ps->type->end_io)
1306 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1308 clear_mapinfo(m, map_context);
1314 * Suspend can't complete until all the I/O is processed so if
1315 * the last path fails we must error any remaining I/O.
1316 * Note that if the freeze_bdev fails while suspending, the
1317 * queue_if_no_path state is lost - userspace should reset it.
1319 static void multipath_presuspend(struct dm_target *ti)
1321 struct multipath *m = (struct multipath *) ti->private;
1323 queue_if_no_path(m, 0, 1);
1326 static void multipath_postsuspend(struct dm_target *ti)
1328 struct multipath *m = ti->private;
1330 mutex_lock(&m->work_mutex);
1331 flush_multipath_work(m);
1332 mutex_unlock(&m->work_mutex);
1336 * Restore the queue_if_no_path setting.
1338 static void multipath_resume(struct dm_target *ti)
1340 struct multipath *m = (struct multipath *) ti->private;
1341 unsigned long flags;
1343 spin_lock_irqsave(&m->lock, flags);
1344 m->queue_if_no_path = m->saved_queue_if_no_path;
1345 spin_unlock_irqrestore(&m->lock, flags);
1349 * Info output has the following format:
1350 * num_multipath_feature_args [multipath_feature_args]*
1351 * num_handler_status_args [handler_status_args]*
1352 * num_groups init_group_number
1353 * [A|D|E num_ps_status_args [ps_status_args]*
1354 * num_paths num_selector_args
1355 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1357 * Table output has the following format (identical to the constructor string):
1358 * num_feature_args [features_args]*
1359 * num_handler_args hw_handler [hw_handler_args]*
1360 * num_groups init_group_number
1361 * [priority selector-name num_ps_args [ps_args]*
1362 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1364 static void multipath_status(struct dm_target *ti, status_type_t type,
1365 unsigned status_flags, char *result, unsigned maxlen)
1368 unsigned long flags;
1369 struct multipath *m = (struct multipath *) ti->private;
1370 struct priority_group *pg;
1375 spin_lock_irqsave(&m->lock, flags);
1378 if (type == STATUSTYPE_INFO)
1379 DMEMIT("2 %u %u ", m->queue_io, m->pg_init_count);
1381 DMEMIT("%u ", m->queue_if_no_path +
1382 (m->pg_init_retries > 0) * 2 +
1383 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1384 m->retain_attached_hw_handler);
1385 if (m->queue_if_no_path)
1386 DMEMIT("queue_if_no_path ");
1387 if (m->pg_init_retries)
1388 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1389 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1390 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1391 if (m->retain_attached_hw_handler)
1392 DMEMIT("retain_attached_hw_handler ");
1395 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1398 DMEMIT("1 %s ", m->hw_handler_name);
1400 DMEMIT("%u ", m->nr_priority_groups);
1403 pg_num = m->next_pg->pg_num;
1404 else if (m->current_pg)
1405 pg_num = m->current_pg->pg_num;
1407 pg_num = (m->nr_priority_groups ? 1 : 0);
1409 DMEMIT("%u ", pg_num);
1412 case STATUSTYPE_INFO:
1413 list_for_each_entry(pg, &m->priority_groups, list) {
1415 state = 'D'; /* Disabled */
1416 else if (pg == m->current_pg)
1417 state = 'A'; /* Currently Active */
1419 state = 'E'; /* Enabled */
1421 DMEMIT("%c ", state);
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->info_args);
1433 list_for_each_entry(p, &pg->pgpaths, list) {
1434 DMEMIT("%s %s %u ", p->path.dev->name,
1435 p->is_active ? "A" : "F",
1437 if (pg->ps.type->status)
1438 sz += pg->ps.type->status(&pg->ps,
1439 &p->path, type, result + sz,
1445 case STATUSTYPE_TABLE:
1446 list_for_each_entry(pg, &m->priority_groups, list) {
1447 DMEMIT("%s ", pg->ps.type->name);
1449 if (pg->ps.type->status)
1450 sz += pg->ps.type->status(&pg->ps, NULL, type,
1456 DMEMIT("%u %u ", pg->nr_pgpaths,
1457 pg->ps.type->table_args);
1459 list_for_each_entry(p, &pg->pgpaths, list) {
1460 DMEMIT("%s ", p->path.dev->name);
1461 if (pg->ps.type->status)
1462 sz += pg->ps.type->status(&pg->ps,
1463 &p->path, type, result + sz,
1470 spin_unlock_irqrestore(&m->lock, flags);
1473 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1477 struct multipath *m = (struct multipath *) ti->private;
1480 mutex_lock(&m->work_mutex);
1482 if (dm_suspended(ti)) {
1488 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1489 r = queue_if_no_path(m, 1, 0);
1491 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1492 r = queue_if_no_path(m, 0, 0);
1498 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1502 if (!strcasecmp(argv[0], "disable_group")) {
1503 r = bypass_pg_num(m, argv[1], 1);
1505 } else if (!strcasecmp(argv[0], "enable_group")) {
1506 r = bypass_pg_num(m, argv[1], 0);
1508 } else if (!strcasecmp(argv[0], "switch_group")) {
1509 r = switch_pg_num(m, argv[1]);
1511 } else if (!strcasecmp(argv[0], "reinstate_path"))
1512 action = reinstate_path;
1513 else if (!strcasecmp(argv[0], "fail_path"))
1516 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1520 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1522 DMWARN("message: error getting device %s",
1527 r = action_dev(m, dev, action);
1529 dm_put_device(ti, dev);
1532 mutex_unlock(&m->work_mutex);
1536 static int multipath_prepare_ioctl(struct dm_target *ti,
1537 struct block_device **bdev, fmode_t *mode)
1539 struct multipath *m = ti->private;
1540 unsigned long flags;
1543 spin_lock_irqsave(&m->lock, flags);
1545 if (!m->current_pgpath)
1546 __choose_pgpath(m, 0);
1548 if (m->current_pgpath) {
1550 *bdev = m->current_pgpath->path.dev->bdev;
1551 *mode = m->current_pgpath->path.dev->mode;
1554 /* pg_init has not started or completed */
1558 /* No path is available */
1559 if (m->queue_if_no_path)
1565 spin_unlock_irqrestore(&m->lock, flags);
1567 if (r == -ENOTCONN) {
1568 spin_lock_irqsave(&m->lock, flags);
1569 if (!m->current_pg) {
1570 /* Path status changed, redo selection */
1571 __choose_pgpath(m, 0);
1573 if (m->pg_init_required)
1574 __pg_init_all_paths(m);
1575 spin_unlock_irqrestore(&m->lock, flags);
1576 dm_table_run_md_queue_async(m->ti->table);
1580 * Only pass ioctls through if the device sizes match exactly.
1582 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1587 static int multipath_iterate_devices(struct dm_target *ti,
1588 iterate_devices_callout_fn fn, void *data)
1590 struct multipath *m = ti->private;
1591 struct priority_group *pg;
1595 list_for_each_entry(pg, &m->priority_groups, list) {
1596 list_for_each_entry(p, &pg->pgpaths, list) {
1597 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1607 static int __pgpath_busy(struct pgpath *pgpath)
1609 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1611 return blk_lld_busy(q);
1615 * We return "busy", only when we can map I/Os but underlying devices
1616 * are busy (so even if we map I/Os now, the I/Os will wait on
1617 * the underlying queue).
1618 * In other words, if we want to kill I/Os or queue them inside us
1619 * due to map unavailability, we don't return "busy". Otherwise,
1620 * dm core won't give us the I/Os and we can't do what we want.
1622 static int multipath_busy(struct dm_target *ti)
1624 int busy = 0, has_active = 0;
1625 struct multipath *m = ti->private;
1626 struct priority_group *pg;
1627 struct pgpath *pgpath;
1628 unsigned long flags;
1630 spin_lock_irqsave(&m->lock, flags);
1632 /* pg_init in progress or no paths available */
1633 if (m->pg_init_in_progress ||
1634 (!m->nr_valid_paths && m->queue_if_no_path)) {
1638 /* Guess which priority_group will be used at next mapping time */
1639 if (unlikely(!m->current_pgpath && m->next_pg))
1641 else if (likely(m->current_pg))
1645 * We don't know which pg will be used at next mapping time.
1646 * We don't call __choose_pgpath() here to avoid to trigger
1647 * pg_init just by busy checking.
1648 * So we don't know whether underlying devices we will be using
1649 * at next mapping time are busy or not. Just try mapping.
1654 * If there is one non-busy active path at least, the path selector
1655 * will be able to select it. So we consider such a pg as not busy.
1658 list_for_each_entry(pgpath, &pg->pgpaths, list)
1659 if (pgpath->is_active) {
1662 if (!__pgpath_busy(pgpath)) {
1670 * No active path in this pg, so this pg won't be used and
1671 * the current_pg will be changed at next mapping time.
1672 * We need to try mapping to determine it.
1677 spin_unlock_irqrestore(&m->lock, flags);
1682 /*-----------------------------------------------------------------
1684 *---------------------------------------------------------------*/
1685 static struct target_type multipath_target = {
1686 .name = "multipath",
1687 .version = {1, 10, 0},
1688 .module = THIS_MODULE,
1689 .ctr = multipath_ctr,
1690 .dtr = multipath_dtr,
1691 .map_rq = multipath_map,
1692 .clone_and_map_rq = multipath_clone_and_map,
1693 .release_clone_rq = multipath_release_clone,
1694 .rq_end_io = multipath_end_io,
1695 .presuspend = multipath_presuspend,
1696 .postsuspend = multipath_postsuspend,
1697 .resume = multipath_resume,
1698 .status = multipath_status,
1699 .message = multipath_message,
1700 .prepare_ioctl = multipath_prepare_ioctl,
1701 .iterate_devices = multipath_iterate_devices,
1702 .busy = multipath_busy,
1705 static int __init dm_multipath_init(void)
1709 /* allocate a slab for the dm_ios */
1710 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1714 r = dm_register_target(&multipath_target);
1716 DMERR("register failed %d", r);
1718 goto bad_register_target;
1721 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1723 DMERR("failed to create workqueue kmpathd");
1725 goto bad_alloc_kmultipathd;
1729 * A separate workqueue is used to handle the device handlers
1730 * to avoid overloading existing workqueue. Overloading the
1731 * old workqueue would also create a bottleneck in the
1732 * path of the storage hardware device activation.
1734 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1736 if (!kmpath_handlerd) {
1737 DMERR("failed to create workqueue kmpath_handlerd");
1739 goto bad_alloc_kmpath_handlerd;
1742 DMINFO("version %u.%u.%u loaded",
1743 multipath_target.version[0], multipath_target.version[1],
1744 multipath_target.version[2]);
1748 bad_alloc_kmpath_handlerd:
1749 destroy_workqueue(kmultipathd);
1750 bad_alloc_kmultipathd:
1751 dm_unregister_target(&multipath_target);
1752 bad_register_target:
1753 kmem_cache_destroy(_mpio_cache);
1758 static void __exit dm_multipath_exit(void)
1760 destroy_workqueue(kmpath_handlerd);
1761 destroy_workqueue(kmultipathd);
1763 dm_unregister_target(&multipath_target);
1764 kmem_cache_destroy(_mpio_cache);
1767 module_init(dm_multipath_init);
1768 module_exit(dm_multipath_exit);
1770 MODULE_DESCRIPTION(DM_NAME " multipath target");
1771 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1772 MODULE_LICENSE("GPL");