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) -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 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 pg_num; /* Reference number */
64 unsigned 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 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 pg_init_retries; /* Number of times to retry pg_init */
89 unsigned 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);
121 /*-----------------------------------------------
122 * Multipath state flags.
123 *-----------------------------------------------*/
125 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
126 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
127 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
128 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
129 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
130 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
131 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
133 static bool mpath_double_check_test_bit(int MPATHF_bit, struct multipath *m)
135 bool r = test_bit(MPATHF_bit, &m->flags);
139 spin_lock_irqsave(&m->lock, flags);
140 r = test_bit(MPATHF_bit, &m->flags);
141 spin_unlock_irqrestore(&m->lock, flags);
147 /*-----------------------------------------------
148 * Allocation routines
149 *-----------------------------------------------*/
151 static struct pgpath *alloc_pgpath(void)
153 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
158 pgpath->is_active = true;
163 static void free_pgpath(struct pgpath *pgpath)
168 static struct priority_group *alloc_priority_group(void)
170 struct priority_group *pg;
172 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
175 INIT_LIST_HEAD(&pg->pgpaths);
180 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
182 struct pgpath *pgpath, *tmp;
184 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
185 list_del(&pgpath->list);
186 dm_put_device(ti, pgpath->path.dev);
191 static void free_priority_group(struct priority_group *pg,
192 struct dm_target *ti)
194 struct path_selector *ps = &pg->ps;
197 ps->type->destroy(ps);
198 dm_put_path_selector(ps->type);
201 free_pgpaths(&pg->pgpaths, ti);
205 static struct multipath *alloc_multipath(struct dm_target *ti)
209 m = kzalloc(sizeof(*m), GFP_KERNEL);
211 INIT_LIST_HEAD(&m->priority_groups);
212 spin_lock_init(&m->lock);
213 atomic_set(&m->nr_valid_paths, 0);
214 INIT_WORK(&m->trigger_event, trigger_event);
215 mutex_init(&m->work_mutex);
217 m->queue_mode = DM_TYPE_NONE;
222 timer_setup(&m->nopath_timer, queue_if_no_path_timeout_work, 0);
228 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
230 if (m->queue_mode == DM_TYPE_NONE) {
231 m->queue_mode = DM_TYPE_REQUEST_BASED;
232 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
233 INIT_WORK(&m->process_queued_bios, process_queued_bios);
235 * bio-based doesn't support any direct scsi_dh management;
236 * it just discovers if a scsi_dh is attached.
238 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
241 dm_table_set_type(ti->table, m->queue_mode);
244 * Init fields that are only used when a scsi_dh is attached
245 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
247 set_bit(MPATHF_QUEUE_IO, &m->flags);
248 atomic_set(&m->pg_init_in_progress, 0);
249 atomic_set(&m->pg_init_count, 0);
250 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
251 init_waitqueue_head(&m->pg_init_wait);
256 static void free_multipath(struct multipath *m)
258 struct priority_group *pg, *tmp;
260 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
262 free_priority_group(pg, m->ti);
265 kfree(m->hw_handler_name);
266 kfree(m->hw_handler_params);
267 mutex_destroy(&m->work_mutex);
271 static struct dm_mpath_io *get_mpio(union map_info *info)
276 static size_t multipath_per_bio_data_size(void)
278 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
281 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
283 return dm_per_bio_data(bio, multipath_per_bio_data_size());
286 static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
288 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
289 void *bio_details = mpio + 1;
293 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
295 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
296 struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
298 mpio->nr_bytes = bio->bi_iter.bi_size;
300 mpio->start_time_ns = 0;
303 dm_bio_record(bio_details, bio);
306 /*-----------------------------------------------
308 *-----------------------------------------------*/
310 static int __pg_init_all_paths(struct multipath *m)
312 struct pgpath *pgpath;
313 unsigned long pg_init_delay = 0;
315 lockdep_assert_held(&m->lock);
317 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
320 atomic_inc(&m->pg_init_count);
321 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
323 /* Check here to reset pg_init_required */
327 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
328 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
329 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
330 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
331 /* Skip failed paths */
332 if (!pgpath->is_active)
334 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
336 atomic_inc(&m->pg_init_in_progress);
338 return atomic_read(&m->pg_init_in_progress);
341 static int pg_init_all_paths(struct multipath *m)
346 spin_lock_irqsave(&m->lock, flags);
347 ret = __pg_init_all_paths(m);
348 spin_unlock_irqrestore(&m->lock, flags);
353 static void __switch_pg(struct multipath *m, struct priority_group *pg)
355 lockdep_assert_held(&m->lock);
359 /* Must we initialise the PG first, and queue I/O till it's ready? */
360 if (m->hw_handler_name) {
361 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
362 set_bit(MPATHF_QUEUE_IO, &m->flags);
364 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
365 clear_bit(MPATHF_QUEUE_IO, &m->flags);
368 atomic_set(&m->pg_init_count, 0);
371 static struct pgpath *choose_path_in_pg(struct multipath *m,
372 struct priority_group *pg,
376 struct dm_path *path;
377 struct pgpath *pgpath;
379 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
381 return ERR_PTR(-ENXIO);
383 pgpath = path_to_pgpath(path);
385 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
386 /* Only update current_pgpath if pg changed */
387 spin_lock_irqsave(&m->lock, flags);
388 m->current_pgpath = pgpath;
390 spin_unlock_irqrestore(&m->lock, flags);
396 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
399 struct priority_group *pg;
400 struct pgpath *pgpath;
401 unsigned bypassed = 1;
403 if (!atomic_read(&m->nr_valid_paths)) {
404 spin_lock_irqsave(&m->lock, flags);
405 clear_bit(MPATHF_QUEUE_IO, &m->flags);
406 spin_unlock_irqrestore(&m->lock, flags);
410 /* Were we instructed to switch PG? */
411 if (READ_ONCE(m->next_pg)) {
412 spin_lock_irqsave(&m->lock, flags);
415 spin_unlock_irqrestore(&m->lock, flags);
416 goto check_current_pg;
419 spin_unlock_irqrestore(&m->lock, flags);
420 pgpath = choose_path_in_pg(m, pg, nr_bytes);
421 if (!IS_ERR_OR_NULL(pgpath))
425 /* Don't change PG until it has no remaining paths */
427 pg = READ_ONCE(m->current_pg);
429 pgpath = choose_path_in_pg(m, pg, nr_bytes);
430 if (!IS_ERR_OR_NULL(pgpath))
435 * Loop through priority groups until we find a valid path.
436 * First time we skip PGs marked 'bypassed'.
437 * Second time we only try the ones we skipped, but set
438 * pg_init_delay_retry so we do not hammer controllers.
441 list_for_each_entry(pg, &m->priority_groups, list) {
442 if (pg->bypassed == !!bypassed)
444 pgpath = choose_path_in_pg(m, pg, nr_bytes);
445 if (!IS_ERR_OR_NULL(pgpath)) {
447 spin_lock_irqsave(&m->lock, flags);
448 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
449 spin_unlock_irqrestore(&m->lock, flags);
454 } while (bypassed--);
457 spin_lock_irqsave(&m->lock, flags);
458 m->current_pgpath = NULL;
459 m->current_pg = NULL;
460 spin_unlock_irqrestore(&m->lock, flags);
466 * dm_report_EIO() is a macro instead of a function to make pr_debug_ratelimited()
467 * report the function name and line number of the function from which
468 * it has been invoked.
470 #define dm_report_EIO(m) \
472 DMDEBUG_LIMIT("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d", \
473 dm_table_device_name((m)->ti->table), \
474 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
475 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
476 dm_noflush_suspending((m)->ti)); \
480 * Check whether bios must be queued in the device-mapper core rather
481 * than here in the target.
483 static bool __must_push_back(struct multipath *m)
485 return dm_noflush_suspending(m->ti);
488 static bool must_push_back_rq(struct multipath *m)
493 spin_lock_irqsave(&m->lock, flags);
494 ret = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) || __must_push_back(m));
495 spin_unlock_irqrestore(&m->lock, flags);
501 * Map cloned requests (request-based multipath)
503 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
504 union map_info *map_context,
505 struct request **__clone)
507 struct multipath *m = ti->private;
508 size_t nr_bytes = blk_rq_bytes(rq);
509 struct pgpath *pgpath;
510 struct block_device *bdev;
511 struct dm_mpath_io *mpio = get_mpio(map_context);
512 struct request_queue *q;
513 struct request *clone;
515 /* Do we need to select a new pgpath? */
516 pgpath = READ_ONCE(m->current_pgpath);
517 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
518 pgpath = choose_pgpath(m, nr_bytes);
521 if (must_push_back_rq(m))
522 return DM_MAPIO_DELAY_REQUEUE;
523 dm_report_EIO(m); /* Failed */
524 return DM_MAPIO_KILL;
525 } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
526 mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
527 pg_init_all_paths(m);
528 return DM_MAPIO_DELAY_REQUEUE;
531 mpio->pgpath = pgpath;
532 mpio->nr_bytes = nr_bytes;
534 bdev = pgpath->path.dev->bdev;
535 q = bdev_get_queue(bdev);
536 clone = blk_mq_alloc_request(q, rq->cmd_flags | REQ_NOMERGE,
539 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
540 if (blk_queue_dying(q)) {
541 atomic_inc(&m->pg_init_in_progress);
542 activate_or_offline_path(pgpath);
543 return DM_MAPIO_DELAY_REQUEUE;
547 * blk-mq's SCHED_RESTART can cover this requeue, so we
548 * needn't deal with it by DELAY_REQUEUE. More importantly,
549 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
550 * get the queue busy feedback (via BLK_STS_RESOURCE),
551 * otherwise I/O merging can suffer.
553 return DM_MAPIO_REQUEUE;
555 clone->bio = clone->biotail = NULL;
556 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
559 if (pgpath->pg->ps.type->start_io)
560 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
563 return DM_MAPIO_REMAPPED;
566 static void multipath_release_clone(struct request *clone,
567 union map_info *map_context)
569 if (unlikely(map_context)) {
571 * non-NULL map_context means caller is still map
572 * method; must undo multipath_clone_and_map()
574 struct dm_mpath_io *mpio = get_mpio(map_context);
575 struct pgpath *pgpath = mpio->pgpath;
577 if (pgpath && pgpath->pg->ps.type->end_io)
578 pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
581 clone->io_start_time_ns);
584 blk_mq_free_request(clone);
588 * Map cloned bios (bio-based multipath)
591 static void __multipath_queue_bio(struct multipath *m, struct bio *bio)
593 /* Queue for the daemon to resubmit */
594 bio_list_add(&m->queued_bios, bio);
595 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
596 queue_work(kmultipathd, &m->process_queued_bios);
599 static void multipath_queue_bio(struct multipath *m, struct bio *bio)
603 spin_lock_irqsave(&m->lock, flags);
604 __multipath_queue_bio(m, bio);
605 spin_unlock_irqrestore(&m->lock, flags);
608 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
610 struct pgpath *pgpath;
613 /* Do we need to select a new pgpath? */
614 pgpath = READ_ONCE(m->current_pgpath);
615 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
616 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
619 spin_lock_irqsave(&m->lock, flags);
620 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
621 __multipath_queue_bio(m, bio);
622 pgpath = ERR_PTR(-EAGAIN);
624 spin_unlock_irqrestore(&m->lock, flags);
626 } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
627 mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
628 multipath_queue_bio(m, bio);
629 pg_init_all_paths(m);
630 return ERR_PTR(-EAGAIN);
636 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
637 struct dm_mpath_io *mpio)
639 struct pgpath *pgpath = __map_bio(m, bio);
642 return DM_MAPIO_SUBMITTED;
645 if (__must_push_back(m))
646 return DM_MAPIO_REQUEUE;
648 return DM_MAPIO_KILL;
651 mpio->pgpath = pgpath;
653 if (dm_ps_use_hr_timer(pgpath->pg->ps.type))
654 mpio->start_time_ns = ktime_get_ns();
657 bio_set_dev(bio, pgpath->path.dev->bdev);
658 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
660 if (pgpath->pg->ps.type->start_io)
661 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
664 return DM_MAPIO_REMAPPED;
667 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
669 struct multipath *m = ti->private;
670 struct dm_mpath_io *mpio = NULL;
672 multipath_init_per_bio_data(bio, &mpio);
673 return __multipath_map_bio(m, bio, mpio);
676 static void process_queued_io_list(struct multipath *m)
678 if (m->queue_mode == DM_TYPE_REQUEST_BASED)
679 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
680 else if (m->queue_mode == DM_TYPE_BIO_BASED)
681 queue_work(kmultipathd, &m->process_queued_bios);
684 static void process_queued_bios(struct work_struct *work)
689 struct bio_list bios;
690 struct blk_plug plug;
691 struct multipath *m =
692 container_of(work, struct multipath, process_queued_bios);
694 bio_list_init(&bios);
696 spin_lock_irqsave(&m->lock, flags);
698 if (bio_list_empty(&m->queued_bios)) {
699 spin_unlock_irqrestore(&m->lock, flags);
703 bio_list_merge(&bios, &m->queued_bios);
704 bio_list_init(&m->queued_bios);
706 spin_unlock_irqrestore(&m->lock, flags);
708 blk_start_plug(&plug);
709 while ((bio = bio_list_pop(&bios))) {
710 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
711 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
712 r = __multipath_map_bio(m, bio, mpio);
715 bio->bi_status = BLK_STS_IOERR;
718 case DM_MAPIO_REQUEUE:
719 bio->bi_status = BLK_STS_DM_REQUEUE;
722 case DM_MAPIO_REMAPPED:
723 submit_bio_noacct(bio);
725 case DM_MAPIO_SUBMITTED:
728 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
731 blk_finish_plug(&plug);
735 * If we run out of usable paths, should we queue I/O or error it?
737 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
738 bool save_old_value, const char *caller)
741 bool queue_if_no_path_bit, saved_queue_if_no_path_bit;
742 const char *dm_dev_name = dm_table_device_name(m->ti->table);
744 DMDEBUG("%s: %s caller=%s queue_if_no_path=%d save_old_value=%d",
745 dm_dev_name, __func__, caller, queue_if_no_path, save_old_value);
747 spin_lock_irqsave(&m->lock, flags);
749 queue_if_no_path_bit = test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
750 saved_queue_if_no_path_bit = test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
752 if (save_old_value) {
753 if (unlikely(!queue_if_no_path_bit && saved_queue_if_no_path_bit)) {
754 DMERR("%s: QIFNP disabled but saved as enabled, saving again loses state, not saving!",
757 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path_bit);
758 } else if (!queue_if_no_path && saved_queue_if_no_path_bit) {
759 /* due to "fail_if_no_path" message, need to honor it. */
760 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
762 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
764 DMDEBUG("%s: after %s changes; QIFNP = %d; SQIFNP = %d; DNFS = %d",
765 dm_dev_name, __func__,
766 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
767 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
768 dm_noflush_suspending(m->ti));
770 spin_unlock_irqrestore(&m->lock, flags);
772 if (!queue_if_no_path) {
773 dm_table_run_md_queue_async(m->ti->table);
774 process_queued_io_list(m);
781 * If the queue_if_no_path timeout fires, turn off queue_if_no_path and
782 * process any queued I/O.
784 static void queue_if_no_path_timeout_work(struct timer_list *t)
786 struct multipath *m = from_timer(m, t, nopath_timer);
788 DMWARN("queue_if_no_path timeout on %s, failing queued IO",
789 dm_table_device_name(m->ti->table));
790 queue_if_no_path(m, false, false, __func__);
794 * Enable the queue_if_no_path timeout if necessary.
795 * Called with m->lock held.
797 static void enable_nopath_timeout(struct multipath *m)
799 unsigned long queue_if_no_path_timeout =
800 READ_ONCE(queue_if_no_path_timeout_secs) * HZ;
802 lockdep_assert_held(&m->lock);
804 if (queue_if_no_path_timeout > 0 &&
805 atomic_read(&m->nr_valid_paths) == 0 &&
806 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
807 mod_timer(&m->nopath_timer,
808 jiffies + queue_if_no_path_timeout);
812 static void disable_nopath_timeout(struct multipath *m)
814 del_timer_sync(&m->nopath_timer);
818 * An event is triggered whenever a path is taken out of use.
819 * Includes path failure and PG bypass.
821 static void trigger_event(struct work_struct *work)
823 struct multipath *m =
824 container_of(work, struct multipath, trigger_event);
826 dm_table_event(m->ti->table);
829 /*-----------------------------------------------------------------
830 * Constructor/argument parsing:
831 * <#multipath feature args> [<arg>]*
832 * <#hw_handler args> [hw_handler [<arg>]*]
834 * <initial priority group>
835 * [<selector> <#selector args> [<arg>]*
836 * <#paths> <#per-path selector args>
837 * [<path> [<arg>]* ]+ ]+
838 *---------------------------------------------------------------*/
839 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
840 struct dm_target *ti)
843 struct path_selector_type *pst;
846 static const struct dm_arg _args[] = {
847 {0, 1024, "invalid number of path selector args"},
850 pst = dm_get_path_selector(dm_shift_arg(as));
852 ti->error = "unknown path selector type";
856 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
858 dm_put_path_selector(pst);
862 r = pst->create(&pg->ps, ps_argc, as->argv);
864 dm_put_path_selector(pst);
865 ti->error = "path selector constructor failed";
870 dm_consume_args(as, ps_argc);
875 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
876 const char **attached_handler_name, char **error)
878 struct request_queue *q = bdev_get_queue(bdev);
881 if (mpath_double_check_test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, m)) {
883 if (*attached_handler_name) {
885 * Clear any hw_handler_params associated with a
886 * handler that isn't already attached.
888 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
889 kfree(m->hw_handler_params);
890 m->hw_handler_params = NULL;
894 * Reset hw_handler_name to match the attached handler
896 * NB. This modifies the table line to show the actual
897 * handler instead of the original table passed in.
899 kfree(m->hw_handler_name);
900 m->hw_handler_name = *attached_handler_name;
901 *attached_handler_name = NULL;
905 if (m->hw_handler_name) {
906 r = scsi_dh_attach(q, m->hw_handler_name);
908 DMINFO("retaining handler on device %pg", bdev);
912 *error = "error attaching hardware handler";
916 if (m->hw_handler_params) {
917 r = scsi_dh_set_params(q, m->hw_handler_params);
919 *error = "unable to set hardware handler parameters";
928 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
929 struct dm_target *ti)
933 struct multipath *m = ti->private;
934 struct request_queue *q;
935 const char *attached_handler_name = NULL;
937 /* we need at least a path arg */
939 ti->error = "no device given";
940 return ERR_PTR(-EINVAL);
945 return ERR_PTR(-ENOMEM);
947 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
950 ti->error = "error getting device";
954 q = bdev_get_queue(p->path.dev->bdev);
955 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
956 if (attached_handler_name || m->hw_handler_name) {
957 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
958 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
959 kfree(attached_handler_name);
961 dm_put_device(ti, p->path.dev);
966 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
968 dm_put_device(ti, p->path.dev);
978 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
981 static const struct dm_arg _args[] = {
982 {1, 1024, "invalid number of paths"},
983 {0, 1024, "invalid number of selector args"}
987 unsigned i, nr_selector_args, nr_args;
988 struct priority_group *pg;
989 struct dm_target *ti = m->ti;
993 ti->error = "not enough priority group arguments";
994 return ERR_PTR(-EINVAL);
997 pg = alloc_priority_group();
999 ti->error = "couldn't allocate priority group";
1000 return ERR_PTR(-ENOMEM);
1004 r = parse_path_selector(as, pg, ti);
1011 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1015 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1019 nr_args = 1 + nr_selector_args;
1020 for (i = 0; i < pg->nr_pgpaths; i++) {
1021 struct pgpath *pgpath;
1022 struct dm_arg_set path_args;
1024 if (as->argc < nr_args) {
1025 ti->error = "not enough path parameters";
1030 path_args.argc = nr_args;
1031 path_args.argv = as->argv;
1033 pgpath = parse_path(&path_args, &pg->ps, ti);
1034 if (IS_ERR(pgpath)) {
1035 r = PTR_ERR(pgpath);
1040 list_add_tail(&pgpath->list, &pg->pgpaths);
1041 dm_consume_args(as, nr_args);
1047 free_priority_group(pg, ti);
1051 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1055 struct dm_target *ti = m->ti;
1057 static const struct dm_arg _args[] = {
1058 {0, 1024, "invalid number of hardware handler args"},
1061 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1067 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1068 dm_consume_args(as, hw_argc);
1069 DMERR("bio-based multipath doesn't allow hardware handler args");
1073 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1074 if (!m->hw_handler_name)
1081 for (i = 0; i <= hw_argc - 2; i++)
1082 len += strlen(as->argv[i]) + 1;
1083 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1085 ti->error = "memory allocation failed";
1089 j = sprintf(p, "%d", hw_argc - 1);
1090 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1091 j = sprintf(p, "%s", as->argv[i]);
1093 dm_consume_args(as, hw_argc - 1);
1097 kfree(m->hw_handler_name);
1098 m->hw_handler_name = NULL;
1102 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1106 struct dm_target *ti = m->ti;
1107 const char *arg_name;
1109 static const struct dm_arg _args[] = {
1110 {0, 8, "invalid number of feature args"},
1111 {1, 50, "pg_init_retries must be between 1 and 50"},
1112 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1115 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1123 arg_name = dm_shift_arg(as);
1126 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1127 r = queue_if_no_path(m, true, false, __func__);
1131 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1132 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1136 if (!strcasecmp(arg_name, "pg_init_retries") &&
1138 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1143 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1145 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1150 if (!strcasecmp(arg_name, "queue_mode") &&
1152 const char *queue_mode_name = dm_shift_arg(as);
1154 if (!strcasecmp(queue_mode_name, "bio"))
1155 m->queue_mode = DM_TYPE_BIO_BASED;
1156 else if (!strcasecmp(queue_mode_name, "rq") ||
1157 !strcasecmp(queue_mode_name, "mq"))
1158 m->queue_mode = DM_TYPE_REQUEST_BASED;
1160 ti->error = "Unknown 'queue_mode' requested";
1167 ti->error = "Unrecognised multipath feature request";
1169 } while (argc && !r);
1174 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1176 /* target arguments */
1177 static const struct dm_arg _args[] = {
1178 {0, 1024, "invalid number of priority groups"},
1179 {0, 1024, "invalid initial priority group number"},
1183 struct multipath *m;
1184 struct dm_arg_set as;
1185 unsigned pg_count = 0;
1186 unsigned next_pg_num;
1187 unsigned long flags;
1192 m = alloc_multipath(ti);
1194 ti->error = "can't allocate multipath";
1198 r = parse_features(&as, m);
1202 r = alloc_multipath_stage2(ti, m);
1206 r = parse_hw_handler(&as, m);
1210 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1214 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1218 if ((!m->nr_priority_groups && next_pg_num) ||
1219 (m->nr_priority_groups && !next_pg_num)) {
1220 ti->error = "invalid initial priority group";
1225 /* parse the priority groups */
1227 struct priority_group *pg;
1228 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1230 pg = parse_priority_group(&as, m);
1236 nr_valid_paths += pg->nr_pgpaths;
1237 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1239 list_add_tail(&pg->list, &m->priority_groups);
1241 pg->pg_num = pg_count;
1246 if (pg_count != m->nr_priority_groups) {
1247 ti->error = "priority group count mismatch";
1252 spin_lock_irqsave(&m->lock, flags);
1253 enable_nopath_timeout(m);
1254 spin_unlock_irqrestore(&m->lock, flags);
1256 ti->num_flush_bios = 1;
1257 ti->num_discard_bios = 1;
1258 ti->num_write_zeroes_bios = 1;
1259 if (m->queue_mode == DM_TYPE_BIO_BASED)
1260 ti->per_io_data_size = multipath_per_bio_data_size();
1262 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1271 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1276 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1278 if (!atomic_read(&m->pg_init_in_progress))
1283 finish_wait(&m->pg_init_wait, &wait);
1286 static void flush_multipath_work(struct multipath *m)
1288 if (m->hw_handler_name) {
1289 unsigned long flags;
1291 if (!atomic_read(&m->pg_init_in_progress))
1294 spin_lock_irqsave(&m->lock, flags);
1295 if (atomic_read(&m->pg_init_in_progress) &&
1296 !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1297 spin_unlock_irqrestore(&m->lock, flags);
1299 flush_workqueue(kmpath_handlerd);
1300 multipath_wait_for_pg_init_completion(m);
1302 spin_lock_irqsave(&m->lock, flags);
1303 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1305 spin_unlock_irqrestore(&m->lock, flags);
1308 if (m->queue_mode == DM_TYPE_BIO_BASED)
1309 flush_work(&m->process_queued_bios);
1310 flush_work(&m->trigger_event);
1313 static void multipath_dtr(struct dm_target *ti)
1315 struct multipath *m = ti->private;
1317 disable_nopath_timeout(m);
1318 flush_multipath_work(m);
1323 * Take a path out of use.
1325 static int fail_path(struct pgpath *pgpath)
1327 unsigned long flags;
1328 struct multipath *m = pgpath->pg->m;
1330 spin_lock_irqsave(&m->lock, flags);
1332 if (!pgpath->is_active)
1335 DMWARN("%s: Failing path %s.",
1336 dm_table_device_name(m->ti->table),
1337 pgpath->path.dev->name);
1339 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1340 pgpath->is_active = false;
1341 pgpath->fail_count++;
1343 atomic_dec(&m->nr_valid_paths);
1345 if (pgpath == m->current_pgpath)
1346 m->current_pgpath = NULL;
1348 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1349 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1351 schedule_work(&m->trigger_event);
1353 enable_nopath_timeout(m);
1356 spin_unlock_irqrestore(&m->lock, flags);
1362 * Reinstate a previously-failed path
1364 static int reinstate_path(struct pgpath *pgpath)
1366 int r = 0, run_queue = 0;
1367 unsigned long flags;
1368 struct multipath *m = pgpath->pg->m;
1369 unsigned nr_valid_paths;
1371 spin_lock_irqsave(&m->lock, flags);
1373 if (pgpath->is_active)
1376 DMWARN("%s: Reinstating path %s.",
1377 dm_table_device_name(m->ti->table),
1378 pgpath->path.dev->name);
1380 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1384 pgpath->is_active = true;
1386 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1387 if (nr_valid_paths == 1) {
1388 m->current_pgpath = NULL;
1390 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1391 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1392 atomic_inc(&m->pg_init_in_progress);
1395 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1396 pgpath->path.dev->name, nr_valid_paths);
1398 schedule_work(&m->trigger_event);
1401 spin_unlock_irqrestore(&m->lock, flags);
1403 dm_table_run_md_queue_async(m->ti->table);
1404 process_queued_io_list(m);
1407 if (pgpath->is_active)
1408 disable_nopath_timeout(m);
1414 * Fail or reinstate all paths that match the provided struct dm_dev.
1416 static int action_dev(struct multipath *m, struct dm_dev *dev,
1420 struct pgpath *pgpath;
1421 struct priority_group *pg;
1423 list_for_each_entry(pg, &m->priority_groups, list) {
1424 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1425 if (pgpath->path.dev == dev)
1434 * Temporarily try to avoid having to use the specified PG
1436 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1439 unsigned long flags;
1441 spin_lock_irqsave(&m->lock, flags);
1443 pg->bypassed = bypassed;
1444 m->current_pgpath = NULL;
1445 m->current_pg = NULL;
1447 spin_unlock_irqrestore(&m->lock, flags);
1449 schedule_work(&m->trigger_event);
1453 * Switch to using the specified PG from the next I/O that gets mapped
1455 static int switch_pg_num(struct multipath *m, const char *pgstr)
1457 struct priority_group *pg;
1459 unsigned long flags;
1462 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1463 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1464 DMWARN("invalid PG number supplied to switch_pg_num");
1468 spin_lock_irqsave(&m->lock, flags);
1469 list_for_each_entry(pg, &m->priority_groups, list) {
1470 pg->bypassed = false;
1474 m->current_pgpath = NULL;
1475 m->current_pg = NULL;
1478 spin_unlock_irqrestore(&m->lock, flags);
1480 schedule_work(&m->trigger_event);
1485 * Set/clear bypassed status of a PG.
1486 * PGs are numbered upwards from 1 in the order they were declared.
1488 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1490 struct priority_group *pg;
1494 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1495 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1496 DMWARN("invalid PG number supplied to bypass_pg");
1500 list_for_each_entry(pg, &m->priority_groups, list) {
1505 bypass_pg(m, pg, bypassed);
1510 * Should we retry pg_init immediately?
1512 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1514 unsigned long flags;
1515 bool limit_reached = false;
1517 spin_lock_irqsave(&m->lock, flags);
1519 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1520 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1521 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1523 limit_reached = true;
1525 spin_unlock_irqrestore(&m->lock, flags);
1527 return limit_reached;
1530 static void pg_init_done(void *data, int errors)
1532 struct pgpath *pgpath = data;
1533 struct priority_group *pg = pgpath->pg;
1534 struct multipath *m = pg->m;
1535 unsigned long flags;
1536 bool delay_retry = false;
1538 /* device or driver problems */
1543 if (!m->hw_handler_name) {
1547 DMERR("Could not failover the device: Handler scsi_dh_%s "
1548 "Error %d.", m->hw_handler_name, errors);
1550 * Fail path for now, so we do not ping pong
1554 case SCSI_DH_DEV_TEMP_BUSY:
1556 * Probably doing something like FW upgrade on the
1557 * controller so try the other pg.
1559 bypass_pg(m, pg, true);
1562 /* Wait before retrying. */
1565 case SCSI_DH_IMM_RETRY:
1566 case SCSI_DH_RES_TEMP_UNAVAIL:
1567 if (pg_init_limit_reached(m, pgpath))
1571 case SCSI_DH_DEV_OFFLINED:
1574 * We probably do not want to fail the path for a device
1575 * error, but this is what the old dm did. In future
1576 * patches we can do more advanced handling.
1581 spin_lock_irqsave(&m->lock, flags);
1583 if (pgpath == m->current_pgpath) {
1584 DMERR("Could not failover device. Error %d.", errors);
1585 m->current_pgpath = NULL;
1586 m->current_pg = NULL;
1588 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1589 pg->bypassed = false;
1591 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1592 /* Activations of other paths are still on going */
1595 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1597 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1599 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1601 if (__pg_init_all_paths(m))
1604 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1606 process_queued_io_list(m);
1609 * Wake up any thread waiting to suspend.
1611 wake_up(&m->pg_init_wait);
1614 spin_unlock_irqrestore(&m->lock, flags);
1617 static void activate_or_offline_path(struct pgpath *pgpath)
1619 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1621 if (pgpath->is_active && !blk_queue_dying(q))
1622 scsi_dh_activate(q, pg_init_done, pgpath);
1624 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1627 static void activate_path_work(struct work_struct *work)
1629 struct pgpath *pgpath =
1630 container_of(work, struct pgpath, activate_path.work);
1632 activate_or_offline_path(pgpath);
1635 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1636 blk_status_t error, union map_info *map_context)
1638 struct dm_mpath_io *mpio = get_mpio(map_context);
1639 struct pgpath *pgpath = mpio->pgpath;
1640 int r = DM_ENDIO_DONE;
1643 * We don't queue any clone request inside the multipath target
1644 * during end I/O handling, since those clone requests don't have
1645 * bio clones. If we queue them inside the multipath target,
1646 * we need to make bio clones, that requires memory allocation.
1647 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1648 * don't have bio clones.)
1649 * Instead of queueing the clone request here, we queue the original
1650 * request into dm core, which will remake a clone request and
1651 * clone bios for it and resubmit it later.
1653 if (error && blk_path_error(error)) {
1654 struct multipath *m = ti->private;
1656 if (error == BLK_STS_RESOURCE)
1657 r = DM_ENDIO_DELAY_REQUEUE;
1659 r = DM_ENDIO_REQUEUE;
1664 if (!atomic_read(&m->nr_valid_paths) &&
1665 !must_push_back_rq(m)) {
1666 if (error == BLK_STS_IOERR)
1668 /* complete with the original error */
1674 struct path_selector *ps = &pgpath->pg->ps;
1676 if (ps->type->end_io)
1677 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1678 clone->io_start_time_ns);
1684 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1685 blk_status_t *error)
1687 struct multipath *m = ti->private;
1688 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1689 struct pgpath *pgpath = mpio->pgpath;
1690 unsigned long flags;
1691 int r = DM_ENDIO_DONE;
1693 if (!*error || !blk_path_error(*error))
1699 if (!atomic_read(&m->nr_valid_paths)) {
1700 spin_lock_irqsave(&m->lock, flags);
1701 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1702 if (__must_push_back(m)) {
1703 r = DM_ENDIO_REQUEUE;
1706 *error = BLK_STS_IOERR;
1708 spin_unlock_irqrestore(&m->lock, flags);
1711 spin_unlock_irqrestore(&m->lock, flags);
1714 multipath_queue_bio(m, clone);
1715 r = DM_ENDIO_INCOMPLETE;
1718 struct path_selector *ps = &pgpath->pg->ps;
1720 if (ps->type->end_io)
1721 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1722 (mpio->start_time_ns ?:
1723 dm_start_time_ns_from_clone(clone)));
1730 * Suspend with flush can't complete until all the I/O is processed
1731 * so if the last path fails we must error any remaining I/O.
1732 * - Note that if the freeze_bdev fails while suspending, the
1733 * queue_if_no_path state is lost - userspace should reset it.
1734 * Otherwise, during noflush suspend, queue_if_no_path will not change.
1736 static void multipath_presuspend(struct dm_target *ti)
1738 struct multipath *m = ti->private;
1740 /* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1741 if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1742 queue_if_no_path(m, false, true, __func__);
1745 static void multipath_postsuspend(struct dm_target *ti)
1747 struct multipath *m = ti->private;
1749 mutex_lock(&m->work_mutex);
1750 flush_multipath_work(m);
1751 mutex_unlock(&m->work_mutex);
1755 * Restore the queue_if_no_path setting.
1757 static void multipath_resume(struct dm_target *ti)
1759 struct multipath *m = ti->private;
1760 unsigned long flags;
1762 spin_lock_irqsave(&m->lock, flags);
1763 if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1764 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1765 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1768 DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1769 dm_table_device_name(m->ti->table), __func__,
1770 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1771 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1773 spin_unlock_irqrestore(&m->lock, flags);
1777 * Info output has the following format:
1778 * num_multipath_feature_args [multipath_feature_args]*
1779 * num_handler_status_args [handler_status_args]*
1780 * num_groups init_group_number
1781 * [A|D|E num_ps_status_args [ps_status_args]*
1782 * num_paths num_selector_args
1783 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1785 * Table output has the following format (identical to the constructor string):
1786 * num_feature_args [features_args]*
1787 * num_handler_args hw_handler [hw_handler_args]*
1788 * num_groups init_group_number
1789 * [priority selector-name num_ps_args [ps_args]*
1790 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1792 static void multipath_status(struct dm_target *ti, status_type_t type,
1793 unsigned status_flags, char *result, unsigned maxlen)
1795 int sz = 0, pg_counter, pgpath_counter;
1796 unsigned long flags;
1797 struct multipath *m = ti->private;
1798 struct priority_group *pg;
1803 spin_lock_irqsave(&m->lock, flags);
1806 if (type == STATUSTYPE_INFO)
1807 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1808 atomic_read(&m->pg_init_count));
1810 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1811 (m->pg_init_retries > 0) * 2 +
1812 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1813 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1814 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1816 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1817 DMEMIT("queue_if_no_path ");
1818 if (m->pg_init_retries)
1819 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1820 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1821 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1822 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1823 DMEMIT("retain_attached_hw_handler ");
1824 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1825 switch(m->queue_mode) {
1826 case DM_TYPE_BIO_BASED:
1827 DMEMIT("queue_mode bio ");
1836 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1839 DMEMIT("1 %s ", m->hw_handler_name);
1841 DMEMIT("%u ", m->nr_priority_groups);
1844 pg_num = m->next_pg->pg_num;
1845 else if (m->current_pg)
1846 pg_num = m->current_pg->pg_num;
1848 pg_num = (m->nr_priority_groups ? 1 : 0);
1850 DMEMIT("%u ", pg_num);
1853 case STATUSTYPE_INFO:
1854 list_for_each_entry(pg, &m->priority_groups, list) {
1856 state = 'D'; /* Disabled */
1857 else if (pg == m->current_pg)
1858 state = 'A'; /* Currently Active */
1860 state = 'E'; /* Enabled */
1862 DMEMIT("%c ", state);
1864 if (pg->ps.type->status)
1865 sz += pg->ps.type->status(&pg->ps, NULL, type,
1871 DMEMIT("%u %u ", pg->nr_pgpaths,
1872 pg->ps.type->info_args);
1874 list_for_each_entry(p, &pg->pgpaths, list) {
1875 DMEMIT("%s %s %u ", p->path.dev->name,
1876 p->is_active ? "A" : "F",
1878 if (pg->ps.type->status)
1879 sz += pg->ps.type->status(&pg->ps,
1880 &p->path, type, result + sz,
1886 case STATUSTYPE_TABLE:
1887 list_for_each_entry(pg, &m->priority_groups, list) {
1888 DMEMIT("%s ", pg->ps.type->name);
1890 if (pg->ps.type->status)
1891 sz += pg->ps.type->status(&pg->ps, NULL, type,
1897 DMEMIT("%u %u ", pg->nr_pgpaths,
1898 pg->ps.type->table_args);
1900 list_for_each_entry(p, &pg->pgpaths, list) {
1901 DMEMIT("%s ", p->path.dev->name);
1902 if (pg->ps.type->status)
1903 sz += pg->ps.type->status(&pg->ps,
1904 &p->path, type, result + sz,
1910 case STATUSTYPE_IMA:
1911 sz = 0; /*reset the result pointer*/
1913 DMEMIT_TARGET_NAME_VERSION(ti->type);
1914 DMEMIT(",nr_priority_groups=%u", m->nr_priority_groups);
1917 list_for_each_entry(pg, &m->priority_groups, list) {
1919 state = 'D'; /* Disabled */
1920 else if (pg == m->current_pg)
1921 state = 'A'; /* Currently Active */
1923 state = 'E'; /* Enabled */
1924 DMEMIT(",pg_state_%d=%c", pg_counter, state);
1925 DMEMIT(",nr_pgpaths_%d=%u", pg_counter, pg->nr_pgpaths);
1926 DMEMIT(",path_selector_name_%d=%s", pg_counter, pg->ps.type->name);
1929 list_for_each_entry(p, &pg->pgpaths, list) {
1930 DMEMIT(",path_name_%d_%d=%s,is_active_%d_%d=%c,fail_count_%d_%d=%u",
1931 pg_counter, pgpath_counter, p->path.dev->name,
1932 pg_counter, pgpath_counter, p->is_active ? 'A' : 'F',
1933 pg_counter, pgpath_counter, p->fail_count);
1934 if (pg->ps.type->status) {
1935 DMEMIT(",path_selector_status_%d_%d=",
1936 pg_counter, pgpath_counter);
1937 sz += pg->ps.type->status(&pg->ps, &p->path,
1949 spin_unlock_irqrestore(&m->lock, flags);
1952 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1953 char *result, unsigned maxlen)
1957 struct multipath *m = ti->private;
1959 unsigned long flags;
1961 mutex_lock(&m->work_mutex);
1963 if (dm_suspended(ti)) {
1969 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1970 r = queue_if_no_path(m, true, false, __func__);
1971 spin_lock_irqsave(&m->lock, flags);
1972 enable_nopath_timeout(m);
1973 spin_unlock_irqrestore(&m->lock, flags);
1975 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1976 r = queue_if_no_path(m, false, false, __func__);
1977 disable_nopath_timeout(m);
1983 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1987 if (!strcasecmp(argv[0], "disable_group")) {
1988 r = bypass_pg_num(m, argv[1], true);
1990 } else if (!strcasecmp(argv[0], "enable_group")) {
1991 r = bypass_pg_num(m, argv[1], false);
1993 } else if (!strcasecmp(argv[0], "switch_group")) {
1994 r = switch_pg_num(m, argv[1]);
1996 } else if (!strcasecmp(argv[0], "reinstate_path"))
1997 action = reinstate_path;
1998 else if (!strcasecmp(argv[0], "fail_path"))
2001 DMWARN("Unrecognised multipath message received: %s", argv[0]);
2005 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
2007 DMWARN("message: error getting device %s",
2012 r = action_dev(m, dev, action);
2014 dm_put_device(ti, dev);
2017 mutex_unlock(&m->work_mutex);
2021 static int multipath_prepare_ioctl(struct dm_target *ti,
2022 struct block_device **bdev)
2024 struct multipath *m = ti->private;
2025 struct pgpath *pgpath;
2026 unsigned long flags;
2029 pgpath = READ_ONCE(m->current_pgpath);
2030 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
2031 pgpath = choose_pgpath(m, 0);
2034 if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
2035 *bdev = pgpath->path.dev->bdev;
2038 /* pg_init has not started or completed */
2042 /* No path is available */
2044 spin_lock_irqsave(&m->lock, flags);
2045 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2047 spin_unlock_irqrestore(&m->lock, flags);
2050 if (r == -ENOTCONN) {
2051 if (!READ_ONCE(m->current_pg)) {
2052 /* Path status changed, redo selection */
2053 (void) choose_pgpath(m, 0);
2055 spin_lock_irqsave(&m->lock, flags);
2056 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2057 (void) __pg_init_all_paths(m);
2058 spin_unlock_irqrestore(&m->lock, flags);
2059 dm_table_run_md_queue_async(m->ti->table);
2060 process_queued_io_list(m);
2064 * Only pass ioctls through if the device sizes match exactly.
2066 if (!r && ti->len != bdev_nr_sectors((*bdev)))
2071 static int multipath_iterate_devices(struct dm_target *ti,
2072 iterate_devices_callout_fn fn, void *data)
2074 struct multipath *m = ti->private;
2075 struct priority_group *pg;
2079 list_for_each_entry(pg, &m->priority_groups, list) {
2080 list_for_each_entry(p, &pg->pgpaths, list) {
2081 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2091 static int pgpath_busy(struct pgpath *pgpath)
2093 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2095 return blk_lld_busy(q);
2099 * We return "busy", only when we can map I/Os but underlying devices
2100 * are busy (so even if we map I/Os now, the I/Os will wait on
2101 * the underlying queue).
2102 * In other words, if we want to kill I/Os or queue them inside us
2103 * due to map unavailability, we don't return "busy". Otherwise,
2104 * dm core won't give us the I/Os and we can't do what we want.
2106 static int multipath_busy(struct dm_target *ti)
2108 bool busy = false, has_active = false;
2109 struct multipath *m = ti->private;
2110 struct priority_group *pg, *next_pg;
2111 struct pgpath *pgpath;
2113 /* pg_init in progress */
2114 if (atomic_read(&m->pg_init_in_progress))
2117 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2118 if (!atomic_read(&m->nr_valid_paths)) {
2119 unsigned long flags;
2120 spin_lock_irqsave(&m->lock, flags);
2121 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2122 spin_unlock_irqrestore(&m->lock, flags);
2123 return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2125 spin_unlock_irqrestore(&m->lock, flags);
2128 /* Guess which priority_group will be used at next mapping time */
2129 pg = READ_ONCE(m->current_pg);
2130 next_pg = READ_ONCE(m->next_pg);
2131 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2136 * We don't know which pg will be used at next mapping time.
2137 * We don't call choose_pgpath() here to avoid to trigger
2138 * pg_init just by busy checking.
2139 * So we don't know whether underlying devices we will be using
2140 * at next mapping time are busy or not. Just try mapping.
2146 * If there is one non-busy active path at least, the path selector
2147 * will be able to select it. So we consider such a pg as not busy.
2150 list_for_each_entry(pgpath, &pg->pgpaths, list) {
2151 if (pgpath->is_active) {
2153 if (!pgpath_busy(pgpath)) {
2162 * No active path in this pg, so this pg won't be used and
2163 * the current_pg will be changed at next mapping time.
2164 * We need to try mapping to determine it.
2172 /*-----------------------------------------------------------------
2174 *---------------------------------------------------------------*/
2175 static struct target_type multipath_target = {
2176 .name = "multipath",
2177 .version = {1, 14, 0},
2178 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2179 DM_TARGET_PASSES_INTEGRITY,
2180 .module = THIS_MODULE,
2181 .ctr = multipath_ctr,
2182 .dtr = multipath_dtr,
2183 .clone_and_map_rq = multipath_clone_and_map,
2184 .release_clone_rq = multipath_release_clone,
2185 .rq_end_io = multipath_end_io,
2186 .map = multipath_map_bio,
2187 .end_io = multipath_end_io_bio,
2188 .presuspend = multipath_presuspend,
2189 .postsuspend = multipath_postsuspend,
2190 .resume = multipath_resume,
2191 .status = multipath_status,
2192 .message = multipath_message,
2193 .prepare_ioctl = multipath_prepare_ioctl,
2194 .iterate_devices = multipath_iterate_devices,
2195 .busy = multipath_busy,
2198 static int __init dm_multipath_init(void)
2202 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2204 DMERR("failed to create workqueue kmpathd");
2206 goto bad_alloc_kmultipathd;
2210 * A separate workqueue is used to handle the device handlers
2211 * to avoid overloading existing workqueue. Overloading the
2212 * old workqueue would also create a bottleneck in the
2213 * path of the storage hardware device activation.
2215 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2217 if (!kmpath_handlerd) {
2218 DMERR("failed to create workqueue kmpath_handlerd");
2220 goto bad_alloc_kmpath_handlerd;
2223 r = dm_register_target(&multipath_target);
2225 DMERR("request-based register failed %d", r);
2227 goto bad_register_target;
2232 bad_register_target:
2233 destroy_workqueue(kmpath_handlerd);
2234 bad_alloc_kmpath_handlerd:
2235 destroy_workqueue(kmultipathd);
2236 bad_alloc_kmultipathd:
2240 static void __exit dm_multipath_exit(void)
2242 destroy_workqueue(kmpath_handlerd);
2243 destroy_workqueue(kmultipathd);
2245 dm_unregister_target(&multipath_target);
2248 module_init(dm_multipath_init);
2249 module_exit(dm_multipath_exit);
2251 module_param_named(queue_if_no_path_timeout_secs,
2252 queue_if_no_path_timeout_secs, ulong, S_IRUGO | S_IWUSR);
2253 MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2255 MODULE_DESCRIPTION(DM_NAME " multipath target");
2256 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2257 MODULE_LICENSE("GPL");