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 <asm/atomic.h>
24 #define DM_MSG_PREFIX "multipath"
25 #define MESG_STR(x) x, sizeof(x)
26 #define DM_PG_INIT_DELAY_MSECS 2000
27 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
31 struct list_head list;
33 struct priority_group *pg; /* Owning PG */
34 unsigned is_active; /* Path status */
35 unsigned fail_count; /* Cumulative failure count */
38 struct delayed_work 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 #define FEATURE_NO_PARTITIONS 1
62 /* Multipath context */
64 struct list_head list;
69 const char *hw_handler_name;
70 char *hw_handler_params;
72 unsigned nr_priority_groups;
73 struct list_head priority_groups;
75 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
77 unsigned pg_init_required; /* pg_init needs calling? */
78 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
79 unsigned pg_init_delay_retry; /* Delay pg_init retry? */
81 unsigned nr_valid_paths; /* Total number of usable paths */
82 struct pgpath *current_pgpath;
83 struct priority_group *current_pg;
84 struct priority_group *next_pg; /* Switch to this PG if set */
85 unsigned repeat_count; /* I/Os left before calling PS again */
87 unsigned queue_io; /* Must we queue all I/O? */
88 unsigned queue_if_no_path; /* Queue I/O if last path fails? */
89 unsigned saved_queue_if_no_path;/* Saved state during suspension */
90 unsigned pg_init_retries; /* Number of times to retry pg_init */
91 unsigned pg_init_count; /* Number of times pg_init called */
92 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
93 unsigned features; /* Additional selected features */
95 struct work_struct process_queued_ios;
96 struct list_head queued_ios;
99 struct work_struct trigger_event;
102 * We must use a mempool of dm_mpath_io structs so that we
103 * can resubmit bios on error.
105 mempool_t *mpio_pool;
107 struct mutex work_mutex;
111 * Context information attached to each bio we process.
114 struct pgpath *pgpath;
118 typedef int (*action_fn) (struct pgpath *pgpath);
120 #define MIN_IOS 256 /* Mempool size */
122 static struct kmem_cache *_mpio_cache;
124 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
125 static void process_queued_ios(struct work_struct *work);
126 static void trigger_event(struct work_struct *work);
127 static void activate_path(struct work_struct *work);
130 /*-----------------------------------------------
131 * Allocation routines
132 *-----------------------------------------------*/
134 static struct pgpath *alloc_pgpath(void)
136 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
139 pgpath->is_active = 1;
140 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
146 static void free_pgpath(struct pgpath *pgpath)
151 static struct priority_group *alloc_priority_group(void)
153 struct priority_group *pg;
155 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
158 INIT_LIST_HEAD(&pg->pgpaths);
163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
165 struct pgpath *pgpath, *tmp;
167 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
168 list_del(&pgpath->list);
169 dm_put_device(ti, pgpath->path.dev);
174 static void free_priority_group(struct priority_group *pg,
175 struct dm_target *ti)
177 struct path_selector *ps = &pg->ps;
180 ps->type->destroy(ps);
181 dm_put_path_selector(ps->type);
184 free_pgpaths(&pg->pgpaths, ti);
188 static struct multipath *alloc_multipath(struct dm_target *ti)
192 m = kzalloc(sizeof(*m), GFP_KERNEL);
194 INIT_LIST_HEAD(&m->priority_groups);
195 INIT_LIST_HEAD(&m->queued_ios);
196 spin_lock_init(&m->lock);
198 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
199 INIT_WORK(&m->process_queued_ios, process_queued_ios);
200 INIT_WORK(&m->trigger_event, trigger_event);
201 init_waitqueue_head(&m->pg_init_wait);
202 mutex_init(&m->work_mutex);
203 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
215 static void free_multipath(struct multipath *m)
217 struct priority_group *pg, *tmp;
219 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
221 free_priority_group(pg, m->ti);
224 kfree(m->hw_handler_name);
225 kfree(m->hw_handler_params);
226 mempool_destroy(m->mpio_pool);
231 /*-----------------------------------------------
233 *-----------------------------------------------*/
235 static void __pg_init_all_paths(struct multipath *m)
237 struct pgpath *pgpath;
238 unsigned long pg_init_delay = 0;
241 m->pg_init_required = 0;
242 if (m->pg_init_delay_retry)
243 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
244 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
245 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
246 /* Skip failed paths */
247 if (!pgpath->is_active)
249 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
251 m->pg_init_in_progress++;
255 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
257 m->current_pg = pgpath->pg;
259 /* Must we initialise the PG first, and queue I/O till it's ready? */
260 if (m->hw_handler_name) {
261 m->pg_init_required = 1;
264 m->pg_init_required = 0;
268 m->pg_init_count = 0;
271 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
274 struct dm_path *path;
276 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
280 m->current_pgpath = path_to_pgpath(path);
282 if (m->current_pg != pg)
283 __switch_pg(m, m->current_pgpath);
288 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
290 struct priority_group *pg;
291 unsigned bypassed = 1;
293 if (!m->nr_valid_paths)
296 /* Were we instructed to switch PG? */
300 if (!__choose_path_in_pg(m, pg, nr_bytes))
304 /* Don't change PG until it has no remaining paths */
305 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
309 * Loop through priority groups until we find a valid path.
310 * First time we skip PGs marked 'bypassed'.
311 * Second time we only try the ones we skipped.
314 list_for_each_entry(pg, &m->priority_groups, list) {
315 if (pg->bypassed == bypassed)
317 if (!__choose_path_in_pg(m, pg, nr_bytes))
320 } while (bypassed--);
323 m->current_pgpath = NULL;
324 m->current_pg = NULL;
328 * Check whether bios must be queued in the device-mapper core rather
329 * than here in the target.
331 * m->lock must be held on entry.
333 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
334 * same value then we are not between multipath_presuspend()
335 * and multipath_resume() calls and we have no need to check
336 * for the DMF_NOFLUSH_SUSPENDING flag.
338 static int __must_push_back(struct multipath *m)
340 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
341 dm_noflush_suspending(m->ti));
344 static int map_io(struct multipath *m, struct request *clone,
345 struct dm_mpath_io *mpio, unsigned was_queued)
347 int r = DM_MAPIO_REMAPPED;
348 size_t nr_bytes = blk_rq_bytes(clone);
350 struct pgpath *pgpath;
351 struct block_device *bdev;
353 spin_lock_irqsave(&m->lock, flags);
355 /* Do we need to select a new pgpath? */
356 if (!m->current_pgpath ||
357 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
358 __choose_pgpath(m, nr_bytes);
360 pgpath = m->current_pgpath;
365 if ((pgpath && m->queue_io) ||
366 (!pgpath && m->queue_if_no_path)) {
367 /* Queue for the daemon to resubmit */
368 list_add_tail(&clone->queuelist, &m->queued_ios);
370 if ((m->pg_init_required && !m->pg_init_in_progress) ||
372 queue_work(kmultipathd, &m->process_queued_ios);
374 r = DM_MAPIO_SUBMITTED;
376 bdev = pgpath->path.dev->bdev;
377 clone->q = bdev_get_queue(bdev);
378 clone->rq_disk = bdev->bd_disk;
379 } else if (__must_push_back(m))
380 r = DM_MAPIO_REQUEUE;
382 r = -EIO; /* Failed */
384 mpio->pgpath = pgpath;
385 mpio->nr_bytes = nr_bytes;
387 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
388 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
391 spin_unlock_irqrestore(&m->lock, flags);
397 * If we run out of usable paths, should we queue I/O or error it?
399 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
400 unsigned save_old_value)
404 spin_lock_irqsave(&m->lock, flags);
407 m->saved_queue_if_no_path = m->queue_if_no_path;
409 m->saved_queue_if_no_path = queue_if_no_path;
410 m->queue_if_no_path = queue_if_no_path;
411 if (!m->queue_if_no_path && m->queue_size)
412 queue_work(kmultipathd, &m->process_queued_ios);
414 spin_unlock_irqrestore(&m->lock, flags);
419 /*-----------------------------------------------------------------
420 * The multipath daemon is responsible for resubmitting queued ios.
421 *---------------------------------------------------------------*/
423 static void dispatch_queued_ios(struct multipath *m)
427 struct dm_mpath_io *mpio;
428 union map_info *info;
429 struct request *clone, *n;
432 spin_lock_irqsave(&m->lock, flags);
433 list_splice_init(&m->queued_ios, &cl);
434 spin_unlock_irqrestore(&m->lock, flags);
436 list_for_each_entry_safe(clone, n, &cl, queuelist) {
437 list_del_init(&clone->queuelist);
439 info = dm_get_rq_mapinfo(clone);
442 r = map_io(m, clone, mpio, 1);
444 mempool_free(mpio, m->mpio_pool);
445 dm_kill_unmapped_request(clone, r);
446 } else if (r == DM_MAPIO_REMAPPED)
447 dm_dispatch_request(clone);
448 else if (r == DM_MAPIO_REQUEUE) {
449 mempool_free(mpio, m->mpio_pool);
450 dm_requeue_unmapped_request(clone);
455 static void process_queued_ios(struct work_struct *work)
457 struct multipath *m =
458 container_of(work, struct multipath, process_queued_ios);
459 struct pgpath *pgpath = NULL;
460 unsigned must_queue = 1;
463 spin_lock_irqsave(&m->lock, flags);
468 if (!m->current_pgpath)
469 __choose_pgpath(m, 0);
471 pgpath = m->current_pgpath;
473 if ((pgpath && !m->queue_io) ||
474 (!pgpath && !m->queue_if_no_path))
477 if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
478 __pg_init_all_paths(m);
481 spin_unlock_irqrestore(&m->lock, flags);
483 dispatch_queued_ios(m);
487 * An event is triggered whenever a path is taken out of use.
488 * Includes path failure and PG bypass.
490 static void trigger_event(struct work_struct *work)
492 struct multipath *m =
493 container_of(work, struct multipath, trigger_event);
495 dm_table_event(m->ti->table);
498 /*-----------------------------------------------------------------
499 * Constructor/argument parsing:
500 * <#multipath feature args> [<arg>]*
501 * <#hw_handler args> [hw_handler [<arg>]*]
503 * <initial priority group>
504 * [<selector> <#selector args> [<arg>]*
505 * <#paths> <#per-path selector args>
506 * [<path> [<arg>]* ]+ ]+
507 *---------------------------------------------------------------*/
514 static int read_param(struct param *param, char *str, unsigned *v, char **error)
517 (sscanf(str, "%u", v) != 1) ||
520 *error = param->error;
532 static char *shift(struct arg_set *as)
546 static void consume(struct arg_set *as, unsigned n)
548 BUG_ON (as->argc < n);
553 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
554 struct dm_target *ti)
557 struct path_selector_type *pst;
560 static struct param _params[] = {
561 {0, 1024, "invalid number of path selector args"},
564 pst = dm_get_path_selector(shift(as));
566 ti->error = "unknown path selector type";
570 r = read_param(_params, shift(as), &ps_argc, &ti->error);
572 dm_put_path_selector(pst);
576 if (ps_argc > as->argc) {
577 dm_put_path_selector(pst);
578 ti->error = "not enough arguments for path selector";
582 r = pst->create(&pg->ps, ps_argc, as->argv);
584 dm_put_path_selector(pst);
585 ti->error = "path selector constructor failed";
590 consume(as, ps_argc);
595 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
596 struct dm_target *ti)
600 struct multipath *m = ti->private;
602 /* we need at least a path arg */
604 ti->error = "no device given";
605 return ERR_PTR(-EINVAL);
610 return ERR_PTR(-ENOMEM);
612 r = dm_get_device(ti, shift(as), dm_table_get_mode(ti->table),
615 ti->error = "error getting device";
619 if (m->hw_handler_name) {
620 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
622 r = scsi_dh_attach(q, m->hw_handler_name);
625 * Already attached to different hw_handler,
626 * try to reattach with correct one.
629 r = scsi_dh_attach(q, m->hw_handler_name);
633 ti->error = "error attaching hardware handler";
634 dm_put_device(ti, p->path.dev);
638 if (m->hw_handler_params) {
639 r = scsi_dh_set_params(q, m->hw_handler_params);
641 ti->error = "unable to set hardware "
642 "handler parameters";
644 dm_put_device(ti, p->path.dev);
650 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
652 dm_put_device(ti, p->path.dev);
663 static struct priority_group *parse_priority_group(struct arg_set *as,
666 static struct param _params[] = {
667 {1, 1024, "invalid number of paths"},
668 {0, 1024, "invalid number of selector args"}
672 unsigned i, nr_selector_args, nr_params;
673 struct priority_group *pg;
674 struct dm_target *ti = m->ti;
678 ti->error = "not enough priority group arguments";
679 return ERR_PTR(-EINVAL);
682 pg = alloc_priority_group();
684 ti->error = "couldn't allocate priority group";
685 return ERR_PTR(-ENOMEM);
689 r = parse_path_selector(as, pg, ti);
696 r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
700 r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
704 nr_params = 1 + nr_selector_args;
705 for (i = 0; i < pg->nr_pgpaths; i++) {
706 struct pgpath *pgpath;
707 struct arg_set path_args;
709 if (as->argc < nr_params) {
710 ti->error = "not enough path parameters";
715 path_args.argc = nr_params;
716 path_args.argv = as->argv;
718 pgpath = parse_path(&path_args, &pg->ps, ti);
719 if (IS_ERR(pgpath)) {
725 list_add_tail(&pgpath->list, &pg->pgpaths);
726 consume(as, nr_params);
732 free_priority_group(pg, ti);
736 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
740 struct dm_target *ti = m->ti;
742 static struct param _params[] = {
743 {0, 1024, "invalid number of hardware handler args"},
746 if (read_param(_params, shift(as), &hw_argc, &ti->error))
752 if (hw_argc > as->argc) {
753 ti->error = "not enough arguments for hardware handler";
757 m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
758 request_module("scsi_dh_%s", m->hw_handler_name);
759 if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
760 ti->error = "unknown hardware handler type";
769 for (i = 0; i <= hw_argc - 2; i++)
770 len += strlen(as->argv[i]) + 1;
771 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
773 ti->error = "memory allocation failed";
777 j = sprintf(p, "%d", hw_argc - 1);
778 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
779 j = sprintf(p, "%s", as->argv[i]);
781 consume(as, hw_argc - 1);
785 kfree(m->hw_handler_name);
786 m->hw_handler_name = NULL;
790 static int parse_features(struct arg_set *as, struct multipath *m)
794 struct dm_target *ti = m->ti;
795 const char *param_name;
797 static struct param _params[] = {
798 {0, 5, "invalid number of feature args"},
799 {1, 50, "pg_init_retries must be between 1 and 50"},
800 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
803 r = read_param(_params, shift(as), &argc, &ti->error);
811 param_name = shift(as);
814 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
815 r = queue_if_no_path(m, 1, 0);
819 if (!strnicmp(param_name, MESG_STR("no_partitions"))) {
820 m->features |= FEATURE_NO_PARTITIONS;
823 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
825 r = read_param(_params + 1, shift(as),
826 &m->pg_init_retries, &ti->error);
831 if (!strnicmp(param_name, MESG_STR("pg_init_delay_msecs")) &&
833 r = read_param(_params + 2, shift(as),
834 &m->pg_init_delay_msecs, &ti->error);
839 ti->error = "Unrecognised multipath feature request";
841 } while (argc && !r);
846 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
849 /* target parameters */
850 static struct param _params[] = {
851 {0, 1024, "invalid number of priority groups"},
852 {0, 1024, "invalid initial priority group number"},
858 unsigned pg_count = 0;
859 unsigned next_pg_num;
864 m = alloc_multipath(ti);
866 ti->error = "can't allocate multipath";
870 r = parse_features(&as, m);
874 r = parse_hw_handler(&as, m);
878 r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
882 r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
886 if ((!m->nr_priority_groups && next_pg_num) ||
887 (m->nr_priority_groups && !next_pg_num)) {
888 ti->error = "invalid initial priority group";
893 /* parse the priority groups */
895 struct priority_group *pg;
897 pg = parse_priority_group(&as, m);
903 m->nr_valid_paths += pg->nr_pgpaths;
904 list_add_tail(&pg->list, &m->priority_groups);
906 pg->pg_num = pg_count;
911 if (pg_count != m->nr_priority_groups) {
912 ti->error = "priority group count mismatch";
917 ti->num_flush_requests = 1;
918 ti->num_discard_requests = 1;
927 static void multipath_wait_for_pg_init_completion(struct multipath *m)
929 DECLARE_WAITQUEUE(wait, current);
932 add_wait_queue(&m->pg_init_wait, &wait);
935 set_current_state(TASK_UNINTERRUPTIBLE);
937 spin_lock_irqsave(&m->lock, flags);
938 if (!m->pg_init_in_progress) {
939 spin_unlock_irqrestore(&m->lock, flags);
942 spin_unlock_irqrestore(&m->lock, flags);
946 set_current_state(TASK_RUNNING);
948 remove_wait_queue(&m->pg_init_wait, &wait);
951 static void flush_multipath_work(struct multipath *m)
953 flush_workqueue(kmpath_handlerd);
954 multipath_wait_for_pg_init_completion(m);
955 flush_workqueue(kmultipathd);
956 flush_work_sync(&m->trigger_event);
959 static void multipath_dtr(struct dm_target *ti)
961 struct multipath *m = ti->private;
963 flush_multipath_work(m);
968 * Map cloned requests
970 static int multipath_map(struct dm_target *ti, struct request *clone,
971 union map_info *map_context)
974 struct dm_mpath_io *mpio;
975 struct multipath *m = (struct multipath *) ti->private;
977 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
979 /* ENOMEM, requeue */
980 return DM_MAPIO_REQUEUE;
981 memset(mpio, 0, sizeof(*mpio));
983 map_context->ptr = mpio;
984 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
985 r = map_io(m, clone, mpio, 0);
986 if (r < 0 || r == DM_MAPIO_REQUEUE)
987 mempool_free(mpio, m->mpio_pool);
993 * Take a path out of use.
995 static int fail_path(struct pgpath *pgpath)
998 struct multipath *m = pgpath->pg->m;
1000 spin_lock_irqsave(&m->lock, flags);
1002 if (!pgpath->is_active)
1005 DMWARN("Failing path %s.", pgpath->path.dev->name);
1007 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1008 pgpath->is_active = 0;
1009 pgpath->fail_count++;
1011 m->nr_valid_paths--;
1013 if (pgpath == m->current_pgpath)
1014 m->current_pgpath = NULL;
1016 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1017 pgpath->path.dev->name, m->nr_valid_paths);
1019 schedule_work(&m->trigger_event);
1022 spin_unlock_irqrestore(&m->lock, flags);
1028 * Reinstate a previously-failed path
1030 static int reinstate_path(struct pgpath *pgpath)
1033 unsigned long flags;
1034 struct multipath *m = pgpath->pg->m;
1036 spin_lock_irqsave(&m->lock, flags);
1038 if (pgpath->is_active)
1041 if (!pgpath->pg->ps.type->reinstate_path) {
1042 DMWARN("Reinstate path not supported by path selector %s",
1043 pgpath->pg->ps.type->name);
1048 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1052 pgpath->is_active = 1;
1054 if (!m->nr_valid_paths++ && m->queue_size) {
1055 m->current_pgpath = NULL;
1056 queue_work(kmultipathd, &m->process_queued_ios);
1057 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1058 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1059 m->pg_init_in_progress++;
1062 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1063 pgpath->path.dev->name, m->nr_valid_paths);
1065 schedule_work(&m->trigger_event);
1068 spin_unlock_irqrestore(&m->lock, flags);
1074 * Fail or reinstate all paths that match the provided struct dm_dev.
1076 static int action_dev(struct multipath *m, struct dm_dev *dev,
1080 struct pgpath *pgpath;
1081 struct priority_group *pg;
1083 list_for_each_entry(pg, &m->priority_groups, list) {
1084 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1085 if (pgpath->path.dev == dev)
1094 * Temporarily try to avoid having to use the specified PG
1096 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1099 unsigned long flags;
1101 spin_lock_irqsave(&m->lock, flags);
1103 pg->bypassed = bypassed;
1104 m->current_pgpath = NULL;
1105 m->current_pg = NULL;
1107 spin_unlock_irqrestore(&m->lock, flags);
1109 schedule_work(&m->trigger_event);
1113 * Switch to using the specified PG from the next I/O that gets mapped
1115 static int switch_pg_num(struct multipath *m, const char *pgstr)
1117 struct priority_group *pg;
1119 unsigned long flags;
1121 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1122 (pgnum > m->nr_priority_groups)) {
1123 DMWARN("invalid PG number supplied to switch_pg_num");
1127 spin_lock_irqsave(&m->lock, flags);
1128 list_for_each_entry(pg, &m->priority_groups, list) {
1133 m->current_pgpath = NULL;
1134 m->current_pg = NULL;
1137 spin_unlock_irqrestore(&m->lock, flags);
1139 schedule_work(&m->trigger_event);
1144 * Set/clear bypassed status of a PG.
1145 * PGs are numbered upwards from 1 in the order they were declared.
1147 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1149 struct priority_group *pg;
1152 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1153 (pgnum > m->nr_priority_groups)) {
1154 DMWARN("invalid PG number supplied to bypass_pg");
1158 list_for_each_entry(pg, &m->priority_groups, list) {
1163 bypass_pg(m, pg, bypassed);
1168 * Should we retry pg_init immediately?
1170 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1172 unsigned long flags;
1173 int limit_reached = 0;
1175 spin_lock_irqsave(&m->lock, flags);
1177 if (m->pg_init_count <= m->pg_init_retries)
1178 m->pg_init_required = 1;
1182 spin_unlock_irqrestore(&m->lock, flags);
1184 return limit_reached;
1187 static void pg_init_done(void *data, int errors)
1189 struct pgpath *pgpath = data;
1190 struct priority_group *pg = pgpath->pg;
1191 struct multipath *m = pg->m;
1192 unsigned long flags;
1193 unsigned delay_retry = 0;
1195 /* device or driver problems */
1200 if (!m->hw_handler_name) {
1204 DMERR("Count not failover device %s: Handler scsi_dh_%s "
1205 "was not loaded.", pgpath->path.dev->name,
1206 m->hw_handler_name);
1208 * Fail path for now, so we do not ping pong
1212 case SCSI_DH_DEV_TEMP_BUSY:
1214 * Probably doing something like FW upgrade on the
1215 * controller so try the other pg.
1217 bypass_pg(m, pg, 1);
1219 case SCSI_DH_DEV_OFFLINED:
1220 DMWARN("Device %s offlined.", pgpath->path.dev->name);
1224 /* Wait before retrying. */
1226 case SCSI_DH_IMM_RETRY:
1227 case SCSI_DH_RES_TEMP_UNAVAIL:
1228 if (pg_init_limit_reached(m, pgpath))
1234 * We probably do not want to fail the path for a device
1235 * error, but this is what the old dm did. In future
1236 * patches we can do more advanced handling.
1241 spin_lock_irqsave(&m->lock, flags);
1243 if (pgpath == m->current_pgpath) {
1244 DMERR("Could not failover device %s, error %d.",
1245 pgpath->path.dev->name, errors);
1246 m->current_pgpath = NULL;
1247 m->current_pg = NULL;
1249 } else if (!m->pg_init_required)
1252 if (--m->pg_init_in_progress)
1253 /* Activations of other paths are still on going */
1256 if (!m->pg_init_required)
1259 m->pg_init_delay_retry = delay_retry;
1260 queue_work(kmultipathd, &m->process_queued_ios);
1263 * Wake up any thread waiting to suspend.
1265 wake_up(&m->pg_init_wait);
1268 spin_unlock_irqrestore(&m->lock, flags);
1271 static void activate_path(struct work_struct *work)
1273 struct pgpath *pgpath =
1274 container_of(work, struct pgpath, activate_path.work);
1276 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1277 pg_init_done, pgpath);
1283 static int do_end_io(struct multipath *m, struct request *clone,
1284 int error, struct dm_mpath_io *mpio)
1287 * We don't queue any clone request inside the multipath target
1288 * during end I/O handling, since those clone requests don't have
1289 * bio clones. If we queue them inside the multipath target,
1290 * we need to make bio clones, that requires memory allocation.
1291 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1292 * don't have bio clones.)
1293 * Instead of queueing the clone request here, we queue the original
1294 * request into dm core, which will remake a clone request and
1295 * clone bios for it and resubmit it later.
1297 int r = DM_ENDIO_REQUEUE;
1298 unsigned long flags;
1300 if (!error && !clone->errors)
1301 return 0; /* I/O complete */
1303 if (error == -EOPNOTSUPP || error == -EREMOTEIO)
1307 fail_path(mpio->pgpath);
1309 spin_lock_irqsave(&m->lock, flags);
1310 if (!m->nr_valid_paths) {
1311 if (!m->queue_if_no_path) {
1312 if (!__must_push_back(m))
1315 if (error == -EBADE)
1319 spin_unlock_irqrestore(&m->lock, flags);
1324 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1325 int error, union map_info *map_context)
1327 struct multipath *m = ti->private;
1328 struct dm_mpath_io *mpio = map_context->ptr;
1329 struct pgpath *pgpath = mpio->pgpath;
1330 struct path_selector *ps;
1333 r = do_end_io(m, clone, error, mpio);
1335 ps = &pgpath->pg->ps;
1336 if (ps->type->end_io)
1337 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1339 mempool_free(mpio, m->mpio_pool);
1345 * Suspend can't complete until all the I/O is processed so if
1346 * the last path fails we must error any remaining I/O.
1347 * Note that if the freeze_bdev fails while suspending, the
1348 * queue_if_no_path state is lost - userspace should reset it.
1350 static void multipath_presuspend(struct dm_target *ti)
1352 struct multipath *m = (struct multipath *) ti->private;
1354 queue_if_no_path(m, 0, 1);
1357 static void multipath_postsuspend(struct dm_target *ti)
1359 struct multipath *m = ti->private;
1361 mutex_lock(&m->work_mutex);
1362 flush_multipath_work(m);
1363 mutex_unlock(&m->work_mutex);
1367 * Restore the queue_if_no_path setting.
1369 static void multipath_resume(struct dm_target *ti)
1371 struct multipath *m = (struct multipath *) ti->private;
1372 unsigned long flags;
1374 spin_lock_irqsave(&m->lock, flags);
1375 m->queue_if_no_path = m->saved_queue_if_no_path;
1376 spin_unlock_irqrestore(&m->lock, flags);
1380 * Info output has the following format:
1381 * num_multipath_feature_args [multipath_feature_args]*
1382 * num_handler_status_args [handler_status_args]*
1383 * num_groups init_group_number
1384 * [A|D|E num_ps_status_args [ps_status_args]*
1385 * num_paths num_selector_args
1386 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1388 * Table output has the following format (identical to the constructor string):
1389 * num_feature_args [features_args]*
1390 * num_handler_args hw_handler [hw_handler_args]*
1391 * num_groups init_group_number
1392 * [priority selector-name num_ps_args [ps_args]*
1393 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1395 static int multipath_status(struct dm_target *ti, status_type_t type,
1396 char *result, unsigned int maxlen)
1399 unsigned long flags;
1400 struct multipath *m = (struct multipath *) ti->private;
1401 struct priority_group *pg;
1406 spin_lock_irqsave(&m->lock, flags);
1409 if (type == STATUSTYPE_INFO)
1410 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1412 DMEMIT("%u ", m->queue_if_no_path +
1413 (m->pg_init_retries > 0) * 2 +
1414 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1415 (m->features & FEATURE_NO_PARTITIONS));
1416 if (m->queue_if_no_path)
1417 DMEMIT("queue_if_no_path ");
1418 if (m->pg_init_retries)
1419 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1420 if (m->features & FEATURE_NO_PARTITIONS)
1421 DMEMIT("no_partitions ");
1422 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1423 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1426 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1429 DMEMIT("1 %s ", m->hw_handler_name);
1431 DMEMIT("%u ", m->nr_priority_groups);
1434 pg_num = m->next_pg->pg_num;
1435 else if (m->current_pg)
1436 pg_num = m->current_pg->pg_num;
1438 pg_num = (m->nr_priority_groups ? 1 : 0);
1440 DMEMIT("%u ", pg_num);
1443 case STATUSTYPE_INFO:
1444 list_for_each_entry(pg, &m->priority_groups, list) {
1446 state = 'D'; /* Disabled */
1447 else if (pg == m->current_pg)
1448 state = 'A'; /* Currently Active */
1450 state = 'E'; /* Enabled */
1452 DMEMIT("%c ", state);
1454 if (pg->ps.type->status)
1455 sz += pg->ps.type->status(&pg->ps, NULL, type,
1461 DMEMIT("%u %u ", pg->nr_pgpaths,
1462 pg->ps.type->info_args);
1464 list_for_each_entry(p, &pg->pgpaths, list) {
1465 DMEMIT("%s %s %u ", p->path.dev->name,
1466 p->is_active ? "A" : "F",
1468 if (pg->ps.type->status)
1469 sz += pg->ps.type->status(&pg->ps,
1470 &p->path, type, result + sz,
1476 case STATUSTYPE_TABLE:
1477 list_for_each_entry(pg, &m->priority_groups, list) {
1478 DMEMIT("%s ", pg->ps.type->name);
1480 if (pg->ps.type->status)
1481 sz += pg->ps.type->status(&pg->ps, NULL, type,
1487 DMEMIT("%u %u ", pg->nr_pgpaths,
1488 pg->ps.type->table_args);
1490 list_for_each_entry(p, &pg->pgpaths, list) {
1491 DMEMIT("%s ", p->path.dev->name);
1492 if (pg->ps.type->status)
1493 sz += pg->ps.type->status(&pg->ps,
1494 &p->path, type, result + sz,
1501 spin_unlock_irqrestore(&m->lock, flags);
1506 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1510 struct multipath *m = (struct multipath *) ti->private;
1513 mutex_lock(&m->work_mutex);
1515 if (dm_suspended(ti)) {
1521 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
1522 r = queue_if_no_path(m, 1, 0);
1524 } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
1525 r = queue_if_no_path(m, 0, 0);
1531 DMWARN("Unrecognised multipath message received.");
1535 if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
1536 r = bypass_pg_num(m, argv[1], 1);
1538 } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
1539 r = bypass_pg_num(m, argv[1], 0);
1541 } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
1542 r = switch_pg_num(m, argv[1]);
1544 } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1545 action = reinstate_path;
1546 else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1549 DMWARN("Unrecognised multipath message received.");
1553 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1555 DMWARN("message: error getting device %s",
1560 r = action_dev(m, dev, action);
1562 dm_put_device(ti, dev);
1565 mutex_unlock(&m->work_mutex);
1569 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1572 struct multipath *m = (struct multipath *) ti->private;
1573 struct block_device *bdev = NULL;
1575 unsigned long flags;
1578 spin_lock_irqsave(&m->lock, flags);
1580 if (!m->current_pgpath)
1581 __choose_pgpath(m, 0);
1583 if (m->current_pgpath) {
1584 bdev = m->current_pgpath->path.dev->bdev;
1585 mode = m->current_pgpath->path.dev->mode;
1593 spin_unlock_irqrestore(&m->lock, flags);
1595 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1598 static int multipath_iterate_devices(struct dm_target *ti,
1599 iterate_devices_callout_fn fn, void *data)
1601 struct multipath *m = ti->private;
1602 struct priority_group *pg;
1606 list_for_each_entry(pg, &m->priority_groups, list) {
1607 list_for_each_entry(p, &pg->pgpaths, list) {
1608 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1618 static int __pgpath_busy(struct pgpath *pgpath)
1620 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1622 return dm_underlying_device_busy(q);
1626 * We return "busy", only when we can map I/Os but underlying devices
1627 * are busy (so even if we map I/Os now, the I/Os will wait on
1628 * the underlying queue).
1629 * In other words, if we want to kill I/Os or queue them inside us
1630 * due to map unavailability, we don't return "busy". Otherwise,
1631 * dm core won't give us the I/Os and we can't do what we want.
1633 static int multipath_busy(struct dm_target *ti)
1635 int busy = 0, has_active = 0;
1636 struct multipath *m = ti->private;
1637 struct priority_group *pg;
1638 struct pgpath *pgpath;
1639 unsigned long flags;
1641 spin_lock_irqsave(&m->lock, flags);
1643 /* Guess which priority_group will be used at next mapping time */
1644 if (unlikely(!m->current_pgpath && m->next_pg))
1646 else if (likely(m->current_pg))
1650 * We don't know which pg will be used at next mapping time.
1651 * We don't call __choose_pgpath() here to avoid to trigger
1652 * pg_init just by busy checking.
1653 * So we don't know whether underlying devices we will be using
1654 * at next mapping time are busy or not. Just try mapping.
1659 * If there is one non-busy active path at least, the path selector
1660 * will be able to select it. So we consider such a pg as not busy.
1663 list_for_each_entry(pgpath, &pg->pgpaths, list)
1664 if (pgpath->is_active) {
1667 if (!__pgpath_busy(pgpath)) {
1675 * No active path in this pg, so this pg won't be used and
1676 * the current_pg will be changed at next mapping time.
1677 * We need to try mapping to determine it.
1682 spin_unlock_irqrestore(&m->lock, flags);
1687 /*-----------------------------------------------------------------
1689 *---------------------------------------------------------------*/
1690 static struct target_type multipath_target = {
1691 .name = "multipath",
1692 .version = {1, 3, 0},
1693 .module = THIS_MODULE,
1694 .ctr = multipath_ctr,
1695 .dtr = multipath_dtr,
1696 .map_rq = multipath_map,
1697 .rq_end_io = multipath_end_io,
1698 .presuspend = multipath_presuspend,
1699 .postsuspend = multipath_postsuspend,
1700 .resume = multipath_resume,
1701 .status = multipath_status,
1702 .message = multipath_message,
1703 .ioctl = multipath_ioctl,
1704 .iterate_devices = multipath_iterate_devices,
1705 .busy = multipath_busy,
1708 static int __init dm_multipath_init(void)
1712 /* allocate a slab for the dm_ios */
1713 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1717 r = dm_register_target(&multipath_target);
1719 DMERR("register failed %d", r);
1720 kmem_cache_destroy(_mpio_cache);
1724 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1726 DMERR("failed to create workqueue kmpathd");
1727 dm_unregister_target(&multipath_target);
1728 kmem_cache_destroy(_mpio_cache);
1733 * A separate workqueue is used to handle the device handlers
1734 * to avoid overloading existing workqueue. Overloading the
1735 * old workqueue would also create a bottleneck in the
1736 * path of the storage hardware device activation.
1738 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1740 if (!kmpath_handlerd) {
1741 DMERR("failed to create workqueue kmpath_handlerd");
1742 destroy_workqueue(kmultipathd);
1743 dm_unregister_target(&multipath_target);
1744 kmem_cache_destroy(_mpio_cache);
1748 DMINFO("version %u.%u.%u loaded",
1749 multipath_target.version[0], multipath_target.version[1],
1750 multipath_target.version[2]);
1755 static void __exit dm_multipath_exit(void)
1757 destroy_workqueue(kmpath_handlerd);
1758 destroy_workqueue(kmultipathd);
1760 dm_unregister_target(&multipath_target);
1761 kmem_cache_destroy(_mpio_cache);
1764 module_init(dm_multipath_init);
1765 module_exit(dm_multipath_exit);
1767 MODULE_DESCRIPTION(DM_NAME " multipath target");
1768 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1769 MODULE_LICENSE("GPL");