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_REQUEUE;
382 size_t nr_bytes = blk_rq_bytes(clone);
384 struct pgpath *pgpath;
385 struct block_device *bdev;
386 struct dm_mpath_io *mpio;
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 if (set_mapinfo(m, map_context) < 0)
400 /* ENOMEM, requeue */
403 bdev = pgpath->path.dev->bdev;
404 clone->q = bdev_get_queue(bdev);
405 clone->rq_disk = bdev->bd_disk;
406 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
407 mpio = map_context->ptr;
408 mpio->pgpath = pgpath;
409 mpio->nr_bytes = nr_bytes;
410 if (pgpath->pg->ps.type->start_io)
411 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
414 r = DM_MAPIO_REMAPPED;
417 __pg_init_all_paths(m);
418 } else if (!__must_push_back(m))
419 r = -EIO; /* Failed */
422 spin_unlock_irqrestore(&m->lock, flags);
428 * If we run out of usable paths, should we queue I/O or error it?
430 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
431 unsigned save_old_value)
435 spin_lock_irqsave(&m->lock, flags);
438 m->saved_queue_if_no_path = m->queue_if_no_path;
440 m->saved_queue_if_no_path = queue_if_no_path;
441 m->queue_if_no_path = queue_if_no_path;
442 if (!m->queue_if_no_path)
443 dm_table_run_md_queue_async(m->ti->table);
445 spin_unlock_irqrestore(&m->lock, flags);
451 * An event is triggered whenever a path is taken out of use.
452 * Includes path failure and PG bypass.
454 static void trigger_event(struct work_struct *work)
456 struct multipath *m =
457 container_of(work, struct multipath, trigger_event);
459 dm_table_event(m->ti->table);
462 /*-----------------------------------------------------------------
463 * Constructor/argument parsing:
464 * <#multipath feature args> [<arg>]*
465 * <#hw_handler args> [hw_handler [<arg>]*]
467 * <initial priority group>
468 * [<selector> <#selector args> [<arg>]*
469 * <#paths> <#per-path selector args>
470 * [<path> [<arg>]* ]+ ]+
471 *---------------------------------------------------------------*/
472 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
473 struct dm_target *ti)
476 struct path_selector_type *pst;
479 static struct dm_arg _args[] = {
480 {0, 1024, "invalid number of path selector args"},
483 pst = dm_get_path_selector(dm_shift_arg(as));
485 ti->error = "unknown path selector type";
489 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
491 dm_put_path_selector(pst);
495 r = pst->create(&pg->ps, ps_argc, as->argv);
497 dm_put_path_selector(pst);
498 ti->error = "path selector constructor failed";
503 dm_consume_args(as, ps_argc);
508 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
509 struct dm_target *ti)
513 struct multipath *m = ti->private;
514 struct request_queue *q = NULL;
515 const char *attached_handler_name;
517 /* we need at least a path arg */
519 ti->error = "no device given";
520 return ERR_PTR(-EINVAL);
525 return ERR_PTR(-ENOMEM);
527 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
530 ti->error = "error getting device";
534 if (m->retain_attached_hw_handler || m->hw_handler_name)
535 q = bdev_get_queue(p->path.dev->bdev);
537 if (m->retain_attached_hw_handler) {
538 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
539 if (attached_handler_name) {
541 * Reset hw_handler_name to match the attached handler
542 * and clear any hw_handler_params associated with the
545 * NB. This modifies the table line to show the actual
546 * handler instead of the original table passed in.
548 kfree(m->hw_handler_name);
549 m->hw_handler_name = attached_handler_name;
551 kfree(m->hw_handler_params);
552 m->hw_handler_params = NULL;
556 if (m->hw_handler_name) {
558 * Increments scsi_dh reference, even when using an
559 * already-attached handler.
561 r = scsi_dh_attach(q, m->hw_handler_name);
564 * Already attached to different hw_handler:
565 * try to reattach with correct one.
568 r = scsi_dh_attach(q, m->hw_handler_name);
572 ti->error = "error attaching hardware handler";
573 dm_put_device(ti, p->path.dev);
577 if (m->hw_handler_params) {
578 r = scsi_dh_set_params(q, m->hw_handler_params);
580 ti->error = "unable to set hardware "
581 "handler parameters";
583 dm_put_device(ti, p->path.dev);
589 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
591 dm_put_device(ti, p->path.dev);
602 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
605 static struct dm_arg _args[] = {
606 {1, 1024, "invalid number of paths"},
607 {0, 1024, "invalid number of selector args"}
611 unsigned i, nr_selector_args, nr_args;
612 struct priority_group *pg;
613 struct dm_target *ti = m->ti;
617 ti->error = "not enough priority group arguments";
618 return ERR_PTR(-EINVAL);
621 pg = alloc_priority_group();
623 ti->error = "couldn't allocate priority group";
624 return ERR_PTR(-ENOMEM);
628 r = parse_path_selector(as, pg, ti);
635 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
639 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
643 nr_args = 1 + nr_selector_args;
644 for (i = 0; i < pg->nr_pgpaths; i++) {
645 struct pgpath *pgpath;
646 struct dm_arg_set path_args;
648 if (as->argc < nr_args) {
649 ti->error = "not enough path parameters";
654 path_args.argc = nr_args;
655 path_args.argv = as->argv;
657 pgpath = parse_path(&path_args, &pg->ps, ti);
658 if (IS_ERR(pgpath)) {
664 list_add_tail(&pgpath->list, &pg->pgpaths);
665 dm_consume_args(as, nr_args);
671 free_priority_group(pg, ti);
675 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
679 struct dm_target *ti = m->ti;
681 static struct dm_arg _args[] = {
682 {0, 1024, "invalid number of hardware handler args"},
685 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
691 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
692 if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
693 "scsi_dh_%s", m->hw_handler_name)) {
694 ti->error = "unknown hardware handler type";
703 for (i = 0; i <= hw_argc - 2; i++)
704 len += strlen(as->argv[i]) + 1;
705 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
707 ti->error = "memory allocation failed";
711 j = sprintf(p, "%d", hw_argc - 1);
712 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
713 j = sprintf(p, "%s", as->argv[i]);
715 dm_consume_args(as, hw_argc - 1);
719 kfree(m->hw_handler_name);
720 m->hw_handler_name = NULL;
724 static int parse_features(struct dm_arg_set *as, struct multipath *m)
728 struct dm_target *ti = m->ti;
729 const char *arg_name;
731 static struct dm_arg _args[] = {
732 {0, 6, "invalid number of feature args"},
733 {1, 50, "pg_init_retries must be between 1 and 50"},
734 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
737 r = dm_read_arg_group(_args, as, &argc, &ti->error);
745 arg_name = dm_shift_arg(as);
748 if (!strcasecmp(arg_name, "queue_if_no_path")) {
749 r = queue_if_no_path(m, 1, 0);
753 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
754 m->retain_attached_hw_handler = 1;
758 if (!strcasecmp(arg_name, "pg_init_retries") &&
760 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
765 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
767 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
772 ti->error = "Unrecognised multipath feature request";
774 } while (argc && !r);
779 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
782 /* target arguments */
783 static struct dm_arg _args[] = {
784 {0, 1024, "invalid number of priority groups"},
785 {0, 1024, "invalid initial priority group number"},
790 struct dm_arg_set as;
791 unsigned pg_count = 0;
792 unsigned next_pg_num;
797 m = alloc_multipath(ti);
799 ti->error = "can't allocate multipath";
803 r = parse_features(&as, m);
807 r = parse_hw_handler(&as, m);
811 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
815 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
819 if ((!m->nr_priority_groups && next_pg_num) ||
820 (m->nr_priority_groups && !next_pg_num)) {
821 ti->error = "invalid initial priority group";
826 /* parse the priority groups */
828 struct priority_group *pg;
830 pg = parse_priority_group(&as, m);
836 m->nr_valid_paths += pg->nr_pgpaths;
837 list_add_tail(&pg->list, &m->priority_groups);
839 pg->pg_num = pg_count;
844 if (pg_count != m->nr_priority_groups) {
845 ti->error = "priority group count mismatch";
850 ti->num_flush_bios = 1;
851 ti->num_discard_bios = 1;
852 ti->num_write_same_bios = 1;
861 static void multipath_wait_for_pg_init_completion(struct multipath *m)
863 DECLARE_WAITQUEUE(wait, current);
866 add_wait_queue(&m->pg_init_wait, &wait);
869 set_current_state(TASK_UNINTERRUPTIBLE);
871 spin_lock_irqsave(&m->lock, flags);
872 if (!m->pg_init_in_progress) {
873 spin_unlock_irqrestore(&m->lock, flags);
876 spin_unlock_irqrestore(&m->lock, flags);
880 set_current_state(TASK_RUNNING);
882 remove_wait_queue(&m->pg_init_wait, &wait);
885 static void flush_multipath_work(struct multipath *m)
889 spin_lock_irqsave(&m->lock, flags);
890 m->pg_init_disabled = 1;
891 spin_unlock_irqrestore(&m->lock, flags);
893 flush_workqueue(kmpath_handlerd);
894 multipath_wait_for_pg_init_completion(m);
895 flush_workqueue(kmultipathd);
896 flush_work(&m->trigger_event);
898 spin_lock_irqsave(&m->lock, flags);
899 m->pg_init_disabled = 0;
900 spin_unlock_irqrestore(&m->lock, flags);
903 static void multipath_dtr(struct dm_target *ti)
905 struct multipath *m = ti->private;
907 flush_multipath_work(m);
912 * Map cloned requests
914 static int multipath_map(struct dm_target *ti, struct request *clone,
915 union map_info *map_context)
917 struct multipath *m = (struct multipath *) ti->private;
919 return map_io(m, clone, map_context);
923 * Take a path out of use.
925 static int fail_path(struct pgpath *pgpath)
928 struct multipath *m = pgpath->pg->m;
930 spin_lock_irqsave(&m->lock, flags);
932 if (!pgpath->is_active)
935 DMWARN("Failing path %s.", pgpath->path.dev->name);
937 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
938 pgpath->is_active = 0;
939 pgpath->fail_count++;
943 if (pgpath == m->current_pgpath)
944 m->current_pgpath = NULL;
946 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
947 pgpath->path.dev->name, m->nr_valid_paths);
949 schedule_work(&m->trigger_event);
952 spin_unlock_irqrestore(&m->lock, flags);
958 * Reinstate a previously-failed path
960 static int reinstate_path(struct pgpath *pgpath)
964 struct multipath *m = pgpath->pg->m;
966 spin_lock_irqsave(&m->lock, flags);
968 if (pgpath->is_active)
971 if (!pgpath->pg->ps.type->reinstate_path) {
972 DMWARN("Reinstate path not supported by path selector %s",
973 pgpath->pg->ps.type->name);
978 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
982 pgpath->is_active = 1;
984 if (!m->nr_valid_paths++) {
985 m->current_pgpath = NULL;
986 dm_table_run_md_queue_async(m->ti->table);
987 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
988 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
989 m->pg_init_in_progress++;
992 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
993 pgpath->path.dev->name, m->nr_valid_paths);
995 schedule_work(&m->trigger_event);
998 spin_unlock_irqrestore(&m->lock, flags);
1004 * Fail or reinstate all paths that match the provided struct dm_dev.
1006 static int action_dev(struct multipath *m, struct dm_dev *dev,
1010 struct pgpath *pgpath;
1011 struct priority_group *pg;
1013 list_for_each_entry(pg, &m->priority_groups, list) {
1014 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1015 if (pgpath->path.dev == dev)
1024 * Temporarily try to avoid having to use the specified PG
1026 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1029 unsigned long flags;
1031 spin_lock_irqsave(&m->lock, flags);
1033 pg->bypassed = bypassed;
1034 m->current_pgpath = NULL;
1035 m->current_pg = NULL;
1037 spin_unlock_irqrestore(&m->lock, flags);
1039 schedule_work(&m->trigger_event);
1043 * Switch to using the specified PG from the next I/O that gets mapped
1045 static int switch_pg_num(struct multipath *m, const char *pgstr)
1047 struct priority_group *pg;
1049 unsigned long flags;
1052 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1053 (pgnum > m->nr_priority_groups)) {
1054 DMWARN("invalid PG number supplied to switch_pg_num");
1058 spin_lock_irqsave(&m->lock, flags);
1059 list_for_each_entry(pg, &m->priority_groups, list) {
1064 m->current_pgpath = NULL;
1065 m->current_pg = NULL;
1068 spin_unlock_irqrestore(&m->lock, flags);
1070 schedule_work(&m->trigger_event);
1075 * Set/clear bypassed status of a PG.
1076 * PGs are numbered upwards from 1 in the order they were declared.
1078 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1080 struct priority_group *pg;
1084 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1085 (pgnum > m->nr_priority_groups)) {
1086 DMWARN("invalid PG number supplied to bypass_pg");
1090 list_for_each_entry(pg, &m->priority_groups, list) {
1095 bypass_pg(m, pg, bypassed);
1100 * Should we retry pg_init immediately?
1102 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1104 unsigned long flags;
1105 int limit_reached = 0;
1107 spin_lock_irqsave(&m->lock, flags);
1109 if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1110 m->pg_init_required = 1;
1114 spin_unlock_irqrestore(&m->lock, flags);
1116 return limit_reached;
1119 static void pg_init_done(void *data, int errors)
1121 struct pgpath *pgpath = data;
1122 struct priority_group *pg = pgpath->pg;
1123 struct multipath *m = pg->m;
1124 unsigned long flags;
1125 unsigned delay_retry = 0;
1127 /* device or driver problems */
1132 if (!m->hw_handler_name) {
1136 DMERR("Could not failover the device: Handler scsi_dh_%s "
1137 "Error %d.", m->hw_handler_name, errors);
1139 * Fail path for now, so we do not ping pong
1143 case SCSI_DH_DEV_TEMP_BUSY:
1145 * Probably doing something like FW upgrade on the
1146 * controller so try the other pg.
1148 bypass_pg(m, pg, 1);
1151 /* Wait before retrying. */
1153 case SCSI_DH_IMM_RETRY:
1154 case SCSI_DH_RES_TEMP_UNAVAIL:
1155 if (pg_init_limit_reached(m, pgpath))
1161 * We probably do not want to fail the path for a device
1162 * error, but this is what the old dm did. In future
1163 * patches we can do more advanced handling.
1168 spin_lock_irqsave(&m->lock, flags);
1170 if (pgpath == m->current_pgpath) {
1171 DMERR("Could not failover device. Error %d.", errors);
1172 m->current_pgpath = NULL;
1173 m->current_pg = NULL;
1175 } else if (!m->pg_init_required)
1178 if (--m->pg_init_in_progress)
1179 /* Activations of other paths are still on going */
1182 if (m->pg_init_required) {
1183 m->pg_init_delay_retry = delay_retry;
1184 if (__pg_init_all_paths(m))
1190 * Wake up any thread waiting to suspend.
1192 wake_up(&m->pg_init_wait);
1195 spin_unlock_irqrestore(&m->lock, flags);
1198 static void activate_path(struct work_struct *work)
1200 struct pgpath *pgpath =
1201 container_of(work, struct pgpath, activate_path.work);
1203 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1204 pg_init_done, pgpath);
1207 static int noretry_error(int error)
1218 /* Anything else could be a path failure, so should be retried */
1225 static int do_end_io(struct multipath *m, struct request *clone,
1226 int error, struct dm_mpath_io *mpio)
1229 * We don't queue any clone request inside the multipath target
1230 * during end I/O handling, since those clone requests don't have
1231 * bio clones. If we queue them inside the multipath target,
1232 * we need to make bio clones, that requires memory allocation.
1233 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1234 * don't have bio clones.)
1235 * Instead of queueing the clone request here, we queue the original
1236 * request into dm core, which will remake a clone request and
1237 * clone bios for it and resubmit it later.
1239 int r = DM_ENDIO_REQUEUE;
1240 unsigned long flags;
1242 if (!error && !clone->errors)
1243 return 0; /* I/O complete */
1245 if (noretry_error(error)) {
1246 if ((clone->cmd_flags & REQ_WRITE_SAME) &&
1247 !clone->q->limits.max_write_same_sectors) {
1248 struct queue_limits *limits;
1250 /* device doesn't really support WRITE SAME, disable it */
1251 limits = dm_get_queue_limits(dm_table_get_md(m->ti->table));
1252 limits->max_write_same_sectors = 0;
1258 fail_path(mpio->pgpath);
1260 spin_lock_irqsave(&m->lock, flags);
1261 if (!m->nr_valid_paths) {
1262 if (!m->queue_if_no_path) {
1263 if (!__must_push_back(m))
1266 if (error == -EBADE)
1270 spin_unlock_irqrestore(&m->lock, flags);
1275 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1276 int error, union map_info *map_context)
1278 struct multipath *m = ti->private;
1279 struct dm_mpath_io *mpio = map_context->ptr;
1280 struct pgpath *pgpath;
1281 struct path_selector *ps;
1286 r = do_end_io(m, clone, error, mpio);
1287 pgpath = mpio->pgpath;
1289 ps = &pgpath->pg->ps;
1290 if (ps->type->end_io)
1291 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1293 clear_mapinfo(m, map_context);
1299 * Suspend can't complete until all the I/O is processed so if
1300 * the last path fails we must error any remaining I/O.
1301 * Note that if the freeze_bdev fails while suspending, the
1302 * queue_if_no_path state is lost - userspace should reset it.
1304 static void multipath_presuspend(struct dm_target *ti)
1306 struct multipath *m = (struct multipath *) ti->private;
1308 queue_if_no_path(m, 0, 1);
1311 static void multipath_postsuspend(struct dm_target *ti)
1313 struct multipath *m = ti->private;
1315 mutex_lock(&m->work_mutex);
1316 flush_multipath_work(m);
1317 mutex_unlock(&m->work_mutex);
1321 * Restore the queue_if_no_path setting.
1323 static void multipath_resume(struct dm_target *ti)
1325 struct multipath *m = (struct multipath *) ti->private;
1326 unsigned long flags;
1328 spin_lock_irqsave(&m->lock, flags);
1329 m->queue_if_no_path = m->saved_queue_if_no_path;
1330 spin_unlock_irqrestore(&m->lock, flags);
1334 * Info output has the following format:
1335 * num_multipath_feature_args [multipath_feature_args]*
1336 * num_handler_status_args [handler_status_args]*
1337 * num_groups init_group_number
1338 * [A|D|E num_ps_status_args [ps_status_args]*
1339 * num_paths num_selector_args
1340 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1342 * Table output has the following format (identical to the constructor string):
1343 * num_feature_args [features_args]*
1344 * num_handler_args hw_handler [hw_handler_args]*
1345 * num_groups init_group_number
1346 * [priority selector-name num_ps_args [ps_args]*
1347 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1349 static void multipath_status(struct dm_target *ti, status_type_t type,
1350 unsigned status_flags, char *result, unsigned maxlen)
1353 unsigned long flags;
1354 struct multipath *m = (struct multipath *) ti->private;
1355 struct priority_group *pg;
1360 spin_lock_irqsave(&m->lock, flags);
1363 if (type == STATUSTYPE_INFO)
1364 DMEMIT("2 %u %u ", m->queue_io, m->pg_init_count);
1366 DMEMIT("%u ", m->queue_if_no_path +
1367 (m->pg_init_retries > 0) * 2 +
1368 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1369 m->retain_attached_hw_handler);
1370 if (m->queue_if_no_path)
1371 DMEMIT("queue_if_no_path ");
1372 if (m->pg_init_retries)
1373 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1374 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1375 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1376 if (m->retain_attached_hw_handler)
1377 DMEMIT("retain_attached_hw_handler ");
1380 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1383 DMEMIT("1 %s ", m->hw_handler_name);
1385 DMEMIT("%u ", m->nr_priority_groups);
1388 pg_num = m->next_pg->pg_num;
1389 else if (m->current_pg)
1390 pg_num = m->current_pg->pg_num;
1392 pg_num = (m->nr_priority_groups ? 1 : 0);
1394 DMEMIT("%u ", pg_num);
1397 case STATUSTYPE_INFO:
1398 list_for_each_entry(pg, &m->priority_groups, list) {
1400 state = 'D'; /* Disabled */
1401 else if (pg == m->current_pg)
1402 state = 'A'; /* Currently Active */
1404 state = 'E'; /* Enabled */
1406 DMEMIT("%c ", state);
1408 if (pg->ps.type->status)
1409 sz += pg->ps.type->status(&pg->ps, NULL, type,
1415 DMEMIT("%u %u ", pg->nr_pgpaths,
1416 pg->ps.type->info_args);
1418 list_for_each_entry(p, &pg->pgpaths, list) {
1419 DMEMIT("%s %s %u ", p->path.dev->name,
1420 p->is_active ? "A" : "F",
1422 if (pg->ps.type->status)
1423 sz += pg->ps.type->status(&pg->ps,
1424 &p->path, type, result + sz,
1430 case STATUSTYPE_TABLE:
1431 list_for_each_entry(pg, &m->priority_groups, list) {
1432 DMEMIT("%s ", pg->ps.type->name);
1434 if (pg->ps.type->status)
1435 sz += pg->ps.type->status(&pg->ps, NULL, type,
1441 DMEMIT("%u %u ", pg->nr_pgpaths,
1442 pg->ps.type->table_args);
1444 list_for_each_entry(p, &pg->pgpaths, list) {
1445 DMEMIT("%s ", p->path.dev->name);
1446 if (pg->ps.type->status)
1447 sz += pg->ps.type->status(&pg->ps,
1448 &p->path, type, result + sz,
1455 spin_unlock_irqrestore(&m->lock, flags);
1458 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1462 struct multipath *m = (struct multipath *) ti->private;
1465 mutex_lock(&m->work_mutex);
1467 if (dm_suspended(ti)) {
1473 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1474 r = queue_if_no_path(m, 1, 0);
1476 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1477 r = queue_if_no_path(m, 0, 0);
1483 DMWARN("Unrecognised multipath message received.");
1487 if (!strcasecmp(argv[0], "disable_group")) {
1488 r = bypass_pg_num(m, argv[1], 1);
1490 } else if (!strcasecmp(argv[0], "enable_group")) {
1491 r = bypass_pg_num(m, argv[1], 0);
1493 } else if (!strcasecmp(argv[0], "switch_group")) {
1494 r = switch_pg_num(m, argv[1]);
1496 } else if (!strcasecmp(argv[0], "reinstate_path"))
1497 action = reinstate_path;
1498 else if (!strcasecmp(argv[0], "fail_path"))
1501 DMWARN("Unrecognised multipath message received.");
1505 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1507 DMWARN("message: error getting device %s",
1512 r = action_dev(m, dev, action);
1514 dm_put_device(ti, dev);
1517 mutex_unlock(&m->work_mutex);
1521 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1524 struct multipath *m = ti->private;
1525 struct pgpath *pgpath;
1526 struct block_device *bdev;
1528 unsigned long flags;
1535 spin_lock_irqsave(&m->lock, flags);
1537 if (!m->current_pgpath)
1538 __choose_pgpath(m, 0);
1540 pgpath = m->current_pgpath;
1543 bdev = pgpath->path.dev->bdev;
1544 mode = pgpath->path.dev->mode;
1547 if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1552 spin_unlock_irqrestore(&m->lock, flags);
1555 * Only pass ioctls through if the device sizes match exactly.
1557 if (!bdev || ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) {
1558 int err = scsi_verify_blk_ioctl(NULL, cmd);
1563 if (r == -ENOTCONN && !fatal_signal_pending(current)) {
1564 spin_lock_irqsave(&m->lock, flags);
1565 if (!m->current_pg) {
1566 /* Path status changed, redo selection */
1567 __choose_pgpath(m, 0);
1569 if (m->pg_init_required)
1570 __pg_init_all_paths(m);
1571 spin_unlock_irqrestore(&m->lock, flags);
1572 dm_table_run_md_queue_async(m->ti->table);
1575 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1578 static int multipath_iterate_devices(struct dm_target *ti,
1579 iterate_devices_callout_fn fn, void *data)
1581 struct multipath *m = ti->private;
1582 struct priority_group *pg;
1586 list_for_each_entry(pg, &m->priority_groups, list) {
1587 list_for_each_entry(p, &pg->pgpaths, list) {
1588 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1598 static int __pgpath_busy(struct pgpath *pgpath)
1600 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1602 return dm_underlying_device_busy(q);
1606 * We return "busy", only when we can map I/Os but underlying devices
1607 * are busy (so even if we map I/Os now, the I/Os will wait on
1608 * the underlying queue).
1609 * In other words, if we want to kill I/Os or queue them inside us
1610 * due to map unavailability, we don't return "busy". Otherwise,
1611 * dm core won't give us the I/Os and we can't do what we want.
1613 static int multipath_busy(struct dm_target *ti)
1615 int busy = 0, has_active = 0;
1616 struct multipath *m = ti->private;
1617 struct priority_group *pg;
1618 struct pgpath *pgpath;
1619 unsigned long flags;
1621 spin_lock_irqsave(&m->lock, flags);
1623 /* pg_init in progress, requeue until done */
1628 /* Guess which priority_group will be used at next mapping time */
1629 if (unlikely(!m->current_pgpath && m->next_pg))
1631 else if (likely(m->current_pg))
1635 * We don't know which pg will be used at next mapping time.
1636 * We don't call __choose_pgpath() here to avoid to trigger
1637 * pg_init just by busy checking.
1638 * So we don't know whether underlying devices we will be using
1639 * at next mapping time are busy or not. Just try mapping.
1644 * If there is one non-busy active path at least, the path selector
1645 * will be able to select it. So we consider such a pg as not busy.
1648 list_for_each_entry(pgpath, &pg->pgpaths, list)
1649 if (pgpath->is_active) {
1652 if (!__pgpath_busy(pgpath)) {
1660 * No active path in this pg, so this pg won't be used and
1661 * the current_pg will be changed at next mapping time.
1662 * We need to try mapping to determine it.
1667 spin_unlock_irqrestore(&m->lock, flags);
1672 /*-----------------------------------------------------------------
1674 *---------------------------------------------------------------*/
1675 static struct target_type multipath_target = {
1676 .name = "multipath",
1677 .version = {1, 7, 0},
1678 .module = THIS_MODULE,
1679 .ctr = multipath_ctr,
1680 .dtr = multipath_dtr,
1681 .map_rq = multipath_map,
1682 .rq_end_io = multipath_end_io,
1683 .presuspend = multipath_presuspend,
1684 .postsuspend = multipath_postsuspend,
1685 .resume = multipath_resume,
1686 .status = multipath_status,
1687 .message = multipath_message,
1688 .ioctl = multipath_ioctl,
1689 .iterate_devices = multipath_iterate_devices,
1690 .busy = multipath_busy,
1693 static int __init dm_multipath_init(void)
1697 /* allocate a slab for the dm_ios */
1698 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1702 r = dm_register_target(&multipath_target);
1704 DMERR("register failed %d", r);
1705 kmem_cache_destroy(_mpio_cache);
1709 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1711 DMERR("failed to create workqueue kmpathd");
1712 dm_unregister_target(&multipath_target);
1713 kmem_cache_destroy(_mpio_cache);
1718 * A separate workqueue is used to handle the device handlers
1719 * to avoid overloading existing workqueue. Overloading the
1720 * old workqueue would also create a bottleneck in the
1721 * path of the storage hardware device activation.
1723 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1725 if (!kmpath_handlerd) {
1726 DMERR("failed to create workqueue kmpath_handlerd");
1727 destroy_workqueue(kmultipathd);
1728 dm_unregister_target(&multipath_target);
1729 kmem_cache_destroy(_mpio_cache);
1733 DMINFO("version %u.%u.%u loaded",
1734 multipath_target.version[0], multipath_target.version[1],
1735 multipath_target.version[2]);
1740 static void __exit dm_multipath_exit(void)
1742 destroy_workqueue(kmpath_handlerd);
1743 destroy_workqueue(kmultipathd);
1745 dm_unregister_target(&multipath_target);
1746 kmem_cache_destroy(_mpio_cache);
1749 module_init(dm_multipath_init);
1750 module_exit(dm_multipath_exit);
1752 MODULE_DESCRIPTION(DM_NAME " multipath target");
1753 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1754 MODULE_LICENSE("GPL");