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)
278 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
279 struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
281 memset(mpio, 0, sizeof(*mpio));
282 memset(bio_details, 0, sizeof(*bio_details));
283 dm_bio_record(bio_details, bio);
289 /*-----------------------------------------------
291 *-----------------------------------------------*/
293 static int __pg_init_all_paths(struct multipath *m)
295 struct pgpath *pgpath;
296 unsigned long pg_init_delay = 0;
298 lockdep_assert_held(&m->lock);
300 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
303 atomic_inc(&m->pg_init_count);
304 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
306 /* Check here to reset pg_init_required */
310 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
311 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
312 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
313 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
314 /* Skip failed paths */
315 if (!pgpath->is_active)
317 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
319 atomic_inc(&m->pg_init_in_progress);
321 return atomic_read(&m->pg_init_in_progress);
324 static int pg_init_all_paths(struct multipath *m)
329 spin_lock_irqsave(&m->lock, flags);
330 ret = __pg_init_all_paths(m);
331 spin_unlock_irqrestore(&m->lock, flags);
336 static void __switch_pg(struct multipath *m, struct priority_group *pg)
340 /* Must we initialise the PG first, and queue I/O till it's ready? */
341 if (m->hw_handler_name) {
342 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
343 set_bit(MPATHF_QUEUE_IO, &m->flags);
345 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
346 clear_bit(MPATHF_QUEUE_IO, &m->flags);
349 atomic_set(&m->pg_init_count, 0);
352 static struct pgpath *choose_path_in_pg(struct multipath *m,
353 struct priority_group *pg,
357 struct dm_path *path;
358 struct pgpath *pgpath;
360 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
362 return ERR_PTR(-ENXIO);
364 pgpath = path_to_pgpath(path);
366 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
367 /* Only update current_pgpath if pg changed */
368 spin_lock_irqsave(&m->lock, flags);
369 m->current_pgpath = pgpath;
371 spin_unlock_irqrestore(&m->lock, flags);
377 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
380 struct priority_group *pg;
381 struct pgpath *pgpath;
382 unsigned bypassed = 1;
384 if (!atomic_read(&m->nr_valid_paths)) {
385 clear_bit(MPATHF_QUEUE_IO, &m->flags);
389 /* Were we instructed to switch PG? */
390 if (READ_ONCE(m->next_pg)) {
391 spin_lock_irqsave(&m->lock, flags);
394 spin_unlock_irqrestore(&m->lock, flags);
395 goto check_current_pg;
398 spin_unlock_irqrestore(&m->lock, flags);
399 pgpath = choose_path_in_pg(m, pg, nr_bytes);
400 if (!IS_ERR_OR_NULL(pgpath))
404 /* Don't change PG until it has no remaining paths */
406 pg = READ_ONCE(m->current_pg);
408 pgpath = choose_path_in_pg(m, pg, nr_bytes);
409 if (!IS_ERR_OR_NULL(pgpath))
414 * Loop through priority groups until we find a valid path.
415 * First time we skip PGs marked 'bypassed'.
416 * Second time we only try the ones we skipped, but set
417 * pg_init_delay_retry so we do not hammer controllers.
420 list_for_each_entry(pg, &m->priority_groups, list) {
421 if (pg->bypassed == !!bypassed)
423 pgpath = choose_path_in_pg(m, pg, nr_bytes);
424 if (!IS_ERR_OR_NULL(pgpath)) {
426 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
430 } while (bypassed--);
433 spin_lock_irqsave(&m->lock, flags);
434 m->current_pgpath = NULL;
435 m->current_pg = NULL;
436 spin_unlock_irqrestore(&m->lock, flags);
442 * dm_report_EIO() is a macro instead of a function to make pr_debug()
443 * report the function name and line number of the function from which
444 * it has been invoked.
446 #define dm_report_EIO(m) \
448 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
450 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
451 dm_device_name(md), \
452 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
453 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
454 dm_noflush_suspending((m)->ti)); \
458 * Check whether bios must be queued in the device-mapper core rather
459 * than here in the target.
461 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
462 * the same value then we are not between multipath_presuspend()
463 * and multipath_resume() calls and we have no need to check
464 * for the DMF_NOFLUSH_SUSPENDING flag.
466 static bool __must_push_back(struct multipath *m, unsigned long flags)
468 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
469 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
470 dm_noflush_suspending(m->ti));
474 * Following functions use READ_ONCE to get atomic access to
475 * all m->flags to avoid taking spinlock
477 static bool must_push_back_rq(struct multipath *m)
479 unsigned long flags = READ_ONCE(m->flags);
480 return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
483 static bool must_push_back_bio(struct multipath *m)
485 unsigned long flags = READ_ONCE(m->flags);
486 return __must_push_back(m, flags);
490 * Map cloned requests (request-based multipath)
492 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
493 union map_info *map_context,
494 struct request **__clone)
496 struct multipath *m = ti->private;
497 size_t nr_bytes = blk_rq_bytes(rq);
498 struct pgpath *pgpath;
499 struct block_device *bdev;
500 struct dm_mpath_io *mpio = get_mpio(map_context);
501 struct request_queue *q;
502 struct request *clone;
504 /* Do we need to select a new pgpath? */
505 pgpath = READ_ONCE(m->current_pgpath);
506 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
507 pgpath = choose_pgpath(m, nr_bytes);
510 if (must_push_back_rq(m))
511 return DM_MAPIO_DELAY_REQUEUE;
512 dm_report_EIO(m); /* Failed */
513 return DM_MAPIO_KILL;
514 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
515 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
516 if (pg_init_all_paths(m))
517 return DM_MAPIO_DELAY_REQUEUE;
518 return DM_MAPIO_REQUEUE;
521 memset(mpio, 0, sizeof(*mpio));
522 mpio->pgpath = pgpath;
523 mpio->nr_bytes = nr_bytes;
525 bdev = pgpath->path.dev->bdev;
526 q = bdev_get_queue(bdev);
527 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
529 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
530 bool queue_dying = blk_queue_dying(q);
532 atomic_inc(&m->pg_init_in_progress);
533 activate_or_offline_path(pgpath);
535 return DM_MAPIO_DELAY_REQUEUE;
537 clone->bio = clone->biotail = NULL;
538 clone->rq_disk = bdev->bd_disk;
539 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
542 if (pgpath->pg->ps.type->start_io)
543 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
546 return DM_MAPIO_REMAPPED;
549 static void multipath_release_clone(struct request *clone)
551 blk_put_request(clone);
555 * Map cloned bios (bio-based multipath)
557 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
559 size_t nr_bytes = bio->bi_iter.bi_size;
560 struct pgpath *pgpath;
564 /* Do we need to select a new pgpath? */
565 pgpath = READ_ONCE(m->current_pgpath);
566 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
567 if (!pgpath || !queue_io)
568 pgpath = choose_pgpath(m, nr_bytes);
570 if ((pgpath && queue_io) ||
571 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
572 /* Queue for the daemon to resubmit */
573 spin_lock_irqsave(&m->lock, flags);
574 bio_list_add(&m->queued_bios, bio);
575 spin_unlock_irqrestore(&m->lock, flags);
576 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
577 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
578 pg_init_all_paths(m);
580 queue_work(kmultipathd, &m->process_queued_bios);
581 return DM_MAPIO_SUBMITTED;
585 if (must_push_back_bio(m))
586 return DM_MAPIO_REQUEUE;
588 return DM_MAPIO_KILL;
591 mpio->pgpath = pgpath;
592 mpio->nr_bytes = nr_bytes;
595 bio_set_dev(bio, pgpath->path.dev->bdev);
596 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
598 if (pgpath->pg->ps.type->start_io)
599 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
602 return DM_MAPIO_REMAPPED;
605 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
607 struct multipath *m = ti->private;
608 struct dm_mpath_io *mpio = NULL;
610 multipath_init_per_bio_data(bio, &mpio);
611 return __multipath_map_bio(m, bio, mpio);
614 static void process_queued_io_list(struct multipath *m)
616 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
617 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
618 else if (m->queue_mode == DM_TYPE_BIO_BASED)
619 queue_work(kmultipathd, &m->process_queued_bios);
622 static void process_queued_bios(struct work_struct *work)
627 struct bio_list bios;
628 struct blk_plug plug;
629 struct multipath *m =
630 container_of(work, struct multipath, process_queued_bios);
632 bio_list_init(&bios);
634 spin_lock_irqsave(&m->lock, flags);
636 if (bio_list_empty(&m->queued_bios)) {
637 spin_unlock_irqrestore(&m->lock, flags);
641 bio_list_merge(&bios, &m->queued_bios);
642 bio_list_init(&m->queued_bios);
644 spin_unlock_irqrestore(&m->lock, flags);
646 blk_start_plug(&plug);
647 while ((bio = bio_list_pop(&bios))) {
648 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
651 bio->bi_status = BLK_STS_IOERR;
654 case DM_MAPIO_REQUEUE:
655 bio->bi_status = BLK_STS_DM_REQUEUE;
658 case DM_MAPIO_REMAPPED:
659 generic_make_request(bio);
664 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
667 blk_finish_plug(&plug);
671 * If we run out of usable paths, should we queue I/O or error it?
673 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
678 spin_lock_irqsave(&m->lock, flags);
679 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
680 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
681 (!save_old_value && queue_if_no_path));
682 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
683 spin_unlock_irqrestore(&m->lock, flags);
685 if (!queue_if_no_path) {
686 dm_table_run_md_queue_async(m->ti->table);
687 process_queued_io_list(m);
694 * An event is triggered whenever a path is taken out of use.
695 * Includes path failure and PG bypass.
697 static void trigger_event(struct work_struct *work)
699 struct multipath *m =
700 container_of(work, struct multipath, trigger_event);
702 dm_table_event(m->ti->table);
705 /*-----------------------------------------------------------------
706 * Constructor/argument parsing:
707 * <#multipath feature args> [<arg>]*
708 * <#hw_handler args> [hw_handler [<arg>]*]
710 * <initial priority group>
711 * [<selector> <#selector args> [<arg>]*
712 * <#paths> <#per-path selector args>
713 * [<path> [<arg>]* ]+ ]+
714 *---------------------------------------------------------------*/
715 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
716 struct dm_target *ti)
719 struct path_selector_type *pst;
722 static const struct dm_arg _args[] = {
723 {0, 1024, "invalid number of path selector args"},
726 pst = dm_get_path_selector(dm_shift_arg(as));
728 ti->error = "unknown path selector type";
732 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
734 dm_put_path_selector(pst);
738 r = pst->create(&pg->ps, ps_argc, as->argv);
740 dm_put_path_selector(pst);
741 ti->error = "path selector constructor failed";
746 dm_consume_args(as, ps_argc);
751 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
752 struct dm_target *ti)
756 struct multipath *m = ti->private;
757 struct request_queue *q = NULL;
758 const char *attached_handler_name;
760 /* we need at least a path arg */
762 ti->error = "no device given";
763 return ERR_PTR(-EINVAL);
768 return ERR_PTR(-ENOMEM);
770 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
773 ti->error = "error getting device";
777 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
778 q = bdev_get_queue(p->path.dev->bdev);
780 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
782 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
783 if (attached_handler_name) {
785 * Clear any hw_handler_params associated with a
786 * handler that isn't already attached.
788 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
789 kfree(m->hw_handler_params);
790 m->hw_handler_params = NULL;
794 * Reset hw_handler_name to match the attached handler
796 * NB. This modifies the table line to show the actual
797 * handler instead of the original table passed in.
799 kfree(m->hw_handler_name);
800 m->hw_handler_name = attached_handler_name;
804 if (m->hw_handler_name) {
805 r = scsi_dh_attach(q, m->hw_handler_name);
807 char b[BDEVNAME_SIZE];
809 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
810 bdevname(p->path.dev->bdev, b));
814 ti->error = "error attaching hardware handler";
815 dm_put_device(ti, p->path.dev);
819 if (m->hw_handler_params) {
820 r = scsi_dh_set_params(q, m->hw_handler_params);
822 ti->error = "unable to set hardware "
823 "handler parameters";
824 dm_put_device(ti, p->path.dev);
830 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
832 dm_put_device(ti, p->path.dev);
843 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
846 static const struct dm_arg _args[] = {
847 {1, 1024, "invalid number of paths"},
848 {0, 1024, "invalid number of selector args"}
852 unsigned i, nr_selector_args, nr_args;
853 struct priority_group *pg;
854 struct dm_target *ti = m->ti;
858 ti->error = "not enough priority group arguments";
859 return ERR_PTR(-EINVAL);
862 pg = alloc_priority_group();
864 ti->error = "couldn't allocate priority group";
865 return ERR_PTR(-ENOMEM);
869 r = parse_path_selector(as, pg, ti);
876 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
880 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
884 nr_args = 1 + nr_selector_args;
885 for (i = 0; i < pg->nr_pgpaths; i++) {
886 struct pgpath *pgpath;
887 struct dm_arg_set path_args;
889 if (as->argc < nr_args) {
890 ti->error = "not enough path parameters";
895 path_args.argc = nr_args;
896 path_args.argv = as->argv;
898 pgpath = parse_path(&path_args, &pg->ps, ti);
899 if (IS_ERR(pgpath)) {
905 list_add_tail(&pgpath->list, &pg->pgpaths);
906 dm_consume_args(as, nr_args);
912 free_priority_group(pg, ti);
916 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
920 struct dm_target *ti = m->ti;
922 static const struct dm_arg _args[] = {
923 {0, 1024, "invalid number of hardware handler args"},
926 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
932 if (m->queue_mode == DM_TYPE_BIO_BASED) {
933 dm_consume_args(as, hw_argc);
934 DMERR("bio-based multipath doesn't allow hardware handler args");
938 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
939 if (!m->hw_handler_name)
946 for (i = 0; i <= hw_argc - 2; i++)
947 len += strlen(as->argv[i]) + 1;
948 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
950 ti->error = "memory allocation failed";
954 j = sprintf(p, "%d", hw_argc - 1);
955 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
956 j = sprintf(p, "%s", as->argv[i]);
958 dm_consume_args(as, hw_argc - 1);
962 kfree(m->hw_handler_name);
963 m->hw_handler_name = NULL;
967 static int parse_features(struct dm_arg_set *as, struct multipath *m)
971 struct dm_target *ti = m->ti;
972 const char *arg_name;
974 static const struct dm_arg _args[] = {
975 {0, 8, "invalid number of feature args"},
976 {1, 50, "pg_init_retries must be between 1 and 50"},
977 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
980 r = dm_read_arg_group(_args, as, &argc, &ti->error);
988 arg_name = dm_shift_arg(as);
991 if (!strcasecmp(arg_name, "queue_if_no_path")) {
992 r = queue_if_no_path(m, true, false);
996 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
997 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1001 if (!strcasecmp(arg_name, "pg_init_retries") &&
1003 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1008 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1010 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1015 if (!strcasecmp(arg_name, "queue_mode") &&
1017 const char *queue_mode_name = dm_shift_arg(as);
1019 if (!strcasecmp(queue_mode_name, "bio"))
1020 m->queue_mode = DM_TYPE_BIO_BASED;
1021 else if (!strcasecmp(queue_mode_name, "rq"))
1022 m->queue_mode = DM_TYPE_REQUEST_BASED;
1023 else if (!strcasecmp(queue_mode_name, "mq"))
1024 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1026 ti->error = "Unknown 'queue_mode' requested";
1033 ti->error = "Unrecognised multipath feature request";
1035 } while (argc && !r);
1040 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1042 /* target arguments */
1043 static const struct dm_arg _args[] = {
1044 {0, 1024, "invalid number of priority groups"},
1045 {0, 1024, "invalid initial priority group number"},
1049 struct multipath *m;
1050 struct dm_arg_set as;
1051 unsigned pg_count = 0;
1052 unsigned next_pg_num;
1057 m = alloc_multipath(ti);
1059 ti->error = "can't allocate multipath";
1063 r = parse_features(&as, m);
1067 r = alloc_multipath_stage2(ti, m);
1071 r = parse_hw_handler(&as, m);
1075 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1079 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1083 if ((!m->nr_priority_groups && next_pg_num) ||
1084 (m->nr_priority_groups && !next_pg_num)) {
1085 ti->error = "invalid initial priority group";
1090 /* parse the priority groups */
1092 struct priority_group *pg;
1093 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1095 pg = parse_priority_group(&as, m);
1101 nr_valid_paths += pg->nr_pgpaths;
1102 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1104 list_add_tail(&pg->list, &m->priority_groups);
1106 pg->pg_num = pg_count;
1111 if (pg_count != m->nr_priority_groups) {
1112 ti->error = "priority group count mismatch";
1117 ti->num_flush_bios = 1;
1118 ti->num_discard_bios = 1;
1119 ti->num_write_same_bios = 1;
1120 ti->num_write_zeroes_bios = 1;
1121 if (m->queue_mode == DM_TYPE_BIO_BASED)
1122 ti->per_io_data_size = multipath_per_bio_data_size();
1124 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1133 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1138 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1140 if (!atomic_read(&m->pg_init_in_progress))
1145 finish_wait(&m->pg_init_wait, &wait);
1148 static void flush_multipath_work(struct multipath *m)
1150 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1151 smp_mb__after_atomic();
1153 flush_workqueue(kmpath_handlerd);
1154 multipath_wait_for_pg_init_completion(m);
1155 flush_workqueue(kmultipathd);
1156 flush_work(&m->trigger_event);
1158 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1159 smp_mb__after_atomic();
1162 static void multipath_dtr(struct dm_target *ti)
1164 struct multipath *m = ti->private;
1166 flush_multipath_work(m);
1171 * Take a path out of use.
1173 static int fail_path(struct pgpath *pgpath)
1175 unsigned long flags;
1176 struct multipath *m = pgpath->pg->m;
1178 spin_lock_irqsave(&m->lock, flags);
1180 if (!pgpath->is_active)
1183 DMWARN("Failing path %s.", pgpath->path.dev->name);
1185 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1186 pgpath->is_active = false;
1187 pgpath->fail_count++;
1189 atomic_dec(&m->nr_valid_paths);
1191 if (pgpath == m->current_pgpath)
1192 m->current_pgpath = NULL;
1194 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1195 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1197 schedule_work(&m->trigger_event);
1200 spin_unlock_irqrestore(&m->lock, flags);
1206 * Reinstate a previously-failed path
1208 static int reinstate_path(struct pgpath *pgpath)
1210 int r = 0, run_queue = 0;
1211 unsigned long flags;
1212 struct multipath *m = pgpath->pg->m;
1213 unsigned nr_valid_paths;
1215 spin_lock_irqsave(&m->lock, flags);
1217 if (pgpath->is_active)
1220 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1222 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1226 pgpath->is_active = true;
1228 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1229 if (nr_valid_paths == 1) {
1230 m->current_pgpath = NULL;
1232 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1233 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1234 atomic_inc(&m->pg_init_in_progress);
1237 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1238 pgpath->path.dev->name, nr_valid_paths);
1240 schedule_work(&m->trigger_event);
1243 spin_unlock_irqrestore(&m->lock, flags);
1245 dm_table_run_md_queue_async(m->ti->table);
1246 process_queued_io_list(m);
1253 * Fail or reinstate all paths that match the provided struct dm_dev.
1255 static int action_dev(struct multipath *m, struct dm_dev *dev,
1259 struct pgpath *pgpath;
1260 struct priority_group *pg;
1262 list_for_each_entry(pg, &m->priority_groups, list) {
1263 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1264 if (pgpath->path.dev == dev)
1273 * Temporarily try to avoid having to use the specified PG
1275 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1278 unsigned long flags;
1280 spin_lock_irqsave(&m->lock, flags);
1282 pg->bypassed = bypassed;
1283 m->current_pgpath = NULL;
1284 m->current_pg = NULL;
1286 spin_unlock_irqrestore(&m->lock, flags);
1288 schedule_work(&m->trigger_event);
1292 * Switch to using the specified PG from the next I/O that gets mapped
1294 static int switch_pg_num(struct multipath *m, const char *pgstr)
1296 struct priority_group *pg;
1298 unsigned long flags;
1301 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1302 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1303 DMWARN("invalid PG number supplied to switch_pg_num");
1307 spin_lock_irqsave(&m->lock, flags);
1308 list_for_each_entry(pg, &m->priority_groups, list) {
1309 pg->bypassed = false;
1313 m->current_pgpath = NULL;
1314 m->current_pg = NULL;
1317 spin_unlock_irqrestore(&m->lock, flags);
1319 schedule_work(&m->trigger_event);
1324 * Set/clear bypassed status of a PG.
1325 * PGs are numbered upwards from 1 in the order they were declared.
1327 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1329 struct priority_group *pg;
1333 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1334 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1335 DMWARN("invalid PG number supplied to bypass_pg");
1339 list_for_each_entry(pg, &m->priority_groups, list) {
1344 bypass_pg(m, pg, bypassed);
1349 * Should we retry pg_init immediately?
1351 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1353 unsigned long flags;
1354 bool limit_reached = false;
1356 spin_lock_irqsave(&m->lock, flags);
1358 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1359 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1360 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1362 limit_reached = true;
1364 spin_unlock_irqrestore(&m->lock, flags);
1366 return limit_reached;
1369 static void pg_init_done(void *data, int errors)
1371 struct pgpath *pgpath = data;
1372 struct priority_group *pg = pgpath->pg;
1373 struct multipath *m = pg->m;
1374 unsigned long flags;
1375 bool delay_retry = false;
1377 /* device or driver problems */
1382 if (!m->hw_handler_name) {
1386 DMERR("Could not failover the device: Handler scsi_dh_%s "
1387 "Error %d.", m->hw_handler_name, errors);
1389 * Fail path for now, so we do not ping pong
1393 case SCSI_DH_DEV_TEMP_BUSY:
1395 * Probably doing something like FW upgrade on the
1396 * controller so try the other pg.
1398 bypass_pg(m, pg, true);
1401 /* Wait before retrying. */
1404 case SCSI_DH_IMM_RETRY:
1405 case SCSI_DH_RES_TEMP_UNAVAIL:
1406 if (pg_init_limit_reached(m, pgpath))
1410 case SCSI_DH_DEV_OFFLINED:
1413 * We probably do not want to fail the path for a device
1414 * error, but this is what the old dm did. In future
1415 * patches we can do more advanced handling.
1420 spin_lock_irqsave(&m->lock, flags);
1422 if (pgpath == m->current_pgpath) {
1423 DMERR("Could not failover device. Error %d.", errors);
1424 m->current_pgpath = NULL;
1425 m->current_pg = NULL;
1427 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1428 pg->bypassed = false;
1430 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1431 /* Activations of other paths are still on going */
1434 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1436 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1438 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1440 if (__pg_init_all_paths(m))
1443 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1445 process_queued_io_list(m);
1448 * Wake up any thread waiting to suspend.
1450 wake_up(&m->pg_init_wait);
1453 spin_unlock_irqrestore(&m->lock, flags);
1456 static void activate_or_offline_path(struct pgpath *pgpath)
1458 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1460 if (pgpath->is_active && !blk_queue_dying(q))
1461 scsi_dh_activate(q, pg_init_done, pgpath);
1463 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1466 static void activate_path_work(struct work_struct *work)
1468 struct pgpath *pgpath =
1469 container_of(work, struct pgpath, activate_path.work);
1471 activate_or_offline_path(pgpath);
1474 static int noretry_error(blk_status_t error)
1477 case BLK_STS_NOTSUPP:
1479 case BLK_STS_TARGET:
1481 case BLK_STS_MEDIUM:
1485 /* Anything else could be a path failure, so should be retried */
1489 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1490 blk_status_t error, union map_info *map_context)
1492 struct dm_mpath_io *mpio = get_mpio(map_context);
1493 struct pgpath *pgpath = mpio->pgpath;
1494 int r = DM_ENDIO_DONE;
1497 * We don't queue any clone request inside the multipath target
1498 * during end I/O handling, since those clone requests don't have
1499 * bio clones. If we queue them inside the multipath target,
1500 * we need to make bio clones, that requires memory allocation.
1501 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1502 * don't have bio clones.)
1503 * Instead of queueing the clone request here, we queue the original
1504 * request into dm core, which will remake a clone request and
1505 * clone bios for it and resubmit it later.
1507 if (error && !noretry_error(error)) {
1508 struct multipath *m = ti->private;
1510 r = DM_ENDIO_REQUEUE;
1515 if (atomic_read(&m->nr_valid_paths) == 0 &&
1516 !must_push_back_rq(m)) {
1517 if (error == BLK_STS_IOERR)
1519 /* complete with the original error */
1525 struct path_selector *ps = &pgpath->pg->ps;
1527 if (ps->type->end_io)
1528 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1534 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1535 blk_status_t *error)
1537 struct multipath *m = ti->private;
1538 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1539 struct pgpath *pgpath = mpio->pgpath;
1540 unsigned long flags;
1541 int r = DM_ENDIO_DONE;
1543 if (!*error || noretry_error(*error))
1549 if (atomic_read(&m->nr_valid_paths) == 0 &&
1550 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1551 if (must_push_back_bio(m)) {
1552 r = DM_ENDIO_REQUEUE;
1555 *error = BLK_STS_IOERR;
1560 /* Queue for the daemon to resubmit */
1561 dm_bio_restore(get_bio_details_from_bio(clone), clone);
1563 spin_lock_irqsave(&m->lock, flags);
1564 bio_list_add(&m->queued_bios, clone);
1565 spin_unlock_irqrestore(&m->lock, flags);
1566 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1567 queue_work(kmultipathd, &m->process_queued_bios);
1569 r = DM_ENDIO_INCOMPLETE;
1572 struct path_selector *ps = &pgpath->pg->ps;
1574 if (ps->type->end_io)
1575 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1582 * Suspend can't complete until all the I/O is processed so if
1583 * the last path fails we must error any remaining I/O.
1584 * Note that if the freeze_bdev fails while suspending, the
1585 * queue_if_no_path state is lost - userspace should reset it.
1587 static void multipath_presuspend(struct dm_target *ti)
1589 struct multipath *m = ti->private;
1591 queue_if_no_path(m, false, true);
1594 static void multipath_postsuspend(struct dm_target *ti)
1596 struct multipath *m = ti->private;
1598 mutex_lock(&m->work_mutex);
1599 flush_multipath_work(m);
1600 mutex_unlock(&m->work_mutex);
1604 * Restore the queue_if_no_path setting.
1606 static void multipath_resume(struct dm_target *ti)
1608 struct multipath *m = ti->private;
1609 unsigned long flags;
1611 spin_lock_irqsave(&m->lock, flags);
1612 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1613 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1614 spin_unlock_irqrestore(&m->lock, flags);
1618 * Info output has the following format:
1619 * num_multipath_feature_args [multipath_feature_args]*
1620 * num_handler_status_args [handler_status_args]*
1621 * num_groups init_group_number
1622 * [A|D|E num_ps_status_args [ps_status_args]*
1623 * num_paths num_selector_args
1624 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1626 * Table output has the following format (identical to the constructor string):
1627 * num_feature_args [features_args]*
1628 * num_handler_args hw_handler [hw_handler_args]*
1629 * num_groups init_group_number
1630 * [priority selector-name num_ps_args [ps_args]*
1631 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1633 static void multipath_status(struct dm_target *ti, status_type_t type,
1634 unsigned status_flags, char *result, unsigned maxlen)
1637 unsigned long flags;
1638 struct multipath *m = ti->private;
1639 struct priority_group *pg;
1644 spin_lock_irqsave(&m->lock, flags);
1647 if (type == STATUSTYPE_INFO)
1648 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1649 atomic_read(&m->pg_init_count));
1651 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1652 (m->pg_init_retries > 0) * 2 +
1653 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1654 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1655 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1657 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1658 DMEMIT("queue_if_no_path ");
1659 if (m->pg_init_retries)
1660 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1661 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1662 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1663 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1664 DMEMIT("retain_attached_hw_handler ");
1665 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1666 switch(m->queue_mode) {
1667 case DM_TYPE_BIO_BASED:
1668 DMEMIT("queue_mode bio ");
1670 case DM_TYPE_MQ_REQUEST_BASED:
1671 DMEMIT("queue_mode mq ");
1680 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1683 DMEMIT("1 %s ", m->hw_handler_name);
1685 DMEMIT("%u ", m->nr_priority_groups);
1688 pg_num = m->next_pg->pg_num;
1689 else if (m->current_pg)
1690 pg_num = m->current_pg->pg_num;
1692 pg_num = (m->nr_priority_groups ? 1 : 0);
1694 DMEMIT("%u ", pg_num);
1697 case STATUSTYPE_INFO:
1698 list_for_each_entry(pg, &m->priority_groups, list) {
1700 state = 'D'; /* Disabled */
1701 else if (pg == m->current_pg)
1702 state = 'A'; /* Currently Active */
1704 state = 'E'; /* Enabled */
1706 DMEMIT("%c ", state);
1708 if (pg->ps.type->status)
1709 sz += pg->ps.type->status(&pg->ps, NULL, type,
1715 DMEMIT("%u %u ", pg->nr_pgpaths,
1716 pg->ps.type->info_args);
1718 list_for_each_entry(p, &pg->pgpaths, list) {
1719 DMEMIT("%s %s %u ", p->path.dev->name,
1720 p->is_active ? "A" : "F",
1722 if (pg->ps.type->status)
1723 sz += pg->ps.type->status(&pg->ps,
1724 &p->path, type, result + sz,
1730 case STATUSTYPE_TABLE:
1731 list_for_each_entry(pg, &m->priority_groups, list) {
1732 DMEMIT("%s ", pg->ps.type->name);
1734 if (pg->ps.type->status)
1735 sz += pg->ps.type->status(&pg->ps, NULL, type,
1741 DMEMIT("%u %u ", pg->nr_pgpaths,
1742 pg->ps.type->table_args);
1744 list_for_each_entry(p, &pg->pgpaths, list) {
1745 DMEMIT("%s ", p->path.dev->name);
1746 if (pg->ps.type->status)
1747 sz += pg->ps.type->status(&pg->ps,
1748 &p->path, type, result + sz,
1755 spin_unlock_irqrestore(&m->lock, flags);
1758 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1762 struct multipath *m = ti->private;
1765 mutex_lock(&m->work_mutex);
1767 if (dm_suspended(ti)) {
1773 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1774 r = queue_if_no_path(m, true, false);
1776 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1777 r = queue_if_no_path(m, false, false);
1783 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1787 if (!strcasecmp(argv[0], "disable_group")) {
1788 r = bypass_pg_num(m, argv[1], true);
1790 } else if (!strcasecmp(argv[0], "enable_group")) {
1791 r = bypass_pg_num(m, argv[1], false);
1793 } else if (!strcasecmp(argv[0], "switch_group")) {
1794 r = switch_pg_num(m, argv[1]);
1796 } else if (!strcasecmp(argv[0], "reinstate_path"))
1797 action = reinstate_path;
1798 else if (!strcasecmp(argv[0], "fail_path"))
1801 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1805 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1807 DMWARN("message: error getting device %s",
1812 r = action_dev(m, dev, action);
1814 dm_put_device(ti, dev);
1817 mutex_unlock(&m->work_mutex);
1821 static int multipath_prepare_ioctl(struct dm_target *ti,
1822 struct block_device **bdev, fmode_t *mode)
1824 struct multipath *m = ti->private;
1825 struct pgpath *current_pgpath;
1828 current_pgpath = READ_ONCE(m->current_pgpath);
1829 if (!current_pgpath)
1830 current_pgpath = choose_pgpath(m, 0);
1832 if (current_pgpath) {
1833 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1834 *bdev = current_pgpath->path.dev->bdev;
1835 *mode = current_pgpath->path.dev->mode;
1838 /* pg_init has not started or completed */
1842 /* No path is available */
1843 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1849 if (r == -ENOTCONN) {
1850 if (!READ_ONCE(m->current_pg)) {
1851 /* Path status changed, redo selection */
1852 (void) choose_pgpath(m, 0);
1854 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1855 pg_init_all_paths(m);
1856 dm_table_run_md_queue_async(m->ti->table);
1857 process_queued_io_list(m);
1861 * Only pass ioctls through if the device sizes match exactly.
1863 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1868 static int multipath_iterate_devices(struct dm_target *ti,
1869 iterate_devices_callout_fn fn, void *data)
1871 struct multipath *m = ti->private;
1872 struct priority_group *pg;
1876 list_for_each_entry(pg, &m->priority_groups, list) {
1877 list_for_each_entry(p, &pg->pgpaths, list) {
1878 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1888 static int pgpath_busy(struct pgpath *pgpath)
1890 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1892 return blk_lld_busy(q);
1896 * We return "busy", only when we can map I/Os but underlying devices
1897 * are busy (so even if we map I/Os now, the I/Os will wait on
1898 * the underlying queue).
1899 * In other words, if we want to kill I/Os or queue them inside us
1900 * due to map unavailability, we don't return "busy". Otherwise,
1901 * dm core won't give us the I/Os and we can't do what we want.
1903 static int multipath_busy(struct dm_target *ti)
1905 bool busy = false, has_active = false;
1906 struct multipath *m = ti->private;
1907 struct priority_group *pg, *next_pg;
1908 struct pgpath *pgpath;
1910 /* pg_init in progress */
1911 if (atomic_read(&m->pg_init_in_progress))
1914 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1915 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1916 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1918 /* Guess which priority_group will be used at next mapping time */
1919 pg = READ_ONCE(m->current_pg);
1920 next_pg = READ_ONCE(m->next_pg);
1921 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1926 * We don't know which pg will be used at next mapping time.
1927 * We don't call choose_pgpath() here to avoid to trigger
1928 * pg_init just by busy checking.
1929 * So we don't know whether underlying devices we will be using
1930 * at next mapping time are busy or not. Just try mapping.
1936 * If there is one non-busy active path at least, the path selector
1937 * will be able to select it. So we consider such a pg as not busy.
1940 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1941 if (pgpath->is_active) {
1943 if (!pgpath_busy(pgpath)) {
1952 * No active path in this pg, so this pg won't be used and
1953 * the current_pg will be changed at next mapping time.
1954 * We need to try mapping to determine it.
1962 /*-----------------------------------------------------------------
1964 *---------------------------------------------------------------*/
1965 static struct target_type multipath_target = {
1966 .name = "multipath",
1967 .version = {1, 12, 0},
1968 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1969 .module = THIS_MODULE,
1970 .ctr = multipath_ctr,
1971 .dtr = multipath_dtr,
1972 .clone_and_map_rq = multipath_clone_and_map,
1973 .release_clone_rq = multipath_release_clone,
1974 .rq_end_io = multipath_end_io,
1975 .map = multipath_map_bio,
1976 .end_io = multipath_end_io_bio,
1977 .presuspend = multipath_presuspend,
1978 .postsuspend = multipath_postsuspend,
1979 .resume = multipath_resume,
1980 .status = multipath_status,
1981 .message = multipath_message,
1982 .prepare_ioctl = multipath_prepare_ioctl,
1983 .iterate_devices = multipath_iterate_devices,
1984 .busy = multipath_busy,
1987 static int __init dm_multipath_init(void)
1991 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1993 DMERR("failed to create workqueue kmpathd");
1995 goto bad_alloc_kmultipathd;
1999 * A separate workqueue is used to handle the device handlers
2000 * to avoid overloading existing workqueue. Overloading the
2001 * old workqueue would also create a bottleneck in the
2002 * path of the storage hardware device activation.
2004 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2006 if (!kmpath_handlerd) {
2007 DMERR("failed to create workqueue kmpath_handlerd");
2009 goto bad_alloc_kmpath_handlerd;
2012 r = dm_register_target(&multipath_target);
2014 DMERR("request-based register failed %d", r);
2016 goto bad_register_target;
2021 bad_register_target:
2022 destroy_workqueue(kmpath_handlerd);
2023 bad_alloc_kmpath_handlerd:
2024 destroy_workqueue(kmultipathd);
2025 bad_alloc_kmultipathd:
2029 static void __exit dm_multipath_exit(void)
2031 destroy_workqueue(kmpath_handlerd);
2032 destroy_workqueue(kmultipathd);
2034 dm_unregister_target(&multipath_target);
2037 module_init(dm_multipath_init);
2038 module_exit(dm_multipath_exit);
2040 MODULE_DESCRIPTION(DM_NAME " multipath target");
2041 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2042 MODULE_LICENSE("GPL");