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>
11 #include "dm-bio-record.h"
12 #include "dm-path-selector.h"
13 #include "dm-uevent.h"
15 #include <linux/blkdev.h>
16 #include <linux/ctype.h>
17 #include <linux/init.h>
18 #include <linux/mempool.h>
19 #include <linux/module.h>
20 #include <linux/pagemap.h>
21 #include <linux/slab.h>
22 #include <linux/time.h>
23 #include <linux/workqueue.h>
24 #include <linux/delay.h>
25 #include <scsi/scsi_dh.h>
26 #include <linux/atomic.h>
27 #include <linux/blk-mq.h>
29 #define DM_MSG_PREFIX "multipath"
30 #define DM_PG_INIT_DELAY_MSECS 2000
31 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
35 struct list_head list;
37 struct priority_group *pg; /* Owning PG */
38 unsigned fail_count; /* Cumulative failure count */
41 struct delayed_work activate_path;
43 bool is_active:1; /* Path status */
46 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
49 * Paths are grouped into Priority Groups and numbered from 1 upwards.
50 * Each has a path selector which controls which path gets used.
52 struct priority_group {
53 struct list_head list;
55 struct multipath *m; /* Owning multipath instance */
56 struct path_selector ps;
58 unsigned pg_num; /* Reference number */
59 unsigned nr_pgpaths; /* Number of paths in PG */
60 struct list_head pgpaths;
62 bool bypassed:1; /* Temporarily bypass this PG? */
65 /* Multipath context */
67 struct list_head list;
70 const char *hw_handler_name;
71 char *hw_handler_params;
75 unsigned nr_priority_groups;
76 struct list_head priority_groups;
78 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
80 struct pgpath *current_pgpath;
81 struct priority_group *current_pg;
82 struct priority_group *next_pg; /* Switch to this PG if set */
84 unsigned long flags; /* Multipath state flags */
86 unsigned pg_init_retries; /* Number of times to retry pg_init */
87 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
89 atomic_t nr_valid_paths; /* Total number of usable paths */
90 atomic_t pg_init_in_progress; /* Only one pg_init allowed at once */
91 atomic_t pg_init_count; /* Number of times pg_init called */
93 enum dm_queue_mode queue_mode;
95 struct mutex work_mutex;
96 struct work_struct trigger_event;
98 struct work_struct process_queued_bios;
99 struct bio_list queued_bios;
103 * Context information attached to each io we process.
106 struct pgpath *pgpath;
110 typedef int (*action_fn) (struct pgpath *pgpath);
112 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
113 static void trigger_event(struct work_struct *work);
114 static void activate_or_offline_path(struct pgpath *pgpath);
115 static void activate_path_work(struct work_struct *work);
116 static void process_queued_bios(struct work_struct *work);
118 /*-----------------------------------------------
119 * Multipath state flags.
120 *-----------------------------------------------*/
122 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
123 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
124 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
125 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
126 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
127 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
128 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
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 = true;
140 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path_work);
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 spin_lock_init(&m->lock);
196 set_bit(MPATHF_QUEUE_IO, &m->flags);
197 atomic_set(&m->nr_valid_paths, 0);
198 atomic_set(&m->pg_init_in_progress, 0);
199 atomic_set(&m->pg_init_count, 0);
200 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
201 INIT_WORK(&m->trigger_event, trigger_event);
202 init_waitqueue_head(&m->pg_init_wait);
203 mutex_init(&m->work_mutex);
205 m->queue_mode = DM_TYPE_NONE;
214 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
216 if (m->queue_mode == DM_TYPE_NONE) {
218 * Default to request-based.
220 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
221 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
223 m->queue_mode = DM_TYPE_REQUEST_BASED;
224 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
225 INIT_WORK(&m->process_queued_bios, process_queued_bios);
227 * bio-based doesn't support any direct scsi_dh management;
228 * it just discovers if a scsi_dh is attached.
230 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
233 dm_table_set_type(ti->table, m->queue_mode);
238 static void free_multipath(struct multipath *m)
240 struct priority_group *pg, *tmp;
242 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
244 free_priority_group(pg, m->ti);
247 kfree(m->hw_handler_name);
248 kfree(m->hw_handler_params);
252 static struct dm_mpath_io *get_mpio(union map_info *info)
257 static size_t multipath_per_bio_data_size(void)
259 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
262 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
264 return dm_per_bio_data(bio, multipath_per_bio_data_size());
267 static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
269 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
270 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
271 void *bio_details = mpio + 1;
276 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p,
277 struct dm_bio_details **bio_details_p)
279 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
280 struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
282 memset(mpio, 0, sizeof(*mpio));
283 memset(bio_details, 0, sizeof(*bio_details));
284 dm_bio_record(bio_details, bio);
289 *bio_details_p = bio_details;
292 /*-----------------------------------------------
294 *-----------------------------------------------*/
296 static int __pg_init_all_paths(struct multipath *m)
298 struct pgpath *pgpath;
299 unsigned long pg_init_delay = 0;
301 lockdep_assert_held(&m->lock);
303 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
306 atomic_inc(&m->pg_init_count);
307 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
309 /* Check here to reset pg_init_required */
313 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
314 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
315 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
316 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
317 /* Skip failed paths */
318 if (!pgpath->is_active)
320 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
322 atomic_inc(&m->pg_init_in_progress);
324 return atomic_read(&m->pg_init_in_progress);
327 static int pg_init_all_paths(struct multipath *m)
332 spin_lock_irqsave(&m->lock, flags);
333 ret = __pg_init_all_paths(m);
334 spin_unlock_irqrestore(&m->lock, flags);
339 static void __switch_pg(struct multipath *m, struct priority_group *pg)
343 /* Must we initialise the PG first, and queue I/O till it's ready? */
344 if (m->hw_handler_name) {
345 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
346 set_bit(MPATHF_QUEUE_IO, &m->flags);
348 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
349 clear_bit(MPATHF_QUEUE_IO, &m->flags);
352 atomic_set(&m->pg_init_count, 0);
355 static struct pgpath *choose_path_in_pg(struct multipath *m,
356 struct priority_group *pg,
360 struct dm_path *path;
361 struct pgpath *pgpath;
363 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
365 return ERR_PTR(-ENXIO);
367 pgpath = path_to_pgpath(path);
369 if (unlikely(lockless_dereference(m->current_pg) != pg)) {
370 /* Only update current_pgpath if pg changed */
371 spin_lock_irqsave(&m->lock, flags);
372 m->current_pgpath = pgpath;
374 spin_unlock_irqrestore(&m->lock, flags);
380 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
383 struct priority_group *pg;
384 struct pgpath *pgpath;
385 unsigned bypassed = 1;
387 if (!atomic_read(&m->nr_valid_paths)) {
388 clear_bit(MPATHF_QUEUE_IO, &m->flags);
392 /* Were we instructed to switch PG? */
393 if (lockless_dereference(m->next_pg)) {
394 spin_lock_irqsave(&m->lock, flags);
397 spin_unlock_irqrestore(&m->lock, flags);
398 goto check_current_pg;
401 spin_unlock_irqrestore(&m->lock, flags);
402 pgpath = choose_path_in_pg(m, pg, nr_bytes);
403 if (!IS_ERR_OR_NULL(pgpath))
407 /* Don't change PG until it has no remaining paths */
409 pg = lockless_dereference(m->current_pg);
411 pgpath = choose_path_in_pg(m, pg, nr_bytes);
412 if (!IS_ERR_OR_NULL(pgpath))
417 * Loop through priority groups until we find a valid path.
418 * First time we skip PGs marked 'bypassed'.
419 * Second time we only try the ones we skipped, but set
420 * pg_init_delay_retry so we do not hammer controllers.
423 list_for_each_entry(pg, &m->priority_groups, list) {
424 if (pg->bypassed == !!bypassed)
426 pgpath = choose_path_in_pg(m, pg, nr_bytes);
427 if (!IS_ERR_OR_NULL(pgpath)) {
429 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
433 } while (bypassed--);
436 spin_lock_irqsave(&m->lock, flags);
437 m->current_pgpath = NULL;
438 m->current_pg = NULL;
439 spin_unlock_irqrestore(&m->lock, flags);
445 * dm_report_EIO() is a macro instead of a function to make pr_debug()
446 * report the function name and line number of the function from which
447 * it has been invoked.
449 #define dm_report_EIO(m) \
451 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
453 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
454 dm_device_name(md), \
455 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
456 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
457 dm_noflush_suspending((m)->ti)); \
461 * Map cloned requests (request-based multipath)
463 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
464 union map_info *map_context,
465 struct request **__clone)
467 struct multipath *m = ti->private;
468 size_t nr_bytes = blk_rq_bytes(rq);
469 struct pgpath *pgpath;
470 struct block_device *bdev;
471 struct dm_mpath_io *mpio = get_mpio(map_context);
472 struct request_queue *q;
473 struct request *clone;
475 /* Do we need to select a new pgpath? */
476 pgpath = lockless_dereference(m->current_pgpath);
477 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
478 pgpath = choose_pgpath(m, nr_bytes);
481 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
482 return DM_MAPIO_DELAY_REQUEUE;
483 dm_report_EIO(m); /* Failed */
484 return DM_MAPIO_KILL;
485 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
486 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
487 if (pg_init_all_paths(m))
488 return DM_MAPIO_DELAY_REQUEUE;
489 return DM_MAPIO_REQUEUE;
492 memset(mpio, 0, sizeof(*mpio));
493 mpio->pgpath = pgpath;
494 mpio->nr_bytes = nr_bytes;
496 bdev = pgpath->path.dev->bdev;
497 q = bdev_get_queue(bdev);
498 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
500 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
501 bool queue_dying = blk_queue_dying(q);
503 atomic_inc(&m->pg_init_in_progress);
504 activate_or_offline_path(pgpath);
506 return DM_MAPIO_DELAY_REQUEUE;
508 clone->bio = clone->biotail = NULL;
509 clone->rq_disk = bdev->bd_disk;
510 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
513 if (pgpath->pg->ps.type->start_io)
514 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
517 return DM_MAPIO_REMAPPED;
520 static void multipath_release_clone(struct request *clone)
522 blk_put_request(clone);
526 * Map cloned bios (bio-based multipath)
528 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
530 size_t nr_bytes = bio->bi_iter.bi_size;
531 struct pgpath *pgpath;
535 /* Do we need to select a new pgpath? */
536 pgpath = lockless_dereference(m->current_pgpath);
537 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
538 if (!pgpath || !queue_io)
539 pgpath = choose_pgpath(m, nr_bytes);
541 if ((pgpath && queue_io) ||
542 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
543 /* Queue for the daemon to resubmit */
544 spin_lock_irqsave(&m->lock, flags);
545 bio_list_add(&m->queued_bios, bio);
546 spin_unlock_irqrestore(&m->lock, flags);
547 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
548 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
549 pg_init_all_paths(m);
551 queue_work(kmultipathd, &m->process_queued_bios);
552 return DM_MAPIO_SUBMITTED;
556 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
557 return DM_MAPIO_REQUEUE;
559 return DM_MAPIO_KILL;
562 mpio->pgpath = pgpath;
563 mpio->nr_bytes = nr_bytes;
566 bio_set_dev(bio, pgpath->path.dev->bdev);
567 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
569 if (pgpath->pg->ps.type->start_io)
570 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
573 return DM_MAPIO_REMAPPED;
576 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
578 struct multipath *m = ti->private;
579 struct dm_mpath_io *mpio = NULL;
581 multipath_init_per_bio_data(bio, &mpio, NULL);
583 return __multipath_map_bio(m, bio, mpio);
586 static void process_queued_io_list(struct multipath *m)
588 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
589 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
590 else if (m->queue_mode == DM_TYPE_BIO_BASED)
591 queue_work(kmultipathd, &m->process_queued_bios);
594 static void process_queued_bios(struct work_struct *work)
599 struct bio_list bios;
600 struct blk_plug plug;
601 struct multipath *m =
602 container_of(work, struct multipath, process_queued_bios);
604 bio_list_init(&bios);
606 spin_lock_irqsave(&m->lock, flags);
608 if (bio_list_empty(&m->queued_bios)) {
609 spin_unlock_irqrestore(&m->lock, flags);
613 bio_list_merge(&bios, &m->queued_bios);
614 bio_list_init(&m->queued_bios);
616 spin_unlock_irqrestore(&m->lock, flags);
618 blk_start_plug(&plug);
619 while ((bio = bio_list_pop(&bios))) {
620 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
623 bio->bi_status = BLK_STS_IOERR;
626 case DM_MAPIO_REQUEUE:
627 bio->bi_status = BLK_STS_DM_REQUEUE;
630 case DM_MAPIO_REMAPPED:
631 generic_make_request(bio);
636 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
639 blk_finish_plug(&plug);
642 static void assign_bit(bool value, long nr, unsigned long *addr)
651 * If we run out of usable paths, should we queue I/O or error it?
653 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
658 spin_lock_irqsave(&m->lock, flags);
659 assign_bit((save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
660 (!save_old_value && queue_if_no_path),
661 MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
662 assign_bit(queue_if_no_path || dm_noflush_suspending(m->ti),
663 MPATHF_QUEUE_IF_NO_PATH, &m->flags);
664 spin_unlock_irqrestore(&m->lock, flags);
666 if (!queue_if_no_path) {
667 dm_table_run_md_queue_async(m->ti->table);
668 process_queued_io_list(m);
675 * An event is triggered whenever a path is taken out of use.
676 * Includes path failure and PG bypass.
678 static void trigger_event(struct work_struct *work)
680 struct multipath *m =
681 container_of(work, struct multipath, trigger_event);
683 dm_table_event(m->ti->table);
686 /*-----------------------------------------------------------------
687 * Constructor/argument parsing:
688 * <#multipath feature args> [<arg>]*
689 * <#hw_handler args> [hw_handler [<arg>]*]
691 * <initial priority group>
692 * [<selector> <#selector args> [<arg>]*
693 * <#paths> <#per-path selector args>
694 * [<path> [<arg>]* ]+ ]+
695 *---------------------------------------------------------------*/
696 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
697 struct dm_target *ti)
700 struct path_selector_type *pst;
703 static const struct dm_arg _args[] = {
704 {0, 1024, "invalid number of path selector args"},
707 pst = dm_get_path_selector(dm_shift_arg(as));
709 ti->error = "unknown path selector type";
713 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
715 dm_put_path_selector(pst);
719 r = pst->create(&pg->ps, ps_argc, as->argv);
721 dm_put_path_selector(pst);
722 ti->error = "path selector constructor failed";
727 dm_consume_args(as, ps_argc);
732 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
733 struct dm_target *ti)
737 struct multipath *m = ti->private;
738 struct request_queue *q = NULL;
739 const char *attached_handler_name;
741 /* we need at least a path arg */
743 ti->error = "no device given";
744 return ERR_PTR(-EINVAL);
749 return ERR_PTR(-ENOMEM);
751 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
754 ti->error = "error getting device";
758 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
759 q = bdev_get_queue(p->path.dev->bdev);
761 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
763 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
764 if (attached_handler_name) {
766 * Clear any hw_handler_params associated with a
767 * handler that isn't already attached.
769 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
770 kfree(m->hw_handler_params);
771 m->hw_handler_params = NULL;
775 * Reset hw_handler_name to match the attached handler
777 * NB. This modifies the table line to show the actual
778 * handler instead of the original table passed in.
780 kfree(m->hw_handler_name);
781 m->hw_handler_name = attached_handler_name;
785 if (m->hw_handler_name) {
786 r = scsi_dh_attach(q, m->hw_handler_name);
788 char b[BDEVNAME_SIZE];
790 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
791 bdevname(p->path.dev->bdev, b));
795 ti->error = "error attaching hardware handler";
796 dm_put_device(ti, p->path.dev);
800 if (m->hw_handler_params) {
801 r = scsi_dh_set_params(q, m->hw_handler_params);
803 ti->error = "unable to set hardware "
804 "handler parameters";
805 dm_put_device(ti, p->path.dev);
811 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
813 dm_put_device(ti, p->path.dev);
824 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
827 static const struct dm_arg _args[] = {
828 {1, 1024, "invalid number of paths"},
829 {0, 1024, "invalid number of selector args"}
833 unsigned i, nr_selector_args, nr_args;
834 struct priority_group *pg;
835 struct dm_target *ti = m->ti;
839 ti->error = "not enough priority group arguments";
840 return ERR_PTR(-EINVAL);
843 pg = alloc_priority_group();
845 ti->error = "couldn't allocate priority group";
846 return ERR_PTR(-ENOMEM);
850 r = parse_path_selector(as, pg, ti);
857 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
861 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
865 nr_args = 1 + nr_selector_args;
866 for (i = 0; i < pg->nr_pgpaths; i++) {
867 struct pgpath *pgpath;
868 struct dm_arg_set path_args;
870 if (as->argc < nr_args) {
871 ti->error = "not enough path parameters";
876 path_args.argc = nr_args;
877 path_args.argv = as->argv;
879 pgpath = parse_path(&path_args, &pg->ps, ti);
880 if (IS_ERR(pgpath)) {
886 list_add_tail(&pgpath->list, &pg->pgpaths);
887 dm_consume_args(as, nr_args);
893 free_priority_group(pg, ti);
897 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
901 struct dm_target *ti = m->ti;
903 static const struct dm_arg _args[] = {
904 {0, 1024, "invalid number of hardware handler args"},
907 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
913 if (m->queue_mode == DM_TYPE_BIO_BASED) {
914 dm_consume_args(as, hw_argc);
915 DMERR("bio-based multipath doesn't allow hardware handler args");
919 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
920 if (!m->hw_handler_name)
927 for (i = 0; i <= hw_argc - 2; i++)
928 len += strlen(as->argv[i]) + 1;
929 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
931 ti->error = "memory allocation failed";
935 j = sprintf(p, "%d", hw_argc - 1);
936 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
937 j = sprintf(p, "%s", as->argv[i]);
939 dm_consume_args(as, hw_argc - 1);
943 kfree(m->hw_handler_name);
944 m->hw_handler_name = NULL;
948 static int parse_features(struct dm_arg_set *as, struct multipath *m)
952 struct dm_target *ti = m->ti;
953 const char *arg_name;
955 static const struct dm_arg _args[] = {
956 {0, 8, "invalid number of feature args"},
957 {1, 50, "pg_init_retries must be between 1 and 50"},
958 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
961 r = dm_read_arg_group(_args, as, &argc, &ti->error);
969 arg_name = dm_shift_arg(as);
972 if (!strcasecmp(arg_name, "queue_if_no_path")) {
973 r = queue_if_no_path(m, true, false);
977 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
978 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
982 if (!strcasecmp(arg_name, "pg_init_retries") &&
984 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
989 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
991 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
996 if (!strcasecmp(arg_name, "queue_mode") &&
998 const char *queue_mode_name = dm_shift_arg(as);
1000 if (!strcasecmp(queue_mode_name, "bio"))
1001 m->queue_mode = DM_TYPE_BIO_BASED;
1002 else if (!strcasecmp(queue_mode_name, "rq"))
1003 m->queue_mode = DM_TYPE_REQUEST_BASED;
1004 else if (!strcasecmp(queue_mode_name, "mq"))
1005 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1007 ti->error = "Unknown 'queue_mode' requested";
1014 ti->error = "Unrecognised multipath feature request";
1016 } while (argc && !r);
1021 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1023 /* target arguments */
1024 static const struct dm_arg _args[] = {
1025 {0, 1024, "invalid number of priority groups"},
1026 {0, 1024, "invalid initial priority group number"},
1030 struct multipath *m;
1031 struct dm_arg_set as;
1032 unsigned pg_count = 0;
1033 unsigned next_pg_num;
1038 m = alloc_multipath(ti);
1040 ti->error = "can't allocate multipath";
1044 r = parse_features(&as, m);
1048 r = alloc_multipath_stage2(ti, m);
1052 r = parse_hw_handler(&as, m);
1056 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1060 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1064 if ((!m->nr_priority_groups && next_pg_num) ||
1065 (m->nr_priority_groups && !next_pg_num)) {
1066 ti->error = "invalid initial priority group";
1071 /* parse the priority groups */
1073 struct priority_group *pg;
1074 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1076 pg = parse_priority_group(&as, m);
1082 nr_valid_paths += pg->nr_pgpaths;
1083 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1085 list_add_tail(&pg->list, &m->priority_groups);
1087 pg->pg_num = pg_count;
1092 if (pg_count != m->nr_priority_groups) {
1093 ti->error = "priority group count mismatch";
1098 ti->num_flush_bios = 1;
1099 ti->num_discard_bios = 1;
1100 ti->num_write_same_bios = 1;
1101 ti->num_write_zeroes_bios = 1;
1102 if (m->queue_mode == DM_TYPE_BIO_BASED)
1103 ti->per_io_data_size = multipath_per_bio_data_size();
1105 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1114 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1119 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1121 if (!atomic_read(&m->pg_init_in_progress))
1126 finish_wait(&m->pg_init_wait, &wait);
1129 static void flush_multipath_work(struct multipath *m)
1131 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1132 smp_mb__after_atomic();
1134 flush_workqueue(kmpath_handlerd);
1135 multipath_wait_for_pg_init_completion(m);
1136 flush_workqueue(kmultipathd);
1137 flush_work(&m->trigger_event);
1139 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1140 smp_mb__after_atomic();
1143 static void multipath_dtr(struct dm_target *ti)
1145 struct multipath *m = ti->private;
1147 flush_multipath_work(m);
1152 * Take a path out of use.
1154 static int fail_path(struct pgpath *pgpath)
1156 unsigned long flags;
1157 struct multipath *m = pgpath->pg->m;
1159 spin_lock_irqsave(&m->lock, flags);
1161 if (!pgpath->is_active)
1164 DMWARN("Failing path %s.", pgpath->path.dev->name);
1166 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1167 pgpath->is_active = false;
1168 pgpath->fail_count++;
1170 atomic_dec(&m->nr_valid_paths);
1172 if (pgpath == m->current_pgpath)
1173 m->current_pgpath = NULL;
1175 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1176 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1178 schedule_work(&m->trigger_event);
1181 spin_unlock_irqrestore(&m->lock, flags);
1187 * Reinstate a previously-failed path
1189 static int reinstate_path(struct pgpath *pgpath)
1191 int r = 0, run_queue = 0;
1192 unsigned long flags;
1193 struct multipath *m = pgpath->pg->m;
1194 unsigned nr_valid_paths;
1196 spin_lock_irqsave(&m->lock, flags);
1198 if (pgpath->is_active)
1201 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1203 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1207 pgpath->is_active = true;
1209 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1210 if (nr_valid_paths == 1) {
1211 m->current_pgpath = NULL;
1213 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1214 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1215 atomic_inc(&m->pg_init_in_progress);
1218 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1219 pgpath->path.dev->name, nr_valid_paths);
1221 schedule_work(&m->trigger_event);
1224 spin_unlock_irqrestore(&m->lock, flags);
1226 dm_table_run_md_queue_async(m->ti->table);
1227 process_queued_io_list(m);
1234 * Fail or reinstate all paths that match the provided struct dm_dev.
1236 static int action_dev(struct multipath *m, struct dm_dev *dev,
1240 struct pgpath *pgpath;
1241 struct priority_group *pg;
1243 list_for_each_entry(pg, &m->priority_groups, list) {
1244 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1245 if (pgpath->path.dev == dev)
1254 * Temporarily try to avoid having to use the specified PG
1256 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1259 unsigned long flags;
1261 spin_lock_irqsave(&m->lock, flags);
1263 pg->bypassed = bypassed;
1264 m->current_pgpath = NULL;
1265 m->current_pg = NULL;
1267 spin_unlock_irqrestore(&m->lock, flags);
1269 schedule_work(&m->trigger_event);
1273 * Switch to using the specified PG from the next I/O that gets mapped
1275 static int switch_pg_num(struct multipath *m, const char *pgstr)
1277 struct priority_group *pg;
1279 unsigned long flags;
1282 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1283 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1284 DMWARN("invalid PG number supplied to switch_pg_num");
1288 spin_lock_irqsave(&m->lock, flags);
1289 list_for_each_entry(pg, &m->priority_groups, list) {
1290 pg->bypassed = false;
1294 m->current_pgpath = NULL;
1295 m->current_pg = NULL;
1298 spin_unlock_irqrestore(&m->lock, flags);
1300 schedule_work(&m->trigger_event);
1305 * Set/clear bypassed status of a PG.
1306 * PGs are numbered upwards from 1 in the order they were declared.
1308 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1310 struct priority_group *pg;
1314 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1315 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1316 DMWARN("invalid PG number supplied to bypass_pg");
1320 list_for_each_entry(pg, &m->priority_groups, list) {
1325 bypass_pg(m, pg, bypassed);
1330 * Should we retry pg_init immediately?
1332 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1334 unsigned long flags;
1335 bool limit_reached = false;
1337 spin_lock_irqsave(&m->lock, flags);
1339 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1340 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1341 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1343 limit_reached = true;
1345 spin_unlock_irqrestore(&m->lock, flags);
1347 return limit_reached;
1350 static void pg_init_done(void *data, int errors)
1352 struct pgpath *pgpath = data;
1353 struct priority_group *pg = pgpath->pg;
1354 struct multipath *m = pg->m;
1355 unsigned long flags;
1356 bool delay_retry = false;
1358 /* device or driver problems */
1363 if (!m->hw_handler_name) {
1367 DMERR("Could not failover the device: Handler scsi_dh_%s "
1368 "Error %d.", m->hw_handler_name, errors);
1370 * Fail path for now, so we do not ping pong
1374 case SCSI_DH_DEV_TEMP_BUSY:
1376 * Probably doing something like FW upgrade on the
1377 * controller so try the other pg.
1379 bypass_pg(m, pg, true);
1382 /* Wait before retrying. */
1385 case SCSI_DH_IMM_RETRY:
1386 case SCSI_DH_RES_TEMP_UNAVAIL:
1387 if (pg_init_limit_reached(m, pgpath))
1391 case SCSI_DH_DEV_OFFLINED:
1394 * We probably do not want to fail the path for a device
1395 * error, but this is what the old dm did. In future
1396 * patches we can do more advanced handling.
1401 spin_lock_irqsave(&m->lock, flags);
1403 if (pgpath == m->current_pgpath) {
1404 DMERR("Could not failover device. Error %d.", errors);
1405 m->current_pgpath = NULL;
1406 m->current_pg = NULL;
1408 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1409 pg->bypassed = false;
1411 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1412 /* Activations of other paths are still on going */
1415 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1417 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1419 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1421 if (__pg_init_all_paths(m))
1424 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1426 process_queued_io_list(m);
1429 * Wake up any thread waiting to suspend.
1431 wake_up(&m->pg_init_wait);
1434 spin_unlock_irqrestore(&m->lock, flags);
1437 static void activate_or_offline_path(struct pgpath *pgpath)
1439 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1441 if (pgpath->is_active && !blk_queue_dying(q))
1442 scsi_dh_activate(q, pg_init_done, pgpath);
1444 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1447 static void activate_path_work(struct work_struct *work)
1449 struct pgpath *pgpath =
1450 container_of(work, struct pgpath, activate_path.work);
1452 activate_or_offline_path(pgpath);
1455 static int noretry_error(blk_status_t error)
1458 case BLK_STS_NOTSUPP:
1460 case BLK_STS_TARGET:
1462 case BLK_STS_MEDIUM:
1466 /* Anything else could be a path failure, so should be retried */
1470 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1471 blk_status_t error, union map_info *map_context)
1473 struct dm_mpath_io *mpio = get_mpio(map_context);
1474 struct pgpath *pgpath = mpio->pgpath;
1475 int r = DM_ENDIO_DONE;
1478 * We don't queue any clone request inside the multipath target
1479 * during end I/O handling, since those clone requests don't have
1480 * bio clones. If we queue them inside the multipath target,
1481 * we need to make bio clones, that requires memory allocation.
1482 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1483 * don't have bio clones.)
1484 * Instead of queueing the clone request here, we queue the original
1485 * request into dm core, which will remake a clone request and
1486 * clone bios for it and resubmit it later.
1488 if (error && !noretry_error(error)) {
1489 struct multipath *m = ti->private;
1491 r = DM_ENDIO_REQUEUE;
1496 if (atomic_read(&m->nr_valid_paths) == 0 &&
1497 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1498 if (error == BLK_STS_IOERR)
1500 /* complete with the original error */
1506 struct path_selector *ps = &pgpath->pg->ps;
1508 if (ps->type->end_io)
1509 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1515 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1516 blk_status_t *error)
1518 struct multipath *m = ti->private;
1519 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1520 struct pgpath *pgpath = mpio->pgpath;
1521 unsigned long flags;
1522 int r = DM_ENDIO_DONE;
1524 if (!*error || noretry_error(*error))
1530 if (atomic_read(&m->nr_valid_paths) == 0 &&
1531 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1533 *error = BLK_STS_IOERR;
1537 /* Queue for the daemon to resubmit */
1538 dm_bio_restore(get_bio_details_from_bio(clone), clone);
1540 spin_lock_irqsave(&m->lock, flags);
1541 bio_list_add(&m->queued_bios, clone);
1542 spin_unlock_irqrestore(&m->lock, flags);
1543 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1544 queue_work(kmultipathd, &m->process_queued_bios);
1546 r = DM_ENDIO_INCOMPLETE;
1549 struct path_selector *ps = &pgpath->pg->ps;
1551 if (ps->type->end_io)
1552 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1559 * Suspend can't complete until all the I/O is processed so if
1560 * the last path fails we must error any remaining I/O.
1561 * Note that if the freeze_bdev fails while suspending, the
1562 * queue_if_no_path state is lost - userspace should reset it.
1564 static void multipath_presuspend(struct dm_target *ti)
1566 struct multipath *m = ti->private;
1568 queue_if_no_path(m, false, true);
1571 static void multipath_postsuspend(struct dm_target *ti)
1573 struct multipath *m = ti->private;
1575 mutex_lock(&m->work_mutex);
1576 flush_multipath_work(m);
1577 mutex_unlock(&m->work_mutex);
1581 * Restore the queue_if_no_path setting.
1583 static void multipath_resume(struct dm_target *ti)
1585 struct multipath *m = ti->private;
1586 unsigned long flags;
1588 spin_lock_irqsave(&m->lock, flags);
1589 assign_bit(test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
1590 MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1591 spin_unlock_irqrestore(&m->lock, flags);
1595 * Info output has the following format:
1596 * num_multipath_feature_args [multipath_feature_args]*
1597 * num_handler_status_args [handler_status_args]*
1598 * num_groups init_group_number
1599 * [A|D|E num_ps_status_args [ps_status_args]*
1600 * num_paths num_selector_args
1601 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1603 * Table output has the following format (identical to the constructor string):
1604 * num_feature_args [features_args]*
1605 * num_handler_args hw_handler [hw_handler_args]*
1606 * num_groups init_group_number
1607 * [priority selector-name num_ps_args [ps_args]*
1608 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1610 static void multipath_status(struct dm_target *ti, status_type_t type,
1611 unsigned status_flags, char *result, unsigned maxlen)
1614 unsigned long flags;
1615 struct multipath *m = ti->private;
1616 struct priority_group *pg;
1621 spin_lock_irqsave(&m->lock, flags);
1624 if (type == STATUSTYPE_INFO)
1625 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1626 atomic_read(&m->pg_init_count));
1628 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1629 (m->pg_init_retries > 0) * 2 +
1630 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1631 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1632 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1634 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1635 DMEMIT("queue_if_no_path ");
1636 if (m->pg_init_retries)
1637 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1638 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1639 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1640 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1641 DMEMIT("retain_attached_hw_handler ");
1642 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1643 switch(m->queue_mode) {
1644 case DM_TYPE_BIO_BASED:
1645 DMEMIT("queue_mode bio ");
1647 case DM_TYPE_MQ_REQUEST_BASED:
1648 DMEMIT("queue_mode mq ");
1657 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1660 DMEMIT("1 %s ", m->hw_handler_name);
1662 DMEMIT("%u ", m->nr_priority_groups);
1665 pg_num = m->next_pg->pg_num;
1666 else if (m->current_pg)
1667 pg_num = m->current_pg->pg_num;
1669 pg_num = (m->nr_priority_groups ? 1 : 0);
1671 DMEMIT("%u ", pg_num);
1674 case STATUSTYPE_INFO:
1675 list_for_each_entry(pg, &m->priority_groups, list) {
1677 state = 'D'; /* Disabled */
1678 else if (pg == m->current_pg)
1679 state = 'A'; /* Currently Active */
1681 state = 'E'; /* Enabled */
1683 DMEMIT("%c ", state);
1685 if (pg->ps.type->status)
1686 sz += pg->ps.type->status(&pg->ps, NULL, type,
1692 DMEMIT("%u %u ", pg->nr_pgpaths,
1693 pg->ps.type->info_args);
1695 list_for_each_entry(p, &pg->pgpaths, list) {
1696 DMEMIT("%s %s %u ", p->path.dev->name,
1697 p->is_active ? "A" : "F",
1699 if (pg->ps.type->status)
1700 sz += pg->ps.type->status(&pg->ps,
1701 &p->path, type, result + sz,
1707 case STATUSTYPE_TABLE:
1708 list_for_each_entry(pg, &m->priority_groups, list) {
1709 DMEMIT("%s ", pg->ps.type->name);
1711 if (pg->ps.type->status)
1712 sz += pg->ps.type->status(&pg->ps, NULL, type,
1718 DMEMIT("%u %u ", pg->nr_pgpaths,
1719 pg->ps.type->table_args);
1721 list_for_each_entry(p, &pg->pgpaths, list) {
1722 DMEMIT("%s ", p->path.dev->name);
1723 if (pg->ps.type->status)
1724 sz += pg->ps.type->status(&pg->ps,
1725 &p->path, type, result + sz,
1732 spin_unlock_irqrestore(&m->lock, flags);
1735 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1739 struct multipath *m = ti->private;
1742 mutex_lock(&m->work_mutex);
1744 if (dm_suspended(ti)) {
1750 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1751 r = queue_if_no_path(m, true, false);
1753 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1754 r = queue_if_no_path(m, false, false);
1760 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1764 if (!strcasecmp(argv[0], "disable_group")) {
1765 r = bypass_pg_num(m, argv[1], true);
1767 } else if (!strcasecmp(argv[0], "enable_group")) {
1768 r = bypass_pg_num(m, argv[1], false);
1770 } else if (!strcasecmp(argv[0], "switch_group")) {
1771 r = switch_pg_num(m, argv[1]);
1773 } else if (!strcasecmp(argv[0], "reinstate_path"))
1774 action = reinstate_path;
1775 else if (!strcasecmp(argv[0], "fail_path"))
1778 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1782 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1784 DMWARN("message: error getting device %s",
1789 r = action_dev(m, dev, action);
1791 dm_put_device(ti, dev);
1794 mutex_unlock(&m->work_mutex);
1798 static int multipath_prepare_ioctl(struct dm_target *ti,
1799 struct block_device **bdev, fmode_t *mode)
1801 struct multipath *m = ti->private;
1802 struct pgpath *current_pgpath;
1805 current_pgpath = lockless_dereference(m->current_pgpath);
1806 if (!current_pgpath)
1807 current_pgpath = choose_pgpath(m, 0);
1809 if (current_pgpath) {
1810 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1811 *bdev = current_pgpath->path.dev->bdev;
1812 *mode = current_pgpath->path.dev->mode;
1815 /* pg_init has not started or completed */
1819 /* No path is available */
1820 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1826 if (r == -ENOTCONN) {
1827 if (!lockless_dereference(m->current_pg)) {
1828 /* Path status changed, redo selection */
1829 (void) choose_pgpath(m, 0);
1831 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1832 pg_init_all_paths(m);
1833 dm_table_run_md_queue_async(m->ti->table);
1834 process_queued_io_list(m);
1838 * Only pass ioctls through if the device sizes match exactly.
1840 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1845 static int multipath_iterate_devices(struct dm_target *ti,
1846 iterate_devices_callout_fn fn, void *data)
1848 struct multipath *m = ti->private;
1849 struct priority_group *pg;
1853 list_for_each_entry(pg, &m->priority_groups, list) {
1854 list_for_each_entry(p, &pg->pgpaths, list) {
1855 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1865 static int pgpath_busy(struct pgpath *pgpath)
1867 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1869 return blk_lld_busy(q);
1873 * We return "busy", only when we can map I/Os but underlying devices
1874 * are busy (so even if we map I/Os now, the I/Os will wait on
1875 * the underlying queue).
1876 * In other words, if we want to kill I/Os or queue them inside us
1877 * due to map unavailability, we don't return "busy". Otherwise,
1878 * dm core won't give us the I/Os and we can't do what we want.
1880 static int multipath_busy(struct dm_target *ti)
1882 bool busy = false, has_active = false;
1883 struct multipath *m = ti->private;
1884 struct priority_group *pg, *next_pg;
1885 struct pgpath *pgpath;
1887 /* pg_init in progress */
1888 if (atomic_read(&m->pg_init_in_progress))
1891 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1892 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1893 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1895 /* Guess which priority_group will be used at next mapping time */
1896 pg = lockless_dereference(m->current_pg);
1897 next_pg = lockless_dereference(m->next_pg);
1898 if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
1903 * We don't know which pg will be used at next mapping time.
1904 * We don't call choose_pgpath() here to avoid to trigger
1905 * pg_init just by busy checking.
1906 * So we don't know whether underlying devices we will be using
1907 * at next mapping time are busy or not. Just try mapping.
1913 * If there is one non-busy active path at least, the path selector
1914 * will be able to select it. So we consider such a pg as not busy.
1917 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1918 if (pgpath->is_active) {
1920 if (!pgpath_busy(pgpath)) {
1929 * No active path in this pg, so this pg won't be used and
1930 * the current_pg will be changed at next mapping time.
1931 * We need to try mapping to determine it.
1939 /*-----------------------------------------------------------------
1941 *---------------------------------------------------------------*/
1942 static struct target_type multipath_target = {
1943 .name = "multipath",
1944 .version = {1, 12, 0},
1945 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1946 .module = THIS_MODULE,
1947 .ctr = multipath_ctr,
1948 .dtr = multipath_dtr,
1949 .clone_and_map_rq = multipath_clone_and_map,
1950 .release_clone_rq = multipath_release_clone,
1951 .rq_end_io = multipath_end_io,
1952 .map = multipath_map_bio,
1953 .end_io = multipath_end_io_bio,
1954 .presuspend = multipath_presuspend,
1955 .postsuspend = multipath_postsuspend,
1956 .resume = multipath_resume,
1957 .status = multipath_status,
1958 .message = multipath_message,
1959 .prepare_ioctl = multipath_prepare_ioctl,
1960 .iterate_devices = multipath_iterate_devices,
1961 .busy = multipath_busy,
1964 static int __init dm_multipath_init(void)
1968 r = dm_register_target(&multipath_target);
1970 DMERR("request-based register failed %d", r);
1972 goto bad_register_target;
1975 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1977 DMERR("failed to create workqueue kmpathd");
1979 goto bad_alloc_kmultipathd;
1983 * A separate workqueue is used to handle the device handlers
1984 * to avoid overloading existing workqueue. Overloading the
1985 * old workqueue would also create a bottleneck in the
1986 * path of the storage hardware device activation.
1988 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1990 if (!kmpath_handlerd) {
1991 DMERR("failed to create workqueue kmpath_handlerd");
1993 goto bad_alloc_kmpath_handlerd;
1998 bad_alloc_kmpath_handlerd:
1999 destroy_workqueue(kmultipathd);
2000 bad_alloc_kmultipathd:
2001 dm_unregister_target(&multipath_target);
2002 bad_register_target:
2006 static void __exit dm_multipath_exit(void)
2008 destroy_workqueue(kmpath_handlerd);
2009 destroy_workqueue(kmultipathd);
2011 dm_unregister_target(&multipath_target);
2014 module_init(dm_multipath_init);
2015 module_exit(dm_multipath_exit);
2017 MODULE_DESCRIPTION(DM_NAME " multipath target");
2018 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2019 MODULE_LICENSE("GPL");