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;
66 const char *hw_handler_name;
67 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; /* Must we queue all I/O? */
85 unsigned queue_if_no_path; /* Queue I/O if last path fails? */
86 unsigned saved_queue_if_no_path;/* Saved state during suspension */
87 unsigned pg_init_retries; /* Number of times to retry pg_init */
88 unsigned pg_init_count; /* Number of times pg_init called */
89 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
91 struct work_struct process_queued_ios;
92 struct list_head queued_ios;
95 struct work_struct trigger_event;
98 * We must use a mempool of dm_mpath_io structs so that we
99 * can resubmit bios on error.
101 mempool_t *mpio_pool;
103 struct mutex work_mutex;
107 * Context information attached to each bio we process.
110 struct pgpath *pgpath;
114 typedef int (*action_fn) (struct pgpath *pgpath);
116 #define MIN_IOS 256 /* Mempool size */
118 static struct kmem_cache *_mpio_cache;
120 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
121 static void process_queued_ios(struct work_struct *work);
122 static void trigger_event(struct work_struct *work);
123 static void activate_path(struct work_struct *work);
126 /*-----------------------------------------------
127 * Allocation routines
128 *-----------------------------------------------*/
130 static struct pgpath *alloc_pgpath(void)
132 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
135 pgpath->is_active = 1;
136 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
142 static void free_pgpath(struct pgpath *pgpath)
147 static struct priority_group *alloc_priority_group(void)
149 struct priority_group *pg;
151 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
154 INIT_LIST_HEAD(&pg->pgpaths);
159 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
161 struct pgpath *pgpath, *tmp;
162 struct multipath *m = ti->private;
164 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
165 list_del(&pgpath->list);
166 if (m->hw_handler_name)
167 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
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);
230 /*-----------------------------------------------
232 *-----------------------------------------------*/
234 static void __pg_init_all_paths(struct multipath *m)
236 struct pgpath *pgpath;
237 unsigned long pg_init_delay = 0;
240 m->pg_init_required = 0;
241 if (m->pg_init_delay_retry)
242 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
243 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
244 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
245 /* Skip failed paths */
246 if (!pgpath->is_active)
248 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
250 m->pg_init_in_progress++;
254 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
256 m->current_pg = pgpath->pg;
258 /* Must we initialise the PG first, and queue I/O till it's ready? */
259 if (m->hw_handler_name) {
260 m->pg_init_required = 1;
263 m->pg_init_required = 0;
267 m->pg_init_count = 0;
270 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
273 struct dm_path *path;
275 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
279 m->current_pgpath = path_to_pgpath(path);
281 if (m->current_pg != pg)
282 __switch_pg(m, m->current_pgpath);
287 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
289 struct priority_group *pg;
290 unsigned bypassed = 1;
292 if (!m->nr_valid_paths)
295 /* Were we instructed to switch PG? */
299 if (!__choose_path_in_pg(m, pg, nr_bytes))
303 /* Don't change PG until it has no remaining paths */
304 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
308 * Loop through priority groups until we find a valid path.
309 * First time we skip PGs marked 'bypassed'.
310 * Second time we only try the ones we skipped.
313 list_for_each_entry(pg, &m->priority_groups, list) {
314 if (pg->bypassed == bypassed)
316 if (!__choose_path_in_pg(m, pg, nr_bytes))
319 } while (bypassed--);
322 m->current_pgpath = NULL;
323 m->current_pg = NULL;
327 * Check whether bios must be queued in the device-mapper core rather
328 * than here in the target.
330 * m->lock must be held on entry.
332 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
333 * same value then we are not between multipath_presuspend()
334 * and multipath_resume() calls and we have no need to check
335 * for the DMF_NOFLUSH_SUSPENDING flag.
337 static int __must_push_back(struct multipath *m)
339 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
340 dm_noflush_suspending(m->ti));
343 static int map_io(struct multipath *m, struct request *clone,
344 struct dm_mpath_io *mpio, unsigned was_queued)
346 int r = DM_MAPIO_REMAPPED;
347 size_t nr_bytes = blk_rq_bytes(clone);
349 struct pgpath *pgpath;
350 struct block_device *bdev;
352 spin_lock_irqsave(&m->lock, flags);
354 /* Do we need to select a new pgpath? */
355 if (!m->current_pgpath ||
356 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
357 __choose_pgpath(m, nr_bytes);
359 pgpath = m->current_pgpath;
364 if ((pgpath && m->queue_io) ||
365 (!pgpath && m->queue_if_no_path)) {
366 /* Queue for the daemon to resubmit */
367 list_add_tail(&clone->queuelist, &m->queued_ios);
369 if ((m->pg_init_required && !m->pg_init_in_progress) ||
371 queue_work(kmultipathd, &m->process_queued_ios);
373 r = DM_MAPIO_SUBMITTED;
375 bdev = pgpath->path.dev->bdev;
376 clone->q = bdev_get_queue(bdev);
377 clone->rq_disk = bdev->bd_disk;
378 } else if (__must_push_back(m))
379 r = DM_MAPIO_REQUEUE;
381 r = -EIO; /* Failed */
383 mpio->pgpath = pgpath;
384 mpio->nr_bytes = nr_bytes;
386 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
387 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
390 spin_unlock_irqrestore(&m->lock, flags);
396 * If we run out of usable paths, should we queue I/O or error it?
398 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
399 unsigned save_old_value)
403 spin_lock_irqsave(&m->lock, flags);
406 m->saved_queue_if_no_path = m->queue_if_no_path;
408 m->saved_queue_if_no_path = queue_if_no_path;
409 m->queue_if_no_path = queue_if_no_path;
410 if (!m->queue_if_no_path && m->queue_size)
411 queue_work(kmultipathd, &m->process_queued_ios);
413 spin_unlock_irqrestore(&m->lock, flags);
418 /*-----------------------------------------------------------------
419 * The multipath daemon is responsible for resubmitting queued ios.
420 *---------------------------------------------------------------*/
422 static void dispatch_queued_ios(struct multipath *m)
426 struct dm_mpath_io *mpio;
427 union map_info *info;
428 struct request *clone, *n;
431 spin_lock_irqsave(&m->lock, flags);
432 list_splice_init(&m->queued_ios, &cl);
433 spin_unlock_irqrestore(&m->lock, flags);
435 list_for_each_entry_safe(clone, n, &cl, queuelist) {
436 list_del_init(&clone->queuelist);
438 info = dm_get_rq_mapinfo(clone);
441 r = map_io(m, clone, mpio, 1);
443 mempool_free(mpio, m->mpio_pool);
444 dm_kill_unmapped_request(clone, r);
445 } else if (r == DM_MAPIO_REMAPPED)
446 dm_dispatch_request(clone);
447 else if (r == DM_MAPIO_REQUEUE) {
448 mempool_free(mpio, m->mpio_pool);
449 dm_requeue_unmapped_request(clone);
454 static void process_queued_ios(struct work_struct *work)
456 struct multipath *m =
457 container_of(work, struct multipath, process_queued_ios);
458 struct pgpath *pgpath = NULL;
459 unsigned must_queue = 1;
462 spin_lock_irqsave(&m->lock, flags);
467 if (!m->current_pgpath)
468 __choose_pgpath(m, 0);
470 pgpath = m->current_pgpath;
472 if ((pgpath && !m->queue_io) ||
473 (!pgpath && !m->queue_if_no_path))
476 if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
477 __pg_init_all_paths(m);
480 spin_unlock_irqrestore(&m->lock, flags);
482 dispatch_queued_ios(m);
486 * An event is triggered whenever a path is taken out of use.
487 * Includes path failure and PG bypass.
489 static void trigger_event(struct work_struct *work)
491 struct multipath *m =
492 container_of(work, struct multipath, trigger_event);
494 dm_table_event(m->ti->table);
497 /*-----------------------------------------------------------------
498 * Constructor/argument parsing:
499 * <#multipath feature args> [<arg>]*
500 * <#hw_handler args> [hw_handler [<arg>]*]
502 * <initial priority group>
503 * [<selector> <#selector args> [<arg>]*
504 * <#paths> <#per-path selector args>
505 * [<path> [<arg>]* ]+ ]+
506 *---------------------------------------------------------------*/
507 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
508 struct dm_target *ti)
511 struct path_selector_type *pst;
514 static struct dm_arg _args[] = {
515 {0, 1024, "invalid number of path selector args"},
518 pst = dm_get_path_selector(dm_shift_arg(as));
520 ti->error = "unknown path selector type";
524 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
526 dm_put_path_selector(pst);
530 r = pst->create(&pg->ps, ps_argc, as->argv);
532 dm_put_path_selector(pst);
533 ti->error = "path selector constructor failed";
538 dm_consume_args(as, ps_argc);
543 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
544 struct dm_target *ti)
548 struct multipath *m = ti->private;
550 /* we need at least a path arg */
552 ti->error = "no device given";
553 return ERR_PTR(-EINVAL);
558 return ERR_PTR(-ENOMEM);
560 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
563 ti->error = "error getting device";
567 if (m->hw_handler_name) {
568 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
570 r = scsi_dh_attach(q, m->hw_handler_name);
573 * Already attached to different hw_handler,
574 * try to reattach with correct one.
577 r = scsi_dh_attach(q, m->hw_handler_name);
581 ti->error = "error attaching hardware handler";
582 dm_put_device(ti, p->path.dev);
586 if (m->hw_handler_params) {
587 r = scsi_dh_set_params(q, m->hw_handler_params);
589 ti->error = "unable to set hardware "
590 "handler parameters";
592 dm_put_device(ti, p->path.dev);
598 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
600 dm_put_device(ti, p->path.dev);
611 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
614 static struct dm_arg _args[] = {
615 {1, 1024, "invalid number of paths"},
616 {0, 1024, "invalid number of selector args"}
620 unsigned i, nr_selector_args, nr_args;
621 struct priority_group *pg;
622 struct dm_target *ti = m->ti;
626 ti->error = "not enough priority group arguments";
627 return ERR_PTR(-EINVAL);
630 pg = alloc_priority_group();
632 ti->error = "couldn't allocate priority group";
633 return ERR_PTR(-ENOMEM);
637 r = parse_path_selector(as, pg, ti);
644 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
648 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
652 nr_args = 1 + nr_selector_args;
653 for (i = 0; i < pg->nr_pgpaths; i++) {
654 struct pgpath *pgpath;
655 struct dm_arg_set path_args;
657 if (as->argc < nr_args) {
658 ti->error = "not enough path parameters";
663 path_args.argc = nr_args;
664 path_args.argv = as->argv;
666 pgpath = parse_path(&path_args, &pg->ps, ti);
667 if (IS_ERR(pgpath)) {
673 list_add_tail(&pgpath->list, &pg->pgpaths);
674 dm_consume_args(as, nr_args);
680 free_priority_group(pg, ti);
684 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
688 struct dm_target *ti = m->ti;
690 static struct dm_arg _args[] = {
691 {0, 1024, "invalid number of hardware handler args"},
694 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
700 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
701 request_module("scsi_dh_%s", m->hw_handler_name);
702 if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
703 ti->error = "unknown hardware handler type";
712 for (i = 0; i <= hw_argc - 2; i++)
713 len += strlen(as->argv[i]) + 1;
714 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
716 ti->error = "memory allocation failed";
720 j = sprintf(p, "%d", hw_argc - 1);
721 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
722 j = sprintf(p, "%s", as->argv[i]);
724 dm_consume_args(as, hw_argc - 1);
728 kfree(m->hw_handler_name);
729 m->hw_handler_name = NULL;
733 static int parse_features(struct dm_arg_set *as, struct multipath *m)
737 struct dm_target *ti = m->ti;
738 const char *arg_name;
740 static struct dm_arg _args[] = {
741 {0, 5, "invalid number of feature args"},
742 {1, 50, "pg_init_retries must be between 1 and 50"},
743 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
746 r = dm_read_arg_group(_args, as, &argc, &ti->error);
754 arg_name = dm_shift_arg(as);
757 if (!strcasecmp(arg_name, "queue_if_no_path")) {
758 r = queue_if_no_path(m, 1, 0);
762 if (!strcasecmp(arg_name, "pg_init_retries") &&
764 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
769 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
771 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
776 ti->error = "Unrecognised multipath feature request";
778 } while (argc && !r);
783 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
786 /* target arguments */
787 static struct dm_arg _args[] = {
788 {0, 1024, "invalid number of priority groups"},
789 {0, 1024, "invalid initial priority group number"},
794 struct dm_arg_set as;
795 unsigned pg_count = 0;
796 unsigned next_pg_num;
801 m = alloc_multipath(ti);
803 ti->error = "can't allocate multipath";
807 r = parse_features(&as, m);
811 r = parse_hw_handler(&as, m);
815 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
819 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
823 if ((!m->nr_priority_groups && next_pg_num) ||
824 (m->nr_priority_groups && !next_pg_num)) {
825 ti->error = "invalid initial priority group";
830 /* parse the priority groups */
832 struct priority_group *pg;
834 pg = parse_priority_group(&as, m);
840 m->nr_valid_paths += pg->nr_pgpaths;
841 list_add_tail(&pg->list, &m->priority_groups);
843 pg->pg_num = pg_count;
848 if (pg_count != m->nr_priority_groups) {
849 ti->error = "priority group count mismatch";
854 ti->num_flush_requests = 1;
855 ti->num_discard_requests = 1;
864 static void multipath_wait_for_pg_init_completion(struct multipath *m)
866 DECLARE_WAITQUEUE(wait, current);
869 add_wait_queue(&m->pg_init_wait, &wait);
872 set_current_state(TASK_UNINTERRUPTIBLE);
874 spin_lock_irqsave(&m->lock, flags);
875 if (!m->pg_init_in_progress) {
876 spin_unlock_irqrestore(&m->lock, flags);
879 spin_unlock_irqrestore(&m->lock, flags);
883 set_current_state(TASK_RUNNING);
885 remove_wait_queue(&m->pg_init_wait, &wait);
888 static void flush_multipath_work(struct multipath *m)
890 flush_workqueue(kmpath_handlerd);
891 multipath_wait_for_pg_init_completion(m);
892 flush_workqueue(kmultipathd);
893 flush_work_sync(&m->trigger_event);
896 static void multipath_dtr(struct dm_target *ti)
898 struct multipath *m = ti->private;
900 flush_multipath_work(m);
905 * Map cloned requests
907 static int multipath_map(struct dm_target *ti, struct request *clone,
908 union map_info *map_context)
911 struct dm_mpath_io *mpio;
912 struct multipath *m = (struct multipath *) ti->private;
914 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
916 /* ENOMEM, requeue */
917 return DM_MAPIO_REQUEUE;
918 memset(mpio, 0, sizeof(*mpio));
920 map_context->ptr = mpio;
921 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
922 r = map_io(m, clone, mpio, 0);
923 if (r < 0 || r == DM_MAPIO_REQUEUE)
924 mempool_free(mpio, m->mpio_pool);
930 * Take a path out of use.
932 static int fail_path(struct pgpath *pgpath)
935 struct multipath *m = pgpath->pg->m;
937 spin_lock_irqsave(&m->lock, flags);
939 if (!pgpath->is_active)
942 DMWARN("Failing path %s.", pgpath->path.dev->name);
944 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
945 pgpath->is_active = 0;
946 pgpath->fail_count++;
950 if (pgpath == m->current_pgpath)
951 m->current_pgpath = NULL;
953 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
954 pgpath->path.dev->name, m->nr_valid_paths);
956 schedule_work(&m->trigger_event);
959 spin_unlock_irqrestore(&m->lock, flags);
965 * Reinstate a previously-failed path
967 static int reinstate_path(struct pgpath *pgpath)
971 struct multipath *m = pgpath->pg->m;
973 spin_lock_irqsave(&m->lock, flags);
975 if (pgpath->is_active)
978 if (!pgpath->pg->ps.type->reinstate_path) {
979 DMWARN("Reinstate path not supported by path selector %s",
980 pgpath->pg->ps.type->name);
985 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
989 pgpath->is_active = 1;
991 if (!m->nr_valid_paths++ && m->queue_size) {
992 m->current_pgpath = NULL;
993 queue_work(kmultipathd, &m->process_queued_ios);
994 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
995 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
996 m->pg_init_in_progress++;
999 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1000 pgpath->path.dev->name, m->nr_valid_paths);
1002 schedule_work(&m->trigger_event);
1005 spin_unlock_irqrestore(&m->lock, flags);
1011 * Fail or reinstate all paths that match the provided struct dm_dev.
1013 static int action_dev(struct multipath *m, struct dm_dev *dev,
1017 struct pgpath *pgpath;
1018 struct priority_group *pg;
1020 list_for_each_entry(pg, &m->priority_groups, list) {
1021 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1022 if (pgpath->path.dev == dev)
1031 * Temporarily try to avoid having to use the specified PG
1033 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1036 unsigned long flags;
1038 spin_lock_irqsave(&m->lock, flags);
1040 pg->bypassed = bypassed;
1041 m->current_pgpath = NULL;
1042 m->current_pg = NULL;
1044 spin_unlock_irqrestore(&m->lock, flags);
1046 schedule_work(&m->trigger_event);
1050 * Switch to using the specified PG from the next I/O that gets mapped
1052 static int switch_pg_num(struct multipath *m, const char *pgstr)
1054 struct priority_group *pg;
1056 unsigned long flags;
1058 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1059 (pgnum > m->nr_priority_groups)) {
1060 DMWARN("invalid PG number supplied to switch_pg_num");
1064 spin_lock_irqsave(&m->lock, flags);
1065 list_for_each_entry(pg, &m->priority_groups, list) {
1070 m->current_pgpath = NULL;
1071 m->current_pg = NULL;
1074 spin_unlock_irqrestore(&m->lock, flags);
1076 schedule_work(&m->trigger_event);
1081 * Set/clear bypassed status of a PG.
1082 * PGs are numbered upwards from 1 in the order they were declared.
1084 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1086 struct priority_group *pg;
1089 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1090 (pgnum > m->nr_priority_groups)) {
1091 DMWARN("invalid PG number supplied to bypass_pg");
1095 list_for_each_entry(pg, &m->priority_groups, list) {
1100 bypass_pg(m, pg, bypassed);
1105 * Should we retry pg_init immediately?
1107 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1109 unsigned long flags;
1110 int limit_reached = 0;
1112 spin_lock_irqsave(&m->lock, flags);
1114 if (m->pg_init_count <= m->pg_init_retries)
1115 m->pg_init_required = 1;
1119 spin_unlock_irqrestore(&m->lock, flags);
1121 return limit_reached;
1124 static void pg_init_done(void *data, int errors)
1126 struct pgpath *pgpath = data;
1127 struct priority_group *pg = pgpath->pg;
1128 struct multipath *m = pg->m;
1129 unsigned long flags;
1130 unsigned delay_retry = 0;
1132 /* device or driver problems */
1137 if (!m->hw_handler_name) {
1141 DMERR("Could not failover the device: Handler scsi_dh_%s "
1142 "Error %d.", m->hw_handler_name, errors);
1144 * Fail path for now, so we do not ping pong
1148 case SCSI_DH_DEV_TEMP_BUSY:
1150 * Probably doing something like FW upgrade on the
1151 * controller so try the other pg.
1153 bypass_pg(m, pg, 1);
1156 /* Wait before retrying. */
1158 case SCSI_DH_IMM_RETRY:
1159 case SCSI_DH_RES_TEMP_UNAVAIL:
1160 if (pg_init_limit_reached(m, pgpath))
1166 * We probably do not want to fail the path for a device
1167 * error, but this is what the old dm did. In future
1168 * patches we can do more advanced handling.
1173 spin_lock_irqsave(&m->lock, flags);
1175 if (pgpath == m->current_pgpath) {
1176 DMERR("Could not failover device. Error %d.", errors);
1177 m->current_pgpath = NULL;
1178 m->current_pg = NULL;
1180 } else if (!m->pg_init_required)
1183 if (--m->pg_init_in_progress)
1184 /* Activations of other paths are still on going */
1187 if (!m->pg_init_required)
1190 m->pg_init_delay_retry = delay_retry;
1191 queue_work(kmultipathd, &m->process_queued_ios);
1194 * Wake up any thread waiting to suspend.
1196 wake_up(&m->pg_init_wait);
1199 spin_unlock_irqrestore(&m->lock, flags);
1202 static void activate_path(struct work_struct *work)
1204 struct pgpath *pgpath =
1205 container_of(work, struct pgpath, activate_path.work);
1207 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1208 pg_init_done, pgpath);
1214 static int do_end_io(struct multipath *m, struct request *clone,
1215 int error, struct dm_mpath_io *mpio)
1218 * We don't queue any clone request inside the multipath target
1219 * during end I/O handling, since those clone requests don't have
1220 * bio clones. If we queue them inside the multipath target,
1221 * we need to make bio clones, that requires memory allocation.
1222 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1223 * don't have bio clones.)
1224 * Instead of queueing the clone request here, we queue the original
1225 * request into dm core, which will remake a clone request and
1226 * clone bios for it and resubmit it later.
1228 int r = DM_ENDIO_REQUEUE;
1229 unsigned long flags;
1231 if (!error && !clone->errors)
1232 return 0; /* I/O complete */
1234 if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1238 fail_path(mpio->pgpath);
1240 spin_lock_irqsave(&m->lock, flags);
1241 if (!m->nr_valid_paths) {
1242 if (!m->queue_if_no_path) {
1243 if (!__must_push_back(m))
1246 if (error == -EBADE)
1250 spin_unlock_irqrestore(&m->lock, flags);
1255 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1256 int error, union map_info *map_context)
1258 struct multipath *m = ti->private;
1259 struct dm_mpath_io *mpio = map_context->ptr;
1260 struct pgpath *pgpath = mpio->pgpath;
1261 struct path_selector *ps;
1264 r = do_end_io(m, clone, error, mpio);
1266 ps = &pgpath->pg->ps;
1267 if (ps->type->end_io)
1268 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1270 mempool_free(mpio, m->mpio_pool);
1276 * Suspend can't complete until all the I/O is processed so if
1277 * the last path fails we must error any remaining I/O.
1278 * Note that if the freeze_bdev fails while suspending, the
1279 * queue_if_no_path state is lost - userspace should reset it.
1281 static void multipath_presuspend(struct dm_target *ti)
1283 struct multipath *m = (struct multipath *) ti->private;
1285 queue_if_no_path(m, 0, 1);
1288 static void multipath_postsuspend(struct dm_target *ti)
1290 struct multipath *m = ti->private;
1292 mutex_lock(&m->work_mutex);
1293 flush_multipath_work(m);
1294 mutex_unlock(&m->work_mutex);
1298 * Restore the queue_if_no_path setting.
1300 static void multipath_resume(struct dm_target *ti)
1302 struct multipath *m = (struct multipath *) ti->private;
1303 unsigned long flags;
1305 spin_lock_irqsave(&m->lock, flags);
1306 m->queue_if_no_path = m->saved_queue_if_no_path;
1307 spin_unlock_irqrestore(&m->lock, flags);
1311 * Info output has the following format:
1312 * num_multipath_feature_args [multipath_feature_args]*
1313 * num_handler_status_args [handler_status_args]*
1314 * num_groups init_group_number
1315 * [A|D|E num_ps_status_args [ps_status_args]*
1316 * num_paths num_selector_args
1317 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1319 * Table output has the following format (identical to the constructor string):
1320 * num_feature_args [features_args]*
1321 * num_handler_args hw_handler [hw_handler_args]*
1322 * num_groups init_group_number
1323 * [priority selector-name num_ps_args [ps_args]*
1324 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1326 static int multipath_status(struct dm_target *ti, status_type_t type,
1327 char *result, unsigned int maxlen)
1330 unsigned long flags;
1331 struct multipath *m = (struct multipath *) ti->private;
1332 struct priority_group *pg;
1337 spin_lock_irqsave(&m->lock, flags);
1340 if (type == STATUSTYPE_INFO)
1341 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1343 DMEMIT("%u ", m->queue_if_no_path +
1344 (m->pg_init_retries > 0) * 2 +
1345 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2);
1346 if (m->queue_if_no_path)
1347 DMEMIT("queue_if_no_path ");
1348 if (m->pg_init_retries)
1349 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1350 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1351 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1354 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1357 DMEMIT("1 %s ", m->hw_handler_name);
1359 DMEMIT("%u ", m->nr_priority_groups);
1362 pg_num = m->next_pg->pg_num;
1363 else if (m->current_pg)
1364 pg_num = m->current_pg->pg_num;
1366 pg_num = (m->nr_priority_groups ? 1 : 0);
1368 DMEMIT("%u ", pg_num);
1371 case STATUSTYPE_INFO:
1372 list_for_each_entry(pg, &m->priority_groups, list) {
1374 state = 'D'; /* Disabled */
1375 else if (pg == m->current_pg)
1376 state = 'A'; /* Currently Active */
1378 state = 'E'; /* Enabled */
1380 DMEMIT("%c ", state);
1382 if (pg->ps.type->status)
1383 sz += pg->ps.type->status(&pg->ps, NULL, type,
1389 DMEMIT("%u %u ", pg->nr_pgpaths,
1390 pg->ps.type->info_args);
1392 list_for_each_entry(p, &pg->pgpaths, list) {
1393 DMEMIT("%s %s %u ", p->path.dev->name,
1394 p->is_active ? "A" : "F",
1396 if (pg->ps.type->status)
1397 sz += pg->ps.type->status(&pg->ps,
1398 &p->path, type, result + sz,
1404 case STATUSTYPE_TABLE:
1405 list_for_each_entry(pg, &m->priority_groups, list) {
1406 DMEMIT("%s ", pg->ps.type->name);
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->table_args);
1418 list_for_each_entry(p, &pg->pgpaths, list) {
1419 DMEMIT("%s ", p->path.dev->name);
1420 if (pg->ps.type->status)
1421 sz += pg->ps.type->status(&pg->ps,
1422 &p->path, type, result + sz,
1429 spin_unlock_irqrestore(&m->lock, flags);
1434 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1438 struct multipath *m = (struct multipath *) ti->private;
1441 mutex_lock(&m->work_mutex);
1443 if (dm_suspended(ti)) {
1449 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1450 r = queue_if_no_path(m, 1, 0);
1452 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1453 r = queue_if_no_path(m, 0, 0);
1459 DMWARN("Unrecognised multipath message received.");
1463 if (!strcasecmp(argv[0], "disable_group")) {
1464 r = bypass_pg_num(m, argv[1], 1);
1466 } else if (!strcasecmp(argv[0], "enable_group")) {
1467 r = bypass_pg_num(m, argv[1], 0);
1469 } else if (!strcasecmp(argv[0], "switch_group")) {
1470 r = switch_pg_num(m, argv[1]);
1472 } else if (!strcasecmp(argv[0], "reinstate_path"))
1473 action = reinstate_path;
1474 else if (!strcasecmp(argv[0], "fail_path"))
1477 DMWARN("Unrecognised multipath message received.");
1481 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1483 DMWARN("message: error getting device %s",
1488 r = action_dev(m, dev, action);
1490 dm_put_device(ti, dev);
1493 mutex_unlock(&m->work_mutex);
1497 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1500 struct multipath *m = (struct multipath *) ti->private;
1501 struct block_device *bdev = NULL;
1503 unsigned long flags;
1506 spin_lock_irqsave(&m->lock, flags);
1508 if (!m->current_pgpath)
1509 __choose_pgpath(m, 0);
1511 if (m->current_pgpath) {
1512 bdev = m->current_pgpath->path.dev->bdev;
1513 mode = m->current_pgpath->path.dev->mode;
1521 spin_unlock_irqrestore(&m->lock, flags);
1523 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1526 static int multipath_iterate_devices(struct dm_target *ti,
1527 iterate_devices_callout_fn fn, void *data)
1529 struct multipath *m = ti->private;
1530 struct priority_group *pg;
1534 list_for_each_entry(pg, &m->priority_groups, list) {
1535 list_for_each_entry(p, &pg->pgpaths, list) {
1536 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1546 static int __pgpath_busy(struct pgpath *pgpath)
1548 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1550 return dm_underlying_device_busy(q);
1554 * We return "busy", only when we can map I/Os but underlying devices
1555 * are busy (so even if we map I/Os now, the I/Os will wait on
1556 * the underlying queue).
1557 * In other words, if we want to kill I/Os or queue them inside us
1558 * due to map unavailability, we don't return "busy". Otherwise,
1559 * dm core won't give us the I/Os and we can't do what we want.
1561 static int multipath_busy(struct dm_target *ti)
1563 int busy = 0, has_active = 0;
1564 struct multipath *m = ti->private;
1565 struct priority_group *pg;
1566 struct pgpath *pgpath;
1567 unsigned long flags;
1569 spin_lock_irqsave(&m->lock, flags);
1571 /* Guess which priority_group will be used at next mapping time */
1572 if (unlikely(!m->current_pgpath && m->next_pg))
1574 else if (likely(m->current_pg))
1578 * We don't know which pg will be used at next mapping time.
1579 * We don't call __choose_pgpath() here to avoid to trigger
1580 * pg_init just by busy checking.
1581 * So we don't know whether underlying devices we will be using
1582 * at next mapping time are busy or not. Just try mapping.
1587 * If there is one non-busy active path at least, the path selector
1588 * will be able to select it. So we consider such a pg as not busy.
1591 list_for_each_entry(pgpath, &pg->pgpaths, list)
1592 if (pgpath->is_active) {
1595 if (!__pgpath_busy(pgpath)) {
1603 * No active path in this pg, so this pg won't be used and
1604 * the current_pg will be changed at next mapping time.
1605 * We need to try mapping to determine it.
1610 spin_unlock_irqrestore(&m->lock, flags);
1615 /*-----------------------------------------------------------------
1617 *---------------------------------------------------------------*/
1618 static struct target_type multipath_target = {
1619 .name = "multipath",
1620 .version = {1, 3, 0},
1621 .module = THIS_MODULE,
1622 .ctr = multipath_ctr,
1623 .dtr = multipath_dtr,
1624 .map_rq = multipath_map,
1625 .rq_end_io = multipath_end_io,
1626 .presuspend = multipath_presuspend,
1627 .postsuspend = multipath_postsuspend,
1628 .resume = multipath_resume,
1629 .status = multipath_status,
1630 .message = multipath_message,
1631 .ioctl = multipath_ioctl,
1632 .iterate_devices = multipath_iterate_devices,
1633 .busy = multipath_busy,
1636 static int __init dm_multipath_init(void)
1640 /* allocate a slab for the dm_ios */
1641 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1645 r = dm_register_target(&multipath_target);
1647 DMERR("register failed %d", r);
1648 kmem_cache_destroy(_mpio_cache);
1652 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1654 DMERR("failed to create workqueue kmpathd");
1655 dm_unregister_target(&multipath_target);
1656 kmem_cache_destroy(_mpio_cache);
1661 * A separate workqueue is used to handle the device handlers
1662 * to avoid overloading existing workqueue. Overloading the
1663 * old workqueue would also create a bottleneck in the
1664 * path of the storage hardware device activation.
1666 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1668 if (!kmpath_handlerd) {
1669 DMERR("failed to create workqueue kmpath_handlerd");
1670 destroy_workqueue(kmultipathd);
1671 dm_unregister_target(&multipath_target);
1672 kmem_cache_destroy(_mpio_cache);
1676 DMINFO("version %u.%u.%u loaded",
1677 multipath_target.version[0], multipath_target.version[1],
1678 multipath_target.version[2]);
1683 static void __exit dm_multipath_exit(void)
1685 destroy_workqueue(kmpath_handlerd);
1686 destroy_workqueue(kmultipathd);
1688 dm_unregister_target(&multipath_target);
1689 kmem_cache_destroy(_mpio_cache);
1692 module_init(dm_multipath_init);
1693 module_exit(dm_multipath_exit);
1695 MODULE_DESCRIPTION(DM_NAME " multipath target");
1696 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1697 MODULE_LICENSE("GPL");