2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
5 * This file is released under the GPL.
8 #include <linux/device-mapper.h>
11 #include "dm-path-selector.h"
12 #include "dm-uevent.h"
14 #include <linux/ctype.h>
15 #include <linux/init.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/pagemap.h>
19 #include <linux/slab.h>
20 #include <linux/time.h>
21 #include <linux/workqueue.h>
22 #include <linux/delay.h>
23 #include <scsi/scsi_dh.h>
24 #include <linux/atomic.h>
26 #define DM_MSG_PREFIX "multipath"
27 #define DM_PG_INIT_DELAY_MSECS 2000
28 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
32 struct list_head list;
34 struct priority_group *pg; /* Owning PG */
35 unsigned is_active; /* Path status */
36 unsigned fail_count; /* Cumulative failure count */
39 struct delayed_work activate_path;
42 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
45 * Paths are grouped into Priority Groups and numbered from 1 upwards.
46 * Each has a path selector which controls which path gets used.
48 struct priority_group {
49 struct list_head list;
51 struct multipath *m; /* Owning multipath instance */
52 struct path_selector ps;
54 unsigned pg_num; /* Reference number */
55 unsigned bypassed; /* Temporarily bypass this PG? */
57 unsigned nr_pgpaths; /* Number of paths in PG */
58 struct list_head pgpaths;
61 /* Multipath context */
63 struct list_head list;
66 const char *hw_handler_name;
67 char *hw_handler_params;
71 unsigned nr_priority_groups;
72 struct list_head priority_groups;
74 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
76 unsigned pg_init_required; /* pg_init needs calling? */
77 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
78 unsigned pg_init_delay_retry; /* Delay pg_init retry? */
80 unsigned nr_valid_paths; /* Total number of usable paths */
81 struct pgpath *current_pgpath;
82 struct priority_group *current_pg;
83 struct priority_group *next_pg; /* Switch to this PG if set */
84 unsigned repeat_count; /* I/Os left before calling PS again */
86 unsigned queue_io:1; /* Must we queue all I/O? */
87 unsigned queue_if_no_path:1; /* Queue I/O if last path fails? */
88 unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
89 unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
90 unsigned pg_init_disabled:1; /* pg_init is not currently allowed */
92 unsigned pg_init_retries; /* Number of times to retry pg_init */
93 unsigned pg_init_count; /* Number of times pg_init called */
94 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
96 struct work_struct trigger_event;
99 * We must use a mempool of dm_mpath_io structs so that we
100 * can resubmit bios on error.
102 mempool_t *mpio_pool;
104 struct mutex work_mutex;
108 * Context information attached to each bio we process.
111 struct pgpath *pgpath;
115 typedef int (*action_fn) (struct pgpath *pgpath);
117 static struct kmem_cache *_mpio_cache;
119 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
120 static void trigger_event(struct work_struct *work);
121 static void activate_path(struct work_struct *work);
122 static int __pgpath_busy(struct pgpath *pgpath);
125 /*-----------------------------------------------
126 * Allocation routines
127 *-----------------------------------------------*/
129 static struct pgpath *alloc_pgpath(void)
131 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
134 pgpath->is_active = 1;
135 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
141 static void free_pgpath(struct pgpath *pgpath)
146 static struct priority_group *alloc_priority_group(void)
148 struct priority_group *pg;
150 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
153 INIT_LIST_HEAD(&pg->pgpaths);
158 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
160 struct pgpath *pgpath, *tmp;
161 struct multipath *m = ti->private;
163 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
164 list_del(&pgpath->list);
165 if (m->hw_handler_name)
166 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
167 dm_put_device(ti, pgpath->path.dev);
172 static void free_priority_group(struct priority_group *pg,
173 struct dm_target *ti)
175 struct path_selector *ps = &pg->ps;
178 ps->type->destroy(ps);
179 dm_put_path_selector(ps->type);
182 free_pgpaths(&pg->pgpaths, ti);
186 static struct multipath *alloc_multipath(struct dm_target *ti)
189 unsigned min_ios = dm_get_reserved_rq_based_ios();
191 m = kzalloc(sizeof(*m), GFP_KERNEL);
193 INIT_LIST_HEAD(&m->priority_groups);
194 spin_lock_init(&m->lock);
196 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
197 INIT_WORK(&m->trigger_event, trigger_event);
198 init_waitqueue_head(&m->pg_init_wait);
199 mutex_init(&m->work_mutex);
200 m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
212 static void free_multipath(struct multipath *m)
214 struct priority_group *pg, *tmp;
216 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
218 free_priority_group(pg, m->ti);
221 kfree(m->hw_handler_name);
222 kfree(m->hw_handler_params);
223 mempool_destroy(m->mpio_pool);
227 static int set_mapinfo(struct multipath *m, union map_info *info)
229 struct dm_mpath_io *mpio;
231 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
235 memset(mpio, 0, sizeof(*mpio));
241 static void clear_mapinfo(struct multipath *m, union map_info *info)
243 struct dm_mpath_io *mpio = info->ptr;
246 mempool_free(mpio, m->mpio_pool);
249 /*-----------------------------------------------
251 *-----------------------------------------------*/
253 static int __pg_init_all_paths(struct multipath *m)
255 struct pgpath *pgpath;
256 unsigned long pg_init_delay = 0;
258 if (m->pg_init_in_progress || m->pg_init_disabled)
262 m->pg_init_required = 0;
264 /* Check here to reset pg_init_required */
268 if (m->pg_init_delay_retry)
269 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
270 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
271 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
272 /* Skip failed paths */
273 if (!pgpath->is_active)
275 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
277 m->pg_init_in_progress++;
279 return m->pg_init_in_progress;
282 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
284 m->current_pg = pgpath->pg;
286 /* Must we initialise the PG first, and queue I/O till it's ready? */
287 if (m->hw_handler_name) {
288 m->pg_init_required = 1;
291 m->pg_init_required = 0;
295 m->pg_init_count = 0;
298 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
301 struct dm_path *path;
303 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
307 m->current_pgpath = path_to_pgpath(path);
309 if (m->current_pg != pg)
310 __switch_pg(m, m->current_pgpath);
315 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
317 struct priority_group *pg;
318 unsigned bypassed = 1;
320 if (!m->nr_valid_paths)
323 /* Were we instructed to switch PG? */
327 if (!__choose_path_in_pg(m, pg, nr_bytes))
331 /* Don't change PG until it has no remaining paths */
332 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
336 * Loop through priority groups until we find a valid path.
337 * First time we skip PGs marked 'bypassed'.
338 * Second time we only try the ones we skipped, but set
339 * pg_init_delay_retry so we do not hammer controllers.
342 list_for_each_entry(pg, &m->priority_groups, list) {
343 if (pg->bypassed == bypassed)
345 if (!__choose_path_in_pg(m, pg, nr_bytes)) {
347 m->pg_init_delay_retry = 1;
351 } while (bypassed--);
354 m->current_pgpath = NULL;
355 m->current_pg = NULL;
359 * Check whether bios must be queued in the device-mapper core rather
360 * than here in the target.
362 * m->lock must be held on entry.
364 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
365 * same value then we are not between multipath_presuspend()
366 * and multipath_resume() calls and we have no need to check
367 * for the DMF_NOFLUSH_SUSPENDING flag.
369 static int __must_push_back(struct multipath *m)
371 return (m->queue_if_no_path ||
372 (m->queue_if_no_path != m->saved_queue_if_no_path &&
373 dm_noflush_suspending(m->ti)));
376 #define pg_ready(m) (!(m)->queue_io && !(m)->pg_init_required)
378 static int map_io(struct multipath *m, struct request *clone,
379 union map_info *map_context)
381 int r = DM_MAPIO_REMAPPED;
382 size_t nr_bytes = blk_rq_bytes(clone);
384 struct pgpath *pgpath;
385 struct block_device *bdev;
386 struct dm_mpath_io *mpio = map_context->ptr;
388 spin_lock_irqsave(&m->lock, flags);
390 /* Do we need to select a new pgpath? */
391 if (!m->current_pgpath ||
392 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
393 __choose_pgpath(m, nr_bytes);
395 pgpath = m->current_pgpath;
399 bdev = pgpath->path.dev->bdev;
400 clone->q = bdev_get_queue(bdev);
401 clone->rq_disk = bdev->bd_disk;
402 mpio->pgpath = pgpath;
403 mpio->nr_bytes = nr_bytes;
404 if (pgpath->pg->ps.type->start_io)
405 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
409 __pg_init_all_paths(m);
410 r = DM_MAPIO_REQUEUE;
414 if (__must_push_back(m))
415 r = DM_MAPIO_REQUEUE;
417 r = -EIO; /* Failed */
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)
441 dm_table_run_md_queue_async(m->ti->table);
443 spin_unlock_irqrestore(&m->lock, flags);
449 * An event is triggered whenever a path is taken out of use.
450 * Includes path failure and PG bypass.
452 static void trigger_event(struct work_struct *work)
454 struct multipath *m =
455 container_of(work, struct multipath, trigger_event);
457 dm_table_event(m->ti->table);
460 /*-----------------------------------------------------------------
461 * Constructor/argument parsing:
462 * <#multipath feature args> [<arg>]*
463 * <#hw_handler args> [hw_handler [<arg>]*]
465 * <initial priority group>
466 * [<selector> <#selector args> [<arg>]*
467 * <#paths> <#per-path selector args>
468 * [<path> [<arg>]* ]+ ]+
469 *---------------------------------------------------------------*/
470 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
471 struct dm_target *ti)
474 struct path_selector_type *pst;
477 static struct dm_arg _args[] = {
478 {0, 1024, "invalid number of path selector args"},
481 pst = dm_get_path_selector(dm_shift_arg(as));
483 ti->error = "unknown path selector type";
487 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
489 dm_put_path_selector(pst);
493 r = pst->create(&pg->ps, ps_argc, as->argv);
495 dm_put_path_selector(pst);
496 ti->error = "path selector constructor failed";
501 dm_consume_args(as, ps_argc);
506 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
507 struct dm_target *ti)
511 struct multipath *m = ti->private;
512 struct request_queue *q = NULL;
513 const char *attached_handler_name;
515 /* we need at least a path arg */
517 ti->error = "no device given";
518 return ERR_PTR(-EINVAL);
523 return ERR_PTR(-ENOMEM);
525 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
528 ti->error = "error getting device";
532 if (m->retain_attached_hw_handler || m->hw_handler_name)
533 q = bdev_get_queue(p->path.dev->bdev);
535 if (m->retain_attached_hw_handler) {
536 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
537 if (attached_handler_name) {
539 * Reset hw_handler_name to match the attached handler
540 * and clear any hw_handler_params associated with the
543 * NB. This modifies the table line to show the actual
544 * handler instead of the original table passed in.
546 kfree(m->hw_handler_name);
547 m->hw_handler_name = attached_handler_name;
549 kfree(m->hw_handler_params);
550 m->hw_handler_params = NULL;
554 if (m->hw_handler_name) {
556 * Increments scsi_dh reference, even when using an
557 * already-attached handler.
559 r = scsi_dh_attach(q, m->hw_handler_name);
562 * Already attached to different hw_handler:
563 * try to reattach with correct one.
566 r = scsi_dh_attach(q, m->hw_handler_name);
570 ti->error = "error attaching hardware handler";
571 dm_put_device(ti, p->path.dev);
575 if (m->hw_handler_params) {
576 r = scsi_dh_set_params(q, m->hw_handler_params);
578 ti->error = "unable to set hardware "
579 "handler parameters";
581 dm_put_device(ti, p->path.dev);
587 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
589 dm_put_device(ti, p->path.dev);
600 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
603 static struct dm_arg _args[] = {
604 {1, 1024, "invalid number of paths"},
605 {0, 1024, "invalid number of selector args"}
609 unsigned i, nr_selector_args, nr_args;
610 struct priority_group *pg;
611 struct dm_target *ti = m->ti;
615 ti->error = "not enough priority group arguments";
616 return ERR_PTR(-EINVAL);
619 pg = alloc_priority_group();
621 ti->error = "couldn't allocate priority group";
622 return ERR_PTR(-ENOMEM);
626 r = parse_path_selector(as, pg, ti);
633 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
637 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
641 nr_args = 1 + nr_selector_args;
642 for (i = 0; i < pg->nr_pgpaths; i++) {
643 struct pgpath *pgpath;
644 struct dm_arg_set path_args;
646 if (as->argc < nr_args) {
647 ti->error = "not enough path parameters";
652 path_args.argc = nr_args;
653 path_args.argv = as->argv;
655 pgpath = parse_path(&path_args, &pg->ps, ti);
656 if (IS_ERR(pgpath)) {
662 list_add_tail(&pgpath->list, &pg->pgpaths);
663 dm_consume_args(as, nr_args);
669 free_priority_group(pg, ti);
673 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
677 struct dm_target *ti = m->ti;
679 static struct dm_arg _args[] = {
680 {0, 1024, "invalid number of hardware handler args"},
683 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
689 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
690 if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
691 "scsi_dh_%s", m->hw_handler_name)) {
692 ti->error = "unknown hardware handler type";
701 for (i = 0; i <= hw_argc - 2; i++)
702 len += strlen(as->argv[i]) + 1;
703 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
705 ti->error = "memory allocation failed";
709 j = sprintf(p, "%d", hw_argc - 1);
710 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
711 j = sprintf(p, "%s", as->argv[i]);
713 dm_consume_args(as, hw_argc - 1);
717 kfree(m->hw_handler_name);
718 m->hw_handler_name = NULL;
722 static int parse_features(struct dm_arg_set *as, struct multipath *m)
726 struct dm_target *ti = m->ti;
727 const char *arg_name;
729 static struct dm_arg _args[] = {
730 {0, 6, "invalid number of feature args"},
731 {1, 50, "pg_init_retries must be between 1 and 50"},
732 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
735 r = dm_read_arg_group(_args, as, &argc, &ti->error);
743 arg_name = dm_shift_arg(as);
746 if (!strcasecmp(arg_name, "queue_if_no_path")) {
747 r = queue_if_no_path(m, 1, 0);
751 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
752 m->retain_attached_hw_handler = 1;
756 if (!strcasecmp(arg_name, "pg_init_retries") &&
758 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
763 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
765 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
770 ti->error = "Unrecognised multipath feature request";
772 } while (argc && !r);
777 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
780 /* target arguments */
781 static struct dm_arg _args[] = {
782 {0, 1024, "invalid number of priority groups"},
783 {0, 1024, "invalid initial priority group number"},
788 struct dm_arg_set as;
789 unsigned pg_count = 0;
790 unsigned next_pg_num;
795 m = alloc_multipath(ti);
797 ti->error = "can't allocate multipath";
801 r = parse_features(&as, m);
805 r = parse_hw_handler(&as, m);
809 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
813 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
817 if ((!m->nr_priority_groups && next_pg_num) ||
818 (m->nr_priority_groups && !next_pg_num)) {
819 ti->error = "invalid initial priority group";
824 /* parse the priority groups */
826 struct priority_group *pg;
828 pg = parse_priority_group(&as, m);
834 m->nr_valid_paths += pg->nr_pgpaths;
835 list_add_tail(&pg->list, &m->priority_groups);
837 pg->pg_num = pg_count;
842 if (pg_count != m->nr_priority_groups) {
843 ti->error = "priority group count mismatch";
848 ti->num_flush_bios = 1;
849 ti->num_discard_bios = 1;
850 ti->num_write_same_bios = 1;
859 static void multipath_wait_for_pg_init_completion(struct multipath *m)
861 DECLARE_WAITQUEUE(wait, current);
864 add_wait_queue(&m->pg_init_wait, &wait);
867 set_current_state(TASK_UNINTERRUPTIBLE);
869 spin_lock_irqsave(&m->lock, flags);
870 if (!m->pg_init_in_progress) {
871 spin_unlock_irqrestore(&m->lock, flags);
874 spin_unlock_irqrestore(&m->lock, flags);
878 set_current_state(TASK_RUNNING);
880 remove_wait_queue(&m->pg_init_wait, &wait);
883 static void flush_multipath_work(struct multipath *m)
887 spin_lock_irqsave(&m->lock, flags);
888 m->pg_init_disabled = 1;
889 spin_unlock_irqrestore(&m->lock, flags);
891 flush_workqueue(kmpath_handlerd);
892 multipath_wait_for_pg_init_completion(m);
893 flush_workqueue(kmultipathd);
894 flush_work(&m->trigger_event);
896 spin_lock_irqsave(&m->lock, flags);
897 m->pg_init_disabled = 0;
898 spin_unlock_irqrestore(&m->lock, flags);
901 static void multipath_dtr(struct dm_target *ti)
903 struct multipath *m = ti->private;
905 flush_multipath_work(m);
910 * Map cloned requests
912 static int multipath_map(struct dm_target *ti, struct request *clone,
913 union map_info *map_context)
916 struct multipath *m = (struct multipath *) ti->private;
918 if (set_mapinfo(m, map_context) < 0)
919 /* ENOMEM, requeue */
920 return DM_MAPIO_REQUEUE;
922 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
923 r = map_io(m, clone, map_context);
924 if (r < 0 || r == DM_MAPIO_REQUEUE)
925 clear_mapinfo(m, map_context);
931 * Take a path out of use.
933 static int fail_path(struct pgpath *pgpath)
936 struct multipath *m = pgpath->pg->m;
938 spin_lock_irqsave(&m->lock, flags);
940 if (!pgpath->is_active)
943 DMWARN("Failing path %s.", pgpath->path.dev->name);
945 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
946 pgpath->is_active = 0;
947 pgpath->fail_count++;
951 if (pgpath == m->current_pgpath)
952 m->current_pgpath = NULL;
954 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
955 pgpath->path.dev->name, m->nr_valid_paths);
957 schedule_work(&m->trigger_event);
960 spin_unlock_irqrestore(&m->lock, flags);
966 * Reinstate a previously-failed path
968 static int reinstate_path(struct pgpath *pgpath)
972 struct multipath *m = pgpath->pg->m;
974 spin_lock_irqsave(&m->lock, flags);
976 if (pgpath->is_active)
979 if (!pgpath->pg->ps.type->reinstate_path) {
980 DMWARN("Reinstate path not supported by path selector %s",
981 pgpath->pg->ps.type->name);
986 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
990 pgpath->is_active = 1;
992 if (!m->nr_valid_paths++) {
993 m->current_pgpath = NULL;
994 dm_table_run_md_queue_async(m->ti->table);
995 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
996 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
997 m->pg_init_in_progress++;
1000 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1001 pgpath->path.dev->name, m->nr_valid_paths);
1003 schedule_work(&m->trigger_event);
1006 spin_unlock_irqrestore(&m->lock, flags);
1012 * Fail or reinstate all paths that match the provided struct dm_dev.
1014 static int action_dev(struct multipath *m, struct dm_dev *dev,
1018 struct pgpath *pgpath;
1019 struct priority_group *pg;
1021 list_for_each_entry(pg, &m->priority_groups, list) {
1022 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1023 if (pgpath->path.dev == dev)
1032 * Temporarily try to avoid having to use the specified PG
1034 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1037 unsigned long flags;
1039 spin_lock_irqsave(&m->lock, flags);
1041 pg->bypassed = bypassed;
1042 m->current_pgpath = NULL;
1043 m->current_pg = NULL;
1045 spin_unlock_irqrestore(&m->lock, flags);
1047 schedule_work(&m->trigger_event);
1051 * Switch to using the specified PG from the next I/O that gets mapped
1053 static int switch_pg_num(struct multipath *m, const char *pgstr)
1055 struct priority_group *pg;
1057 unsigned long flags;
1060 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1061 (pgnum > m->nr_priority_groups)) {
1062 DMWARN("invalid PG number supplied to switch_pg_num");
1066 spin_lock_irqsave(&m->lock, flags);
1067 list_for_each_entry(pg, &m->priority_groups, list) {
1072 m->current_pgpath = NULL;
1073 m->current_pg = NULL;
1076 spin_unlock_irqrestore(&m->lock, flags);
1078 schedule_work(&m->trigger_event);
1083 * Set/clear bypassed status of a PG.
1084 * PGs are numbered upwards from 1 in the order they were declared.
1086 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1088 struct priority_group *pg;
1092 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1093 (pgnum > m->nr_priority_groups)) {
1094 DMWARN("invalid PG number supplied to bypass_pg");
1098 list_for_each_entry(pg, &m->priority_groups, list) {
1103 bypass_pg(m, pg, bypassed);
1108 * Should we retry pg_init immediately?
1110 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1112 unsigned long flags;
1113 int limit_reached = 0;
1115 spin_lock_irqsave(&m->lock, flags);
1117 if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1118 m->pg_init_required = 1;
1122 spin_unlock_irqrestore(&m->lock, flags);
1124 return limit_reached;
1127 static void pg_init_done(void *data, int errors)
1129 struct pgpath *pgpath = data;
1130 struct priority_group *pg = pgpath->pg;
1131 struct multipath *m = pg->m;
1132 unsigned long flags;
1133 unsigned delay_retry = 0;
1135 /* device or driver problems */
1140 if (!m->hw_handler_name) {
1144 DMERR("Could not failover the device: Handler scsi_dh_%s "
1145 "Error %d.", m->hw_handler_name, errors);
1147 * Fail path for now, so we do not ping pong
1151 case SCSI_DH_DEV_TEMP_BUSY:
1153 * Probably doing something like FW upgrade on the
1154 * controller so try the other pg.
1156 bypass_pg(m, pg, 1);
1159 /* Wait before retrying. */
1161 case SCSI_DH_IMM_RETRY:
1162 case SCSI_DH_RES_TEMP_UNAVAIL:
1163 if (pg_init_limit_reached(m, pgpath))
1169 * We probably do not want to fail the path for a device
1170 * error, but this is what the old dm did. In future
1171 * patches we can do more advanced handling.
1176 spin_lock_irqsave(&m->lock, flags);
1178 if (pgpath == m->current_pgpath) {
1179 DMERR("Could not failover device. Error %d.", errors);
1180 m->current_pgpath = NULL;
1181 m->current_pg = NULL;
1183 } else if (!m->pg_init_required)
1186 if (--m->pg_init_in_progress)
1187 /* Activations of other paths are still on going */
1190 if (m->pg_init_required) {
1191 m->pg_init_delay_retry = delay_retry;
1192 if (__pg_init_all_paths(m))
1198 * Wake up any thread waiting to suspend.
1200 wake_up(&m->pg_init_wait);
1203 spin_unlock_irqrestore(&m->lock, flags);
1206 static void activate_path(struct work_struct *work)
1208 struct pgpath *pgpath =
1209 container_of(work, struct pgpath, activate_path.work);
1211 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1212 pg_init_done, pgpath);
1215 static int noretry_error(int error)
1226 /* Anything else could be a path failure, so should be retried */
1233 static int do_end_io(struct multipath *m, struct request *clone,
1234 int error, struct dm_mpath_io *mpio)
1237 * We don't queue any clone request inside the multipath target
1238 * during end I/O handling, since those clone requests don't have
1239 * bio clones. If we queue them inside the multipath target,
1240 * we need to make bio clones, that requires memory allocation.
1241 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1242 * don't have bio clones.)
1243 * Instead of queueing the clone request here, we queue the original
1244 * request into dm core, which will remake a clone request and
1245 * clone bios for it and resubmit it later.
1247 int r = DM_ENDIO_REQUEUE;
1248 unsigned long flags;
1250 if (!error && !clone->errors)
1251 return 0; /* I/O complete */
1253 if (noretry_error(error)) {
1254 if ((clone->cmd_flags & REQ_WRITE_SAME) &&
1255 !clone->q->limits.max_write_same_sectors) {
1256 struct queue_limits *limits;
1258 /* device doesn't really support WRITE SAME, disable it */
1259 limits = dm_get_queue_limits(dm_table_get_md(m->ti->table));
1260 limits->max_write_same_sectors = 0;
1266 fail_path(mpio->pgpath);
1268 spin_lock_irqsave(&m->lock, flags);
1269 if (!m->nr_valid_paths) {
1270 if (!m->queue_if_no_path) {
1271 if (!__must_push_back(m))
1274 if (error == -EBADE)
1278 spin_unlock_irqrestore(&m->lock, flags);
1283 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1284 int error, union map_info *map_context)
1286 struct multipath *m = ti->private;
1287 struct dm_mpath_io *mpio = map_context->ptr;
1288 struct pgpath *pgpath;
1289 struct path_selector *ps;
1294 r = do_end_io(m, clone, error, mpio);
1295 pgpath = mpio->pgpath;
1297 ps = &pgpath->pg->ps;
1298 if (ps->type->end_io)
1299 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1301 clear_mapinfo(m, map_context);
1307 * Suspend can't complete until all the I/O is processed so if
1308 * the last path fails we must error any remaining I/O.
1309 * Note that if the freeze_bdev fails while suspending, the
1310 * queue_if_no_path state is lost - userspace should reset it.
1312 static void multipath_presuspend(struct dm_target *ti)
1314 struct multipath *m = (struct multipath *) ti->private;
1316 queue_if_no_path(m, 0, 1);
1319 static void multipath_postsuspend(struct dm_target *ti)
1321 struct multipath *m = ti->private;
1323 mutex_lock(&m->work_mutex);
1324 flush_multipath_work(m);
1325 mutex_unlock(&m->work_mutex);
1329 * Restore the queue_if_no_path setting.
1331 static void multipath_resume(struct dm_target *ti)
1333 struct multipath *m = (struct multipath *) ti->private;
1334 unsigned long flags;
1336 spin_lock_irqsave(&m->lock, flags);
1337 m->queue_if_no_path = m->saved_queue_if_no_path;
1338 spin_unlock_irqrestore(&m->lock, flags);
1342 * Info output has the following format:
1343 * num_multipath_feature_args [multipath_feature_args]*
1344 * num_handler_status_args [handler_status_args]*
1345 * num_groups init_group_number
1346 * [A|D|E num_ps_status_args [ps_status_args]*
1347 * num_paths num_selector_args
1348 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1350 * Table output has the following format (identical to the constructor string):
1351 * num_feature_args [features_args]*
1352 * num_handler_args hw_handler [hw_handler_args]*
1353 * num_groups init_group_number
1354 * [priority selector-name num_ps_args [ps_args]*
1355 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1357 static void multipath_status(struct dm_target *ti, status_type_t type,
1358 unsigned status_flags, char *result, unsigned maxlen)
1361 unsigned long flags;
1362 struct multipath *m = (struct multipath *) ti->private;
1363 struct priority_group *pg;
1368 spin_lock_irqsave(&m->lock, flags);
1371 if (type == STATUSTYPE_INFO)
1372 DMEMIT("2 %u %u ", m->queue_io, m->pg_init_count);
1374 DMEMIT("%u ", m->queue_if_no_path +
1375 (m->pg_init_retries > 0) * 2 +
1376 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1377 m->retain_attached_hw_handler);
1378 if (m->queue_if_no_path)
1379 DMEMIT("queue_if_no_path ");
1380 if (m->pg_init_retries)
1381 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1382 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1383 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1384 if (m->retain_attached_hw_handler)
1385 DMEMIT("retain_attached_hw_handler ");
1388 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1391 DMEMIT("1 %s ", m->hw_handler_name);
1393 DMEMIT("%u ", m->nr_priority_groups);
1396 pg_num = m->next_pg->pg_num;
1397 else if (m->current_pg)
1398 pg_num = m->current_pg->pg_num;
1400 pg_num = (m->nr_priority_groups ? 1 : 0);
1402 DMEMIT("%u ", pg_num);
1405 case STATUSTYPE_INFO:
1406 list_for_each_entry(pg, &m->priority_groups, list) {
1408 state = 'D'; /* Disabled */
1409 else if (pg == m->current_pg)
1410 state = 'A'; /* Currently Active */
1412 state = 'E'; /* Enabled */
1414 DMEMIT("%c ", state);
1416 if (pg->ps.type->status)
1417 sz += pg->ps.type->status(&pg->ps, NULL, type,
1423 DMEMIT("%u %u ", pg->nr_pgpaths,
1424 pg->ps.type->info_args);
1426 list_for_each_entry(p, &pg->pgpaths, list) {
1427 DMEMIT("%s %s %u ", p->path.dev->name,
1428 p->is_active ? "A" : "F",
1430 if (pg->ps.type->status)
1431 sz += pg->ps.type->status(&pg->ps,
1432 &p->path, type, result + sz,
1438 case STATUSTYPE_TABLE:
1439 list_for_each_entry(pg, &m->priority_groups, list) {
1440 DMEMIT("%s ", pg->ps.type->name);
1442 if (pg->ps.type->status)
1443 sz += pg->ps.type->status(&pg->ps, NULL, type,
1449 DMEMIT("%u %u ", pg->nr_pgpaths,
1450 pg->ps.type->table_args);
1452 list_for_each_entry(p, &pg->pgpaths, list) {
1453 DMEMIT("%s ", p->path.dev->name);
1454 if (pg->ps.type->status)
1455 sz += pg->ps.type->status(&pg->ps,
1456 &p->path, type, result + sz,
1463 spin_unlock_irqrestore(&m->lock, flags);
1466 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1470 struct multipath *m = (struct multipath *) ti->private;
1473 mutex_lock(&m->work_mutex);
1475 if (dm_suspended(ti)) {
1481 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1482 r = queue_if_no_path(m, 1, 0);
1484 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1485 r = queue_if_no_path(m, 0, 0);
1491 DMWARN("Unrecognised multipath message received.");
1495 if (!strcasecmp(argv[0], "disable_group")) {
1496 r = bypass_pg_num(m, argv[1], 1);
1498 } else if (!strcasecmp(argv[0], "enable_group")) {
1499 r = bypass_pg_num(m, argv[1], 0);
1501 } else if (!strcasecmp(argv[0], "switch_group")) {
1502 r = switch_pg_num(m, argv[1]);
1504 } else if (!strcasecmp(argv[0], "reinstate_path"))
1505 action = reinstate_path;
1506 else if (!strcasecmp(argv[0], "fail_path"))
1509 DMWARN("Unrecognised multipath message received.");
1513 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1515 DMWARN("message: error getting device %s",
1520 r = action_dev(m, dev, action);
1522 dm_put_device(ti, dev);
1525 mutex_unlock(&m->work_mutex);
1529 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1532 struct multipath *m = ti->private;
1533 struct pgpath *pgpath;
1534 struct block_device *bdev;
1536 unsigned long flags;
1543 spin_lock_irqsave(&m->lock, flags);
1545 if (!m->current_pgpath)
1546 __choose_pgpath(m, 0);
1548 pgpath = m->current_pgpath;
1551 bdev = pgpath->path.dev->bdev;
1552 mode = pgpath->path.dev->mode;
1555 if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1560 spin_unlock_irqrestore(&m->lock, flags);
1563 * Only pass ioctls through if the device sizes match exactly.
1565 if (!bdev || ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) {
1566 int err = scsi_verify_blk_ioctl(NULL, cmd);
1571 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
1572 spin_lock_irqsave(&m->lock, flags);
1573 if (!m->current_pg) {
1574 /* Path status changed, redo selection */
1575 __choose_pgpath(m, 0);
1577 if (m->pg_init_required)
1578 __pg_init_all_paths(m);
1579 spin_unlock_irqrestore(&m->lock, flags);
1580 dm_table_run_md_queue_async(m->ti->table);
1583 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1586 static int multipath_iterate_devices(struct dm_target *ti,
1587 iterate_devices_callout_fn fn, void *data)
1589 struct multipath *m = ti->private;
1590 struct priority_group *pg;
1594 list_for_each_entry(pg, &m->priority_groups, list) {
1595 list_for_each_entry(p, &pg->pgpaths, list) {
1596 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1606 static int __pgpath_busy(struct pgpath *pgpath)
1608 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1610 return dm_underlying_device_busy(q);
1614 * We return "busy", only when we can map I/Os but underlying devices
1615 * are busy (so even if we map I/Os now, the I/Os will wait on
1616 * the underlying queue).
1617 * In other words, if we want to kill I/Os or queue them inside us
1618 * due to map unavailability, we don't return "busy". Otherwise,
1619 * dm core won't give us the I/Os and we can't do what we want.
1621 static int multipath_busy(struct dm_target *ti)
1623 int busy = 0, has_active = 0;
1624 struct multipath *m = ti->private;
1625 struct priority_group *pg;
1626 struct pgpath *pgpath;
1627 unsigned long flags;
1629 spin_lock_irqsave(&m->lock, flags);
1631 /* pg_init in progress, requeue until done */
1636 /* Guess which priority_group will be used at next mapping time */
1637 if (unlikely(!m->current_pgpath && m->next_pg))
1639 else if (likely(m->current_pg))
1643 * We don't know which pg will be used at next mapping time.
1644 * We don't call __choose_pgpath() here to avoid to trigger
1645 * pg_init just by busy checking.
1646 * So we don't know whether underlying devices we will be using
1647 * at next mapping time are busy or not. Just try mapping.
1652 * If there is one non-busy active path at least, the path selector
1653 * will be able to select it. So we consider such a pg as not busy.
1656 list_for_each_entry(pgpath, &pg->pgpaths, list)
1657 if (pgpath->is_active) {
1660 if (!__pgpath_busy(pgpath)) {
1668 * No active path in this pg, so this pg won't be used and
1669 * the current_pg will be changed at next mapping time.
1670 * We need to try mapping to determine it.
1675 spin_unlock_irqrestore(&m->lock, flags);
1680 /*-----------------------------------------------------------------
1682 *---------------------------------------------------------------*/
1683 static struct target_type multipath_target = {
1684 .name = "multipath",
1685 .version = {1, 7, 0},
1686 .module = THIS_MODULE,
1687 .ctr = multipath_ctr,
1688 .dtr = multipath_dtr,
1689 .map_rq = multipath_map,
1690 .rq_end_io = multipath_end_io,
1691 .presuspend = multipath_presuspend,
1692 .postsuspend = multipath_postsuspend,
1693 .resume = multipath_resume,
1694 .status = multipath_status,
1695 .message = multipath_message,
1696 .ioctl = multipath_ioctl,
1697 .iterate_devices = multipath_iterate_devices,
1698 .busy = multipath_busy,
1701 static int __init dm_multipath_init(void)
1705 /* allocate a slab for the dm_ios */
1706 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1710 r = dm_register_target(&multipath_target);
1712 DMERR("register failed %d", r);
1713 kmem_cache_destroy(_mpio_cache);
1717 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1719 DMERR("failed to create workqueue kmpathd");
1720 dm_unregister_target(&multipath_target);
1721 kmem_cache_destroy(_mpio_cache);
1726 * A separate workqueue is used to handle the device handlers
1727 * to avoid overloading existing workqueue. Overloading the
1728 * old workqueue would also create a bottleneck in the
1729 * path of the storage hardware device activation.
1731 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1733 if (!kmpath_handlerd) {
1734 DMERR("failed to create workqueue kmpath_handlerd");
1735 destroy_workqueue(kmultipathd);
1736 dm_unregister_target(&multipath_target);
1737 kmem_cache_destroy(_mpio_cache);
1741 DMINFO("version %u.%u.%u loaded",
1742 multipath_target.version[0], multipath_target.version[1],
1743 multipath_target.version[2]);
1748 static void __exit dm_multipath_exit(void)
1750 destroy_workqueue(kmpath_handlerd);
1751 destroy_workqueue(kmultipathd);
1753 dm_unregister_target(&multipath_target);
1754 kmem_cache_destroy(_mpio_cache);
1757 module_init(dm_multipath_init);
1758 module_exit(dm_multipath_exit);
1760 MODULE_DESCRIPTION(DM_NAME " multipath target");
1761 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1762 MODULE_LICENSE("GPL");