1 /*
2 * Copyright (C) 2003 Sistina Software Limited.
3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
4 *
5 * This file is released under the GPL.
6 */
8 #include <linux/device-mapper.h>
10 #include "dm-rq.h"
11 #include "dm-bio-record.h"
12 #include "dm-path-selector.h"
13 #include "dm-uevent.h"
15 #include <linux/blkdev.h>
16 #include <linux/ctype.h>
17 #include <linux/init.h>
18 #include <linux/mempool.h>
19 #include <linux/module.h>
20 #include <linux/pagemap.h>
21 #include <linux/slab.h>
22 #include <linux/time.h>
23 #include <linux/workqueue.h>
24 #include <linux/delay.h>
25 #include <scsi/scsi_dh.h>
26 #include <linux/atomic.h>
27 #include <linux/blk-mq.h>
29 #define DM_MSG_PREFIX "multipath"
30 #define DM_PG_INIT_DELAY_MSECS 2000
31 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
33 /* Path properties */
34 struct pgpath {
35 struct list_head list;
37 struct priority_group *pg; /* Owning PG */
38 unsigned fail_count; /* Cumulative failure count */
40 struct dm_path path;
41 struct delayed_work activate_path;
43 bool is_active:1; /* Path status */
44 };
46 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
48 /*
49 * Paths are grouped into Priority Groups and numbered from 1 upwards.
50 * Each has a path selector which controls which path gets used.
51 */
52 struct priority_group {
53 struct list_head list;
55 struct multipath *m; /* Owning multipath instance */
56 struct path_selector ps;
58 unsigned pg_num; /* Reference number */
59 unsigned nr_pgpaths; /* Number of paths in PG */
60 struct list_head pgpaths;
62 bool bypassed:1; /* Temporarily bypass this PG? */
63 };
65 /* Multipath context */
66 struct multipath {
67 unsigned long flags; /* Multipath state flags */
69 spinlock_t lock;
70 enum dm_queue_mode queue_mode;
72 struct pgpath *current_pgpath;
73 struct priority_group *current_pg;
74 struct priority_group *next_pg; /* Switch to this PG if set */
76 atomic_t nr_valid_paths; /* Total number of usable paths */
77 unsigned nr_priority_groups;
78 struct list_head priority_groups;
80 const char *hw_handler_name;
81 char *hw_handler_params;
82 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
83 unsigned pg_init_retries; /* Number of times to retry pg_init */
84 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
85 atomic_t pg_init_in_progress; /* Only one pg_init allowed at once */
86 atomic_t pg_init_count; /* Number of times pg_init called */
88 struct mutex work_mutex;
89 struct work_struct trigger_event;
90 struct dm_target *ti;
92 struct work_struct process_queued_bios;
93 struct bio_list queued_bios;
94 };
96 /*
97 * Context information attached to each io we process.
98 */
99 struct dm_mpath_io {
100 struct pgpath *pgpath;
101 size_t nr_bytes;
102 };
104 typedef int (*action_fn) (struct pgpath *pgpath);
106 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
107 static void trigger_event(struct work_struct *work);
108 static void activate_or_offline_path(struct pgpath *pgpath);
109 static void activate_path_work(struct work_struct *work);
110 static void process_queued_bios(struct work_struct *work);
112 /*-----------------------------------------------
113 * Multipath state flags.
114 *-----------------------------------------------*/
116 #define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
117 #define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
118 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
119 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
120 #define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
121 #define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
122 #define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
124 /*-----------------------------------------------
125 * Allocation routines
126 *-----------------------------------------------*/
128 static struct pgpath *alloc_pgpath(void)
129 {
130 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
132 if (!pgpath)
133 return NULL;
135 pgpath->is_active = true;
137 return pgpath;
138 }
140 static void free_pgpath(struct pgpath *pgpath)
141 {
142 kfree(pgpath);
143 }
145 static struct priority_group *alloc_priority_group(void)
146 {
147 struct priority_group *pg;
149 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
151 if (pg)
152 INIT_LIST_HEAD(&pg->pgpaths);
154 return pg;
155 }
157 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
158 {
159 struct pgpath *pgpath, *tmp;
161 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
162 list_del(&pgpath->list);
163 dm_put_device(ti, pgpath->path.dev);
164 free_pgpath(pgpath);
165 }
166 }
168 static void free_priority_group(struct priority_group *pg,
169 struct dm_target *ti)
170 {
171 struct path_selector *ps = &pg->ps;
173 if (ps->type) {
174 ps->type->destroy(ps);
175 dm_put_path_selector(ps->type);
176 }
178 free_pgpaths(&pg->pgpaths, ti);
179 kfree(pg);
180 }
182 static struct multipath *alloc_multipath(struct dm_target *ti)
183 {
184 struct multipath *m;
186 m = kzalloc(sizeof(*m), GFP_KERNEL);
187 if (m) {
188 INIT_LIST_HEAD(&m->priority_groups);
189 spin_lock_init(&m->lock);
190 atomic_set(&m->nr_valid_paths, 0);
191 INIT_WORK(&m->trigger_event, trigger_event);
192 mutex_init(&m->work_mutex);
194 m->queue_mode = DM_TYPE_NONE;
196 m->ti = ti;
197 ti->private = m;
198 }
200 return m;
201 }
203 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
204 {
205 if (m->queue_mode == DM_TYPE_NONE) {
206 /*
207 * Default to request-based.
208 */
209 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
210 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
211 else
212 m->queue_mode = DM_TYPE_REQUEST_BASED;
214 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
215 INIT_WORK(&m->process_queued_bios, process_queued_bios);
216 /*
217 * bio-based doesn't support any direct scsi_dh management;
218 * it just discovers if a scsi_dh is attached.
219 */
220 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
221 }
223 dm_table_set_type(ti->table, m->queue_mode);
225 /*
226 * Init fields that are only used when a scsi_dh is attached
227 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
228 */
229 set_bit(MPATHF_QUEUE_IO, &m->flags);
230 atomic_set(&m->pg_init_in_progress, 0);
231 atomic_set(&m->pg_init_count, 0);
232 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
233 init_waitqueue_head(&m->pg_init_wait);
235 return 0;
236 }
238 static void free_multipath(struct multipath *m)
239 {
240 struct priority_group *pg, *tmp;
242 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
243 list_del(&pg->list);
244 free_priority_group(pg, m->ti);
245 }
247 kfree(m->hw_handler_name);
248 kfree(m->hw_handler_params);
249 mutex_destroy(&m->work_mutex);
250 kfree(m);
251 }
253 static struct dm_mpath_io *get_mpio(union map_info *info)
254 {
255 return info->ptr;
256 }
258 static size_t multipath_per_bio_data_size(void)
259 {
260 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
261 }
263 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
264 {
265 return dm_per_bio_data(bio, multipath_per_bio_data_size());
266 }
268 static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
269 {
270 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
271 void *bio_details = mpio + 1;
272 return bio_details;
273 }
275 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
276 {
277 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
278 struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
280 mpio->nr_bytes = bio->bi_iter.bi_size;
281 mpio->pgpath = NULL;
282 *mpio_p = mpio;
284 dm_bio_record(bio_details, bio);
285 }
287 /*-----------------------------------------------
288 * Path selection
289 *-----------------------------------------------*/
291 static int __pg_init_all_paths(struct multipath *m)
292 {
293 struct pgpath *pgpath;
294 unsigned long pg_init_delay = 0;
296 lockdep_assert_held(&m->lock);
298 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
299 return 0;
301 atomic_inc(&m->pg_init_count);
302 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
304 /* Check here to reset pg_init_required */
305 if (!m->current_pg)
306 return 0;
308 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
309 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
310 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
311 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
312 /* Skip failed paths */
313 if (!pgpath->is_active)
314 continue;
315 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
316 pg_init_delay))
317 atomic_inc(&m->pg_init_in_progress);
318 }
319 return atomic_read(&m->pg_init_in_progress);
320 }
322 static int pg_init_all_paths(struct multipath *m)
323 {
324 int ret;
325 unsigned long flags;
327 spin_lock_irqsave(&m->lock, flags);
328 ret = __pg_init_all_paths(m);
329 spin_unlock_irqrestore(&m->lock, flags);
331 return ret;
332 }
334 static void __switch_pg(struct multipath *m, struct priority_group *pg)
335 {
336 m->current_pg = pg;
338 /* Must we initialise the PG first, and queue I/O till it's ready? */
339 if (m->hw_handler_name) {
340 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
341 set_bit(MPATHF_QUEUE_IO, &m->flags);
342 } else {
343 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
344 clear_bit(MPATHF_QUEUE_IO, &m->flags);
345 }
347 atomic_set(&m->pg_init_count, 0);
348 }
350 static struct pgpath *choose_path_in_pg(struct multipath *m,
351 struct priority_group *pg,
352 size_t nr_bytes)
353 {
354 unsigned long flags;
355 struct dm_path *path;
356 struct pgpath *pgpath;
358 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
359 if (!path)
360 return ERR_PTR(-ENXIO);
362 pgpath = path_to_pgpath(path);
364 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
365 /* Only update current_pgpath if pg changed */
366 spin_lock_irqsave(&m->lock, flags);
367 m->current_pgpath = pgpath;
368 __switch_pg(m, pg);
369 spin_unlock_irqrestore(&m->lock, flags);
370 }
372 return pgpath;
373 }
375 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
376 {
377 unsigned long flags;
378 struct priority_group *pg;
379 struct pgpath *pgpath;
380 unsigned bypassed = 1;
382 if (!atomic_read(&m->nr_valid_paths)) {
383 clear_bit(MPATHF_QUEUE_IO, &m->flags);
384 goto failed;
385 }
387 /* Were we instructed to switch PG? */
388 if (READ_ONCE(m->next_pg)) {
389 spin_lock_irqsave(&m->lock, flags);
390 pg = m->next_pg;
391 if (!pg) {
392 spin_unlock_irqrestore(&m->lock, flags);
393 goto check_current_pg;
394 }
395 m->next_pg = NULL;
396 spin_unlock_irqrestore(&m->lock, flags);
397 pgpath = choose_path_in_pg(m, pg, nr_bytes);
398 if (!IS_ERR_OR_NULL(pgpath))
399 return pgpath;
400 }
402 /* Don't change PG until it has no remaining paths */
403 check_current_pg:
404 pg = READ_ONCE(m->current_pg);
405 if (pg) {
406 pgpath = choose_path_in_pg(m, pg, nr_bytes);
407 if (!IS_ERR_OR_NULL(pgpath))
408 return pgpath;
409 }
411 /*
412 * Loop through priority groups until we find a valid path.
413 * First time we skip PGs marked 'bypassed'.
414 * Second time we only try the ones we skipped, but set
415 * pg_init_delay_retry so we do not hammer controllers.
416 */
417 do {
418 list_for_each_entry(pg, &m->priority_groups, list) {
419 if (pg->bypassed == !!bypassed)
420 continue;
421 pgpath = choose_path_in_pg(m, pg, nr_bytes);
422 if (!IS_ERR_OR_NULL(pgpath)) {
423 if (!bypassed)
424 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
425 return pgpath;
426 }
427 }
428 } while (bypassed--);
430 failed:
431 spin_lock_irqsave(&m->lock, flags);
432 m->current_pgpath = NULL;
433 m->current_pg = NULL;
434 spin_unlock_irqrestore(&m->lock, flags);
436 return NULL;
437 }
439 /*
440 * dm_report_EIO() is a macro instead of a function to make pr_debug()
441 * report the function name and line number of the function from which
442 * it has been invoked.
443 */
444 #define dm_report_EIO(m) \
445 do { \
446 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
447 \
448 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
449 dm_device_name(md), \
450 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
451 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
452 dm_noflush_suspending((m)->ti)); \
453 } while (0)
455 /*
456 * Check whether bios must be queued in the device-mapper core rather
457 * than here in the target.
458 *
459 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
460 * the same value then we are not between multipath_presuspend()
461 * and multipath_resume() calls and we have no need to check
462 * for the DMF_NOFLUSH_SUSPENDING flag.
463 */
464 static bool __must_push_back(struct multipath *m, unsigned long flags)
465 {
466 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
467 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
468 dm_noflush_suspending(m->ti));
469 }
471 /*
472 * Following functions use READ_ONCE to get atomic access to
473 * all m->flags to avoid taking spinlock
474 */
475 static bool must_push_back_rq(struct multipath *m)
476 {
477 unsigned long flags = READ_ONCE(m->flags);
478 return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
479 }
481 static bool must_push_back_bio(struct multipath *m)
482 {
483 unsigned long flags = READ_ONCE(m->flags);
484 return __must_push_back(m, flags);
485 }
487 /*
488 * Map cloned requests (request-based multipath)
489 */
490 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
491 union map_info *map_context,
492 struct request **__clone)
493 {
494 struct multipath *m = ti->private;
495 size_t nr_bytes = blk_rq_bytes(rq);
496 struct pgpath *pgpath;
497 struct block_device *bdev;
498 struct dm_mpath_io *mpio = get_mpio(map_context);
499 struct request_queue *q;
500 struct request *clone;
502 /* Do we need to select a new pgpath? */
503 pgpath = READ_ONCE(m->current_pgpath);
504 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
505 pgpath = choose_pgpath(m, nr_bytes);
507 if (!pgpath) {
508 if (must_push_back_rq(m))
509 return DM_MAPIO_DELAY_REQUEUE;
510 dm_report_EIO(m); /* Failed */
511 return DM_MAPIO_KILL;
512 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
513 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
514 pg_init_all_paths(m);
515 return DM_MAPIO_DELAY_REQUEUE;
516 }
518 mpio->pgpath = pgpath;
519 mpio->nr_bytes = nr_bytes;
521 bdev = pgpath->path.dev->bdev;
522 q = bdev_get_queue(bdev);
523 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE,
524 BLK_MQ_REQ_NOWAIT);
525 if (IS_ERR(clone)) {
526 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
527 if (blk_queue_dying(q)) {
528 atomic_inc(&m->pg_init_in_progress);
529 activate_or_offline_path(pgpath);
530 return DM_MAPIO_DELAY_REQUEUE;
531 }
533 /*
534 * blk-mq's SCHED_RESTART can cover this requeue, so we
535 * needn't deal with it by DELAY_REQUEUE. More importantly,
536 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
537 * get the queue busy feedback (via BLK_STS_RESOURCE),
538 * otherwise I/O merging can suffer.
539 */
540 if (q->mq_ops)
541 return DM_MAPIO_REQUEUE;
542 else
543 return DM_MAPIO_DELAY_REQUEUE;
544 }
545 clone->bio = clone->biotail = NULL;
546 clone->rq_disk = bdev->bd_disk;
547 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
548 *__clone = clone;
550 if (pgpath->pg->ps.type->start_io)
551 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
552 &pgpath->path,
553 nr_bytes);
554 return DM_MAPIO_REMAPPED;
555 }
557 static void multipath_release_clone(struct request *clone)
558 {
559 blk_put_request(clone);
560 }
562 /*
563 * Map cloned bios (bio-based multipath)
564 */
566 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
567 {
568 struct pgpath *pgpath;
569 unsigned long flags;
570 bool queue_io;
572 /* Do we need to select a new pgpath? */
573 pgpath = READ_ONCE(m->current_pgpath);
574 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
575 if (!pgpath || !queue_io)
576 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
578 if ((pgpath && queue_io) ||
579 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
580 /* Queue for the daemon to resubmit */
581 spin_lock_irqsave(&m->lock, flags);
582 bio_list_add(&m->queued_bios, bio);
583 spin_unlock_irqrestore(&m->lock, flags);
585 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
586 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
587 pg_init_all_paths(m);
588 else if (!queue_io)
589 queue_work(kmultipathd, &m->process_queued_bios);
591 return ERR_PTR(-EAGAIN);
592 }
594 return pgpath;
595 }
597 static struct pgpath *__map_bio_fast(struct multipath *m, struct bio *bio)
598 {
599 struct pgpath *pgpath;
600 unsigned long flags;
602 /* Do we need to select a new pgpath? */
603 /*
604 * FIXME: currently only switching path if no path (due to failure, etc)
605 * - which negates the point of using a path selector
606 */
607 pgpath = READ_ONCE(m->current_pgpath);
608 if (!pgpath)
609 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
611 if (!pgpath) {
612 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
613 /* Queue for the daemon to resubmit */
614 spin_lock_irqsave(&m->lock, flags);
615 bio_list_add(&m->queued_bios, bio);
616 spin_unlock_irqrestore(&m->lock, flags);
617 queue_work(kmultipathd, &m->process_queued_bios);
619 return ERR_PTR(-EAGAIN);
620 }
621 return NULL;
622 }
624 return pgpath;
625 }
627 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
628 struct dm_mpath_io *mpio)
629 {
630 struct pgpath *pgpath;
632 if (!m->hw_handler_name)
633 pgpath = __map_bio_fast(m, bio);
634 else
635 pgpath = __map_bio(m, bio);
637 if (IS_ERR(pgpath))
638 return DM_MAPIO_SUBMITTED;
640 if (!pgpath) {
641 if (must_push_back_bio(m))
642 return DM_MAPIO_REQUEUE;
643 dm_report_EIO(m);
644 return DM_MAPIO_KILL;
645 }
647 mpio->pgpath = pgpath;
649 bio->bi_status = 0;
650 bio_set_dev(bio, pgpath->path.dev->bdev);
651 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
653 if (pgpath->pg->ps.type->start_io)
654 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
655 &pgpath->path,
656 mpio->nr_bytes);
657 return DM_MAPIO_REMAPPED;
658 }
660 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
661 {
662 struct multipath *m = ti->private;
663 struct dm_mpath_io *mpio = NULL;
665 multipath_init_per_bio_data(bio, &mpio);
666 return __multipath_map_bio(m, bio, mpio);
667 }
669 static void process_queued_io_list(struct multipath *m)
670 {
671 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
672 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
673 else if (m->queue_mode == DM_TYPE_BIO_BASED)
674 queue_work(kmultipathd, &m->process_queued_bios);
675 }
677 static void process_queued_bios(struct work_struct *work)
678 {
679 int r;
680 unsigned long flags;
681 struct bio *bio;
682 struct bio_list bios;
683 struct blk_plug plug;
684 struct multipath *m =
685 container_of(work, struct multipath, process_queued_bios);
687 bio_list_init(&bios);
689 spin_lock_irqsave(&m->lock, flags);
691 if (bio_list_empty(&m->queued_bios)) {
692 spin_unlock_irqrestore(&m->lock, flags);
693 return;
694 }
696 bio_list_merge(&bios, &m->queued_bios);
697 bio_list_init(&m->queued_bios);
699 spin_unlock_irqrestore(&m->lock, flags);
701 blk_start_plug(&plug);
702 while ((bio = bio_list_pop(&bios))) {
703 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
704 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
705 r = __multipath_map_bio(m, bio, mpio);
706 switch (r) {
707 case DM_MAPIO_KILL:
708 bio->bi_status = BLK_STS_IOERR;
709 bio_endio(bio);
710 break;
711 case DM_MAPIO_REQUEUE:
712 bio->bi_status = BLK_STS_DM_REQUEUE;
713 bio_endio(bio);
714 break;
715 case DM_MAPIO_REMAPPED:
716 generic_make_request(bio);
717 break;
718 case DM_MAPIO_SUBMITTED:
719 break;
720 default:
721 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
722 }
723 }
724 blk_finish_plug(&plug);
725 }
727 /*
728 * If we run out of usable paths, should we queue I/O or error it?
729 */
730 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
731 bool save_old_value)
732 {
733 unsigned long flags;
735 spin_lock_irqsave(&m->lock, flags);
736 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
737 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
738 (!save_old_value && queue_if_no_path));
739 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
740 spin_unlock_irqrestore(&m->lock, flags);
742 if (!queue_if_no_path) {
743 dm_table_run_md_queue_async(m->ti->table);
744 process_queued_io_list(m);
745 }
747 return 0;
748 }
750 /*
751 * An event is triggered whenever a path is taken out of use.
752 * Includes path failure and PG bypass.
753 */
754 static void trigger_event(struct work_struct *work)
755 {
756 struct multipath *m =
757 container_of(work, struct multipath, trigger_event);
759 dm_table_event(m->ti->table);
760 }
762 /*-----------------------------------------------------------------
763 * Constructor/argument parsing:
764 * <#multipath feature args> [<arg>]*
765 * <#hw_handler args> [hw_handler [<arg>]*]
766 * <#priority groups>
767 * <initial priority group>
768 * [<selector> <#selector args> [<arg>]*
769 * <#paths> <#per-path selector args>
770 * [<path> [<arg>]* ]+ ]+
771 *---------------------------------------------------------------*/
772 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
773 struct dm_target *ti)
774 {
775 int r;
776 struct path_selector_type *pst;
777 unsigned ps_argc;
779 static const struct dm_arg _args[] = {
780 {0, 1024, "invalid number of path selector args"},
781 };
783 pst = dm_get_path_selector(dm_shift_arg(as));
784 if (!pst) {
785 ti->error = "unknown path selector type";
786 return -EINVAL;
787 }
789 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
790 if (r) {
791 dm_put_path_selector(pst);
792 return -EINVAL;
793 }
795 r = pst->create(&pg->ps, ps_argc, as->argv);
796 if (r) {
797 dm_put_path_selector(pst);
798 ti->error = "path selector constructor failed";
799 return r;
800 }
802 pg->ps.type = pst;
803 dm_consume_args(as, ps_argc);
805 return 0;
806 }
808 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
809 const char **attached_handler_name, char **error)
810 {
811 struct request_queue *q = bdev_get_queue(bdev);
812 int r;
814 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
815 retain:
816 if (*attached_handler_name) {
817 /*
818 * Clear any hw_handler_params associated with a
819 * handler that isn't already attached.
820 */
821 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
822 kfree(m->hw_handler_params);
823 m->hw_handler_params = NULL;
824 }
826 /*
827 * Reset hw_handler_name to match the attached handler
828 *
829 * NB. This modifies the table line to show the actual
830 * handler instead of the original table passed in.
831 */
832 kfree(m->hw_handler_name);
833 m->hw_handler_name = *attached_handler_name;
834 *attached_handler_name = NULL;
835 }
836 }
838 if (m->hw_handler_name) {
839 r = scsi_dh_attach(q, m->hw_handler_name);
840 if (r == -EBUSY) {
841 char b[BDEVNAME_SIZE];
843 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
844 bdevname(bdev, b));
845 goto retain;
846 }
847 if (r < 0) {
848 *error = "error attaching hardware handler";
849 return r;
850 }
852 if (m->hw_handler_params) {
853 r = scsi_dh_set_params(q, m->hw_handler_params);
854 if (r < 0) {
855 *error = "unable to set hardware handler parameters";
856 return r;
857 }
858 }
859 }
861 return 0;
862 }
864 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
865 struct dm_target *ti)
866 {
867 int r;
868 struct pgpath *p;
869 struct multipath *m = ti->private;
870 struct request_queue *q;
871 const char *attached_handler_name = NULL;
873 /* we need at least a path arg */
874 if (as->argc < 1) {
875 ti->error = "no device given";
876 return ERR_PTR(-EINVAL);
877 }
879 p = alloc_pgpath();
880 if (!p)
881 return ERR_PTR(-ENOMEM);
883 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
884 &p->path.dev);
885 if (r) {
886 ti->error = "error getting device";
887 goto bad;
888 }
890 q = bdev_get_queue(p->path.dev->bdev);
891 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
892 if (attached_handler_name || m->hw_handler_name) {
893 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
894 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
895 if (r) {
896 dm_put_device(ti, p->path.dev);
897 goto bad;
898 }
899 }
901 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
902 if (r) {
903 dm_put_device(ti, p->path.dev);
904 goto bad;
905 }
907 return p;
908 bad:
909 kfree(attached_handler_name);
910 free_pgpath(p);
911 return ERR_PTR(r);
912 }
914 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
915 struct multipath *m)
916 {
917 static const struct dm_arg _args[] = {
918 {1, 1024, "invalid number of paths"},
919 {0, 1024, "invalid number of selector args"}
920 };
922 int r;
923 unsigned i, nr_selector_args, nr_args;
924 struct priority_group *pg;
925 struct dm_target *ti = m->ti;
927 if (as->argc < 2) {
928 as->argc = 0;
929 ti->error = "not enough priority group arguments";
930 return ERR_PTR(-EINVAL);
931 }
933 pg = alloc_priority_group();
934 if (!pg) {
935 ti->error = "couldn't allocate priority group";
936 return ERR_PTR(-ENOMEM);
937 }
938 pg->m = m;
940 r = parse_path_selector(as, pg, ti);
941 if (r)
942 goto bad;
944 /*
945 * read the paths
946 */
947 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
948 if (r)
949 goto bad;
951 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
952 if (r)
953 goto bad;
955 nr_args = 1 + nr_selector_args;
956 for (i = 0; i < pg->nr_pgpaths; i++) {
957 struct pgpath *pgpath;
958 struct dm_arg_set path_args;
960 if (as->argc < nr_args) {
961 ti->error = "not enough path parameters";
962 r = -EINVAL;
963 goto bad;
964 }
966 path_args.argc = nr_args;
967 path_args.argv = as->argv;
969 pgpath = parse_path(&path_args, &pg->ps, ti);
970 if (IS_ERR(pgpath)) {
971 r = PTR_ERR(pgpath);
972 goto bad;
973 }
975 pgpath->pg = pg;
976 list_add_tail(&pgpath->list, &pg->pgpaths);
977 dm_consume_args(as, nr_args);
978 }
980 return pg;
982 bad:
983 free_priority_group(pg, ti);
984 return ERR_PTR(r);
985 }
987 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
988 {
989 unsigned hw_argc;
990 int ret;
991 struct dm_target *ti = m->ti;
993 static const struct dm_arg _args[] = {
994 {0, 1024, "invalid number of hardware handler args"},
995 };
997 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
998 return -EINVAL;
1000 if (!hw_argc)
1001 return 0;
1003 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1004 dm_consume_args(as, hw_argc);
1005 DMERR("bio-based multipath doesn't allow hardware handler args");
1006 return 0;
1007 }
1009 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1010 if (!m->hw_handler_name)
1011 return -EINVAL;
1013 if (hw_argc > 1) {
1014 char *p;
1015 int i, j, len = 4;
1017 for (i = 0; i <= hw_argc - 2; i++)
1018 len += strlen(as->argv[i]) + 1;
1019 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1020 if (!p) {
1021 ti->error = "memory allocation failed";
1022 ret = -ENOMEM;
1023 goto fail;
1024 }
1025 j = sprintf(p, "%d", hw_argc - 1);
1026 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1027 j = sprintf(p, "%s", as->argv[i]);
1028 }
1029 dm_consume_args(as, hw_argc - 1);
1031 return 0;
1032 fail:
1033 kfree(m->hw_handler_name);
1034 m->hw_handler_name = NULL;
1035 return ret;
1036 }
1038 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1039 {
1040 int r;
1041 unsigned argc;
1042 struct dm_target *ti = m->ti;
1043 const char *arg_name;
1045 static const struct dm_arg _args[] = {
1046 {0, 8, "invalid number of feature args"},
1047 {1, 50, "pg_init_retries must be between 1 and 50"},
1048 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1049 };
1051 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1052 if (r)
1053 return -EINVAL;
1055 if (!argc)
1056 return 0;
1058 do {
1059 arg_name = dm_shift_arg(as);
1060 argc--;
1062 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1063 r = queue_if_no_path(m, true, false);
1064 continue;
1065 }
1067 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1068 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1069 continue;
1070 }
1072 if (!strcasecmp(arg_name, "pg_init_retries") &&
1073 (argc >= 1)) {
1074 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1075 argc--;
1076 continue;
1077 }
1079 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1080 (argc >= 1)) {
1081 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1082 argc--;
1083 continue;
1084 }
1086 if (!strcasecmp(arg_name, "queue_mode") &&
1087 (argc >= 1)) {
1088 const char *queue_mode_name = dm_shift_arg(as);
1090 if (!strcasecmp(queue_mode_name, "bio"))
1091 m->queue_mode = DM_TYPE_BIO_BASED;
1092 else if (!strcasecmp(queue_mode_name, "rq"))
1093 m->queue_mode = DM_TYPE_REQUEST_BASED;
1094 else if (!strcasecmp(queue_mode_name, "mq"))
1095 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1096 else {
1097 ti->error = "Unknown 'queue_mode' requested";
1098 r = -EINVAL;
1099 }
1100 argc--;
1101 continue;
1102 }
1104 ti->error = "Unrecognised multipath feature request";
1105 r = -EINVAL;
1106 } while (argc && !r);
1108 return r;
1109 }
1111 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1112 {
1113 /* target arguments */
1114 static const struct dm_arg _args[] = {
1115 {0, 1024, "invalid number of priority groups"},
1116 {0, 1024, "invalid initial priority group number"},
1117 };
1119 int r;
1120 struct multipath *m;
1121 struct dm_arg_set as;
1122 unsigned pg_count = 0;
1123 unsigned next_pg_num;
1125 as.argc = argc;
1126 as.argv = argv;
1128 m = alloc_multipath(ti);
1129 if (!m) {
1130 ti->error = "can't allocate multipath";
1131 return -EINVAL;
1132 }
1134 r = parse_features(&as, m);
1135 if (r)
1136 goto bad;
1138 r = alloc_multipath_stage2(ti, m);
1139 if (r)
1140 goto bad;
1142 r = parse_hw_handler(&as, m);
1143 if (r)
1144 goto bad;
1146 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1147 if (r)
1148 goto bad;
1150 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1151 if (r)
1152 goto bad;
1154 if ((!m->nr_priority_groups && next_pg_num) ||
1155 (m->nr_priority_groups && !next_pg_num)) {
1156 ti->error = "invalid initial priority group";
1157 r = -EINVAL;
1158 goto bad;
1159 }
1161 /* parse the priority groups */
1162 while (as.argc) {
1163 struct priority_group *pg;
1164 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1166 pg = parse_priority_group(&as, m);
1167 if (IS_ERR(pg)) {
1168 r = PTR_ERR(pg);
1169 goto bad;
1170 }
1172 nr_valid_paths += pg->nr_pgpaths;
1173 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1175 list_add_tail(&pg->list, &m->priority_groups);
1176 pg_count++;
1177 pg->pg_num = pg_count;
1178 if (!--next_pg_num)
1179 m->next_pg = pg;
1180 }
1182 if (pg_count != m->nr_priority_groups) {
1183 ti->error = "priority group count mismatch";
1184 r = -EINVAL;
1185 goto bad;
1186 }
1188 ti->num_flush_bios = 1;
1189 ti->num_discard_bios = 1;
1190 ti->num_write_same_bios = 1;
1191 ti->num_write_zeroes_bios = 1;
1192 if (m->queue_mode == DM_TYPE_BIO_BASED)
1193 ti->per_io_data_size = multipath_per_bio_data_size();
1194 else
1195 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1197 return 0;
1199 bad:
1200 free_multipath(m);
1201 return r;
1202 }
1204 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1205 {
1206 DEFINE_WAIT(wait);
1208 while (1) {
1209 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1211 if (!atomic_read(&m->pg_init_in_progress))
1212 break;
1214 io_schedule();
1215 }
1216 finish_wait(&m->pg_init_wait, &wait);
1217 }
1219 static void flush_multipath_work(struct multipath *m)
1220 {
1221 if (m->hw_handler_name) {
1222 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1223 smp_mb__after_atomic();
1225 flush_workqueue(kmpath_handlerd);
1226 multipath_wait_for_pg_init_completion(m);
1228 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1229 smp_mb__after_atomic();
1230 }
1232 flush_workqueue(kmultipathd);
1233 flush_work(&m->trigger_event);
1234 }
1236 static void multipath_dtr(struct dm_target *ti)
1237 {
1238 struct multipath *m = ti->private;
1240 flush_multipath_work(m);
1241 free_multipath(m);
1242 }
1244 /*
1245 * Take a path out of use.
1246 */
1247 static int fail_path(struct pgpath *pgpath)
1248 {
1249 unsigned long flags;
1250 struct multipath *m = pgpath->pg->m;
1252 spin_lock_irqsave(&m->lock, flags);
1254 if (!pgpath->is_active)
1255 goto out;
1257 DMWARN("Failing path %s.", pgpath->path.dev->name);
1259 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1260 pgpath->is_active = false;
1261 pgpath->fail_count++;
1263 atomic_dec(&m->nr_valid_paths);
1265 if (pgpath == m->current_pgpath)
1266 m->current_pgpath = NULL;
1268 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1269 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1271 schedule_work(&m->trigger_event);
1273 out:
1274 spin_unlock_irqrestore(&m->lock, flags);
1276 return 0;
1277 }
1279 /*
1280 * Reinstate a previously-failed path
1281 */
1282 static int reinstate_path(struct pgpath *pgpath)
1283 {
1284 int r = 0, run_queue = 0;
1285 unsigned long flags;
1286 struct multipath *m = pgpath->pg->m;
1287 unsigned nr_valid_paths;
1289 spin_lock_irqsave(&m->lock, flags);
1291 if (pgpath->is_active)
1292 goto out;
1294 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1296 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1297 if (r)
1298 goto out;
1300 pgpath->is_active = true;
1302 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1303 if (nr_valid_paths == 1) {
1304 m->current_pgpath = NULL;
1305 run_queue = 1;
1306 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1307 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1308 atomic_inc(&m->pg_init_in_progress);
1309 }
1311 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1312 pgpath->path.dev->name, nr_valid_paths);
1314 schedule_work(&m->trigger_event);
1316 out:
1317 spin_unlock_irqrestore(&m->lock, flags);
1318 if (run_queue) {
1319 dm_table_run_md_queue_async(m->ti->table);
1320 process_queued_io_list(m);
1321 }
1323 return r;
1324 }
1326 /*
1327 * Fail or reinstate all paths that match the provided struct dm_dev.
1328 */
1329 static int action_dev(struct multipath *m, struct dm_dev *dev,
1330 action_fn action)
1331 {
1332 int r = -EINVAL;
1333 struct pgpath *pgpath;
1334 struct priority_group *pg;
1336 list_for_each_entry(pg, &m->priority_groups, list) {
1337 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1338 if (pgpath->path.dev == dev)
1339 r = action(pgpath);
1340 }
1341 }
1343 return r;
1344 }
1346 /*
1347 * Temporarily try to avoid having to use the specified PG
1348 */
1349 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1350 bool bypassed)
1351 {
1352 unsigned long flags;
1354 spin_lock_irqsave(&m->lock, flags);
1356 pg->bypassed = bypassed;
1357 m->current_pgpath = NULL;
1358 m->current_pg = NULL;
1360 spin_unlock_irqrestore(&m->lock, flags);
1362 schedule_work(&m->trigger_event);
1363 }
1365 /*
1366 * Switch to using the specified PG from the next I/O that gets mapped
1367 */
1368 static int switch_pg_num(struct multipath *m, const char *pgstr)
1369 {
1370 struct priority_group *pg;
1371 unsigned pgnum;
1372 unsigned long flags;
1373 char dummy;
1375 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1376 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1377 DMWARN("invalid PG number supplied to switch_pg_num");
1378 return -EINVAL;
1379 }
1381 spin_lock_irqsave(&m->lock, flags);
1382 list_for_each_entry(pg, &m->priority_groups, list) {
1383 pg->bypassed = false;
1384 if (--pgnum)
1385 continue;
1387 m->current_pgpath = NULL;
1388 m->current_pg = NULL;
1389 m->next_pg = pg;
1390 }
1391 spin_unlock_irqrestore(&m->lock, flags);
1393 schedule_work(&m->trigger_event);
1394 return 0;
1395 }
1397 /*
1398 * Set/clear bypassed status of a PG.
1399 * PGs are numbered upwards from 1 in the order they were declared.
1400 */
1401 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1402 {
1403 struct priority_group *pg;
1404 unsigned pgnum;
1405 char dummy;
1407 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1408 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1409 DMWARN("invalid PG number supplied to bypass_pg");
1410 return -EINVAL;
1411 }
1413 list_for_each_entry(pg, &m->priority_groups, list) {
1414 if (!--pgnum)
1415 break;
1416 }
1418 bypass_pg(m, pg, bypassed);
1419 return 0;
1420 }
1422 /*
1423 * Should we retry pg_init immediately?
1424 */
1425 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1426 {
1427 unsigned long flags;
1428 bool limit_reached = false;
1430 spin_lock_irqsave(&m->lock, flags);
1432 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1433 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1434 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1435 else
1436 limit_reached = true;
1438 spin_unlock_irqrestore(&m->lock, flags);
1440 return limit_reached;
1441 }
1443 static void pg_init_done(void *data, int errors)
1444 {
1445 struct pgpath *pgpath = data;
1446 struct priority_group *pg = pgpath->pg;
1447 struct multipath *m = pg->m;
1448 unsigned long flags;
1449 bool delay_retry = false;
1451 /* device or driver problems */
1452 switch (errors) {
1453 case SCSI_DH_OK:
1454 break;
1455 case SCSI_DH_NOSYS:
1456 if (!m->hw_handler_name) {
1457 errors = 0;
1458 break;
1459 }
1460 DMERR("Could not failover the device: Handler scsi_dh_%s "
1461 "Error %d.", m->hw_handler_name, errors);
1462 /*
1463 * Fail path for now, so we do not ping pong
1464 */
1465 fail_path(pgpath);
1466 break;
1467 case SCSI_DH_DEV_TEMP_BUSY:
1468 /*
1469 * Probably doing something like FW upgrade on the
1470 * controller so try the other pg.
1471 */
1472 bypass_pg(m, pg, true);
1473 break;
1474 case SCSI_DH_RETRY:
1475 /* Wait before retrying. */
1476 delay_retry = 1;
1477 /* fall through */
1478 case SCSI_DH_IMM_RETRY:
1479 case SCSI_DH_RES_TEMP_UNAVAIL:
1480 if (pg_init_limit_reached(m, pgpath))
1481 fail_path(pgpath);
1482 errors = 0;
1483 break;
1484 case SCSI_DH_DEV_OFFLINED:
1485 default:
1486 /*
1487 * We probably do not want to fail the path for a device
1488 * error, but this is what the old dm did. In future
1489 * patches we can do more advanced handling.
1490 */
1491 fail_path(pgpath);
1492 }
1494 spin_lock_irqsave(&m->lock, flags);
1495 if (errors) {
1496 if (pgpath == m->current_pgpath) {
1497 DMERR("Could not failover device. Error %d.", errors);
1498 m->current_pgpath = NULL;
1499 m->current_pg = NULL;
1500 }
1501 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1502 pg->bypassed = false;
1504 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1505 /* Activations of other paths are still on going */
1506 goto out;
1508 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1509 if (delay_retry)
1510 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1511 else
1512 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1514 if (__pg_init_all_paths(m))
1515 goto out;
1516 }
1517 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1519 process_queued_io_list(m);
1521 /*
1522 * Wake up any thread waiting to suspend.
1523 */
1524 wake_up(&m->pg_init_wait);
1526 out:
1527 spin_unlock_irqrestore(&m->lock, flags);
1528 }
1530 static void activate_or_offline_path(struct pgpath *pgpath)
1531 {
1532 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1534 if (pgpath->is_active && !blk_queue_dying(q))
1535 scsi_dh_activate(q, pg_init_done, pgpath);
1536 else
1537 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1538 }
1540 static void activate_path_work(struct work_struct *work)
1541 {
1542 struct pgpath *pgpath =
1543 container_of(work, struct pgpath, activate_path.work);
1545 activate_or_offline_path(pgpath);
1546 }
1548 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1549 blk_status_t error, union map_info *map_context)
1550 {
1551 struct dm_mpath_io *mpio = get_mpio(map_context);
1552 struct pgpath *pgpath = mpio->pgpath;
1553 int r = DM_ENDIO_DONE;
1555 /*
1556 * We don't queue any clone request inside the multipath target
1557 * during end I/O handling, since those clone requests don't have
1558 * bio clones. If we queue them inside the multipath target,
1559 * we need to make bio clones, that requires memory allocation.
1560 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1561 * don't have bio clones.)
1562 * Instead of queueing the clone request here, we queue the original
1563 * request into dm core, which will remake a clone request and
1564 * clone bios for it and resubmit it later.
1565 */
1566 if (error && blk_path_error(error)) {
1567 struct multipath *m = ti->private;
1569 if (error == BLK_STS_RESOURCE)
1570 r = DM_ENDIO_DELAY_REQUEUE;
1571 else
1572 r = DM_ENDIO_REQUEUE;
1574 if (pgpath)
1575 fail_path(pgpath);
1577 if (atomic_read(&m->nr_valid_paths) == 0 &&
1578 !must_push_back_rq(m)) {
1579 if (error == BLK_STS_IOERR)
1580 dm_report_EIO(m);
1581 /* complete with the original error */
1582 r = DM_ENDIO_DONE;
1583 }
1584 }
1586 if (pgpath) {
1587 struct path_selector *ps = &pgpath->pg->ps;
1589 if (ps->type->end_io)
1590 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1591 }
1593 return r;
1594 }
1596 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1597 blk_status_t *error)
1598 {
1599 struct multipath *m = ti->private;
1600 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1601 struct pgpath *pgpath = mpio->pgpath;
1602 unsigned long flags;
1603 int r = DM_ENDIO_DONE;
1605 if (!*error || !blk_path_error(*error))
1606 goto done;
1608 if (pgpath)
1609 fail_path(pgpath);
1611 if (atomic_read(&m->nr_valid_paths) == 0 &&
1612 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1613 if (must_push_back_bio(m)) {
1614 r = DM_ENDIO_REQUEUE;
1615 } else {
1616 dm_report_EIO(m);
1617 *error = BLK_STS_IOERR;
1618 }
1619 goto done;
1620 }
1622 spin_lock_irqsave(&m->lock, flags);
1623 bio_list_add(&m->queued_bios, clone);
1624 spin_unlock_irqrestore(&m->lock, flags);
1625 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1626 queue_work(kmultipathd, &m->process_queued_bios);
1628 r = DM_ENDIO_INCOMPLETE;
1629 done:
1630 if (pgpath) {
1631 struct path_selector *ps = &pgpath->pg->ps;
1633 if (ps->type->end_io)
1634 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1635 }
1637 return r;
1638 }
1640 /*
1641 * Suspend can't complete until all the I/O is processed so if
1642 * the last path fails we must error any remaining I/O.
1643 * Note that if the freeze_bdev fails while suspending, the
1644 * queue_if_no_path state is lost - userspace should reset it.
1645 */
1646 static void multipath_presuspend(struct dm_target *ti)
1647 {
1648 struct multipath *m = ti->private;
1650 queue_if_no_path(m, false, true);
1651 }
1653 static void multipath_postsuspend(struct dm_target *ti)
1654 {
1655 struct multipath *m = ti->private;
1657 mutex_lock(&m->work_mutex);
1658 flush_multipath_work(m);
1659 mutex_unlock(&m->work_mutex);
1660 }
1662 /*
1663 * Restore the queue_if_no_path setting.
1664 */
1665 static void multipath_resume(struct dm_target *ti)
1666 {
1667 struct multipath *m = ti->private;
1668 unsigned long flags;
1670 spin_lock_irqsave(&m->lock, flags);
1671 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1672 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1673 spin_unlock_irqrestore(&m->lock, flags);
1674 }
1676 /*
1677 * Info output has the following format:
1678 * num_multipath_feature_args [multipath_feature_args]*
1679 * num_handler_status_args [handler_status_args]*
1680 * num_groups init_group_number
1681 * [A|D|E num_ps_status_args [ps_status_args]*
1682 * num_paths num_selector_args
1683 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1684 *
1685 * Table output has the following format (identical to the constructor string):
1686 * num_feature_args [features_args]*
1687 * num_handler_args hw_handler [hw_handler_args]*
1688 * num_groups init_group_number
1689 * [priority selector-name num_ps_args [ps_args]*
1690 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1691 */
1692 static void multipath_status(struct dm_target *ti, status_type_t type,
1693 unsigned status_flags, char *result, unsigned maxlen)
1694 {
1695 int sz = 0;
1696 unsigned long flags;
1697 struct multipath *m = ti->private;
1698 struct priority_group *pg;
1699 struct pgpath *p;
1700 unsigned pg_num;
1701 char state;
1703 spin_lock_irqsave(&m->lock, flags);
1705 /* Features */
1706 if (type == STATUSTYPE_INFO)
1707 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1708 atomic_read(&m->pg_init_count));
1709 else {
1710 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1711 (m->pg_init_retries > 0) * 2 +
1712 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1713 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1714 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1716 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1717 DMEMIT("queue_if_no_path ");
1718 if (m->pg_init_retries)
1719 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1720 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1721 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1722 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1723 DMEMIT("retain_attached_hw_handler ");
1724 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1725 switch(m->queue_mode) {
1726 case DM_TYPE_BIO_BASED:
1727 DMEMIT("queue_mode bio ");
1728 break;
1729 case DM_TYPE_MQ_REQUEST_BASED:
1730 DMEMIT("queue_mode mq ");
1731 break;
1732 default:
1733 WARN_ON_ONCE(true);
1734 break;
1735 }
1736 }
1737 }
1739 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1740 DMEMIT("0 ");
1741 else
1742 DMEMIT("1 %s ", m->hw_handler_name);
1744 DMEMIT("%u ", m->nr_priority_groups);
1746 if (m->next_pg)
1747 pg_num = m->next_pg->pg_num;
1748 else if (m->current_pg)
1749 pg_num = m->current_pg->pg_num;
1750 else
1751 pg_num = (m->nr_priority_groups ? 1 : 0);
1753 DMEMIT("%u ", pg_num);
1755 switch (type) {
1756 case STATUSTYPE_INFO:
1757 list_for_each_entry(pg, &m->priority_groups, list) {
1758 if (pg->bypassed)
1759 state = 'D'; /* Disabled */
1760 else if (pg == m->current_pg)
1761 state = 'A'; /* Currently Active */
1762 else
1763 state = 'E'; /* Enabled */
1765 DMEMIT("%c ", state);
1767 if (pg->ps.type->status)
1768 sz += pg->ps.type->status(&pg->ps, NULL, type,
1769 result + sz,
1770 maxlen - sz);
1771 else
1772 DMEMIT("0 ");
1774 DMEMIT("%u %u ", pg->nr_pgpaths,
1775 pg->ps.type->info_args);
1777 list_for_each_entry(p, &pg->pgpaths, list) {
1778 DMEMIT("%s %s %u ", p->path.dev->name,
1779 p->is_active ? "A" : "F",
1780 p->fail_count);
1781 if (pg->ps.type->status)
1782 sz += pg->ps.type->status(&pg->ps,
1783 &p->path, type, result + sz,
1784 maxlen - sz);
1785 }
1786 }
1787 break;
1789 case STATUSTYPE_TABLE:
1790 list_for_each_entry(pg, &m->priority_groups, list) {
1791 DMEMIT("%s ", pg->ps.type->name);
1793 if (pg->ps.type->status)
1794 sz += pg->ps.type->status(&pg->ps, NULL, type,
1795 result + sz,
1796 maxlen - sz);
1797 else
1798 DMEMIT("0 ");
1800 DMEMIT("%u %u ", pg->nr_pgpaths,
1801 pg->ps.type->table_args);
1803 list_for_each_entry(p, &pg->pgpaths, list) {
1804 DMEMIT("%s ", p->path.dev->name);
1805 if (pg->ps.type->status)
1806 sz += pg->ps.type->status(&pg->ps,
1807 &p->path, type, result + sz,
1808 maxlen - sz);
1809 }
1810 }
1811 break;
1812 }
1814 spin_unlock_irqrestore(&m->lock, flags);
1815 }
1817 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1818 char *result, unsigned maxlen)
1819 {
1820 int r = -EINVAL;
1821 struct dm_dev *dev;
1822 struct multipath *m = ti->private;
1823 action_fn action;
1825 mutex_lock(&m->work_mutex);
1827 if (dm_suspended(ti)) {
1828 r = -EBUSY;
1829 goto out;
1830 }
1832 if (argc == 1) {
1833 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1834 r = queue_if_no_path(m, true, false);
1835 goto out;
1836 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1837 r = queue_if_no_path(m, false, false);
1838 goto out;
1839 }
1840 }
1842 if (argc != 2) {
1843 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1844 goto out;
1845 }
1847 if (!strcasecmp(argv[0], "disable_group")) {
1848 r = bypass_pg_num(m, argv[1], true);
1849 goto out;
1850 } else if (!strcasecmp(argv[0], "enable_group")) {
1851 r = bypass_pg_num(m, argv[1], false);
1852 goto out;
1853 } else if (!strcasecmp(argv[0], "switch_group")) {
1854 r = switch_pg_num(m, argv[1]);
1855 goto out;
1856 } else if (!strcasecmp(argv[0], "reinstate_path"))
1857 action = reinstate_path;
1858 else if (!strcasecmp(argv[0], "fail_path"))
1859 action = fail_path;
1860 else {
1861 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1862 goto out;
1863 }
1865 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1866 if (r) {
1867 DMWARN("message: error getting device %s",
1868 argv[1]);
1869 goto out;
1870 }
1872 r = action_dev(m, dev, action);
1874 dm_put_device(ti, dev);
1876 out:
1877 mutex_unlock(&m->work_mutex);
1878 return r;
1879 }
1881 static int multipath_prepare_ioctl(struct dm_target *ti,
1882 struct block_device **bdev)
1883 {
1884 struct multipath *m = ti->private;
1885 struct pgpath *current_pgpath;
1886 int r;
1888 current_pgpath = READ_ONCE(m->current_pgpath);
1889 if (!current_pgpath)
1890 current_pgpath = choose_pgpath(m, 0);
1892 if (current_pgpath) {
1893 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1894 *bdev = current_pgpath->path.dev->bdev;
1895 r = 0;
1896 } else {
1897 /* pg_init has not started or completed */
1898 r = -ENOTCONN;
1899 }
1900 } else {
1901 /* No path is available */
1902 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1903 r = -ENOTCONN;
1904 else
1905 r = -EIO;
1906 }
1908 if (r == -ENOTCONN) {
1909 if (!READ_ONCE(m->current_pg)) {
1910 /* Path status changed, redo selection */
1911 (void) choose_pgpath(m, 0);
1912 }
1913 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1914 pg_init_all_paths(m);
1915 dm_table_run_md_queue_async(m->ti->table);
1916 process_queued_io_list(m);
1917 }
1919 /*
1920 * Only pass ioctls through if the device sizes match exactly.
1921 */
1922 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1923 return 1;
1924 return r;
1925 }
1927 static int multipath_iterate_devices(struct dm_target *ti,
1928 iterate_devices_callout_fn fn, void *data)
1929 {
1930 struct multipath *m = ti->private;
1931 struct priority_group *pg;
1932 struct pgpath *p;
1933 int ret = 0;
1935 list_for_each_entry(pg, &m->priority_groups, list) {
1936 list_for_each_entry(p, &pg->pgpaths, list) {
1937 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1938 if (ret)
1939 goto out;
1940 }
1941 }
1943 out:
1944 return ret;
1945 }
1947 static int pgpath_busy(struct pgpath *pgpath)
1948 {
1949 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1951 return blk_lld_busy(q);
1952 }
1954 /*
1955 * We return "busy", only when we can map I/Os but underlying devices
1956 * are busy (so even if we map I/Os now, the I/Os will wait on
1957 * the underlying queue).
1958 * In other words, if we want to kill I/Os or queue them inside us
1959 * due to map unavailability, we don't return "busy". Otherwise,
1960 * dm core won't give us the I/Os and we can't do what we want.
1961 */
1962 static int multipath_busy(struct dm_target *ti)
1963 {
1964 bool busy = false, has_active = false;
1965 struct multipath *m = ti->private;
1966 struct priority_group *pg, *next_pg;
1967 struct pgpath *pgpath;
1969 /* pg_init in progress */
1970 if (atomic_read(&m->pg_init_in_progress))
1971 return true;
1973 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1974 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1975 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1977 /* Guess which priority_group will be used at next mapping time */
1978 pg = READ_ONCE(m->current_pg);
1979 next_pg = READ_ONCE(m->next_pg);
1980 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1981 pg = next_pg;
1983 if (!pg) {
1984 /*
1985 * We don't know which pg will be used at next mapping time.
1986 * We don't call choose_pgpath() here to avoid to trigger
1987 * pg_init just by busy checking.
1988 * So we don't know whether underlying devices we will be using
1989 * at next mapping time are busy or not. Just try mapping.
1990 */
1991 return busy;
1992 }
1994 /*
1995 * If there is one non-busy active path at least, the path selector
1996 * will be able to select it. So we consider such a pg as not busy.
1997 */
1998 busy = true;
1999 list_for_each_entry(pgpath, &pg->pgpaths, list) {
2000 if (pgpath->is_active) {
2001 has_active = true;
2002 if (!pgpath_busy(pgpath)) {
2003 busy = false;
2004 break;
2005 }
2006 }
2007 }
2009 if (!has_active) {
2010 /*
2011 * No active path in this pg, so this pg won't be used and
2012 * the current_pg will be changed at next mapping time.
2013 * We need to try mapping to determine it.
2014 */
2015 busy = false;
2016 }
2018 return busy;
2019 }
2021 /*-----------------------------------------------------------------
2022 * Module setup
2023 *---------------------------------------------------------------*/
2024 static struct target_type multipath_target = {
2025 .name = "multipath",
2026 .version = {1, 13, 0},
2027 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2028 DM_TARGET_PASSES_INTEGRITY,
2029 .module = THIS_MODULE,
2030 .ctr = multipath_ctr,
2031 .dtr = multipath_dtr,
2032 .clone_and_map_rq = multipath_clone_and_map,
2033 .release_clone_rq = multipath_release_clone,
2034 .rq_end_io = multipath_end_io,
2035 .map = multipath_map_bio,
2036 .end_io = multipath_end_io_bio,
2037 .presuspend = multipath_presuspend,
2038 .postsuspend = multipath_postsuspend,
2039 .resume = multipath_resume,
2040 .status = multipath_status,
2041 .message = multipath_message,
2042 .prepare_ioctl = multipath_prepare_ioctl,
2043 .iterate_devices = multipath_iterate_devices,
2044 .busy = multipath_busy,
2045 };
2047 static int __init dm_multipath_init(void)
2048 {
2049 int r;
2051 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2052 if (!kmultipathd) {
2053 DMERR("failed to create workqueue kmpathd");
2054 r = -ENOMEM;
2055 goto bad_alloc_kmultipathd;
2056 }
2058 /*
2059 * A separate workqueue is used to handle the device handlers
2060 * to avoid overloading existing workqueue. Overloading the
2061 * old workqueue would also create a bottleneck in the
2062 * path of the storage hardware device activation.
2063 */
2064 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2065 WQ_MEM_RECLAIM);
2066 if (!kmpath_handlerd) {
2067 DMERR("failed to create workqueue kmpath_handlerd");
2068 r = -ENOMEM;
2069 goto bad_alloc_kmpath_handlerd;
2070 }
2072 r = dm_register_target(&multipath_target);
2073 if (r < 0) {
2074 DMERR("request-based register failed %d", r);
2075 r = -EINVAL;
2076 goto bad_register_target;
2077 }
2079 return 0;
2081 bad_register_target:
2082 destroy_workqueue(kmpath_handlerd);
2083 bad_alloc_kmpath_handlerd:
2084 destroy_workqueue(kmultipathd);
2085 bad_alloc_kmultipathd:
2086 return r;
2087 }
2089 static void __exit dm_multipath_exit(void)
2090 {
2091 destroy_workqueue(kmpath_handlerd);
2092 destroy_workqueue(kmultipathd);
2094 dm_unregister_target(&multipath_target);
2095 }
2097 module_init(dm_multipath_init);
2098 module_exit(dm_multipath_exit);
2100 MODULE_DESCRIPTION(DM_NAME " multipath target");
2101 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2102 MODULE_LICENSE("GPL");