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 */
7
8#include <linux/device-mapper.h>
9
10#include "dm-rq.h"
11#include "dm-bio-record.h"
12#include "dm-path-selector.h"
13#include "dm-uevent.h"
14
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>
28
29#define DM_MSG_PREFIX "multipath"
30#define DM_PG_INIT_DELAY_MSECS 2000
31#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
32
33/* Path properties */
34struct pgpath {
35 struct list_head list;
36
37 struct priority_group *pg; /* Owning PG */
38 unsigned fail_count; /* Cumulative failure count */
39
40 struct dm_path path;
41 struct delayed_work activate_path;
42
43 bool is_active:1; /* Path status */
44};
45
46#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
47
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 */
52struct priority_group {
53 struct list_head list;
54
55 struct multipath *m; /* Owning multipath instance */
56 struct path_selector ps;
57
58 unsigned pg_num; /* Reference number */
59 unsigned nr_pgpaths; /* Number of paths in PG */
60 struct list_head pgpaths;
61
62 bool bypassed:1; /* Temporarily bypass this PG? */
63};
64
65/* Multipath context */
66struct multipath {
67 unsigned long flags; /* Multipath state flags */
68
69 spinlock_t lock;
70 enum dm_queue_mode queue_mode;
71
72 struct pgpath *current_pgpath;
73 struct priority_group *current_pg;
74 struct priority_group *next_pg; /* Switch to this PG if set */
75
76 atomic_t nr_valid_paths; /* Total number of usable paths */
77 unsigned nr_priority_groups;
78 struct list_head priority_groups;
79
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 */
87
88 struct mutex work_mutex;
89 struct work_struct trigger_event;
90 struct dm_target *ti;
91
92 struct work_struct process_queued_bios;
93 struct bio_list queued_bios;
94};
95
96/*
97 * Context information attached to each io we process.
98 */
99struct dm_mpath_io {
100 struct pgpath *pgpath;
101 size_t nr_bytes;
102};
103
104typedef int (*action_fn) (struct pgpath *pgpath);
105
106static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
107static void trigger_event(struct work_struct *work);
108static void activate_or_offline_path(struct pgpath *pgpath);
109static void activate_path_work(struct work_struct *work);
110static void process_queued_bios(struct work_struct *work);
111
112/*-----------------------------------------------
113 * Multipath state flags.
114 *-----------------------------------------------*/
115
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? */
123
124/*-----------------------------------------------
125 * Allocation routines
126 *-----------------------------------------------*/
127
128static struct pgpath *alloc_pgpath(void)
129{
130 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
131
132 if (!pgpath)
133 return NULL;
134
135 pgpath->is_active = true;
136
137 return pgpath;
138}
139
140static void free_pgpath(struct pgpath *pgpath)
141{
142 kfree(pgpath);
143}
144
145static struct priority_group *alloc_priority_group(void)
146{
147 struct priority_group *pg;
148
149 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
150
151 if (pg)
152 INIT_LIST_HEAD(&pg->pgpaths);
153
154 return pg;
155}
156
157static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
158{
159 struct pgpath *pgpath, *tmp;
160
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}
167
168static void free_priority_group(struct priority_group *pg,
169 struct dm_target *ti)
170{
171 struct path_selector *ps = &pg->ps;
172
173 if (ps->type) {
174 ps->type->destroy(ps);
175 dm_put_path_selector(ps->type);
176 }
177
178 free_pgpaths(&pg->pgpaths, ti);
179 kfree(pg);
180}
181
182static struct multipath *alloc_multipath(struct dm_target *ti)
183{
184 struct multipath *m;
185
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);
193
194 m->queue_mode = DM_TYPE_NONE;
195
196 m->ti = ti;
197 ti->private = m;
198 }
199
200 return m;
201}
202
203static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
204{
205 if (m->queue_mode == DM_TYPE_NONE) {
206 m->queue_mode = DM_TYPE_REQUEST_BASED;
207 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
208 INIT_WORK(&m->process_queued_bios, process_queued_bios);
209 /*
210 * bio-based doesn't support any direct scsi_dh management;
211 * it just discovers if a scsi_dh is attached.
212 */
213 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
214 }
215
216 dm_table_set_type(ti->table, m->queue_mode);
217
218 /*
219 * Init fields that are only used when a scsi_dh is attached
220 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
221 */
222 set_bit(MPATHF_QUEUE_IO, &m->flags);
223 atomic_set(&m->pg_init_in_progress, 0);
224 atomic_set(&m->pg_init_count, 0);
225 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
226 init_waitqueue_head(&m->pg_init_wait);
227
228 return 0;
229}
230
231static void free_multipath(struct multipath *m)
232{
233 struct priority_group *pg, *tmp;
234
235 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
236 list_del(&pg->list);
237 free_priority_group(pg, m->ti);
238 }
239
240 kfree(m->hw_handler_name);
241 kfree(m->hw_handler_params);
242 mutex_destroy(&m->work_mutex);
243 kfree(m);
244}
245
246static struct dm_mpath_io *get_mpio(union map_info *info)
247{
248 return info->ptr;
249}
250
251static size_t multipath_per_bio_data_size(void)
252{
253 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
254}
255
256static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
257{
258 return dm_per_bio_data(bio, multipath_per_bio_data_size());
259}
260
261static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
262{
263 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
264 void *bio_details = mpio + 1;
265 return bio_details;
266}
267
268static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
269{
270 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
271 struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
272
273 mpio->nr_bytes = bio->bi_iter.bi_size;
274 mpio->pgpath = NULL;
275 *mpio_p = mpio;
276
277 dm_bio_record(bio_details, bio);
278}
279
280/*-----------------------------------------------
281 * Path selection
282 *-----------------------------------------------*/
283
284static int __pg_init_all_paths(struct multipath *m)
285{
286 struct pgpath *pgpath;
287 unsigned long pg_init_delay = 0;
288
289 lockdep_assert_held(&m->lock);
290
291 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
292 return 0;
293
294 atomic_inc(&m->pg_init_count);
295 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
296
297 /* Check here to reset pg_init_required */
298 if (!m->current_pg)
299 return 0;
300
301 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
302 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
303 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
304 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
305 /* Skip failed paths */
306 if (!pgpath->is_active)
307 continue;
308 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
309 pg_init_delay))
310 atomic_inc(&m->pg_init_in_progress);
311 }
312 return atomic_read(&m->pg_init_in_progress);
313}
314
315static int pg_init_all_paths(struct multipath *m)
316{
317 int ret;
318 unsigned long flags;
319
320 spin_lock_irqsave(&m->lock, flags);
321 ret = __pg_init_all_paths(m);
322 spin_unlock_irqrestore(&m->lock, flags);
323
324 return ret;
325}
326
327static void __switch_pg(struct multipath *m, struct priority_group *pg)
328{
329 m->current_pg = pg;
330
331 /* Must we initialise the PG first, and queue I/O till it's ready? */
332 if (m->hw_handler_name) {
333 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
334 set_bit(MPATHF_QUEUE_IO, &m->flags);
335 } else {
336 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
337 clear_bit(MPATHF_QUEUE_IO, &m->flags);
338 }
339
340 atomic_set(&m->pg_init_count, 0);
341}
342
343static struct pgpath *choose_path_in_pg(struct multipath *m,
344 struct priority_group *pg,
345 size_t nr_bytes)
346{
347 unsigned long flags;
348 struct dm_path *path;
349 struct pgpath *pgpath;
350
351 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
352 if (!path)
353 return ERR_PTR(-ENXIO);
354
355 pgpath = path_to_pgpath(path);
356
357 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
358 /* Only update current_pgpath if pg changed */
359 spin_lock_irqsave(&m->lock, flags);
360 m->current_pgpath = pgpath;
361 __switch_pg(m, pg);
362 spin_unlock_irqrestore(&m->lock, flags);
363 }
364
365 return pgpath;
366}
367
368static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
369{
370 unsigned long flags;
371 struct priority_group *pg;
372 struct pgpath *pgpath;
373 unsigned bypassed = 1;
374
375 if (!atomic_read(&m->nr_valid_paths)) {
376 clear_bit(MPATHF_QUEUE_IO, &m->flags);
377 goto failed;
378 }
379
380 /* Were we instructed to switch PG? */
381 if (READ_ONCE(m->next_pg)) {
382 spin_lock_irqsave(&m->lock, flags);
383 pg = m->next_pg;
384 if (!pg) {
385 spin_unlock_irqrestore(&m->lock, flags);
386 goto check_current_pg;
387 }
388 m->next_pg = NULL;
389 spin_unlock_irqrestore(&m->lock, flags);
390 pgpath = choose_path_in_pg(m, pg, nr_bytes);
391 if (!IS_ERR_OR_NULL(pgpath))
392 return pgpath;
393 }
394
395 /* Don't change PG until it has no remaining paths */
396check_current_pg:
397 pg = READ_ONCE(m->current_pg);
398 if (pg) {
399 pgpath = choose_path_in_pg(m, pg, nr_bytes);
400 if (!IS_ERR_OR_NULL(pgpath))
401 return pgpath;
402 }
403
404 /*
405 * Loop through priority groups until we find a valid path.
406 * First time we skip PGs marked 'bypassed'.
407 * Second time we only try the ones we skipped, but set
408 * pg_init_delay_retry so we do not hammer controllers.
409 */
410 do {
411 list_for_each_entry(pg, &m->priority_groups, list) {
412 if (pg->bypassed == !!bypassed)
413 continue;
414 pgpath = choose_path_in_pg(m, pg, nr_bytes);
415 if (!IS_ERR_OR_NULL(pgpath)) {
416 if (!bypassed)
417 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
418 return pgpath;
419 }
420 }
421 } while (bypassed--);
422
423failed:
424 spin_lock_irqsave(&m->lock, flags);
425 m->current_pgpath = NULL;
426 m->current_pg = NULL;
427 spin_unlock_irqrestore(&m->lock, flags);
428
429 return NULL;
430}
431
432/*
433 * dm_report_EIO() is a macro instead of a function to make pr_debug()
434 * report the function name and line number of the function from which
435 * it has been invoked.
436 */
437#define dm_report_EIO(m) \
438do { \
439 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
440 \
441 pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
442 dm_device_name(md), \
443 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
444 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
445 dm_noflush_suspending((m)->ti)); \
446} while (0)
447
448/*
449 * Check whether bios must be queued in the device-mapper core rather
450 * than here in the target.
451 *
452 * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
453 * the same value then we are not between multipath_presuspend()
454 * and multipath_resume() calls and we have no need to check
455 * for the DMF_NOFLUSH_SUSPENDING flag.
456 */
457static bool __must_push_back(struct multipath *m, unsigned long flags)
458{
459 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
460 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
461 dm_noflush_suspending(m->ti));
462}
463
464/*
465 * Following functions use READ_ONCE to get atomic access to
466 * all m->flags to avoid taking spinlock
467 */
468static bool must_push_back_rq(struct multipath *m)
469{
470 unsigned long flags = READ_ONCE(m->flags);
471 return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
472}
473
474static bool must_push_back_bio(struct multipath *m)
475{
476 unsigned long flags = READ_ONCE(m->flags);
477 return __must_push_back(m, flags);
478}
479
480/*
481 * Map cloned requests (request-based multipath)
482 */
483static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
484 union map_info *map_context,
485 struct request **__clone)
486{
487 struct multipath *m = ti->private;
488 size_t nr_bytes = blk_rq_bytes(rq);
489 struct pgpath *pgpath;
490 struct block_device *bdev;
491 struct dm_mpath_io *mpio = get_mpio(map_context);
492 struct request_queue *q;
493 struct request *clone;
494
495 /* Do we need to select a new pgpath? */
496 pgpath = READ_ONCE(m->current_pgpath);
497 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
498 pgpath = choose_pgpath(m, nr_bytes);
499
500 if (!pgpath) {
501 if (must_push_back_rq(m))
502 return DM_MAPIO_DELAY_REQUEUE;
503 dm_report_EIO(m); /* Failed */
504 return DM_MAPIO_KILL;
505 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
506 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
507 pg_init_all_paths(m);
508 return DM_MAPIO_DELAY_REQUEUE;
509 }
510
511 mpio->pgpath = pgpath;
512 mpio->nr_bytes = nr_bytes;
513
514 bdev = pgpath->path.dev->bdev;
515 q = bdev_get_queue(bdev);
516 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE,
517 BLK_MQ_REQ_NOWAIT);
518 if (IS_ERR(clone)) {
519 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
520 if (blk_queue_dying(q)) {
521 atomic_inc(&m->pg_init_in_progress);
522 activate_or_offline_path(pgpath);
523 return DM_MAPIO_DELAY_REQUEUE;
524 }
525
526 /*
527 * blk-mq's SCHED_RESTART can cover this requeue, so we
528 * needn't deal with it by DELAY_REQUEUE. More importantly,
529 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
530 * get the queue busy feedback (via BLK_STS_RESOURCE),
531 * otherwise I/O merging can suffer.
532 */
533 return DM_MAPIO_REQUEUE;
534 }
535 clone->bio = clone->biotail = NULL;
536 clone->rq_disk = bdev->bd_disk;
537 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
538 *__clone = clone;
539
540 if (pgpath->pg->ps.type->start_io)
541 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
542 &pgpath->path,
543 nr_bytes);
544 return DM_MAPIO_REMAPPED;
545}
546
547static void multipath_release_clone(struct request *clone)
548{
549 blk_put_request(clone);
550}
551
552/*
553 * Map cloned bios (bio-based multipath)
554 */
555
556static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
557{
558 struct pgpath *pgpath;
559 unsigned long flags;
560 bool queue_io;
561
562 /* Do we need to select a new pgpath? */
563 pgpath = READ_ONCE(m->current_pgpath);
564 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
565 if (!pgpath || !queue_io)
566 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
567
568 if ((pgpath && queue_io) ||
569 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
570 /* Queue for the daemon to resubmit */
571 spin_lock_irqsave(&m->lock, flags);
572 bio_list_add(&m->queued_bios, bio);
573 spin_unlock_irqrestore(&m->lock, flags);
574
575 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
576 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
577 pg_init_all_paths(m);
578 else if (!queue_io)
579 queue_work(kmultipathd, &m->process_queued_bios);
580
581 return ERR_PTR(-EAGAIN);
582 }
583
584 return pgpath;
585}
586
587static struct pgpath *__map_bio_fast(struct multipath *m, struct bio *bio)
588{
589 struct pgpath *pgpath;
590 unsigned long flags;
591
592 /* Do we need to select a new pgpath? */
593 /*
594 * FIXME: currently only switching path if no path (due to failure, etc)
595 * - which negates the point of using a path selector
596 */
597 pgpath = READ_ONCE(m->current_pgpath);
598 if (!pgpath)
599 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
600
601 if (!pgpath) {
602 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
603 /* Queue for the daemon to resubmit */
604 spin_lock_irqsave(&m->lock, flags);
605 bio_list_add(&m->queued_bios, bio);
606 spin_unlock_irqrestore(&m->lock, flags);
607 queue_work(kmultipathd, &m->process_queued_bios);
608
609 return ERR_PTR(-EAGAIN);
610 }
611 return NULL;
612 }
613
614 return pgpath;
615}
616
617static int __multipath_map_bio(struct multipath *m, struct bio *bio,
618 struct dm_mpath_io *mpio)
619{
620 struct pgpath *pgpath;
621
622 if (!m->hw_handler_name)
623 pgpath = __map_bio_fast(m, bio);
624 else
625 pgpath = __map_bio(m, bio);
626
627 if (IS_ERR(pgpath))
628 return DM_MAPIO_SUBMITTED;
629
630 if (!pgpath) {
631 if (must_push_back_bio(m))
632 return DM_MAPIO_REQUEUE;
633 dm_report_EIO(m);
634 return DM_MAPIO_KILL;
635 }
636
637 mpio->pgpath = pgpath;
638
639 bio->bi_status = 0;
640 bio_set_dev(bio, pgpath->path.dev->bdev);
641 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
642
643 if (pgpath->pg->ps.type->start_io)
644 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
645 &pgpath->path,
646 mpio->nr_bytes);
647 return DM_MAPIO_REMAPPED;
648}
649
650static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
651{
652 struct multipath *m = ti->private;
653 struct dm_mpath_io *mpio = NULL;
654
655 multipath_init_per_bio_data(bio, &mpio);
656 return __multipath_map_bio(m, bio, mpio);
657}
658
659static void process_queued_io_list(struct multipath *m)
660{
661 if (m->queue_mode == DM_TYPE_REQUEST_BASED)
662 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
663 else if (m->queue_mode == DM_TYPE_BIO_BASED)
664 queue_work(kmultipathd, &m->process_queued_bios);
665}
666
667static void process_queued_bios(struct work_struct *work)
668{
669 int r;
670 unsigned long flags;
671 struct bio *bio;
672 struct bio_list bios;
673 struct blk_plug plug;
674 struct multipath *m =
675 container_of(work, struct multipath, process_queued_bios);
676
677 bio_list_init(&bios);
678
679 spin_lock_irqsave(&m->lock, flags);
680
681 if (bio_list_empty(&m->queued_bios)) {
682 spin_unlock_irqrestore(&m->lock, flags);
683 return;
684 }
685
686 bio_list_merge(&bios, &m->queued_bios);
687 bio_list_init(&m->queued_bios);
688
689 spin_unlock_irqrestore(&m->lock, flags);
690
691 blk_start_plug(&plug);
692 while ((bio = bio_list_pop(&bios))) {
693 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
694 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
695 r = __multipath_map_bio(m, bio, mpio);
696 switch (r) {
697 case DM_MAPIO_KILL:
698 bio->bi_status = BLK_STS_IOERR;
699 bio_endio(bio);
700 break;
701 case DM_MAPIO_REQUEUE:
702 bio->bi_status = BLK_STS_DM_REQUEUE;
703 bio_endio(bio);
704 break;
705 case DM_MAPIO_REMAPPED:
706 generic_make_request(bio);
707 break;
708 case DM_MAPIO_SUBMITTED:
709 break;
710 default:
711 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
712 }
713 }
714 blk_finish_plug(&plug);
715}
716
717/*
718 * If we run out of usable paths, should we queue I/O or error it?
719 */
720static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
721 bool save_old_value)
722{
723 unsigned long flags;
724
725 spin_lock_irqsave(&m->lock, flags);
726 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
727 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
728 (!save_old_value && queue_if_no_path));
729 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
730 spin_unlock_irqrestore(&m->lock, flags);
731
732 if (!queue_if_no_path) {
733 dm_table_run_md_queue_async(m->ti->table);
734 process_queued_io_list(m);
735 }
736
737 return 0;
738}
739
740/*
741 * An event is triggered whenever a path is taken out of use.
742 * Includes path failure and PG bypass.
743 */
744static void trigger_event(struct work_struct *work)
745{
746 struct multipath *m =
747 container_of(work, struct multipath, trigger_event);
748
749 dm_table_event(m->ti->table);
750}
751
752/*-----------------------------------------------------------------
753 * Constructor/argument parsing:
754 * <#multipath feature args> [<arg>]*
755 * <#hw_handler args> [hw_handler [<arg>]*]
756 * <#priority groups>
757 * <initial priority group>
758 * [<selector> <#selector args> [<arg>]*
759 * <#paths> <#per-path selector args>
760 * [<path> [<arg>]* ]+ ]+
761 *---------------------------------------------------------------*/
762static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
763 struct dm_target *ti)
764{
765 int r;
766 struct path_selector_type *pst;
767 unsigned ps_argc;
768
769 static const struct dm_arg _args[] = {
770 {0, 1024, "invalid number of path selector args"},
771 };
772
773 pst = dm_get_path_selector(dm_shift_arg(as));
774 if (!pst) {
775 ti->error = "unknown path selector type";
776 return -EINVAL;
777 }
778
779 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
780 if (r) {
781 dm_put_path_selector(pst);
782 return -EINVAL;
783 }
784
785 r = pst->create(&pg->ps, ps_argc, as->argv);
786 if (r) {
787 dm_put_path_selector(pst);
788 ti->error = "path selector constructor failed";
789 return r;
790 }
791
792 pg->ps.type = pst;
793 dm_consume_args(as, ps_argc);
794
795 return 0;
796}
797
798static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
799 const char **attached_handler_name, char **error)
800{
801 struct request_queue *q = bdev_get_queue(bdev);
802 int r;
803
804 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
805retain:
806 if (*attached_handler_name) {
807 /*
808 * Clear any hw_handler_params associated with a
809 * handler that isn't already attached.
810 */
811 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
812 kfree(m->hw_handler_params);
813 m->hw_handler_params = NULL;
814 }
815
816 /*
817 * Reset hw_handler_name to match the attached handler
818 *
819 * NB. This modifies the table line to show the actual
820 * handler instead of the original table passed in.
821 */
822 kfree(m->hw_handler_name);
823 m->hw_handler_name = *attached_handler_name;
824 *attached_handler_name = NULL;
825 }
826 }
827
828 if (m->hw_handler_name) {
829 r = scsi_dh_attach(q, m->hw_handler_name);
830 if (r == -EBUSY) {
831 char b[BDEVNAME_SIZE];
832
833 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
834 bdevname(bdev, b));
835 goto retain;
836 }
837 if (r < 0) {
838 *error = "error attaching hardware handler";
839 return r;
840 }
841
842 if (m->hw_handler_params) {
843 r = scsi_dh_set_params(q, m->hw_handler_params);
844 if (r < 0) {
845 *error = "unable to set hardware handler parameters";
846 return r;
847 }
848 }
849 }
850
851 return 0;
852}
853
854static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
855 struct dm_target *ti)
856{
857 int r;
858 struct pgpath *p;
859 struct multipath *m = ti->private;
860 struct request_queue *q;
861 const char *attached_handler_name = NULL;
862
863 /* we need at least a path arg */
864 if (as->argc < 1) {
865 ti->error = "no device given";
866 return ERR_PTR(-EINVAL);
867 }
868
869 p = alloc_pgpath();
870 if (!p)
871 return ERR_PTR(-ENOMEM);
872
873 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
874 &p->path.dev);
875 if (r) {
876 ti->error = "error getting device";
877 goto bad;
878 }
879
880 q = bdev_get_queue(p->path.dev->bdev);
881 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
882 if (attached_handler_name || m->hw_handler_name) {
883 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
884 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
885 if (r) {
886 dm_put_device(ti, p->path.dev);
887 goto bad;
888 }
889 }
890
891 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
892 if (r) {
893 dm_put_device(ti, p->path.dev);
894 goto bad;
895 }
896
897 return p;
898 bad:
899 kfree(attached_handler_name);
900 free_pgpath(p);
901 return ERR_PTR(r);
902}
903
904static struct priority_group *parse_priority_group(struct dm_arg_set *as,
905 struct multipath *m)
906{
907 static const struct dm_arg _args[] = {
908 {1, 1024, "invalid number of paths"},
909 {0, 1024, "invalid number of selector args"}
910 };
911
912 int r;
913 unsigned i, nr_selector_args, nr_args;
914 struct priority_group *pg;
915 struct dm_target *ti = m->ti;
916
917 if (as->argc < 2) {
918 as->argc = 0;
919 ti->error = "not enough priority group arguments";
920 return ERR_PTR(-EINVAL);
921 }
922
923 pg = alloc_priority_group();
924 if (!pg) {
925 ti->error = "couldn't allocate priority group";
926 return ERR_PTR(-ENOMEM);
927 }
928 pg->m = m;
929
930 r = parse_path_selector(as, pg, ti);
931 if (r)
932 goto bad;
933
934 /*
935 * read the paths
936 */
937 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
938 if (r)
939 goto bad;
940
941 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
942 if (r)
943 goto bad;
944
945 nr_args = 1 + nr_selector_args;
946 for (i = 0; i < pg->nr_pgpaths; i++) {
947 struct pgpath *pgpath;
948 struct dm_arg_set path_args;
949
950 if (as->argc < nr_args) {
951 ti->error = "not enough path parameters";
952 r = -EINVAL;
953 goto bad;
954 }
955
956 path_args.argc = nr_args;
957 path_args.argv = as->argv;
958
959 pgpath = parse_path(&path_args, &pg->ps, ti);
960 if (IS_ERR(pgpath)) {
961 r = PTR_ERR(pgpath);
962 goto bad;
963 }
964
965 pgpath->pg = pg;
966 list_add_tail(&pgpath->list, &pg->pgpaths);
967 dm_consume_args(as, nr_args);
968 }
969
970 return pg;
971
972 bad:
973 free_priority_group(pg, ti);
974 return ERR_PTR(r);
975}
976
977static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
978{
979 unsigned hw_argc;
980 int ret;
981 struct dm_target *ti = m->ti;
982
983 static const struct dm_arg _args[] = {
984 {0, 1024, "invalid number of hardware handler args"},
985 };
986
987 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
988 return -EINVAL;
989
990 if (!hw_argc)
991 return 0;
992
993 if (m->queue_mode == DM_TYPE_BIO_BASED) {
994 dm_consume_args(as, hw_argc);
995 DMERR("bio-based multipath doesn't allow hardware handler args");
996 return 0;
997 }
998
999 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1000 if (!m->hw_handler_name)
1001 return -EINVAL;
1002
1003 if (hw_argc > 1) {
1004 char *p;
1005 int i, j, len = 4;
1006
1007 for (i = 0; i <= hw_argc - 2; i++)
1008 len += strlen(as->argv[i]) + 1;
1009 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1010 if (!p) {
1011 ti->error = "memory allocation failed";
1012 ret = -ENOMEM;
1013 goto fail;
1014 }
1015 j = sprintf(p, "%d", hw_argc - 1);
1016 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1017 j = sprintf(p, "%s", as->argv[i]);
1018 }
1019 dm_consume_args(as, hw_argc - 1);
1020
1021 return 0;
1022fail:
1023 kfree(m->hw_handler_name);
1024 m->hw_handler_name = NULL;
1025 return ret;
1026}
1027
1028static int parse_features(struct dm_arg_set *as, struct multipath *m)
1029{
1030 int r;
1031 unsigned argc;
1032 struct dm_target *ti = m->ti;
1033 const char *arg_name;
1034
1035 static const struct dm_arg _args[] = {
1036 {0, 8, "invalid number of feature args"},
1037 {1, 50, "pg_init_retries must be between 1 and 50"},
1038 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1039 };
1040
1041 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1042 if (r)
1043 return -EINVAL;
1044
1045 if (!argc)
1046 return 0;
1047
1048 do {
1049 arg_name = dm_shift_arg(as);
1050 argc--;
1051
1052 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1053 r = queue_if_no_path(m, true, false);
1054 continue;
1055 }
1056
1057 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1058 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1059 continue;
1060 }
1061
1062 if (!strcasecmp(arg_name, "pg_init_retries") &&
1063 (argc >= 1)) {
1064 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1065 argc--;
1066 continue;
1067 }
1068
1069 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1070 (argc >= 1)) {
1071 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1072 argc--;
1073 continue;
1074 }
1075
1076 if (!strcasecmp(arg_name, "queue_mode") &&
1077 (argc >= 1)) {
1078 const char *queue_mode_name = dm_shift_arg(as);
1079
1080 if (!strcasecmp(queue_mode_name, "bio"))
1081 m->queue_mode = DM_TYPE_BIO_BASED;
1082 else if (!strcasecmp(queue_mode_name, "rq") ||
1083 !strcasecmp(queue_mode_name, "mq"))
1084 m->queue_mode = DM_TYPE_REQUEST_BASED;
1085 else {
1086 ti->error = "Unknown 'queue_mode' requested";
1087 r = -EINVAL;
1088 }
1089 argc--;
1090 continue;
1091 }
1092
1093 ti->error = "Unrecognised multipath feature request";
1094 r = -EINVAL;
1095 } while (argc && !r);
1096
1097 return r;
1098}
1099
1100static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1101{
1102 /* target arguments */
1103 static const struct dm_arg _args[] = {
1104 {0, 1024, "invalid number of priority groups"},
1105 {0, 1024, "invalid initial priority group number"},
1106 };
1107
1108 int r;
1109 struct multipath *m;
1110 struct dm_arg_set as;
1111 unsigned pg_count = 0;
1112 unsigned next_pg_num;
1113
1114 as.argc = argc;
1115 as.argv = argv;
1116
1117 m = alloc_multipath(ti);
1118 if (!m) {
1119 ti->error = "can't allocate multipath";
1120 return -EINVAL;
1121 }
1122
1123 r = parse_features(&as, m);
1124 if (r)
1125 goto bad;
1126
1127 r = alloc_multipath_stage2(ti, m);
1128 if (r)
1129 goto bad;
1130
1131 r = parse_hw_handler(&as, m);
1132 if (r)
1133 goto bad;
1134
1135 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1136 if (r)
1137 goto bad;
1138
1139 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1140 if (r)
1141 goto bad;
1142
1143 if ((!m->nr_priority_groups && next_pg_num) ||
1144 (m->nr_priority_groups && !next_pg_num)) {
1145 ti->error = "invalid initial priority group";
1146 r = -EINVAL;
1147 goto bad;
1148 }
1149
1150 /* parse the priority groups */
1151 while (as.argc) {
1152 struct priority_group *pg;
1153 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1154
1155 pg = parse_priority_group(&as, m);
1156 if (IS_ERR(pg)) {
1157 r = PTR_ERR(pg);
1158 goto bad;
1159 }
1160
1161 nr_valid_paths += pg->nr_pgpaths;
1162 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1163
1164 list_add_tail(&pg->list, &m->priority_groups);
1165 pg_count++;
1166 pg->pg_num = pg_count;
1167 if (!--next_pg_num)
1168 m->next_pg = pg;
1169 }
1170
1171 if (pg_count != m->nr_priority_groups) {
1172 ti->error = "priority group count mismatch";
1173 r = -EINVAL;
1174 goto bad;
1175 }
1176
1177 ti->num_flush_bios = 1;
1178 ti->num_discard_bios = 1;
1179 ti->num_write_same_bios = 1;
1180 ti->num_write_zeroes_bios = 1;
1181 if (m->queue_mode == DM_TYPE_BIO_BASED)
1182 ti->per_io_data_size = multipath_per_bio_data_size();
1183 else
1184 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1185
1186 return 0;
1187
1188 bad:
1189 free_multipath(m);
1190 return r;
1191}
1192
1193static void multipath_wait_for_pg_init_completion(struct multipath *m)
1194{
1195 DEFINE_WAIT(wait);
1196
1197 while (1) {
1198 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1199
1200 if (!atomic_read(&m->pg_init_in_progress))
1201 break;
1202
1203 io_schedule();
1204 }
1205 finish_wait(&m->pg_init_wait, &wait);
1206}
1207
1208static void flush_multipath_work(struct multipath *m)
1209{
1210 if (m->hw_handler_name) {
1211 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1212 smp_mb__after_atomic();
1213
1214 if (atomic_read(&m->pg_init_in_progress))
1215 flush_workqueue(kmpath_handlerd);
1216 multipath_wait_for_pg_init_completion(m);
1217
1218 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1219 smp_mb__after_atomic();
1220 }
1221
1222 if (m->queue_mode == DM_TYPE_BIO_BASED)
1223 flush_work(&m->process_queued_bios);
1224 flush_work(&m->trigger_event);
1225}
1226
1227static void multipath_dtr(struct dm_target *ti)
1228{
1229 struct multipath *m = ti->private;
1230
1231 flush_multipath_work(m);
1232 free_multipath(m);
1233}
1234
1235/*
1236 * Take a path out of use.
1237 */
1238static int fail_path(struct pgpath *pgpath)
1239{
1240 unsigned long flags;
1241 struct multipath *m = pgpath->pg->m;
1242
1243 spin_lock_irqsave(&m->lock, flags);
1244
1245 if (!pgpath->is_active)
1246 goto out;
1247
1248 DMWARN("Failing path %s.", pgpath->path.dev->name);
1249
1250 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1251 pgpath->is_active = false;
1252 pgpath->fail_count++;
1253
1254 atomic_dec(&m->nr_valid_paths);
1255
1256 if (pgpath == m->current_pgpath)
1257 m->current_pgpath = NULL;
1258
1259 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1260 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1261
1262 schedule_work(&m->trigger_event);
1263
1264out:
1265 spin_unlock_irqrestore(&m->lock, flags);
1266
1267 return 0;
1268}
1269
1270/*
1271 * Reinstate a previously-failed path
1272 */
1273static int reinstate_path(struct pgpath *pgpath)
1274{
1275 int r = 0, run_queue = 0;
1276 unsigned long flags;
1277 struct multipath *m = pgpath->pg->m;
1278 unsigned nr_valid_paths;
1279
1280 spin_lock_irqsave(&m->lock, flags);
1281
1282 if (pgpath->is_active)
1283 goto out;
1284
1285 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1286
1287 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1288 if (r)
1289 goto out;
1290
1291 pgpath->is_active = true;
1292
1293 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1294 if (nr_valid_paths == 1) {
1295 m->current_pgpath = NULL;
1296 run_queue = 1;
1297 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1298 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1299 atomic_inc(&m->pg_init_in_progress);
1300 }
1301
1302 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1303 pgpath->path.dev->name, nr_valid_paths);
1304
1305 schedule_work(&m->trigger_event);
1306
1307out:
1308 spin_unlock_irqrestore(&m->lock, flags);
1309 if (run_queue) {
1310 dm_table_run_md_queue_async(m->ti->table);
1311 process_queued_io_list(m);
1312 }
1313
1314 return r;
1315}
1316
1317/*
1318 * Fail or reinstate all paths that match the provided struct dm_dev.
1319 */
1320static int action_dev(struct multipath *m, struct dm_dev *dev,
1321 action_fn action)
1322{
1323 int r = -EINVAL;
1324 struct pgpath *pgpath;
1325 struct priority_group *pg;
1326
1327 list_for_each_entry(pg, &m->priority_groups, list) {
1328 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1329 if (pgpath->path.dev == dev)
1330 r = action(pgpath);
1331 }
1332 }
1333
1334 return r;
1335}
1336
1337/*
1338 * Temporarily try to avoid having to use the specified PG
1339 */
1340static void bypass_pg(struct multipath *m, struct priority_group *pg,
1341 bool bypassed)
1342{
1343 unsigned long flags;
1344
1345 spin_lock_irqsave(&m->lock, flags);
1346
1347 pg->bypassed = bypassed;
1348 m->current_pgpath = NULL;
1349 m->current_pg = NULL;
1350
1351 spin_unlock_irqrestore(&m->lock, flags);
1352
1353 schedule_work(&m->trigger_event);
1354}
1355
1356/*
1357 * Switch to using the specified PG from the next I/O that gets mapped
1358 */
1359static int switch_pg_num(struct multipath *m, const char *pgstr)
1360{
1361 struct priority_group *pg;
1362 unsigned pgnum;
1363 unsigned long flags;
1364 char dummy;
1365
1366 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1367 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1368 DMWARN("invalid PG number supplied to switch_pg_num");
1369 return -EINVAL;
1370 }
1371
1372 spin_lock_irqsave(&m->lock, flags);
1373 list_for_each_entry(pg, &m->priority_groups, list) {
1374 pg->bypassed = false;
1375 if (--pgnum)
1376 continue;
1377
1378 m->current_pgpath = NULL;
1379 m->current_pg = NULL;
1380 m->next_pg = pg;
1381 }
1382 spin_unlock_irqrestore(&m->lock, flags);
1383
1384 schedule_work(&m->trigger_event);
1385 return 0;
1386}
1387
1388/*
1389 * Set/clear bypassed status of a PG.
1390 * PGs are numbered upwards from 1 in the order they were declared.
1391 */
1392static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1393{
1394 struct priority_group *pg;
1395 unsigned pgnum;
1396 char dummy;
1397
1398 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1399 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1400 DMWARN("invalid PG number supplied to bypass_pg");
1401 return -EINVAL;
1402 }
1403
1404 list_for_each_entry(pg, &m->priority_groups, list) {
1405 if (!--pgnum)
1406 break;
1407 }
1408
1409 bypass_pg(m, pg, bypassed);
1410 return 0;
1411}
1412
1413/*
1414 * Should we retry pg_init immediately?
1415 */
1416static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1417{
1418 unsigned long flags;
1419 bool limit_reached = false;
1420
1421 spin_lock_irqsave(&m->lock, flags);
1422
1423 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1424 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1425 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1426 else
1427 limit_reached = true;
1428
1429 spin_unlock_irqrestore(&m->lock, flags);
1430
1431 return limit_reached;
1432}
1433
1434static void pg_init_done(void *data, int errors)
1435{
1436 struct pgpath *pgpath = data;
1437 struct priority_group *pg = pgpath->pg;
1438 struct multipath *m = pg->m;
1439 unsigned long flags;
1440 bool delay_retry = false;
1441
1442 /* device or driver problems */
1443 switch (errors) {
1444 case SCSI_DH_OK:
1445 break;
1446 case SCSI_DH_NOSYS:
1447 if (!m->hw_handler_name) {
1448 errors = 0;
1449 break;
1450 }
1451 DMERR("Could not failover the device: Handler scsi_dh_%s "
1452 "Error %d.", m->hw_handler_name, errors);
1453 /*
1454 * Fail path for now, so we do not ping pong
1455 */
1456 fail_path(pgpath);
1457 break;
1458 case SCSI_DH_DEV_TEMP_BUSY:
1459 /*
1460 * Probably doing something like FW upgrade on the
1461 * controller so try the other pg.
1462 */
1463 bypass_pg(m, pg, true);
1464 break;
1465 case SCSI_DH_RETRY:
1466 /* Wait before retrying. */
1467 delay_retry = 1;
1468 /* fall through */
1469 case SCSI_DH_IMM_RETRY:
1470 case SCSI_DH_RES_TEMP_UNAVAIL:
1471 if (pg_init_limit_reached(m, pgpath))
1472 fail_path(pgpath);
1473 errors = 0;
1474 break;
1475 case SCSI_DH_DEV_OFFLINED:
1476 default:
1477 /*
1478 * We probably do not want to fail the path for a device
1479 * error, but this is what the old dm did. In future
1480 * patches we can do more advanced handling.
1481 */
1482 fail_path(pgpath);
1483 }
1484
1485 spin_lock_irqsave(&m->lock, flags);
1486 if (errors) {
1487 if (pgpath == m->current_pgpath) {
1488 DMERR("Could not failover device. Error %d.", errors);
1489 m->current_pgpath = NULL;
1490 m->current_pg = NULL;
1491 }
1492 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1493 pg->bypassed = false;
1494
1495 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1496 /* Activations of other paths are still on going */
1497 goto out;
1498
1499 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1500 if (delay_retry)
1501 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1502 else
1503 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1504
1505 if (__pg_init_all_paths(m))
1506 goto out;
1507 }
1508 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1509
1510 process_queued_io_list(m);
1511
1512 /*
1513 * Wake up any thread waiting to suspend.
1514 */
1515 wake_up(&m->pg_init_wait);
1516
1517out:
1518 spin_unlock_irqrestore(&m->lock, flags);
1519}
1520
1521static void activate_or_offline_path(struct pgpath *pgpath)
1522{
1523 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1524
1525 if (pgpath->is_active && !blk_queue_dying(q))
1526 scsi_dh_activate(q, pg_init_done, pgpath);
1527 else
1528 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1529}
1530
1531static void activate_path_work(struct work_struct *work)
1532{
1533 struct pgpath *pgpath =
1534 container_of(work, struct pgpath, activate_path.work);
1535
1536 activate_or_offline_path(pgpath);
1537}
1538
1539static int multipath_end_io(struct dm_target *ti, struct request *clone,
1540 blk_status_t error, union map_info *map_context)
1541{
1542 struct dm_mpath_io *mpio = get_mpio(map_context);
1543 struct pgpath *pgpath = mpio->pgpath;
1544 int r = DM_ENDIO_DONE;
1545
1546 /*
1547 * We don't queue any clone request inside the multipath target
1548 * during end I/O handling, since those clone requests don't have
1549 * bio clones. If we queue them inside the multipath target,
1550 * we need to make bio clones, that requires memory allocation.
1551 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1552 * don't have bio clones.)
1553 * Instead of queueing the clone request here, we queue the original
1554 * request into dm core, which will remake a clone request and
1555 * clone bios for it and resubmit it later.
1556 */
1557 if (error && blk_path_error(error)) {
1558 struct multipath *m = ti->private;
1559
1560 if (error == BLK_STS_RESOURCE)
1561 r = DM_ENDIO_DELAY_REQUEUE;
1562 else
1563 r = DM_ENDIO_REQUEUE;
1564
1565 if (pgpath)
1566 fail_path(pgpath);
1567
1568 if (atomic_read(&m->nr_valid_paths) == 0 &&
1569 !must_push_back_rq(m)) {
1570 if (error == BLK_STS_IOERR)
1571 dm_report_EIO(m);
1572 /* complete with the original error */
1573 r = DM_ENDIO_DONE;
1574 }
1575 }
1576
1577 if (pgpath) {
1578 struct path_selector *ps = &pgpath->pg->ps;
1579
1580 if (ps->type->end_io)
1581 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1582 }
1583
1584 return r;
1585}
1586
1587static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1588 blk_status_t *error)
1589{
1590 struct multipath *m = ti->private;
1591 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1592 struct pgpath *pgpath = mpio->pgpath;
1593 unsigned long flags;
1594 int r = DM_ENDIO_DONE;
1595
1596 if (!*error || !blk_path_error(*error))
1597 goto done;
1598
1599 if (pgpath)
1600 fail_path(pgpath);
1601
1602 if (atomic_read(&m->nr_valid_paths) == 0 &&
1603 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1604 if (must_push_back_bio(m)) {
1605 r = DM_ENDIO_REQUEUE;
1606 } else {
1607 dm_report_EIO(m);
1608 *error = BLK_STS_IOERR;
1609 }
1610 goto done;
1611 }
1612
1613 spin_lock_irqsave(&m->lock, flags);
1614 bio_list_add(&m->queued_bios, clone);
1615 spin_unlock_irqrestore(&m->lock, flags);
1616 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1617 queue_work(kmultipathd, &m->process_queued_bios);
1618
1619 r = DM_ENDIO_INCOMPLETE;
1620done:
1621 if (pgpath) {
1622 struct path_selector *ps = &pgpath->pg->ps;
1623
1624 if (ps->type->end_io)
1625 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1626 }
1627
1628 return r;
1629}
1630
1631/*
1632 * Suspend can't complete until all the I/O is processed so if
1633 * the last path fails we must error any remaining I/O.
1634 * Note that if the freeze_bdev fails while suspending, the
1635 * queue_if_no_path state is lost - userspace should reset it.
1636 */
1637static void multipath_presuspend(struct dm_target *ti)
1638{
1639 struct multipath *m = ti->private;
1640
1641 queue_if_no_path(m, false, true);
1642}
1643
1644static void multipath_postsuspend(struct dm_target *ti)
1645{
1646 struct multipath *m = ti->private;
1647
1648 mutex_lock(&m->work_mutex);
1649 flush_multipath_work(m);
1650 mutex_unlock(&m->work_mutex);
1651}
1652
1653/*
1654 * Restore the queue_if_no_path setting.
1655 */
1656static void multipath_resume(struct dm_target *ti)
1657{
1658 struct multipath *m = ti->private;
1659 unsigned long flags;
1660
1661 spin_lock_irqsave(&m->lock, flags);
1662 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1663 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1664 spin_unlock_irqrestore(&m->lock, flags);
1665}
1666
1667/*
1668 * Info output has the following format:
1669 * num_multipath_feature_args [multipath_feature_args]*
1670 * num_handler_status_args [handler_status_args]*
1671 * num_groups init_group_number
1672 * [A|D|E num_ps_status_args [ps_status_args]*
1673 * num_paths num_selector_args
1674 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1675 *
1676 * Table output has the following format (identical to the constructor string):
1677 * num_feature_args [features_args]*
1678 * num_handler_args hw_handler [hw_handler_args]*
1679 * num_groups init_group_number
1680 * [priority selector-name num_ps_args [ps_args]*
1681 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1682 */
1683static void multipath_status(struct dm_target *ti, status_type_t type,
1684 unsigned status_flags, char *result, unsigned maxlen)
1685{
1686 int sz = 0;
1687 unsigned long flags;
1688 struct multipath *m = ti->private;
1689 struct priority_group *pg;
1690 struct pgpath *p;
1691 unsigned pg_num;
1692 char state;
1693
1694 spin_lock_irqsave(&m->lock, flags);
1695
1696 /* Features */
1697 if (type == STATUSTYPE_INFO)
1698 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1699 atomic_read(&m->pg_init_count));
1700 else {
1701 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1702 (m->pg_init_retries > 0) * 2 +
1703 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1704 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1705 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1706
1707 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1708 DMEMIT("queue_if_no_path ");
1709 if (m->pg_init_retries)
1710 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1711 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1712 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1713 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1714 DMEMIT("retain_attached_hw_handler ");
1715 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1716 switch(m->queue_mode) {
1717 case DM_TYPE_BIO_BASED:
1718 DMEMIT("queue_mode bio ");
1719 break;
1720 default:
1721 WARN_ON_ONCE(true);
1722 break;
1723 }
1724 }
1725 }
1726
1727 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1728 DMEMIT("0 ");
1729 else
1730 DMEMIT("1 %s ", m->hw_handler_name);
1731
1732 DMEMIT("%u ", m->nr_priority_groups);
1733
1734 if (m->next_pg)
1735 pg_num = m->next_pg->pg_num;
1736 else if (m->current_pg)
1737 pg_num = m->current_pg->pg_num;
1738 else
1739 pg_num = (m->nr_priority_groups ? 1 : 0);
1740
1741 DMEMIT("%u ", pg_num);
1742
1743 switch (type) {
1744 case STATUSTYPE_INFO:
1745 list_for_each_entry(pg, &m->priority_groups, list) {
1746 if (pg->bypassed)
1747 state = 'D'; /* Disabled */
1748 else if (pg == m->current_pg)
1749 state = 'A'; /* Currently Active */
1750 else
1751 state = 'E'; /* Enabled */
1752
1753 DMEMIT("%c ", state);
1754
1755 if (pg->ps.type->status)
1756 sz += pg->ps.type->status(&pg->ps, NULL, type,
1757 result + sz,
1758 maxlen - sz);
1759 else
1760 DMEMIT("0 ");
1761
1762 DMEMIT("%u %u ", pg->nr_pgpaths,
1763 pg->ps.type->info_args);
1764
1765 list_for_each_entry(p, &pg->pgpaths, list) {
1766 DMEMIT("%s %s %u ", p->path.dev->name,
1767 p->is_active ? "A" : "F",
1768 p->fail_count);
1769 if (pg->ps.type->status)
1770 sz += pg->ps.type->status(&pg->ps,
1771 &p->path, type, result + sz,
1772 maxlen - sz);
1773 }
1774 }
1775 break;
1776
1777 case STATUSTYPE_TABLE:
1778 list_for_each_entry(pg, &m->priority_groups, list) {
1779 DMEMIT("%s ", pg->ps.type->name);
1780
1781 if (pg->ps.type->status)
1782 sz += pg->ps.type->status(&pg->ps, NULL, type,
1783 result + sz,
1784 maxlen - sz);
1785 else
1786 DMEMIT("0 ");
1787
1788 DMEMIT("%u %u ", pg->nr_pgpaths,
1789 pg->ps.type->table_args);
1790
1791 list_for_each_entry(p, &pg->pgpaths, list) {
1792 DMEMIT("%s ", p->path.dev->name);
1793 if (pg->ps.type->status)
1794 sz += pg->ps.type->status(&pg->ps,
1795 &p->path, type, result + sz,
1796 maxlen - sz);
1797 }
1798 }
1799 break;
1800 }
1801
1802 spin_unlock_irqrestore(&m->lock, flags);
1803}
1804
1805static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1806 char *result, unsigned maxlen)
1807{
1808 int r = -EINVAL;
1809 struct dm_dev *dev;
1810 struct multipath *m = ti->private;
1811 action_fn action;
1812
1813 mutex_lock(&m->work_mutex);
1814
1815 if (dm_suspended(ti)) {
1816 r = -EBUSY;
1817 goto out;
1818 }
1819
1820 if (argc == 1) {
1821 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1822 r = queue_if_no_path(m, true, false);
1823 goto out;
1824 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1825 r = queue_if_no_path(m, false, false);
1826 goto out;
1827 }
1828 }
1829
1830 if (argc != 2) {
1831 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1832 goto out;
1833 }
1834
1835 if (!strcasecmp(argv[0], "disable_group")) {
1836 r = bypass_pg_num(m, argv[1], true);
1837 goto out;
1838 } else if (!strcasecmp(argv[0], "enable_group")) {
1839 r = bypass_pg_num(m, argv[1], false);
1840 goto out;
1841 } else if (!strcasecmp(argv[0], "switch_group")) {
1842 r = switch_pg_num(m, argv[1]);
1843 goto out;
1844 } else if (!strcasecmp(argv[0], "reinstate_path"))
1845 action = reinstate_path;
1846 else if (!strcasecmp(argv[0], "fail_path"))
1847 action = fail_path;
1848 else {
1849 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1850 goto out;
1851 }
1852
1853 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1854 if (r) {
1855 DMWARN("message: error getting device %s",
1856 argv[1]);
1857 goto out;
1858 }
1859
1860 r = action_dev(m, dev, action);
1861
1862 dm_put_device(ti, dev);
1863
1864out:
1865 mutex_unlock(&m->work_mutex);
1866 return r;
1867}
1868
1869static int multipath_prepare_ioctl(struct dm_target *ti,
1870 struct block_device **bdev)
1871{
1872 struct multipath *m = ti->private;
1873 struct pgpath *current_pgpath;
1874 int r;
1875
1876 current_pgpath = READ_ONCE(m->current_pgpath);
1877 if (!current_pgpath)
1878 current_pgpath = choose_pgpath(m, 0);
1879
1880 if (current_pgpath) {
1881 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1882 *bdev = current_pgpath->path.dev->bdev;
1883 r = 0;
1884 } else {
1885 /* pg_init has not started or completed */
1886 r = -ENOTCONN;
1887 }
1888 } else {
1889 /* No path is available */
1890 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1891 r = -ENOTCONN;
1892 else
1893 r = -EIO;
1894 }
1895
1896 if (r == -ENOTCONN) {
1897 if (!READ_ONCE(m->current_pg)) {
1898 /* Path status changed, redo selection */
1899 (void) choose_pgpath(m, 0);
1900 }
1901 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1902 pg_init_all_paths(m);
1903 dm_table_run_md_queue_async(m->ti->table);
1904 process_queued_io_list(m);
1905 }
1906
1907 /*
1908 * Only pass ioctls through if the device sizes match exactly.
1909 */
1910 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1911 return 1;
1912 return r;
1913}
1914
1915static int multipath_iterate_devices(struct dm_target *ti,
1916 iterate_devices_callout_fn fn, void *data)
1917{
1918 struct multipath *m = ti->private;
1919 struct priority_group *pg;
1920 struct pgpath *p;
1921 int ret = 0;
1922
1923 list_for_each_entry(pg, &m->priority_groups, list) {
1924 list_for_each_entry(p, &pg->pgpaths, list) {
1925 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1926 if (ret)
1927 goto out;
1928 }
1929 }
1930
1931out:
1932 return ret;
1933}
1934
1935static int pgpath_busy(struct pgpath *pgpath)
1936{
1937 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1938
1939 return blk_lld_busy(q);
1940}
1941
1942/*
1943 * We return "busy", only when we can map I/Os but underlying devices
1944 * are busy (so even if we map I/Os now, the I/Os will wait on
1945 * the underlying queue).
1946 * In other words, if we want to kill I/Os or queue them inside us
1947 * due to map unavailability, we don't return "busy". Otherwise,
1948 * dm core won't give us the I/Os and we can't do what we want.
1949 */
1950static int multipath_busy(struct dm_target *ti)
1951{
1952 bool busy = false, has_active = false;
1953 struct multipath *m = ti->private;
1954 struct priority_group *pg, *next_pg;
1955 struct pgpath *pgpath;
1956
1957 /* pg_init in progress */
1958 if (atomic_read(&m->pg_init_in_progress))
1959 return true;
1960
1961 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1962 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1963 return (m->queue_mode != DM_TYPE_REQUEST_BASED);
1964
1965 /* Guess which priority_group will be used at next mapping time */
1966 pg = READ_ONCE(m->current_pg);
1967 next_pg = READ_ONCE(m->next_pg);
1968 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1969 pg = next_pg;
1970
1971 if (!pg) {
1972 /*
1973 * We don't know which pg will be used at next mapping time.
1974 * We don't call choose_pgpath() here to avoid to trigger
1975 * pg_init just by busy checking.
1976 * So we don't know whether underlying devices we will be using
1977 * at next mapping time are busy or not. Just try mapping.
1978 */
1979 return busy;
1980 }
1981
1982 /*
1983 * If there is one non-busy active path at least, the path selector
1984 * will be able to select it. So we consider such a pg as not busy.
1985 */
1986 busy = true;
1987 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1988 if (pgpath->is_active) {
1989 has_active = true;
1990 if (!pgpath_busy(pgpath)) {
1991 busy = false;
1992 break;
1993 }
1994 }
1995 }
1996
1997 if (!has_active) {
1998 /*
1999 * No active path in this pg, so this pg won't be used and
2000 * the current_pg will be changed at next mapping time.
2001 * We need to try mapping to determine it.
2002 */
2003 busy = false;
2004 }
2005
2006 return busy;
2007}
2008
2009/*-----------------------------------------------------------------
2010 * Module setup
2011 *---------------------------------------------------------------*/
2012static struct target_type multipath_target = {
2013 .name = "multipath",
2014 .version = {1, 13, 0},
2015 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2016 DM_TARGET_PASSES_INTEGRITY,
2017 .module = THIS_MODULE,
2018 .ctr = multipath_ctr,
2019 .dtr = multipath_dtr,
2020 .clone_and_map_rq = multipath_clone_and_map,
2021 .release_clone_rq = multipath_release_clone,
2022 .rq_end_io = multipath_end_io,
2023 .map = multipath_map_bio,
2024 .end_io = multipath_end_io_bio,
2025 .presuspend = multipath_presuspend,
2026 .postsuspend = multipath_postsuspend,
2027 .resume = multipath_resume,
2028 .status = multipath_status,
2029 .message = multipath_message,
2030 .prepare_ioctl = multipath_prepare_ioctl,
2031 .iterate_devices = multipath_iterate_devices,
2032 .busy = multipath_busy,
2033};
2034
2035static int __init dm_multipath_init(void)
2036{
2037 int r;
2038
2039 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2040 if (!kmultipathd) {
2041 DMERR("failed to create workqueue kmpathd");
2042 r = -ENOMEM;
2043 goto bad_alloc_kmultipathd;
2044 }
2045
2046 /*
2047 * A separate workqueue is used to handle the device handlers
2048 * to avoid overloading existing workqueue. Overloading the
2049 * old workqueue would also create a bottleneck in the
2050 * path of the storage hardware device activation.
2051 */
2052 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2053 WQ_MEM_RECLAIM);
2054 if (!kmpath_handlerd) {
2055 DMERR("failed to create workqueue kmpath_handlerd");
2056 r = -ENOMEM;
2057 goto bad_alloc_kmpath_handlerd;
2058 }
2059
2060 r = dm_register_target(&multipath_target);
2061 if (r < 0) {
2062 DMERR("request-based register failed %d", r);
2063 r = -EINVAL;
2064 goto bad_register_target;
2065 }
2066
2067 return 0;
2068
2069bad_register_target:
2070 destroy_workqueue(kmpath_handlerd);
2071bad_alloc_kmpath_handlerd:
2072 destroy_workqueue(kmultipathd);
2073bad_alloc_kmultipathd:
2074 return r;
2075}
2076
2077static void __exit dm_multipath_exit(void)
2078{
2079 destroy_workqueue(kmpath_handlerd);
2080 destroy_workqueue(kmultipathd);
2081
2082 dm_unregister_target(&multipath_target);
2083}
2084
2085module_init(dm_multipath_init);
2086module_exit(dm_multipath_exit);
2087
2088MODULE_DESCRIPTION(DM_NAME " multipath target");
2089MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2090MODULE_LICENSE("GPL");
2091