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.
9 #include "dm-path-selector.h"
10 #include "dm-bio-list.h"
11 #include "dm-bio-record.h"
12 #include "dm-uevent.h"
14 #include <linux/ctype.h>
15 #include <linux/init.h>
16 #include <linux/mempool.h>
17 #include <linux/module.h>
18 #include <linux/pagemap.h>
19 #include <linux/slab.h>
20 #include <linux/time.h>
21 #include <linux/workqueue.h>
22 #include <scsi/scsi_dh.h>
23 #include <asm/atomic.h>
25 #define DM_MSG_PREFIX "multipath"
26 #define MESG_STR(x) x, sizeof(x)
30 struct list_head list;
32 struct priority_group *pg; /* Owning PG */
33 unsigned fail_count; /* Cumulative failure count */
38 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
41 * Paths are grouped into Priority Groups and numbered from 1 upwards.
42 * Each has a path selector which controls which path gets used.
44 struct priority_group {
45 struct list_head list;
47 struct multipath *m; /* Owning multipath instance */
48 struct path_selector ps;
50 unsigned pg_num; /* Reference number */
51 unsigned bypassed; /* Temporarily bypass this PG? */
53 unsigned nr_pgpaths; /* Number of paths in PG */
54 struct list_head pgpaths;
57 /* Multipath context */
59 struct list_head list;
64 const char *hw_handler_name;
65 struct work_struct activate_path;
66 struct pgpath *pgpath_to_activate;
67 unsigned nr_priority_groups;
68 struct list_head priority_groups;
69 unsigned pg_init_required; /* pg_init needs calling? */
70 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
72 unsigned nr_valid_paths; /* Total number of usable paths */
73 struct pgpath *current_pgpath;
74 struct priority_group *current_pg;
75 struct priority_group *next_pg; /* Switch to this PG if set */
76 unsigned repeat_count; /* I/Os left before calling PS again */
78 unsigned queue_io; /* Must we queue all I/O? */
79 unsigned queue_if_no_path; /* Queue I/O if last path fails? */
80 unsigned saved_queue_if_no_path;/* Saved state during suspension */
81 unsigned pg_init_retries; /* Number of times to retry pg_init */
82 unsigned pg_init_count; /* Number of times pg_init called */
84 struct work_struct process_queued_ios;
85 struct bio_list queued_ios;
88 struct work_struct trigger_event;
91 * We must use a mempool of dm_mpath_io structs so that we
92 * can resubmit bios on error.
98 * Context information attached to each bio we process.
101 struct pgpath *pgpath;
102 struct dm_bio_details details;
105 typedef int (*action_fn) (struct pgpath *pgpath);
107 #define MIN_IOS 256 /* Mempool size */
109 static struct kmem_cache *_mpio_cache;
111 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
112 static void process_queued_ios(struct work_struct *work);
113 static void trigger_event(struct work_struct *work);
114 static void activate_path(struct work_struct *work);
117 /*-----------------------------------------------
118 * Allocation routines
119 *-----------------------------------------------*/
121 static struct pgpath *alloc_pgpath(void)
123 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
126 pgpath->path.is_active = 1;
131 static void free_pgpath(struct pgpath *pgpath)
136 static struct priority_group *alloc_priority_group(void)
138 struct priority_group *pg;
140 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
143 INIT_LIST_HEAD(&pg->pgpaths);
148 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
151 struct pgpath *pgpath, *tmp;
152 struct multipath *m = ti->private;
154 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
155 list_del(&pgpath->list);
156 if (m->hw_handler_name)
157 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
158 dm_put_device(ti, pgpath->path.dev);
159 spin_lock_irqsave(&m->lock, flags);
160 if (m->pgpath_to_activate == pgpath)
161 m->pgpath_to_activate = NULL;
162 spin_unlock_irqrestore(&m->lock, flags);
167 static void free_priority_group(struct priority_group *pg,
168 struct dm_target *ti)
170 struct path_selector *ps = &pg->ps;
173 ps->type->destroy(ps);
174 dm_put_path_selector(ps->type);
177 free_pgpaths(&pg->pgpaths, ti);
181 static struct multipath *alloc_multipath(struct dm_target *ti)
185 m = kzalloc(sizeof(*m), GFP_KERNEL);
187 INIT_LIST_HEAD(&m->priority_groups);
188 spin_lock_init(&m->lock);
190 INIT_WORK(&m->process_queued_ios, process_queued_ios);
191 INIT_WORK(&m->trigger_event, trigger_event);
192 INIT_WORK(&m->activate_path, activate_path);
193 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
205 static void free_multipath(struct multipath *m)
207 struct priority_group *pg, *tmp;
209 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
211 free_priority_group(pg, m->ti);
214 kfree(m->hw_handler_name);
215 mempool_destroy(m->mpio_pool);
220 /*-----------------------------------------------
222 *-----------------------------------------------*/
224 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
226 m->current_pg = pgpath->pg;
228 /* Must we initialise the PG first, and queue I/O till it's ready? */
229 if (m->hw_handler_name) {
230 m->pg_init_required = 1;
233 m->pg_init_required = 0;
237 m->pg_init_count = 0;
240 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
242 struct dm_path *path;
244 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
248 m->current_pgpath = path_to_pgpath(path);
250 if (m->current_pg != pg)
251 __switch_pg(m, m->current_pgpath);
256 static void __choose_pgpath(struct multipath *m)
258 struct priority_group *pg;
259 unsigned bypassed = 1;
261 if (!m->nr_valid_paths)
264 /* Were we instructed to switch PG? */
268 if (!__choose_path_in_pg(m, pg))
272 /* Don't change PG until it has no remaining paths */
273 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
277 * Loop through priority groups until we find a valid path.
278 * First time we skip PGs marked 'bypassed'.
279 * Second time we only try the ones we skipped.
282 list_for_each_entry(pg, &m->priority_groups, list) {
283 if (pg->bypassed == bypassed)
285 if (!__choose_path_in_pg(m, pg))
288 } while (bypassed--);
291 m->current_pgpath = NULL;
292 m->current_pg = NULL;
296 * Check whether bios must be queued in the device-mapper core rather
297 * than here in the target.
299 * m->lock must be held on entry.
301 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
302 * same value then we are not between multipath_presuspend()
303 * and multipath_resume() calls and we have no need to check
304 * for the DMF_NOFLUSH_SUSPENDING flag.
306 static int __must_push_back(struct multipath *m)
308 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
309 dm_noflush_suspending(m->ti));
312 static int map_io(struct multipath *m, struct bio *bio,
313 struct dm_mpath_io *mpio, unsigned was_queued)
315 int r = DM_MAPIO_REMAPPED;
317 struct pgpath *pgpath;
319 spin_lock_irqsave(&m->lock, flags);
321 /* Do we need to select a new pgpath? */
322 if (!m->current_pgpath ||
323 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
326 pgpath = m->current_pgpath;
331 if ((pgpath && m->queue_io) ||
332 (!pgpath && m->queue_if_no_path)) {
333 /* Queue for the daemon to resubmit */
334 bio_list_add(&m->queued_ios, bio);
336 if ((m->pg_init_required && !m->pg_init_in_progress) ||
338 queue_work(kmultipathd, &m->process_queued_ios);
340 r = DM_MAPIO_SUBMITTED;
342 bio->bi_bdev = pgpath->path.dev->bdev;
343 else if (__must_push_back(m))
344 r = DM_MAPIO_REQUEUE;
346 r = -EIO; /* Failed */
348 mpio->pgpath = pgpath;
350 spin_unlock_irqrestore(&m->lock, flags);
356 * If we run out of usable paths, should we queue I/O or error it?
358 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
359 unsigned save_old_value)
363 spin_lock_irqsave(&m->lock, flags);
366 m->saved_queue_if_no_path = m->queue_if_no_path;
368 m->saved_queue_if_no_path = queue_if_no_path;
369 m->queue_if_no_path = queue_if_no_path;
370 if (!m->queue_if_no_path && m->queue_size)
371 queue_work(kmultipathd, &m->process_queued_ios);
373 spin_unlock_irqrestore(&m->lock, flags);
378 /*-----------------------------------------------------------------
379 * The multipath daemon is responsible for resubmitting queued ios.
380 *---------------------------------------------------------------*/
382 static void dispatch_queued_ios(struct multipath *m)
386 struct bio *bio = NULL, *next;
387 struct dm_mpath_io *mpio;
388 union map_info *info;
390 spin_lock_irqsave(&m->lock, flags);
391 bio = bio_list_get(&m->queued_ios);
392 spin_unlock_irqrestore(&m->lock, flags);
398 info = dm_get_mapinfo(bio);
401 r = map_io(m, bio, mpio, 1);
404 else if (r == DM_MAPIO_REMAPPED)
405 generic_make_request(bio);
406 else if (r == DM_MAPIO_REQUEUE)
407 bio_endio(bio, -EIO);
413 static void process_queued_ios(struct work_struct *work)
415 struct multipath *m =
416 container_of(work, struct multipath, process_queued_ios);
417 struct pgpath *pgpath = NULL;
418 unsigned init_required = 0, must_queue = 1;
421 spin_lock_irqsave(&m->lock, flags);
426 if (!m->current_pgpath)
429 pgpath = m->current_pgpath;
430 m->pgpath_to_activate = m->current_pgpath;
432 if ((pgpath && !m->queue_io) ||
433 (!pgpath && !m->queue_if_no_path))
436 if (m->pg_init_required && !m->pg_init_in_progress) {
438 m->pg_init_required = 0;
439 m->pg_init_in_progress = 1;
444 spin_unlock_irqrestore(&m->lock, flags);
447 queue_work(kmpath_handlerd, &m->activate_path);
450 dispatch_queued_ios(m);
454 * An event is triggered whenever a path is taken out of use.
455 * Includes path failure and PG bypass.
457 static void trigger_event(struct work_struct *work)
459 struct multipath *m =
460 container_of(work, struct multipath, trigger_event);
462 dm_table_event(m->ti->table);
465 /*-----------------------------------------------------------------
466 * Constructor/argument parsing:
467 * <#multipath feature args> [<arg>]*
468 * <#hw_handler args> [hw_handler [<arg>]*]
470 * <initial priority group>
471 * [<selector> <#selector args> [<arg>]*
472 * <#paths> <#per-path selector args>
473 * [<path> [<arg>]* ]+ ]+
474 *---------------------------------------------------------------*/
481 static int read_param(struct param *param, char *str, unsigned *v, char **error)
484 (sscanf(str, "%u", v) != 1) ||
487 *error = param->error;
499 static char *shift(struct arg_set *as)
513 static void consume(struct arg_set *as, unsigned n)
515 BUG_ON (as->argc < n);
520 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
521 struct dm_target *ti)
524 struct path_selector_type *pst;
527 static struct param _params[] = {
528 {0, 1024, "invalid number of path selector args"},
531 pst = dm_get_path_selector(shift(as));
533 ti->error = "unknown path selector type";
537 r = read_param(_params, shift(as), &ps_argc, &ti->error);
539 dm_put_path_selector(pst);
543 r = pst->create(&pg->ps, ps_argc, as->argv);
545 dm_put_path_selector(pst);
546 ti->error = "path selector constructor failed";
551 consume(as, ps_argc);
556 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
557 struct dm_target *ti)
561 struct multipath *m = ti->private;
563 /* we need at least a path arg */
565 ti->error = "no device given";
573 r = dm_get_device(ti, shift(as), ti->begin, ti->len,
574 dm_table_get_mode(ti->table), &p->path.dev);
576 ti->error = "error getting device";
580 if (m->hw_handler_name) {
581 r = scsi_dh_attach(bdev_get_queue(p->path.dev->bdev),
584 dm_put_device(ti, p->path.dev);
589 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
591 dm_put_device(ti, p->path.dev);
602 static struct priority_group *parse_priority_group(struct arg_set *as,
605 static struct param _params[] = {
606 {1, 1024, "invalid number of paths"},
607 {0, 1024, "invalid number of selector args"}
611 unsigned i, nr_selector_args, nr_params;
612 struct priority_group *pg;
613 struct dm_target *ti = m->ti;
617 ti->error = "not enough priority group aruments";
621 pg = alloc_priority_group();
623 ti->error = "couldn't allocate priority group";
628 r = parse_path_selector(as, pg, ti);
635 r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
639 r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
643 nr_params = 1 + nr_selector_args;
644 for (i = 0; i < pg->nr_pgpaths; i++) {
645 struct pgpath *pgpath;
646 struct arg_set path_args;
648 if (as->argc < nr_params) {
649 ti->error = "not enough path parameters";
653 path_args.argc = nr_params;
654 path_args.argv = as->argv;
656 pgpath = parse_path(&path_args, &pg->ps, ti);
661 list_add_tail(&pgpath->list, &pg->pgpaths);
662 consume(as, nr_params);
668 free_priority_group(pg, ti);
672 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
675 struct dm_target *ti = m->ti;
677 static struct param _params[] = {
678 {0, 1024, "invalid number of hardware handler args"},
681 if (read_param(_params, shift(as), &hw_argc, &ti->error))
687 m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
688 request_module("scsi_dh_%s", m->hw_handler_name);
689 if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
690 ti->error = "unknown hardware handler type";
691 kfree(m->hw_handler_name);
692 m->hw_handler_name = NULL;
695 consume(as, hw_argc - 1);
700 static int parse_features(struct arg_set *as, struct multipath *m)
704 struct dm_target *ti = m->ti;
705 const char *param_name;
707 static struct param _params[] = {
708 {0, 3, "invalid number of feature args"},
709 {1, 50, "pg_init_retries must be between 1 and 50"},
712 r = read_param(_params, shift(as), &argc, &ti->error);
720 param_name = shift(as);
723 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
724 r = queue_if_no_path(m, 1, 0);
728 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
730 r = read_param(_params + 1, shift(as),
731 &m->pg_init_retries, &ti->error);
736 ti->error = "Unrecognised multipath feature request";
738 } while (argc && !r);
743 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
746 /* target parameters */
747 static struct param _params[] = {
748 {1, 1024, "invalid number of priority groups"},
749 {1, 1024, "invalid initial priority group number"},
755 unsigned pg_count = 0;
756 unsigned next_pg_num;
761 m = alloc_multipath(ti);
763 ti->error = "can't allocate multipath";
767 r = parse_features(&as, m);
771 r = parse_hw_handler(&as, m);
775 r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
779 r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
783 /* parse the priority groups */
785 struct priority_group *pg;
787 pg = parse_priority_group(&as, m);
793 m->nr_valid_paths += pg->nr_pgpaths;
794 list_add_tail(&pg->list, &m->priority_groups);
796 pg->pg_num = pg_count;
801 if (pg_count != m->nr_priority_groups) {
802 ti->error = "priority group count mismatch";
814 static void multipath_dtr(struct dm_target *ti)
816 struct multipath *m = (struct multipath *) ti->private;
818 flush_workqueue(kmpath_handlerd);
819 flush_workqueue(kmultipathd);
824 * Map bios, recording original fields for later in case we have to resubmit
826 static int multipath_map(struct dm_target *ti, struct bio *bio,
827 union map_info *map_context)
830 struct dm_mpath_io *mpio;
831 struct multipath *m = (struct multipath *) ti->private;
833 mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
834 dm_bio_record(&mpio->details, bio);
836 map_context->ptr = mpio;
837 bio->bi_rw |= (1 << BIO_RW_FAILFAST);
838 r = map_io(m, bio, mpio, 0);
839 if (r < 0 || r == DM_MAPIO_REQUEUE)
840 mempool_free(mpio, m->mpio_pool);
846 * Take a path out of use.
848 static int fail_path(struct pgpath *pgpath)
851 struct multipath *m = pgpath->pg->m;
853 spin_lock_irqsave(&m->lock, flags);
855 if (!pgpath->path.is_active)
858 DMWARN("Failing path %s.", pgpath->path.dev->name);
860 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
861 pgpath->path.is_active = 0;
862 pgpath->fail_count++;
866 if (pgpath == m->current_pgpath)
867 m->current_pgpath = NULL;
869 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
870 pgpath->path.dev->name, m->nr_valid_paths);
872 queue_work(kmultipathd, &m->trigger_event);
875 spin_unlock_irqrestore(&m->lock, flags);
881 * Reinstate a previously-failed path
883 static int reinstate_path(struct pgpath *pgpath)
887 struct multipath *m = pgpath->pg->m;
889 spin_lock_irqsave(&m->lock, flags);
891 if (pgpath->path.is_active)
894 if (!pgpath->pg->ps.type->reinstate_path) {
895 DMWARN("Reinstate path not supported by path selector %s",
896 pgpath->pg->ps.type->name);
901 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
905 pgpath->path.is_active = 1;
907 m->current_pgpath = NULL;
908 if (!m->nr_valid_paths++ && m->queue_size)
909 queue_work(kmultipathd, &m->process_queued_ios);
911 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
912 pgpath->path.dev->name, m->nr_valid_paths);
914 queue_work(kmultipathd, &m->trigger_event);
917 spin_unlock_irqrestore(&m->lock, flags);
923 * Fail or reinstate all paths that match the provided struct dm_dev.
925 static int action_dev(struct multipath *m, struct dm_dev *dev,
929 struct pgpath *pgpath;
930 struct priority_group *pg;
932 list_for_each_entry(pg, &m->priority_groups, list) {
933 list_for_each_entry(pgpath, &pg->pgpaths, list) {
934 if (pgpath->path.dev == dev)
943 * Temporarily try to avoid having to use the specified PG
945 static void bypass_pg(struct multipath *m, struct priority_group *pg,
950 spin_lock_irqsave(&m->lock, flags);
952 pg->bypassed = bypassed;
953 m->current_pgpath = NULL;
954 m->current_pg = NULL;
956 spin_unlock_irqrestore(&m->lock, flags);
958 queue_work(kmultipathd, &m->trigger_event);
962 * Switch to using the specified PG from the next I/O that gets mapped
964 static int switch_pg_num(struct multipath *m, const char *pgstr)
966 struct priority_group *pg;
970 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
971 (pgnum > m->nr_priority_groups)) {
972 DMWARN("invalid PG number supplied to switch_pg_num");
976 spin_lock_irqsave(&m->lock, flags);
977 list_for_each_entry(pg, &m->priority_groups, list) {
982 m->current_pgpath = NULL;
983 m->current_pg = NULL;
986 spin_unlock_irqrestore(&m->lock, flags);
988 queue_work(kmultipathd, &m->trigger_event);
993 * Set/clear bypassed status of a PG.
994 * PGs are numbered upwards from 1 in the order they were declared.
996 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
998 struct priority_group *pg;
1001 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1002 (pgnum > m->nr_priority_groups)) {
1003 DMWARN("invalid PG number supplied to bypass_pg");
1007 list_for_each_entry(pg, &m->priority_groups, list) {
1012 bypass_pg(m, pg, bypassed);
1017 * Should we retry pg_init immediately?
1019 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1021 unsigned long flags;
1022 int limit_reached = 0;
1024 spin_lock_irqsave(&m->lock, flags);
1026 if (m->pg_init_count <= m->pg_init_retries)
1027 m->pg_init_required = 1;
1031 spin_unlock_irqrestore(&m->lock, flags);
1033 return limit_reached;
1036 static void pg_init_done(struct dm_path *path, int errors)
1038 struct pgpath *pgpath = path_to_pgpath(path);
1039 struct priority_group *pg = pgpath->pg;
1040 struct multipath *m = pg->m;
1041 unsigned long flags;
1043 /* device or driver problems */
1048 if (!m->hw_handler_name) {
1052 DMERR("Cannot failover device because scsi_dh_%s was not "
1053 "loaded.", m->hw_handler_name);
1055 * Fail path for now, so we do not ping pong
1059 case SCSI_DH_DEV_TEMP_BUSY:
1061 * Probably doing something like FW upgrade on the
1062 * controller so try the other pg.
1064 bypass_pg(m, pg, 1);
1066 /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1068 case SCSI_DH_IMM_RETRY:
1069 case SCSI_DH_RES_TEMP_UNAVAIL:
1070 if (pg_init_limit_reached(m, pgpath))
1076 * We probably do not want to fail the path for a device
1077 * error, but this is what the old dm did. In future
1078 * patches we can do more advanced handling.
1083 spin_lock_irqsave(&m->lock, flags);
1085 DMERR("Could not failover device. Error %d.", errors);
1086 m->current_pgpath = NULL;
1087 m->current_pg = NULL;
1088 } else if (!m->pg_init_required) {
1093 m->pg_init_in_progress = 0;
1094 queue_work(kmultipathd, &m->process_queued_ios);
1095 spin_unlock_irqrestore(&m->lock, flags);
1098 static void activate_path(struct work_struct *work)
1101 struct multipath *m =
1102 container_of(work, struct multipath, activate_path);
1103 struct dm_path *path;
1104 unsigned long flags;
1106 spin_lock_irqsave(&m->lock, flags);
1107 path = &m->pgpath_to_activate->path;
1108 m->pgpath_to_activate = NULL;
1109 spin_unlock_irqrestore(&m->lock, flags);
1112 ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
1113 pg_init_done(path, ret);
1119 static int do_end_io(struct multipath *m, struct bio *bio,
1120 int error, struct dm_mpath_io *mpio)
1122 unsigned long flags;
1125 return 0; /* I/O complete */
1127 if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1130 if (error == -EOPNOTSUPP)
1133 spin_lock_irqsave(&m->lock, flags);
1134 if (!m->nr_valid_paths) {
1135 if (__must_push_back(m)) {
1136 spin_unlock_irqrestore(&m->lock, flags);
1137 return DM_ENDIO_REQUEUE;
1138 } else if (!m->queue_if_no_path) {
1139 spin_unlock_irqrestore(&m->lock, flags);
1142 spin_unlock_irqrestore(&m->lock, flags);
1146 spin_unlock_irqrestore(&m->lock, flags);
1149 fail_path(mpio->pgpath);
1152 dm_bio_restore(&mpio->details, bio);
1154 /* queue for the daemon to resubmit or fail */
1155 spin_lock_irqsave(&m->lock, flags);
1156 bio_list_add(&m->queued_ios, bio);
1159 queue_work(kmultipathd, &m->process_queued_ios);
1160 spin_unlock_irqrestore(&m->lock, flags);
1162 return DM_ENDIO_INCOMPLETE; /* io not complete */
1165 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1166 int error, union map_info *map_context)
1168 struct multipath *m = ti->private;
1169 struct dm_mpath_io *mpio = map_context->ptr;
1170 struct pgpath *pgpath = mpio->pgpath;
1171 struct path_selector *ps;
1174 r = do_end_io(m, bio, error, mpio);
1176 ps = &pgpath->pg->ps;
1177 if (ps->type->end_io)
1178 ps->type->end_io(ps, &pgpath->path);
1180 if (r != DM_ENDIO_INCOMPLETE)
1181 mempool_free(mpio, m->mpio_pool);
1187 * Suspend can't complete until all the I/O is processed so if
1188 * the last path fails we must error any remaining I/O.
1189 * Note that if the freeze_bdev fails while suspending, the
1190 * queue_if_no_path state is lost - userspace should reset it.
1192 static void multipath_presuspend(struct dm_target *ti)
1194 struct multipath *m = (struct multipath *) ti->private;
1196 queue_if_no_path(m, 0, 1);
1200 * Restore the queue_if_no_path setting.
1202 static void multipath_resume(struct dm_target *ti)
1204 struct multipath *m = (struct multipath *) ti->private;
1205 unsigned long flags;
1207 spin_lock_irqsave(&m->lock, flags);
1208 m->queue_if_no_path = m->saved_queue_if_no_path;
1209 spin_unlock_irqrestore(&m->lock, flags);
1213 * Info output has the following format:
1214 * num_multipath_feature_args [multipath_feature_args]*
1215 * num_handler_status_args [handler_status_args]*
1216 * num_groups init_group_number
1217 * [A|D|E num_ps_status_args [ps_status_args]*
1218 * num_paths num_selector_args
1219 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1221 * Table output has the following format (identical to the constructor string):
1222 * num_feature_args [features_args]*
1223 * num_handler_args hw_handler [hw_handler_args]*
1224 * num_groups init_group_number
1225 * [priority selector-name num_ps_args [ps_args]*
1226 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1228 static int multipath_status(struct dm_target *ti, status_type_t type,
1229 char *result, unsigned int maxlen)
1232 unsigned long flags;
1233 struct multipath *m = (struct multipath *) ti->private;
1234 struct priority_group *pg;
1239 spin_lock_irqsave(&m->lock, flags);
1242 if (type == STATUSTYPE_INFO)
1243 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1245 DMEMIT("%u ", m->queue_if_no_path +
1246 (m->pg_init_retries > 0) * 2);
1247 if (m->queue_if_no_path)
1248 DMEMIT("queue_if_no_path ");
1249 if (m->pg_init_retries)
1250 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1253 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1256 DMEMIT("1 %s ", m->hw_handler_name);
1258 DMEMIT("%u ", m->nr_priority_groups);
1261 pg_num = m->next_pg->pg_num;
1262 else if (m->current_pg)
1263 pg_num = m->current_pg->pg_num;
1267 DMEMIT("%u ", pg_num);
1270 case STATUSTYPE_INFO:
1271 list_for_each_entry(pg, &m->priority_groups, list) {
1273 state = 'D'; /* Disabled */
1274 else if (pg == m->current_pg)
1275 state = 'A'; /* Currently Active */
1277 state = 'E'; /* Enabled */
1279 DMEMIT("%c ", state);
1281 if (pg->ps.type->status)
1282 sz += pg->ps.type->status(&pg->ps, NULL, type,
1288 DMEMIT("%u %u ", pg->nr_pgpaths,
1289 pg->ps.type->info_args);
1291 list_for_each_entry(p, &pg->pgpaths, list) {
1292 DMEMIT("%s %s %u ", p->path.dev->name,
1293 p->path.is_active ? "A" : "F",
1295 if (pg->ps.type->status)
1296 sz += pg->ps.type->status(&pg->ps,
1297 &p->path, type, result + sz,
1303 case STATUSTYPE_TABLE:
1304 list_for_each_entry(pg, &m->priority_groups, list) {
1305 DMEMIT("%s ", pg->ps.type->name);
1307 if (pg->ps.type->status)
1308 sz += pg->ps.type->status(&pg->ps, NULL, type,
1314 DMEMIT("%u %u ", pg->nr_pgpaths,
1315 pg->ps.type->table_args);
1317 list_for_each_entry(p, &pg->pgpaths, list) {
1318 DMEMIT("%s ", p->path.dev->name);
1319 if (pg->ps.type->status)
1320 sz += pg->ps.type->status(&pg->ps,
1321 &p->path, type, result + sz,
1328 spin_unlock_irqrestore(&m->lock, flags);
1333 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1337 struct multipath *m = (struct multipath *) ti->private;
1341 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1342 return queue_if_no_path(m, 1, 0);
1343 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1344 return queue_if_no_path(m, 0, 0);
1350 if (!strnicmp(argv[0], MESG_STR("disable_group")))
1351 return bypass_pg_num(m, argv[1], 1);
1352 else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1353 return bypass_pg_num(m, argv[1], 0);
1354 else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1355 return switch_pg_num(m, argv[1]);
1356 else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1357 action = reinstate_path;
1358 else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1363 r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1364 dm_table_get_mode(ti->table), &dev);
1366 DMWARN("message: error getting device %s",
1371 r = action_dev(m, dev, action);
1373 dm_put_device(ti, dev);
1378 DMWARN("Unrecognised multipath message received.");
1382 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1383 struct file *filp, unsigned int cmd,
1386 struct multipath *m = (struct multipath *) ti->private;
1387 struct block_device *bdev = NULL;
1388 unsigned long flags;
1389 struct file fake_file = {};
1390 struct dentry fake_dentry = {};
1393 fake_file.f_path.dentry = &fake_dentry;
1395 spin_lock_irqsave(&m->lock, flags);
1397 if (!m->current_pgpath)
1400 if (m->current_pgpath) {
1401 bdev = m->current_pgpath->path.dev->bdev;
1402 fake_dentry.d_inode = bdev->bd_inode;
1403 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1411 spin_unlock_irqrestore(&m->lock, flags);
1413 return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1414 bdev->bd_disk, cmd, arg);
1417 /*-----------------------------------------------------------------
1419 *---------------------------------------------------------------*/
1420 static struct target_type multipath_target = {
1421 .name = "multipath",
1422 .version = {1, 0, 5},
1423 .module = THIS_MODULE,
1424 .ctr = multipath_ctr,
1425 .dtr = multipath_dtr,
1426 .map = multipath_map,
1427 .end_io = multipath_end_io,
1428 .presuspend = multipath_presuspend,
1429 .resume = multipath_resume,
1430 .status = multipath_status,
1431 .message = multipath_message,
1432 .ioctl = multipath_ioctl,
1435 static int __init dm_multipath_init(void)
1439 /* allocate a slab for the dm_ios */
1440 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1444 r = dm_register_target(&multipath_target);
1446 DMERR("register failed %d", r);
1447 kmem_cache_destroy(_mpio_cache);
1451 kmultipathd = create_workqueue("kmpathd");
1453 DMERR("failed to create workqueue kmpathd");
1454 dm_unregister_target(&multipath_target);
1455 kmem_cache_destroy(_mpio_cache);
1460 * A separate workqueue is used to handle the device handlers
1461 * to avoid overloading existing workqueue. Overloading the
1462 * old workqueue would also create a bottleneck in the
1463 * path of the storage hardware device activation.
1465 kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1466 if (!kmpath_handlerd) {
1467 DMERR("failed to create workqueue kmpath_handlerd");
1468 destroy_workqueue(kmultipathd);
1469 dm_unregister_target(&multipath_target);
1470 kmem_cache_destroy(_mpio_cache);
1474 DMINFO("version %u.%u.%u loaded",
1475 multipath_target.version[0], multipath_target.version[1],
1476 multipath_target.version[2]);
1481 static void __exit dm_multipath_exit(void)
1485 destroy_workqueue(kmpath_handlerd);
1486 destroy_workqueue(kmultipathd);
1488 r = dm_unregister_target(&multipath_target);
1490 DMERR("target unregister failed %d", r);
1491 kmem_cache_destroy(_mpio_cache);
1494 module_init(dm_multipath_init);
1495 module_exit(dm_multipath_exit);
1497 MODULE_DESCRIPTION(DM_NAME " multipath target");
1498 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1499 MODULE_LICENSE("GPL");