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-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 is_active; /* Path status */
34 unsigned fail_count; /* Cumulative failure count */
37 struct work_struct deactivate_path;
38 struct work_struct activate_path;
41 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
44 * Paths are grouped into Priority Groups and numbered from 1 upwards.
45 * Each has a path selector which controls which path gets used.
47 struct priority_group {
48 struct list_head list;
50 struct multipath *m; /* Owning multipath instance */
51 struct path_selector ps;
53 unsigned pg_num; /* Reference number */
54 unsigned bypassed; /* Temporarily bypass this PG? */
56 unsigned nr_pgpaths; /* Number of paths in PG */
57 struct list_head pgpaths;
60 /* Multipath context */
62 struct list_head list;
67 const char *hw_handler_name;
68 unsigned nr_priority_groups;
69 struct list_head priority_groups;
70 unsigned pg_init_required; /* pg_init needs calling? */
71 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
73 unsigned nr_valid_paths; /* Total number of usable paths */
74 struct pgpath *current_pgpath;
75 struct priority_group *current_pg;
76 struct priority_group *next_pg; /* Switch to this PG if set */
77 unsigned repeat_count; /* I/Os left before calling PS again */
79 unsigned queue_io; /* Must we queue all I/O? */
80 unsigned queue_if_no_path; /* Queue I/O if last path fails? */
81 unsigned saved_queue_if_no_path;/* Saved state during suspension */
82 unsigned pg_init_retries; /* Number of times to retry pg_init */
83 unsigned pg_init_count; /* Number of times pg_init called */
85 struct work_struct process_queued_ios;
86 struct bio_list queued_ios;
89 struct work_struct trigger_event;
92 * We must use a mempool of dm_mpath_io structs so that we
93 * can resubmit bios on error.
99 * Context information attached to each bio we process.
102 struct pgpath *pgpath;
103 struct dm_bio_details details;
106 typedef int (*action_fn) (struct pgpath *pgpath);
108 #define MIN_IOS 256 /* Mempool size */
110 static struct kmem_cache *_mpio_cache;
112 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
113 static void process_queued_ios(struct work_struct *work);
114 static void trigger_event(struct work_struct *work);
115 static void activate_path(struct work_struct *work);
116 static void deactivate_path(struct work_struct *work);
119 /*-----------------------------------------------
120 * Allocation routines
121 *-----------------------------------------------*/
123 static struct pgpath *alloc_pgpath(void)
125 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
128 pgpath->is_active = 1;
129 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
130 INIT_WORK(&pgpath->activate_path, activate_path);
136 static void free_pgpath(struct pgpath *pgpath)
141 static void deactivate_path(struct work_struct *work)
143 struct pgpath *pgpath =
144 container_of(work, struct pgpath, deactivate_path);
146 blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
149 static struct priority_group *alloc_priority_group(void)
151 struct priority_group *pg;
153 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
156 INIT_LIST_HEAD(&pg->pgpaths);
161 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
163 struct pgpath *pgpath, *tmp;
164 struct multipath *m = ti->private;
166 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
167 list_del(&pgpath->list);
168 if (m->hw_handler_name)
169 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
170 dm_put_device(ti, pgpath->path.dev);
175 static void free_priority_group(struct priority_group *pg,
176 struct dm_target *ti)
178 struct path_selector *ps = &pg->ps;
181 ps->type->destroy(ps);
182 dm_put_path_selector(ps->type);
185 free_pgpaths(&pg->pgpaths, ti);
189 static struct multipath *alloc_multipath(struct dm_target *ti)
193 m = kzalloc(sizeof(*m), GFP_KERNEL);
195 INIT_LIST_HEAD(&m->priority_groups);
196 spin_lock_init(&m->lock);
198 INIT_WORK(&m->process_queued_ios, process_queued_ios);
199 INIT_WORK(&m->trigger_event, trigger_event);
200 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
212 static void free_multipath(struct multipath *m)
214 struct priority_group *pg, *tmp;
216 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
218 free_priority_group(pg, m->ti);
221 kfree(m->hw_handler_name);
222 mempool_destroy(m->mpio_pool);
227 /*-----------------------------------------------
229 *-----------------------------------------------*/
231 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
233 m->current_pg = pgpath->pg;
235 /* Must we initialise the PG first, and queue I/O till it's ready? */
236 if (m->hw_handler_name) {
237 m->pg_init_required = 1;
240 m->pg_init_required = 0;
244 m->pg_init_count = 0;
247 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
249 struct dm_path *path;
251 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
255 m->current_pgpath = path_to_pgpath(path);
257 if (m->current_pg != pg)
258 __switch_pg(m, m->current_pgpath);
263 static void __choose_pgpath(struct multipath *m)
265 struct priority_group *pg;
266 unsigned bypassed = 1;
268 if (!m->nr_valid_paths)
271 /* Were we instructed to switch PG? */
275 if (!__choose_path_in_pg(m, pg))
279 /* Don't change PG until it has no remaining paths */
280 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
284 * Loop through priority groups until we find a valid path.
285 * First time we skip PGs marked 'bypassed'.
286 * Second time we only try the ones we skipped.
289 list_for_each_entry(pg, &m->priority_groups, list) {
290 if (pg->bypassed == bypassed)
292 if (!__choose_path_in_pg(m, pg))
295 } while (bypassed--);
298 m->current_pgpath = NULL;
299 m->current_pg = NULL;
303 * Check whether bios must be queued in the device-mapper core rather
304 * than here in the target.
306 * m->lock must be held on entry.
308 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
309 * same value then we are not between multipath_presuspend()
310 * and multipath_resume() calls and we have no need to check
311 * for the DMF_NOFLUSH_SUSPENDING flag.
313 static int __must_push_back(struct multipath *m)
315 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
316 dm_noflush_suspending(m->ti));
319 static int map_io(struct multipath *m, struct bio *bio,
320 struct dm_mpath_io *mpio, unsigned was_queued)
322 int r = DM_MAPIO_REMAPPED;
324 struct pgpath *pgpath;
326 spin_lock_irqsave(&m->lock, flags);
328 /* Do we need to select a new pgpath? */
329 if (!m->current_pgpath ||
330 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
333 pgpath = m->current_pgpath;
338 if ((pgpath && m->queue_io) ||
339 (!pgpath && m->queue_if_no_path)) {
340 /* Queue for the daemon to resubmit */
341 bio_list_add(&m->queued_ios, bio);
343 if ((m->pg_init_required && !m->pg_init_in_progress) ||
345 queue_work(kmultipathd, &m->process_queued_ios);
347 r = DM_MAPIO_SUBMITTED;
349 bio->bi_bdev = pgpath->path.dev->bdev;
350 else if (__must_push_back(m))
351 r = DM_MAPIO_REQUEUE;
353 r = -EIO; /* Failed */
355 mpio->pgpath = pgpath;
357 spin_unlock_irqrestore(&m->lock, flags);
363 * If we run out of usable paths, should we queue I/O or error it?
365 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
366 unsigned save_old_value)
370 spin_lock_irqsave(&m->lock, flags);
373 m->saved_queue_if_no_path = m->queue_if_no_path;
375 m->saved_queue_if_no_path = queue_if_no_path;
376 m->queue_if_no_path = queue_if_no_path;
377 if (!m->queue_if_no_path && m->queue_size)
378 queue_work(kmultipathd, &m->process_queued_ios);
380 spin_unlock_irqrestore(&m->lock, flags);
385 /*-----------------------------------------------------------------
386 * The multipath daemon is responsible for resubmitting queued ios.
387 *---------------------------------------------------------------*/
389 static void dispatch_queued_ios(struct multipath *m)
393 struct bio *bio = NULL, *next;
394 struct dm_mpath_io *mpio;
395 union map_info *info;
397 spin_lock_irqsave(&m->lock, flags);
398 bio = bio_list_get(&m->queued_ios);
399 spin_unlock_irqrestore(&m->lock, flags);
405 info = dm_get_mapinfo(bio);
408 r = map_io(m, bio, mpio, 1);
411 else if (r == DM_MAPIO_REMAPPED)
412 generic_make_request(bio);
413 else if (r == DM_MAPIO_REQUEUE)
414 bio_endio(bio, -EIO);
420 static void process_queued_ios(struct work_struct *work)
422 struct multipath *m =
423 container_of(work, struct multipath, process_queued_ios);
424 struct pgpath *pgpath = NULL, *tmp;
425 unsigned must_queue = 1;
428 spin_lock_irqsave(&m->lock, flags);
433 if (!m->current_pgpath)
436 pgpath = m->current_pgpath;
438 if ((pgpath && !m->queue_io) ||
439 (!pgpath && !m->queue_if_no_path))
442 if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
444 m->pg_init_required = 0;
445 list_for_each_entry(tmp, &pgpath->pg->pgpaths, list) {
446 if (queue_work(kmpath_handlerd, &tmp->activate_path))
447 m->pg_init_in_progress++;
451 spin_unlock_irqrestore(&m->lock, flags);
453 dispatch_queued_ios(m);
457 * An event is triggered whenever a path is taken out of use.
458 * Includes path failure and PG bypass.
460 static void trigger_event(struct work_struct *work)
462 struct multipath *m =
463 container_of(work, struct multipath, trigger_event);
465 dm_table_event(m->ti->table);
468 /*-----------------------------------------------------------------
469 * Constructor/argument parsing:
470 * <#multipath feature args> [<arg>]*
471 * <#hw_handler args> [hw_handler [<arg>]*]
473 * <initial priority group>
474 * [<selector> <#selector args> [<arg>]*
475 * <#paths> <#per-path selector args>
476 * [<path> [<arg>]* ]+ ]+
477 *---------------------------------------------------------------*/
484 static int read_param(struct param *param, char *str, unsigned *v, char **error)
487 (sscanf(str, "%u", v) != 1) ||
490 *error = param->error;
502 static char *shift(struct arg_set *as)
516 static void consume(struct arg_set *as, unsigned n)
518 BUG_ON (as->argc < n);
523 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
524 struct dm_target *ti)
527 struct path_selector_type *pst;
530 static struct param _params[] = {
531 {0, 1024, "invalid number of path selector args"},
534 pst = dm_get_path_selector(shift(as));
536 ti->error = "unknown path selector type";
540 r = read_param(_params, shift(as), &ps_argc, &ti->error);
542 dm_put_path_selector(pst);
546 if (ps_argc > as->argc) {
547 dm_put_path_selector(pst);
548 ti->error = "not enough arguments for path selector";
552 r = pst->create(&pg->ps, ps_argc, as->argv);
554 dm_put_path_selector(pst);
555 ti->error = "path selector constructor failed";
560 consume(as, ps_argc);
565 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
566 struct dm_target *ti)
570 struct multipath *m = ti->private;
572 /* we need at least a path arg */
574 ti->error = "no device given";
575 return ERR_PTR(-EINVAL);
580 return ERR_PTR(-ENOMEM);
582 r = dm_get_device(ti, shift(as), ti->begin, ti->len,
583 dm_table_get_mode(ti->table), &p->path.dev);
585 ti->error = "error getting device";
589 if (m->hw_handler_name) {
590 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
592 r = scsi_dh_attach(q, m->hw_handler_name);
595 * Already attached to different hw_handler,
596 * try to reattach with correct one.
599 r = scsi_dh_attach(q, m->hw_handler_name);
603 ti->error = "error attaching hardware handler";
604 dm_put_device(ti, p->path.dev);
609 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
611 dm_put_device(ti, p->path.dev);
622 static struct priority_group *parse_priority_group(struct arg_set *as,
625 static struct param _params[] = {
626 {1, 1024, "invalid number of paths"},
627 {0, 1024, "invalid number of selector args"}
631 unsigned i, nr_selector_args, nr_params;
632 struct priority_group *pg;
633 struct dm_target *ti = m->ti;
637 ti->error = "not enough priority group arguments";
638 return ERR_PTR(-EINVAL);
641 pg = alloc_priority_group();
643 ti->error = "couldn't allocate priority group";
644 return ERR_PTR(-ENOMEM);
648 r = parse_path_selector(as, pg, ti);
655 r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
659 r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
663 nr_params = 1 + nr_selector_args;
664 for (i = 0; i < pg->nr_pgpaths; i++) {
665 struct pgpath *pgpath;
666 struct arg_set path_args;
668 if (as->argc < nr_params) {
669 ti->error = "not enough path parameters";
673 path_args.argc = nr_params;
674 path_args.argv = as->argv;
676 pgpath = parse_path(&path_args, &pg->ps, ti);
677 if (IS_ERR(pgpath)) {
683 list_add_tail(&pgpath->list, &pg->pgpaths);
684 consume(as, nr_params);
690 free_priority_group(pg, ti);
694 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
697 struct dm_target *ti = m->ti;
699 static struct param _params[] = {
700 {0, 1024, "invalid number of hardware handler args"},
703 if (read_param(_params, shift(as), &hw_argc, &ti->error))
709 if (hw_argc > as->argc) {
710 ti->error = "not enough arguments for hardware handler";
714 m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
715 request_module("scsi_dh_%s", m->hw_handler_name);
716 if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
717 ti->error = "unknown hardware handler type";
718 kfree(m->hw_handler_name);
719 m->hw_handler_name = NULL;
724 DMWARN("Ignoring user-specified arguments for "
725 "hardware handler \"%s\"", m->hw_handler_name);
726 consume(as, hw_argc - 1);
731 static int parse_features(struct arg_set *as, struct multipath *m)
735 struct dm_target *ti = m->ti;
736 const char *param_name;
738 static struct param _params[] = {
739 {0, 3, "invalid number of feature args"},
740 {1, 50, "pg_init_retries must be between 1 and 50"},
743 r = read_param(_params, shift(as), &argc, &ti->error);
751 param_name = shift(as);
754 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
755 r = queue_if_no_path(m, 1, 0);
759 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
761 r = read_param(_params + 1, shift(as),
762 &m->pg_init_retries, &ti->error);
767 ti->error = "Unrecognised multipath feature request";
769 } while (argc && !r);
774 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
777 /* target parameters */
778 static struct param _params[] = {
779 {1, 1024, "invalid number of priority groups"},
780 {1, 1024, "invalid initial priority group number"},
786 unsigned pg_count = 0;
787 unsigned next_pg_num;
792 m = alloc_multipath(ti);
794 ti->error = "can't allocate multipath";
798 r = parse_features(&as, m);
802 r = parse_hw_handler(&as, m);
806 r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
810 r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
814 /* parse the priority groups */
816 struct priority_group *pg;
818 pg = parse_priority_group(&as, m);
824 m->nr_valid_paths += pg->nr_pgpaths;
825 list_add_tail(&pg->list, &m->priority_groups);
827 pg->pg_num = pg_count;
832 if (pg_count != m->nr_priority_groups) {
833 ti->error = "priority group count mismatch";
838 ti->num_flush_requests = 1;
847 static void multipath_dtr(struct dm_target *ti)
849 struct multipath *m = (struct multipath *) ti->private;
851 flush_workqueue(kmpath_handlerd);
852 flush_workqueue(kmultipathd);
853 flush_scheduled_work();
858 * Map bios, recording original fields for later in case we have to resubmit
860 static int multipath_map(struct dm_target *ti, struct bio *bio,
861 union map_info *map_context)
864 struct dm_mpath_io *mpio;
865 struct multipath *m = (struct multipath *) ti->private;
867 mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
868 dm_bio_record(&mpio->details, bio);
870 map_context->ptr = mpio;
871 bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
872 r = map_io(m, bio, mpio, 0);
873 if (r < 0 || r == DM_MAPIO_REQUEUE)
874 mempool_free(mpio, m->mpio_pool);
880 * Take a path out of use.
882 static int fail_path(struct pgpath *pgpath)
885 struct multipath *m = pgpath->pg->m;
887 spin_lock_irqsave(&m->lock, flags);
889 if (!pgpath->is_active)
892 DMWARN("Failing path %s.", pgpath->path.dev->name);
894 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
895 pgpath->is_active = 0;
896 pgpath->fail_count++;
900 if (pgpath == m->current_pgpath)
901 m->current_pgpath = NULL;
903 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
904 pgpath->path.dev->name, m->nr_valid_paths);
906 schedule_work(&m->trigger_event);
907 queue_work(kmultipathd, &pgpath->deactivate_path);
910 spin_unlock_irqrestore(&m->lock, flags);
916 * Reinstate a previously-failed path
918 static int reinstate_path(struct pgpath *pgpath)
922 struct multipath *m = pgpath->pg->m;
924 spin_lock_irqsave(&m->lock, flags);
926 if (pgpath->is_active)
929 if (!pgpath->pg->ps.type->reinstate_path) {
930 DMWARN("Reinstate path not supported by path selector %s",
931 pgpath->pg->ps.type->name);
936 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
940 pgpath->is_active = 1;
942 if (!m->nr_valid_paths++ && m->queue_size) {
943 m->current_pgpath = NULL;
944 queue_work(kmultipathd, &m->process_queued_ios);
945 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
946 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
947 m->pg_init_in_progress++;
950 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
951 pgpath->path.dev->name, m->nr_valid_paths);
953 schedule_work(&m->trigger_event);
956 spin_unlock_irqrestore(&m->lock, flags);
962 * Fail or reinstate all paths that match the provided struct dm_dev.
964 static int action_dev(struct multipath *m, struct dm_dev *dev,
968 struct pgpath *pgpath;
969 struct priority_group *pg;
971 list_for_each_entry(pg, &m->priority_groups, list) {
972 list_for_each_entry(pgpath, &pg->pgpaths, list) {
973 if (pgpath->path.dev == dev)
982 * Temporarily try to avoid having to use the specified PG
984 static void bypass_pg(struct multipath *m, struct priority_group *pg,
989 spin_lock_irqsave(&m->lock, flags);
991 pg->bypassed = bypassed;
992 m->current_pgpath = NULL;
993 m->current_pg = NULL;
995 spin_unlock_irqrestore(&m->lock, flags);
997 schedule_work(&m->trigger_event);
1001 * Switch to using the specified PG from the next I/O that gets mapped
1003 static int switch_pg_num(struct multipath *m, const char *pgstr)
1005 struct priority_group *pg;
1007 unsigned long flags;
1009 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1010 (pgnum > m->nr_priority_groups)) {
1011 DMWARN("invalid PG number supplied to switch_pg_num");
1015 spin_lock_irqsave(&m->lock, flags);
1016 list_for_each_entry(pg, &m->priority_groups, list) {
1021 m->current_pgpath = NULL;
1022 m->current_pg = NULL;
1025 spin_unlock_irqrestore(&m->lock, flags);
1027 schedule_work(&m->trigger_event);
1032 * Set/clear bypassed status of a PG.
1033 * PGs are numbered upwards from 1 in the order they were declared.
1035 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1037 struct priority_group *pg;
1040 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1041 (pgnum > m->nr_priority_groups)) {
1042 DMWARN("invalid PG number supplied to bypass_pg");
1046 list_for_each_entry(pg, &m->priority_groups, list) {
1051 bypass_pg(m, pg, bypassed);
1056 * Should we retry pg_init immediately?
1058 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1060 unsigned long flags;
1061 int limit_reached = 0;
1063 spin_lock_irqsave(&m->lock, flags);
1065 if (m->pg_init_count <= m->pg_init_retries)
1066 m->pg_init_required = 1;
1070 spin_unlock_irqrestore(&m->lock, flags);
1072 return limit_reached;
1075 static void pg_init_done(struct dm_path *path, int errors)
1077 struct pgpath *pgpath = path_to_pgpath(path);
1078 struct priority_group *pg = pgpath->pg;
1079 struct multipath *m = pg->m;
1080 unsigned long flags;
1082 /* device or driver problems */
1087 if (!m->hw_handler_name) {
1091 DMERR("Cannot failover device because scsi_dh_%s was not "
1092 "loaded.", m->hw_handler_name);
1094 * Fail path for now, so we do not ping pong
1098 case SCSI_DH_DEV_TEMP_BUSY:
1100 * Probably doing something like FW upgrade on the
1101 * controller so try the other pg.
1103 bypass_pg(m, pg, 1);
1105 /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1107 case SCSI_DH_IMM_RETRY:
1108 case SCSI_DH_RES_TEMP_UNAVAIL:
1109 if (pg_init_limit_reached(m, pgpath))
1115 * We probably do not want to fail the path for a device
1116 * error, but this is what the old dm did. In future
1117 * patches we can do more advanced handling.
1122 spin_lock_irqsave(&m->lock, flags);
1124 if (pgpath == m->current_pgpath) {
1125 DMERR("Could not failover device. Error %d.", errors);
1126 m->current_pgpath = NULL;
1127 m->current_pg = NULL;
1129 } else if (!m->pg_init_required) {
1134 m->pg_init_in_progress--;
1135 if (!m->pg_init_in_progress)
1136 queue_work(kmultipathd, &m->process_queued_ios);
1137 spin_unlock_irqrestore(&m->lock, flags);
1140 static void activate_path(struct work_struct *work)
1143 struct pgpath *pgpath =
1144 container_of(work, struct pgpath, activate_path);
1146 ret = scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev));
1147 pg_init_done(&pgpath->path, ret);
1153 static int do_end_io(struct multipath *m, struct bio *bio,
1154 int error, struct dm_mpath_io *mpio)
1156 unsigned long flags;
1159 return 0; /* I/O complete */
1161 if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1164 if (error == -EOPNOTSUPP)
1167 spin_lock_irqsave(&m->lock, flags);
1168 if (!m->nr_valid_paths) {
1169 if (__must_push_back(m)) {
1170 spin_unlock_irqrestore(&m->lock, flags);
1171 return DM_ENDIO_REQUEUE;
1172 } else if (!m->queue_if_no_path) {
1173 spin_unlock_irqrestore(&m->lock, flags);
1176 spin_unlock_irqrestore(&m->lock, flags);
1180 spin_unlock_irqrestore(&m->lock, flags);
1183 fail_path(mpio->pgpath);
1186 dm_bio_restore(&mpio->details, bio);
1188 /* queue for the daemon to resubmit or fail */
1189 spin_lock_irqsave(&m->lock, flags);
1190 bio_list_add(&m->queued_ios, bio);
1193 queue_work(kmultipathd, &m->process_queued_ios);
1194 spin_unlock_irqrestore(&m->lock, flags);
1196 return DM_ENDIO_INCOMPLETE; /* io not complete */
1199 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1200 int error, union map_info *map_context)
1202 struct multipath *m = ti->private;
1203 struct dm_mpath_io *mpio = map_context->ptr;
1204 struct pgpath *pgpath = mpio->pgpath;
1205 struct path_selector *ps;
1208 r = do_end_io(m, bio, error, mpio);
1210 ps = &pgpath->pg->ps;
1211 if (ps->type->end_io)
1212 ps->type->end_io(ps, &pgpath->path);
1214 if (r != DM_ENDIO_INCOMPLETE)
1215 mempool_free(mpio, m->mpio_pool);
1221 * Suspend can't complete until all the I/O is processed so if
1222 * the last path fails we must error any remaining I/O.
1223 * Note that if the freeze_bdev fails while suspending, the
1224 * queue_if_no_path state is lost - userspace should reset it.
1226 static void multipath_presuspend(struct dm_target *ti)
1228 struct multipath *m = (struct multipath *) ti->private;
1230 queue_if_no_path(m, 0, 1);
1234 * Restore the queue_if_no_path setting.
1236 static void multipath_resume(struct dm_target *ti)
1238 struct multipath *m = (struct multipath *) ti->private;
1239 unsigned long flags;
1241 spin_lock_irqsave(&m->lock, flags);
1242 m->queue_if_no_path = m->saved_queue_if_no_path;
1243 spin_unlock_irqrestore(&m->lock, flags);
1247 * Info output has the following format:
1248 * num_multipath_feature_args [multipath_feature_args]*
1249 * num_handler_status_args [handler_status_args]*
1250 * num_groups init_group_number
1251 * [A|D|E num_ps_status_args [ps_status_args]*
1252 * num_paths num_selector_args
1253 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1255 * Table output has the following format (identical to the constructor string):
1256 * num_feature_args [features_args]*
1257 * num_handler_args hw_handler [hw_handler_args]*
1258 * num_groups init_group_number
1259 * [priority selector-name num_ps_args [ps_args]*
1260 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1262 static int multipath_status(struct dm_target *ti, status_type_t type,
1263 char *result, unsigned int maxlen)
1266 unsigned long flags;
1267 struct multipath *m = (struct multipath *) ti->private;
1268 struct priority_group *pg;
1273 spin_lock_irqsave(&m->lock, flags);
1276 if (type == STATUSTYPE_INFO)
1277 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1279 DMEMIT("%u ", m->queue_if_no_path +
1280 (m->pg_init_retries > 0) * 2);
1281 if (m->queue_if_no_path)
1282 DMEMIT("queue_if_no_path ");
1283 if (m->pg_init_retries)
1284 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1287 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1290 DMEMIT("1 %s ", m->hw_handler_name);
1292 DMEMIT("%u ", m->nr_priority_groups);
1295 pg_num = m->next_pg->pg_num;
1296 else if (m->current_pg)
1297 pg_num = m->current_pg->pg_num;
1301 DMEMIT("%u ", pg_num);
1304 case STATUSTYPE_INFO:
1305 list_for_each_entry(pg, &m->priority_groups, list) {
1307 state = 'D'; /* Disabled */
1308 else if (pg == m->current_pg)
1309 state = 'A'; /* Currently Active */
1311 state = 'E'; /* Enabled */
1313 DMEMIT("%c ", state);
1315 if (pg->ps.type->status)
1316 sz += pg->ps.type->status(&pg->ps, NULL, type,
1322 DMEMIT("%u %u ", pg->nr_pgpaths,
1323 pg->ps.type->info_args);
1325 list_for_each_entry(p, &pg->pgpaths, list) {
1326 DMEMIT("%s %s %u ", p->path.dev->name,
1327 p->is_active ? "A" : "F",
1329 if (pg->ps.type->status)
1330 sz += pg->ps.type->status(&pg->ps,
1331 &p->path, type, result + sz,
1337 case STATUSTYPE_TABLE:
1338 list_for_each_entry(pg, &m->priority_groups, list) {
1339 DMEMIT("%s ", pg->ps.type->name);
1341 if (pg->ps.type->status)
1342 sz += pg->ps.type->status(&pg->ps, NULL, type,
1348 DMEMIT("%u %u ", pg->nr_pgpaths,
1349 pg->ps.type->table_args);
1351 list_for_each_entry(p, &pg->pgpaths, list) {
1352 DMEMIT("%s ", p->path.dev->name);
1353 if (pg->ps.type->status)
1354 sz += pg->ps.type->status(&pg->ps,
1355 &p->path, type, result + sz,
1362 spin_unlock_irqrestore(&m->lock, flags);
1367 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1371 struct multipath *m = (struct multipath *) ti->private;
1375 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1376 return queue_if_no_path(m, 1, 0);
1377 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1378 return queue_if_no_path(m, 0, 0);
1384 if (!strnicmp(argv[0], MESG_STR("disable_group")))
1385 return bypass_pg_num(m, argv[1], 1);
1386 else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1387 return bypass_pg_num(m, argv[1], 0);
1388 else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1389 return switch_pg_num(m, argv[1]);
1390 else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1391 action = reinstate_path;
1392 else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1397 r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1398 dm_table_get_mode(ti->table), &dev);
1400 DMWARN("message: error getting device %s",
1405 r = action_dev(m, dev, action);
1407 dm_put_device(ti, dev);
1412 DMWARN("Unrecognised multipath message received.");
1416 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1419 struct multipath *m = (struct multipath *) ti->private;
1420 struct block_device *bdev = NULL;
1422 unsigned long flags;
1425 spin_lock_irqsave(&m->lock, flags);
1427 if (!m->current_pgpath)
1430 if (m->current_pgpath) {
1431 bdev = m->current_pgpath->path.dev->bdev;
1432 mode = m->current_pgpath->path.dev->mode;
1440 spin_unlock_irqrestore(&m->lock, flags);
1442 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1445 /*-----------------------------------------------------------------
1447 *---------------------------------------------------------------*/
1448 static struct target_type multipath_target = {
1449 .name = "multipath",
1450 .version = {1, 0, 5},
1451 .module = THIS_MODULE,
1452 .ctr = multipath_ctr,
1453 .dtr = multipath_dtr,
1454 .map = multipath_map,
1455 .end_io = multipath_end_io,
1456 .presuspend = multipath_presuspend,
1457 .resume = multipath_resume,
1458 .status = multipath_status,
1459 .message = multipath_message,
1460 .ioctl = multipath_ioctl,
1463 static int __init dm_multipath_init(void)
1467 /* allocate a slab for the dm_ios */
1468 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1472 r = dm_register_target(&multipath_target);
1474 DMERR("register failed %d", r);
1475 kmem_cache_destroy(_mpio_cache);
1479 kmultipathd = create_workqueue("kmpathd");
1481 DMERR("failed to create workqueue kmpathd");
1482 dm_unregister_target(&multipath_target);
1483 kmem_cache_destroy(_mpio_cache);
1488 * A separate workqueue is used to handle the device handlers
1489 * to avoid overloading existing workqueue. Overloading the
1490 * old workqueue would also create a bottleneck in the
1491 * path of the storage hardware device activation.
1493 kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1494 if (!kmpath_handlerd) {
1495 DMERR("failed to create workqueue kmpath_handlerd");
1496 destroy_workqueue(kmultipathd);
1497 dm_unregister_target(&multipath_target);
1498 kmem_cache_destroy(_mpio_cache);
1502 DMINFO("version %u.%u.%u loaded",
1503 multipath_target.version[0], multipath_target.version[1],
1504 multipath_target.version[2]);
1509 static void __exit dm_multipath_exit(void)
1511 destroy_workqueue(kmpath_handlerd);
1512 destroy_workqueue(kmultipathd);
1514 dm_unregister_target(&multipath_target);
1515 kmem_cache_destroy(_mpio_cache);
1518 module_init(dm_multipath_init);
1519 module_exit(dm_multipath_exit);
1521 MODULE_DESCRIPTION(DM_NAME " multipath target");
1522 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1523 MODULE_LICENSE("GPL");