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 unsigned long flags; /* Multipath state flags */
70 enum dm_queue_mode queue_mode;
72 struct pgpath *current_pgpath;
73 struct priority_group *current_pg;
74 struct priority_group *next_pg; /* Switch to this PG if set */
76 atomic_t nr_valid_paths; /* Total number of usable paths */
77 unsigned nr_priority_groups;
78 struct list_head priority_groups;
80 const char *hw_handler_name;
81 char *hw_handler_params;
82 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
83 unsigned pg_init_retries; /* Number of times to retry pg_init */
84 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
85 atomic_t pg_init_in_progress; /* Only one pg_init allowed at once */
86 atomic_t pg_init_count; /* Number of times pg_init called */
88 struct mutex work_mutex;
89 struct work_struct trigger_event;
92 struct work_struct process_queued_bios;
93 struct bio_list queued_bios;
97 * Context information attached to each io we process.
100 struct pgpath *pgpath;
104 typedef int (*action_fn) (struct pgpath *pgpath);
106 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
107 static void trigger_event(struct work_struct *work);
108 static void activate_or_offline_path(struct pgpath *pgpath);
109 static void activate_path_work(struct work_struct *work);
110 static void process_queued_bios(struct work_struct *work);
112 /*-----------------------------------------------
113 * Multipath state flags.
114 *-----------------------------------------------*/
116 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
117 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
118 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
119 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
120 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
121 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
122 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
124 /*-----------------------------------------------
125 * Allocation routines
126 *-----------------------------------------------*/
128 static struct pgpath *alloc_pgpath(void)
130 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
135 pgpath->is_active = true;
140 static void free_pgpath(struct pgpath *pgpath)
145 static struct priority_group *alloc_priority_group(void)
147 struct priority_group *pg;
149 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
152 INIT_LIST_HEAD(&pg->pgpaths);
157 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
159 struct pgpath *pgpath, *tmp;
161 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
162 list_del(&pgpath->list);
163 dm_put_device(ti, pgpath->path.dev);
168 static void free_priority_group(struct priority_group *pg,
169 struct dm_target *ti)
171 struct path_selector *ps = &pg->ps;
174 ps->type->destroy(ps);
175 dm_put_path_selector(ps->type);
178 free_pgpaths(&pg->pgpaths, ti);
182 static struct multipath *alloc_multipath(struct dm_target *ti)
186 m = kzalloc(sizeof(*m), GFP_KERNEL);
188 INIT_LIST_HEAD(&m->priority_groups);
189 spin_lock_init(&m->lock);
190 atomic_set(&m->nr_valid_paths, 0);
191 INIT_WORK(&m->trigger_event, trigger_event);
192 mutex_init(&m->work_mutex);
194 m->queue_mode = DM_TYPE_NONE;
203 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
205 if (m->queue_mode == DM_TYPE_NONE) {
207 * Default to request-based.
209 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
210 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
212 m->queue_mode = DM_TYPE_REQUEST_BASED;
214 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
215 INIT_WORK(&m->process_queued_bios, process_queued_bios);
217 * bio-based doesn't support any direct scsi_dh management;
218 * it just discovers if a scsi_dh is attached.
220 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
223 dm_table_set_type(ti->table, m->queue_mode);
226 * Init fields that are only used when a scsi_dh is attached
227 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
229 set_bit(MPATHF_QUEUE_IO, &m->flags);
230 atomic_set(&m->pg_init_in_progress, 0);
231 atomic_set(&m->pg_init_count, 0);
232 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
233 init_waitqueue_head(&m->pg_init_wait);
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);
249 mutex_destroy(&m->work_mutex);
253 static struct dm_mpath_io *get_mpio(union map_info *info)
258 static size_t multipath_per_bio_data_size(void)
260 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
263 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
265 return dm_per_bio_data(bio, multipath_per_bio_data_size());
268 static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
270 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
271 void *bio_details = mpio + 1;
275 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
277 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
278 struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
280 mpio->nr_bytes = bio->bi_iter.bi_size;
284 dm_bio_record(bio_details, bio);
287 /*-----------------------------------------------
289 *-----------------------------------------------*/
291 static int __pg_init_all_paths(struct multipath *m)
293 struct pgpath *pgpath;
294 unsigned long pg_init_delay = 0;
296 lockdep_assert_held(&m->lock);
298 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
301 atomic_inc(&m->pg_init_count);
302 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
304 /* Check here to reset pg_init_required */
308 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
309 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
310 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
311 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
312 /* Skip failed paths */
313 if (!pgpath->is_active)
315 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
317 atomic_inc(&m->pg_init_in_progress);
319 return atomic_read(&m->pg_init_in_progress);
322 static int pg_init_all_paths(struct multipath *m)
327 spin_lock_irqsave(&m->lock, flags);
328 ret = __pg_init_all_paths(m);
329 spin_unlock_irqrestore(&m->lock, flags);
334 static void __switch_pg(struct multipath *m, struct priority_group *pg)
338 /* Must we initialise the PG first, and queue I/O till it's ready? */
339 if (m->hw_handler_name) {
340 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
341 set_bit(MPATHF_QUEUE_IO, &m->flags);
343 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
344 clear_bit(MPATHF_QUEUE_IO, &m->flags);
347 atomic_set(&m->pg_init_count, 0);
350 static struct pgpath *choose_path_in_pg(struct multipath *m,
351 struct priority_group *pg,
355 struct dm_path *path;
356 struct pgpath *pgpath;
358 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
360 return ERR_PTR(-ENXIO);
362 pgpath = path_to_pgpath(path);
364 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
365 /* Only update current_pgpath if pg changed */
366 spin_lock_irqsave(&m->lock, flags);
367 m->current_pgpath = pgpath;
369 spin_unlock_irqrestore(&m->lock, flags);
375 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
378 struct priority_group *pg;
379 struct pgpath *pgpath;
380 unsigned bypassed = 1;
382 if (!atomic_read(&m->nr_valid_paths)) {
383 clear_bit(MPATHF_QUEUE_IO, &m->flags);
387 /* Were we instructed to switch PG? */
388 if (READ_ONCE(m->next_pg)) {
389 spin_lock_irqsave(&m->lock, flags);
392 spin_unlock_irqrestore(&m->lock, flags);
393 goto check_current_pg;
396 spin_unlock_irqrestore(&m->lock, flags);
397 pgpath = choose_path_in_pg(m, pg, nr_bytes);
398 if (!IS_ERR_OR_NULL(pgpath))
402 /* Don't change PG until it has no remaining paths */
404 pg = READ_ONCE(m->current_pg);
406 pgpath = choose_path_in_pg(m, pg, nr_bytes);
407 if (!IS_ERR_OR_NULL(pgpath))
412 * Loop through priority groups until we find a valid path.
413 * First time we skip PGs marked 'bypassed'.
414 * Second time we only try the ones we skipped, but set
415 * pg_init_delay_retry so we do not hammer controllers.
418 list_for_each_entry(pg, &m->priority_groups, list) {
419 if (pg->bypassed == !!bypassed)
421 pgpath = choose_path_in_pg(m, pg, nr_bytes);
422 if (!IS_ERR_OR_NULL(pgpath)) {
424 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
428 } while (bypassed--);
431 spin_lock_irqsave(&m->lock, flags);
432 m->current_pgpath = NULL;
433 m->current_pg = NULL;
434 spin_unlock_irqrestore(&m->lock, flags);
440 * dm_report_EIO() is a macro instead of a function to make pr_debug()
441 * report the function name and line number of the function from which
442 * it has been invoked.
444 #define dm_report_EIO(m) \
446 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
448 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
449 dm_device_name(md), \
450 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
451 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
452 dm_noflush_suspending((m)->ti)); \
456 * Check whether bios must be queued in the device-mapper core rather
457 * than here in the target.
459 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
460 * the same value then we are not between multipath_presuspend()
461 * and multipath_resume() calls and we have no need to check
462 * for the DMF_NOFLUSH_SUSPENDING flag.
464 static bool __must_push_back(struct multipath *m, unsigned long flags)
466 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
467 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
468 dm_noflush_suspending(m->ti));
472 * Following functions use READ_ONCE to get atomic access to
473 * all m->flags to avoid taking spinlock
475 static bool must_push_back_rq(struct multipath *m)
477 unsigned long flags = READ_ONCE(m->flags);
478 return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
481 static bool must_push_back_bio(struct multipath *m)
483 unsigned long flags = READ_ONCE(m->flags);
484 return __must_push_back(m, flags);
488 * Map cloned requests (request-based multipath)
490 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
491 union map_info *map_context,
492 struct request **__clone)
494 struct multipath *m = ti->private;
495 size_t nr_bytes = blk_rq_bytes(rq);
496 struct pgpath *pgpath;
497 struct block_device *bdev;
498 struct dm_mpath_io *mpio = get_mpio(map_context);
499 struct request_queue *q;
500 struct request *clone;
502 /* Do we need to select a new pgpath? */
503 pgpath = READ_ONCE(m->current_pgpath);
504 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
505 pgpath = choose_pgpath(m, nr_bytes);
508 if (must_push_back_rq(m))
509 return DM_MAPIO_DELAY_REQUEUE;
510 dm_report_EIO(m); /* Failed */
511 return DM_MAPIO_KILL;
512 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
513 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
514 pg_init_all_paths(m);
515 return DM_MAPIO_DELAY_REQUEUE;
518 mpio->pgpath = pgpath;
519 mpio->nr_bytes = nr_bytes;
521 bdev = pgpath->path.dev->bdev;
522 q = bdev_get_queue(bdev);
523 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE,
526 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
527 if (blk_queue_dying(q)) {
528 atomic_inc(&m->pg_init_in_progress);
529 activate_or_offline_path(pgpath);
530 return DM_MAPIO_DELAY_REQUEUE;
534 * blk-mq's SCHED_RESTART can cover this requeue, so we
535 * needn't deal with it by DELAY_REQUEUE. More importantly,
536 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
537 * get the queue busy feedback (via BLK_STS_RESOURCE),
538 * otherwise I/O merging can suffer.
541 return DM_MAPIO_REQUEUE;
543 return DM_MAPIO_DELAY_REQUEUE;
545 clone->bio = clone->biotail = NULL;
546 clone->rq_disk = bdev->bd_disk;
547 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
550 if (pgpath->pg->ps.type->start_io)
551 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
554 return DM_MAPIO_REMAPPED;
557 static void multipath_release_clone(struct request *clone,
558 union map_info *map_context)
560 if (unlikely(map_context)) {
562 * non-NULL map_context means caller is still map
563 * method; must undo multipath_clone_and_map()
565 struct dm_mpath_io *mpio = get_mpio(map_context);
566 struct pgpath *pgpath = mpio->pgpath;
568 if (pgpath && pgpath->pg->ps.type->end_io)
569 pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
574 blk_put_request(clone);
578 * Map cloned bios (bio-based multipath)
581 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
583 struct pgpath *pgpath;
587 /* Do we need to select a new pgpath? */
588 pgpath = READ_ONCE(m->current_pgpath);
589 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
590 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
592 /* MPATHF_QUEUE_IO might have been cleared by choose_pgpath. */
593 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
595 if ((pgpath && queue_io) ||
596 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
597 /* Queue for the daemon to resubmit */
598 spin_lock_irqsave(&m->lock, flags);
599 bio_list_add(&m->queued_bios, bio);
600 spin_unlock_irqrestore(&m->lock, flags);
602 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
603 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
604 pg_init_all_paths(m);
606 queue_work(kmultipathd, &m->process_queued_bios);
608 return ERR_PTR(-EAGAIN);
614 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
615 struct dm_mpath_io *mpio)
617 struct pgpath *pgpath = __map_bio(m, bio);
620 return DM_MAPIO_SUBMITTED;
623 if (must_push_back_bio(m))
624 return DM_MAPIO_REQUEUE;
626 return DM_MAPIO_KILL;
629 mpio->pgpath = pgpath;
632 bio_set_dev(bio, pgpath->path.dev->bdev);
633 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
635 if (pgpath->pg->ps.type->start_io)
636 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
639 return DM_MAPIO_REMAPPED;
642 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
644 struct multipath *m = ti->private;
645 struct dm_mpath_io *mpio = NULL;
647 multipath_init_per_bio_data(bio, &mpio);
648 return __multipath_map_bio(m, bio, mpio);
651 static void process_queued_io_list(struct multipath *m)
653 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
654 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
655 else if (m->queue_mode == DM_TYPE_BIO_BASED)
656 queue_work(kmultipathd, &m->process_queued_bios);
659 static void process_queued_bios(struct work_struct *work)
664 struct bio_list bios;
665 struct blk_plug plug;
666 struct multipath *m =
667 container_of(work, struct multipath, process_queued_bios);
669 bio_list_init(&bios);
671 spin_lock_irqsave(&m->lock, flags);
673 if (bio_list_empty(&m->queued_bios)) {
674 spin_unlock_irqrestore(&m->lock, flags);
678 bio_list_merge(&bios, &m->queued_bios);
679 bio_list_init(&m->queued_bios);
681 spin_unlock_irqrestore(&m->lock, flags);
683 blk_start_plug(&plug);
684 while ((bio = bio_list_pop(&bios))) {
685 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
686 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
687 r = __multipath_map_bio(m, bio, mpio);
690 bio->bi_status = BLK_STS_IOERR;
693 case DM_MAPIO_REQUEUE:
694 bio->bi_status = BLK_STS_DM_REQUEUE;
697 case DM_MAPIO_REMAPPED:
698 generic_make_request(bio);
700 case DM_MAPIO_SUBMITTED:
703 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
706 blk_finish_plug(&plug);
710 * If we run out of usable paths, should we queue I/O or error it?
712 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
717 spin_lock_irqsave(&m->lock, flags);
718 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
719 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
720 (!save_old_value && queue_if_no_path));
721 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
722 spin_unlock_irqrestore(&m->lock, flags);
724 if (!queue_if_no_path) {
725 dm_table_run_md_queue_async(m->ti->table);
726 process_queued_io_list(m);
733 * An event is triggered whenever a path is taken out of use.
734 * Includes path failure and PG bypass.
736 static void trigger_event(struct work_struct *work)
738 struct multipath *m =
739 container_of(work, struct multipath, trigger_event);
741 dm_table_event(m->ti->table);
744 /*-----------------------------------------------------------------
745 * Constructor/argument parsing:
746 * <#multipath feature args> [<arg>]*
747 * <#hw_handler args> [hw_handler [<arg>]*]
749 * <initial priority group>
750 * [<selector> <#selector args> [<arg>]*
751 * <#paths> <#per-path selector args>
752 * [<path> [<arg>]* ]+ ]+
753 *---------------------------------------------------------------*/
754 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
755 struct dm_target *ti)
758 struct path_selector_type *pst;
761 static const struct dm_arg _args[] = {
762 {0, 1024, "invalid number of path selector args"},
765 pst = dm_get_path_selector(dm_shift_arg(as));
767 ti->error = "unknown path selector type";
771 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
773 dm_put_path_selector(pst);
777 r = pst->create(&pg->ps, ps_argc, as->argv);
779 dm_put_path_selector(pst);
780 ti->error = "path selector constructor failed";
785 dm_consume_args(as, ps_argc);
790 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
791 const char **attached_handler_name, char **error)
793 struct request_queue *q = bdev_get_queue(bdev);
796 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
798 if (*attached_handler_name) {
800 * Clear any hw_handler_params associated with a
801 * handler that isn't already attached.
803 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
804 kfree(m->hw_handler_params);
805 m->hw_handler_params = NULL;
809 * Reset hw_handler_name to match the attached handler
811 * NB. This modifies the table line to show the actual
812 * handler instead of the original table passed in.
814 kfree(m->hw_handler_name);
815 m->hw_handler_name = *attached_handler_name;
816 *attached_handler_name = NULL;
820 if (m->hw_handler_name) {
821 r = scsi_dh_attach(q, m->hw_handler_name);
823 char b[BDEVNAME_SIZE];
825 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
830 *error = "error attaching hardware handler";
834 if (m->hw_handler_params) {
835 r = scsi_dh_set_params(q, m->hw_handler_params);
837 *error = "unable to set hardware handler parameters";
846 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
847 struct dm_target *ti)
851 struct multipath *m = ti->private;
852 struct request_queue *q;
853 const char *attached_handler_name = NULL;
855 /* we need at least a path arg */
857 ti->error = "no device given";
858 return ERR_PTR(-EINVAL);
863 return ERR_PTR(-ENOMEM);
865 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
868 ti->error = "error getting device";
872 q = bdev_get_queue(p->path.dev->bdev);
873 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
874 if (attached_handler_name || m->hw_handler_name) {
875 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
876 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
877 kfree(attached_handler_name);
879 dm_put_device(ti, p->path.dev);
884 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
886 dm_put_device(ti, p->path.dev);
896 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
899 static const struct dm_arg _args[] = {
900 {1, 1024, "invalid number of paths"},
901 {0, 1024, "invalid number of selector args"}
905 unsigned i, nr_selector_args, nr_args;
906 struct priority_group *pg;
907 struct dm_target *ti = m->ti;
911 ti->error = "not enough priority group arguments";
912 return ERR_PTR(-EINVAL);
915 pg = alloc_priority_group();
917 ti->error = "couldn't allocate priority group";
918 return ERR_PTR(-ENOMEM);
922 r = parse_path_selector(as, pg, ti);
929 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
933 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
937 nr_args = 1 + nr_selector_args;
938 for (i = 0; i < pg->nr_pgpaths; i++) {
939 struct pgpath *pgpath;
940 struct dm_arg_set path_args;
942 if (as->argc < nr_args) {
943 ti->error = "not enough path parameters";
948 path_args.argc = nr_args;
949 path_args.argv = as->argv;
951 pgpath = parse_path(&path_args, &pg->ps, ti);
952 if (IS_ERR(pgpath)) {
958 list_add_tail(&pgpath->list, &pg->pgpaths);
959 dm_consume_args(as, nr_args);
965 free_priority_group(pg, ti);
969 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
973 struct dm_target *ti = m->ti;
975 static const struct dm_arg _args[] = {
976 {0, 1024, "invalid number of hardware handler args"},
979 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
985 if (m->queue_mode == DM_TYPE_BIO_BASED) {
986 dm_consume_args(as, hw_argc);
987 DMERR("bio-based multipath doesn't allow hardware handler args");
991 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
992 if (!m->hw_handler_name)
999 for (i = 0; i <= hw_argc - 2; i++)
1000 len += strlen(as->argv[i]) + 1;
1001 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1003 ti->error = "memory allocation failed";
1007 j = sprintf(p, "%d", hw_argc - 1);
1008 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1009 j = sprintf(p, "%s", as->argv[i]);
1011 dm_consume_args(as, hw_argc - 1);
1015 kfree(m->hw_handler_name);
1016 m->hw_handler_name = NULL;
1020 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1024 struct dm_target *ti = m->ti;
1025 const char *arg_name;
1027 static const struct dm_arg _args[] = {
1028 {0, 8, "invalid number of feature args"},
1029 {1, 50, "pg_init_retries must be between 1 and 50"},
1030 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1033 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1041 arg_name = dm_shift_arg(as);
1044 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1045 r = queue_if_no_path(m, true, false);
1049 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1050 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1054 if (!strcasecmp(arg_name, "pg_init_retries") &&
1056 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1061 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1063 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1068 if (!strcasecmp(arg_name, "queue_mode") &&
1070 const char *queue_mode_name = dm_shift_arg(as);
1072 if (!strcasecmp(queue_mode_name, "bio"))
1073 m->queue_mode = DM_TYPE_BIO_BASED;
1074 else if (!strcasecmp(queue_mode_name, "rq"))
1075 m->queue_mode = DM_TYPE_REQUEST_BASED;
1076 else if (!strcasecmp(queue_mode_name, "mq"))
1077 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1079 ti->error = "Unknown 'queue_mode' requested";
1086 ti->error = "Unrecognised multipath feature request";
1088 } while (argc && !r);
1093 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1095 /* target arguments */
1096 static const struct dm_arg _args[] = {
1097 {0, 1024, "invalid number of priority groups"},
1098 {0, 1024, "invalid initial priority group number"},
1102 struct multipath *m;
1103 struct dm_arg_set as;
1104 unsigned pg_count = 0;
1105 unsigned next_pg_num;
1110 m = alloc_multipath(ti);
1112 ti->error = "can't allocate multipath";
1116 r = parse_features(&as, m);
1120 r = alloc_multipath_stage2(ti, m);
1124 r = parse_hw_handler(&as, m);
1128 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1132 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1136 if ((!m->nr_priority_groups && next_pg_num) ||
1137 (m->nr_priority_groups && !next_pg_num)) {
1138 ti->error = "invalid initial priority group";
1143 /* parse the priority groups */
1145 struct priority_group *pg;
1146 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1148 pg = parse_priority_group(&as, m);
1154 nr_valid_paths += pg->nr_pgpaths;
1155 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1157 list_add_tail(&pg->list, &m->priority_groups);
1159 pg->pg_num = pg_count;
1164 if (pg_count != m->nr_priority_groups) {
1165 ti->error = "priority group count mismatch";
1170 ti->num_flush_bios = 1;
1171 ti->num_discard_bios = 1;
1172 ti->num_write_same_bios = 1;
1173 ti->num_write_zeroes_bios = 1;
1174 if (m->queue_mode == DM_TYPE_BIO_BASED)
1175 ti->per_io_data_size = multipath_per_bio_data_size();
1177 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1186 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1191 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1193 if (!atomic_read(&m->pg_init_in_progress))
1198 finish_wait(&m->pg_init_wait, &wait);
1201 static void flush_multipath_work(struct multipath *m)
1203 if (m->hw_handler_name) {
1204 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1205 smp_mb__after_atomic();
1207 flush_workqueue(kmpath_handlerd);
1208 multipath_wait_for_pg_init_completion(m);
1210 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1211 smp_mb__after_atomic();
1214 flush_workqueue(kmultipathd);
1215 flush_work(&m->trigger_event);
1218 static void multipath_dtr(struct dm_target *ti)
1220 struct multipath *m = ti->private;
1222 flush_multipath_work(m);
1227 * Take a path out of use.
1229 static int fail_path(struct pgpath *pgpath)
1231 unsigned long flags;
1232 struct multipath *m = pgpath->pg->m;
1234 spin_lock_irqsave(&m->lock, flags);
1236 if (!pgpath->is_active)
1239 DMWARN("Failing path %s.", pgpath->path.dev->name);
1241 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1242 pgpath->is_active = false;
1243 pgpath->fail_count++;
1245 atomic_dec(&m->nr_valid_paths);
1247 if (pgpath == m->current_pgpath)
1248 m->current_pgpath = NULL;
1250 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1251 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1253 schedule_work(&m->trigger_event);
1256 spin_unlock_irqrestore(&m->lock, flags);
1262 * Reinstate a previously-failed path
1264 static int reinstate_path(struct pgpath *pgpath)
1266 int r = 0, run_queue = 0;
1267 unsigned long flags;
1268 struct multipath *m = pgpath->pg->m;
1269 unsigned nr_valid_paths;
1271 spin_lock_irqsave(&m->lock, flags);
1273 if (pgpath->is_active)
1276 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1278 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1282 pgpath->is_active = true;
1284 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1285 if (nr_valid_paths == 1) {
1286 m->current_pgpath = NULL;
1288 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1289 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1290 atomic_inc(&m->pg_init_in_progress);
1293 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1294 pgpath->path.dev->name, nr_valid_paths);
1296 schedule_work(&m->trigger_event);
1299 spin_unlock_irqrestore(&m->lock, flags);
1301 dm_table_run_md_queue_async(m->ti->table);
1302 process_queued_io_list(m);
1309 * Fail or reinstate all paths that match the provided struct dm_dev.
1311 static int action_dev(struct multipath *m, struct dm_dev *dev,
1315 struct pgpath *pgpath;
1316 struct priority_group *pg;
1318 list_for_each_entry(pg, &m->priority_groups, list) {
1319 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1320 if (pgpath->path.dev == dev)
1329 * Temporarily try to avoid having to use the specified PG
1331 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1334 unsigned long flags;
1336 spin_lock_irqsave(&m->lock, flags);
1338 pg->bypassed = bypassed;
1339 m->current_pgpath = NULL;
1340 m->current_pg = NULL;
1342 spin_unlock_irqrestore(&m->lock, flags);
1344 schedule_work(&m->trigger_event);
1348 * Switch to using the specified PG from the next I/O that gets mapped
1350 static int switch_pg_num(struct multipath *m, const char *pgstr)
1352 struct priority_group *pg;
1354 unsigned long flags;
1357 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1358 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1359 DMWARN("invalid PG number supplied to switch_pg_num");
1363 spin_lock_irqsave(&m->lock, flags);
1364 list_for_each_entry(pg, &m->priority_groups, list) {
1365 pg->bypassed = false;
1369 m->current_pgpath = NULL;
1370 m->current_pg = NULL;
1373 spin_unlock_irqrestore(&m->lock, flags);
1375 schedule_work(&m->trigger_event);
1380 * Set/clear bypassed status of a PG.
1381 * PGs are numbered upwards from 1 in the order they were declared.
1383 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1385 struct priority_group *pg;
1389 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1390 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1391 DMWARN("invalid PG number supplied to bypass_pg");
1395 list_for_each_entry(pg, &m->priority_groups, list) {
1400 bypass_pg(m, pg, bypassed);
1405 * Should we retry pg_init immediately?
1407 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1409 unsigned long flags;
1410 bool limit_reached = false;
1412 spin_lock_irqsave(&m->lock, flags);
1414 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1415 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1416 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1418 limit_reached = true;
1420 spin_unlock_irqrestore(&m->lock, flags);
1422 return limit_reached;
1425 static void pg_init_done(void *data, int errors)
1427 struct pgpath *pgpath = data;
1428 struct priority_group *pg = pgpath->pg;
1429 struct multipath *m = pg->m;
1430 unsigned long flags;
1431 bool delay_retry = false;
1433 /* device or driver problems */
1438 if (!m->hw_handler_name) {
1442 DMERR("Could not failover the device: Handler scsi_dh_%s "
1443 "Error %d.", m->hw_handler_name, errors);
1445 * Fail path for now, so we do not ping pong
1449 case SCSI_DH_DEV_TEMP_BUSY:
1451 * Probably doing something like FW upgrade on the
1452 * controller so try the other pg.
1454 bypass_pg(m, pg, true);
1457 /* Wait before retrying. */
1460 case SCSI_DH_IMM_RETRY:
1461 case SCSI_DH_RES_TEMP_UNAVAIL:
1462 if (pg_init_limit_reached(m, pgpath))
1466 case SCSI_DH_DEV_OFFLINED:
1469 * We probably do not want to fail the path for a device
1470 * error, but this is what the old dm did. In future
1471 * patches we can do more advanced handling.
1476 spin_lock_irqsave(&m->lock, flags);
1478 if (pgpath == m->current_pgpath) {
1479 DMERR("Could not failover device. Error %d.", errors);
1480 m->current_pgpath = NULL;
1481 m->current_pg = NULL;
1483 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1484 pg->bypassed = false;
1486 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1487 /* Activations of other paths are still on going */
1490 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1492 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1494 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1496 if (__pg_init_all_paths(m))
1499 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1501 process_queued_io_list(m);
1504 * Wake up any thread waiting to suspend.
1506 wake_up(&m->pg_init_wait);
1509 spin_unlock_irqrestore(&m->lock, flags);
1512 static void activate_or_offline_path(struct pgpath *pgpath)
1514 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1516 if (pgpath->is_active && !blk_queue_dying(q))
1517 scsi_dh_activate(q, pg_init_done, pgpath);
1519 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1522 static void activate_path_work(struct work_struct *work)
1524 struct pgpath *pgpath =
1525 container_of(work, struct pgpath, activate_path.work);
1527 activate_or_offline_path(pgpath);
1530 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1531 blk_status_t error, union map_info *map_context)
1533 struct dm_mpath_io *mpio = get_mpio(map_context);
1534 struct pgpath *pgpath = mpio->pgpath;
1535 int r = DM_ENDIO_DONE;
1538 * We don't queue any clone request inside the multipath target
1539 * during end I/O handling, since those clone requests don't have
1540 * bio clones. If we queue them inside the multipath target,
1541 * we need to make bio clones, that requires memory allocation.
1542 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1543 * don't have bio clones.)
1544 * Instead of queueing the clone request here, we queue the original
1545 * request into dm core, which will remake a clone request and
1546 * clone bios for it and resubmit it later.
1548 if (error && blk_path_error(error)) {
1549 struct multipath *m = ti->private;
1551 if (error == BLK_STS_RESOURCE)
1552 r = DM_ENDIO_DELAY_REQUEUE;
1554 r = DM_ENDIO_REQUEUE;
1559 if (atomic_read(&m->nr_valid_paths) == 0 &&
1560 !must_push_back_rq(m)) {
1561 if (error == BLK_STS_IOERR)
1563 /* complete with the original error */
1569 struct path_selector *ps = &pgpath->pg->ps;
1571 if (ps->type->end_io)
1572 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1578 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1579 blk_status_t *error)
1581 struct multipath *m = ti->private;
1582 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1583 struct pgpath *pgpath = mpio->pgpath;
1584 unsigned long flags;
1585 int r = DM_ENDIO_DONE;
1587 if (!*error || !blk_path_error(*error))
1593 if (atomic_read(&m->nr_valid_paths) == 0 &&
1594 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1595 if (must_push_back_bio(m)) {
1596 r = DM_ENDIO_REQUEUE;
1599 *error = BLK_STS_IOERR;
1604 spin_lock_irqsave(&m->lock, flags);
1605 bio_list_add(&m->queued_bios, clone);
1606 spin_unlock_irqrestore(&m->lock, flags);
1607 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1608 queue_work(kmultipathd, &m->process_queued_bios);
1610 r = DM_ENDIO_INCOMPLETE;
1613 struct path_selector *ps = &pgpath->pg->ps;
1615 if (ps->type->end_io)
1616 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1623 * Suspend can't complete until all the I/O is processed so if
1624 * the last path fails we must error any remaining I/O.
1625 * Note that if the freeze_bdev fails while suspending, the
1626 * queue_if_no_path state is lost - userspace should reset it.
1628 static void multipath_presuspend(struct dm_target *ti)
1630 struct multipath *m = ti->private;
1632 queue_if_no_path(m, false, true);
1635 static void multipath_postsuspend(struct dm_target *ti)
1637 struct multipath *m = ti->private;
1639 mutex_lock(&m->work_mutex);
1640 flush_multipath_work(m);
1641 mutex_unlock(&m->work_mutex);
1645 * Restore the queue_if_no_path setting.
1647 static void multipath_resume(struct dm_target *ti)
1649 struct multipath *m = ti->private;
1650 unsigned long flags;
1652 spin_lock_irqsave(&m->lock, flags);
1653 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1654 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1655 spin_unlock_irqrestore(&m->lock, flags);
1659 * Info output has the following format:
1660 * num_multipath_feature_args [multipath_feature_args]*
1661 * num_handler_status_args [handler_status_args]*
1662 * num_groups init_group_number
1663 * [A|D|E num_ps_status_args [ps_status_args]*
1664 * num_paths num_selector_args
1665 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1667 * Table output has the following format (identical to the constructor string):
1668 * num_feature_args [features_args]*
1669 * num_handler_args hw_handler [hw_handler_args]*
1670 * num_groups init_group_number
1671 * [priority selector-name num_ps_args [ps_args]*
1672 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1674 static void multipath_status(struct dm_target *ti, status_type_t type,
1675 unsigned status_flags, char *result, unsigned maxlen)
1678 unsigned long flags;
1679 struct multipath *m = ti->private;
1680 struct priority_group *pg;
1685 spin_lock_irqsave(&m->lock, flags);
1688 if (type == STATUSTYPE_INFO)
1689 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1690 atomic_read(&m->pg_init_count));
1692 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1693 (m->pg_init_retries > 0) * 2 +
1694 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1695 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1696 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1698 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1699 DMEMIT("queue_if_no_path ");
1700 if (m->pg_init_retries)
1701 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1702 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1703 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1704 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1705 DMEMIT("retain_attached_hw_handler ");
1706 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1707 switch(m->queue_mode) {
1708 case DM_TYPE_BIO_BASED:
1709 DMEMIT("queue_mode bio ");
1711 case DM_TYPE_MQ_REQUEST_BASED:
1712 DMEMIT("queue_mode mq ");
1721 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1724 DMEMIT("1 %s ", m->hw_handler_name);
1726 DMEMIT("%u ", m->nr_priority_groups);
1729 pg_num = m->next_pg->pg_num;
1730 else if (m->current_pg)
1731 pg_num = m->current_pg->pg_num;
1733 pg_num = (m->nr_priority_groups ? 1 : 0);
1735 DMEMIT("%u ", pg_num);
1738 case STATUSTYPE_INFO:
1739 list_for_each_entry(pg, &m->priority_groups, list) {
1741 state = 'D'; /* Disabled */
1742 else if (pg == m->current_pg)
1743 state = 'A'; /* Currently Active */
1745 state = 'E'; /* Enabled */
1747 DMEMIT("%c ", state);
1749 if (pg->ps.type->status)
1750 sz += pg->ps.type->status(&pg->ps, NULL, type,
1756 DMEMIT("%u %u ", pg->nr_pgpaths,
1757 pg->ps.type->info_args);
1759 list_for_each_entry(p, &pg->pgpaths, list) {
1760 DMEMIT("%s %s %u ", p->path.dev->name,
1761 p->is_active ? "A" : "F",
1763 if (pg->ps.type->status)
1764 sz += pg->ps.type->status(&pg->ps,
1765 &p->path, type, result + sz,
1771 case STATUSTYPE_TABLE:
1772 list_for_each_entry(pg, &m->priority_groups, list) {
1773 DMEMIT("%s ", pg->ps.type->name);
1775 if (pg->ps.type->status)
1776 sz += pg->ps.type->status(&pg->ps, NULL, type,
1782 DMEMIT("%u %u ", pg->nr_pgpaths,
1783 pg->ps.type->table_args);
1785 list_for_each_entry(p, &pg->pgpaths, list) {
1786 DMEMIT("%s ", p->path.dev->name);
1787 if (pg->ps.type->status)
1788 sz += pg->ps.type->status(&pg->ps,
1789 &p->path, type, result + sz,
1796 spin_unlock_irqrestore(&m->lock, flags);
1799 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1800 char *result, unsigned maxlen)
1804 struct multipath *m = ti->private;
1807 mutex_lock(&m->work_mutex);
1809 if (dm_suspended(ti)) {
1815 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1816 r = queue_if_no_path(m, true, false);
1818 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1819 r = queue_if_no_path(m, false, false);
1825 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1829 if (!strcasecmp(argv[0], "disable_group")) {
1830 r = bypass_pg_num(m, argv[1], true);
1832 } else if (!strcasecmp(argv[0], "enable_group")) {
1833 r = bypass_pg_num(m, argv[1], false);
1835 } else if (!strcasecmp(argv[0], "switch_group")) {
1836 r = switch_pg_num(m, argv[1]);
1838 } else if (!strcasecmp(argv[0], "reinstate_path"))
1839 action = reinstate_path;
1840 else if (!strcasecmp(argv[0], "fail_path"))
1843 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1847 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1849 DMWARN("message: error getting device %s",
1854 r = action_dev(m, dev, action);
1856 dm_put_device(ti, dev);
1859 mutex_unlock(&m->work_mutex);
1863 static int multipath_prepare_ioctl(struct dm_target *ti,
1864 struct block_device **bdev)
1866 struct multipath *m = ti->private;
1867 struct pgpath *current_pgpath;
1870 current_pgpath = READ_ONCE(m->current_pgpath);
1871 if (!current_pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
1872 current_pgpath = choose_pgpath(m, 0);
1874 if (current_pgpath) {
1875 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1876 *bdev = current_pgpath->path.dev->bdev;
1879 /* pg_init has not started or completed */
1883 /* No path is available */
1884 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1890 if (r == -ENOTCONN) {
1891 if (!READ_ONCE(m->current_pg)) {
1892 /* Path status changed, redo selection */
1893 (void) choose_pgpath(m, 0);
1895 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1896 pg_init_all_paths(m);
1897 dm_table_run_md_queue_async(m->ti->table);
1898 process_queued_io_list(m);
1902 * Only pass ioctls through if the device sizes match exactly.
1904 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1909 static int multipath_iterate_devices(struct dm_target *ti,
1910 iterate_devices_callout_fn fn, void *data)
1912 struct multipath *m = ti->private;
1913 struct priority_group *pg;
1917 list_for_each_entry(pg, &m->priority_groups, list) {
1918 list_for_each_entry(p, &pg->pgpaths, list) {
1919 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1929 static int pgpath_busy(struct pgpath *pgpath)
1931 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1933 return blk_lld_busy(q);
1937 * We return "busy", only when we can map I/Os but underlying devices
1938 * are busy (so even if we map I/Os now, the I/Os will wait on
1939 * the underlying queue).
1940 * In other words, if we want to kill I/Os or queue them inside us
1941 * due to map unavailability, we don't return "busy". Otherwise,
1942 * dm core won't give us the I/Os and we can't do what we want.
1944 static int multipath_busy(struct dm_target *ti)
1946 bool busy = false, has_active = false;
1947 struct multipath *m = ti->private;
1948 struct priority_group *pg, *next_pg;
1949 struct pgpath *pgpath;
1951 /* pg_init in progress */
1952 if (atomic_read(&m->pg_init_in_progress))
1955 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1956 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1957 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1959 /* Guess which priority_group will be used at next mapping time */
1960 pg = READ_ONCE(m->current_pg);
1961 next_pg = READ_ONCE(m->next_pg);
1962 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1967 * We don't know which pg will be used at next mapping time.
1968 * We don't call choose_pgpath() here to avoid to trigger
1969 * pg_init just by busy checking.
1970 * So we don't know whether underlying devices we will be using
1971 * at next mapping time are busy or not. Just try mapping.
1977 * If there is one non-busy active path at least, the path selector
1978 * will be able to select it. So we consider such a pg as not busy.
1981 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1982 if (pgpath->is_active) {
1984 if (!pgpath_busy(pgpath)) {
1993 * No active path in this pg, so this pg won't be used and
1994 * the current_pg will be changed at next mapping time.
1995 * We need to try mapping to determine it.
2003 /*-----------------------------------------------------------------
2005 *---------------------------------------------------------------*/
2006 static struct target_type multipath_target = {
2007 .name = "multipath",
2008 .version = {1, 13, 0},
2009 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2010 DM_TARGET_PASSES_INTEGRITY,
2011 .module = THIS_MODULE,
2012 .ctr = multipath_ctr,
2013 .dtr = multipath_dtr,
2014 .clone_and_map_rq = multipath_clone_and_map,
2015 .release_clone_rq = multipath_release_clone,
2016 .rq_end_io = multipath_end_io,
2017 .map = multipath_map_bio,
2018 .end_io = multipath_end_io_bio,
2019 .presuspend = multipath_presuspend,
2020 .postsuspend = multipath_postsuspend,
2021 .resume = multipath_resume,
2022 .status = multipath_status,
2023 .message = multipath_message,
2024 .prepare_ioctl = multipath_prepare_ioctl,
2025 .iterate_devices = multipath_iterate_devices,
2026 .busy = multipath_busy,
2029 static int __init dm_multipath_init(void)
2033 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2035 DMERR("failed to create workqueue kmpathd");
2037 goto bad_alloc_kmultipathd;
2041 * A separate workqueue is used to handle the device handlers
2042 * to avoid overloading existing workqueue. Overloading the
2043 * old workqueue would also create a bottleneck in the
2044 * path of the storage hardware device activation.
2046 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2048 if (!kmpath_handlerd) {
2049 DMERR("failed to create workqueue kmpath_handlerd");
2051 goto bad_alloc_kmpath_handlerd;
2054 r = dm_register_target(&multipath_target);
2056 DMERR("request-based register failed %d", r);
2058 goto bad_register_target;
2063 bad_register_target:
2064 destroy_workqueue(kmpath_handlerd);
2065 bad_alloc_kmpath_handlerd:
2066 destroy_workqueue(kmultipathd);
2067 bad_alloc_kmultipathd:
2071 static void __exit dm_multipath_exit(void)
2073 destroy_workqueue(kmpath_handlerd);
2074 destroy_workqueue(kmultipathd);
2076 dm_unregister_target(&multipath_target);
2079 module_init(dm_multipath_init);
2080 module_exit(dm_multipath_exit);
2082 MODULE_DESCRIPTION(DM_NAME " multipath target");
2083 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2084 MODULE_LICENSE("GPL");