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";
566 return ERR_PTR(-EINVAL);
571 return ERR_PTR(-ENOMEM);
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 arguments";
618 return ERR_PTR(-EINVAL);
621 pg = alloc_priority_group();
623 ti->error = "couldn't allocate priority group";
624 return ERR_PTR(-ENOMEM);
628 r = parse_path_selector(as, pg, ti);
635 r = 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);
657 if (IS_ERR(pgpath)) {
663 list_add_tail(&pgpath->list, &pg->pgpaths);
664 consume(as, nr_params);
670 free_priority_group(pg, ti);
674 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
677 struct dm_target *ti = m->ti;
679 static struct param _params[] = {
680 {0, 1024, "invalid number of hardware handler args"},
683 if (read_param(_params, shift(as), &hw_argc, &ti->error))
689 m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
690 request_module("scsi_dh_%s", m->hw_handler_name);
691 if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
692 ti->error = "unknown hardware handler type";
693 kfree(m->hw_handler_name);
694 m->hw_handler_name = NULL;
697 consume(as, hw_argc - 1);
702 static int parse_features(struct arg_set *as, struct multipath *m)
706 struct dm_target *ti = m->ti;
707 const char *param_name;
709 static struct param _params[] = {
710 {0, 3, "invalid number of feature args"},
711 {1, 50, "pg_init_retries must be between 1 and 50"},
714 r = read_param(_params, shift(as), &argc, &ti->error);
722 param_name = shift(as);
725 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
726 r = queue_if_no_path(m, 1, 0);
730 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
732 r = read_param(_params + 1, shift(as),
733 &m->pg_init_retries, &ti->error);
738 ti->error = "Unrecognised multipath feature request";
740 } while (argc && !r);
745 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
748 /* target parameters */
749 static struct param _params[] = {
750 {1, 1024, "invalid number of priority groups"},
751 {1, 1024, "invalid initial priority group number"},
757 unsigned pg_count = 0;
758 unsigned next_pg_num;
763 m = alloc_multipath(ti);
765 ti->error = "can't allocate multipath";
769 r = parse_features(&as, m);
773 r = parse_hw_handler(&as, m);
777 r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
781 r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
785 /* parse the priority groups */
787 struct priority_group *pg;
789 pg = parse_priority_group(&as, m);
795 m->nr_valid_paths += pg->nr_pgpaths;
796 list_add_tail(&pg->list, &m->priority_groups);
798 pg->pg_num = pg_count;
803 if (pg_count != m->nr_priority_groups) {
804 ti->error = "priority group count mismatch";
816 static void multipath_dtr(struct dm_target *ti)
818 struct multipath *m = (struct multipath *) ti->private;
820 flush_workqueue(kmpath_handlerd);
821 flush_workqueue(kmultipathd);
826 * Map bios, recording original fields for later in case we have to resubmit
828 static int multipath_map(struct dm_target *ti, struct bio *bio,
829 union map_info *map_context)
832 struct dm_mpath_io *mpio;
833 struct multipath *m = (struct multipath *) ti->private;
835 mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
836 dm_bio_record(&mpio->details, bio);
838 map_context->ptr = mpio;
839 bio->bi_rw |= (1 << BIO_RW_FAILFAST);
840 r = map_io(m, bio, mpio, 0);
841 if (r < 0 || r == DM_MAPIO_REQUEUE)
842 mempool_free(mpio, m->mpio_pool);
848 * Take a path out of use.
850 static int fail_path(struct pgpath *pgpath)
853 struct multipath *m = pgpath->pg->m;
855 spin_lock_irqsave(&m->lock, flags);
857 if (!pgpath->path.is_active)
860 DMWARN("Failing path %s.", pgpath->path.dev->name);
862 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
863 pgpath->path.is_active = 0;
864 pgpath->fail_count++;
868 if (pgpath == m->current_pgpath)
869 m->current_pgpath = NULL;
871 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
872 pgpath->path.dev->name, m->nr_valid_paths);
874 queue_work(kmultipathd, &m->trigger_event);
877 spin_unlock_irqrestore(&m->lock, flags);
883 * Reinstate a previously-failed path
885 static int reinstate_path(struct pgpath *pgpath)
889 struct multipath *m = pgpath->pg->m;
891 spin_lock_irqsave(&m->lock, flags);
893 if (pgpath->path.is_active)
896 if (!pgpath->pg->ps.type->reinstate_path) {
897 DMWARN("Reinstate path not supported by path selector %s",
898 pgpath->pg->ps.type->name);
903 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
907 pgpath->path.is_active = 1;
909 m->current_pgpath = NULL;
910 if (!m->nr_valid_paths++ && m->queue_size)
911 queue_work(kmultipathd, &m->process_queued_ios);
913 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
914 pgpath->path.dev->name, m->nr_valid_paths);
916 queue_work(kmultipathd, &m->trigger_event);
919 spin_unlock_irqrestore(&m->lock, flags);
925 * Fail or reinstate all paths that match the provided struct dm_dev.
927 static int action_dev(struct multipath *m, struct dm_dev *dev,
931 struct pgpath *pgpath;
932 struct priority_group *pg;
934 list_for_each_entry(pg, &m->priority_groups, list) {
935 list_for_each_entry(pgpath, &pg->pgpaths, list) {
936 if (pgpath->path.dev == dev)
945 * Temporarily try to avoid having to use the specified PG
947 static void bypass_pg(struct multipath *m, struct priority_group *pg,
952 spin_lock_irqsave(&m->lock, flags);
954 pg->bypassed = bypassed;
955 m->current_pgpath = NULL;
956 m->current_pg = NULL;
958 spin_unlock_irqrestore(&m->lock, flags);
960 queue_work(kmultipathd, &m->trigger_event);
964 * Switch to using the specified PG from the next I/O that gets mapped
966 static int switch_pg_num(struct multipath *m, const char *pgstr)
968 struct priority_group *pg;
972 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
973 (pgnum > m->nr_priority_groups)) {
974 DMWARN("invalid PG number supplied to switch_pg_num");
978 spin_lock_irqsave(&m->lock, flags);
979 list_for_each_entry(pg, &m->priority_groups, list) {
984 m->current_pgpath = NULL;
985 m->current_pg = NULL;
988 spin_unlock_irqrestore(&m->lock, flags);
990 queue_work(kmultipathd, &m->trigger_event);
995 * Set/clear bypassed status of a PG.
996 * PGs are numbered upwards from 1 in the order they were declared.
998 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1000 struct priority_group *pg;
1003 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1004 (pgnum > m->nr_priority_groups)) {
1005 DMWARN("invalid PG number supplied to bypass_pg");
1009 list_for_each_entry(pg, &m->priority_groups, list) {
1014 bypass_pg(m, pg, bypassed);
1019 * Should we retry pg_init immediately?
1021 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1023 unsigned long flags;
1024 int limit_reached = 0;
1026 spin_lock_irqsave(&m->lock, flags);
1028 if (m->pg_init_count <= m->pg_init_retries)
1029 m->pg_init_required = 1;
1033 spin_unlock_irqrestore(&m->lock, flags);
1035 return limit_reached;
1038 static void pg_init_done(struct dm_path *path, int errors)
1040 struct pgpath *pgpath = path_to_pgpath(path);
1041 struct priority_group *pg = pgpath->pg;
1042 struct multipath *m = pg->m;
1043 unsigned long flags;
1045 /* device or driver problems */
1050 if (!m->hw_handler_name) {
1054 DMERR("Cannot failover device because scsi_dh_%s was not "
1055 "loaded.", m->hw_handler_name);
1057 * Fail path for now, so we do not ping pong
1061 case SCSI_DH_DEV_TEMP_BUSY:
1063 * Probably doing something like FW upgrade on the
1064 * controller so try the other pg.
1066 bypass_pg(m, pg, 1);
1068 /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1070 case SCSI_DH_IMM_RETRY:
1071 case SCSI_DH_RES_TEMP_UNAVAIL:
1072 if (pg_init_limit_reached(m, pgpath))
1078 * We probably do not want to fail the path for a device
1079 * error, but this is what the old dm did. In future
1080 * patches we can do more advanced handling.
1085 spin_lock_irqsave(&m->lock, flags);
1087 DMERR("Could not failover device. Error %d.", errors);
1088 m->current_pgpath = NULL;
1089 m->current_pg = NULL;
1090 } else if (!m->pg_init_required) {
1095 m->pg_init_in_progress = 0;
1096 queue_work(kmultipathd, &m->process_queued_ios);
1097 spin_unlock_irqrestore(&m->lock, flags);
1100 static void activate_path(struct work_struct *work)
1103 struct multipath *m =
1104 container_of(work, struct multipath, activate_path);
1105 struct dm_path *path;
1106 unsigned long flags;
1108 spin_lock_irqsave(&m->lock, flags);
1109 path = &m->pgpath_to_activate->path;
1110 m->pgpath_to_activate = NULL;
1111 spin_unlock_irqrestore(&m->lock, flags);
1114 ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
1115 pg_init_done(path, ret);
1121 static int do_end_io(struct multipath *m, struct bio *bio,
1122 int error, struct dm_mpath_io *mpio)
1124 unsigned long flags;
1127 return 0; /* I/O complete */
1129 if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1132 if (error == -EOPNOTSUPP)
1135 spin_lock_irqsave(&m->lock, flags);
1136 if (!m->nr_valid_paths) {
1137 if (__must_push_back(m)) {
1138 spin_unlock_irqrestore(&m->lock, flags);
1139 return DM_ENDIO_REQUEUE;
1140 } else if (!m->queue_if_no_path) {
1141 spin_unlock_irqrestore(&m->lock, flags);
1144 spin_unlock_irqrestore(&m->lock, flags);
1148 spin_unlock_irqrestore(&m->lock, flags);
1151 fail_path(mpio->pgpath);
1154 dm_bio_restore(&mpio->details, bio);
1156 /* queue for the daemon to resubmit or fail */
1157 spin_lock_irqsave(&m->lock, flags);
1158 bio_list_add(&m->queued_ios, bio);
1161 queue_work(kmultipathd, &m->process_queued_ios);
1162 spin_unlock_irqrestore(&m->lock, flags);
1164 return DM_ENDIO_INCOMPLETE; /* io not complete */
1167 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1168 int error, union map_info *map_context)
1170 struct multipath *m = ti->private;
1171 struct dm_mpath_io *mpio = map_context->ptr;
1172 struct pgpath *pgpath = mpio->pgpath;
1173 struct path_selector *ps;
1176 r = do_end_io(m, bio, error, mpio);
1178 ps = &pgpath->pg->ps;
1179 if (ps->type->end_io)
1180 ps->type->end_io(ps, &pgpath->path);
1182 if (r != DM_ENDIO_INCOMPLETE)
1183 mempool_free(mpio, m->mpio_pool);
1189 * Suspend can't complete until all the I/O is processed so if
1190 * the last path fails we must error any remaining I/O.
1191 * Note that if the freeze_bdev fails while suspending, the
1192 * queue_if_no_path state is lost - userspace should reset it.
1194 static void multipath_presuspend(struct dm_target *ti)
1196 struct multipath *m = (struct multipath *) ti->private;
1198 queue_if_no_path(m, 0, 1);
1202 * Restore the queue_if_no_path setting.
1204 static void multipath_resume(struct dm_target *ti)
1206 struct multipath *m = (struct multipath *) ti->private;
1207 unsigned long flags;
1209 spin_lock_irqsave(&m->lock, flags);
1210 m->queue_if_no_path = m->saved_queue_if_no_path;
1211 spin_unlock_irqrestore(&m->lock, flags);
1215 * Info output has the following format:
1216 * num_multipath_feature_args [multipath_feature_args]*
1217 * num_handler_status_args [handler_status_args]*
1218 * num_groups init_group_number
1219 * [A|D|E num_ps_status_args [ps_status_args]*
1220 * num_paths num_selector_args
1221 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1223 * Table output has the following format (identical to the constructor string):
1224 * num_feature_args [features_args]*
1225 * num_handler_args hw_handler [hw_handler_args]*
1226 * num_groups init_group_number
1227 * [priority selector-name num_ps_args [ps_args]*
1228 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1230 static int multipath_status(struct dm_target *ti, status_type_t type,
1231 char *result, unsigned int maxlen)
1234 unsigned long flags;
1235 struct multipath *m = (struct multipath *) ti->private;
1236 struct priority_group *pg;
1241 spin_lock_irqsave(&m->lock, flags);
1244 if (type == STATUSTYPE_INFO)
1245 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1247 DMEMIT("%u ", m->queue_if_no_path +
1248 (m->pg_init_retries > 0) * 2);
1249 if (m->queue_if_no_path)
1250 DMEMIT("queue_if_no_path ");
1251 if (m->pg_init_retries)
1252 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1255 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1258 DMEMIT("1 %s ", m->hw_handler_name);
1260 DMEMIT("%u ", m->nr_priority_groups);
1263 pg_num = m->next_pg->pg_num;
1264 else if (m->current_pg)
1265 pg_num = m->current_pg->pg_num;
1269 DMEMIT("%u ", pg_num);
1272 case STATUSTYPE_INFO:
1273 list_for_each_entry(pg, &m->priority_groups, list) {
1275 state = 'D'; /* Disabled */
1276 else if (pg == m->current_pg)
1277 state = 'A'; /* Currently Active */
1279 state = 'E'; /* Enabled */
1281 DMEMIT("%c ", state);
1283 if (pg->ps.type->status)
1284 sz += pg->ps.type->status(&pg->ps, NULL, type,
1290 DMEMIT("%u %u ", pg->nr_pgpaths,
1291 pg->ps.type->info_args);
1293 list_for_each_entry(p, &pg->pgpaths, list) {
1294 DMEMIT("%s %s %u ", p->path.dev->name,
1295 p->path.is_active ? "A" : "F",
1297 if (pg->ps.type->status)
1298 sz += pg->ps.type->status(&pg->ps,
1299 &p->path, type, result + sz,
1305 case STATUSTYPE_TABLE:
1306 list_for_each_entry(pg, &m->priority_groups, list) {
1307 DMEMIT("%s ", pg->ps.type->name);
1309 if (pg->ps.type->status)
1310 sz += pg->ps.type->status(&pg->ps, NULL, type,
1316 DMEMIT("%u %u ", pg->nr_pgpaths,
1317 pg->ps.type->table_args);
1319 list_for_each_entry(p, &pg->pgpaths, list) {
1320 DMEMIT("%s ", p->path.dev->name);
1321 if (pg->ps.type->status)
1322 sz += pg->ps.type->status(&pg->ps,
1323 &p->path, type, result + sz,
1330 spin_unlock_irqrestore(&m->lock, flags);
1335 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1339 struct multipath *m = (struct multipath *) ti->private;
1343 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1344 return queue_if_no_path(m, 1, 0);
1345 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1346 return queue_if_no_path(m, 0, 0);
1352 if (!strnicmp(argv[0], MESG_STR("disable_group")))
1353 return bypass_pg_num(m, argv[1], 1);
1354 else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1355 return bypass_pg_num(m, argv[1], 0);
1356 else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1357 return switch_pg_num(m, argv[1]);
1358 else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1359 action = reinstate_path;
1360 else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1365 r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1366 dm_table_get_mode(ti->table), &dev);
1368 DMWARN("message: error getting device %s",
1373 r = action_dev(m, dev, action);
1375 dm_put_device(ti, dev);
1380 DMWARN("Unrecognised multipath message received.");
1384 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1385 struct file *filp, unsigned int cmd,
1388 struct multipath *m = (struct multipath *) ti->private;
1389 struct block_device *bdev = NULL;
1390 unsigned long flags;
1391 struct file fake_file = {};
1392 struct dentry fake_dentry = {};
1395 fake_file.f_path.dentry = &fake_dentry;
1397 spin_lock_irqsave(&m->lock, flags);
1399 if (!m->current_pgpath)
1402 if (m->current_pgpath) {
1403 bdev = m->current_pgpath->path.dev->bdev;
1404 fake_dentry.d_inode = bdev->bd_inode;
1405 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1413 spin_unlock_irqrestore(&m->lock, flags);
1415 return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1416 bdev->bd_disk, cmd, arg);
1419 /*-----------------------------------------------------------------
1421 *---------------------------------------------------------------*/
1422 static struct target_type multipath_target = {
1423 .name = "multipath",
1424 .version = {1, 0, 5},
1425 .module = THIS_MODULE,
1426 .ctr = multipath_ctr,
1427 .dtr = multipath_dtr,
1428 .map = multipath_map,
1429 .end_io = multipath_end_io,
1430 .presuspend = multipath_presuspend,
1431 .resume = multipath_resume,
1432 .status = multipath_status,
1433 .message = multipath_message,
1434 .ioctl = multipath_ioctl,
1437 static int __init dm_multipath_init(void)
1441 /* allocate a slab for the dm_ios */
1442 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1446 r = dm_register_target(&multipath_target);
1448 DMERR("register failed %d", r);
1449 kmem_cache_destroy(_mpio_cache);
1453 kmultipathd = create_workqueue("kmpathd");
1455 DMERR("failed to create workqueue kmpathd");
1456 dm_unregister_target(&multipath_target);
1457 kmem_cache_destroy(_mpio_cache);
1462 * A separate workqueue is used to handle the device handlers
1463 * to avoid overloading existing workqueue. Overloading the
1464 * old workqueue would also create a bottleneck in the
1465 * path of the storage hardware device activation.
1467 kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1468 if (!kmpath_handlerd) {
1469 DMERR("failed to create workqueue kmpath_handlerd");
1470 destroy_workqueue(kmultipathd);
1471 dm_unregister_target(&multipath_target);
1472 kmem_cache_destroy(_mpio_cache);
1476 DMINFO("version %u.%u.%u loaded",
1477 multipath_target.version[0], multipath_target.version[1],
1478 multipath_target.version[2]);
1483 static void __exit dm_multipath_exit(void)
1487 destroy_workqueue(kmpath_handlerd);
1488 destroy_workqueue(kmultipathd);
1490 r = dm_unregister_target(&multipath_target);
1492 DMERR("target unregister failed %d", r);
1493 kmem_cache_destroy(_mpio_cache);
1496 module_init(dm_multipath_init);
1497 module_exit(dm_multipath_exit);
1499 MODULE_DESCRIPTION(DM_NAME " multipath target");
1500 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1501 MODULE_LICENSE("GPL");