1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2003 Sistina Software Limited.
4 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
6 * This file is released under the GPL.
9 #include <linux/device-mapper.h>
12 #include "dm-bio-record.h"
13 #include "dm-path-selector.h"
14 #include "dm-uevent.h"
16 #include <linux/blkdev.h>
17 #include <linux/ctype.h>
18 #include <linux/init.h>
19 #include <linux/mempool.h>
20 #include <linux/module.h>
21 #include <linux/pagemap.h>
22 #include <linux/slab.h>
23 #include <linux/time.h>
24 #include <linux/timer.h>
25 #include <linux/workqueue.h>
26 #include <linux/delay.h>
27 #include <scsi/scsi_dh.h>
28 #include <linux/atomic.h>
29 #include <linux/blk-mq.h>
31 #define DM_MSG_PREFIX "multipath"
32 #define DM_PG_INIT_DELAY_MSECS 2000
33 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned int) -1)
34 #define QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT 0
36 static unsigned long queue_if_no_path_timeout_secs = QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT;
40 struct list_head list;
42 struct priority_group *pg; /* Owning PG */
43 unsigned int fail_count; /* Cumulative failure count */
46 struct delayed_work activate_path;
48 bool is_active:1; /* Path status */
51 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
54 * Paths are grouped into Priority Groups and numbered from 1 upwards.
55 * Each has a path selector which controls which path gets used.
57 struct priority_group {
58 struct list_head list;
60 struct multipath *m; /* Owning multipath instance */
61 struct path_selector ps;
63 unsigned int pg_num; /* Reference number */
64 unsigned int nr_pgpaths; /* Number of paths in PG */
65 struct list_head pgpaths;
67 bool bypassed:1; /* Temporarily bypass this PG? */
70 /* Multipath context */
72 unsigned long flags; /* Multipath state flags */
75 enum dm_queue_mode queue_mode;
77 struct pgpath *current_pgpath;
78 struct priority_group *current_pg;
79 struct priority_group *next_pg; /* Switch to this PG if set */
81 atomic_t nr_valid_paths; /* Total number of usable paths */
82 unsigned int nr_priority_groups;
83 struct list_head priority_groups;
85 const char *hw_handler_name;
86 char *hw_handler_params;
87 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
88 unsigned int pg_init_retries; /* Number of times to retry pg_init */
89 unsigned int pg_init_delay_msecs; /* Number of msecs before pg_init retry */
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 struct mutex work_mutex;
94 struct work_struct trigger_event;
97 struct work_struct process_queued_bios;
98 struct bio_list queued_bios;
100 struct timer_list nopath_timer; /* Timeout for queue_if_no_path */
104 * Context information attached to each io we process.
107 struct pgpath *pgpath;
112 typedef int (*action_fn) (struct pgpath *pgpath);
114 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
115 static void trigger_event(struct work_struct *work);
116 static void activate_or_offline_path(struct pgpath *pgpath);
117 static void activate_path_work(struct work_struct *work);
118 static void process_queued_bios(struct work_struct *work);
119 static void queue_if_no_path_timeout_work(struct timer_list *t);
122 *-----------------------------------------------
123 * Multipath state flags.
124 *-----------------------------------------------
126 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
127 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
128 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
129 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
130 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
131 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
132 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
134 static bool mpath_double_check_test_bit(int MPATHF_bit, struct multipath *m)
136 bool r = test_bit(MPATHF_bit, &m->flags);
140 spin_lock_irqsave(&m->lock, flags);
141 r = test_bit(MPATHF_bit, &m->flags);
142 spin_unlock_irqrestore(&m->lock, flags);
149 *-----------------------------------------------
150 * Allocation routines
151 *-----------------------------------------------
153 static struct pgpath *alloc_pgpath(void)
155 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
160 pgpath->is_active = true;
165 static void free_pgpath(struct pgpath *pgpath)
170 static struct priority_group *alloc_priority_group(void)
172 struct priority_group *pg;
174 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
177 INIT_LIST_HEAD(&pg->pgpaths);
182 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
184 struct pgpath *pgpath, *tmp;
186 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
187 list_del(&pgpath->list);
188 dm_put_device(ti, pgpath->path.dev);
193 static void free_priority_group(struct priority_group *pg,
194 struct dm_target *ti)
196 struct path_selector *ps = &pg->ps;
199 ps->type->destroy(ps);
200 dm_put_path_selector(ps->type);
203 free_pgpaths(&pg->pgpaths, ti);
207 static struct multipath *alloc_multipath(struct dm_target *ti)
211 m = kzalloc(sizeof(*m), GFP_KERNEL);
213 INIT_LIST_HEAD(&m->priority_groups);
214 spin_lock_init(&m->lock);
215 atomic_set(&m->nr_valid_paths, 0);
216 INIT_WORK(&m->trigger_event, trigger_event);
217 mutex_init(&m->work_mutex);
219 m->queue_mode = DM_TYPE_NONE;
224 timer_setup(&m->nopath_timer, queue_if_no_path_timeout_work, 0);
230 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
232 if (m->queue_mode == DM_TYPE_NONE) {
233 m->queue_mode = DM_TYPE_REQUEST_BASED;
234 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
235 INIT_WORK(&m->process_queued_bios, process_queued_bios);
237 * bio-based doesn't support any direct scsi_dh management;
238 * it just discovers if a scsi_dh is attached.
240 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
243 dm_table_set_type(ti->table, m->queue_mode);
246 * Init fields that are only used when a scsi_dh is attached
247 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
249 set_bit(MPATHF_QUEUE_IO, &m->flags);
250 atomic_set(&m->pg_init_in_progress, 0);
251 atomic_set(&m->pg_init_count, 0);
252 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
253 init_waitqueue_head(&m->pg_init_wait);
258 static void free_multipath(struct multipath *m)
260 struct priority_group *pg, *tmp;
262 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
264 free_priority_group(pg, m->ti);
267 kfree(m->hw_handler_name);
268 kfree(m->hw_handler_params);
269 mutex_destroy(&m->work_mutex);
273 static struct dm_mpath_io *get_mpio(union map_info *info)
278 static size_t multipath_per_bio_data_size(void)
280 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
283 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
285 return dm_per_bio_data(bio, multipath_per_bio_data_size());
288 static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
290 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
291 void *bio_details = mpio + 1;
295 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
297 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
298 struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
300 mpio->nr_bytes = bio->bi_iter.bi_size;
302 mpio->start_time_ns = 0;
305 dm_bio_record(bio_details, bio);
309 *-----------------------------------------------
311 *-----------------------------------------------
313 static int __pg_init_all_paths(struct multipath *m)
315 struct pgpath *pgpath;
316 unsigned long pg_init_delay = 0;
318 lockdep_assert_held(&m->lock);
320 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
323 atomic_inc(&m->pg_init_count);
324 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
326 /* Check here to reset pg_init_required */
330 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
331 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
332 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
333 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
334 /* Skip failed paths */
335 if (!pgpath->is_active)
337 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
339 atomic_inc(&m->pg_init_in_progress);
341 return atomic_read(&m->pg_init_in_progress);
344 static int pg_init_all_paths(struct multipath *m)
349 spin_lock_irqsave(&m->lock, flags);
350 ret = __pg_init_all_paths(m);
351 spin_unlock_irqrestore(&m->lock, flags);
356 static void __switch_pg(struct multipath *m, struct priority_group *pg)
358 lockdep_assert_held(&m->lock);
362 /* Must we initialise the PG first, and queue I/O till it's ready? */
363 if (m->hw_handler_name) {
364 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
365 set_bit(MPATHF_QUEUE_IO, &m->flags);
367 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
368 clear_bit(MPATHF_QUEUE_IO, &m->flags);
371 atomic_set(&m->pg_init_count, 0);
374 static struct pgpath *choose_path_in_pg(struct multipath *m,
375 struct priority_group *pg,
379 struct dm_path *path;
380 struct pgpath *pgpath;
382 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
384 return ERR_PTR(-ENXIO);
386 pgpath = path_to_pgpath(path);
388 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
389 /* Only update current_pgpath if pg changed */
390 spin_lock_irqsave(&m->lock, flags);
391 m->current_pgpath = pgpath;
393 spin_unlock_irqrestore(&m->lock, flags);
399 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
402 struct priority_group *pg;
403 struct pgpath *pgpath;
404 unsigned int bypassed = 1;
406 if (!atomic_read(&m->nr_valid_paths)) {
407 spin_lock_irqsave(&m->lock, flags);
408 clear_bit(MPATHF_QUEUE_IO, &m->flags);
409 spin_unlock_irqrestore(&m->lock, flags);
413 /* Were we instructed to switch PG? */
414 if (READ_ONCE(m->next_pg)) {
415 spin_lock_irqsave(&m->lock, flags);
418 spin_unlock_irqrestore(&m->lock, flags);
419 goto check_current_pg;
422 spin_unlock_irqrestore(&m->lock, flags);
423 pgpath = choose_path_in_pg(m, pg, nr_bytes);
424 if (!IS_ERR_OR_NULL(pgpath))
428 /* Don't change PG until it has no remaining paths */
430 pg = READ_ONCE(m->current_pg);
432 pgpath = choose_path_in_pg(m, pg, nr_bytes);
433 if (!IS_ERR_OR_NULL(pgpath))
438 * Loop through priority groups until we find a valid path.
439 * First time we skip PGs marked 'bypassed'.
440 * Second time we only try the ones we skipped, but set
441 * pg_init_delay_retry so we do not hammer controllers.
444 list_for_each_entry(pg, &m->priority_groups, list) {
445 if (pg->bypassed == !!bypassed)
447 pgpath = choose_path_in_pg(m, pg, nr_bytes);
448 if (!IS_ERR_OR_NULL(pgpath)) {
450 spin_lock_irqsave(&m->lock, flags);
451 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
452 spin_unlock_irqrestore(&m->lock, flags);
457 } while (bypassed--);
460 spin_lock_irqsave(&m->lock, flags);
461 m->current_pgpath = NULL;
462 m->current_pg = NULL;
463 spin_unlock_irqrestore(&m->lock, flags);
469 * dm_report_EIO() is a macro instead of a function to make pr_debug_ratelimited()
470 * report the function name and line number of the function from which
471 * it has been invoked.
473 #define dm_report_EIO(m) \
475 DMDEBUG_LIMIT("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d", \
476 dm_table_device_name((m)->ti->table), \
477 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
478 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
479 dm_noflush_suspending((m)->ti)); \
483 * Check whether bios must be queued in the device-mapper core rather
484 * than here in the target.
486 static bool __must_push_back(struct multipath *m)
488 return dm_noflush_suspending(m->ti);
491 static bool must_push_back_rq(struct multipath *m)
496 spin_lock_irqsave(&m->lock, flags);
497 ret = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) || __must_push_back(m));
498 spin_unlock_irqrestore(&m->lock, flags);
504 * Map cloned requests (request-based multipath)
506 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
507 union map_info *map_context,
508 struct request **__clone)
510 struct multipath *m = ti->private;
511 size_t nr_bytes = blk_rq_bytes(rq);
512 struct pgpath *pgpath;
513 struct block_device *bdev;
514 struct dm_mpath_io *mpio = get_mpio(map_context);
515 struct request_queue *q;
516 struct request *clone;
518 /* Do we need to select a new pgpath? */
519 pgpath = READ_ONCE(m->current_pgpath);
520 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
521 pgpath = choose_pgpath(m, nr_bytes);
524 if (must_push_back_rq(m))
525 return DM_MAPIO_DELAY_REQUEUE;
526 dm_report_EIO(m); /* Failed */
527 return DM_MAPIO_KILL;
528 } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
529 mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
530 pg_init_all_paths(m);
531 return DM_MAPIO_DELAY_REQUEUE;
534 mpio->pgpath = pgpath;
535 mpio->nr_bytes = nr_bytes;
537 bdev = pgpath->path.dev->bdev;
538 q = bdev_get_queue(bdev);
539 clone = blk_mq_alloc_request(q, rq->cmd_flags | REQ_NOMERGE,
542 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
543 if (blk_queue_dying(q)) {
544 atomic_inc(&m->pg_init_in_progress);
545 activate_or_offline_path(pgpath);
546 return DM_MAPIO_DELAY_REQUEUE;
550 * blk-mq's SCHED_RESTART can cover this requeue, so we
551 * needn't deal with it by DELAY_REQUEUE. More importantly,
552 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
553 * get the queue busy feedback (via BLK_STS_RESOURCE),
554 * otherwise I/O merging can suffer.
556 return DM_MAPIO_REQUEUE;
558 clone->bio = clone->biotail = NULL;
559 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
562 if (pgpath->pg->ps.type->start_io)
563 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
566 return DM_MAPIO_REMAPPED;
569 static void multipath_release_clone(struct request *clone,
570 union map_info *map_context)
572 if (unlikely(map_context)) {
574 * non-NULL map_context means caller is still map
575 * method; must undo multipath_clone_and_map()
577 struct dm_mpath_io *mpio = get_mpio(map_context);
578 struct pgpath *pgpath = mpio->pgpath;
580 if (pgpath && pgpath->pg->ps.type->end_io)
581 pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
584 clone->io_start_time_ns);
587 blk_mq_free_request(clone);
591 * Map cloned bios (bio-based multipath)
594 static void __multipath_queue_bio(struct multipath *m, struct bio *bio)
596 /* Queue for the daemon to resubmit */
597 bio_list_add(&m->queued_bios, bio);
598 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
599 queue_work(kmultipathd, &m->process_queued_bios);
602 static void multipath_queue_bio(struct multipath *m, struct bio *bio)
606 spin_lock_irqsave(&m->lock, flags);
607 __multipath_queue_bio(m, bio);
608 spin_unlock_irqrestore(&m->lock, flags);
611 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
613 struct pgpath *pgpath;
616 /* Do we need to select a new pgpath? */
617 pgpath = READ_ONCE(m->current_pgpath);
618 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
619 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
622 spin_lock_irqsave(&m->lock, flags);
623 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
624 __multipath_queue_bio(m, bio);
625 pgpath = ERR_PTR(-EAGAIN);
627 spin_unlock_irqrestore(&m->lock, flags);
629 } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
630 mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
631 multipath_queue_bio(m, bio);
632 pg_init_all_paths(m);
633 return ERR_PTR(-EAGAIN);
639 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
640 struct dm_mpath_io *mpio)
642 struct pgpath *pgpath = __map_bio(m, bio);
645 return DM_MAPIO_SUBMITTED;
648 if (__must_push_back(m))
649 return DM_MAPIO_REQUEUE;
651 return DM_MAPIO_KILL;
654 mpio->pgpath = pgpath;
656 if (dm_ps_use_hr_timer(pgpath->pg->ps.type))
657 mpio->start_time_ns = ktime_get_ns();
660 bio_set_dev(bio, pgpath->path.dev->bdev);
661 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
663 if (pgpath->pg->ps.type->start_io)
664 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
667 return DM_MAPIO_REMAPPED;
670 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
672 struct multipath *m = ti->private;
673 struct dm_mpath_io *mpio = NULL;
675 multipath_init_per_bio_data(bio, &mpio);
676 return __multipath_map_bio(m, bio, mpio);
679 static void process_queued_io_list(struct multipath *m)
681 if (m->queue_mode == DM_TYPE_REQUEST_BASED)
682 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
683 else if (m->queue_mode == DM_TYPE_BIO_BASED)
684 queue_work(kmultipathd, &m->process_queued_bios);
687 static void process_queued_bios(struct work_struct *work)
692 struct bio_list bios;
693 struct blk_plug plug;
694 struct multipath *m =
695 container_of(work, struct multipath, process_queued_bios);
697 bio_list_init(&bios);
699 spin_lock_irqsave(&m->lock, flags);
701 if (bio_list_empty(&m->queued_bios)) {
702 spin_unlock_irqrestore(&m->lock, flags);
706 bio_list_merge(&bios, &m->queued_bios);
707 bio_list_init(&m->queued_bios);
709 spin_unlock_irqrestore(&m->lock, flags);
711 blk_start_plug(&plug);
712 while ((bio = bio_list_pop(&bios))) {
713 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
714 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
715 r = __multipath_map_bio(m, bio, mpio);
718 bio->bi_status = BLK_STS_IOERR;
721 case DM_MAPIO_REQUEUE:
722 bio->bi_status = BLK_STS_DM_REQUEUE;
725 case DM_MAPIO_REMAPPED:
726 submit_bio_noacct(bio);
728 case DM_MAPIO_SUBMITTED:
731 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
734 blk_finish_plug(&plug);
738 * If we run out of usable paths, should we queue I/O or error it?
740 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
741 bool save_old_value, const char *caller)
744 bool queue_if_no_path_bit, saved_queue_if_no_path_bit;
745 const char *dm_dev_name = dm_table_device_name(m->ti->table);
747 DMDEBUG("%s: %s caller=%s queue_if_no_path=%d save_old_value=%d",
748 dm_dev_name, __func__, caller, queue_if_no_path, save_old_value);
750 spin_lock_irqsave(&m->lock, flags);
752 queue_if_no_path_bit = test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
753 saved_queue_if_no_path_bit = test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
755 if (save_old_value) {
756 if (unlikely(!queue_if_no_path_bit && saved_queue_if_no_path_bit)) {
757 DMERR("%s: QIFNP disabled but saved as enabled, saving again loses state, not saving!",
760 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path_bit);
761 } else if (!queue_if_no_path && saved_queue_if_no_path_bit) {
762 /* due to "fail_if_no_path" message, need to honor it. */
763 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
765 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
767 DMDEBUG("%s: after %s changes; QIFNP = %d; SQIFNP = %d; DNFS = %d",
768 dm_dev_name, __func__,
769 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
770 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
771 dm_noflush_suspending(m->ti));
773 spin_unlock_irqrestore(&m->lock, flags);
775 if (!queue_if_no_path) {
776 dm_table_run_md_queue_async(m->ti->table);
777 process_queued_io_list(m);
784 * If the queue_if_no_path timeout fires, turn off queue_if_no_path and
785 * process any queued I/O.
787 static void queue_if_no_path_timeout_work(struct timer_list *t)
789 struct multipath *m = from_timer(m, t, nopath_timer);
791 DMWARN("queue_if_no_path timeout on %s, failing queued IO",
792 dm_table_device_name(m->ti->table));
793 queue_if_no_path(m, false, false, __func__);
797 * Enable the queue_if_no_path timeout if necessary.
798 * Called with m->lock held.
800 static void enable_nopath_timeout(struct multipath *m)
802 unsigned long queue_if_no_path_timeout =
803 READ_ONCE(queue_if_no_path_timeout_secs) * HZ;
805 lockdep_assert_held(&m->lock);
807 if (queue_if_no_path_timeout > 0 &&
808 atomic_read(&m->nr_valid_paths) == 0 &&
809 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
810 mod_timer(&m->nopath_timer,
811 jiffies + queue_if_no_path_timeout);
815 static void disable_nopath_timeout(struct multipath *m)
817 del_timer_sync(&m->nopath_timer);
821 * An event is triggered whenever a path is taken out of use.
822 * Includes path failure and PG bypass.
824 static void trigger_event(struct work_struct *work)
826 struct multipath *m =
827 container_of(work, struct multipath, trigger_event);
829 dm_table_event(m->ti->table);
833 *---------------------------------------------------------------
834 * Constructor/argument parsing:
835 * <#multipath feature args> [<arg>]*
836 * <#hw_handler args> [hw_handler [<arg>]*]
838 * <initial priority group>
839 * [<selector> <#selector args> [<arg>]*
840 * <#paths> <#per-path selector args>
841 * [<path> [<arg>]* ]+ ]+
842 *---------------------------------------------------------------
844 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
845 struct dm_target *ti)
848 struct path_selector_type *pst;
849 unsigned int ps_argc;
851 static const struct dm_arg _args[] = {
852 {0, 1024, "invalid number of path selector args"},
855 pst = dm_get_path_selector(dm_shift_arg(as));
857 ti->error = "unknown path selector type";
861 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
863 dm_put_path_selector(pst);
867 r = pst->create(&pg->ps, ps_argc, as->argv);
869 dm_put_path_selector(pst);
870 ti->error = "path selector constructor failed";
875 dm_consume_args(as, ps_argc);
880 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
881 const char **attached_handler_name, char **error)
883 struct request_queue *q = bdev_get_queue(bdev);
886 if (mpath_double_check_test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, m)) {
888 if (*attached_handler_name) {
890 * Clear any hw_handler_params associated with a
891 * handler that isn't already attached.
893 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
894 kfree(m->hw_handler_params);
895 m->hw_handler_params = NULL;
899 * Reset hw_handler_name to match the attached handler
901 * NB. This modifies the table line to show the actual
902 * handler instead of the original table passed in.
904 kfree(m->hw_handler_name);
905 m->hw_handler_name = *attached_handler_name;
906 *attached_handler_name = NULL;
910 if (m->hw_handler_name) {
911 r = scsi_dh_attach(q, m->hw_handler_name);
913 DMINFO("retaining handler on device %pg", bdev);
917 *error = "error attaching hardware handler";
921 if (m->hw_handler_params) {
922 r = scsi_dh_set_params(q, m->hw_handler_params);
924 *error = "unable to set hardware handler parameters";
933 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
934 struct dm_target *ti)
938 struct multipath *m = ti->private;
939 struct request_queue *q;
940 const char *attached_handler_name = NULL;
942 /* we need at least a path arg */
944 ti->error = "no device given";
945 return ERR_PTR(-EINVAL);
950 return ERR_PTR(-ENOMEM);
952 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
955 ti->error = "error getting device";
959 q = bdev_get_queue(p->path.dev->bdev);
960 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
961 if (attached_handler_name || m->hw_handler_name) {
962 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
963 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
964 kfree(attached_handler_name);
966 dm_put_device(ti, p->path.dev);
971 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
973 dm_put_device(ti, p->path.dev);
983 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
986 static const struct dm_arg _args[] = {
987 {1, 1024, "invalid number of paths"},
988 {0, 1024, "invalid number of selector args"}
992 unsigned int i, nr_selector_args, nr_args;
993 struct priority_group *pg;
994 struct dm_target *ti = m->ti;
998 ti->error = "not enough priority group arguments";
999 return ERR_PTR(-EINVAL);
1002 pg = alloc_priority_group();
1004 ti->error = "couldn't allocate priority group";
1005 return ERR_PTR(-ENOMEM);
1009 r = parse_path_selector(as, pg, ti);
1016 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1020 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1024 nr_args = 1 + nr_selector_args;
1025 for (i = 0; i < pg->nr_pgpaths; i++) {
1026 struct pgpath *pgpath;
1027 struct dm_arg_set path_args;
1029 if (as->argc < nr_args) {
1030 ti->error = "not enough path parameters";
1035 path_args.argc = nr_args;
1036 path_args.argv = as->argv;
1038 pgpath = parse_path(&path_args, &pg->ps, ti);
1039 if (IS_ERR(pgpath)) {
1040 r = PTR_ERR(pgpath);
1045 list_add_tail(&pgpath->list, &pg->pgpaths);
1046 dm_consume_args(as, nr_args);
1052 free_priority_group(pg, ti);
1056 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1058 unsigned int hw_argc;
1060 struct dm_target *ti = m->ti;
1062 static const struct dm_arg _args[] = {
1063 {0, 1024, "invalid number of hardware handler args"},
1066 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1072 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1073 dm_consume_args(as, hw_argc);
1074 DMERR("bio-based multipath doesn't allow hardware handler args");
1078 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1079 if (!m->hw_handler_name)
1086 for (i = 0; i <= hw_argc - 2; i++)
1087 len += strlen(as->argv[i]) + 1;
1088 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1090 ti->error = "memory allocation failed";
1094 j = sprintf(p, "%d", hw_argc - 1);
1095 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1096 j = sprintf(p, "%s", as->argv[i]);
1098 dm_consume_args(as, hw_argc - 1);
1102 kfree(m->hw_handler_name);
1103 m->hw_handler_name = NULL;
1107 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1111 struct dm_target *ti = m->ti;
1112 const char *arg_name;
1114 static const struct dm_arg _args[] = {
1115 {0, 8, "invalid number of feature args"},
1116 {1, 50, "pg_init_retries must be between 1 and 50"},
1117 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1120 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1128 arg_name = dm_shift_arg(as);
1131 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1132 r = queue_if_no_path(m, true, false, __func__);
1136 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1137 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1141 if (!strcasecmp(arg_name, "pg_init_retries") &&
1143 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1148 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1150 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1155 if (!strcasecmp(arg_name, "queue_mode") &&
1157 const char *queue_mode_name = dm_shift_arg(as);
1159 if (!strcasecmp(queue_mode_name, "bio"))
1160 m->queue_mode = DM_TYPE_BIO_BASED;
1161 else if (!strcasecmp(queue_mode_name, "rq") ||
1162 !strcasecmp(queue_mode_name, "mq"))
1163 m->queue_mode = DM_TYPE_REQUEST_BASED;
1165 ti->error = "Unknown 'queue_mode' requested";
1172 ti->error = "Unrecognised multipath feature request";
1174 } while (argc && !r);
1179 static int multipath_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1181 /* target arguments */
1182 static const struct dm_arg _args[] = {
1183 {0, 1024, "invalid number of priority groups"},
1184 {0, 1024, "invalid initial priority group number"},
1188 struct multipath *m;
1189 struct dm_arg_set as;
1190 unsigned int pg_count = 0;
1191 unsigned int next_pg_num;
1192 unsigned long flags;
1197 m = alloc_multipath(ti);
1199 ti->error = "can't allocate multipath";
1203 r = parse_features(&as, m);
1207 r = alloc_multipath_stage2(ti, m);
1211 r = parse_hw_handler(&as, m);
1215 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1219 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1223 if ((!m->nr_priority_groups && next_pg_num) ||
1224 (m->nr_priority_groups && !next_pg_num)) {
1225 ti->error = "invalid initial priority group";
1230 /* parse the priority groups */
1232 struct priority_group *pg;
1233 unsigned int nr_valid_paths = atomic_read(&m->nr_valid_paths);
1235 pg = parse_priority_group(&as, m);
1241 nr_valid_paths += pg->nr_pgpaths;
1242 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1244 list_add_tail(&pg->list, &m->priority_groups);
1246 pg->pg_num = pg_count;
1251 if (pg_count != m->nr_priority_groups) {
1252 ti->error = "priority group count mismatch";
1257 spin_lock_irqsave(&m->lock, flags);
1258 enable_nopath_timeout(m);
1259 spin_unlock_irqrestore(&m->lock, flags);
1261 ti->num_flush_bios = 1;
1262 ti->num_discard_bios = 1;
1263 ti->num_write_zeroes_bios = 1;
1264 if (m->queue_mode == DM_TYPE_BIO_BASED)
1265 ti->per_io_data_size = multipath_per_bio_data_size();
1267 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1276 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1281 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1283 if (!atomic_read(&m->pg_init_in_progress))
1288 finish_wait(&m->pg_init_wait, &wait);
1291 static void flush_multipath_work(struct multipath *m)
1293 if (m->hw_handler_name) {
1294 unsigned long flags;
1296 if (!atomic_read(&m->pg_init_in_progress))
1299 spin_lock_irqsave(&m->lock, flags);
1300 if (atomic_read(&m->pg_init_in_progress) &&
1301 !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1302 spin_unlock_irqrestore(&m->lock, flags);
1304 flush_workqueue(kmpath_handlerd);
1305 multipath_wait_for_pg_init_completion(m);
1307 spin_lock_irqsave(&m->lock, flags);
1308 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1310 spin_unlock_irqrestore(&m->lock, flags);
1313 if (m->queue_mode == DM_TYPE_BIO_BASED)
1314 flush_work(&m->process_queued_bios);
1315 flush_work(&m->trigger_event);
1318 static void multipath_dtr(struct dm_target *ti)
1320 struct multipath *m = ti->private;
1322 disable_nopath_timeout(m);
1323 flush_multipath_work(m);
1328 * Take a path out of use.
1330 static int fail_path(struct pgpath *pgpath)
1332 unsigned long flags;
1333 struct multipath *m = pgpath->pg->m;
1335 spin_lock_irqsave(&m->lock, flags);
1337 if (!pgpath->is_active)
1340 DMWARN("%s: Failing path %s.",
1341 dm_table_device_name(m->ti->table),
1342 pgpath->path.dev->name);
1344 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1345 pgpath->is_active = false;
1346 pgpath->fail_count++;
1348 atomic_dec(&m->nr_valid_paths);
1350 if (pgpath == m->current_pgpath)
1351 m->current_pgpath = NULL;
1353 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1354 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1356 schedule_work(&m->trigger_event);
1358 enable_nopath_timeout(m);
1361 spin_unlock_irqrestore(&m->lock, flags);
1367 * Reinstate a previously-failed path
1369 static int reinstate_path(struct pgpath *pgpath)
1371 int r = 0, run_queue = 0;
1372 unsigned long flags;
1373 struct multipath *m = pgpath->pg->m;
1374 unsigned int nr_valid_paths;
1376 spin_lock_irqsave(&m->lock, flags);
1378 if (pgpath->is_active)
1381 DMWARN("%s: Reinstating path %s.",
1382 dm_table_device_name(m->ti->table),
1383 pgpath->path.dev->name);
1385 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1389 pgpath->is_active = true;
1391 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1392 if (nr_valid_paths == 1) {
1393 m->current_pgpath = NULL;
1395 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1396 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1397 atomic_inc(&m->pg_init_in_progress);
1400 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1401 pgpath->path.dev->name, nr_valid_paths);
1403 schedule_work(&m->trigger_event);
1406 spin_unlock_irqrestore(&m->lock, flags);
1408 dm_table_run_md_queue_async(m->ti->table);
1409 process_queued_io_list(m);
1412 if (pgpath->is_active)
1413 disable_nopath_timeout(m);
1419 * Fail or reinstate all paths that match the provided struct dm_dev.
1421 static int action_dev(struct multipath *m, struct dm_dev *dev,
1425 struct pgpath *pgpath;
1426 struct priority_group *pg;
1428 list_for_each_entry(pg, &m->priority_groups, list) {
1429 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1430 if (pgpath->path.dev == dev)
1439 * Temporarily try to avoid having to use the specified PG
1441 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1444 unsigned long flags;
1446 spin_lock_irqsave(&m->lock, flags);
1448 pg->bypassed = bypassed;
1449 m->current_pgpath = NULL;
1450 m->current_pg = NULL;
1452 spin_unlock_irqrestore(&m->lock, flags);
1454 schedule_work(&m->trigger_event);
1458 * Switch to using the specified PG from the next I/O that gets mapped
1460 static int switch_pg_num(struct multipath *m, const char *pgstr)
1462 struct priority_group *pg;
1464 unsigned long flags;
1467 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1468 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1469 DMWARN("invalid PG number supplied to switch_pg_num");
1473 spin_lock_irqsave(&m->lock, flags);
1474 list_for_each_entry(pg, &m->priority_groups, list) {
1475 pg->bypassed = false;
1479 m->current_pgpath = NULL;
1480 m->current_pg = NULL;
1483 spin_unlock_irqrestore(&m->lock, flags);
1485 schedule_work(&m->trigger_event);
1490 * Set/clear bypassed status of a PG.
1491 * PGs are numbered upwards from 1 in the order they were declared.
1493 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1495 struct priority_group *pg;
1499 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1500 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1501 DMWARN("invalid PG number supplied to bypass_pg");
1505 list_for_each_entry(pg, &m->priority_groups, list) {
1510 bypass_pg(m, pg, bypassed);
1515 * Should we retry pg_init immediately?
1517 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1519 unsigned long flags;
1520 bool limit_reached = false;
1522 spin_lock_irqsave(&m->lock, flags);
1524 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1525 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1526 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1528 limit_reached = true;
1530 spin_unlock_irqrestore(&m->lock, flags);
1532 return limit_reached;
1535 static void pg_init_done(void *data, int errors)
1537 struct pgpath *pgpath = data;
1538 struct priority_group *pg = pgpath->pg;
1539 struct multipath *m = pg->m;
1540 unsigned long flags;
1541 bool delay_retry = false;
1543 /* device or driver problems */
1548 if (!m->hw_handler_name) {
1552 DMERR("Could not failover the device: Handler scsi_dh_%s "
1553 "Error %d.", m->hw_handler_name, errors);
1555 * Fail path for now, so we do not ping pong
1559 case SCSI_DH_DEV_TEMP_BUSY:
1561 * Probably doing something like FW upgrade on the
1562 * controller so try the other pg.
1564 bypass_pg(m, pg, true);
1567 /* Wait before retrying. */
1570 case SCSI_DH_IMM_RETRY:
1571 case SCSI_DH_RES_TEMP_UNAVAIL:
1572 if (pg_init_limit_reached(m, pgpath))
1576 case SCSI_DH_DEV_OFFLINED:
1579 * We probably do not want to fail the path for a device
1580 * error, but this is what the old dm did. In future
1581 * patches we can do more advanced handling.
1586 spin_lock_irqsave(&m->lock, flags);
1588 if (pgpath == m->current_pgpath) {
1589 DMERR("Could not failover device. Error %d.", errors);
1590 m->current_pgpath = NULL;
1591 m->current_pg = NULL;
1593 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1594 pg->bypassed = false;
1596 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1597 /* Activations of other paths are still on going */
1600 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1602 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1604 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1606 if (__pg_init_all_paths(m))
1609 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1611 process_queued_io_list(m);
1614 * Wake up any thread waiting to suspend.
1616 wake_up(&m->pg_init_wait);
1619 spin_unlock_irqrestore(&m->lock, flags);
1622 static void activate_or_offline_path(struct pgpath *pgpath)
1624 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1626 if (pgpath->is_active && !blk_queue_dying(q))
1627 scsi_dh_activate(q, pg_init_done, pgpath);
1629 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1632 static void activate_path_work(struct work_struct *work)
1634 struct pgpath *pgpath =
1635 container_of(work, struct pgpath, activate_path.work);
1637 activate_or_offline_path(pgpath);
1640 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1641 blk_status_t error, union map_info *map_context)
1643 struct dm_mpath_io *mpio = get_mpio(map_context);
1644 struct pgpath *pgpath = mpio->pgpath;
1645 int r = DM_ENDIO_DONE;
1648 * We don't queue any clone request inside the multipath target
1649 * during end I/O handling, since those clone requests don't have
1650 * bio clones. If we queue them inside the multipath target,
1651 * we need to make bio clones, that requires memory allocation.
1652 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1653 * don't have bio clones.)
1654 * Instead of queueing the clone request here, we queue the original
1655 * request into dm core, which will remake a clone request and
1656 * clone bios for it and resubmit it later.
1658 if (error && blk_path_error(error)) {
1659 struct multipath *m = ti->private;
1661 if (error == BLK_STS_RESOURCE)
1662 r = DM_ENDIO_DELAY_REQUEUE;
1664 r = DM_ENDIO_REQUEUE;
1669 if (!atomic_read(&m->nr_valid_paths) &&
1670 !must_push_back_rq(m)) {
1671 if (error == BLK_STS_IOERR)
1673 /* complete with the original error */
1679 struct path_selector *ps = &pgpath->pg->ps;
1681 if (ps->type->end_io)
1682 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1683 clone->io_start_time_ns);
1689 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1690 blk_status_t *error)
1692 struct multipath *m = ti->private;
1693 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1694 struct pgpath *pgpath = mpio->pgpath;
1695 unsigned long flags;
1696 int r = DM_ENDIO_DONE;
1698 if (!*error || !blk_path_error(*error))
1704 if (!atomic_read(&m->nr_valid_paths)) {
1705 spin_lock_irqsave(&m->lock, flags);
1706 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1707 if (__must_push_back(m)) {
1708 r = DM_ENDIO_REQUEUE;
1711 *error = BLK_STS_IOERR;
1713 spin_unlock_irqrestore(&m->lock, flags);
1716 spin_unlock_irqrestore(&m->lock, flags);
1719 multipath_queue_bio(m, clone);
1720 r = DM_ENDIO_INCOMPLETE;
1723 struct path_selector *ps = &pgpath->pg->ps;
1725 if (ps->type->end_io)
1726 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1727 (mpio->start_time_ns ?:
1728 dm_start_time_ns_from_clone(clone)));
1735 * Suspend with flush can't complete until all the I/O is processed
1736 * so if the last path fails we must error any remaining I/O.
1737 * - Note that if the freeze_bdev fails while suspending, the
1738 * queue_if_no_path state is lost - userspace should reset it.
1739 * Otherwise, during noflush suspend, queue_if_no_path will not change.
1741 static void multipath_presuspend(struct dm_target *ti)
1743 struct multipath *m = ti->private;
1745 /* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1746 if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1747 queue_if_no_path(m, false, true, __func__);
1750 static void multipath_postsuspend(struct dm_target *ti)
1752 struct multipath *m = ti->private;
1754 mutex_lock(&m->work_mutex);
1755 flush_multipath_work(m);
1756 mutex_unlock(&m->work_mutex);
1760 * Restore the queue_if_no_path setting.
1762 static void multipath_resume(struct dm_target *ti)
1764 struct multipath *m = ti->private;
1765 unsigned long flags;
1767 spin_lock_irqsave(&m->lock, flags);
1768 if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1769 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1770 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1773 DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1774 dm_table_device_name(m->ti->table), __func__,
1775 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1776 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1778 spin_unlock_irqrestore(&m->lock, flags);
1782 * Info output has the following format:
1783 * num_multipath_feature_args [multipath_feature_args]*
1784 * num_handler_status_args [handler_status_args]*
1785 * num_groups init_group_number
1786 * [A|D|E num_ps_status_args [ps_status_args]*
1787 * num_paths num_selector_args
1788 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1790 * Table output has the following format (identical to the constructor string):
1791 * num_feature_args [features_args]*
1792 * num_handler_args hw_handler [hw_handler_args]*
1793 * num_groups init_group_number
1794 * [priority selector-name num_ps_args [ps_args]*
1795 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1797 static void multipath_status(struct dm_target *ti, status_type_t type,
1798 unsigned int status_flags, char *result, unsigned int maxlen)
1800 int sz = 0, pg_counter, pgpath_counter;
1801 unsigned long flags;
1802 struct multipath *m = ti->private;
1803 struct priority_group *pg;
1805 unsigned int pg_num;
1808 spin_lock_irqsave(&m->lock, flags);
1811 if (type == STATUSTYPE_INFO)
1812 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1813 atomic_read(&m->pg_init_count));
1815 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1816 (m->pg_init_retries > 0) * 2 +
1817 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1818 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1819 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1821 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1822 DMEMIT("queue_if_no_path ");
1823 if (m->pg_init_retries)
1824 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1825 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1826 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1827 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1828 DMEMIT("retain_attached_hw_handler ");
1829 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1830 switch (m->queue_mode) {
1831 case DM_TYPE_BIO_BASED:
1832 DMEMIT("queue_mode bio ");
1841 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1844 DMEMIT("1 %s ", m->hw_handler_name);
1846 DMEMIT("%u ", m->nr_priority_groups);
1849 pg_num = m->next_pg->pg_num;
1850 else if (m->current_pg)
1851 pg_num = m->current_pg->pg_num;
1853 pg_num = (m->nr_priority_groups ? 1 : 0);
1855 DMEMIT("%u ", pg_num);
1858 case STATUSTYPE_INFO:
1859 list_for_each_entry(pg, &m->priority_groups, list) {
1861 state = 'D'; /* Disabled */
1862 else if (pg == m->current_pg)
1863 state = 'A'; /* Currently Active */
1865 state = 'E'; /* Enabled */
1867 DMEMIT("%c ", state);
1869 if (pg->ps.type->status)
1870 sz += pg->ps.type->status(&pg->ps, NULL, type,
1876 DMEMIT("%u %u ", pg->nr_pgpaths,
1877 pg->ps.type->info_args);
1879 list_for_each_entry(p, &pg->pgpaths, list) {
1880 DMEMIT("%s %s %u ", p->path.dev->name,
1881 p->is_active ? "A" : "F",
1883 if (pg->ps.type->status)
1884 sz += pg->ps.type->status(&pg->ps,
1885 &p->path, type, result + sz,
1891 case STATUSTYPE_TABLE:
1892 list_for_each_entry(pg, &m->priority_groups, list) {
1893 DMEMIT("%s ", pg->ps.type->name);
1895 if (pg->ps.type->status)
1896 sz += pg->ps.type->status(&pg->ps, NULL, type,
1902 DMEMIT("%u %u ", pg->nr_pgpaths,
1903 pg->ps.type->table_args);
1905 list_for_each_entry(p, &pg->pgpaths, list) {
1906 DMEMIT("%s ", p->path.dev->name);
1907 if (pg->ps.type->status)
1908 sz += pg->ps.type->status(&pg->ps,
1909 &p->path, type, result + sz,
1915 case STATUSTYPE_IMA:
1916 sz = 0; /*reset the result pointer*/
1918 DMEMIT_TARGET_NAME_VERSION(ti->type);
1919 DMEMIT(",nr_priority_groups=%u", m->nr_priority_groups);
1922 list_for_each_entry(pg, &m->priority_groups, list) {
1924 state = 'D'; /* Disabled */
1925 else if (pg == m->current_pg)
1926 state = 'A'; /* Currently Active */
1928 state = 'E'; /* Enabled */
1929 DMEMIT(",pg_state_%d=%c", pg_counter, state);
1930 DMEMIT(",nr_pgpaths_%d=%u", pg_counter, pg->nr_pgpaths);
1931 DMEMIT(",path_selector_name_%d=%s", pg_counter, pg->ps.type->name);
1934 list_for_each_entry(p, &pg->pgpaths, list) {
1935 DMEMIT(",path_name_%d_%d=%s,is_active_%d_%d=%c,fail_count_%d_%d=%u",
1936 pg_counter, pgpath_counter, p->path.dev->name,
1937 pg_counter, pgpath_counter, p->is_active ? 'A' : 'F',
1938 pg_counter, pgpath_counter, p->fail_count);
1939 if (pg->ps.type->status) {
1940 DMEMIT(",path_selector_status_%d_%d=",
1941 pg_counter, pgpath_counter);
1942 sz += pg->ps.type->status(&pg->ps, &p->path,
1954 spin_unlock_irqrestore(&m->lock, flags);
1957 static int multipath_message(struct dm_target *ti, unsigned int argc, char **argv,
1958 char *result, unsigned int maxlen)
1962 struct multipath *m = ti->private;
1964 unsigned long flags;
1966 mutex_lock(&m->work_mutex);
1968 if (dm_suspended(ti)) {
1974 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1975 r = queue_if_no_path(m, true, false, __func__);
1976 spin_lock_irqsave(&m->lock, flags);
1977 enable_nopath_timeout(m);
1978 spin_unlock_irqrestore(&m->lock, flags);
1980 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1981 r = queue_if_no_path(m, false, false, __func__);
1982 disable_nopath_timeout(m);
1988 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1992 if (!strcasecmp(argv[0], "disable_group")) {
1993 r = bypass_pg_num(m, argv[1], true);
1995 } else if (!strcasecmp(argv[0], "enable_group")) {
1996 r = bypass_pg_num(m, argv[1], false);
1998 } else if (!strcasecmp(argv[0], "switch_group")) {
1999 r = switch_pg_num(m, argv[1]);
2001 } else if (!strcasecmp(argv[0], "reinstate_path"))
2002 action = reinstate_path;
2003 else if (!strcasecmp(argv[0], "fail_path"))
2006 DMWARN("Unrecognised multipath message received: %s", argv[0]);
2010 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
2012 DMWARN("message: error getting device %s",
2017 r = action_dev(m, dev, action);
2019 dm_put_device(ti, dev);
2022 mutex_unlock(&m->work_mutex);
2026 static int multipath_prepare_ioctl(struct dm_target *ti,
2027 struct block_device **bdev)
2029 struct multipath *m = ti->private;
2030 struct pgpath *pgpath;
2031 unsigned long flags;
2034 pgpath = READ_ONCE(m->current_pgpath);
2035 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
2036 pgpath = choose_pgpath(m, 0);
2039 if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
2040 *bdev = pgpath->path.dev->bdev;
2043 /* pg_init has not started or completed */
2047 /* No path is available */
2049 spin_lock_irqsave(&m->lock, flags);
2050 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2052 spin_unlock_irqrestore(&m->lock, flags);
2055 if (r == -ENOTCONN) {
2056 if (!READ_ONCE(m->current_pg)) {
2057 /* Path status changed, redo selection */
2058 (void) choose_pgpath(m, 0);
2060 spin_lock_irqsave(&m->lock, flags);
2061 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2062 (void) __pg_init_all_paths(m);
2063 spin_unlock_irqrestore(&m->lock, flags);
2064 dm_table_run_md_queue_async(m->ti->table);
2065 process_queued_io_list(m);
2069 * Only pass ioctls through if the device sizes match exactly.
2071 if (!r && ti->len != bdev_nr_sectors((*bdev)))
2076 static int multipath_iterate_devices(struct dm_target *ti,
2077 iterate_devices_callout_fn fn, void *data)
2079 struct multipath *m = ti->private;
2080 struct priority_group *pg;
2084 list_for_each_entry(pg, &m->priority_groups, list) {
2085 list_for_each_entry(p, &pg->pgpaths, list) {
2086 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2096 static int pgpath_busy(struct pgpath *pgpath)
2098 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2100 return blk_lld_busy(q);
2104 * We return "busy", only when we can map I/Os but underlying devices
2105 * are busy (so even if we map I/Os now, the I/Os will wait on
2106 * the underlying queue).
2107 * In other words, if we want to kill I/Os or queue them inside us
2108 * due to map unavailability, we don't return "busy". Otherwise,
2109 * dm core won't give us the I/Os and we can't do what we want.
2111 static int multipath_busy(struct dm_target *ti)
2113 bool busy = false, has_active = false;
2114 struct multipath *m = ti->private;
2115 struct priority_group *pg, *next_pg;
2116 struct pgpath *pgpath;
2118 /* pg_init in progress */
2119 if (atomic_read(&m->pg_init_in_progress))
2122 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2123 if (!atomic_read(&m->nr_valid_paths)) {
2124 unsigned long flags;
2125 spin_lock_irqsave(&m->lock, flags);
2126 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2127 spin_unlock_irqrestore(&m->lock, flags);
2128 return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2130 spin_unlock_irqrestore(&m->lock, flags);
2133 /* Guess which priority_group will be used at next mapping time */
2134 pg = READ_ONCE(m->current_pg);
2135 next_pg = READ_ONCE(m->next_pg);
2136 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2141 * We don't know which pg will be used at next mapping time.
2142 * We don't call choose_pgpath() here to avoid to trigger
2143 * pg_init just by busy checking.
2144 * So we don't know whether underlying devices we will be using
2145 * at next mapping time are busy or not. Just try mapping.
2151 * If there is one non-busy active path at least, the path selector
2152 * will be able to select it. So we consider such a pg as not busy.
2155 list_for_each_entry(pgpath, &pg->pgpaths, list) {
2156 if (pgpath->is_active) {
2158 if (!pgpath_busy(pgpath)) {
2167 * No active path in this pg, so this pg won't be used and
2168 * the current_pg will be changed at next mapping time.
2169 * We need to try mapping to determine it.
2178 *---------------------------------------------------------------
2180 *---------------------------------------------------------------
2182 static struct target_type multipath_target = {
2183 .name = "multipath",
2184 .version = {1, 14, 0},
2185 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2186 DM_TARGET_PASSES_INTEGRITY,
2187 .module = THIS_MODULE,
2188 .ctr = multipath_ctr,
2189 .dtr = multipath_dtr,
2190 .clone_and_map_rq = multipath_clone_and_map,
2191 .release_clone_rq = multipath_release_clone,
2192 .rq_end_io = multipath_end_io,
2193 .map = multipath_map_bio,
2194 .end_io = multipath_end_io_bio,
2195 .presuspend = multipath_presuspend,
2196 .postsuspend = multipath_postsuspend,
2197 .resume = multipath_resume,
2198 .status = multipath_status,
2199 .message = multipath_message,
2200 .prepare_ioctl = multipath_prepare_ioctl,
2201 .iterate_devices = multipath_iterate_devices,
2202 .busy = multipath_busy,
2205 static int __init dm_multipath_init(void)
2209 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2211 DMERR("failed to create workqueue kmpathd");
2213 goto bad_alloc_kmultipathd;
2217 * A separate workqueue is used to handle the device handlers
2218 * to avoid overloading existing workqueue. Overloading the
2219 * old workqueue would also create a bottleneck in the
2220 * path of the storage hardware device activation.
2222 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2224 if (!kmpath_handlerd) {
2225 DMERR("failed to create workqueue kmpath_handlerd");
2227 goto bad_alloc_kmpath_handlerd;
2230 r = dm_register_target(&multipath_target);
2232 DMERR("request-based register failed %d", r);
2234 goto bad_register_target;
2239 bad_register_target:
2240 destroy_workqueue(kmpath_handlerd);
2241 bad_alloc_kmpath_handlerd:
2242 destroy_workqueue(kmultipathd);
2243 bad_alloc_kmultipathd:
2247 static void __exit dm_multipath_exit(void)
2249 destroy_workqueue(kmpath_handlerd);
2250 destroy_workqueue(kmultipathd);
2252 dm_unregister_target(&multipath_target);
2255 module_init(dm_multipath_init);
2256 module_exit(dm_multipath_exit);
2258 module_param_named(queue_if_no_path_timeout_secs,
2259 queue_if_no_path_timeout_secs, ulong, S_IRUGO | S_IWUSR);
2260 MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2262 MODULE_DESCRIPTION(DM_NAME " multipath target");
2263 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2264 MODULE_LICENSE("GPL");