Loading...
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-path-selector.h"
11#include "dm-uevent.h"
12
13#include <linux/ctype.h>
14#include <linux/init.h>
15#include <linux/mempool.h>
16#include <linux/module.h>
17#include <linux/pagemap.h>
18#include <linux/slab.h>
19#include <linux/time.h>
20#include <linux/workqueue.h>
21#include <linux/delay.h>
22#include <scsi/scsi_dh.h>
23#include <linux/atomic.h>
24
25#define DM_MSG_PREFIX "multipath"
26#define DM_PG_INIT_DELAY_MSECS 2000
27#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
28
29/* Path properties */
30struct pgpath {
31 struct list_head list;
32
33 struct priority_group *pg; /* Owning PG */
34 unsigned is_active; /* Path status */
35 unsigned fail_count; /* Cumulative failure count */
36
37 struct dm_path path;
38 struct delayed_work activate_path;
39};
40
41#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
42
43/*
44 * Paths are grouped into Priority Groups and numbered from 1 upwards.
45 * Each has a path selector which controls which path gets used.
46 */
47struct priority_group {
48 struct list_head list;
49
50 struct multipath *m; /* Owning multipath instance */
51 struct path_selector ps;
52
53 unsigned pg_num; /* Reference number */
54 unsigned bypassed; /* Temporarily bypass this PG? */
55
56 unsigned nr_pgpaths; /* Number of paths in PG */
57 struct list_head pgpaths;
58};
59
60/* Multipath context */
61struct multipath {
62 struct list_head list;
63 struct dm_target *ti;
64
65 const char *hw_handler_name;
66 char *hw_handler_params;
67
68 spinlock_t lock;
69
70 unsigned nr_priority_groups;
71 struct list_head priority_groups;
72
73 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
74
75 unsigned pg_init_required; /* pg_init needs calling? */
76 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
77 unsigned pg_init_delay_retry; /* Delay pg_init retry? */
78
79 unsigned nr_valid_paths; /* Total number of usable paths */
80 struct pgpath *current_pgpath;
81 struct priority_group *current_pg;
82 struct priority_group *next_pg; /* Switch to this PG if set */
83 unsigned repeat_count; /* I/Os left before calling PS again */
84
85 unsigned queue_io:1; /* Must we queue all I/O? */
86 unsigned queue_if_no_path:1; /* Queue I/O if last path fails? */
87 unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
88
89 unsigned pg_init_retries; /* Number of times to retry pg_init */
90 unsigned pg_init_count; /* Number of times pg_init called */
91 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
92
93 unsigned queue_size;
94 struct work_struct process_queued_ios;
95 struct list_head queued_ios;
96
97 struct work_struct trigger_event;
98
99 /*
100 * We must use a mempool of dm_mpath_io structs so that we
101 * can resubmit bios on error.
102 */
103 mempool_t *mpio_pool;
104
105 struct mutex work_mutex;
106};
107
108/*
109 * Context information attached to each bio we process.
110 */
111struct dm_mpath_io {
112 struct pgpath *pgpath;
113 size_t nr_bytes;
114};
115
116typedef int (*action_fn) (struct pgpath *pgpath);
117
118#define MIN_IOS 256 /* Mempool size */
119
120static struct kmem_cache *_mpio_cache;
121
122static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
123static void process_queued_ios(struct work_struct *work);
124static void trigger_event(struct work_struct *work);
125static void activate_path(struct work_struct *work);
126
127
128/*-----------------------------------------------
129 * Allocation routines
130 *-----------------------------------------------*/
131
132static struct pgpath *alloc_pgpath(void)
133{
134 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
135
136 if (pgpath) {
137 pgpath->is_active = 1;
138 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
139 }
140
141 return pgpath;
142}
143
144static void free_pgpath(struct pgpath *pgpath)
145{
146 kfree(pgpath);
147}
148
149static struct priority_group *alloc_priority_group(void)
150{
151 struct priority_group *pg;
152
153 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
154
155 if (pg)
156 INIT_LIST_HEAD(&pg->pgpaths);
157
158 return pg;
159}
160
161static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
162{
163 struct pgpath *pgpath, *tmp;
164 struct multipath *m = ti->private;
165
166 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
167 list_del(&pgpath->list);
168 if (m->hw_handler_name)
169 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
170 dm_put_device(ti, pgpath->path.dev);
171 free_pgpath(pgpath);
172 }
173}
174
175static void free_priority_group(struct priority_group *pg,
176 struct dm_target *ti)
177{
178 struct path_selector *ps = &pg->ps;
179
180 if (ps->type) {
181 ps->type->destroy(ps);
182 dm_put_path_selector(ps->type);
183 }
184
185 free_pgpaths(&pg->pgpaths, ti);
186 kfree(pg);
187}
188
189static struct multipath *alloc_multipath(struct dm_target *ti)
190{
191 struct multipath *m;
192
193 m = kzalloc(sizeof(*m), GFP_KERNEL);
194 if (m) {
195 INIT_LIST_HEAD(&m->priority_groups);
196 INIT_LIST_HEAD(&m->queued_ios);
197 spin_lock_init(&m->lock);
198 m->queue_io = 1;
199 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
200 INIT_WORK(&m->process_queued_ios, process_queued_ios);
201 INIT_WORK(&m->trigger_event, trigger_event);
202 init_waitqueue_head(&m->pg_init_wait);
203 mutex_init(&m->work_mutex);
204 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
205 if (!m->mpio_pool) {
206 kfree(m);
207 return NULL;
208 }
209 m->ti = ti;
210 ti->private = m;
211 }
212
213 return m;
214}
215
216static void free_multipath(struct multipath *m)
217{
218 struct priority_group *pg, *tmp;
219
220 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
221 list_del(&pg->list);
222 free_priority_group(pg, m->ti);
223 }
224
225 kfree(m->hw_handler_name);
226 kfree(m->hw_handler_params);
227 mempool_destroy(m->mpio_pool);
228 kfree(m);
229}
230
231static int set_mapinfo(struct multipath *m, union map_info *info)
232{
233 struct dm_mpath_io *mpio;
234
235 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
236 if (!mpio)
237 return -ENOMEM;
238
239 memset(mpio, 0, sizeof(*mpio));
240 info->ptr = mpio;
241
242 return 0;
243}
244
245static void clear_mapinfo(struct multipath *m, union map_info *info)
246{
247 struct dm_mpath_io *mpio = info->ptr;
248
249 info->ptr = NULL;
250 mempool_free(mpio, m->mpio_pool);
251}
252
253/*-----------------------------------------------
254 * Path selection
255 *-----------------------------------------------*/
256
257static void __pg_init_all_paths(struct multipath *m)
258{
259 struct pgpath *pgpath;
260 unsigned long pg_init_delay = 0;
261
262 m->pg_init_count++;
263 m->pg_init_required = 0;
264 if (m->pg_init_delay_retry)
265 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
266 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
267 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
268 /* Skip failed paths */
269 if (!pgpath->is_active)
270 continue;
271 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
272 pg_init_delay))
273 m->pg_init_in_progress++;
274 }
275}
276
277static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
278{
279 m->current_pg = pgpath->pg;
280
281 /* Must we initialise the PG first, and queue I/O till it's ready? */
282 if (m->hw_handler_name) {
283 m->pg_init_required = 1;
284 m->queue_io = 1;
285 } else {
286 m->pg_init_required = 0;
287 m->queue_io = 0;
288 }
289
290 m->pg_init_count = 0;
291}
292
293static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
294 size_t nr_bytes)
295{
296 struct dm_path *path;
297
298 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
299 if (!path)
300 return -ENXIO;
301
302 m->current_pgpath = path_to_pgpath(path);
303
304 if (m->current_pg != pg)
305 __switch_pg(m, m->current_pgpath);
306
307 return 0;
308}
309
310static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
311{
312 struct priority_group *pg;
313 unsigned bypassed = 1;
314
315 if (!m->nr_valid_paths)
316 goto failed;
317
318 /* Were we instructed to switch PG? */
319 if (m->next_pg) {
320 pg = m->next_pg;
321 m->next_pg = NULL;
322 if (!__choose_path_in_pg(m, pg, nr_bytes))
323 return;
324 }
325
326 /* Don't change PG until it has no remaining paths */
327 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
328 return;
329
330 /*
331 * Loop through priority groups until we find a valid path.
332 * First time we skip PGs marked 'bypassed'.
333 * Second time we only try the ones we skipped, but set
334 * pg_init_delay_retry so we do not hammer controllers.
335 */
336 do {
337 list_for_each_entry(pg, &m->priority_groups, list) {
338 if (pg->bypassed == bypassed)
339 continue;
340 if (!__choose_path_in_pg(m, pg, nr_bytes)) {
341 if (!bypassed)
342 m->pg_init_delay_retry = 1;
343 return;
344 }
345 }
346 } while (bypassed--);
347
348failed:
349 m->current_pgpath = NULL;
350 m->current_pg = NULL;
351}
352
353/*
354 * Check whether bios must be queued in the device-mapper core rather
355 * than here in the target.
356 *
357 * m->lock must be held on entry.
358 *
359 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
360 * same value then we are not between multipath_presuspend()
361 * and multipath_resume() calls and we have no need to check
362 * for the DMF_NOFLUSH_SUSPENDING flag.
363 */
364static int __must_push_back(struct multipath *m)
365{
366 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
367 dm_noflush_suspending(m->ti));
368}
369
370static int map_io(struct multipath *m, struct request *clone,
371 union map_info *map_context, unsigned was_queued)
372{
373 int r = DM_MAPIO_REMAPPED;
374 size_t nr_bytes = blk_rq_bytes(clone);
375 unsigned long flags;
376 struct pgpath *pgpath;
377 struct block_device *bdev;
378 struct dm_mpath_io *mpio = map_context->ptr;
379
380 spin_lock_irqsave(&m->lock, flags);
381
382 /* Do we need to select a new pgpath? */
383 if (!m->current_pgpath ||
384 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
385 __choose_pgpath(m, nr_bytes);
386
387 pgpath = m->current_pgpath;
388
389 if (was_queued)
390 m->queue_size--;
391
392 if ((pgpath && m->queue_io) ||
393 (!pgpath && m->queue_if_no_path)) {
394 /* Queue for the daemon to resubmit */
395 list_add_tail(&clone->queuelist, &m->queued_ios);
396 m->queue_size++;
397 if ((m->pg_init_required && !m->pg_init_in_progress) ||
398 !m->queue_io)
399 queue_work(kmultipathd, &m->process_queued_ios);
400 pgpath = NULL;
401 r = DM_MAPIO_SUBMITTED;
402 } else if (pgpath) {
403 bdev = pgpath->path.dev->bdev;
404 clone->q = bdev_get_queue(bdev);
405 clone->rq_disk = bdev->bd_disk;
406 } else if (__must_push_back(m))
407 r = DM_MAPIO_REQUEUE;
408 else
409 r = -EIO; /* Failed */
410
411 mpio->pgpath = pgpath;
412 mpio->nr_bytes = nr_bytes;
413
414 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
415 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
416 nr_bytes);
417
418 spin_unlock_irqrestore(&m->lock, flags);
419
420 return r;
421}
422
423/*
424 * If we run out of usable paths, should we queue I/O or error it?
425 */
426static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
427 unsigned save_old_value)
428{
429 unsigned long flags;
430
431 spin_lock_irqsave(&m->lock, flags);
432
433 if (save_old_value)
434 m->saved_queue_if_no_path = m->queue_if_no_path;
435 else
436 m->saved_queue_if_no_path = queue_if_no_path;
437 m->queue_if_no_path = queue_if_no_path;
438 if (!m->queue_if_no_path && m->queue_size)
439 queue_work(kmultipathd, &m->process_queued_ios);
440
441 spin_unlock_irqrestore(&m->lock, flags);
442
443 return 0;
444}
445
446/*-----------------------------------------------------------------
447 * The multipath daemon is responsible for resubmitting queued ios.
448 *---------------------------------------------------------------*/
449
450static void dispatch_queued_ios(struct multipath *m)
451{
452 int r;
453 unsigned long flags;
454 union map_info *info;
455 struct request *clone, *n;
456 LIST_HEAD(cl);
457
458 spin_lock_irqsave(&m->lock, flags);
459 list_splice_init(&m->queued_ios, &cl);
460 spin_unlock_irqrestore(&m->lock, flags);
461
462 list_for_each_entry_safe(clone, n, &cl, queuelist) {
463 list_del_init(&clone->queuelist);
464
465 info = dm_get_rq_mapinfo(clone);
466
467 r = map_io(m, clone, info, 1);
468 if (r < 0) {
469 clear_mapinfo(m, info);
470 dm_kill_unmapped_request(clone, r);
471 } else if (r == DM_MAPIO_REMAPPED)
472 dm_dispatch_request(clone);
473 else if (r == DM_MAPIO_REQUEUE) {
474 clear_mapinfo(m, info);
475 dm_requeue_unmapped_request(clone);
476 }
477 }
478}
479
480static void process_queued_ios(struct work_struct *work)
481{
482 struct multipath *m =
483 container_of(work, struct multipath, process_queued_ios);
484 struct pgpath *pgpath = NULL;
485 unsigned must_queue = 1;
486 unsigned long flags;
487
488 spin_lock_irqsave(&m->lock, flags);
489
490 if (!m->current_pgpath)
491 __choose_pgpath(m, 0);
492
493 pgpath = m->current_pgpath;
494
495 if ((pgpath && !m->queue_io) ||
496 (!pgpath && !m->queue_if_no_path))
497 must_queue = 0;
498
499 if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
500 __pg_init_all_paths(m);
501
502 spin_unlock_irqrestore(&m->lock, flags);
503 if (!must_queue)
504 dispatch_queued_ios(m);
505}
506
507/*
508 * An event is triggered whenever a path is taken out of use.
509 * Includes path failure and PG bypass.
510 */
511static void trigger_event(struct work_struct *work)
512{
513 struct multipath *m =
514 container_of(work, struct multipath, trigger_event);
515
516 dm_table_event(m->ti->table);
517}
518
519/*-----------------------------------------------------------------
520 * Constructor/argument parsing:
521 * <#multipath feature args> [<arg>]*
522 * <#hw_handler args> [hw_handler [<arg>]*]
523 * <#priority groups>
524 * <initial priority group>
525 * [<selector> <#selector args> [<arg>]*
526 * <#paths> <#per-path selector args>
527 * [<path> [<arg>]* ]+ ]+
528 *---------------------------------------------------------------*/
529static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
530 struct dm_target *ti)
531{
532 int r;
533 struct path_selector_type *pst;
534 unsigned ps_argc;
535
536 static struct dm_arg _args[] = {
537 {0, 1024, "invalid number of path selector args"},
538 };
539
540 pst = dm_get_path_selector(dm_shift_arg(as));
541 if (!pst) {
542 ti->error = "unknown path selector type";
543 return -EINVAL;
544 }
545
546 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
547 if (r) {
548 dm_put_path_selector(pst);
549 return -EINVAL;
550 }
551
552 r = pst->create(&pg->ps, ps_argc, as->argv);
553 if (r) {
554 dm_put_path_selector(pst);
555 ti->error = "path selector constructor failed";
556 return r;
557 }
558
559 pg->ps.type = pst;
560 dm_consume_args(as, ps_argc);
561
562 return 0;
563}
564
565static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
566 struct dm_target *ti)
567{
568 int r;
569 struct pgpath *p;
570 struct multipath *m = ti->private;
571
572 /* we need at least a path arg */
573 if (as->argc < 1) {
574 ti->error = "no device given";
575 return ERR_PTR(-EINVAL);
576 }
577
578 p = alloc_pgpath();
579 if (!p)
580 return ERR_PTR(-ENOMEM);
581
582 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
583 &p->path.dev);
584 if (r) {
585 ti->error = "error getting device";
586 goto bad;
587 }
588
589 if (m->hw_handler_name) {
590 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
591
592 r = scsi_dh_attach(q, m->hw_handler_name);
593 if (r == -EBUSY) {
594 /*
595 * Already attached to different hw_handler,
596 * try to reattach with correct one.
597 */
598 scsi_dh_detach(q);
599 r = scsi_dh_attach(q, m->hw_handler_name);
600 }
601
602 if (r < 0) {
603 ti->error = "error attaching hardware handler";
604 dm_put_device(ti, p->path.dev);
605 goto bad;
606 }
607
608 if (m->hw_handler_params) {
609 r = scsi_dh_set_params(q, m->hw_handler_params);
610 if (r < 0) {
611 ti->error = "unable to set hardware "
612 "handler parameters";
613 scsi_dh_detach(q);
614 dm_put_device(ti, p->path.dev);
615 goto bad;
616 }
617 }
618 }
619
620 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
621 if (r) {
622 dm_put_device(ti, p->path.dev);
623 goto bad;
624 }
625
626 return p;
627
628 bad:
629 free_pgpath(p);
630 return ERR_PTR(r);
631}
632
633static struct priority_group *parse_priority_group(struct dm_arg_set *as,
634 struct multipath *m)
635{
636 static struct dm_arg _args[] = {
637 {1, 1024, "invalid number of paths"},
638 {0, 1024, "invalid number of selector args"}
639 };
640
641 int r;
642 unsigned i, nr_selector_args, nr_args;
643 struct priority_group *pg;
644 struct dm_target *ti = m->ti;
645
646 if (as->argc < 2) {
647 as->argc = 0;
648 ti->error = "not enough priority group arguments";
649 return ERR_PTR(-EINVAL);
650 }
651
652 pg = alloc_priority_group();
653 if (!pg) {
654 ti->error = "couldn't allocate priority group";
655 return ERR_PTR(-ENOMEM);
656 }
657 pg->m = m;
658
659 r = parse_path_selector(as, pg, ti);
660 if (r)
661 goto bad;
662
663 /*
664 * read the paths
665 */
666 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
667 if (r)
668 goto bad;
669
670 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
671 if (r)
672 goto bad;
673
674 nr_args = 1 + nr_selector_args;
675 for (i = 0; i < pg->nr_pgpaths; i++) {
676 struct pgpath *pgpath;
677 struct dm_arg_set path_args;
678
679 if (as->argc < nr_args) {
680 ti->error = "not enough path parameters";
681 r = -EINVAL;
682 goto bad;
683 }
684
685 path_args.argc = nr_args;
686 path_args.argv = as->argv;
687
688 pgpath = parse_path(&path_args, &pg->ps, ti);
689 if (IS_ERR(pgpath)) {
690 r = PTR_ERR(pgpath);
691 goto bad;
692 }
693
694 pgpath->pg = pg;
695 list_add_tail(&pgpath->list, &pg->pgpaths);
696 dm_consume_args(as, nr_args);
697 }
698
699 return pg;
700
701 bad:
702 free_priority_group(pg, ti);
703 return ERR_PTR(r);
704}
705
706static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
707{
708 unsigned hw_argc;
709 int ret;
710 struct dm_target *ti = m->ti;
711
712 static struct dm_arg _args[] = {
713 {0, 1024, "invalid number of hardware handler args"},
714 };
715
716 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
717 return -EINVAL;
718
719 if (!hw_argc)
720 return 0;
721
722 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
723 if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
724 "scsi_dh_%s", m->hw_handler_name)) {
725 ti->error = "unknown hardware handler type";
726 ret = -EINVAL;
727 goto fail;
728 }
729
730 if (hw_argc > 1) {
731 char *p;
732 int i, j, len = 4;
733
734 for (i = 0; i <= hw_argc - 2; i++)
735 len += strlen(as->argv[i]) + 1;
736 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
737 if (!p) {
738 ti->error = "memory allocation failed";
739 ret = -ENOMEM;
740 goto fail;
741 }
742 j = sprintf(p, "%d", hw_argc - 1);
743 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
744 j = sprintf(p, "%s", as->argv[i]);
745 }
746 dm_consume_args(as, hw_argc - 1);
747
748 return 0;
749fail:
750 kfree(m->hw_handler_name);
751 m->hw_handler_name = NULL;
752 return ret;
753}
754
755static int parse_features(struct dm_arg_set *as, struct multipath *m)
756{
757 int r;
758 unsigned argc;
759 struct dm_target *ti = m->ti;
760 const char *arg_name;
761
762 static struct dm_arg _args[] = {
763 {0, 5, "invalid number of feature args"},
764 {1, 50, "pg_init_retries must be between 1 and 50"},
765 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
766 };
767
768 r = dm_read_arg_group(_args, as, &argc, &ti->error);
769 if (r)
770 return -EINVAL;
771
772 if (!argc)
773 return 0;
774
775 do {
776 arg_name = dm_shift_arg(as);
777 argc--;
778
779 if (!strcasecmp(arg_name, "queue_if_no_path")) {
780 r = queue_if_no_path(m, 1, 0);
781 continue;
782 }
783
784 if (!strcasecmp(arg_name, "pg_init_retries") &&
785 (argc >= 1)) {
786 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
787 argc--;
788 continue;
789 }
790
791 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
792 (argc >= 1)) {
793 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
794 argc--;
795 continue;
796 }
797
798 ti->error = "Unrecognised multipath feature request";
799 r = -EINVAL;
800 } while (argc && !r);
801
802 return r;
803}
804
805static int multipath_ctr(struct dm_target *ti, unsigned int argc,
806 char **argv)
807{
808 /* target arguments */
809 static struct dm_arg _args[] = {
810 {0, 1024, "invalid number of priority groups"},
811 {0, 1024, "invalid initial priority group number"},
812 };
813
814 int r;
815 struct multipath *m;
816 struct dm_arg_set as;
817 unsigned pg_count = 0;
818 unsigned next_pg_num;
819
820 as.argc = argc;
821 as.argv = argv;
822
823 m = alloc_multipath(ti);
824 if (!m) {
825 ti->error = "can't allocate multipath";
826 return -EINVAL;
827 }
828
829 r = parse_features(&as, m);
830 if (r)
831 goto bad;
832
833 r = parse_hw_handler(&as, m);
834 if (r)
835 goto bad;
836
837 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
838 if (r)
839 goto bad;
840
841 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
842 if (r)
843 goto bad;
844
845 if ((!m->nr_priority_groups && next_pg_num) ||
846 (m->nr_priority_groups && !next_pg_num)) {
847 ti->error = "invalid initial priority group";
848 r = -EINVAL;
849 goto bad;
850 }
851
852 /* parse the priority groups */
853 while (as.argc) {
854 struct priority_group *pg;
855
856 pg = parse_priority_group(&as, m);
857 if (IS_ERR(pg)) {
858 r = PTR_ERR(pg);
859 goto bad;
860 }
861
862 m->nr_valid_paths += pg->nr_pgpaths;
863 list_add_tail(&pg->list, &m->priority_groups);
864 pg_count++;
865 pg->pg_num = pg_count;
866 if (!--next_pg_num)
867 m->next_pg = pg;
868 }
869
870 if (pg_count != m->nr_priority_groups) {
871 ti->error = "priority group count mismatch";
872 r = -EINVAL;
873 goto bad;
874 }
875
876 ti->num_flush_requests = 1;
877 ti->num_discard_requests = 1;
878
879 return 0;
880
881 bad:
882 free_multipath(m);
883 return r;
884}
885
886static void multipath_wait_for_pg_init_completion(struct multipath *m)
887{
888 DECLARE_WAITQUEUE(wait, current);
889 unsigned long flags;
890
891 add_wait_queue(&m->pg_init_wait, &wait);
892
893 while (1) {
894 set_current_state(TASK_UNINTERRUPTIBLE);
895
896 spin_lock_irqsave(&m->lock, flags);
897 if (!m->pg_init_in_progress) {
898 spin_unlock_irqrestore(&m->lock, flags);
899 break;
900 }
901 spin_unlock_irqrestore(&m->lock, flags);
902
903 io_schedule();
904 }
905 set_current_state(TASK_RUNNING);
906
907 remove_wait_queue(&m->pg_init_wait, &wait);
908}
909
910static void flush_multipath_work(struct multipath *m)
911{
912 flush_workqueue(kmpath_handlerd);
913 multipath_wait_for_pg_init_completion(m);
914 flush_workqueue(kmultipathd);
915 flush_work_sync(&m->trigger_event);
916}
917
918static void multipath_dtr(struct dm_target *ti)
919{
920 struct multipath *m = ti->private;
921
922 flush_multipath_work(m);
923 free_multipath(m);
924}
925
926/*
927 * Map cloned requests
928 */
929static int multipath_map(struct dm_target *ti, struct request *clone,
930 union map_info *map_context)
931{
932 int r;
933 struct multipath *m = (struct multipath *) ti->private;
934
935 if (set_mapinfo(m, map_context) < 0)
936 /* ENOMEM, requeue */
937 return DM_MAPIO_REQUEUE;
938
939 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
940 r = map_io(m, clone, map_context, 0);
941 if (r < 0 || r == DM_MAPIO_REQUEUE)
942 clear_mapinfo(m, map_context);
943
944 return r;
945}
946
947/*
948 * Take a path out of use.
949 */
950static int fail_path(struct pgpath *pgpath)
951{
952 unsigned long flags;
953 struct multipath *m = pgpath->pg->m;
954
955 spin_lock_irqsave(&m->lock, flags);
956
957 if (!pgpath->is_active)
958 goto out;
959
960 DMWARN("Failing path %s.", pgpath->path.dev->name);
961
962 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
963 pgpath->is_active = 0;
964 pgpath->fail_count++;
965
966 m->nr_valid_paths--;
967
968 if (pgpath == m->current_pgpath)
969 m->current_pgpath = NULL;
970
971 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
972 pgpath->path.dev->name, m->nr_valid_paths);
973
974 schedule_work(&m->trigger_event);
975
976out:
977 spin_unlock_irqrestore(&m->lock, flags);
978
979 return 0;
980}
981
982/*
983 * Reinstate a previously-failed path
984 */
985static int reinstate_path(struct pgpath *pgpath)
986{
987 int r = 0;
988 unsigned long flags;
989 struct multipath *m = pgpath->pg->m;
990
991 spin_lock_irqsave(&m->lock, flags);
992
993 if (pgpath->is_active)
994 goto out;
995
996 if (!pgpath->pg->ps.type->reinstate_path) {
997 DMWARN("Reinstate path not supported by path selector %s",
998 pgpath->pg->ps.type->name);
999 r = -EINVAL;
1000 goto out;
1001 }
1002
1003 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1004 if (r)
1005 goto out;
1006
1007 pgpath->is_active = 1;
1008
1009 if (!m->nr_valid_paths++ && m->queue_size) {
1010 m->current_pgpath = NULL;
1011 queue_work(kmultipathd, &m->process_queued_ios);
1012 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1013 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1014 m->pg_init_in_progress++;
1015 }
1016
1017 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1018 pgpath->path.dev->name, m->nr_valid_paths);
1019
1020 schedule_work(&m->trigger_event);
1021
1022out:
1023 spin_unlock_irqrestore(&m->lock, flags);
1024
1025 return r;
1026}
1027
1028/*
1029 * Fail or reinstate all paths that match the provided struct dm_dev.
1030 */
1031static int action_dev(struct multipath *m, struct dm_dev *dev,
1032 action_fn action)
1033{
1034 int r = -EINVAL;
1035 struct pgpath *pgpath;
1036 struct priority_group *pg;
1037
1038 list_for_each_entry(pg, &m->priority_groups, list) {
1039 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1040 if (pgpath->path.dev == dev)
1041 r = action(pgpath);
1042 }
1043 }
1044
1045 return r;
1046}
1047
1048/*
1049 * Temporarily try to avoid having to use the specified PG
1050 */
1051static void bypass_pg(struct multipath *m, struct priority_group *pg,
1052 int bypassed)
1053{
1054 unsigned long flags;
1055
1056 spin_lock_irqsave(&m->lock, flags);
1057
1058 pg->bypassed = bypassed;
1059 m->current_pgpath = NULL;
1060 m->current_pg = NULL;
1061
1062 spin_unlock_irqrestore(&m->lock, flags);
1063
1064 schedule_work(&m->trigger_event);
1065}
1066
1067/*
1068 * Switch to using the specified PG from the next I/O that gets mapped
1069 */
1070static int switch_pg_num(struct multipath *m, const char *pgstr)
1071{
1072 struct priority_group *pg;
1073 unsigned pgnum;
1074 unsigned long flags;
1075 char dummy;
1076
1077 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1078 (pgnum > m->nr_priority_groups)) {
1079 DMWARN("invalid PG number supplied to switch_pg_num");
1080 return -EINVAL;
1081 }
1082
1083 spin_lock_irqsave(&m->lock, flags);
1084 list_for_each_entry(pg, &m->priority_groups, list) {
1085 pg->bypassed = 0;
1086 if (--pgnum)
1087 continue;
1088
1089 m->current_pgpath = NULL;
1090 m->current_pg = NULL;
1091 m->next_pg = pg;
1092 }
1093 spin_unlock_irqrestore(&m->lock, flags);
1094
1095 schedule_work(&m->trigger_event);
1096 return 0;
1097}
1098
1099/*
1100 * Set/clear bypassed status of a PG.
1101 * PGs are numbered upwards from 1 in the order they were declared.
1102 */
1103static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1104{
1105 struct priority_group *pg;
1106 unsigned pgnum;
1107 char dummy;
1108
1109 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1110 (pgnum > m->nr_priority_groups)) {
1111 DMWARN("invalid PG number supplied to bypass_pg");
1112 return -EINVAL;
1113 }
1114
1115 list_for_each_entry(pg, &m->priority_groups, list) {
1116 if (!--pgnum)
1117 break;
1118 }
1119
1120 bypass_pg(m, pg, bypassed);
1121 return 0;
1122}
1123
1124/*
1125 * Should we retry pg_init immediately?
1126 */
1127static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1128{
1129 unsigned long flags;
1130 int limit_reached = 0;
1131
1132 spin_lock_irqsave(&m->lock, flags);
1133
1134 if (m->pg_init_count <= m->pg_init_retries)
1135 m->pg_init_required = 1;
1136 else
1137 limit_reached = 1;
1138
1139 spin_unlock_irqrestore(&m->lock, flags);
1140
1141 return limit_reached;
1142}
1143
1144static void pg_init_done(void *data, int errors)
1145{
1146 struct pgpath *pgpath = data;
1147 struct priority_group *pg = pgpath->pg;
1148 struct multipath *m = pg->m;
1149 unsigned long flags;
1150 unsigned delay_retry = 0;
1151
1152 /* device or driver problems */
1153 switch (errors) {
1154 case SCSI_DH_OK:
1155 break;
1156 case SCSI_DH_NOSYS:
1157 if (!m->hw_handler_name) {
1158 errors = 0;
1159 break;
1160 }
1161 DMERR("Could not failover the device: Handler scsi_dh_%s "
1162 "Error %d.", m->hw_handler_name, errors);
1163 /*
1164 * Fail path for now, so we do not ping pong
1165 */
1166 fail_path(pgpath);
1167 break;
1168 case SCSI_DH_DEV_TEMP_BUSY:
1169 /*
1170 * Probably doing something like FW upgrade on the
1171 * controller so try the other pg.
1172 */
1173 bypass_pg(m, pg, 1);
1174 break;
1175 case SCSI_DH_RETRY:
1176 /* Wait before retrying. */
1177 delay_retry = 1;
1178 case SCSI_DH_IMM_RETRY:
1179 case SCSI_DH_RES_TEMP_UNAVAIL:
1180 if (pg_init_limit_reached(m, pgpath))
1181 fail_path(pgpath);
1182 errors = 0;
1183 break;
1184 default:
1185 /*
1186 * We probably do not want to fail the path for a device
1187 * error, but this is what the old dm did. In future
1188 * patches we can do more advanced handling.
1189 */
1190 fail_path(pgpath);
1191 }
1192
1193 spin_lock_irqsave(&m->lock, flags);
1194 if (errors) {
1195 if (pgpath == m->current_pgpath) {
1196 DMERR("Could not failover device. Error %d.", errors);
1197 m->current_pgpath = NULL;
1198 m->current_pg = NULL;
1199 }
1200 } else if (!m->pg_init_required)
1201 pg->bypassed = 0;
1202
1203 if (--m->pg_init_in_progress)
1204 /* Activations of other paths are still on going */
1205 goto out;
1206
1207 if (!m->pg_init_required)
1208 m->queue_io = 0;
1209
1210 m->pg_init_delay_retry = delay_retry;
1211 queue_work(kmultipathd, &m->process_queued_ios);
1212
1213 /*
1214 * Wake up any thread waiting to suspend.
1215 */
1216 wake_up(&m->pg_init_wait);
1217
1218out:
1219 spin_unlock_irqrestore(&m->lock, flags);
1220}
1221
1222static void activate_path(struct work_struct *work)
1223{
1224 struct pgpath *pgpath =
1225 container_of(work, struct pgpath, activate_path.work);
1226
1227 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1228 pg_init_done, pgpath);
1229}
1230
1231/*
1232 * end_io handling
1233 */
1234static int do_end_io(struct multipath *m, struct request *clone,
1235 int error, struct dm_mpath_io *mpio)
1236{
1237 /*
1238 * We don't queue any clone request inside the multipath target
1239 * during end I/O handling, since those clone requests don't have
1240 * bio clones. If we queue them inside the multipath target,
1241 * we need to make bio clones, that requires memory allocation.
1242 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1243 * don't have bio clones.)
1244 * Instead of queueing the clone request here, we queue the original
1245 * request into dm core, which will remake a clone request and
1246 * clone bios for it and resubmit it later.
1247 */
1248 int r = DM_ENDIO_REQUEUE;
1249 unsigned long flags;
1250
1251 if (!error && !clone->errors)
1252 return 0; /* I/O complete */
1253
1254 if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1255 return error;
1256
1257 if (mpio->pgpath)
1258 fail_path(mpio->pgpath);
1259
1260 spin_lock_irqsave(&m->lock, flags);
1261 if (!m->nr_valid_paths) {
1262 if (!m->queue_if_no_path) {
1263 if (!__must_push_back(m))
1264 r = -EIO;
1265 } else {
1266 if (error == -EBADE)
1267 r = error;
1268 }
1269 }
1270 spin_unlock_irqrestore(&m->lock, flags);
1271
1272 return r;
1273}
1274
1275static int multipath_end_io(struct dm_target *ti, struct request *clone,
1276 int error, union map_info *map_context)
1277{
1278 struct multipath *m = ti->private;
1279 struct dm_mpath_io *mpio = map_context->ptr;
1280 struct pgpath *pgpath = mpio->pgpath;
1281 struct path_selector *ps;
1282 int r;
1283
1284 BUG_ON(!mpio);
1285
1286 r = do_end_io(m, clone, error, mpio);
1287 if (pgpath) {
1288 ps = &pgpath->pg->ps;
1289 if (ps->type->end_io)
1290 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1291 }
1292 clear_mapinfo(m, map_context);
1293
1294 return r;
1295}
1296
1297/*
1298 * Suspend can't complete until all the I/O is processed so if
1299 * the last path fails we must error any remaining I/O.
1300 * Note that if the freeze_bdev fails while suspending, the
1301 * queue_if_no_path state is lost - userspace should reset it.
1302 */
1303static void multipath_presuspend(struct dm_target *ti)
1304{
1305 struct multipath *m = (struct multipath *) ti->private;
1306
1307 queue_if_no_path(m, 0, 1);
1308}
1309
1310static void multipath_postsuspend(struct dm_target *ti)
1311{
1312 struct multipath *m = ti->private;
1313
1314 mutex_lock(&m->work_mutex);
1315 flush_multipath_work(m);
1316 mutex_unlock(&m->work_mutex);
1317}
1318
1319/*
1320 * Restore the queue_if_no_path setting.
1321 */
1322static void multipath_resume(struct dm_target *ti)
1323{
1324 struct multipath *m = (struct multipath *) ti->private;
1325 unsigned long flags;
1326
1327 spin_lock_irqsave(&m->lock, flags);
1328 m->queue_if_no_path = m->saved_queue_if_no_path;
1329 spin_unlock_irqrestore(&m->lock, flags);
1330}
1331
1332/*
1333 * Info output has the following format:
1334 * num_multipath_feature_args [multipath_feature_args]*
1335 * num_handler_status_args [handler_status_args]*
1336 * num_groups init_group_number
1337 * [A|D|E num_ps_status_args [ps_status_args]*
1338 * num_paths num_selector_args
1339 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1340 *
1341 * Table output has the following format (identical to the constructor string):
1342 * num_feature_args [features_args]*
1343 * num_handler_args hw_handler [hw_handler_args]*
1344 * num_groups init_group_number
1345 * [priority selector-name num_ps_args [ps_args]*
1346 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1347 */
1348static int multipath_status(struct dm_target *ti, status_type_t type,
1349 char *result, unsigned int maxlen)
1350{
1351 int sz = 0;
1352 unsigned long flags;
1353 struct multipath *m = (struct multipath *) ti->private;
1354 struct priority_group *pg;
1355 struct pgpath *p;
1356 unsigned pg_num;
1357 char state;
1358
1359 spin_lock_irqsave(&m->lock, flags);
1360
1361 /* Features */
1362 if (type == STATUSTYPE_INFO)
1363 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1364 else {
1365 DMEMIT("%u ", m->queue_if_no_path +
1366 (m->pg_init_retries > 0) * 2 +
1367 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2);
1368 if (m->queue_if_no_path)
1369 DMEMIT("queue_if_no_path ");
1370 if (m->pg_init_retries)
1371 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1372 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1373 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1374 }
1375
1376 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1377 DMEMIT("0 ");
1378 else
1379 DMEMIT("1 %s ", m->hw_handler_name);
1380
1381 DMEMIT("%u ", m->nr_priority_groups);
1382
1383 if (m->next_pg)
1384 pg_num = m->next_pg->pg_num;
1385 else if (m->current_pg)
1386 pg_num = m->current_pg->pg_num;
1387 else
1388 pg_num = (m->nr_priority_groups ? 1 : 0);
1389
1390 DMEMIT("%u ", pg_num);
1391
1392 switch (type) {
1393 case STATUSTYPE_INFO:
1394 list_for_each_entry(pg, &m->priority_groups, list) {
1395 if (pg->bypassed)
1396 state = 'D'; /* Disabled */
1397 else if (pg == m->current_pg)
1398 state = 'A'; /* Currently Active */
1399 else
1400 state = 'E'; /* Enabled */
1401
1402 DMEMIT("%c ", state);
1403
1404 if (pg->ps.type->status)
1405 sz += pg->ps.type->status(&pg->ps, NULL, type,
1406 result + sz,
1407 maxlen - sz);
1408 else
1409 DMEMIT("0 ");
1410
1411 DMEMIT("%u %u ", pg->nr_pgpaths,
1412 pg->ps.type->info_args);
1413
1414 list_for_each_entry(p, &pg->pgpaths, list) {
1415 DMEMIT("%s %s %u ", p->path.dev->name,
1416 p->is_active ? "A" : "F",
1417 p->fail_count);
1418 if (pg->ps.type->status)
1419 sz += pg->ps.type->status(&pg->ps,
1420 &p->path, type, result + sz,
1421 maxlen - sz);
1422 }
1423 }
1424 break;
1425
1426 case STATUSTYPE_TABLE:
1427 list_for_each_entry(pg, &m->priority_groups, list) {
1428 DMEMIT("%s ", pg->ps.type->name);
1429
1430 if (pg->ps.type->status)
1431 sz += pg->ps.type->status(&pg->ps, NULL, type,
1432 result + sz,
1433 maxlen - sz);
1434 else
1435 DMEMIT("0 ");
1436
1437 DMEMIT("%u %u ", pg->nr_pgpaths,
1438 pg->ps.type->table_args);
1439
1440 list_for_each_entry(p, &pg->pgpaths, list) {
1441 DMEMIT("%s ", p->path.dev->name);
1442 if (pg->ps.type->status)
1443 sz += pg->ps.type->status(&pg->ps,
1444 &p->path, type, result + sz,
1445 maxlen - sz);
1446 }
1447 }
1448 break;
1449 }
1450
1451 spin_unlock_irqrestore(&m->lock, flags);
1452
1453 return 0;
1454}
1455
1456static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1457{
1458 int r = -EINVAL;
1459 struct dm_dev *dev;
1460 struct multipath *m = (struct multipath *) ti->private;
1461 action_fn action;
1462
1463 mutex_lock(&m->work_mutex);
1464
1465 if (dm_suspended(ti)) {
1466 r = -EBUSY;
1467 goto out;
1468 }
1469
1470 if (argc == 1) {
1471 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1472 r = queue_if_no_path(m, 1, 0);
1473 goto out;
1474 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1475 r = queue_if_no_path(m, 0, 0);
1476 goto out;
1477 }
1478 }
1479
1480 if (argc != 2) {
1481 DMWARN("Unrecognised multipath message received.");
1482 goto out;
1483 }
1484
1485 if (!strcasecmp(argv[0], "disable_group")) {
1486 r = bypass_pg_num(m, argv[1], 1);
1487 goto out;
1488 } else if (!strcasecmp(argv[0], "enable_group")) {
1489 r = bypass_pg_num(m, argv[1], 0);
1490 goto out;
1491 } else if (!strcasecmp(argv[0], "switch_group")) {
1492 r = switch_pg_num(m, argv[1]);
1493 goto out;
1494 } else if (!strcasecmp(argv[0], "reinstate_path"))
1495 action = reinstate_path;
1496 else if (!strcasecmp(argv[0], "fail_path"))
1497 action = fail_path;
1498 else {
1499 DMWARN("Unrecognised multipath message received.");
1500 goto out;
1501 }
1502
1503 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1504 if (r) {
1505 DMWARN("message: error getting device %s",
1506 argv[1]);
1507 goto out;
1508 }
1509
1510 r = action_dev(m, dev, action);
1511
1512 dm_put_device(ti, dev);
1513
1514out:
1515 mutex_unlock(&m->work_mutex);
1516 return r;
1517}
1518
1519static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1520 unsigned long arg)
1521{
1522 struct multipath *m = ti->private;
1523 struct pgpath *pgpath;
1524 struct block_device *bdev;
1525 fmode_t mode;
1526 unsigned long flags;
1527 int r;
1528
1529again:
1530 bdev = NULL;
1531 mode = 0;
1532 r = 0;
1533
1534 spin_lock_irqsave(&m->lock, flags);
1535
1536 if (!m->current_pgpath)
1537 __choose_pgpath(m, 0);
1538
1539 pgpath = m->current_pgpath;
1540
1541 if (pgpath) {
1542 bdev = pgpath->path.dev->bdev;
1543 mode = pgpath->path.dev->mode;
1544 }
1545
1546 if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1547 r = -EAGAIN;
1548 else if (!bdev)
1549 r = -EIO;
1550
1551 spin_unlock_irqrestore(&m->lock, flags);
1552
1553 /*
1554 * Only pass ioctls through if the device sizes match exactly.
1555 */
1556 if (!r && ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT)
1557 r = scsi_verify_blk_ioctl(NULL, cmd);
1558
1559 if (r == -EAGAIN && !fatal_signal_pending(current)) {
1560 queue_work(kmultipathd, &m->process_queued_ios);
1561 msleep(10);
1562 goto again;
1563 }
1564
1565 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1566}
1567
1568static int multipath_iterate_devices(struct dm_target *ti,
1569 iterate_devices_callout_fn fn, void *data)
1570{
1571 struct multipath *m = ti->private;
1572 struct priority_group *pg;
1573 struct pgpath *p;
1574 int ret = 0;
1575
1576 list_for_each_entry(pg, &m->priority_groups, list) {
1577 list_for_each_entry(p, &pg->pgpaths, list) {
1578 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1579 if (ret)
1580 goto out;
1581 }
1582 }
1583
1584out:
1585 return ret;
1586}
1587
1588static int __pgpath_busy(struct pgpath *pgpath)
1589{
1590 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1591
1592 return dm_underlying_device_busy(q);
1593}
1594
1595/*
1596 * We return "busy", only when we can map I/Os but underlying devices
1597 * are busy (so even if we map I/Os now, the I/Os will wait on
1598 * the underlying queue).
1599 * In other words, if we want to kill I/Os or queue them inside us
1600 * due to map unavailability, we don't return "busy". Otherwise,
1601 * dm core won't give us the I/Os and we can't do what we want.
1602 */
1603static int multipath_busy(struct dm_target *ti)
1604{
1605 int busy = 0, has_active = 0;
1606 struct multipath *m = ti->private;
1607 struct priority_group *pg;
1608 struct pgpath *pgpath;
1609 unsigned long flags;
1610
1611 spin_lock_irqsave(&m->lock, flags);
1612
1613 /* Guess which priority_group will be used at next mapping time */
1614 if (unlikely(!m->current_pgpath && m->next_pg))
1615 pg = m->next_pg;
1616 else if (likely(m->current_pg))
1617 pg = m->current_pg;
1618 else
1619 /*
1620 * We don't know which pg will be used at next mapping time.
1621 * We don't call __choose_pgpath() here to avoid to trigger
1622 * pg_init just by busy checking.
1623 * So we don't know whether underlying devices we will be using
1624 * at next mapping time are busy or not. Just try mapping.
1625 */
1626 goto out;
1627
1628 /*
1629 * If there is one non-busy active path at least, the path selector
1630 * will be able to select it. So we consider such a pg as not busy.
1631 */
1632 busy = 1;
1633 list_for_each_entry(pgpath, &pg->pgpaths, list)
1634 if (pgpath->is_active) {
1635 has_active = 1;
1636
1637 if (!__pgpath_busy(pgpath)) {
1638 busy = 0;
1639 break;
1640 }
1641 }
1642
1643 if (!has_active)
1644 /*
1645 * No active path in this pg, so this pg won't be used and
1646 * the current_pg will be changed at next mapping time.
1647 * We need to try mapping to determine it.
1648 */
1649 busy = 0;
1650
1651out:
1652 spin_unlock_irqrestore(&m->lock, flags);
1653
1654 return busy;
1655}
1656
1657/*-----------------------------------------------------------------
1658 * Module setup
1659 *---------------------------------------------------------------*/
1660static struct target_type multipath_target = {
1661 .name = "multipath",
1662 .version = {1, 4, 0},
1663 .module = THIS_MODULE,
1664 .ctr = multipath_ctr,
1665 .dtr = multipath_dtr,
1666 .map_rq = multipath_map,
1667 .rq_end_io = multipath_end_io,
1668 .presuspend = multipath_presuspend,
1669 .postsuspend = multipath_postsuspend,
1670 .resume = multipath_resume,
1671 .status = multipath_status,
1672 .message = multipath_message,
1673 .ioctl = multipath_ioctl,
1674 .iterate_devices = multipath_iterate_devices,
1675 .busy = multipath_busy,
1676};
1677
1678static int __init dm_multipath_init(void)
1679{
1680 int r;
1681
1682 /* allocate a slab for the dm_ios */
1683 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1684 if (!_mpio_cache)
1685 return -ENOMEM;
1686
1687 r = dm_register_target(&multipath_target);
1688 if (r < 0) {
1689 DMERR("register failed %d", r);
1690 kmem_cache_destroy(_mpio_cache);
1691 return -EINVAL;
1692 }
1693
1694 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1695 if (!kmultipathd) {
1696 DMERR("failed to create workqueue kmpathd");
1697 dm_unregister_target(&multipath_target);
1698 kmem_cache_destroy(_mpio_cache);
1699 return -ENOMEM;
1700 }
1701
1702 /*
1703 * A separate workqueue is used to handle the device handlers
1704 * to avoid overloading existing workqueue. Overloading the
1705 * old workqueue would also create a bottleneck in the
1706 * path of the storage hardware device activation.
1707 */
1708 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1709 WQ_MEM_RECLAIM);
1710 if (!kmpath_handlerd) {
1711 DMERR("failed to create workqueue kmpath_handlerd");
1712 destroy_workqueue(kmultipathd);
1713 dm_unregister_target(&multipath_target);
1714 kmem_cache_destroy(_mpio_cache);
1715 return -ENOMEM;
1716 }
1717
1718 DMINFO("version %u.%u.%u loaded",
1719 multipath_target.version[0], multipath_target.version[1],
1720 multipath_target.version[2]);
1721
1722 return r;
1723}
1724
1725static void __exit dm_multipath_exit(void)
1726{
1727 destroy_workqueue(kmpath_handlerd);
1728 destroy_workqueue(kmultipathd);
1729
1730 dm_unregister_target(&multipath_target);
1731 kmem_cache_destroy(_mpio_cache);
1732}
1733
1734module_init(dm_multipath_init);
1735module_exit(dm_multipath_exit);
1736
1737MODULE_DESCRIPTION(DM_NAME " multipath target");
1738MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1739MODULE_LICENSE("GPL");
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 union map_info *map_context)
549{
550 if (unlikely(map_context)) {
551 /*
552 * non-NULL map_context means caller is still map
553 * method; must undo multipath_clone_and_map()
554 */
555 struct dm_mpath_io *mpio = get_mpio(map_context);
556 struct pgpath *pgpath = mpio->pgpath;
557
558 if (pgpath && pgpath->pg->ps.type->end_io)
559 pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
560 &pgpath->path,
561 mpio->nr_bytes);
562 }
563
564 blk_put_request(clone);
565}
566
567/*
568 * Map cloned bios (bio-based multipath)
569 */
570
571static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
572{
573 struct pgpath *pgpath;
574 unsigned long flags;
575 bool queue_io;
576
577 /* Do we need to select a new pgpath? */
578 pgpath = READ_ONCE(m->current_pgpath);
579 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
580 if (!pgpath || !queue_io)
581 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
582
583 if ((pgpath && queue_io) ||
584 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
585 /* Queue for the daemon to resubmit */
586 spin_lock_irqsave(&m->lock, flags);
587 bio_list_add(&m->queued_bios, bio);
588 spin_unlock_irqrestore(&m->lock, flags);
589
590 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
591 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
592 pg_init_all_paths(m);
593 else if (!queue_io)
594 queue_work(kmultipathd, &m->process_queued_bios);
595
596 return ERR_PTR(-EAGAIN);
597 }
598
599 return pgpath;
600}
601
602static struct pgpath *__map_bio_fast(struct multipath *m, struct bio *bio)
603{
604 struct pgpath *pgpath;
605 unsigned long flags;
606
607 /* Do we need to select a new pgpath? */
608 /*
609 * FIXME: currently only switching path if no path (due to failure, etc)
610 * - which negates the point of using a path selector
611 */
612 pgpath = READ_ONCE(m->current_pgpath);
613 if (!pgpath)
614 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
615
616 if (!pgpath) {
617 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
618 /* Queue for the daemon to resubmit */
619 spin_lock_irqsave(&m->lock, flags);
620 bio_list_add(&m->queued_bios, bio);
621 spin_unlock_irqrestore(&m->lock, flags);
622 queue_work(kmultipathd, &m->process_queued_bios);
623
624 return ERR_PTR(-EAGAIN);
625 }
626 return NULL;
627 }
628
629 return pgpath;
630}
631
632static int __multipath_map_bio(struct multipath *m, struct bio *bio,
633 struct dm_mpath_io *mpio)
634{
635 struct pgpath *pgpath;
636
637 if (!m->hw_handler_name)
638 pgpath = __map_bio_fast(m, bio);
639 else
640 pgpath = __map_bio(m, bio);
641
642 if (IS_ERR(pgpath))
643 return DM_MAPIO_SUBMITTED;
644
645 if (!pgpath) {
646 if (must_push_back_bio(m))
647 return DM_MAPIO_REQUEUE;
648 dm_report_EIO(m);
649 return DM_MAPIO_KILL;
650 }
651
652 mpio->pgpath = pgpath;
653
654 bio->bi_status = 0;
655 bio_set_dev(bio, pgpath->path.dev->bdev);
656 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
657
658 if (pgpath->pg->ps.type->start_io)
659 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
660 &pgpath->path,
661 mpio->nr_bytes);
662 return DM_MAPIO_REMAPPED;
663}
664
665static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
666{
667 struct multipath *m = ti->private;
668 struct dm_mpath_io *mpio = NULL;
669
670 multipath_init_per_bio_data(bio, &mpio);
671 return __multipath_map_bio(m, bio, mpio);
672}
673
674static void process_queued_io_list(struct multipath *m)
675{
676 if (m->queue_mode == DM_TYPE_REQUEST_BASED)
677 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
678 else if (m->queue_mode == DM_TYPE_BIO_BASED)
679 queue_work(kmultipathd, &m->process_queued_bios);
680}
681
682static void process_queued_bios(struct work_struct *work)
683{
684 int r;
685 unsigned long flags;
686 struct bio *bio;
687 struct bio_list bios;
688 struct blk_plug plug;
689 struct multipath *m =
690 container_of(work, struct multipath, process_queued_bios);
691
692 bio_list_init(&bios);
693
694 spin_lock_irqsave(&m->lock, flags);
695
696 if (bio_list_empty(&m->queued_bios)) {
697 spin_unlock_irqrestore(&m->lock, flags);
698 return;
699 }
700
701 bio_list_merge(&bios, &m->queued_bios);
702 bio_list_init(&m->queued_bios);
703
704 spin_unlock_irqrestore(&m->lock, flags);
705
706 blk_start_plug(&plug);
707 while ((bio = bio_list_pop(&bios))) {
708 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
709 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
710 r = __multipath_map_bio(m, bio, mpio);
711 switch (r) {
712 case DM_MAPIO_KILL:
713 bio->bi_status = BLK_STS_IOERR;
714 bio_endio(bio);
715 break;
716 case DM_MAPIO_REQUEUE:
717 bio->bi_status = BLK_STS_DM_REQUEUE;
718 bio_endio(bio);
719 break;
720 case DM_MAPIO_REMAPPED:
721 generic_make_request(bio);
722 break;
723 case DM_MAPIO_SUBMITTED:
724 break;
725 default:
726 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
727 }
728 }
729 blk_finish_plug(&plug);
730}
731
732/*
733 * If we run out of usable paths, should we queue I/O or error it?
734 */
735static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
736 bool save_old_value)
737{
738 unsigned long flags;
739
740 spin_lock_irqsave(&m->lock, flags);
741 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
742 (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
743 (!save_old_value && queue_if_no_path));
744 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
745 spin_unlock_irqrestore(&m->lock, flags);
746
747 if (!queue_if_no_path) {
748 dm_table_run_md_queue_async(m->ti->table);
749 process_queued_io_list(m);
750 }
751
752 return 0;
753}
754
755/*
756 * An event is triggered whenever a path is taken out of use.
757 * Includes path failure and PG bypass.
758 */
759static void trigger_event(struct work_struct *work)
760{
761 struct multipath *m =
762 container_of(work, struct multipath, trigger_event);
763
764 dm_table_event(m->ti->table);
765}
766
767/*-----------------------------------------------------------------
768 * Constructor/argument parsing:
769 * <#multipath feature args> [<arg>]*
770 * <#hw_handler args> [hw_handler [<arg>]*]
771 * <#priority groups>
772 * <initial priority group>
773 * [<selector> <#selector args> [<arg>]*
774 * <#paths> <#per-path selector args>
775 * [<path> [<arg>]* ]+ ]+
776 *---------------------------------------------------------------*/
777static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
778 struct dm_target *ti)
779{
780 int r;
781 struct path_selector_type *pst;
782 unsigned ps_argc;
783
784 static const struct dm_arg _args[] = {
785 {0, 1024, "invalid number of path selector args"},
786 };
787
788 pst = dm_get_path_selector(dm_shift_arg(as));
789 if (!pst) {
790 ti->error = "unknown path selector type";
791 return -EINVAL;
792 }
793
794 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
795 if (r) {
796 dm_put_path_selector(pst);
797 return -EINVAL;
798 }
799
800 r = pst->create(&pg->ps, ps_argc, as->argv);
801 if (r) {
802 dm_put_path_selector(pst);
803 ti->error = "path selector constructor failed";
804 return r;
805 }
806
807 pg->ps.type = pst;
808 dm_consume_args(as, ps_argc);
809
810 return 0;
811}
812
813static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
814 const char **attached_handler_name, char **error)
815{
816 struct request_queue *q = bdev_get_queue(bdev);
817 int r;
818
819 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
820retain:
821 if (*attached_handler_name) {
822 /*
823 * Clear any hw_handler_params associated with a
824 * handler that isn't already attached.
825 */
826 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
827 kfree(m->hw_handler_params);
828 m->hw_handler_params = NULL;
829 }
830
831 /*
832 * Reset hw_handler_name to match the attached handler
833 *
834 * NB. This modifies the table line to show the actual
835 * handler instead of the original table passed in.
836 */
837 kfree(m->hw_handler_name);
838 m->hw_handler_name = *attached_handler_name;
839 *attached_handler_name = NULL;
840 }
841 }
842
843 if (m->hw_handler_name) {
844 r = scsi_dh_attach(q, m->hw_handler_name);
845 if (r == -EBUSY) {
846 char b[BDEVNAME_SIZE];
847
848 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
849 bdevname(bdev, b));
850 goto retain;
851 }
852 if (r < 0) {
853 *error = "error attaching hardware handler";
854 return r;
855 }
856
857 if (m->hw_handler_params) {
858 r = scsi_dh_set_params(q, m->hw_handler_params);
859 if (r < 0) {
860 *error = "unable to set hardware handler parameters";
861 return r;
862 }
863 }
864 }
865
866 return 0;
867}
868
869static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
870 struct dm_target *ti)
871{
872 int r;
873 struct pgpath *p;
874 struct multipath *m = ti->private;
875 struct request_queue *q;
876 const char *attached_handler_name = NULL;
877
878 /* we need at least a path arg */
879 if (as->argc < 1) {
880 ti->error = "no device given";
881 return ERR_PTR(-EINVAL);
882 }
883
884 p = alloc_pgpath();
885 if (!p)
886 return ERR_PTR(-ENOMEM);
887
888 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
889 &p->path.dev);
890 if (r) {
891 ti->error = "error getting device";
892 goto bad;
893 }
894
895 q = bdev_get_queue(p->path.dev->bdev);
896 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
897 if (attached_handler_name || m->hw_handler_name) {
898 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
899 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
900 kfree(attached_handler_name);
901 if (r) {
902 dm_put_device(ti, p->path.dev);
903 goto bad;
904 }
905 }
906
907 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
908 if (r) {
909 dm_put_device(ti, p->path.dev);
910 goto bad;
911 }
912
913 return p;
914 bad:
915 free_pgpath(p);
916 return ERR_PTR(r);
917}
918
919static struct priority_group *parse_priority_group(struct dm_arg_set *as,
920 struct multipath *m)
921{
922 static const struct dm_arg _args[] = {
923 {1, 1024, "invalid number of paths"},
924 {0, 1024, "invalid number of selector args"}
925 };
926
927 int r;
928 unsigned i, nr_selector_args, nr_args;
929 struct priority_group *pg;
930 struct dm_target *ti = m->ti;
931
932 if (as->argc < 2) {
933 as->argc = 0;
934 ti->error = "not enough priority group arguments";
935 return ERR_PTR(-EINVAL);
936 }
937
938 pg = alloc_priority_group();
939 if (!pg) {
940 ti->error = "couldn't allocate priority group";
941 return ERR_PTR(-ENOMEM);
942 }
943 pg->m = m;
944
945 r = parse_path_selector(as, pg, ti);
946 if (r)
947 goto bad;
948
949 /*
950 * read the paths
951 */
952 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
953 if (r)
954 goto bad;
955
956 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
957 if (r)
958 goto bad;
959
960 nr_args = 1 + nr_selector_args;
961 for (i = 0; i < pg->nr_pgpaths; i++) {
962 struct pgpath *pgpath;
963 struct dm_arg_set path_args;
964
965 if (as->argc < nr_args) {
966 ti->error = "not enough path parameters";
967 r = -EINVAL;
968 goto bad;
969 }
970
971 path_args.argc = nr_args;
972 path_args.argv = as->argv;
973
974 pgpath = parse_path(&path_args, &pg->ps, ti);
975 if (IS_ERR(pgpath)) {
976 r = PTR_ERR(pgpath);
977 goto bad;
978 }
979
980 pgpath->pg = pg;
981 list_add_tail(&pgpath->list, &pg->pgpaths);
982 dm_consume_args(as, nr_args);
983 }
984
985 return pg;
986
987 bad:
988 free_priority_group(pg, ti);
989 return ERR_PTR(r);
990}
991
992static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
993{
994 unsigned hw_argc;
995 int ret;
996 struct dm_target *ti = m->ti;
997
998 static const struct dm_arg _args[] = {
999 {0, 1024, "invalid number of hardware handler args"},
1000 };
1001
1002 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1003 return -EINVAL;
1004
1005 if (!hw_argc)
1006 return 0;
1007
1008 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1009 dm_consume_args(as, hw_argc);
1010 DMERR("bio-based multipath doesn't allow hardware handler args");
1011 return 0;
1012 }
1013
1014 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1015 if (!m->hw_handler_name)
1016 return -EINVAL;
1017
1018 if (hw_argc > 1) {
1019 char *p;
1020 int i, j, len = 4;
1021
1022 for (i = 0; i <= hw_argc - 2; i++)
1023 len += strlen(as->argv[i]) + 1;
1024 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1025 if (!p) {
1026 ti->error = "memory allocation failed";
1027 ret = -ENOMEM;
1028 goto fail;
1029 }
1030 j = sprintf(p, "%d", hw_argc - 1);
1031 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1032 j = sprintf(p, "%s", as->argv[i]);
1033 }
1034 dm_consume_args(as, hw_argc - 1);
1035
1036 return 0;
1037fail:
1038 kfree(m->hw_handler_name);
1039 m->hw_handler_name = NULL;
1040 return ret;
1041}
1042
1043static int parse_features(struct dm_arg_set *as, struct multipath *m)
1044{
1045 int r;
1046 unsigned argc;
1047 struct dm_target *ti = m->ti;
1048 const char *arg_name;
1049
1050 static const struct dm_arg _args[] = {
1051 {0, 8, "invalid number of feature args"},
1052 {1, 50, "pg_init_retries must be between 1 and 50"},
1053 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1054 };
1055
1056 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1057 if (r)
1058 return -EINVAL;
1059
1060 if (!argc)
1061 return 0;
1062
1063 do {
1064 arg_name = dm_shift_arg(as);
1065 argc--;
1066
1067 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1068 r = queue_if_no_path(m, true, false);
1069 continue;
1070 }
1071
1072 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1073 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1074 continue;
1075 }
1076
1077 if (!strcasecmp(arg_name, "pg_init_retries") &&
1078 (argc >= 1)) {
1079 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1080 argc--;
1081 continue;
1082 }
1083
1084 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1085 (argc >= 1)) {
1086 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1087 argc--;
1088 continue;
1089 }
1090
1091 if (!strcasecmp(arg_name, "queue_mode") &&
1092 (argc >= 1)) {
1093 const char *queue_mode_name = dm_shift_arg(as);
1094
1095 if (!strcasecmp(queue_mode_name, "bio"))
1096 m->queue_mode = DM_TYPE_BIO_BASED;
1097 else if (!strcasecmp(queue_mode_name, "rq") ||
1098 !strcasecmp(queue_mode_name, "mq"))
1099 m->queue_mode = DM_TYPE_REQUEST_BASED;
1100 else {
1101 ti->error = "Unknown 'queue_mode' requested";
1102 r = -EINVAL;
1103 }
1104 argc--;
1105 continue;
1106 }
1107
1108 ti->error = "Unrecognised multipath feature request";
1109 r = -EINVAL;
1110 } while (argc && !r);
1111
1112 return r;
1113}
1114
1115static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1116{
1117 /* target arguments */
1118 static const struct dm_arg _args[] = {
1119 {0, 1024, "invalid number of priority groups"},
1120 {0, 1024, "invalid initial priority group number"},
1121 };
1122
1123 int r;
1124 struct multipath *m;
1125 struct dm_arg_set as;
1126 unsigned pg_count = 0;
1127 unsigned next_pg_num;
1128
1129 as.argc = argc;
1130 as.argv = argv;
1131
1132 m = alloc_multipath(ti);
1133 if (!m) {
1134 ti->error = "can't allocate multipath";
1135 return -EINVAL;
1136 }
1137
1138 r = parse_features(&as, m);
1139 if (r)
1140 goto bad;
1141
1142 r = alloc_multipath_stage2(ti, m);
1143 if (r)
1144 goto bad;
1145
1146 r = parse_hw_handler(&as, m);
1147 if (r)
1148 goto bad;
1149
1150 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1151 if (r)
1152 goto bad;
1153
1154 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1155 if (r)
1156 goto bad;
1157
1158 if ((!m->nr_priority_groups && next_pg_num) ||
1159 (m->nr_priority_groups && !next_pg_num)) {
1160 ti->error = "invalid initial priority group";
1161 r = -EINVAL;
1162 goto bad;
1163 }
1164
1165 /* parse the priority groups */
1166 while (as.argc) {
1167 struct priority_group *pg;
1168 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1169
1170 pg = parse_priority_group(&as, m);
1171 if (IS_ERR(pg)) {
1172 r = PTR_ERR(pg);
1173 goto bad;
1174 }
1175
1176 nr_valid_paths += pg->nr_pgpaths;
1177 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1178
1179 list_add_tail(&pg->list, &m->priority_groups);
1180 pg_count++;
1181 pg->pg_num = pg_count;
1182 if (!--next_pg_num)
1183 m->next_pg = pg;
1184 }
1185
1186 if (pg_count != m->nr_priority_groups) {
1187 ti->error = "priority group count mismatch";
1188 r = -EINVAL;
1189 goto bad;
1190 }
1191
1192 ti->num_flush_bios = 1;
1193 ti->num_discard_bios = 1;
1194 ti->num_write_same_bios = 1;
1195 ti->num_write_zeroes_bios = 1;
1196 if (m->queue_mode == DM_TYPE_BIO_BASED)
1197 ti->per_io_data_size = multipath_per_bio_data_size();
1198 else
1199 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1200
1201 return 0;
1202
1203 bad:
1204 free_multipath(m);
1205 return r;
1206}
1207
1208static void multipath_wait_for_pg_init_completion(struct multipath *m)
1209{
1210 DEFINE_WAIT(wait);
1211
1212 while (1) {
1213 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1214
1215 if (!atomic_read(&m->pg_init_in_progress))
1216 break;
1217
1218 io_schedule();
1219 }
1220 finish_wait(&m->pg_init_wait, &wait);
1221}
1222
1223static void flush_multipath_work(struct multipath *m)
1224{
1225 if (m->hw_handler_name) {
1226 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1227 smp_mb__after_atomic();
1228
1229 if (atomic_read(&m->pg_init_in_progress))
1230 flush_workqueue(kmpath_handlerd);
1231 multipath_wait_for_pg_init_completion(m);
1232
1233 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1234 smp_mb__after_atomic();
1235 }
1236
1237 if (m->queue_mode == DM_TYPE_BIO_BASED)
1238 flush_work(&m->process_queued_bios);
1239 flush_work(&m->trigger_event);
1240}
1241
1242static void multipath_dtr(struct dm_target *ti)
1243{
1244 struct multipath *m = ti->private;
1245
1246 flush_multipath_work(m);
1247 free_multipath(m);
1248}
1249
1250/*
1251 * Take a path out of use.
1252 */
1253static int fail_path(struct pgpath *pgpath)
1254{
1255 unsigned long flags;
1256 struct multipath *m = pgpath->pg->m;
1257
1258 spin_lock_irqsave(&m->lock, flags);
1259
1260 if (!pgpath->is_active)
1261 goto out;
1262
1263 DMWARN("Failing path %s.", pgpath->path.dev->name);
1264
1265 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1266 pgpath->is_active = false;
1267 pgpath->fail_count++;
1268
1269 atomic_dec(&m->nr_valid_paths);
1270
1271 if (pgpath == m->current_pgpath)
1272 m->current_pgpath = NULL;
1273
1274 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1275 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1276
1277 schedule_work(&m->trigger_event);
1278
1279out:
1280 spin_unlock_irqrestore(&m->lock, flags);
1281
1282 return 0;
1283}
1284
1285/*
1286 * Reinstate a previously-failed path
1287 */
1288static int reinstate_path(struct pgpath *pgpath)
1289{
1290 int r = 0, run_queue = 0;
1291 unsigned long flags;
1292 struct multipath *m = pgpath->pg->m;
1293 unsigned nr_valid_paths;
1294
1295 spin_lock_irqsave(&m->lock, flags);
1296
1297 if (pgpath->is_active)
1298 goto out;
1299
1300 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1301
1302 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1303 if (r)
1304 goto out;
1305
1306 pgpath->is_active = true;
1307
1308 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1309 if (nr_valid_paths == 1) {
1310 m->current_pgpath = NULL;
1311 run_queue = 1;
1312 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1313 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1314 atomic_inc(&m->pg_init_in_progress);
1315 }
1316
1317 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1318 pgpath->path.dev->name, nr_valid_paths);
1319
1320 schedule_work(&m->trigger_event);
1321
1322out:
1323 spin_unlock_irqrestore(&m->lock, flags);
1324 if (run_queue) {
1325 dm_table_run_md_queue_async(m->ti->table);
1326 process_queued_io_list(m);
1327 }
1328
1329 return r;
1330}
1331
1332/*
1333 * Fail or reinstate all paths that match the provided struct dm_dev.
1334 */
1335static int action_dev(struct multipath *m, struct dm_dev *dev,
1336 action_fn action)
1337{
1338 int r = -EINVAL;
1339 struct pgpath *pgpath;
1340 struct priority_group *pg;
1341
1342 list_for_each_entry(pg, &m->priority_groups, list) {
1343 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1344 if (pgpath->path.dev == dev)
1345 r = action(pgpath);
1346 }
1347 }
1348
1349 return r;
1350}
1351
1352/*
1353 * Temporarily try to avoid having to use the specified PG
1354 */
1355static void bypass_pg(struct multipath *m, struct priority_group *pg,
1356 bool bypassed)
1357{
1358 unsigned long flags;
1359
1360 spin_lock_irqsave(&m->lock, flags);
1361
1362 pg->bypassed = bypassed;
1363 m->current_pgpath = NULL;
1364 m->current_pg = NULL;
1365
1366 spin_unlock_irqrestore(&m->lock, flags);
1367
1368 schedule_work(&m->trigger_event);
1369}
1370
1371/*
1372 * Switch to using the specified PG from the next I/O that gets mapped
1373 */
1374static int switch_pg_num(struct multipath *m, const char *pgstr)
1375{
1376 struct priority_group *pg;
1377 unsigned pgnum;
1378 unsigned long flags;
1379 char dummy;
1380
1381 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1382 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1383 DMWARN("invalid PG number supplied to switch_pg_num");
1384 return -EINVAL;
1385 }
1386
1387 spin_lock_irqsave(&m->lock, flags);
1388 list_for_each_entry(pg, &m->priority_groups, list) {
1389 pg->bypassed = false;
1390 if (--pgnum)
1391 continue;
1392
1393 m->current_pgpath = NULL;
1394 m->current_pg = NULL;
1395 m->next_pg = pg;
1396 }
1397 spin_unlock_irqrestore(&m->lock, flags);
1398
1399 schedule_work(&m->trigger_event);
1400 return 0;
1401}
1402
1403/*
1404 * Set/clear bypassed status of a PG.
1405 * PGs are numbered upwards from 1 in the order they were declared.
1406 */
1407static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1408{
1409 struct priority_group *pg;
1410 unsigned pgnum;
1411 char dummy;
1412
1413 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1414 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1415 DMWARN("invalid PG number supplied to bypass_pg");
1416 return -EINVAL;
1417 }
1418
1419 list_for_each_entry(pg, &m->priority_groups, list) {
1420 if (!--pgnum)
1421 break;
1422 }
1423
1424 bypass_pg(m, pg, bypassed);
1425 return 0;
1426}
1427
1428/*
1429 * Should we retry pg_init immediately?
1430 */
1431static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1432{
1433 unsigned long flags;
1434 bool limit_reached = false;
1435
1436 spin_lock_irqsave(&m->lock, flags);
1437
1438 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1439 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1440 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1441 else
1442 limit_reached = true;
1443
1444 spin_unlock_irqrestore(&m->lock, flags);
1445
1446 return limit_reached;
1447}
1448
1449static void pg_init_done(void *data, int errors)
1450{
1451 struct pgpath *pgpath = data;
1452 struct priority_group *pg = pgpath->pg;
1453 struct multipath *m = pg->m;
1454 unsigned long flags;
1455 bool delay_retry = false;
1456
1457 /* device or driver problems */
1458 switch (errors) {
1459 case SCSI_DH_OK:
1460 break;
1461 case SCSI_DH_NOSYS:
1462 if (!m->hw_handler_name) {
1463 errors = 0;
1464 break;
1465 }
1466 DMERR("Could not failover the device: Handler scsi_dh_%s "
1467 "Error %d.", m->hw_handler_name, errors);
1468 /*
1469 * Fail path for now, so we do not ping pong
1470 */
1471 fail_path(pgpath);
1472 break;
1473 case SCSI_DH_DEV_TEMP_BUSY:
1474 /*
1475 * Probably doing something like FW upgrade on the
1476 * controller so try the other pg.
1477 */
1478 bypass_pg(m, pg, true);
1479 break;
1480 case SCSI_DH_RETRY:
1481 /* Wait before retrying. */
1482 delay_retry = 1;
1483 /* fall through */
1484 case SCSI_DH_IMM_RETRY:
1485 case SCSI_DH_RES_TEMP_UNAVAIL:
1486 if (pg_init_limit_reached(m, pgpath))
1487 fail_path(pgpath);
1488 errors = 0;
1489 break;
1490 case SCSI_DH_DEV_OFFLINED:
1491 default:
1492 /*
1493 * We probably do not want to fail the path for a device
1494 * error, but this is what the old dm did. In future
1495 * patches we can do more advanced handling.
1496 */
1497 fail_path(pgpath);
1498 }
1499
1500 spin_lock_irqsave(&m->lock, flags);
1501 if (errors) {
1502 if (pgpath == m->current_pgpath) {
1503 DMERR("Could not failover device. Error %d.", errors);
1504 m->current_pgpath = NULL;
1505 m->current_pg = NULL;
1506 }
1507 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1508 pg->bypassed = false;
1509
1510 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1511 /* Activations of other paths are still on going */
1512 goto out;
1513
1514 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1515 if (delay_retry)
1516 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1517 else
1518 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1519
1520 if (__pg_init_all_paths(m))
1521 goto out;
1522 }
1523 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1524
1525 process_queued_io_list(m);
1526
1527 /*
1528 * Wake up any thread waiting to suspend.
1529 */
1530 wake_up(&m->pg_init_wait);
1531
1532out:
1533 spin_unlock_irqrestore(&m->lock, flags);
1534}
1535
1536static void activate_or_offline_path(struct pgpath *pgpath)
1537{
1538 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1539
1540 if (pgpath->is_active && !blk_queue_dying(q))
1541 scsi_dh_activate(q, pg_init_done, pgpath);
1542 else
1543 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1544}
1545
1546static void activate_path_work(struct work_struct *work)
1547{
1548 struct pgpath *pgpath =
1549 container_of(work, struct pgpath, activate_path.work);
1550
1551 activate_or_offline_path(pgpath);
1552}
1553
1554static int multipath_end_io(struct dm_target *ti, struct request *clone,
1555 blk_status_t error, union map_info *map_context)
1556{
1557 struct dm_mpath_io *mpio = get_mpio(map_context);
1558 struct pgpath *pgpath = mpio->pgpath;
1559 int r = DM_ENDIO_DONE;
1560
1561 /*
1562 * We don't queue any clone request inside the multipath target
1563 * during end I/O handling, since those clone requests don't have
1564 * bio clones. If we queue them inside the multipath target,
1565 * we need to make bio clones, that requires memory allocation.
1566 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1567 * don't have bio clones.)
1568 * Instead of queueing the clone request here, we queue the original
1569 * request into dm core, which will remake a clone request and
1570 * clone bios for it and resubmit it later.
1571 */
1572 if (error && blk_path_error(error)) {
1573 struct multipath *m = ti->private;
1574
1575 if (error == BLK_STS_RESOURCE)
1576 r = DM_ENDIO_DELAY_REQUEUE;
1577 else
1578 r = DM_ENDIO_REQUEUE;
1579
1580 if (pgpath)
1581 fail_path(pgpath);
1582
1583 if (atomic_read(&m->nr_valid_paths) == 0 &&
1584 !must_push_back_rq(m)) {
1585 if (error == BLK_STS_IOERR)
1586 dm_report_EIO(m);
1587 /* complete with the original error */
1588 r = DM_ENDIO_DONE;
1589 }
1590 }
1591
1592 if (pgpath) {
1593 struct path_selector *ps = &pgpath->pg->ps;
1594
1595 if (ps->type->end_io)
1596 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1597 }
1598
1599 return r;
1600}
1601
1602static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1603 blk_status_t *error)
1604{
1605 struct multipath *m = ti->private;
1606 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1607 struct pgpath *pgpath = mpio->pgpath;
1608 unsigned long flags;
1609 int r = DM_ENDIO_DONE;
1610
1611 if (!*error || !blk_path_error(*error))
1612 goto done;
1613
1614 if (pgpath)
1615 fail_path(pgpath);
1616
1617 if (atomic_read(&m->nr_valid_paths) == 0 &&
1618 !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1619 if (must_push_back_bio(m)) {
1620 r = DM_ENDIO_REQUEUE;
1621 } else {
1622 dm_report_EIO(m);
1623 *error = BLK_STS_IOERR;
1624 }
1625 goto done;
1626 }
1627
1628 spin_lock_irqsave(&m->lock, flags);
1629 bio_list_add(&m->queued_bios, clone);
1630 spin_unlock_irqrestore(&m->lock, flags);
1631 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1632 queue_work(kmultipathd, &m->process_queued_bios);
1633
1634 r = DM_ENDIO_INCOMPLETE;
1635done:
1636 if (pgpath) {
1637 struct path_selector *ps = &pgpath->pg->ps;
1638
1639 if (ps->type->end_io)
1640 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1641 }
1642
1643 return r;
1644}
1645
1646/*
1647 * Suspend can't complete until all the I/O is processed so if
1648 * the last path fails we must error any remaining I/O.
1649 * Note that if the freeze_bdev fails while suspending, the
1650 * queue_if_no_path state is lost - userspace should reset it.
1651 */
1652static void multipath_presuspend(struct dm_target *ti)
1653{
1654 struct multipath *m = ti->private;
1655
1656 queue_if_no_path(m, false, true);
1657}
1658
1659static void multipath_postsuspend(struct dm_target *ti)
1660{
1661 struct multipath *m = ti->private;
1662
1663 mutex_lock(&m->work_mutex);
1664 flush_multipath_work(m);
1665 mutex_unlock(&m->work_mutex);
1666}
1667
1668/*
1669 * Restore the queue_if_no_path setting.
1670 */
1671static void multipath_resume(struct dm_target *ti)
1672{
1673 struct multipath *m = ti->private;
1674 unsigned long flags;
1675
1676 spin_lock_irqsave(&m->lock, flags);
1677 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1678 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1679 spin_unlock_irqrestore(&m->lock, flags);
1680}
1681
1682/*
1683 * Info output has the following format:
1684 * num_multipath_feature_args [multipath_feature_args]*
1685 * num_handler_status_args [handler_status_args]*
1686 * num_groups init_group_number
1687 * [A|D|E num_ps_status_args [ps_status_args]*
1688 * num_paths num_selector_args
1689 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1690 *
1691 * Table output has the following format (identical to the constructor string):
1692 * num_feature_args [features_args]*
1693 * num_handler_args hw_handler [hw_handler_args]*
1694 * num_groups init_group_number
1695 * [priority selector-name num_ps_args [ps_args]*
1696 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1697 */
1698static void multipath_status(struct dm_target *ti, status_type_t type,
1699 unsigned status_flags, char *result, unsigned maxlen)
1700{
1701 int sz = 0;
1702 unsigned long flags;
1703 struct multipath *m = ti->private;
1704 struct priority_group *pg;
1705 struct pgpath *p;
1706 unsigned pg_num;
1707 char state;
1708
1709 spin_lock_irqsave(&m->lock, flags);
1710
1711 /* Features */
1712 if (type == STATUSTYPE_INFO)
1713 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1714 atomic_read(&m->pg_init_count));
1715 else {
1716 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1717 (m->pg_init_retries > 0) * 2 +
1718 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1719 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1720 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1721
1722 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1723 DMEMIT("queue_if_no_path ");
1724 if (m->pg_init_retries)
1725 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1726 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1727 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1728 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1729 DMEMIT("retain_attached_hw_handler ");
1730 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1731 switch(m->queue_mode) {
1732 case DM_TYPE_BIO_BASED:
1733 DMEMIT("queue_mode bio ");
1734 break;
1735 default:
1736 WARN_ON_ONCE(true);
1737 break;
1738 }
1739 }
1740 }
1741
1742 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1743 DMEMIT("0 ");
1744 else
1745 DMEMIT("1 %s ", m->hw_handler_name);
1746
1747 DMEMIT("%u ", m->nr_priority_groups);
1748
1749 if (m->next_pg)
1750 pg_num = m->next_pg->pg_num;
1751 else if (m->current_pg)
1752 pg_num = m->current_pg->pg_num;
1753 else
1754 pg_num = (m->nr_priority_groups ? 1 : 0);
1755
1756 DMEMIT("%u ", pg_num);
1757
1758 switch (type) {
1759 case STATUSTYPE_INFO:
1760 list_for_each_entry(pg, &m->priority_groups, list) {
1761 if (pg->bypassed)
1762 state = 'D'; /* Disabled */
1763 else if (pg == m->current_pg)
1764 state = 'A'; /* Currently Active */
1765 else
1766 state = 'E'; /* Enabled */
1767
1768 DMEMIT("%c ", state);
1769
1770 if (pg->ps.type->status)
1771 sz += pg->ps.type->status(&pg->ps, NULL, type,
1772 result + sz,
1773 maxlen - sz);
1774 else
1775 DMEMIT("0 ");
1776
1777 DMEMIT("%u %u ", pg->nr_pgpaths,
1778 pg->ps.type->info_args);
1779
1780 list_for_each_entry(p, &pg->pgpaths, list) {
1781 DMEMIT("%s %s %u ", p->path.dev->name,
1782 p->is_active ? "A" : "F",
1783 p->fail_count);
1784 if (pg->ps.type->status)
1785 sz += pg->ps.type->status(&pg->ps,
1786 &p->path, type, result + sz,
1787 maxlen - sz);
1788 }
1789 }
1790 break;
1791
1792 case STATUSTYPE_TABLE:
1793 list_for_each_entry(pg, &m->priority_groups, list) {
1794 DMEMIT("%s ", pg->ps.type->name);
1795
1796 if (pg->ps.type->status)
1797 sz += pg->ps.type->status(&pg->ps, NULL, type,
1798 result + sz,
1799 maxlen - sz);
1800 else
1801 DMEMIT("0 ");
1802
1803 DMEMIT("%u %u ", pg->nr_pgpaths,
1804 pg->ps.type->table_args);
1805
1806 list_for_each_entry(p, &pg->pgpaths, list) {
1807 DMEMIT("%s ", p->path.dev->name);
1808 if (pg->ps.type->status)
1809 sz += pg->ps.type->status(&pg->ps,
1810 &p->path, type, result + sz,
1811 maxlen - sz);
1812 }
1813 }
1814 break;
1815 }
1816
1817 spin_unlock_irqrestore(&m->lock, flags);
1818}
1819
1820static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1821 char *result, unsigned maxlen)
1822{
1823 int r = -EINVAL;
1824 struct dm_dev *dev;
1825 struct multipath *m = ti->private;
1826 action_fn action;
1827
1828 mutex_lock(&m->work_mutex);
1829
1830 if (dm_suspended(ti)) {
1831 r = -EBUSY;
1832 goto out;
1833 }
1834
1835 if (argc == 1) {
1836 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1837 r = queue_if_no_path(m, true, false);
1838 goto out;
1839 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1840 r = queue_if_no_path(m, false, false);
1841 goto out;
1842 }
1843 }
1844
1845 if (argc != 2) {
1846 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1847 goto out;
1848 }
1849
1850 if (!strcasecmp(argv[0], "disable_group")) {
1851 r = bypass_pg_num(m, argv[1], true);
1852 goto out;
1853 } else if (!strcasecmp(argv[0], "enable_group")) {
1854 r = bypass_pg_num(m, argv[1], false);
1855 goto out;
1856 } else if (!strcasecmp(argv[0], "switch_group")) {
1857 r = switch_pg_num(m, argv[1]);
1858 goto out;
1859 } else if (!strcasecmp(argv[0], "reinstate_path"))
1860 action = reinstate_path;
1861 else if (!strcasecmp(argv[0], "fail_path"))
1862 action = fail_path;
1863 else {
1864 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1865 goto out;
1866 }
1867
1868 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1869 if (r) {
1870 DMWARN("message: error getting device %s",
1871 argv[1]);
1872 goto out;
1873 }
1874
1875 r = action_dev(m, dev, action);
1876
1877 dm_put_device(ti, dev);
1878
1879out:
1880 mutex_unlock(&m->work_mutex);
1881 return r;
1882}
1883
1884static int multipath_prepare_ioctl(struct dm_target *ti,
1885 struct block_device **bdev)
1886{
1887 struct multipath *m = ti->private;
1888 struct pgpath *current_pgpath;
1889 int r;
1890
1891 current_pgpath = READ_ONCE(m->current_pgpath);
1892 if (!current_pgpath)
1893 current_pgpath = choose_pgpath(m, 0);
1894
1895 if (current_pgpath) {
1896 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1897 *bdev = current_pgpath->path.dev->bdev;
1898 r = 0;
1899 } else {
1900 /* pg_init has not started or completed */
1901 r = -ENOTCONN;
1902 }
1903 } else {
1904 /* No path is available */
1905 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1906 r = -ENOTCONN;
1907 else
1908 r = -EIO;
1909 }
1910
1911 if (r == -ENOTCONN) {
1912 if (!READ_ONCE(m->current_pg)) {
1913 /* Path status changed, redo selection */
1914 (void) choose_pgpath(m, 0);
1915 }
1916 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1917 pg_init_all_paths(m);
1918 dm_table_run_md_queue_async(m->ti->table);
1919 process_queued_io_list(m);
1920 }
1921
1922 /*
1923 * Only pass ioctls through if the device sizes match exactly.
1924 */
1925 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1926 return 1;
1927 return r;
1928}
1929
1930static int multipath_iterate_devices(struct dm_target *ti,
1931 iterate_devices_callout_fn fn, void *data)
1932{
1933 struct multipath *m = ti->private;
1934 struct priority_group *pg;
1935 struct pgpath *p;
1936 int ret = 0;
1937
1938 list_for_each_entry(pg, &m->priority_groups, list) {
1939 list_for_each_entry(p, &pg->pgpaths, list) {
1940 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1941 if (ret)
1942 goto out;
1943 }
1944 }
1945
1946out:
1947 return ret;
1948}
1949
1950static int pgpath_busy(struct pgpath *pgpath)
1951{
1952 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1953
1954 return blk_lld_busy(q);
1955}
1956
1957/*
1958 * We return "busy", only when we can map I/Os but underlying devices
1959 * are busy (so even if we map I/Os now, the I/Os will wait on
1960 * the underlying queue).
1961 * In other words, if we want to kill I/Os or queue them inside us
1962 * due to map unavailability, we don't return "busy". Otherwise,
1963 * dm core won't give us the I/Os and we can't do what we want.
1964 */
1965static int multipath_busy(struct dm_target *ti)
1966{
1967 bool busy = false, has_active = false;
1968 struct multipath *m = ti->private;
1969 struct priority_group *pg, *next_pg;
1970 struct pgpath *pgpath;
1971
1972 /* pg_init in progress */
1973 if (atomic_read(&m->pg_init_in_progress))
1974 return true;
1975
1976 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1977 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1978 return (m->queue_mode != DM_TYPE_REQUEST_BASED);
1979
1980 /* Guess which priority_group will be used at next mapping time */
1981 pg = READ_ONCE(m->current_pg);
1982 next_pg = READ_ONCE(m->next_pg);
1983 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1984 pg = next_pg;
1985
1986 if (!pg) {
1987 /*
1988 * We don't know which pg will be used at next mapping time.
1989 * We don't call choose_pgpath() here to avoid to trigger
1990 * pg_init just by busy checking.
1991 * So we don't know whether underlying devices we will be using
1992 * at next mapping time are busy or not. Just try mapping.
1993 */
1994 return busy;
1995 }
1996
1997 /*
1998 * If there is one non-busy active path at least, the path selector
1999 * will be able to select it. So we consider such a pg as not busy.
2000 */
2001 busy = true;
2002 list_for_each_entry(pgpath, &pg->pgpaths, list) {
2003 if (pgpath->is_active) {
2004 has_active = true;
2005 if (!pgpath_busy(pgpath)) {
2006 busy = false;
2007 break;
2008 }
2009 }
2010 }
2011
2012 if (!has_active) {
2013 /*
2014 * No active path in this pg, so this pg won't be used and
2015 * the current_pg will be changed at next mapping time.
2016 * We need to try mapping to determine it.
2017 */
2018 busy = false;
2019 }
2020
2021 return busy;
2022}
2023
2024/*-----------------------------------------------------------------
2025 * Module setup
2026 *---------------------------------------------------------------*/
2027static struct target_type multipath_target = {
2028 .name = "multipath",
2029 .version = {1, 13, 0},
2030 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2031 DM_TARGET_PASSES_INTEGRITY,
2032 .module = THIS_MODULE,
2033 .ctr = multipath_ctr,
2034 .dtr = multipath_dtr,
2035 .clone_and_map_rq = multipath_clone_and_map,
2036 .release_clone_rq = multipath_release_clone,
2037 .rq_end_io = multipath_end_io,
2038 .map = multipath_map_bio,
2039 .end_io = multipath_end_io_bio,
2040 .presuspend = multipath_presuspend,
2041 .postsuspend = multipath_postsuspend,
2042 .resume = multipath_resume,
2043 .status = multipath_status,
2044 .message = multipath_message,
2045 .prepare_ioctl = multipath_prepare_ioctl,
2046 .iterate_devices = multipath_iterate_devices,
2047 .busy = multipath_busy,
2048};
2049
2050static int __init dm_multipath_init(void)
2051{
2052 int r;
2053
2054 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2055 if (!kmultipathd) {
2056 DMERR("failed to create workqueue kmpathd");
2057 r = -ENOMEM;
2058 goto bad_alloc_kmultipathd;
2059 }
2060
2061 /*
2062 * A separate workqueue is used to handle the device handlers
2063 * to avoid overloading existing workqueue. Overloading the
2064 * old workqueue would also create a bottleneck in the
2065 * path of the storage hardware device activation.
2066 */
2067 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2068 WQ_MEM_RECLAIM);
2069 if (!kmpath_handlerd) {
2070 DMERR("failed to create workqueue kmpath_handlerd");
2071 r = -ENOMEM;
2072 goto bad_alloc_kmpath_handlerd;
2073 }
2074
2075 r = dm_register_target(&multipath_target);
2076 if (r < 0) {
2077 DMERR("request-based register failed %d", r);
2078 r = -EINVAL;
2079 goto bad_register_target;
2080 }
2081
2082 return 0;
2083
2084bad_register_target:
2085 destroy_workqueue(kmpath_handlerd);
2086bad_alloc_kmpath_handlerd:
2087 destroy_workqueue(kmultipathd);
2088bad_alloc_kmultipathd:
2089 return r;
2090}
2091
2092static void __exit dm_multipath_exit(void)
2093{
2094 destroy_workqueue(kmpath_handlerd);
2095 destroy_workqueue(kmultipathd);
2096
2097 dm_unregister_target(&multipath_target);
2098}
2099
2100module_init(dm_multipath_init);
2101module_exit(dm_multipath_exit);
2102
2103MODULE_DESCRIPTION(DM_NAME " multipath target");
2104MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2105MODULE_LICENSE("GPL");