2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
5 * This file is released under the GPL.
8 #include <linux/device-mapper.h>
10 #include "dm-path-selector.h"
11 #include "dm-uevent.h"
13 #include <linux/ctype.h>
14 #include <linux/init.h>
15 #include <linux/mempool.h>
16 #include <linux/module.h>
17 #include <linux/pagemap.h>
18 #include <linux/slab.h>
19 #include <linux/time.h>
20 #include <linux/workqueue.h>
21 #include <scsi/scsi_dh.h>
22 #include <linux/atomic.h>
24 #define DM_MSG_PREFIX "multipath"
25 #define DM_PG_INIT_DELAY_MSECS 2000
26 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
30 struct list_head list;
32 struct priority_group *pg; /* Owning PG */
33 unsigned is_active; /* Path status */
34 unsigned fail_count; /* Cumulative failure count */
37 struct delayed_work activate_path;
40 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
43 * Paths are grouped into Priority Groups and numbered from 1 upwards.
44 * Each has a path selector which controls which path gets used.
46 struct priority_group {
47 struct list_head list;
49 struct multipath *m; /* Owning multipath instance */
50 struct path_selector ps;
52 unsigned pg_num; /* Reference number */
53 unsigned bypassed; /* Temporarily bypass this PG? */
55 unsigned nr_pgpaths; /* Number of paths in PG */
56 struct list_head pgpaths;
59 /* Multipath context */
61 struct list_head list;
64 const char *hw_handler_name;
65 char *hw_handler_params;
69 unsigned nr_priority_groups;
70 struct list_head priority_groups;
72 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
74 unsigned pg_init_required; /* pg_init needs calling? */
75 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
76 unsigned pg_init_delay_retry; /* Delay pg_init retry? */
78 unsigned nr_valid_paths; /* Total number of usable paths */
79 struct pgpath *current_pgpath;
80 struct priority_group *current_pg;
81 struct priority_group *next_pg; /* Switch to this PG if set */
82 unsigned repeat_count; /* I/Os left before calling PS again */
84 unsigned queue_io:1; /* Must we queue all I/O? */
85 unsigned queue_if_no_path:1; /* Queue I/O if last path fails? */
86 unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
88 unsigned pg_init_retries; /* Number of times to retry pg_init */
89 unsigned pg_init_count; /* Number of times pg_init called */
90 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
93 struct work_struct process_queued_ios;
94 struct list_head queued_ios;
96 struct work_struct trigger_event;
99 * We must use a mempool of dm_mpath_io structs so that we
100 * can resubmit bios on error.
102 mempool_t *mpio_pool;
104 struct mutex work_mutex;
108 * Context information attached to each bio we process.
111 struct pgpath *pgpath;
115 typedef int (*action_fn) (struct pgpath *pgpath);
117 #define MIN_IOS 256 /* Mempool size */
119 static struct kmem_cache *_mpio_cache;
121 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
122 static void process_queued_ios(struct work_struct *work);
123 static void trigger_event(struct work_struct *work);
124 static void activate_path(struct work_struct *work);
127 /*-----------------------------------------------
128 * Allocation routines
129 *-----------------------------------------------*/
131 static struct pgpath *alloc_pgpath(void)
133 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
136 pgpath->is_active = 1;
137 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
143 static void free_pgpath(struct pgpath *pgpath)
148 static struct priority_group *alloc_priority_group(void)
150 struct priority_group *pg;
152 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
155 INIT_LIST_HEAD(&pg->pgpaths);
160 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
162 struct pgpath *pgpath, *tmp;
163 struct multipath *m = ti->private;
165 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
166 list_del(&pgpath->list);
167 if (m->hw_handler_name)
168 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
169 dm_put_device(ti, pgpath->path.dev);
174 static void free_priority_group(struct priority_group *pg,
175 struct dm_target *ti)
177 struct path_selector *ps = &pg->ps;
180 ps->type->destroy(ps);
181 dm_put_path_selector(ps->type);
184 free_pgpaths(&pg->pgpaths, ti);
188 static struct multipath *alloc_multipath(struct dm_target *ti)
192 m = kzalloc(sizeof(*m), GFP_KERNEL);
194 INIT_LIST_HEAD(&m->priority_groups);
195 INIT_LIST_HEAD(&m->queued_ios);
196 spin_lock_init(&m->lock);
198 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
199 INIT_WORK(&m->process_queued_ios, process_queued_ios);
200 INIT_WORK(&m->trigger_event, trigger_event);
201 init_waitqueue_head(&m->pg_init_wait);
202 mutex_init(&m->work_mutex);
203 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
215 static void free_multipath(struct multipath *m)
217 struct priority_group *pg, *tmp;
219 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
221 free_priority_group(pg, m->ti);
224 kfree(m->hw_handler_name);
225 kfree(m->hw_handler_params);
226 mempool_destroy(m->mpio_pool);
230 static int set_mapinfo(struct multipath *m, union map_info *info)
232 struct dm_mpath_io *mpio;
234 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
238 memset(mpio, 0, sizeof(*mpio));
244 static void clear_mapinfo(struct multipath *m, union map_info *info)
246 struct dm_mpath_io *mpio = info->ptr;
249 mempool_free(mpio, m->mpio_pool);
252 /*-----------------------------------------------
254 *-----------------------------------------------*/
256 static void __pg_init_all_paths(struct multipath *m)
258 struct pgpath *pgpath;
259 unsigned long pg_init_delay = 0;
262 m->pg_init_required = 0;
263 if (m->pg_init_delay_retry)
264 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
265 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
266 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
267 /* Skip failed paths */
268 if (!pgpath->is_active)
270 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
272 m->pg_init_in_progress++;
276 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
278 m->current_pg = pgpath->pg;
280 /* Must we initialise the PG first, and queue I/O till it's ready? */
281 if (m->hw_handler_name) {
282 m->pg_init_required = 1;
285 m->pg_init_required = 0;
289 m->pg_init_count = 0;
292 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
295 struct dm_path *path;
297 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
301 m->current_pgpath = path_to_pgpath(path);
303 if (m->current_pg != pg)
304 __switch_pg(m, m->current_pgpath);
309 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
311 struct priority_group *pg;
312 unsigned bypassed = 1;
314 if (!m->nr_valid_paths)
317 /* Were we instructed to switch PG? */
321 if (!__choose_path_in_pg(m, pg, nr_bytes))
325 /* Don't change PG until it has no remaining paths */
326 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
330 * Loop through priority groups until we find a valid path.
331 * First time we skip PGs marked 'bypassed'.
332 * Second time we only try the ones we skipped.
335 list_for_each_entry(pg, &m->priority_groups, list) {
336 if (pg->bypassed == bypassed)
338 if (!__choose_path_in_pg(m, pg, nr_bytes))
341 } while (bypassed--);
344 m->current_pgpath = NULL;
345 m->current_pg = NULL;
349 * Check whether bios must be queued in the device-mapper core rather
350 * than here in the target.
352 * m->lock must be held on entry.
354 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
355 * same value then we are not between multipath_presuspend()
356 * and multipath_resume() calls and we have no need to check
357 * for the DMF_NOFLUSH_SUSPENDING flag.
359 static int __must_push_back(struct multipath *m)
361 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
362 dm_noflush_suspending(m->ti));
365 static int map_io(struct multipath *m, struct request *clone,
366 union map_info *map_context, unsigned was_queued)
368 int r = DM_MAPIO_REMAPPED;
369 size_t nr_bytes = blk_rq_bytes(clone);
371 struct pgpath *pgpath;
372 struct block_device *bdev;
373 struct dm_mpath_io *mpio = map_context->ptr;
375 spin_lock_irqsave(&m->lock, flags);
377 /* Do we need to select a new pgpath? */
378 if (!m->current_pgpath ||
379 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
380 __choose_pgpath(m, nr_bytes);
382 pgpath = m->current_pgpath;
387 if ((pgpath && m->queue_io) ||
388 (!pgpath && m->queue_if_no_path)) {
389 /* Queue for the daemon to resubmit */
390 list_add_tail(&clone->queuelist, &m->queued_ios);
392 if ((m->pg_init_required && !m->pg_init_in_progress) ||
394 queue_work(kmultipathd, &m->process_queued_ios);
396 r = DM_MAPIO_SUBMITTED;
398 bdev = pgpath->path.dev->bdev;
399 clone->q = bdev_get_queue(bdev);
400 clone->rq_disk = bdev->bd_disk;
401 } else if (__must_push_back(m))
402 r = DM_MAPIO_REQUEUE;
404 r = -EIO; /* Failed */
406 mpio->pgpath = pgpath;
407 mpio->nr_bytes = nr_bytes;
409 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
410 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
413 spin_unlock_irqrestore(&m->lock, flags);
419 * If we run out of usable paths, should we queue I/O or error it?
421 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
422 unsigned save_old_value)
426 spin_lock_irqsave(&m->lock, flags);
429 m->saved_queue_if_no_path = m->queue_if_no_path;
431 m->saved_queue_if_no_path = queue_if_no_path;
432 m->queue_if_no_path = queue_if_no_path;
433 if (!m->queue_if_no_path && m->queue_size)
434 queue_work(kmultipathd, &m->process_queued_ios);
436 spin_unlock_irqrestore(&m->lock, flags);
441 /*-----------------------------------------------------------------
442 * The multipath daemon is responsible for resubmitting queued ios.
443 *---------------------------------------------------------------*/
445 static void dispatch_queued_ios(struct multipath *m)
449 union map_info *info;
450 struct request *clone, *n;
453 spin_lock_irqsave(&m->lock, flags);
454 list_splice_init(&m->queued_ios, &cl);
455 spin_unlock_irqrestore(&m->lock, flags);
457 list_for_each_entry_safe(clone, n, &cl, queuelist) {
458 list_del_init(&clone->queuelist);
460 info = dm_get_rq_mapinfo(clone);
462 r = map_io(m, clone, info, 1);
464 clear_mapinfo(m, info);
465 dm_kill_unmapped_request(clone, r);
466 } else if (r == DM_MAPIO_REMAPPED)
467 dm_dispatch_request(clone);
468 else if (r == DM_MAPIO_REQUEUE) {
469 clear_mapinfo(m, info);
470 dm_requeue_unmapped_request(clone);
475 static void process_queued_ios(struct work_struct *work)
477 struct multipath *m =
478 container_of(work, struct multipath, process_queued_ios);
479 struct pgpath *pgpath = NULL;
480 unsigned must_queue = 1;
483 spin_lock_irqsave(&m->lock, flags);
488 if (!m->current_pgpath)
489 __choose_pgpath(m, 0);
491 pgpath = m->current_pgpath;
493 if ((pgpath && !m->queue_io) ||
494 (!pgpath && !m->queue_if_no_path))
497 if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
498 __pg_init_all_paths(m);
501 spin_unlock_irqrestore(&m->lock, flags);
503 dispatch_queued_ios(m);
507 * An event is triggered whenever a path is taken out of use.
508 * Includes path failure and PG bypass.
510 static void trigger_event(struct work_struct *work)
512 struct multipath *m =
513 container_of(work, struct multipath, trigger_event);
515 dm_table_event(m->ti->table);
518 /*-----------------------------------------------------------------
519 * Constructor/argument parsing:
520 * <#multipath feature args> [<arg>]*
521 * <#hw_handler args> [hw_handler [<arg>]*]
523 * <initial priority group>
524 * [<selector> <#selector args> [<arg>]*
525 * <#paths> <#per-path selector args>
526 * [<path> [<arg>]* ]+ ]+
527 *---------------------------------------------------------------*/
528 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
529 struct dm_target *ti)
532 struct path_selector_type *pst;
535 static struct dm_arg _args[] = {
536 {0, 1024, "invalid number of path selector args"},
539 pst = dm_get_path_selector(dm_shift_arg(as));
541 ti->error = "unknown path selector type";
545 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
547 dm_put_path_selector(pst);
551 r = pst->create(&pg->ps, ps_argc, as->argv);
553 dm_put_path_selector(pst);
554 ti->error = "path selector constructor failed";
559 dm_consume_args(as, ps_argc);
564 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
565 struct dm_target *ti)
569 struct multipath *m = ti->private;
571 /* we need at least a path arg */
573 ti->error = "no device given";
574 return ERR_PTR(-EINVAL);
579 return ERR_PTR(-ENOMEM);
581 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
584 ti->error = "error getting device";
588 if (m->hw_handler_name) {
589 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
591 r = scsi_dh_attach(q, m->hw_handler_name);
594 * Already attached to different hw_handler,
595 * try to reattach with correct one.
598 r = scsi_dh_attach(q, m->hw_handler_name);
602 ti->error = "error attaching hardware handler";
603 dm_put_device(ti, p->path.dev);
607 if (m->hw_handler_params) {
608 r = scsi_dh_set_params(q, m->hw_handler_params);
610 ti->error = "unable to set hardware "
611 "handler parameters";
613 dm_put_device(ti, p->path.dev);
619 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
621 dm_put_device(ti, p->path.dev);
632 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
635 static struct dm_arg _args[] = {
636 {1, 1024, "invalid number of paths"},
637 {0, 1024, "invalid number of selector args"}
641 unsigned i, nr_selector_args, nr_args;
642 struct priority_group *pg;
643 struct dm_target *ti = m->ti;
647 ti->error = "not enough priority group arguments";
648 return ERR_PTR(-EINVAL);
651 pg = alloc_priority_group();
653 ti->error = "couldn't allocate priority group";
654 return ERR_PTR(-ENOMEM);
658 r = parse_path_selector(as, pg, ti);
665 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
669 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
673 nr_args = 1 + nr_selector_args;
674 for (i = 0; i < pg->nr_pgpaths; i++) {
675 struct pgpath *pgpath;
676 struct dm_arg_set path_args;
678 if (as->argc < nr_args) {
679 ti->error = "not enough path parameters";
684 path_args.argc = nr_args;
685 path_args.argv = as->argv;
687 pgpath = parse_path(&path_args, &pg->ps, ti);
688 if (IS_ERR(pgpath)) {
694 list_add_tail(&pgpath->list, &pg->pgpaths);
695 dm_consume_args(as, nr_args);
701 free_priority_group(pg, ti);
705 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
709 struct dm_target *ti = m->ti;
711 static struct dm_arg _args[] = {
712 {0, 1024, "invalid number of hardware handler args"},
715 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
721 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
722 if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
723 "scsi_dh_%s", m->hw_handler_name)) {
724 ti->error = "unknown hardware handler type";
733 for (i = 0; i <= hw_argc - 2; i++)
734 len += strlen(as->argv[i]) + 1;
735 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
737 ti->error = "memory allocation failed";
741 j = sprintf(p, "%d", hw_argc - 1);
742 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
743 j = sprintf(p, "%s", as->argv[i]);
745 dm_consume_args(as, hw_argc - 1);
749 kfree(m->hw_handler_name);
750 m->hw_handler_name = NULL;
754 static int parse_features(struct dm_arg_set *as, struct multipath *m)
758 struct dm_target *ti = m->ti;
759 const char *arg_name;
761 static struct dm_arg _args[] = {
762 {0, 5, "invalid number of feature args"},
763 {1, 50, "pg_init_retries must be between 1 and 50"},
764 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
767 r = dm_read_arg_group(_args, as, &argc, &ti->error);
775 arg_name = dm_shift_arg(as);
778 if (!strcasecmp(arg_name, "queue_if_no_path")) {
779 r = queue_if_no_path(m, 1, 0);
783 if (!strcasecmp(arg_name, "pg_init_retries") &&
785 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
790 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
792 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
797 ti->error = "Unrecognised multipath feature request";
799 } while (argc && !r);
804 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
807 /* target arguments */
808 static struct dm_arg _args[] = {
809 {0, 1024, "invalid number of priority groups"},
810 {0, 1024, "invalid initial priority group number"},
815 struct dm_arg_set as;
816 unsigned pg_count = 0;
817 unsigned next_pg_num;
822 m = alloc_multipath(ti);
824 ti->error = "can't allocate multipath";
828 r = parse_features(&as, m);
832 r = parse_hw_handler(&as, m);
836 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
840 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
844 if ((!m->nr_priority_groups && next_pg_num) ||
845 (m->nr_priority_groups && !next_pg_num)) {
846 ti->error = "invalid initial priority group";
851 /* parse the priority groups */
853 struct priority_group *pg;
855 pg = parse_priority_group(&as, m);
861 m->nr_valid_paths += pg->nr_pgpaths;
862 list_add_tail(&pg->list, &m->priority_groups);
864 pg->pg_num = pg_count;
869 if (pg_count != m->nr_priority_groups) {
870 ti->error = "priority group count mismatch";
875 ti->num_flush_requests = 1;
876 ti->num_discard_requests = 1;
885 static void multipath_wait_for_pg_init_completion(struct multipath *m)
887 DECLARE_WAITQUEUE(wait, current);
890 add_wait_queue(&m->pg_init_wait, &wait);
893 set_current_state(TASK_UNINTERRUPTIBLE);
895 spin_lock_irqsave(&m->lock, flags);
896 if (!m->pg_init_in_progress) {
897 spin_unlock_irqrestore(&m->lock, flags);
900 spin_unlock_irqrestore(&m->lock, flags);
904 set_current_state(TASK_RUNNING);
906 remove_wait_queue(&m->pg_init_wait, &wait);
909 static void flush_multipath_work(struct multipath *m)
911 flush_workqueue(kmpath_handlerd);
912 multipath_wait_for_pg_init_completion(m);
913 flush_workqueue(kmultipathd);
914 flush_work_sync(&m->trigger_event);
917 static void multipath_dtr(struct dm_target *ti)
919 struct multipath *m = ti->private;
921 flush_multipath_work(m);
926 * Map cloned requests
928 static int multipath_map(struct dm_target *ti, struct request *clone,
929 union map_info *map_context)
932 struct multipath *m = (struct multipath *) ti->private;
934 if (set_mapinfo(m, map_context) < 0)
935 /* ENOMEM, requeue */
936 return DM_MAPIO_REQUEUE;
938 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
939 r = map_io(m, clone, map_context, 0);
940 if (r < 0 || r == DM_MAPIO_REQUEUE)
941 clear_mapinfo(m, map_context);
947 * Take a path out of use.
949 static int fail_path(struct pgpath *pgpath)
952 struct multipath *m = pgpath->pg->m;
954 spin_lock_irqsave(&m->lock, flags);
956 if (!pgpath->is_active)
959 DMWARN("Failing path %s.", pgpath->path.dev->name);
961 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
962 pgpath->is_active = 0;
963 pgpath->fail_count++;
967 if (pgpath == m->current_pgpath)
968 m->current_pgpath = NULL;
970 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
971 pgpath->path.dev->name, m->nr_valid_paths);
973 schedule_work(&m->trigger_event);
976 spin_unlock_irqrestore(&m->lock, flags);
982 * Reinstate a previously-failed path
984 static int reinstate_path(struct pgpath *pgpath)
988 struct multipath *m = pgpath->pg->m;
990 spin_lock_irqsave(&m->lock, flags);
992 if (pgpath->is_active)
995 if (!pgpath->pg->ps.type->reinstate_path) {
996 DMWARN("Reinstate path not supported by path selector %s",
997 pgpath->pg->ps.type->name);
1002 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1006 pgpath->is_active = 1;
1008 if (!m->nr_valid_paths++ && m->queue_size) {
1009 m->current_pgpath = NULL;
1010 queue_work(kmultipathd, &m->process_queued_ios);
1011 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1012 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1013 m->pg_init_in_progress++;
1016 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1017 pgpath->path.dev->name, m->nr_valid_paths);
1019 schedule_work(&m->trigger_event);
1022 spin_unlock_irqrestore(&m->lock, flags);
1028 * Fail or reinstate all paths that match the provided struct dm_dev.
1030 static int action_dev(struct multipath *m, struct dm_dev *dev,
1034 struct pgpath *pgpath;
1035 struct priority_group *pg;
1037 list_for_each_entry(pg, &m->priority_groups, list) {
1038 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1039 if (pgpath->path.dev == dev)
1048 * Temporarily try to avoid having to use the specified PG
1050 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1053 unsigned long flags;
1055 spin_lock_irqsave(&m->lock, flags);
1057 pg->bypassed = bypassed;
1058 m->current_pgpath = NULL;
1059 m->current_pg = NULL;
1061 spin_unlock_irqrestore(&m->lock, flags);
1063 schedule_work(&m->trigger_event);
1067 * Switch to using the specified PG from the next I/O that gets mapped
1069 static int switch_pg_num(struct multipath *m, const char *pgstr)
1071 struct priority_group *pg;
1073 unsigned long flags;
1076 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1077 (pgnum > m->nr_priority_groups)) {
1078 DMWARN("invalid PG number supplied to switch_pg_num");
1082 spin_lock_irqsave(&m->lock, flags);
1083 list_for_each_entry(pg, &m->priority_groups, list) {
1088 m->current_pgpath = NULL;
1089 m->current_pg = NULL;
1092 spin_unlock_irqrestore(&m->lock, flags);
1094 schedule_work(&m->trigger_event);
1099 * Set/clear bypassed status of a PG.
1100 * PGs are numbered upwards from 1 in the order they were declared.
1102 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1104 struct priority_group *pg;
1108 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1109 (pgnum > m->nr_priority_groups)) {
1110 DMWARN("invalid PG number supplied to bypass_pg");
1114 list_for_each_entry(pg, &m->priority_groups, list) {
1119 bypass_pg(m, pg, bypassed);
1124 * Should we retry pg_init immediately?
1126 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1128 unsigned long flags;
1129 int limit_reached = 0;
1131 spin_lock_irqsave(&m->lock, flags);
1133 if (m->pg_init_count <= m->pg_init_retries)
1134 m->pg_init_required = 1;
1138 spin_unlock_irqrestore(&m->lock, flags);
1140 return limit_reached;
1143 static void pg_init_done(void *data, int errors)
1145 struct pgpath *pgpath = data;
1146 struct priority_group *pg = pgpath->pg;
1147 struct multipath *m = pg->m;
1148 unsigned long flags;
1149 unsigned delay_retry = 0;
1151 /* device or driver problems */
1156 if (!m->hw_handler_name) {
1160 DMERR("Could not failover the device: Handler scsi_dh_%s "
1161 "Error %d.", m->hw_handler_name, errors);
1163 * Fail path for now, so we do not ping pong
1167 case SCSI_DH_DEV_TEMP_BUSY:
1169 * Probably doing something like FW upgrade on the
1170 * controller so try the other pg.
1172 bypass_pg(m, pg, 1);
1175 /* Wait before retrying. */
1177 case SCSI_DH_IMM_RETRY:
1178 case SCSI_DH_RES_TEMP_UNAVAIL:
1179 if (pg_init_limit_reached(m, pgpath))
1185 * We probably do not want to fail the path for a device
1186 * error, but this is what the old dm did. In future
1187 * patches we can do more advanced handling.
1192 spin_lock_irqsave(&m->lock, flags);
1194 if (pgpath == m->current_pgpath) {
1195 DMERR("Could not failover device. Error %d.", errors);
1196 m->current_pgpath = NULL;
1197 m->current_pg = NULL;
1199 } else if (!m->pg_init_required)
1202 if (--m->pg_init_in_progress)
1203 /* Activations of other paths are still on going */
1206 if (!m->pg_init_required)
1209 m->pg_init_delay_retry = delay_retry;
1210 queue_work(kmultipathd, &m->process_queued_ios);
1213 * Wake up any thread waiting to suspend.
1215 wake_up(&m->pg_init_wait);
1218 spin_unlock_irqrestore(&m->lock, flags);
1221 static void activate_path(struct work_struct *work)
1223 struct pgpath *pgpath =
1224 container_of(work, struct pgpath, activate_path.work);
1226 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1227 pg_init_done, pgpath);
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 (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1257 fail_path(mpio->pgpath);
1259 spin_lock_irqsave(&m->lock, flags);
1260 if (!m->nr_valid_paths) {
1261 if (!m->queue_if_no_path) {
1262 if (!__must_push_back(m))
1265 if (error == -EBADE)
1269 spin_unlock_irqrestore(&m->lock, flags);
1274 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1275 int error, union map_info *map_context)
1277 struct multipath *m = ti->private;
1278 struct dm_mpath_io *mpio = map_context->ptr;
1279 struct pgpath *pgpath = mpio->pgpath;
1280 struct path_selector *ps;
1285 r = do_end_io(m, clone, error, mpio);
1287 ps = &pgpath->pg->ps;
1288 if (ps->type->end_io)
1289 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1291 clear_mapinfo(m, map_context);
1297 * Suspend can't complete until all the I/O is processed so if
1298 * the last path fails we must error any remaining I/O.
1299 * Note that if the freeze_bdev fails while suspending, the
1300 * queue_if_no_path state is lost - userspace should reset it.
1302 static void multipath_presuspend(struct dm_target *ti)
1304 struct multipath *m = (struct multipath *) ti->private;
1306 queue_if_no_path(m, 0, 1);
1309 static void multipath_postsuspend(struct dm_target *ti)
1311 struct multipath *m = ti->private;
1313 mutex_lock(&m->work_mutex);
1314 flush_multipath_work(m);
1315 mutex_unlock(&m->work_mutex);
1319 * Restore the queue_if_no_path setting.
1321 static void multipath_resume(struct dm_target *ti)
1323 struct multipath *m = (struct multipath *) ti->private;
1324 unsigned long flags;
1326 spin_lock_irqsave(&m->lock, flags);
1327 m->queue_if_no_path = m->saved_queue_if_no_path;
1328 spin_unlock_irqrestore(&m->lock, flags);
1332 * Info output has the following format:
1333 * num_multipath_feature_args [multipath_feature_args]*
1334 * num_handler_status_args [handler_status_args]*
1335 * num_groups init_group_number
1336 * [A|D|E num_ps_status_args [ps_status_args]*
1337 * num_paths num_selector_args
1338 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1340 * Table output has the following format (identical to the constructor string):
1341 * num_feature_args [features_args]*
1342 * num_handler_args hw_handler [hw_handler_args]*
1343 * num_groups init_group_number
1344 * [priority selector-name num_ps_args [ps_args]*
1345 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1347 static int multipath_status(struct dm_target *ti, status_type_t type,
1348 char *result, unsigned int maxlen)
1351 unsigned long flags;
1352 struct multipath *m = (struct multipath *) ti->private;
1353 struct priority_group *pg;
1358 spin_lock_irqsave(&m->lock, flags);
1361 if (type == STATUSTYPE_INFO)
1362 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1364 DMEMIT("%u ", m->queue_if_no_path +
1365 (m->pg_init_retries > 0) * 2 +
1366 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2);
1367 if (m->queue_if_no_path)
1368 DMEMIT("queue_if_no_path ");
1369 if (m->pg_init_retries)
1370 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1371 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1372 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1375 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1378 DMEMIT("1 %s ", m->hw_handler_name);
1380 DMEMIT("%u ", m->nr_priority_groups);
1383 pg_num = m->next_pg->pg_num;
1384 else if (m->current_pg)
1385 pg_num = m->current_pg->pg_num;
1387 pg_num = (m->nr_priority_groups ? 1 : 0);
1389 DMEMIT("%u ", pg_num);
1392 case STATUSTYPE_INFO:
1393 list_for_each_entry(pg, &m->priority_groups, list) {
1395 state = 'D'; /* Disabled */
1396 else if (pg == m->current_pg)
1397 state = 'A'; /* Currently Active */
1399 state = 'E'; /* Enabled */
1401 DMEMIT("%c ", state);
1403 if (pg->ps.type->status)
1404 sz += pg->ps.type->status(&pg->ps, NULL, type,
1410 DMEMIT("%u %u ", pg->nr_pgpaths,
1411 pg->ps.type->info_args);
1413 list_for_each_entry(p, &pg->pgpaths, list) {
1414 DMEMIT("%s %s %u ", p->path.dev->name,
1415 p->is_active ? "A" : "F",
1417 if (pg->ps.type->status)
1418 sz += pg->ps.type->status(&pg->ps,
1419 &p->path, type, result + sz,
1425 case STATUSTYPE_TABLE:
1426 list_for_each_entry(pg, &m->priority_groups, list) {
1427 DMEMIT("%s ", pg->ps.type->name);
1429 if (pg->ps.type->status)
1430 sz += pg->ps.type->status(&pg->ps, NULL, type,
1436 DMEMIT("%u %u ", pg->nr_pgpaths,
1437 pg->ps.type->table_args);
1439 list_for_each_entry(p, &pg->pgpaths, list) {
1440 DMEMIT("%s ", p->path.dev->name);
1441 if (pg->ps.type->status)
1442 sz += pg->ps.type->status(&pg->ps,
1443 &p->path, type, result + sz,
1450 spin_unlock_irqrestore(&m->lock, flags);
1455 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1459 struct multipath *m = (struct multipath *) ti->private;
1462 mutex_lock(&m->work_mutex);
1464 if (dm_suspended(ti)) {
1470 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1471 r = queue_if_no_path(m, 1, 0);
1473 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1474 r = queue_if_no_path(m, 0, 0);
1480 DMWARN("Unrecognised multipath message received.");
1484 if (!strcasecmp(argv[0], "disable_group")) {
1485 r = bypass_pg_num(m, argv[1], 1);
1487 } else if (!strcasecmp(argv[0], "enable_group")) {
1488 r = bypass_pg_num(m, argv[1], 0);
1490 } else if (!strcasecmp(argv[0], "switch_group")) {
1491 r = switch_pg_num(m, argv[1]);
1493 } else if (!strcasecmp(argv[0], "reinstate_path"))
1494 action = reinstate_path;
1495 else if (!strcasecmp(argv[0], "fail_path"))
1498 DMWARN("Unrecognised multipath message received.");
1502 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1504 DMWARN("message: error getting device %s",
1509 r = action_dev(m, dev, action);
1511 dm_put_device(ti, dev);
1514 mutex_unlock(&m->work_mutex);
1518 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1521 struct multipath *m = (struct multipath *) ti->private;
1522 struct block_device *bdev = NULL;
1524 unsigned long flags;
1527 spin_lock_irqsave(&m->lock, flags);
1529 if (!m->current_pgpath)
1530 __choose_pgpath(m, 0);
1532 if (m->current_pgpath) {
1533 bdev = m->current_pgpath->path.dev->bdev;
1534 mode = m->current_pgpath->path.dev->mode;
1542 spin_unlock_irqrestore(&m->lock, flags);
1545 * Only pass ioctls through if the device sizes match exactly.
1547 if (!r && ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT)
1548 r = scsi_verify_blk_ioctl(NULL, cmd);
1550 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1553 static int multipath_iterate_devices(struct dm_target *ti,
1554 iterate_devices_callout_fn fn, void *data)
1556 struct multipath *m = ti->private;
1557 struct priority_group *pg;
1561 list_for_each_entry(pg, &m->priority_groups, list) {
1562 list_for_each_entry(p, &pg->pgpaths, list) {
1563 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1573 static int __pgpath_busy(struct pgpath *pgpath)
1575 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1577 return dm_underlying_device_busy(q);
1581 * We return "busy", only when we can map I/Os but underlying devices
1582 * are busy (so even if we map I/Os now, the I/Os will wait on
1583 * the underlying queue).
1584 * In other words, if we want to kill I/Os or queue them inside us
1585 * due to map unavailability, we don't return "busy". Otherwise,
1586 * dm core won't give us the I/Os and we can't do what we want.
1588 static int multipath_busy(struct dm_target *ti)
1590 int busy = 0, has_active = 0;
1591 struct multipath *m = ti->private;
1592 struct priority_group *pg;
1593 struct pgpath *pgpath;
1594 unsigned long flags;
1596 spin_lock_irqsave(&m->lock, flags);
1598 /* Guess which priority_group will be used at next mapping time */
1599 if (unlikely(!m->current_pgpath && m->next_pg))
1601 else if (likely(m->current_pg))
1605 * We don't know which pg will be used at next mapping time.
1606 * We don't call __choose_pgpath() here to avoid to trigger
1607 * pg_init just by busy checking.
1608 * So we don't know whether underlying devices we will be using
1609 * at next mapping time are busy or not. Just try mapping.
1614 * If there is one non-busy active path at least, the path selector
1615 * will be able to select it. So we consider such a pg as not busy.
1618 list_for_each_entry(pgpath, &pg->pgpaths, list)
1619 if (pgpath->is_active) {
1622 if (!__pgpath_busy(pgpath)) {
1630 * No active path in this pg, so this pg won't be used and
1631 * the current_pg will be changed at next mapping time.
1632 * We need to try mapping to determine it.
1637 spin_unlock_irqrestore(&m->lock, flags);
1642 /*-----------------------------------------------------------------
1644 *---------------------------------------------------------------*/
1645 static struct target_type multipath_target = {
1646 .name = "multipath",
1647 .version = {1, 3, 0},
1648 .module = THIS_MODULE,
1649 .ctr = multipath_ctr,
1650 .dtr = multipath_dtr,
1651 .map_rq = multipath_map,
1652 .rq_end_io = multipath_end_io,
1653 .presuspend = multipath_presuspend,
1654 .postsuspend = multipath_postsuspend,
1655 .resume = multipath_resume,
1656 .status = multipath_status,
1657 .message = multipath_message,
1658 .ioctl = multipath_ioctl,
1659 .iterate_devices = multipath_iterate_devices,
1660 .busy = multipath_busy,
1663 static int __init dm_multipath_init(void)
1667 /* allocate a slab for the dm_ios */
1668 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1672 r = dm_register_target(&multipath_target);
1674 DMERR("register failed %d", r);
1675 kmem_cache_destroy(_mpio_cache);
1679 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1681 DMERR("failed to create workqueue kmpathd");
1682 dm_unregister_target(&multipath_target);
1683 kmem_cache_destroy(_mpio_cache);
1688 * A separate workqueue is used to handle the device handlers
1689 * to avoid overloading existing workqueue. Overloading the
1690 * old workqueue would also create a bottleneck in the
1691 * path of the storage hardware device activation.
1693 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1695 if (!kmpath_handlerd) {
1696 DMERR("failed to create workqueue kmpath_handlerd");
1697 destroy_workqueue(kmultipathd);
1698 dm_unregister_target(&multipath_target);
1699 kmem_cache_destroy(_mpio_cache);
1703 DMINFO("version %u.%u.%u loaded",
1704 multipath_target.version[0], multipath_target.version[1],
1705 multipath_target.version[2]);
1710 static void __exit dm_multipath_exit(void)
1712 destroy_workqueue(kmpath_handlerd);
1713 destroy_workqueue(kmultipathd);
1715 dm_unregister_target(&multipath_target);
1716 kmem_cache_destroy(_mpio_cache);
1719 module_init(dm_multipath_init);
1720 module_exit(dm_multipath_exit);
1722 MODULE_DESCRIPTION(DM_NAME " multipath target");
1723 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1724 MODULE_LICENSE("GPL");