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 * Check whether bios must be queued in the device-mapper core rather
446 * than here in the target.
448 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
449 * same value then we are not between multipath_presuspend()
450 * and multipath_resume() calls and we have no need to check
451 * for the DMF_NOFLUSH_SUSPENDING flag.
453 static bool __must_push_back(struct multipath *m)
455 return ((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));
460 static bool must_push_back_rq(struct multipath *m)
465 spin_lock_irqsave(&m->lock, flags);
466 r = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) ||
467 __must_push_back(m));
468 spin_unlock_irqrestore(&m->lock, flags);
473 static bool must_push_back_bio(struct multipath *m)
478 spin_lock_irqsave(&m->lock, flags);
479 r = __must_push_back(m);
480 spin_unlock_irqrestore(&m->lock, flags);
486 * Map cloned requests (request-based multipath)
488 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
489 union map_info *map_context,
490 struct request **__clone)
492 struct multipath *m = ti->private;
493 size_t nr_bytes = blk_rq_bytes(rq);
494 struct pgpath *pgpath;
495 struct block_device *bdev;
496 struct dm_mpath_io *mpio = get_mpio(map_context);
497 struct request_queue *q;
498 struct request *clone;
500 /* Do we need to select a new pgpath? */
501 pgpath = lockless_dereference(m->current_pgpath);
502 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
503 pgpath = choose_pgpath(m, nr_bytes);
506 if (must_push_back_rq(m))
507 return DM_MAPIO_DELAY_REQUEUE;
508 return -EIO; /* Failed */
509 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
510 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
511 if (pg_init_all_paths(m))
512 return DM_MAPIO_DELAY_REQUEUE;
513 return DM_MAPIO_REQUEUE;
516 memset(mpio, 0, sizeof(*mpio));
517 mpio->pgpath = pgpath;
518 mpio->nr_bytes = nr_bytes;
520 bdev = pgpath->path.dev->bdev;
521 q = bdev_get_queue(bdev);
522 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
524 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
525 bool queue_dying = blk_queue_dying(q);
526 DMERR_LIMIT("blk_get_request() returned %ld%s - requeuing",
527 PTR_ERR(clone), queue_dying ? " (path offline)" : "");
529 atomic_inc(&m->pg_init_in_progress);
530 activate_or_offline_path(pgpath);
531 return DM_MAPIO_REQUEUE;
533 return DM_MAPIO_DELAY_REQUEUE;
535 clone->bio = clone->biotail = NULL;
536 clone->rq_disk = bdev->bd_disk;
537 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
540 if (pgpath->pg->ps.type->start_io)
541 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
544 return DM_MAPIO_REMAPPED;
547 static void multipath_release_clone(struct request *clone)
549 blk_put_request(clone);
553 * Map cloned bios (bio-based multipath)
555 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
557 size_t nr_bytes = bio->bi_iter.bi_size;
558 struct pgpath *pgpath;
562 /* Do we need to select a new pgpath? */
563 pgpath = lockless_dereference(m->current_pgpath);
564 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
565 if (!pgpath || !queue_io)
566 pgpath = choose_pgpath(m, nr_bytes);
568 if ((pgpath && queue_io) ||
569 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
570 /* Queue for the daemon to resubmit */
571 spin_lock_irqsave(&m->lock, flags);
572 bio_list_add(&m->queued_bios, bio);
573 spin_unlock_irqrestore(&m->lock, flags);
574 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
575 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
576 pg_init_all_paths(m);
578 queue_work(kmultipathd, &m->process_queued_bios);
579 return DM_MAPIO_SUBMITTED;
583 if (!must_push_back_bio(m))
585 return DM_MAPIO_REQUEUE;
588 mpio->pgpath = pgpath;
589 mpio->nr_bytes = nr_bytes;
592 bio->bi_bdev = pgpath->path.dev->bdev;
593 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
595 if (pgpath->pg->ps.type->start_io)
596 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
599 return DM_MAPIO_REMAPPED;
602 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
604 struct multipath *m = ti->private;
605 struct dm_mpath_io *mpio = NULL;
607 multipath_init_per_bio_data(bio, &mpio, NULL);
609 return __multipath_map_bio(m, bio, mpio);
612 static void process_queued_io_list(struct multipath *m)
614 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
615 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
616 else if (m->queue_mode == DM_TYPE_BIO_BASED)
617 queue_work(kmultipathd, &m->process_queued_bios);
620 static void process_queued_bios(struct work_struct *work)
625 struct bio_list bios;
626 struct blk_plug plug;
627 struct multipath *m =
628 container_of(work, struct multipath, process_queued_bios);
630 bio_list_init(&bios);
632 spin_lock_irqsave(&m->lock, flags);
634 if (bio_list_empty(&m->queued_bios)) {
635 spin_unlock_irqrestore(&m->lock, flags);
639 bio_list_merge(&bios, &m->queued_bios);
640 bio_list_init(&m->queued_bios);
642 spin_unlock_irqrestore(&m->lock, flags);
644 blk_start_plug(&plug);
645 while ((bio = bio_list_pop(&bios))) {
646 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
647 if (r < 0 || r == DM_MAPIO_REQUEUE) {
650 } else if (r == DM_MAPIO_REMAPPED)
651 generic_make_request(bio);
653 blk_finish_plug(&plug);
657 * If we run out of usable paths, should we queue I/O or error it?
659 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
664 spin_lock_irqsave(&m->lock, flags);
666 if (save_old_value) {
667 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
668 set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
670 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
672 if (queue_if_no_path)
673 set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
675 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
677 if (queue_if_no_path)
678 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
680 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
682 spin_unlock_irqrestore(&m->lock, flags);
684 if (!queue_if_no_path) {
685 dm_table_run_md_queue_async(m->ti->table);
686 process_queued_io_list(m);
693 * An event is triggered whenever a path is taken out of use.
694 * Includes path failure and PG bypass.
696 static void trigger_event(struct work_struct *work)
698 struct multipath *m =
699 container_of(work, struct multipath, trigger_event);
701 dm_table_event(m->ti->table);
704 /*-----------------------------------------------------------------
705 * Constructor/argument parsing:
706 * <#multipath feature args> [<arg>]*
707 * <#hw_handler args> [hw_handler [<arg>]*]
709 * <initial priority group>
710 * [<selector> <#selector args> [<arg>]*
711 * <#paths> <#per-path selector args>
712 * [<path> [<arg>]* ]+ ]+
713 *---------------------------------------------------------------*/
714 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
715 struct dm_target *ti)
718 struct path_selector_type *pst;
721 static struct dm_arg _args[] = {
722 {0, 1024, "invalid number of path selector args"},
725 pst = dm_get_path_selector(dm_shift_arg(as));
727 ti->error = "unknown path selector type";
731 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
733 dm_put_path_selector(pst);
737 r = pst->create(&pg->ps, ps_argc, as->argv);
739 dm_put_path_selector(pst);
740 ti->error = "path selector constructor failed";
745 dm_consume_args(as, ps_argc);
750 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
751 struct dm_target *ti)
755 struct multipath *m = ti->private;
756 struct request_queue *q = NULL;
757 const char *attached_handler_name;
759 /* we need at least a path arg */
761 ti->error = "no device given";
762 return ERR_PTR(-EINVAL);
767 return ERR_PTR(-ENOMEM);
769 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
772 ti->error = "error getting device";
776 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
777 q = bdev_get_queue(p->path.dev->bdev);
779 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
781 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
782 if (attached_handler_name) {
784 * Clear any hw_handler_params associated with a
785 * handler that isn't already attached.
787 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
788 kfree(m->hw_handler_params);
789 m->hw_handler_params = NULL;
793 * Reset hw_handler_name to match the attached handler
795 * NB. This modifies the table line to show the actual
796 * handler instead of the original table passed in.
798 kfree(m->hw_handler_name);
799 m->hw_handler_name = attached_handler_name;
803 if (m->hw_handler_name) {
804 r = scsi_dh_attach(q, m->hw_handler_name);
806 char b[BDEVNAME_SIZE];
808 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
809 bdevname(p->path.dev->bdev, b));
813 ti->error = "error attaching hardware handler";
814 dm_put_device(ti, p->path.dev);
818 if (m->hw_handler_params) {
819 r = scsi_dh_set_params(q, m->hw_handler_params);
821 ti->error = "unable to set hardware "
822 "handler parameters";
823 dm_put_device(ti, p->path.dev);
829 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
831 dm_put_device(ti, p->path.dev);
842 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
845 static struct dm_arg _args[] = {
846 {1, 1024, "invalid number of paths"},
847 {0, 1024, "invalid number of selector args"}
851 unsigned i, nr_selector_args, nr_args;
852 struct priority_group *pg;
853 struct dm_target *ti = m->ti;
857 ti->error = "not enough priority group arguments";
858 return ERR_PTR(-EINVAL);
861 pg = alloc_priority_group();
863 ti->error = "couldn't allocate priority group";
864 return ERR_PTR(-ENOMEM);
868 r = parse_path_selector(as, pg, ti);
875 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
879 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
883 nr_args = 1 + nr_selector_args;
884 for (i = 0; i < pg->nr_pgpaths; i++) {
885 struct pgpath *pgpath;
886 struct dm_arg_set path_args;
888 if (as->argc < nr_args) {
889 ti->error = "not enough path parameters";
894 path_args.argc = nr_args;
895 path_args.argv = as->argv;
897 pgpath = parse_path(&path_args, &pg->ps, ti);
898 if (IS_ERR(pgpath)) {
904 list_add_tail(&pgpath->list, &pg->pgpaths);
905 dm_consume_args(as, nr_args);
911 free_priority_group(pg, ti);
915 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
919 struct dm_target *ti = m->ti;
921 static struct dm_arg _args[] = {
922 {0, 1024, "invalid number of hardware handler args"},
925 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
931 if (m->queue_mode == DM_TYPE_BIO_BASED) {
932 dm_consume_args(as, hw_argc);
933 DMERR("bio-based multipath doesn't allow hardware handler args");
937 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
938 if (!m->hw_handler_name)
945 for (i = 0; i <= hw_argc - 2; i++)
946 len += strlen(as->argv[i]) + 1;
947 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
949 ti->error = "memory allocation failed";
953 j = sprintf(p, "%d", hw_argc - 1);
954 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
955 j = sprintf(p, "%s", as->argv[i]);
957 dm_consume_args(as, hw_argc - 1);
961 kfree(m->hw_handler_name);
962 m->hw_handler_name = NULL;
966 static int parse_features(struct dm_arg_set *as, struct multipath *m)
970 struct dm_target *ti = m->ti;
971 const char *arg_name;
973 static struct dm_arg _args[] = {
974 {0, 8, "invalid number of feature args"},
975 {1, 50, "pg_init_retries must be between 1 and 50"},
976 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
979 r = dm_read_arg_group(_args, as, &argc, &ti->error);
987 arg_name = dm_shift_arg(as);
990 if (!strcasecmp(arg_name, "queue_if_no_path")) {
991 r = queue_if_no_path(m, true, false);
995 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
996 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1000 if (!strcasecmp(arg_name, "pg_init_retries") &&
1002 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1007 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1009 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1014 if (!strcasecmp(arg_name, "queue_mode") &&
1016 const char *queue_mode_name = dm_shift_arg(as);
1018 if (!strcasecmp(queue_mode_name, "bio"))
1019 m->queue_mode = DM_TYPE_BIO_BASED;
1020 else if (!strcasecmp(queue_mode_name, "rq"))
1021 m->queue_mode = DM_TYPE_REQUEST_BASED;
1022 else if (!strcasecmp(queue_mode_name, "mq"))
1023 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1025 ti->error = "Unknown 'queue_mode' requested";
1032 ti->error = "Unrecognised multipath feature request";
1034 } while (argc && !r);
1039 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1041 /* target arguments */
1042 static struct dm_arg _args[] = {
1043 {0, 1024, "invalid number of priority groups"},
1044 {0, 1024, "invalid initial priority group number"},
1048 struct multipath *m;
1049 struct dm_arg_set as;
1050 unsigned pg_count = 0;
1051 unsigned next_pg_num;
1056 m = alloc_multipath(ti);
1058 ti->error = "can't allocate multipath";
1062 r = parse_features(&as, m);
1066 r = alloc_multipath_stage2(ti, m);
1070 r = parse_hw_handler(&as, m);
1074 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1078 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1082 if ((!m->nr_priority_groups && next_pg_num) ||
1083 (m->nr_priority_groups && !next_pg_num)) {
1084 ti->error = "invalid initial priority group";
1089 /* parse the priority groups */
1091 struct priority_group *pg;
1092 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1094 pg = parse_priority_group(&as, m);
1100 nr_valid_paths += pg->nr_pgpaths;
1101 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1103 list_add_tail(&pg->list, &m->priority_groups);
1105 pg->pg_num = pg_count;
1110 if (pg_count != m->nr_priority_groups) {
1111 ti->error = "priority group count mismatch";
1116 ti->num_flush_bios = 1;
1117 ti->num_discard_bios = 1;
1118 ti->num_write_same_bios = 1;
1119 if (m->queue_mode == DM_TYPE_BIO_BASED)
1120 ti->per_io_data_size = multipath_per_bio_data_size();
1122 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1131 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1136 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1138 if (!atomic_read(&m->pg_init_in_progress))
1143 finish_wait(&m->pg_init_wait, &wait);
1146 static void flush_multipath_work(struct multipath *m)
1148 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1149 smp_mb__after_atomic();
1151 flush_workqueue(kmpath_handlerd);
1152 multipath_wait_for_pg_init_completion(m);
1153 flush_workqueue(kmultipathd);
1154 flush_work(&m->trigger_event);
1156 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1157 smp_mb__after_atomic();
1160 static void multipath_dtr(struct dm_target *ti)
1162 struct multipath *m = ti->private;
1164 flush_multipath_work(m);
1169 * Take a path out of use.
1171 static int fail_path(struct pgpath *pgpath)
1173 unsigned long flags;
1174 struct multipath *m = pgpath->pg->m;
1176 spin_lock_irqsave(&m->lock, flags);
1178 if (!pgpath->is_active)
1181 DMWARN("Failing path %s.", pgpath->path.dev->name);
1183 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1184 pgpath->is_active = false;
1185 pgpath->fail_count++;
1187 atomic_dec(&m->nr_valid_paths);
1189 if (pgpath == m->current_pgpath)
1190 m->current_pgpath = NULL;
1192 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1193 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1195 schedule_work(&m->trigger_event);
1198 spin_unlock_irqrestore(&m->lock, flags);
1204 * Reinstate a previously-failed path
1206 static int reinstate_path(struct pgpath *pgpath)
1208 int r = 0, run_queue = 0;
1209 unsigned long flags;
1210 struct multipath *m = pgpath->pg->m;
1211 unsigned nr_valid_paths;
1213 spin_lock_irqsave(&m->lock, flags);
1215 if (pgpath->is_active)
1218 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1220 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1224 pgpath->is_active = true;
1226 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1227 if (nr_valid_paths == 1) {
1228 m->current_pgpath = NULL;
1230 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1231 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1232 atomic_inc(&m->pg_init_in_progress);
1235 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1236 pgpath->path.dev->name, nr_valid_paths);
1238 schedule_work(&m->trigger_event);
1241 spin_unlock_irqrestore(&m->lock, flags);
1243 dm_table_run_md_queue_async(m->ti->table);
1244 process_queued_io_list(m);
1251 * Fail or reinstate all paths that match the provided struct dm_dev.
1253 static int action_dev(struct multipath *m, struct dm_dev *dev,
1257 struct pgpath *pgpath;
1258 struct priority_group *pg;
1260 list_for_each_entry(pg, &m->priority_groups, list) {
1261 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1262 if (pgpath->path.dev == dev)
1271 * Temporarily try to avoid having to use the specified PG
1273 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1276 unsigned long flags;
1278 spin_lock_irqsave(&m->lock, flags);
1280 pg->bypassed = bypassed;
1281 m->current_pgpath = NULL;
1282 m->current_pg = NULL;
1284 spin_unlock_irqrestore(&m->lock, flags);
1286 schedule_work(&m->trigger_event);
1290 * Switch to using the specified PG from the next I/O that gets mapped
1292 static int switch_pg_num(struct multipath *m, const char *pgstr)
1294 struct priority_group *pg;
1296 unsigned long flags;
1299 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1300 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1301 DMWARN("invalid PG number supplied to switch_pg_num");
1305 spin_lock_irqsave(&m->lock, flags);
1306 list_for_each_entry(pg, &m->priority_groups, list) {
1307 pg->bypassed = false;
1311 m->current_pgpath = NULL;
1312 m->current_pg = NULL;
1315 spin_unlock_irqrestore(&m->lock, flags);
1317 schedule_work(&m->trigger_event);
1322 * Set/clear bypassed status of a PG.
1323 * PGs are numbered upwards from 1 in the order they were declared.
1325 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1327 struct priority_group *pg;
1331 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1332 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1333 DMWARN("invalid PG number supplied to bypass_pg");
1337 list_for_each_entry(pg, &m->priority_groups, list) {
1342 bypass_pg(m, pg, bypassed);
1347 * Should we retry pg_init immediately?
1349 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1351 unsigned long flags;
1352 bool limit_reached = false;
1354 spin_lock_irqsave(&m->lock, flags);
1356 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1357 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1358 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1360 limit_reached = true;
1362 spin_unlock_irqrestore(&m->lock, flags);
1364 return limit_reached;
1367 static void pg_init_done(void *data, int errors)
1369 struct pgpath *pgpath = data;
1370 struct priority_group *pg = pgpath->pg;
1371 struct multipath *m = pg->m;
1372 unsigned long flags;
1373 bool delay_retry = false;
1375 /* device or driver problems */
1380 if (!m->hw_handler_name) {
1384 DMERR("Could not failover the device: Handler scsi_dh_%s "
1385 "Error %d.", m->hw_handler_name, errors);
1387 * Fail path for now, so we do not ping pong
1391 case SCSI_DH_DEV_TEMP_BUSY:
1393 * Probably doing something like FW upgrade on the
1394 * controller so try the other pg.
1396 bypass_pg(m, pg, true);
1399 /* Wait before retrying. */
1401 case SCSI_DH_IMM_RETRY:
1402 case SCSI_DH_RES_TEMP_UNAVAIL:
1403 if (pg_init_limit_reached(m, pgpath))
1407 case SCSI_DH_DEV_OFFLINED:
1410 * We probably do not want to fail the path for a device
1411 * error, but this is what the old dm did. In future
1412 * patches we can do more advanced handling.
1417 spin_lock_irqsave(&m->lock, flags);
1419 if (pgpath == m->current_pgpath) {
1420 DMERR("Could not failover device. Error %d.", errors);
1421 m->current_pgpath = NULL;
1422 m->current_pg = NULL;
1424 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1425 pg->bypassed = false;
1427 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1428 /* Activations of other paths are still on going */
1431 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1433 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1435 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1437 if (__pg_init_all_paths(m))
1440 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1442 process_queued_io_list(m);
1445 * Wake up any thread waiting to suspend.
1447 wake_up(&m->pg_init_wait);
1450 spin_unlock_irqrestore(&m->lock, flags);
1453 static void activate_or_offline_path(struct pgpath *pgpath)
1455 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1457 if (pgpath->is_active && !blk_queue_dying(q))
1458 scsi_dh_activate(q, pg_init_done, pgpath);
1460 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1463 static void activate_path_work(struct work_struct *work)
1465 struct pgpath *pgpath =
1466 container_of(work, struct pgpath, activate_path.work);
1468 activate_or_offline_path(pgpath);
1471 static int noretry_error(int error)
1476 * EBADE signals an reservation conflict.
1477 * We shouldn't fail the path here as we can communicate with
1478 * the target. We should failover to the next path, but in
1479 * doing so we might be causing a ping-pong between paths.
1480 * So just return the reservation conflict error.
1490 /* Anything else could be a path failure, so should be retried */
1497 static int do_end_io(struct multipath *m, struct request *clone,
1498 int error, struct dm_mpath_io *mpio)
1501 * We don't queue any clone request inside the multipath target
1502 * during end I/O handling, since those clone requests don't have
1503 * bio clones. If we queue them inside the multipath target,
1504 * we need to make bio clones, that requires memory allocation.
1505 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1506 * don't have bio clones.)
1507 * Instead of queueing the clone request here, we queue the original
1508 * request into dm core, which will remake a clone request and
1509 * clone bios for it and resubmit it later.
1511 int r = DM_ENDIO_REQUEUE;
1513 if (!error && !clone->errors)
1514 return 0; /* I/O complete */
1516 if (noretry_error(error))
1520 fail_path(mpio->pgpath);
1522 if (!atomic_read(&m->nr_valid_paths)) {
1523 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1524 if (!must_push_back_rq(m))
1532 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1533 int error, union map_info *map_context)
1535 struct multipath *m = ti->private;
1536 struct dm_mpath_io *mpio = get_mpio(map_context);
1537 struct pgpath *pgpath;
1538 struct path_selector *ps;
1543 r = do_end_io(m, clone, error, mpio);
1544 pgpath = mpio->pgpath;
1546 ps = &pgpath->pg->ps;
1547 if (ps->type->end_io)
1548 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1554 static int do_end_io_bio(struct multipath *m, struct bio *clone,
1555 int error, struct dm_mpath_io *mpio)
1557 unsigned long flags;
1560 return 0; /* I/O complete */
1562 if (noretry_error(error))
1566 fail_path(mpio->pgpath);
1568 if (!atomic_read(&m->nr_valid_paths)) {
1569 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1570 if (!must_push_back_bio(m))
1572 return DM_ENDIO_REQUEUE;
1576 /* Queue for the daemon to resubmit */
1577 dm_bio_restore(get_bio_details_from_bio(clone), clone);
1579 spin_lock_irqsave(&m->lock, flags);
1580 bio_list_add(&m->queued_bios, clone);
1581 spin_unlock_irqrestore(&m->lock, flags);
1582 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1583 queue_work(kmultipathd, &m->process_queued_bios);
1585 return DM_ENDIO_INCOMPLETE;
1588 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone, int error)
1590 struct multipath *m = ti->private;
1591 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1592 struct pgpath *pgpath;
1593 struct path_selector *ps;
1598 r = do_end_io_bio(m, clone, error, mpio);
1599 pgpath = mpio->pgpath;
1601 ps = &pgpath->pg->ps;
1602 if (ps->type->end_io)
1603 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1610 * Suspend can't complete until all the I/O is processed so if
1611 * the last path fails we must error any remaining I/O.
1612 * Note that if the freeze_bdev fails while suspending, the
1613 * queue_if_no_path state is lost - userspace should reset it.
1615 static void multipath_presuspend(struct dm_target *ti)
1617 struct multipath *m = ti->private;
1619 queue_if_no_path(m, false, true);
1622 static void multipath_postsuspend(struct dm_target *ti)
1624 struct multipath *m = ti->private;
1626 mutex_lock(&m->work_mutex);
1627 flush_multipath_work(m);
1628 mutex_unlock(&m->work_mutex);
1632 * Restore the queue_if_no_path setting.
1634 static void multipath_resume(struct dm_target *ti)
1636 struct multipath *m = ti->private;
1637 unsigned long flags;
1639 spin_lock_irqsave(&m->lock, flags);
1640 if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags))
1641 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1643 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1644 spin_unlock_irqrestore(&m->lock, flags);
1648 * Info output has the following format:
1649 * num_multipath_feature_args [multipath_feature_args]*
1650 * num_handler_status_args [handler_status_args]*
1651 * num_groups init_group_number
1652 * [A|D|E num_ps_status_args [ps_status_args]*
1653 * num_paths num_selector_args
1654 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1656 * Table output has the following format (identical to the constructor string):
1657 * num_feature_args [features_args]*
1658 * num_handler_args hw_handler [hw_handler_args]*
1659 * num_groups init_group_number
1660 * [priority selector-name num_ps_args [ps_args]*
1661 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1663 static void multipath_status(struct dm_target *ti, status_type_t type,
1664 unsigned status_flags, char *result, unsigned maxlen)
1667 unsigned long flags;
1668 struct multipath *m = ti->private;
1669 struct priority_group *pg;
1674 spin_lock_irqsave(&m->lock, flags);
1677 if (type == STATUSTYPE_INFO)
1678 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1679 atomic_read(&m->pg_init_count));
1681 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1682 (m->pg_init_retries > 0) * 2 +
1683 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1684 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1685 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1687 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1688 DMEMIT("queue_if_no_path ");
1689 if (m->pg_init_retries)
1690 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1691 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1692 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1693 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1694 DMEMIT("retain_attached_hw_handler ");
1695 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1696 switch(m->queue_mode) {
1697 case DM_TYPE_BIO_BASED:
1698 DMEMIT("queue_mode bio ");
1700 case DM_TYPE_MQ_REQUEST_BASED:
1701 DMEMIT("queue_mode mq ");
1710 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1713 DMEMIT("1 %s ", m->hw_handler_name);
1715 DMEMIT("%u ", m->nr_priority_groups);
1718 pg_num = m->next_pg->pg_num;
1719 else if (m->current_pg)
1720 pg_num = m->current_pg->pg_num;
1722 pg_num = (m->nr_priority_groups ? 1 : 0);
1724 DMEMIT("%u ", pg_num);
1727 case STATUSTYPE_INFO:
1728 list_for_each_entry(pg, &m->priority_groups, list) {
1730 state = 'D'; /* Disabled */
1731 else if (pg == m->current_pg)
1732 state = 'A'; /* Currently Active */
1734 state = 'E'; /* Enabled */
1736 DMEMIT("%c ", state);
1738 if (pg->ps.type->status)
1739 sz += pg->ps.type->status(&pg->ps, NULL, type,
1745 DMEMIT("%u %u ", pg->nr_pgpaths,
1746 pg->ps.type->info_args);
1748 list_for_each_entry(p, &pg->pgpaths, list) {
1749 DMEMIT("%s %s %u ", p->path.dev->name,
1750 p->is_active ? "A" : "F",
1752 if (pg->ps.type->status)
1753 sz += pg->ps.type->status(&pg->ps,
1754 &p->path, type, result + sz,
1760 case STATUSTYPE_TABLE:
1761 list_for_each_entry(pg, &m->priority_groups, list) {
1762 DMEMIT("%s ", pg->ps.type->name);
1764 if (pg->ps.type->status)
1765 sz += pg->ps.type->status(&pg->ps, NULL, type,
1771 DMEMIT("%u %u ", pg->nr_pgpaths,
1772 pg->ps.type->table_args);
1774 list_for_each_entry(p, &pg->pgpaths, list) {
1775 DMEMIT("%s ", p->path.dev->name);
1776 if (pg->ps.type->status)
1777 sz += pg->ps.type->status(&pg->ps,
1778 &p->path, type, result + sz,
1785 spin_unlock_irqrestore(&m->lock, flags);
1788 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1792 struct multipath *m = ti->private;
1795 mutex_lock(&m->work_mutex);
1797 if (dm_suspended(ti)) {
1803 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1804 r = queue_if_no_path(m, true, false);
1806 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1807 r = queue_if_no_path(m, false, false);
1813 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1817 if (!strcasecmp(argv[0], "disable_group")) {
1818 r = bypass_pg_num(m, argv[1], true);
1820 } else if (!strcasecmp(argv[0], "enable_group")) {
1821 r = bypass_pg_num(m, argv[1], false);
1823 } else if (!strcasecmp(argv[0], "switch_group")) {
1824 r = switch_pg_num(m, argv[1]);
1826 } else if (!strcasecmp(argv[0], "reinstate_path"))
1827 action = reinstate_path;
1828 else if (!strcasecmp(argv[0], "fail_path"))
1831 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1835 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1837 DMWARN("message: error getting device %s",
1842 r = action_dev(m, dev, action);
1844 dm_put_device(ti, dev);
1847 mutex_unlock(&m->work_mutex);
1851 static int multipath_prepare_ioctl(struct dm_target *ti,
1852 struct block_device **bdev, fmode_t *mode)
1854 struct multipath *m = ti->private;
1855 struct pgpath *current_pgpath;
1858 current_pgpath = lockless_dereference(m->current_pgpath);
1859 if (!current_pgpath)
1860 current_pgpath = choose_pgpath(m, 0);
1862 if (current_pgpath) {
1863 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1864 *bdev = current_pgpath->path.dev->bdev;
1865 *mode = current_pgpath->path.dev->mode;
1868 /* pg_init has not started or completed */
1872 /* No path is available */
1873 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1879 if (r == -ENOTCONN) {
1880 if (!lockless_dereference(m->current_pg)) {
1881 /* Path status changed, redo selection */
1882 (void) choose_pgpath(m, 0);
1884 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1885 pg_init_all_paths(m);
1886 dm_table_run_md_queue_async(m->ti->table);
1887 process_queued_io_list(m);
1891 * Only pass ioctls through if the device sizes match exactly.
1893 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1898 static int multipath_iterate_devices(struct dm_target *ti,
1899 iterate_devices_callout_fn fn, void *data)
1901 struct multipath *m = ti->private;
1902 struct priority_group *pg;
1906 list_for_each_entry(pg, &m->priority_groups, list) {
1907 list_for_each_entry(p, &pg->pgpaths, list) {
1908 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1918 static int pgpath_busy(struct pgpath *pgpath)
1920 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1922 return blk_lld_busy(q);
1926 * We return "busy", only when we can map I/Os but underlying devices
1927 * are busy (so even if we map I/Os now, the I/Os will wait on
1928 * the underlying queue).
1929 * In other words, if we want to kill I/Os or queue them inside us
1930 * due to map unavailability, we don't return "busy". Otherwise,
1931 * dm core won't give us the I/Os and we can't do what we want.
1933 static int multipath_busy(struct dm_target *ti)
1935 bool busy = false, has_active = false;
1936 struct multipath *m = ti->private;
1937 struct priority_group *pg, *next_pg;
1938 struct pgpath *pgpath;
1940 /* pg_init in progress */
1941 if (atomic_read(&m->pg_init_in_progress))
1944 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1945 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1946 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1948 /* Guess which priority_group will be used at next mapping time */
1949 pg = lockless_dereference(m->current_pg);
1950 next_pg = lockless_dereference(m->next_pg);
1951 if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
1956 * We don't know which pg will be used at next mapping time.
1957 * We don't call choose_pgpath() here to avoid to trigger
1958 * pg_init just by busy checking.
1959 * So we don't know whether underlying devices we will be using
1960 * at next mapping time are busy or not. Just try mapping.
1966 * If there is one non-busy active path at least, the path selector
1967 * will be able to select it. So we consider such a pg as not busy.
1970 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1971 if (pgpath->is_active) {
1973 if (!pgpath_busy(pgpath)) {
1982 * No active path in this pg, so this pg won't be used and
1983 * the current_pg will be changed at next mapping time.
1984 * We need to try mapping to determine it.
1992 /*-----------------------------------------------------------------
1994 *---------------------------------------------------------------*/
1995 static struct target_type multipath_target = {
1996 .name = "multipath",
1997 .version = {1, 12, 0},
1998 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1999 .module = THIS_MODULE,
2000 .ctr = multipath_ctr,
2001 .dtr = multipath_dtr,
2002 .clone_and_map_rq = multipath_clone_and_map,
2003 .release_clone_rq = multipath_release_clone,
2004 .rq_end_io = multipath_end_io,
2005 .map = multipath_map_bio,
2006 .end_io = multipath_end_io_bio,
2007 .presuspend = multipath_presuspend,
2008 .postsuspend = multipath_postsuspend,
2009 .resume = multipath_resume,
2010 .status = multipath_status,
2011 .message = multipath_message,
2012 .prepare_ioctl = multipath_prepare_ioctl,
2013 .iterate_devices = multipath_iterate_devices,
2014 .busy = multipath_busy,
2017 static int __init dm_multipath_init(void)
2021 r = dm_register_target(&multipath_target);
2023 DMERR("request-based register failed %d", r);
2025 goto bad_register_target;
2028 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2030 DMERR("failed to create workqueue kmpathd");
2032 goto bad_alloc_kmultipathd;
2036 * A separate workqueue is used to handle the device handlers
2037 * to avoid overloading existing workqueue. Overloading the
2038 * old workqueue would also create a bottleneck in the
2039 * path of the storage hardware device activation.
2041 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2043 if (!kmpath_handlerd) {
2044 DMERR("failed to create workqueue kmpath_handlerd");
2046 goto bad_alloc_kmpath_handlerd;
2051 bad_alloc_kmpath_handlerd:
2052 destroy_workqueue(kmultipathd);
2053 bad_alloc_kmultipathd:
2054 dm_unregister_target(&multipath_target);
2055 bad_register_target:
2059 static void __exit dm_multipath_exit(void)
2061 destroy_workqueue(kmpath_handlerd);
2062 destroy_workqueue(kmultipathd);
2064 dm_unregister_target(&multipath_target);
2067 module_init(dm_multipath_init);
2068 module_exit(dm_multipath_exit);
2070 MODULE_DESCRIPTION(DM_NAME " multipath target");
2071 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2072 MODULE_LICENSE("GPL");