2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
5 * This file is released under the GPL.
8 #include <linux/device-mapper.h>
10 #include "dm-path-selector.h"
11 #include "dm-uevent.h"
13 #include <linux/ctype.h>
14 #include <linux/init.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/pagemap.h>
18 #include <linux/slab.h>
19 #include <linux/time.h>
20 #include <linux/workqueue.h>
21 #include <scsi/scsi_dh.h>
22 #include <linux/atomic.h>
24 #define DM_MSG_PREFIX "multipath"
25 #define DM_PG_INIT_DELAY_MSECS 2000
26 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
30 struct list_head list;
32 struct priority_group *pg; /* Owning PG */
33 unsigned is_active; /* Path status */
34 unsigned fail_count; /* Cumulative failure count */
37 struct delayed_work activate_path;
40 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
43 * Paths are grouped into Priority Groups and numbered from 1 upwards.
44 * Each has a path selector which controls which path gets used.
46 struct priority_group {
47 struct list_head list;
49 struct multipath *m; /* Owning multipath instance */
50 struct path_selector ps;
52 unsigned pg_num; /* Reference number */
53 unsigned bypassed; /* Temporarily bypass this PG? */
55 unsigned nr_pgpaths; /* Number of paths in PG */
56 struct list_head pgpaths;
59 #define FEATURE_NO_PARTITIONS 1
61 /* Multipath context */
63 struct list_head list;
68 const char *hw_handler_name;
69 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; /* Must we queue all I/O? */
87 unsigned queue_if_no_path; /* Queue I/O if last path fails? */
88 unsigned saved_queue_if_no_path;/* Saved state during suspension */
89 unsigned pg_init_retries; /* Number of times to retry pg_init */
90 unsigned pg_init_count; /* Number of times pg_init called */
91 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
92 unsigned features; /* Additional selected features */
94 struct work_struct process_queued_ios;
95 struct list_head queued_ios;
98 struct work_struct trigger_event;
101 * We must use a mempool of dm_mpath_io structs so that we
102 * can resubmit bios on error.
104 mempool_t *mpio_pool;
106 struct mutex work_mutex;
110 * Context information attached to each bio we process.
113 struct pgpath *pgpath;
117 typedef int (*action_fn) (struct pgpath *pgpath);
119 #define MIN_IOS 256 /* Mempool size */
121 static struct kmem_cache *_mpio_cache;
123 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
124 static void process_queued_ios(struct work_struct *work);
125 static void trigger_event(struct work_struct *work);
126 static void activate_path(struct work_struct *work);
129 /*-----------------------------------------------
130 * Allocation routines
131 *-----------------------------------------------*/
133 static struct pgpath *alloc_pgpath(void)
135 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
138 pgpath->is_active = 1;
139 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
145 static void free_pgpath(struct pgpath *pgpath)
150 static struct priority_group *alloc_priority_group(void)
152 struct priority_group *pg;
154 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
157 INIT_LIST_HEAD(&pg->pgpaths);
162 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
164 struct pgpath *pgpath, *tmp;
166 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
167 list_del(&pgpath->list);
168 dm_put_device(ti, pgpath->path.dev);
173 static void free_priority_group(struct priority_group *pg,
174 struct dm_target *ti)
176 struct path_selector *ps = &pg->ps;
179 ps->type->destroy(ps);
180 dm_put_path_selector(ps->type);
183 free_pgpaths(&pg->pgpaths, ti);
187 static struct multipath *alloc_multipath(struct dm_target *ti)
191 m = kzalloc(sizeof(*m), GFP_KERNEL);
193 INIT_LIST_HEAD(&m->priority_groups);
194 INIT_LIST_HEAD(&m->queued_ios);
195 spin_lock_init(&m->lock);
197 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
198 INIT_WORK(&m->process_queued_ios, process_queued_ios);
199 INIT_WORK(&m->trigger_event, trigger_event);
200 init_waitqueue_head(&m->pg_init_wait);
201 mutex_init(&m->work_mutex);
202 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
214 static void free_multipath(struct multipath *m)
216 struct priority_group *pg, *tmp;
218 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
220 free_priority_group(pg, m->ti);
223 kfree(m->hw_handler_name);
224 kfree(m->hw_handler_params);
225 mempool_destroy(m->mpio_pool);
229 static int set_mapinfo(struct multipath *m, union map_info *info)
231 struct dm_mpath_io *mpio;
233 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
237 memset(mpio, 0, sizeof(*mpio));
243 static void clear_mapinfo(struct multipath *m, union map_info *info)
245 struct dm_mpath_io *mpio = info->ptr;
248 mempool_free(mpio, m->mpio_pool);
251 /*-----------------------------------------------
253 *-----------------------------------------------*/
255 static void __pg_init_all_paths(struct multipath *m)
257 struct pgpath *pgpath;
258 unsigned long pg_init_delay = 0;
261 m->pg_init_required = 0;
262 if (m->pg_init_delay_retry)
263 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
264 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
265 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
266 /* Skip failed paths */
267 if (!pgpath->is_active)
269 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
271 m->pg_init_in_progress++;
275 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
277 m->current_pg = pgpath->pg;
279 /* Must we initialise the PG first, and queue I/O till it's ready? */
280 if (m->hw_handler_name) {
281 m->pg_init_required = 1;
284 m->pg_init_required = 0;
288 m->pg_init_count = 0;
291 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
294 struct dm_path *path;
296 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
300 m->current_pgpath = path_to_pgpath(path);
302 if (!m->current_pgpath->path.dev) {
303 m->current_pgpath = NULL;
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)
321 /* Were we instructed to switch PG? */
325 if (!__choose_path_in_pg(m, pg, nr_bytes))
329 /* Don't change PG until it has no remaining paths */
330 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
334 * Loop through priority groups until we find a valid path.
335 * First time we skip PGs marked 'bypassed'.
336 * Second time we only try the ones we skipped.
339 list_for_each_entry(pg, &m->priority_groups, list) {
340 if (pg->bypassed == bypassed)
342 if (!__choose_path_in_pg(m, pg, nr_bytes))
345 } while (bypassed--);
348 m->current_pgpath = NULL;
349 m->current_pg = NULL;
353 * Check whether bios must be queued in the device-mapper core rather
354 * than here in the target.
356 * m->lock must be held on entry.
358 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
359 * same value then we are not between multipath_presuspend()
360 * and multipath_resume() calls and we have no need to check
361 * for the DMF_NOFLUSH_SUSPENDING flag.
363 static int __must_push_back(struct multipath *m)
365 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
366 dm_noflush_suspending(m->ti));
369 static int map_io(struct multipath *m, struct request *clone,
370 union map_info *map_context, unsigned was_queued)
372 int r = DM_MAPIO_REMAPPED;
373 size_t nr_bytes = blk_rq_bytes(clone);
375 struct pgpath *pgpath;
376 struct block_device *bdev;
377 struct dm_mpath_io *mpio = map_context->ptr;
379 spin_lock_irqsave(&m->lock, flags);
381 /* Do we need to select a new pgpath? */
382 if (!m->current_pgpath ||
383 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
384 __choose_pgpath(m, nr_bytes);
386 pgpath = m->current_pgpath;
391 if ((pgpath && m->queue_io) ||
392 (!pgpath && m->queue_if_no_path)) {
393 /* Queue for the daemon to resubmit */
394 list_add_tail(&clone->queuelist, &m->queued_ios);
396 if ((m->pg_init_required && !m->pg_init_in_progress) ||
398 queue_work(kmultipathd, &m->process_queued_ios);
400 r = DM_MAPIO_SUBMITTED;
402 bdev = pgpath->path.dev->bdev;
403 clone->q = bdev_get_queue(bdev);
404 clone->rq_disk = bdev->bd_disk;
405 } else if (__must_push_back(m))
406 r = DM_MAPIO_REQUEUE;
408 r = -EIO; /* Failed */
410 mpio->pgpath = pgpath;
411 mpio->nr_bytes = nr_bytes;
413 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
414 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
417 spin_unlock_irqrestore(&m->lock, flags);
423 * If we run out of usable paths, should we queue I/O or error it?
425 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
426 unsigned save_old_value)
430 spin_lock_irqsave(&m->lock, flags);
433 m->saved_queue_if_no_path = m->queue_if_no_path;
435 m->saved_queue_if_no_path = queue_if_no_path;
436 m->queue_if_no_path = queue_if_no_path;
437 if (!m->queue_if_no_path && m->queue_size)
438 queue_work(kmultipathd, &m->process_queued_ios);
440 spin_unlock_irqrestore(&m->lock, flags);
445 /*-----------------------------------------------------------------
446 * The multipath daemon is responsible for resubmitting queued ios.
447 *---------------------------------------------------------------*/
449 static void dispatch_queued_ios(struct multipath *m)
453 union map_info *info;
454 struct request *clone, *n;
457 spin_lock_irqsave(&m->lock, flags);
458 list_splice_init(&m->queued_ios, &cl);
459 spin_unlock_irqrestore(&m->lock, flags);
461 list_for_each_entry_safe(clone, n, &cl, queuelist) {
462 list_del_init(&clone->queuelist);
464 info = dm_get_rq_mapinfo(clone);
466 r = map_io(m, clone, info, 1);
468 clear_mapinfo(m, info);
469 dm_kill_unmapped_request(clone, r);
470 } else if (r == DM_MAPIO_REMAPPED)
471 dm_dispatch_request(clone);
472 else if (r == DM_MAPIO_REQUEUE) {
473 clear_mapinfo(m, info);
474 dm_requeue_unmapped_request(clone);
479 static void process_queued_ios(struct work_struct *work)
481 struct multipath *m =
482 container_of(work, struct multipath, process_queued_ios);
483 struct pgpath *pgpath = NULL;
484 unsigned must_queue = 1;
487 spin_lock_irqsave(&m->lock, flags);
492 if (!m->current_pgpath)
493 __choose_pgpath(m, 0);
495 pgpath = m->current_pgpath;
497 if ((pgpath && !m->queue_io) ||
498 (!pgpath && !m->queue_if_no_path))
501 if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
502 __pg_init_all_paths(m);
505 spin_unlock_irqrestore(&m->lock, flags);
507 dispatch_queued_ios(m);
511 * An event is triggered whenever a path is taken out of use.
512 * Includes path failure and PG bypass.
514 static void trigger_event(struct work_struct *work)
516 struct multipath *m =
517 container_of(work, struct multipath, trigger_event);
519 dm_table_event(m->ti->table);
522 /*-----------------------------------------------------------------
523 * Constructor/argument parsing:
524 * <#multipath feature args> [<arg>]*
525 * <#hw_handler args> [hw_handler [<arg>]*]
527 * <initial priority group>
528 * [<selector> <#selector args> [<arg>]*
529 * <#paths> <#per-path selector args>
530 * [<path> [<arg>]* ]+ ]+
531 *---------------------------------------------------------------*/
532 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
533 struct dm_target *ti)
536 struct path_selector_type *pst;
539 static struct dm_arg _args[] = {
540 {0, 1024, "invalid number of path selector args"},
543 pst = dm_get_path_selector(dm_shift_arg(as));
545 ti->error = "unknown path selector type";
549 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
551 dm_put_path_selector(pst);
555 r = pst->create(&pg->ps, ps_argc, as->argv);
557 dm_put_path_selector(pst);
558 ti->error = "path selector constructor failed";
563 dm_consume_args(as, ps_argc);
568 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
569 struct dm_target *ti)
574 struct multipath *m = ti->private;
576 /* we need at least a path arg */
578 ti->error = "no device given";
579 return ERR_PTR(-EINVAL);
584 return ERR_PTR(-ENOMEM);
586 path = dm_shift_arg(as);
587 r = dm_get_device(ti, path, dm_table_get_mode(ti->table),
590 unsigned major, minor;
592 /* Try to add a failed device */
593 if (r == -ENXIO && sscanf(path, "%u:%u", &major, &minor) == 2) {
596 /* Extract the major/minor numbers */
597 dev = MKDEV(major, minor);
598 if (MAJOR(dev) != major || MINOR(dev) != minor) {
599 /* Nice try, didn't work */
600 DMWARN("Invalid device path %s", path);
601 ti->error = "error converting devnum";
604 DMWARN("adding disabled device %d:%d", major, minor);
606 format_dev_t(p->path.pdev, dev);
609 ti->error = "error getting device";
613 memcpy(p->path.pdev, p->path.dev->name, 16);
617 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
619 if (m->hw_handler_name) {
620 r = scsi_dh_attach(q, m->hw_handler_name);
623 * Already attached to different hw_handler,
624 * try to reattach with correct one.
627 r = scsi_dh_attach(q, m->hw_handler_name);
630 ti->error = "error attaching hardware handler";
631 dm_put_device(ti, p->path.dev);
635 /* Play safe and detach hardware handler */
639 if (m->hw_handler_params) {
640 r = scsi_dh_set_params(q, m->hw_handler_params);
642 ti->error = "unable to set hardware "
643 "handler parameters";
645 dm_put_device(ti, p->path.dev);
651 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
653 dm_put_device(ti, p->path.dev);
658 ps->type->fail_path(ps, &p->path);
669 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
672 static struct dm_arg _args[] = {
673 {1, 1024, "invalid number of paths"},
674 {0, 1024, "invalid number of selector args"}
678 unsigned i, nr_selector_args, nr_args;
679 struct priority_group *pg;
680 struct dm_target *ti = m->ti;
684 ti->error = "not enough priority group arguments";
685 return ERR_PTR(-EINVAL);
688 pg = alloc_priority_group();
690 ti->error = "couldn't allocate priority group";
691 return ERR_PTR(-ENOMEM);
695 r = parse_path_selector(as, pg, ti);
702 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
706 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
710 nr_args = 1 + nr_selector_args;
711 for (i = 0; i < pg->nr_pgpaths; i++) {
712 struct pgpath *pgpath;
713 struct dm_arg_set path_args;
715 if (as->argc < nr_args) {
716 ti->error = "not enough path parameters";
721 path_args.argc = nr_args;
722 path_args.argv = as->argv;
724 pgpath = parse_path(&path_args, &pg->ps, ti);
725 if (IS_ERR(pgpath)) {
731 list_add_tail(&pgpath->list, &pg->pgpaths);
732 dm_consume_args(as, nr_args);
738 free_priority_group(pg, ti);
742 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
746 struct dm_target *ti = m->ti;
748 static struct dm_arg _args[] = {
749 {0, 1024, "invalid number of hardware handler args"},
752 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
758 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
759 if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
760 "scsi_dh_%s", m->hw_handler_name)) {
761 ti->error = "unknown hardware handler type";
770 for (i = 0; i <= hw_argc - 2; i++)
771 len += strlen(as->argv[i]) + 1;
772 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
774 ti->error = "memory allocation failed";
778 j = sprintf(p, "%d", hw_argc - 1);
779 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
780 j = sprintf(p, "%s", as->argv[i]);
782 dm_consume_args(as, hw_argc - 1);
786 kfree(m->hw_handler_name);
787 m->hw_handler_name = NULL;
791 static int parse_features(struct dm_arg_set *as, struct multipath *m)
795 struct dm_target *ti = m->ti;
796 const char *arg_name;
798 static struct dm_arg _args[] = {
799 {0, 5, "invalid number of feature args"},
800 {1, 50, "pg_init_retries must be between 1 and 50"},
801 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
804 r = dm_read_arg_group(_args, as, &argc, &ti->error);
812 arg_name = dm_shift_arg(as);
815 if (!strcasecmp(arg_name, "queue_if_no_path")) {
816 r = queue_if_no_path(m, 1, 0);
820 if (!strcasecmp(arg_name, "no_partitions")) {
821 m->features |= FEATURE_NO_PARTITIONS;
824 if (!strcasecmp(arg_name, "pg_init_retries") &&
826 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
831 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
833 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
838 ti->error = "Unrecognised multipath feature request";
840 } while (argc && !r);
845 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
848 /* target arguments */
849 static struct dm_arg _args[] = {
850 {0, 1024, "invalid number of priority groups"},
851 {0, 1024, "invalid initial priority group number"},
856 struct dm_arg_set as;
857 unsigned pg_count = 0;
858 unsigned next_pg_num;
863 m = alloc_multipath(ti);
865 ti->error = "can't allocate multipath";
869 r = parse_features(&as, m);
873 r = parse_hw_handler(&as, m);
877 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
881 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
885 if ((!m->nr_priority_groups && next_pg_num) ||
886 (m->nr_priority_groups && !next_pg_num)) {
887 ti->error = "invalid initial priority group";
892 /* parse the priority groups */
894 struct priority_group *pg;
896 pg = parse_priority_group(&as, m);
902 m->nr_valid_paths += pg->nr_pgpaths;
903 list_add_tail(&pg->list, &m->priority_groups);
905 pg->pg_num = pg_count;
910 if (pg_count != m->nr_priority_groups) {
911 ti->error = "priority group count mismatch";
916 ti->num_flush_requests = 1;
917 ti->num_discard_requests = 1;
926 static void multipath_wait_for_pg_init_completion(struct multipath *m)
928 DECLARE_WAITQUEUE(wait, current);
931 add_wait_queue(&m->pg_init_wait, &wait);
934 set_current_state(TASK_UNINTERRUPTIBLE);
936 spin_lock_irqsave(&m->lock, flags);
937 if (!m->pg_init_in_progress) {
938 spin_unlock_irqrestore(&m->lock, flags);
941 spin_unlock_irqrestore(&m->lock, flags);
945 set_current_state(TASK_RUNNING);
947 remove_wait_queue(&m->pg_init_wait, &wait);
950 static void flush_multipath_work(struct multipath *m)
952 flush_workqueue(kmpath_handlerd);
953 multipath_wait_for_pg_init_completion(m);
954 flush_workqueue(kmultipathd);
955 flush_work_sync(&m->trigger_event);
958 static void multipath_dtr(struct dm_target *ti)
960 struct multipath *m = ti->private;
962 flush_multipath_work(m);
967 * Map cloned requests
969 static int multipath_map(struct dm_target *ti, struct request *clone,
970 union map_info *map_context)
973 struct multipath *m = (struct multipath *) ti->private;
975 if (set_mapinfo(m, map_context) < 0)
976 /* ENOMEM, requeue */
977 return DM_MAPIO_REQUEUE;
979 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
980 r = map_io(m, clone, map_context, 0);
981 if (r < 0 || r == DM_MAPIO_REQUEUE)
982 clear_mapinfo(m, map_context);
988 * Take a path out of use.
990 static int fail_path(struct pgpath *pgpath)
993 struct multipath *m = pgpath->pg->m;
995 spin_lock_irqsave(&m->lock, flags);
997 if (!pgpath->is_active)
1000 DMWARN("Failing path %s.", pgpath->path.pdev);
1002 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1003 pgpath->is_active = 0;
1004 pgpath->fail_count++;
1006 m->nr_valid_paths--;
1008 if (pgpath == m->current_pgpath)
1009 m->current_pgpath = NULL;
1011 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1012 pgpath->path.pdev, m->nr_valid_paths);
1014 schedule_work(&m->trigger_event);
1017 spin_unlock_irqrestore(&m->lock, flags);
1023 * Reinstate a previously-failed path
1025 static int reinstate_path(struct pgpath *pgpath)
1028 unsigned long flags;
1029 struct multipath *m = pgpath->pg->m;
1031 spin_lock_irqsave(&m->lock, flags);
1033 if (pgpath->is_active)
1036 if (!pgpath->path.dev) {
1037 DMWARN("Cannot reinstate disabled path %s", pgpath->path.pdev);
1042 if (!pgpath->pg->ps.type->reinstate_path) {
1043 DMWARN("Reinstate path not supported by path selector %s",
1044 pgpath->pg->ps.type->name);
1049 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1053 pgpath->is_active = 1;
1055 if (!m->nr_valid_paths++ && m->queue_size) {
1056 m->current_pgpath = NULL;
1057 queue_work(kmultipathd, &m->process_queued_ios);
1058 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1059 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1060 m->pg_init_in_progress++;
1063 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1064 pgpath->path.pdev, m->nr_valid_paths);
1066 schedule_work(&m->trigger_event);
1069 spin_unlock_irqrestore(&m->lock, flags);
1075 * Fail or reinstate all paths that match the provided struct dm_dev.
1077 static int action_dev(struct multipath *m, struct dm_dev *dev,
1081 struct pgpath *pgpath;
1082 struct priority_group *pg;
1087 list_for_each_entry(pg, &m->priority_groups, list) {
1088 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1089 if (pgpath->path.dev == dev)
1098 * Temporarily try to avoid having to use the specified PG
1100 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1103 unsigned long flags;
1105 spin_lock_irqsave(&m->lock, flags);
1107 pg->bypassed = bypassed;
1108 m->current_pgpath = NULL;
1109 m->current_pg = NULL;
1111 spin_unlock_irqrestore(&m->lock, flags);
1113 schedule_work(&m->trigger_event);
1117 * Switch to using the specified PG from the next I/O that gets mapped
1119 static int switch_pg_num(struct multipath *m, const char *pgstr)
1121 struct priority_group *pg;
1123 unsigned long flags;
1126 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1127 (pgnum > m->nr_priority_groups)) {
1128 DMWARN("invalid PG number supplied to switch_pg_num");
1132 spin_lock_irqsave(&m->lock, flags);
1133 list_for_each_entry(pg, &m->priority_groups, list) {
1138 m->current_pgpath = NULL;
1139 m->current_pg = NULL;
1142 spin_unlock_irqrestore(&m->lock, flags);
1144 schedule_work(&m->trigger_event);
1149 * Set/clear bypassed status of a PG.
1150 * PGs are numbered upwards from 1 in the order they were declared.
1152 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1154 struct priority_group *pg;
1158 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1159 (pgnum > m->nr_priority_groups)) {
1160 DMWARN("invalid PG number supplied to bypass_pg");
1164 list_for_each_entry(pg, &m->priority_groups, list) {
1169 bypass_pg(m, pg, bypassed);
1174 * Should we retry pg_init immediately?
1176 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1178 unsigned long flags;
1179 int limit_reached = 0;
1181 spin_lock_irqsave(&m->lock, flags);
1183 if (m->pg_init_count <= m->pg_init_retries)
1184 m->pg_init_required = 1;
1188 spin_unlock_irqrestore(&m->lock, flags);
1190 return limit_reached;
1193 static void pg_init_done(void *data, int errors)
1195 struct pgpath *pgpath = data;
1196 struct priority_group *pg = pgpath->pg;
1197 struct multipath *m = pg->m;
1198 unsigned long flags;
1199 unsigned delay_retry = 0;
1201 /* device or driver problems */
1206 if (!m->hw_handler_name) {
1210 DMERR("Count not failover device %s: Handler scsi_dh_%s "
1211 "was not loaded.", pgpath->path.dev->name,
1212 m->hw_handler_name);
1214 * Fail path for now, so we do not ping pong
1218 case SCSI_DH_DEV_TEMP_BUSY:
1220 * Probably doing something like FW upgrade on the
1221 * controller so try the other pg.
1223 bypass_pg(m, pg, 1);
1225 case SCSI_DH_DEV_OFFLINED:
1226 DMWARN("Device %s offlined.", pgpath->path.dev->name);
1230 /* Wait before retrying. */
1232 case SCSI_DH_IMM_RETRY:
1233 case SCSI_DH_RES_TEMP_UNAVAIL:
1234 if (pg_init_limit_reached(m, pgpath))
1240 * We probably do not want to fail the path for a device
1241 * error, but this is what the old dm did. In future
1242 * patches we can do more advanced handling.
1247 spin_lock_irqsave(&m->lock, flags);
1249 if (pgpath == m->current_pgpath) {
1250 DMERR("Could not failover device %s, error %d.",
1251 pgpath->path.dev->name, errors);
1252 m->current_pgpath = NULL;
1253 m->current_pg = NULL;
1255 } else if (!m->pg_init_required)
1258 if (--m->pg_init_in_progress)
1259 /* Activations of other paths are still on going */
1262 if (!m->pg_init_required)
1265 m->pg_init_delay_retry = delay_retry;
1266 queue_work(kmultipathd, &m->process_queued_ios);
1269 * Wake up any thread waiting to suspend.
1271 wake_up(&m->pg_init_wait);
1274 spin_unlock_irqrestore(&m->lock, flags);
1277 static void activate_path(struct work_struct *work)
1279 struct pgpath *pgpath =
1280 container_of(work, struct pgpath, activate_path.work);
1282 if (pgpath->path.dev)
1283 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1284 pg_init_done, pgpath);
1290 static int do_end_io(struct multipath *m, struct request *clone,
1291 int error, struct dm_mpath_io *mpio)
1294 * We don't queue any clone request inside the multipath target
1295 * during end I/O handling, since those clone requests don't have
1296 * bio clones. If we queue them inside the multipath target,
1297 * we need to make bio clones, that requires memory allocation.
1298 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1299 * don't have bio clones.)
1300 * Instead of queueing the clone request here, we queue the original
1301 * request into dm core, which will remake a clone request and
1302 * clone bios for it and resubmit it later.
1304 int r = DM_ENDIO_REQUEUE;
1305 unsigned long flags;
1307 if (!error && !clone->errors)
1308 return 0; /* I/O complete */
1310 if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1314 fail_path(mpio->pgpath);
1316 spin_lock_irqsave(&m->lock, flags);
1317 if (!m->nr_valid_paths) {
1318 if (!m->queue_if_no_path) {
1319 if (!__must_push_back(m))
1322 if (error == -EBADE)
1326 spin_unlock_irqrestore(&m->lock, flags);
1331 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1332 int error, union map_info *map_context)
1334 struct multipath *m = ti->private;
1335 struct dm_mpath_io *mpio = map_context->ptr;
1336 struct pgpath *pgpath = mpio->pgpath;
1337 struct path_selector *ps;
1342 r = do_end_io(m, clone, error, mpio);
1344 ps = &pgpath->pg->ps;
1345 if (ps->type->end_io)
1346 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1348 clear_mapinfo(m, map_context);
1354 * Suspend can't complete until all the I/O is processed so if
1355 * the last path fails we must error any remaining I/O.
1356 * Note that if the freeze_bdev fails while suspending, the
1357 * queue_if_no_path state is lost - userspace should reset it.
1359 static void multipath_presuspend(struct dm_target *ti)
1361 struct multipath *m = (struct multipath *) ti->private;
1363 queue_if_no_path(m, 0, 1);
1366 static void multipath_postsuspend(struct dm_target *ti)
1368 struct multipath *m = ti->private;
1370 mutex_lock(&m->work_mutex);
1371 flush_multipath_work(m);
1372 mutex_unlock(&m->work_mutex);
1376 * Restore the queue_if_no_path setting.
1378 static void multipath_resume(struct dm_target *ti)
1380 struct multipath *m = (struct multipath *) ti->private;
1381 unsigned long flags;
1383 spin_lock_irqsave(&m->lock, flags);
1384 m->queue_if_no_path = m->saved_queue_if_no_path;
1385 spin_unlock_irqrestore(&m->lock, flags);
1389 * Info output has the following format:
1390 * num_multipath_feature_args [multipath_feature_args]*
1391 * num_handler_status_args [handler_status_args]*
1392 * num_groups init_group_number
1393 * [A|D|E num_ps_status_args [ps_status_args]*
1394 * num_paths num_selector_args
1395 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1397 * Table output has the following format (identical to the constructor string):
1398 * num_feature_args [features_args]*
1399 * num_handler_args hw_handler [hw_handler_args]*
1400 * num_groups init_group_number
1401 * [priority selector-name num_ps_args [ps_args]*
1402 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1404 static int multipath_status(struct dm_target *ti, status_type_t type,
1405 char *result, unsigned int maxlen)
1408 unsigned long flags;
1409 struct multipath *m = (struct multipath *) ti->private;
1410 struct priority_group *pg;
1415 spin_lock_irqsave(&m->lock, flags);
1418 if (type == STATUSTYPE_INFO)
1419 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1421 DMEMIT("%u ", m->queue_if_no_path +
1422 (m->pg_init_retries > 0) * 2 +
1423 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1424 (m->features & FEATURE_NO_PARTITIONS));
1425 if (m->queue_if_no_path)
1426 DMEMIT("queue_if_no_path ");
1427 if (m->pg_init_retries)
1428 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1429 if (m->features & FEATURE_NO_PARTITIONS)
1430 DMEMIT("no_partitions ");
1431 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1432 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1435 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1438 DMEMIT("1 %s ", m->hw_handler_name);
1440 DMEMIT("%u ", m->nr_priority_groups);
1443 pg_num = m->next_pg->pg_num;
1444 else if (m->current_pg)
1445 pg_num = m->current_pg->pg_num;
1447 pg_num = (m->nr_priority_groups ? 1 : 0);
1449 DMEMIT("%u ", pg_num);
1452 case STATUSTYPE_INFO:
1453 list_for_each_entry(pg, &m->priority_groups, list) {
1455 state = 'D'; /* Disabled */
1456 else if (pg == m->current_pg)
1457 state = 'A'; /* Currently Active */
1459 state = 'E'; /* Enabled */
1461 DMEMIT("%c ", state);
1463 if (pg->ps.type->status)
1464 sz += pg->ps.type->status(&pg->ps, NULL, type,
1470 DMEMIT("%u %u ", pg->nr_pgpaths,
1471 pg->ps.type->info_args);
1473 list_for_each_entry(p, &pg->pgpaths, list) {
1474 DMEMIT("%s %s %u ", p->path.pdev,
1475 p->is_active ? "A" : "F",
1477 if (pg->ps.type->status)
1478 sz += pg->ps.type->status(&pg->ps,
1479 &p->path, type, result + sz,
1485 case STATUSTYPE_TABLE:
1486 list_for_each_entry(pg, &m->priority_groups, list) {
1487 DMEMIT("%s ", pg->ps.type->name);
1489 if (pg->ps.type->status)
1490 sz += pg->ps.type->status(&pg->ps, NULL, type,
1496 DMEMIT("%u %u ", pg->nr_pgpaths,
1497 pg->ps.type->table_args);
1499 list_for_each_entry(p, &pg->pgpaths, list) {
1500 DMEMIT("%s ", p->path.pdev);
1501 if (pg->ps.type->status)
1502 sz += pg->ps.type->status(&pg->ps,
1503 &p->path, type, result + sz,
1510 spin_unlock_irqrestore(&m->lock, flags);
1515 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1519 struct multipath *m = (struct multipath *) ti->private;
1522 mutex_lock(&m->work_mutex);
1524 if (dm_suspended(ti)) {
1530 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1531 r = queue_if_no_path(m, 1, 0);
1533 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1534 r = queue_if_no_path(m, 0, 0);
1540 DMWARN("Unrecognised multipath message received.");
1544 if (!strcasecmp(argv[0], "disable_group")) {
1545 r = bypass_pg_num(m, argv[1], 1);
1547 } else if (!strcasecmp(argv[0], "enable_group")) {
1548 r = bypass_pg_num(m, argv[1], 0);
1550 } else if (!strcasecmp(argv[0], "switch_group")) {
1551 r = switch_pg_num(m, argv[1]);
1553 } else if (!strcasecmp(argv[0], "reinstate_path"))
1554 action = reinstate_path;
1555 else if (!strcasecmp(argv[0], "fail_path"))
1558 DMWARN("Unrecognised multipath message received.");
1562 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1564 DMWARN("message: error getting device %s",
1569 r = action_dev(m, dev, action);
1571 dm_put_device(ti, dev);
1574 mutex_unlock(&m->work_mutex);
1578 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1581 struct multipath *m = (struct multipath *) ti->private;
1582 struct block_device *bdev = NULL;
1584 unsigned long flags;
1587 spin_lock_irqsave(&m->lock, flags);
1589 if (!m->current_pgpath)
1590 __choose_pgpath(m, 0);
1592 if (m->current_pgpath && m->current_pgpath->path.dev) {
1593 bdev = m->current_pgpath->path.dev->bdev;
1594 mode = m->current_pgpath->path.dev->mode;
1602 spin_unlock_irqrestore(&m->lock, flags);
1605 * Only pass ioctls through if the device sizes match exactly.
1607 if (!r && ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT)
1608 r = scsi_verify_blk_ioctl(NULL, cmd);
1610 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1613 static int multipath_iterate_devices(struct dm_target *ti,
1614 iterate_devices_callout_fn fn, void *data)
1616 struct multipath *m = ti->private;
1617 struct priority_group *pg;
1621 list_for_each_entry(pg, &m->priority_groups, list) {
1622 list_for_each_entry(p, &pg->pgpaths, list) {
1623 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1633 static int __pgpath_busy(struct pgpath *pgpath)
1635 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1637 return dm_underlying_device_busy(q);
1641 * We return "busy", only when we can map I/Os but underlying devices
1642 * are busy (so even if we map I/Os now, the I/Os will wait on
1643 * the underlying queue).
1644 * In other words, if we want to kill I/Os or queue them inside us
1645 * due to map unavailability, we don't return "busy". Otherwise,
1646 * dm core won't give us the I/Os and we can't do what we want.
1648 static int multipath_busy(struct dm_target *ti)
1650 int busy = 0, has_active = 0;
1651 struct multipath *m = ti->private;
1652 struct priority_group *pg;
1653 struct pgpath *pgpath;
1654 unsigned long flags;
1656 spin_lock_irqsave(&m->lock, flags);
1658 /* Guess which priority_group will be used at next mapping time */
1659 if (unlikely(!m->current_pgpath && m->next_pg))
1661 else if (likely(m->current_pg))
1665 * We don't know which pg will be used at next mapping time.
1666 * We don't call __choose_pgpath() here to avoid to trigger
1667 * pg_init just by busy checking.
1668 * So we don't know whether underlying devices we will be using
1669 * at next mapping time are busy or not. Just try mapping.
1674 * If there is one non-busy active path at least, the path selector
1675 * will be able to select it. So we consider such a pg as not busy.
1678 list_for_each_entry(pgpath, &pg->pgpaths, list)
1679 if (pgpath->is_active) {
1682 if (!__pgpath_busy(pgpath)) {
1690 * No active path in this pg, so this pg won't be used and
1691 * the current_pg will be changed at next mapping time.
1692 * We need to try mapping to determine it.
1697 spin_unlock_irqrestore(&m->lock, flags);
1702 /*-----------------------------------------------------------------
1704 *---------------------------------------------------------------*/
1705 static struct target_type multipath_target = {
1706 .name = "multipath",
1707 .version = {1, 3, 0},
1708 .module = THIS_MODULE,
1709 .ctr = multipath_ctr,
1710 .dtr = multipath_dtr,
1711 .map_rq = multipath_map,
1712 .rq_end_io = multipath_end_io,
1713 .presuspend = multipath_presuspend,
1714 .postsuspend = multipath_postsuspend,
1715 .resume = multipath_resume,
1716 .status = multipath_status,
1717 .message = multipath_message,
1718 .ioctl = multipath_ioctl,
1719 .iterate_devices = multipath_iterate_devices,
1720 .busy = multipath_busy,
1723 static int __init dm_multipath_init(void)
1727 /* allocate a slab for the dm_ios */
1728 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1732 r = dm_register_target(&multipath_target);
1734 DMERR("register failed %d", r);
1735 kmem_cache_destroy(_mpio_cache);
1739 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1741 DMERR("failed to create workqueue kmpathd");
1742 dm_unregister_target(&multipath_target);
1743 kmem_cache_destroy(_mpio_cache);
1748 * A separate workqueue is used to handle the device handlers
1749 * to avoid overloading existing workqueue. Overloading the
1750 * old workqueue would also create a bottleneck in the
1751 * path of the storage hardware device activation.
1753 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1755 if (!kmpath_handlerd) {
1756 DMERR("failed to create workqueue kmpath_handlerd");
1757 destroy_workqueue(kmultipathd);
1758 dm_unregister_target(&multipath_target);
1759 kmem_cache_destroy(_mpio_cache);
1763 DMINFO("version %u.%u.%u loaded",
1764 multipath_target.version[0], multipath_target.version[1],
1765 multipath_target.version[2]);
1770 static void __exit dm_multipath_exit(void)
1772 destroy_workqueue(kmpath_handlerd);
1773 destroy_workqueue(kmultipathd);
1775 dm_unregister_target(&multipath_target);
1776 kmem_cache_destroy(_mpio_cache);
1779 module_init(dm_multipath_init);
1780 module_exit(dm_multipath_exit);
1782 MODULE_DESCRIPTION(DM_NAME " multipath target");
1783 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1784 MODULE_LICENSE("GPL");