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 <scsi/scsi_dh.h>
22#include <linux/atomic.h>
23
24#define DM_MSG_PREFIX "multipath"
25#define DM_PG_INIT_DELAY_MSECS 2000
26#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
27
28/* Path properties */
29struct pgpath {
30 struct list_head list;
31
32 struct priority_group *pg; /* Owning PG */
33 unsigned is_active; /* Path status */
34 unsigned fail_count; /* Cumulative failure count */
35
36 struct dm_path path;
37 struct delayed_work activate_path;
38};
39
40#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
41
42/*
43 * Paths are grouped into Priority Groups and numbered from 1 upwards.
44 * Each has a path selector which controls which path gets used.
45 */
46struct priority_group {
47 struct list_head list;
48
49 struct multipath *m; /* Owning multipath instance */
50 struct path_selector ps;
51
52 unsigned pg_num; /* Reference number */
53 unsigned bypassed; /* Temporarily bypass this PG? */
54
55 unsigned nr_pgpaths; /* Number of paths in PG */
56 struct list_head pgpaths;
57};
58
59/* Multipath context */
60struct multipath {
61 struct list_head list;
62 struct dm_target *ti;
63
64 spinlock_t lock;
65
66 const char *hw_handler_name;
67 char *hw_handler_params;
68
69 unsigned nr_priority_groups;
70 struct list_head priority_groups;
71
72 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
73
74 unsigned pg_init_required; /* pg_init needs calling? */
75 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
76 unsigned pg_init_delay_retry; /* Delay pg_init retry? */
77
78 unsigned nr_valid_paths; /* Total number of usable paths */
79 struct pgpath *current_pgpath;
80 struct priority_group *current_pg;
81 struct priority_group *next_pg; /* Switch to this PG if set */
82 unsigned repeat_count; /* I/Os left before calling PS again */
83
84 unsigned queue_io; /* Must we queue all I/O? */
85 unsigned queue_if_no_path; /* Queue I/O if last path fails? */
86 unsigned saved_queue_if_no_path;/* Saved state during suspension */
87 unsigned pg_init_retries; /* Number of times to retry pg_init */
88 unsigned pg_init_count; /* Number of times pg_init called */
89 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
90
91 struct work_struct process_queued_ios;
92 struct list_head queued_ios;
93 unsigned queue_size;
94
95 struct work_struct trigger_event;
96
97 /*
98 * We must use a mempool of dm_mpath_io structs so that we
99 * can resubmit bios on error.
100 */
101 mempool_t *mpio_pool;
102
103 struct mutex work_mutex;
104};
105
106/*
107 * Context information attached to each bio we process.
108 */
109struct dm_mpath_io {
110 struct pgpath *pgpath;
111 size_t nr_bytes;
112};
113
114typedef int (*action_fn) (struct pgpath *pgpath);
115
116#define MIN_IOS 256 /* Mempool size */
117
118static struct kmem_cache *_mpio_cache;
119
120static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
121static void process_queued_ios(struct work_struct *work);
122static void trigger_event(struct work_struct *work);
123static void activate_path(struct work_struct *work);
124
125
126/*-----------------------------------------------
127 * Allocation routines
128 *-----------------------------------------------*/
129
130static struct pgpath *alloc_pgpath(void)
131{
132 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
133
134 if (pgpath) {
135 pgpath->is_active = 1;
136 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
137 }
138
139 return pgpath;
140}
141
142static void free_pgpath(struct pgpath *pgpath)
143{
144 kfree(pgpath);
145}
146
147static struct priority_group *alloc_priority_group(void)
148{
149 struct priority_group *pg;
150
151 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
152
153 if (pg)
154 INIT_LIST_HEAD(&pg->pgpaths);
155
156 return pg;
157}
158
159static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
160{
161 struct pgpath *pgpath, *tmp;
162 struct multipath *m = ti->private;
163
164 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
165 list_del(&pgpath->list);
166 if (m->hw_handler_name)
167 scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
168 dm_put_device(ti, pgpath->path.dev);
169 free_pgpath(pgpath);
170 }
171}
172
173static void free_priority_group(struct priority_group *pg,
174 struct dm_target *ti)
175{
176 struct path_selector *ps = &pg->ps;
177
178 if (ps->type) {
179 ps->type->destroy(ps);
180 dm_put_path_selector(ps->type);
181 }
182
183 free_pgpaths(&pg->pgpaths, ti);
184 kfree(pg);
185}
186
187static struct multipath *alloc_multipath(struct dm_target *ti)
188{
189 struct multipath *m;
190
191 m = kzalloc(sizeof(*m), GFP_KERNEL);
192 if (m) {
193 INIT_LIST_HEAD(&m->priority_groups);
194 INIT_LIST_HEAD(&m->queued_ios);
195 spin_lock_init(&m->lock);
196 m->queue_io = 1;
197 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
198 INIT_WORK(&m->process_queued_ios, process_queued_ios);
199 INIT_WORK(&m->trigger_event, trigger_event);
200 init_waitqueue_head(&m->pg_init_wait);
201 mutex_init(&m->work_mutex);
202 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
203 if (!m->mpio_pool) {
204 kfree(m);
205 return NULL;
206 }
207 m->ti = ti;
208 ti->private = m;
209 }
210
211 return m;
212}
213
214static void free_multipath(struct multipath *m)
215{
216 struct priority_group *pg, *tmp;
217
218 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
219 list_del(&pg->list);
220 free_priority_group(pg, m->ti);
221 }
222
223 kfree(m->hw_handler_name);
224 kfree(m->hw_handler_params);
225 mempool_destroy(m->mpio_pool);
226 kfree(m);
227}
228
229
230/*-----------------------------------------------
231 * Path selection
232 *-----------------------------------------------*/
233
234static void __pg_init_all_paths(struct multipath *m)
235{
236 struct pgpath *pgpath;
237 unsigned long pg_init_delay = 0;
238
239 m->pg_init_count++;
240 m->pg_init_required = 0;
241 if (m->pg_init_delay_retry)
242 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
243 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
244 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
245 /* Skip failed paths */
246 if (!pgpath->is_active)
247 continue;
248 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
249 pg_init_delay))
250 m->pg_init_in_progress++;
251 }
252}
253
254static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
255{
256 m->current_pg = pgpath->pg;
257
258 /* Must we initialise the PG first, and queue I/O till it's ready? */
259 if (m->hw_handler_name) {
260 m->pg_init_required = 1;
261 m->queue_io = 1;
262 } else {
263 m->pg_init_required = 0;
264 m->queue_io = 0;
265 }
266
267 m->pg_init_count = 0;
268}
269
270static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
271 size_t nr_bytes)
272{
273 struct dm_path *path;
274
275 path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
276 if (!path)
277 return -ENXIO;
278
279 m->current_pgpath = path_to_pgpath(path);
280
281 if (m->current_pg != pg)
282 __switch_pg(m, m->current_pgpath);
283
284 return 0;
285}
286
287static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
288{
289 struct priority_group *pg;
290 unsigned bypassed = 1;
291
292 if (!m->nr_valid_paths)
293 goto failed;
294
295 /* Were we instructed to switch PG? */
296 if (m->next_pg) {
297 pg = m->next_pg;
298 m->next_pg = NULL;
299 if (!__choose_path_in_pg(m, pg, nr_bytes))
300 return;
301 }
302
303 /* Don't change PG until it has no remaining paths */
304 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
305 return;
306
307 /*
308 * Loop through priority groups until we find a valid path.
309 * First time we skip PGs marked 'bypassed'.
310 * Second time we only try the ones we skipped.
311 */
312 do {
313 list_for_each_entry(pg, &m->priority_groups, list) {
314 if (pg->bypassed == bypassed)
315 continue;
316 if (!__choose_path_in_pg(m, pg, nr_bytes))
317 return;
318 }
319 } while (bypassed--);
320
321failed:
322 m->current_pgpath = NULL;
323 m->current_pg = NULL;
324}
325
326/*
327 * Check whether bios must be queued in the device-mapper core rather
328 * than here in the target.
329 *
330 * m->lock must be held on entry.
331 *
332 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
333 * same value then we are not between multipath_presuspend()
334 * and multipath_resume() calls and we have no need to check
335 * for the DMF_NOFLUSH_SUSPENDING flag.
336 */
337static int __must_push_back(struct multipath *m)
338{
339 return (m->queue_if_no_path != m->saved_queue_if_no_path &&
340 dm_noflush_suspending(m->ti));
341}
342
343static int map_io(struct multipath *m, struct request *clone,
344 struct dm_mpath_io *mpio, unsigned was_queued)
345{
346 int r = DM_MAPIO_REMAPPED;
347 size_t nr_bytes = blk_rq_bytes(clone);
348 unsigned long flags;
349 struct pgpath *pgpath;
350 struct block_device *bdev;
351
352 spin_lock_irqsave(&m->lock, flags);
353
354 /* Do we need to select a new pgpath? */
355 if (!m->current_pgpath ||
356 (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
357 __choose_pgpath(m, nr_bytes);
358
359 pgpath = m->current_pgpath;
360
361 if (was_queued)
362 m->queue_size--;
363
364 if ((pgpath && m->queue_io) ||
365 (!pgpath && m->queue_if_no_path)) {
366 /* Queue for the daemon to resubmit */
367 list_add_tail(&clone->queuelist, &m->queued_ios);
368 m->queue_size++;
369 if ((m->pg_init_required && !m->pg_init_in_progress) ||
370 !m->queue_io)
371 queue_work(kmultipathd, &m->process_queued_ios);
372 pgpath = NULL;
373 r = DM_MAPIO_SUBMITTED;
374 } else if (pgpath) {
375 bdev = pgpath->path.dev->bdev;
376 clone->q = bdev_get_queue(bdev);
377 clone->rq_disk = bdev->bd_disk;
378 } else if (__must_push_back(m))
379 r = DM_MAPIO_REQUEUE;
380 else
381 r = -EIO; /* Failed */
382
383 mpio->pgpath = pgpath;
384 mpio->nr_bytes = nr_bytes;
385
386 if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
387 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
388 nr_bytes);
389
390 spin_unlock_irqrestore(&m->lock, flags);
391
392 return r;
393}
394
395/*
396 * If we run out of usable paths, should we queue I/O or error it?
397 */
398static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
399 unsigned save_old_value)
400{
401 unsigned long flags;
402
403 spin_lock_irqsave(&m->lock, flags);
404
405 if (save_old_value)
406 m->saved_queue_if_no_path = m->queue_if_no_path;
407 else
408 m->saved_queue_if_no_path = queue_if_no_path;
409 m->queue_if_no_path = queue_if_no_path;
410 if (!m->queue_if_no_path && m->queue_size)
411 queue_work(kmultipathd, &m->process_queued_ios);
412
413 spin_unlock_irqrestore(&m->lock, flags);
414
415 return 0;
416}
417
418/*-----------------------------------------------------------------
419 * The multipath daemon is responsible for resubmitting queued ios.
420 *---------------------------------------------------------------*/
421
422static void dispatch_queued_ios(struct multipath *m)
423{
424 int r;
425 unsigned long flags;
426 struct dm_mpath_io *mpio;
427 union map_info *info;
428 struct request *clone, *n;
429 LIST_HEAD(cl);
430
431 spin_lock_irqsave(&m->lock, flags);
432 list_splice_init(&m->queued_ios, &cl);
433 spin_unlock_irqrestore(&m->lock, flags);
434
435 list_for_each_entry_safe(clone, n, &cl, queuelist) {
436 list_del_init(&clone->queuelist);
437
438 info = dm_get_rq_mapinfo(clone);
439 mpio = info->ptr;
440
441 r = map_io(m, clone, mpio, 1);
442 if (r < 0) {
443 mempool_free(mpio, m->mpio_pool);
444 dm_kill_unmapped_request(clone, r);
445 } else if (r == DM_MAPIO_REMAPPED)
446 dm_dispatch_request(clone);
447 else if (r == DM_MAPIO_REQUEUE) {
448 mempool_free(mpio, m->mpio_pool);
449 dm_requeue_unmapped_request(clone);
450 }
451 }
452}
453
454static void process_queued_ios(struct work_struct *work)
455{
456 struct multipath *m =
457 container_of(work, struct multipath, process_queued_ios);
458 struct pgpath *pgpath = NULL;
459 unsigned must_queue = 1;
460 unsigned long flags;
461
462 spin_lock_irqsave(&m->lock, flags);
463
464 if (!m->queue_size)
465 goto out;
466
467 if (!m->current_pgpath)
468 __choose_pgpath(m, 0);
469
470 pgpath = m->current_pgpath;
471
472 if ((pgpath && !m->queue_io) ||
473 (!pgpath && !m->queue_if_no_path))
474 must_queue = 0;
475
476 if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
477 __pg_init_all_paths(m);
478
479out:
480 spin_unlock_irqrestore(&m->lock, flags);
481 if (!must_queue)
482 dispatch_queued_ios(m);
483}
484
485/*
486 * An event is triggered whenever a path is taken out of use.
487 * Includes path failure and PG bypass.
488 */
489static void trigger_event(struct work_struct *work)
490{
491 struct multipath *m =
492 container_of(work, struct multipath, trigger_event);
493
494 dm_table_event(m->ti->table);
495}
496
497/*-----------------------------------------------------------------
498 * Constructor/argument parsing:
499 * <#multipath feature args> [<arg>]*
500 * <#hw_handler args> [hw_handler [<arg>]*]
501 * <#priority groups>
502 * <initial priority group>
503 * [<selector> <#selector args> [<arg>]*
504 * <#paths> <#per-path selector args>
505 * [<path> [<arg>]* ]+ ]+
506 *---------------------------------------------------------------*/
507static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
508 struct dm_target *ti)
509{
510 int r;
511 struct path_selector_type *pst;
512 unsigned ps_argc;
513
514 static struct dm_arg _args[] = {
515 {0, 1024, "invalid number of path selector args"},
516 };
517
518 pst = dm_get_path_selector(dm_shift_arg(as));
519 if (!pst) {
520 ti->error = "unknown path selector type";
521 return -EINVAL;
522 }
523
524 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
525 if (r) {
526 dm_put_path_selector(pst);
527 return -EINVAL;
528 }
529
530 r = pst->create(&pg->ps, ps_argc, as->argv);
531 if (r) {
532 dm_put_path_selector(pst);
533 ti->error = "path selector constructor failed";
534 return r;
535 }
536
537 pg->ps.type = pst;
538 dm_consume_args(as, ps_argc);
539
540 return 0;
541}
542
543static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
544 struct dm_target *ti)
545{
546 int r;
547 struct pgpath *p;
548 struct multipath *m = ti->private;
549
550 /* we need at least a path arg */
551 if (as->argc < 1) {
552 ti->error = "no device given";
553 return ERR_PTR(-EINVAL);
554 }
555
556 p = alloc_pgpath();
557 if (!p)
558 return ERR_PTR(-ENOMEM);
559
560 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
561 &p->path.dev);
562 if (r) {
563 ti->error = "error getting device";
564 goto bad;
565 }
566
567 if (m->hw_handler_name) {
568 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
569
570 r = scsi_dh_attach(q, m->hw_handler_name);
571 if (r == -EBUSY) {
572 /*
573 * Already attached to different hw_handler,
574 * try to reattach with correct one.
575 */
576 scsi_dh_detach(q);
577 r = scsi_dh_attach(q, m->hw_handler_name);
578 }
579
580 if (r < 0) {
581 ti->error = "error attaching hardware handler";
582 dm_put_device(ti, p->path.dev);
583 goto bad;
584 }
585
586 if (m->hw_handler_params) {
587 r = scsi_dh_set_params(q, m->hw_handler_params);
588 if (r < 0) {
589 ti->error = "unable to set hardware "
590 "handler parameters";
591 scsi_dh_detach(q);
592 dm_put_device(ti, p->path.dev);
593 goto bad;
594 }
595 }
596 }
597
598 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
599 if (r) {
600 dm_put_device(ti, p->path.dev);
601 goto bad;
602 }
603
604 return p;
605
606 bad:
607 free_pgpath(p);
608 return ERR_PTR(r);
609}
610
611static struct priority_group *parse_priority_group(struct dm_arg_set *as,
612 struct multipath *m)
613{
614 static struct dm_arg _args[] = {
615 {1, 1024, "invalid number of paths"},
616 {0, 1024, "invalid number of selector args"}
617 };
618
619 int r;
620 unsigned i, nr_selector_args, nr_args;
621 struct priority_group *pg;
622 struct dm_target *ti = m->ti;
623
624 if (as->argc < 2) {
625 as->argc = 0;
626 ti->error = "not enough priority group arguments";
627 return ERR_PTR(-EINVAL);
628 }
629
630 pg = alloc_priority_group();
631 if (!pg) {
632 ti->error = "couldn't allocate priority group";
633 return ERR_PTR(-ENOMEM);
634 }
635 pg->m = m;
636
637 r = parse_path_selector(as, pg, ti);
638 if (r)
639 goto bad;
640
641 /*
642 * read the paths
643 */
644 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
645 if (r)
646 goto bad;
647
648 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
649 if (r)
650 goto bad;
651
652 nr_args = 1 + nr_selector_args;
653 for (i = 0; i < pg->nr_pgpaths; i++) {
654 struct pgpath *pgpath;
655 struct dm_arg_set path_args;
656
657 if (as->argc < nr_args) {
658 ti->error = "not enough path parameters";
659 r = -EINVAL;
660 goto bad;
661 }
662
663 path_args.argc = nr_args;
664 path_args.argv = as->argv;
665
666 pgpath = parse_path(&path_args, &pg->ps, ti);
667 if (IS_ERR(pgpath)) {
668 r = PTR_ERR(pgpath);
669 goto bad;
670 }
671
672 pgpath->pg = pg;
673 list_add_tail(&pgpath->list, &pg->pgpaths);
674 dm_consume_args(as, nr_args);
675 }
676
677 return pg;
678
679 bad:
680 free_priority_group(pg, ti);
681 return ERR_PTR(r);
682}
683
684static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
685{
686 unsigned hw_argc;
687 int ret;
688 struct dm_target *ti = m->ti;
689
690 static struct dm_arg _args[] = {
691 {0, 1024, "invalid number of hardware handler args"},
692 };
693
694 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
695 return -EINVAL;
696
697 if (!hw_argc)
698 return 0;
699
700 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
701 request_module("scsi_dh_%s", m->hw_handler_name);
702 if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
703 ti->error = "unknown hardware handler type";
704 ret = -EINVAL;
705 goto fail;
706 }
707
708 if (hw_argc > 1) {
709 char *p;
710 int i, j, len = 4;
711
712 for (i = 0; i <= hw_argc - 2; i++)
713 len += strlen(as->argv[i]) + 1;
714 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
715 if (!p) {
716 ti->error = "memory allocation failed";
717 ret = -ENOMEM;
718 goto fail;
719 }
720 j = sprintf(p, "%d", hw_argc - 1);
721 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
722 j = sprintf(p, "%s", as->argv[i]);
723 }
724 dm_consume_args(as, hw_argc - 1);
725
726 return 0;
727fail:
728 kfree(m->hw_handler_name);
729 m->hw_handler_name = NULL;
730 return ret;
731}
732
733static int parse_features(struct dm_arg_set *as, struct multipath *m)
734{
735 int r;
736 unsigned argc;
737 struct dm_target *ti = m->ti;
738 const char *arg_name;
739
740 static struct dm_arg _args[] = {
741 {0, 5, "invalid number of feature args"},
742 {1, 50, "pg_init_retries must be between 1 and 50"},
743 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
744 };
745
746 r = dm_read_arg_group(_args, as, &argc, &ti->error);
747 if (r)
748 return -EINVAL;
749
750 if (!argc)
751 return 0;
752
753 do {
754 arg_name = dm_shift_arg(as);
755 argc--;
756
757 if (!strcasecmp(arg_name, "queue_if_no_path")) {
758 r = queue_if_no_path(m, 1, 0);
759 continue;
760 }
761
762 if (!strcasecmp(arg_name, "pg_init_retries") &&
763 (argc >= 1)) {
764 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
765 argc--;
766 continue;
767 }
768
769 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
770 (argc >= 1)) {
771 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
772 argc--;
773 continue;
774 }
775
776 ti->error = "Unrecognised multipath feature request";
777 r = -EINVAL;
778 } while (argc && !r);
779
780 return r;
781}
782
783static int multipath_ctr(struct dm_target *ti, unsigned int argc,
784 char **argv)
785{
786 /* target arguments */
787 static struct dm_arg _args[] = {
788 {0, 1024, "invalid number of priority groups"},
789 {0, 1024, "invalid initial priority group number"},
790 };
791
792 int r;
793 struct multipath *m;
794 struct dm_arg_set as;
795 unsigned pg_count = 0;
796 unsigned next_pg_num;
797
798 as.argc = argc;
799 as.argv = argv;
800
801 m = alloc_multipath(ti);
802 if (!m) {
803 ti->error = "can't allocate multipath";
804 return -EINVAL;
805 }
806
807 r = parse_features(&as, m);
808 if (r)
809 goto bad;
810
811 r = parse_hw_handler(&as, m);
812 if (r)
813 goto bad;
814
815 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
816 if (r)
817 goto bad;
818
819 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
820 if (r)
821 goto bad;
822
823 if ((!m->nr_priority_groups && next_pg_num) ||
824 (m->nr_priority_groups && !next_pg_num)) {
825 ti->error = "invalid initial priority group";
826 r = -EINVAL;
827 goto bad;
828 }
829
830 /* parse the priority groups */
831 while (as.argc) {
832 struct priority_group *pg;
833
834 pg = parse_priority_group(&as, m);
835 if (IS_ERR(pg)) {
836 r = PTR_ERR(pg);
837 goto bad;
838 }
839
840 m->nr_valid_paths += pg->nr_pgpaths;
841 list_add_tail(&pg->list, &m->priority_groups);
842 pg_count++;
843 pg->pg_num = pg_count;
844 if (!--next_pg_num)
845 m->next_pg = pg;
846 }
847
848 if (pg_count != m->nr_priority_groups) {
849 ti->error = "priority group count mismatch";
850 r = -EINVAL;
851 goto bad;
852 }
853
854 ti->num_flush_requests = 1;
855 ti->num_discard_requests = 1;
856
857 return 0;
858
859 bad:
860 free_multipath(m);
861 return r;
862}
863
864static void multipath_wait_for_pg_init_completion(struct multipath *m)
865{
866 DECLARE_WAITQUEUE(wait, current);
867 unsigned long flags;
868
869 add_wait_queue(&m->pg_init_wait, &wait);
870
871 while (1) {
872 set_current_state(TASK_UNINTERRUPTIBLE);
873
874 spin_lock_irqsave(&m->lock, flags);
875 if (!m->pg_init_in_progress) {
876 spin_unlock_irqrestore(&m->lock, flags);
877 break;
878 }
879 spin_unlock_irqrestore(&m->lock, flags);
880
881 io_schedule();
882 }
883 set_current_state(TASK_RUNNING);
884
885 remove_wait_queue(&m->pg_init_wait, &wait);
886}
887
888static void flush_multipath_work(struct multipath *m)
889{
890 flush_workqueue(kmpath_handlerd);
891 multipath_wait_for_pg_init_completion(m);
892 flush_workqueue(kmultipathd);
893 flush_work_sync(&m->trigger_event);
894}
895
896static void multipath_dtr(struct dm_target *ti)
897{
898 struct multipath *m = ti->private;
899
900 flush_multipath_work(m);
901 free_multipath(m);
902}
903
904/*
905 * Map cloned requests
906 */
907static int multipath_map(struct dm_target *ti, struct request *clone,
908 union map_info *map_context)
909{
910 int r;
911 struct dm_mpath_io *mpio;
912 struct multipath *m = (struct multipath *) ti->private;
913
914 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
915 if (!mpio)
916 /* ENOMEM, requeue */
917 return DM_MAPIO_REQUEUE;
918 memset(mpio, 0, sizeof(*mpio));
919
920 map_context->ptr = mpio;
921 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
922 r = map_io(m, clone, mpio, 0);
923 if (r < 0 || r == DM_MAPIO_REQUEUE)
924 mempool_free(mpio, m->mpio_pool);
925
926 return r;
927}
928
929/*
930 * Take a path out of use.
931 */
932static int fail_path(struct pgpath *pgpath)
933{
934 unsigned long flags;
935 struct multipath *m = pgpath->pg->m;
936
937 spin_lock_irqsave(&m->lock, flags);
938
939 if (!pgpath->is_active)
940 goto out;
941
942 DMWARN("Failing path %s.", pgpath->path.dev->name);
943
944 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
945 pgpath->is_active = 0;
946 pgpath->fail_count++;
947
948 m->nr_valid_paths--;
949
950 if (pgpath == m->current_pgpath)
951 m->current_pgpath = NULL;
952
953 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
954 pgpath->path.dev->name, m->nr_valid_paths);
955
956 schedule_work(&m->trigger_event);
957
958out:
959 spin_unlock_irqrestore(&m->lock, flags);
960
961 return 0;
962}
963
964/*
965 * Reinstate a previously-failed path
966 */
967static int reinstate_path(struct pgpath *pgpath)
968{
969 int r = 0;
970 unsigned long flags;
971 struct multipath *m = pgpath->pg->m;
972
973 spin_lock_irqsave(&m->lock, flags);
974
975 if (pgpath->is_active)
976 goto out;
977
978 if (!pgpath->pg->ps.type->reinstate_path) {
979 DMWARN("Reinstate path not supported by path selector %s",
980 pgpath->pg->ps.type->name);
981 r = -EINVAL;
982 goto out;
983 }
984
985 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
986 if (r)
987 goto out;
988
989 pgpath->is_active = 1;
990
991 if (!m->nr_valid_paths++ && m->queue_size) {
992 m->current_pgpath = NULL;
993 queue_work(kmultipathd, &m->process_queued_ios);
994 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
995 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
996 m->pg_init_in_progress++;
997 }
998
999 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1000 pgpath->path.dev->name, m->nr_valid_paths);
1001
1002 schedule_work(&m->trigger_event);
1003
1004out:
1005 spin_unlock_irqrestore(&m->lock, flags);
1006
1007 return r;
1008}
1009
1010/*
1011 * Fail or reinstate all paths that match the provided struct dm_dev.
1012 */
1013static int action_dev(struct multipath *m, struct dm_dev *dev,
1014 action_fn action)
1015{
1016 int r = -EINVAL;
1017 struct pgpath *pgpath;
1018 struct priority_group *pg;
1019
1020 list_for_each_entry(pg, &m->priority_groups, list) {
1021 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1022 if (pgpath->path.dev == dev)
1023 r = action(pgpath);
1024 }
1025 }
1026
1027 return r;
1028}
1029
1030/*
1031 * Temporarily try to avoid having to use the specified PG
1032 */
1033static void bypass_pg(struct multipath *m, struct priority_group *pg,
1034 int bypassed)
1035{
1036 unsigned long flags;
1037
1038 spin_lock_irqsave(&m->lock, flags);
1039
1040 pg->bypassed = bypassed;
1041 m->current_pgpath = NULL;
1042 m->current_pg = NULL;
1043
1044 spin_unlock_irqrestore(&m->lock, flags);
1045
1046 schedule_work(&m->trigger_event);
1047}
1048
1049/*
1050 * Switch to using the specified PG from the next I/O that gets mapped
1051 */
1052static int switch_pg_num(struct multipath *m, const char *pgstr)
1053{
1054 struct priority_group *pg;
1055 unsigned pgnum;
1056 unsigned long flags;
1057
1058 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1059 (pgnum > m->nr_priority_groups)) {
1060 DMWARN("invalid PG number supplied to switch_pg_num");
1061 return -EINVAL;
1062 }
1063
1064 spin_lock_irqsave(&m->lock, flags);
1065 list_for_each_entry(pg, &m->priority_groups, list) {
1066 pg->bypassed = 0;
1067 if (--pgnum)
1068 continue;
1069
1070 m->current_pgpath = NULL;
1071 m->current_pg = NULL;
1072 m->next_pg = pg;
1073 }
1074 spin_unlock_irqrestore(&m->lock, flags);
1075
1076 schedule_work(&m->trigger_event);
1077 return 0;
1078}
1079
1080/*
1081 * Set/clear bypassed status of a PG.
1082 * PGs are numbered upwards from 1 in the order they were declared.
1083 */
1084static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1085{
1086 struct priority_group *pg;
1087 unsigned pgnum;
1088
1089 if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1090 (pgnum > m->nr_priority_groups)) {
1091 DMWARN("invalid PG number supplied to bypass_pg");
1092 return -EINVAL;
1093 }
1094
1095 list_for_each_entry(pg, &m->priority_groups, list) {
1096 if (!--pgnum)
1097 break;
1098 }
1099
1100 bypass_pg(m, pg, bypassed);
1101 return 0;
1102}
1103
1104/*
1105 * Should we retry pg_init immediately?
1106 */
1107static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1108{
1109 unsigned long flags;
1110 int limit_reached = 0;
1111
1112 spin_lock_irqsave(&m->lock, flags);
1113
1114 if (m->pg_init_count <= m->pg_init_retries)
1115 m->pg_init_required = 1;
1116 else
1117 limit_reached = 1;
1118
1119 spin_unlock_irqrestore(&m->lock, flags);
1120
1121 return limit_reached;
1122}
1123
1124static void pg_init_done(void *data, int errors)
1125{
1126 struct pgpath *pgpath = data;
1127 struct priority_group *pg = pgpath->pg;
1128 struct multipath *m = pg->m;
1129 unsigned long flags;
1130 unsigned delay_retry = 0;
1131
1132 /* device or driver problems */
1133 switch (errors) {
1134 case SCSI_DH_OK:
1135 break;
1136 case SCSI_DH_NOSYS:
1137 if (!m->hw_handler_name) {
1138 errors = 0;
1139 break;
1140 }
1141 DMERR("Could not failover the device: Handler scsi_dh_%s "
1142 "Error %d.", m->hw_handler_name, errors);
1143 /*
1144 * Fail path for now, so we do not ping pong
1145 */
1146 fail_path(pgpath);
1147 break;
1148 case SCSI_DH_DEV_TEMP_BUSY:
1149 /*
1150 * Probably doing something like FW upgrade on the
1151 * controller so try the other pg.
1152 */
1153 bypass_pg(m, pg, 1);
1154 break;
1155 case SCSI_DH_RETRY:
1156 /* Wait before retrying. */
1157 delay_retry = 1;
1158 case SCSI_DH_IMM_RETRY:
1159 case SCSI_DH_RES_TEMP_UNAVAIL:
1160 if (pg_init_limit_reached(m, pgpath))
1161 fail_path(pgpath);
1162 errors = 0;
1163 break;
1164 default:
1165 /*
1166 * We probably do not want to fail the path for a device
1167 * error, but this is what the old dm did. In future
1168 * patches we can do more advanced handling.
1169 */
1170 fail_path(pgpath);
1171 }
1172
1173 spin_lock_irqsave(&m->lock, flags);
1174 if (errors) {
1175 if (pgpath == m->current_pgpath) {
1176 DMERR("Could not failover device. Error %d.", errors);
1177 m->current_pgpath = NULL;
1178 m->current_pg = NULL;
1179 }
1180 } else if (!m->pg_init_required)
1181 pg->bypassed = 0;
1182
1183 if (--m->pg_init_in_progress)
1184 /* Activations of other paths are still on going */
1185 goto out;
1186
1187 if (!m->pg_init_required)
1188 m->queue_io = 0;
1189
1190 m->pg_init_delay_retry = delay_retry;
1191 queue_work(kmultipathd, &m->process_queued_ios);
1192
1193 /*
1194 * Wake up any thread waiting to suspend.
1195 */
1196 wake_up(&m->pg_init_wait);
1197
1198out:
1199 spin_unlock_irqrestore(&m->lock, flags);
1200}
1201
1202static void activate_path(struct work_struct *work)
1203{
1204 struct pgpath *pgpath =
1205 container_of(work, struct pgpath, activate_path.work);
1206
1207 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1208 pg_init_done, pgpath);
1209}
1210
1211/*
1212 * end_io handling
1213 */
1214static int do_end_io(struct multipath *m, struct request *clone,
1215 int error, struct dm_mpath_io *mpio)
1216{
1217 /*
1218 * We don't queue any clone request inside the multipath target
1219 * during end I/O handling, since those clone requests don't have
1220 * bio clones. If we queue them inside the multipath target,
1221 * we need to make bio clones, that requires memory allocation.
1222 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1223 * don't have bio clones.)
1224 * Instead of queueing the clone request here, we queue the original
1225 * request into dm core, which will remake a clone request and
1226 * clone bios for it and resubmit it later.
1227 */
1228 int r = DM_ENDIO_REQUEUE;
1229 unsigned long flags;
1230
1231 if (!error && !clone->errors)
1232 return 0; /* I/O complete */
1233
1234 if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1235 return error;
1236
1237 if (mpio->pgpath)
1238 fail_path(mpio->pgpath);
1239
1240 spin_lock_irqsave(&m->lock, flags);
1241 if (!m->nr_valid_paths) {
1242 if (!m->queue_if_no_path) {
1243 if (!__must_push_back(m))
1244 r = -EIO;
1245 } else {
1246 if (error == -EBADE)
1247 r = error;
1248 }
1249 }
1250 spin_unlock_irqrestore(&m->lock, flags);
1251
1252 return r;
1253}
1254
1255static int multipath_end_io(struct dm_target *ti, struct request *clone,
1256 int error, union map_info *map_context)
1257{
1258 struct multipath *m = ti->private;
1259 struct dm_mpath_io *mpio = map_context->ptr;
1260 struct pgpath *pgpath = mpio->pgpath;
1261 struct path_selector *ps;
1262 int r;
1263
1264 r = do_end_io(m, clone, error, mpio);
1265 if (pgpath) {
1266 ps = &pgpath->pg->ps;
1267 if (ps->type->end_io)
1268 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1269 }
1270 mempool_free(mpio, m->mpio_pool);
1271
1272 return r;
1273}
1274
1275/*
1276 * Suspend can't complete until all the I/O is processed so if
1277 * the last path fails we must error any remaining I/O.
1278 * Note that if the freeze_bdev fails while suspending, the
1279 * queue_if_no_path state is lost - userspace should reset it.
1280 */
1281static void multipath_presuspend(struct dm_target *ti)
1282{
1283 struct multipath *m = (struct multipath *) ti->private;
1284
1285 queue_if_no_path(m, 0, 1);
1286}
1287
1288static void multipath_postsuspend(struct dm_target *ti)
1289{
1290 struct multipath *m = ti->private;
1291
1292 mutex_lock(&m->work_mutex);
1293 flush_multipath_work(m);
1294 mutex_unlock(&m->work_mutex);
1295}
1296
1297/*
1298 * Restore the queue_if_no_path setting.
1299 */
1300static void multipath_resume(struct dm_target *ti)
1301{
1302 struct multipath *m = (struct multipath *) ti->private;
1303 unsigned long flags;
1304
1305 spin_lock_irqsave(&m->lock, flags);
1306 m->queue_if_no_path = m->saved_queue_if_no_path;
1307 spin_unlock_irqrestore(&m->lock, flags);
1308}
1309
1310/*
1311 * Info output has the following format:
1312 * num_multipath_feature_args [multipath_feature_args]*
1313 * num_handler_status_args [handler_status_args]*
1314 * num_groups init_group_number
1315 * [A|D|E num_ps_status_args [ps_status_args]*
1316 * num_paths num_selector_args
1317 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1318 *
1319 * Table output has the following format (identical to the constructor string):
1320 * num_feature_args [features_args]*
1321 * num_handler_args hw_handler [hw_handler_args]*
1322 * num_groups init_group_number
1323 * [priority selector-name num_ps_args [ps_args]*
1324 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1325 */
1326static int multipath_status(struct dm_target *ti, status_type_t type,
1327 char *result, unsigned int maxlen)
1328{
1329 int sz = 0;
1330 unsigned long flags;
1331 struct multipath *m = (struct multipath *) ti->private;
1332 struct priority_group *pg;
1333 struct pgpath *p;
1334 unsigned pg_num;
1335 char state;
1336
1337 spin_lock_irqsave(&m->lock, flags);
1338
1339 /* Features */
1340 if (type == STATUSTYPE_INFO)
1341 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1342 else {
1343 DMEMIT("%u ", m->queue_if_no_path +
1344 (m->pg_init_retries > 0) * 2 +
1345 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2);
1346 if (m->queue_if_no_path)
1347 DMEMIT("queue_if_no_path ");
1348 if (m->pg_init_retries)
1349 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1350 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1351 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1352 }
1353
1354 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1355 DMEMIT("0 ");
1356 else
1357 DMEMIT("1 %s ", m->hw_handler_name);
1358
1359 DMEMIT("%u ", m->nr_priority_groups);
1360
1361 if (m->next_pg)
1362 pg_num = m->next_pg->pg_num;
1363 else if (m->current_pg)
1364 pg_num = m->current_pg->pg_num;
1365 else
1366 pg_num = (m->nr_priority_groups ? 1 : 0);
1367
1368 DMEMIT("%u ", pg_num);
1369
1370 switch (type) {
1371 case STATUSTYPE_INFO:
1372 list_for_each_entry(pg, &m->priority_groups, list) {
1373 if (pg->bypassed)
1374 state = 'D'; /* Disabled */
1375 else if (pg == m->current_pg)
1376 state = 'A'; /* Currently Active */
1377 else
1378 state = 'E'; /* Enabled */
1379
1380 DMEMIT("%c ", state);
1381
1382 if (pg->ps.type->status)
1383 sz += pg->ps.type->status(&pg->ps, NULL, type,
1384 result + sz,
1385 maxlen - sz);
1386 else
1387 DMEMIT("0 ");
1388
1389 DMEMIT("%u %u ", pg->nr_pgpaths,
1390 pg->ps.type->info_args);
1391
1392 list_for_each_entry(p, &pg->pgpaths, list) {
1393 DMEMIT("%s %s %u ", p->path.dev->name,
1394 p->is_active ? "A" : "F",
1395 p->fail_count);
1396 if (pg->ps.type->status)
1397 sz += pg->ps.type->status(&pg->ps,
1398 &p->path, type, result + sz,
1399 maxlen - sz);
1400 }
1401 }
1402 break;
1403
1404 case STATUSTYPE_TABLE:
1405 list_for_each_entry(pg, &m->priority_groups, list) {
1406 DMEMIT("%s ", pg->ps.type->name);
1407
1408 if (pg->ps.type->status)
1409 sz += pg->ps.type->status(&pg->ps, NULL, type,
1410 result + sz,
1411 maxlen - sz);
1412 else
1413 DMEMIT("0 ");
1414
1415 DMEMIT("%u %u ", pg->nr_pgpaths,
1416 pg->ps.type->table_args);
1417
1418 list_for_each_entry(p, &pg->pgpaths, list) {
1419 DMEMIT("%s ", p->path.dev->name);
1420 if (pg->ps.type->status)
1421 sz += pg->ps.type->status(&pg->ps,
1422 &p->path, type, result + sz,
1423 maxlen - sz);
1424 }
1425 }
1426 break;
1427 }
1428
1429 spin_unlock_irqrestore(&m->lock, flags);
1430
1431 return 0;
1432}
1433
1434static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1435{
1436 int r = -EINVAL;
1437 struct dm_dev *dev;
1438 struct multipath *m = (struct multipath *) ti->private;
1439 action_fn action;
1440
1441 mutex_lock(&m->work_mutex);
1442
1443 if (dm_suspended(ti)) {
1444 r = -EBUSY;
1445 goto out;
1446 }
1447
1448 if (argc == 1) {
1449 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1450 r = queue_if_no_path(m, 1, 0);
1451 goto out;
1452 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1453 r = queue_if_no_path(m, 0, 0);
1454 goto out;
1455 }
1456 }
1457
1458 if (argc != 2) {
1459 DMWARN("Unrecognised multipath message received.");
1460 goto out;
1461 }
1462
1463 if (!strcasecmp(argv[0], "disable_group")) {
1464 r = bypass_pg_num(m, argv[1], 1);
1465 goto out;
1466 } else if (!strcasecmp(argv[0], "enable_group")) {
1467 r = bypass_pg_num(m, argv[1], 0);
1468 goto out;
1469 } else if (!strcasecmp(argv[0], "switch_group")) {
1470 r = switch_pg_num(m, argv[1]);
1471 goto out;
1472 } else if (!strcasecmp(argv[0], "reinstate_path"))
1473 action = reinstate_path;
1474 else if (!strcasecmp(argv[0], "fail_path"))
1475 action = fail_path;
1476 else {
1477 DMWARN("Unrecognised multipath message received.");
1478 goto out;
1479 }
1480
1481 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1482 if (r) {
1483 DMWARN("message: error getting device %s",
1484 argv[1]);
1485 goto out;
1486 }
1487
1488 r = action_dev(m, dev, action);
1489
1490 dm_put_device(ti, dev);
1491
1492out:
1493 mutex_unlock(&m->work_mutex);
1494 return r;
1495}
1496
1497static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1498 unsigned long arg)
1499{
1500 struct multipath *m = (struct multipath *) ti->private;
1501 struct block_device *bdev = NULL;
1502 fmode_t mode = 0;
1503 unsigned long flags;
1504 int r = 0;
1505
1506 spin_lock_irqsave(&m->lock, flags);
1507
1508 if (!m->current_pgpath)
1509 __choose_pgpath(m, 0);
1510
1511 if (m->current_pgpath) {
1512 bdev = m->current_pgpath->path.dev->bdev;
1513 mode = m->current_pgpath->path.dev->mode;
1514 }
1515
1516 if (m->queue_io)
1517 r = -EAGAIN;
1518 else if (!bdev)
1519 r = -EIO;
1520
1521 spin_unlock_irqrestore(&m->lock, flags);
1522
1523 return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1524}
1525
1526static int multipath_iterate_devices(struct dm_target *ti,
1527 iterate_devices_callout_fn fn, void *data)
1528{
1529 struct multipath *m = ti->private;
1530 struct priority_group *pg;
1531 struct pgpath *p;
1532 int ret = 0;
1533
1534 list_for_each_entry(pg, &m->priority_groups, list) {
1535 list_for_each_entry(p, &pg->pgpaths, list) {
1536 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1537 if (ret)
1538 goto out;
1539 }
1540 }
1541
1542out:
1543 return ret;
1544}
1545
1546static int __pgpath_busy(struct pgpath *pgpath)
1547{
1548 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1549
1550 return dm_underlying_device_busy(q);
1551}
1552
1553/*
1554 * We return "busy", only when we can map I/Os but underlying devices
1555 * are busy (so even if we map I/Os now, the I/Os will wait on
1556 * the underlying queue).
1557 * In other words, if we want to kill I/Os or queue them inside us
1558 * due to map unavailability, we don't return "busy". Otherwise,
1559 * dm core won't give us the I/Os and we can't do what we want.
1560 */
1561static int multipath_busy(struct dm_target *ti)
1562{
1563 int busy = 0, has_active = 0;
1564 struct multipath *m = ti->private;
1565 struct priority_group *pg;
1566 struct pgpath *pgpath;
1567 unsigned long flags;
1568
1569 spin_lock_irqsave(&m->lock, flags);
1570
1571 /* Guess which priority_group will be used at next mapping time */
1572 if (unlikely(!m->current_pgpath && m->next_pg))
1573 pg = m->next_pg;
1574 else if (likely(m->current_pg))
1575 pg = m->current_pg;
1576 else
1577 /*
1578 * We don't know which pg will be used at next mapping time.
1579 * We don't call __choose_pgpath() here to avoid to trigger
1580 * pg_init just by busy checking.
1581 * So we don't know whether underlying devices we will be using
1582 * at next mapping time are busy or not. Just try mapping.
1583 */
1584 goto out;
1585
1586 /*
1587 * If there is one non-busy active path at least, the path selector
1588 * will be able to select it. So we consider such a pg as not busy.
1589 */
1590 busy = 1;
1591 list_for_each_entry(pgpath, &pg->pgpaths, list)
1592 if (pgpath->is_active) {
1593 has_active = 1;
1594
1595 if (!__pgpath_busy(pgpath)) {
1596 busy = 0;
1597 break;
1598 }
1599 }
1600
1601 if (!has_active)
1602 /*
1603 * No active path in this pg, so this pg won't be used and
1604 * the current_pg will be changed at next mapping time.
1605 * We need to try mapping to determine it.
1606 */
1607 busy = 0;
1608
1609out:
1610 spin_unlock_irqrestore(&m->lock, flags);
1611
1612 return busy;
1613}
1614
1615/*-----------------------------------------------------------------
1616 * Module setup
1617 *---------------------------------------------------------------*/
1618static struct target_type multipath_target = {
1619 .name = "multipath",
1620 .version = {1, 3, 0},
1621 .module = THIS_MODULE,
1622 .ctr = multipath_ctr,
1623 .dtr = multipath_dtr,
1624 .map_rq = multipath_map,
1625 .rq_end_io = multipath_end_io,
1626 .presuspend = multipath_presuspend,
1627 .postsuspend = multipath_postsuspend,
1628 .resume = multipath_resume,
1629 .status = multipath_status,
1630 .message = multipath_message,
1631 .ioctl = multipath_ioctl,
1632 .iterate_devices = multipath_iterate_devices,
1633 .busy = multipath_busy,
1634};
1635
1636static int __init dm_multipath_init(void)
1637{
1638 int r;
1639
1640 /* allocate a slab for the dm_ios */
1641 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1642 if (!_mpio_cache)
1643 return -ENOMEM;
1644
1645 r = dm_register_target(&multipath_target);
1646 if (r < 0) {
1647 DMERR("register failed %d", r);
1648 kmem_cache_destroy(_mpio_cache);
1649 return -EINVAL;
1650 }
1651
1652 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1653 if (!kmultipathd) {
1654 DMERR("failed to create workqueue kmpathd");
1655 dm_unregister_target(&multipath_target);
1656 kmem_cache_destroy(_mpio_cache);
1657 return -ENOMEM;
1658 }
1659
1660 /*
1661 * A separate workqueue is used to handle the device handlers
1662 * to avoid overloading existing workqueue. Overloading the
1663 * old workqueue would also create a bottleneck in the
1664 * path of the storage hardware device activation.
1665 */
1666 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1667 WQ_MEM_RECLAIM);
1668 if (!kmpath_handlerd) {
1669 DMERR("failed to create workqueue kmpath_handlerd");
1670 destroy_workqueue(kmultipathd);
1671 dm_unregister_target(&multipath_target);
1672 kmem_cache_destroy(_mpio_cache);
1673 return -ENOMEM;
1674 }
1675
1676 DMINFO("version %u.%u.%u loaded",
1677 multipath_target.version[0], multipath_target.version[1],
1678 multipath_target.version[2]);
1679
1680 return r;
1681}
1682
1683static void __exit dm_multipath_exit(void)
1684{
1685 destroy_workqueue(kmpath_handlerd);
1686 destroy_workqueue(kmultipathd);
1687
1688 dm_unregister_target(&multipath_target);
1689 kmem_cache_destroy(_mpio_cache);
1690}
1691
1692module_init(dm_multipath_init);
1693module_exit(dm_multipath_exit);
1694
1695MODULE_DESCRIPTION(DM_NAME " multipath target");
1696MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1697MODULE_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/timer.h>
24#include <linux/workqueue.h>
25#include <linux/delay.h>
26#include <scsi/scsi_dh.h>
27#include <linux/atomic.h>
28#include <linux/blk-mq.h>
29
30#define DM_MSG_PREFIX "multipath"
31#define DM_PG_INIT_DELAY_MSECS 2000
32#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
33#define QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT 0
34
35static unsigned long queue_if_no_path_timeout_secs = QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT;
36
37/* Path properties */
38struct pgpath {
39 struct list_head list;
40
41 struct priority_group *pg; /* Owning PG */
42 unsigned fail_count; /* Cumulative failure count */
43
44 struct dm_path path;
45 struct delayed_work activate_path;
46
47 bool is_active:1; /* Path status */
48};
49
50#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
51
52/*
53 * Paths are grouped into Priority Groups and numbered from 1 upwards.
54 * Each has a path selector which controls which path gets used.
55 */
56struct priority_group {
57 struct list_head list;
58
59 struct multipath *m; /* Owning multipath instance */
60 struct path_selector ps;
61
62 unsigned pg_num; /* Reference number */
63 unsigned nr_pgpaths; /* Number of paths in PG */
64 struct list_head pgpaths;
65
66 bool bypassed:1; /* Temporarily bypass this PG? */
67};
68
69/* Multipath context */
70struct multipath {
71 unsigned long flags; /* Multipath state flags */
72
73 spinlock_t lock;
74 enum dm_queue_mode queue_mode;
75
76 struct pgpath *current_pgpath;
77 struct priority_group *current_pg;
78 struct priority_group *next_pg; /* Switch to this PG if set */
79
80 atomic_t nr_valid_paths; /* Total number of usable paths */
81 unsigned nr_priority_groups;
82 struct list_head priority_groups;
83
84 const char *hw_handler_name;
85 char *hw_handler_params;
86 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
87 unsigned pg_init_retries; /* Number of times to retry pg_init */
88 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
89 atomic_t pg_init_in_progress; /* Only one pg_init allowed at once */
90 atomic_t pg_init_count; /* Number of times pg_init called */
91
92 struct mutex work_mutex;
93 struct work_struct trigger_event;
94 struct dm_target *ti;
95
96 struct work_struct process_queued_bios;
97 struct bio_list queued_bios;
98
99 struct timer_list nopath_timer; /* Timeout for queue_if_no_path */
100};
101
102/*
103 * Context information attached to each io we process.
104 */
105struct dm_mpath_io {
106 struct pgpath *pgpath;
107 size_t nr_bytes;
108};
109
110typedef int (*action_fn) (struct pgpath *pgpath);
111
112static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
113static void trigger_event(struct work_struct *work);
114static void activate_or_offline_path(struct pgpath *pgpath);
115static void activate_path_work(struct work_struct *work);
116static void process_queued_bios(struct work_struct *work);
117static void queue_if_no_path_timeout_work(struct timer_list *t);
118
119/*-----------------------------------------------
120 * Multipath state flags.
121 *-----------------------------------------------*/
122
123#define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
124#define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
125#define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
126#define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
127#define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
128#define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
129#define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
130
131static bool mpath_double_check_test_bit(int MPATHF_bit, struct multipath *m)
132{
133 bool r = test_bit(MPATHF_bit, &m->flags);
134
135 if (r) {
136 unsigned long flags;
137 spin_lock_irqsave(&m->lock, flags);
138 r = test_bit(MPATHF_bit, &m->flags);
139 spin_unlock_irqrestore(&m->lock, flags);
140 }
141
142 return r;
143}
144
145/*-----------------------------------------------
146 * Allocation routines
147 *-----------------------------------------------*/
148
149static struct pgpath *alloc_pgpath(void)
150{
151 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
152
153 if (!pgpath)
154 return NULL;
155
156 pgpath->is_active = true;
157
158 return pgpath;
159}
160
161static void free_pgpath(struct pgpath *pgpath)
162{
163 kfree(pgpath);
164}
165
166static struct priority_group *alloc_priority_group(void)
167{
168 struct priority_group *pg;
169
170 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
171
172 if (pg)
173 INIT_LIST_HEAD(&pg->pgpaths);
174
175 return pg;
176}
177
178static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
179{
180 struct pgpath *pgpath, *tmp;
181
182 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
183 list_del(&pgpath->list);
184 dm_put_device(ti, pgpath->path.dev);
185 free_pgpath(pgpath);
186 }
187}
188
189static void free_priority_group(struct priority_group *pg,
190 struct dm_target *ti)
191{
192 struct path_selector *ps = &pg->ps;
193
194 if (ps->type) {
195 ps->type->destroy(ps);
196 dm_put_path_selector(ps->type);
197 }
198
199 free_pgpaths(&pg->pgpaths, ti);
200 kfree(pg);
201}
202
203static struct multipath *alloc_multipath(struct dm_target *ti)
204{
205 struct multipath *m;
206
207 m = kzalloc(sizeof(*m), GFP_KERNEL);
208 if (m) {
209 INIT_LIST_HEAD(&m->priority_groups);
210 spin_lock_init(&m->lock);
211 atomic_set(&m->nr_valid_paths, 0);
212 INIT_WORK(&m->trigger_event, trigger_event);
213 mutex_init(&m->work_mutex);
214
215 m->queue_mode = DM_TYPE_NONE;
216
217 m->ti = ti;
218 ti->private = m;
219
220 timer_setup(&m->nopath_timer, queue_if_no_path_timeout_work, 0);
221 }
222
223 return m;
224}
225
226static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
227{
228 if (m->queue_mode == DM_TYPE_NONE) {
229 m->queue_mode = DM_TYPE_REQUEST_BASED;
230 } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
231 INIT_WORK(&m->process_queued_bios, process_queued_bios);
232 /*
233 * bio-based doesn't support any direct scsi_dh management;
234 * it just discovers if a scsi_dh is attached.
235 */
236 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
237 }
238
239 dm_table_set_type(ti->table, m->queue_mode);
240
241 /*
242 * Init fields that are only used when a scsi_dh is attached
243 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
244 */
245 set_bit(MPATHF_QUEUE_IO, &m->flags);
246 atomic_set(&m->pg_init_in_progress, 0);
247 atomic_set(&m->pg_init_count, 0);
248 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
249 init_waitqueue_head(&m->pg_init_wait);
250
251 return 0;
252}
253
254static void free_multipath(struct multipath *m)
255{
256 struct priority_group *pg, *tmp;
257
258 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
259 list_del(&pg->list);
260 free_priority_group(pg, m->ti);
261 }
262
263 kfree(m->hw_handler_name);
264 kfree(m->hw_handler_params);
265 mutex_destroy(&m->work_mutex);
266 kfree(m);
267}
268
269static struct dm_mpath_io *get_mpio(union map_info *info)
270{
271 return info->ptr;
272}
273
274static size_t multipath_per_bio_data_size(void)
275{
276 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
277}
278
279static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
280{
281 return dm_per_bio_data(bio, multipath_per_bio_data_size());
282}
283
284static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
285{
286 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
287 void *bio_details = mpio + 1;
288 return bio_details;
289}
290
291static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
292{
293 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
294 struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
295
296 mpio->nr_bytes = bio->bi_iter.bi_size;
297 mpio->pgpath = NULL;
298 *mpio_p = mpio;
299
300 dm_bio_record(bio_details, bio);
301}
302
303/*-----------------------------------------------
304 * Path selection
305 *-----------------------------------------------*/
306
307static int __pg_init_all_paths(struct multipath *m)
308{
309 struct pgpath *pgpath;
310 unsigned long pg_init_delay = 0;
311
312 lockdep_assert_held(&m->lock);
313
314 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
315 return 0;
316
317 atomic_inc(&m->pg_init_count);
318 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
319
320 /* Check here to reset pg_init_required */
321 if (!m->current_pg)
322 return 0;
323
324 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
325 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
326 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
327 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
328 /* Skip failed paths */
329 if (!pgpath->is_active)
330 continue;
331 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
332 pg_init_delay))
333 atomic_inc(&m->pg_init_in_progress);
334 }
335 return atomic_read(&m->pg_init_in_progress);
336}
337
338static int pg_init_all_paths(struct multipath *m)
339{
340 int ret;
341 unsigned long flags;
342
343 spin_lock_irqsave(&m->lock, flags);
344 ret = __pg_init_all_paths(m);
345 spin_unlock_irqrestore(&m->lock, flags);
346
347 return ret;
348}
349
350static void __switch_pg(struct multipath *m, struct priority_group *pg)
351{
352 lockdep_assert_held(&m->lock);
353
354 m->current_pg = pg;
355
356 /* Must we initialise the PG first, and queue I/O till it's ready? */
357 if (m->hw_handler_name) {
358 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
359 set_bit(MPATHF_QUEUE_IO, &m->flags);
360 } else {
361 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
362 clear_bit(MPATHF_QUEUE_IO, &m->flags);
363 }
364
365 atomic_set(&m->pg_init_count, 0);
366}
367
368static struct pgpath *choose_path_in_pg(struct multipath *m,
369 struct priority_group *pg,
370 size_t nr_bytes)
371{
372 unsigned long flags;
373 struct dm_path *path;
374 struct pgpath *pgpath;
375
376 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
377 if (!path)
378 return ERR_PTR(-ENXIO);
379
380 pgpath = path_to_pgpath(path);
381
382 if (unlikely(READ_ONCE(m->current_pg) != pg)) {
383 /* Only update current_pgpath if pg changed */
384 spin_lock_irqsave(&m->lock, flags);
385 m->current_pgpath = pgpath;
386 __switch_pg(m, pg);
387 spin_unlock_irqrestore(&m->lock, flags);
388 }
389
390 return pgpath;
391}
392
393static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
394{
395 unsigned long flags;
396 struct priority_group *pg;
397 struct pgpath *pgpath;
398 unsigned bypassed = 1;
399
400 if (!atomic_read(&m->nr_valid_paths)) {
401 spin_lock_irqsave(&m->lock, flags);
402 clear_bit(MPATHF_QUEUE_IO, &m->flags);
403 spin_unlock_irqrestore(&m->lock, flags);
404 goto failed;
405 }
406
407 /* Were we instructed to switch PG? */
408 if (READ_ONCE(m->next_pg)) {
409 spin_lock_irqsave(&m->lock, flags);
410 pg = m->next_pg;
411 if (!pg) {
412 spin_unlock_irqrestore(&m->lock, flags);
413 goto check_current_pg;
414 }
415 m->next_pg = NULL;
416 spin_unlock_irqrestore(&m->lock, flags);
417 pgpath = choose_path_in_pg(m, pg, nr_bytes);
418 if (!IS_ERR_OR_NULL(pgpath))
419 return pgpath;
420 }
421
422 /* Don't change PG until it has no remaining paths */
423check_current_pg:
424 pg = READ_ONCE(m->current_pg);
425 if (pg) {
426 pgpath = choose_path_in_pg(m, pg, nr_bytes);
427 if (!IS_ERR_OR_NULL(pgpath))
428 return pgpath;
429 }
430
431 /*
432 * Loop through priority groups until we find a valid path.
433 * First time we skip PGs marked 'bypassed'.
434 * Second time we only try the ones we skipped, but set
435 * pg_init_delay_retry so we do not hammer controllers.
436 */
437 do {
438 list_for_each_entry(pg, &m->priority_groups, list) {
439 if (pg->bypassed == !!bypassed)
440 continue;
441 pgpath = choose_path_in_pg(m, pg, nr_bytes);
442 if (!IS_ERR_OR_NULL(pgpath)) {
443 if (!bypassed) {
444 spin_lock_irqsave(&m->lock, flags);
445 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
446 spin_unlock_irqrestore(&m->lock, flags);
447 }
448 return pgpath;
449 }
450 }
451 } while (bypassed--);
452
453failed:
454 spin_lock_irqsave(&m->lock, flags);
455 m->current_pgpath = NULL;
456 m->current_pg = NULL;
457 spin_unlock_irqrestore(&m->lock, flags);
458
459 return NULL;
460}
461
462/*
463 * dm_report_EIO() is a macro instead of a function to make pr_debug_ratelimited()
464 * report the function name and line number of the function from which
465 * it has been invoked.
466 */
467#define dm_report_EIO(m) \
468do { \
469 struct mapped_device *md = dm_table_get_md((m)->ti->table); \
470 \
471 DMDEBUG_LIMIT("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d", \
472 dm_device_name(md), \
473 test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags), \
474 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
475 dm_noflush_suspending((m)->ti)); \
476} while (0)
477
478/*
479 * Check whether bios must be queued in the device-mapper core rather
480 * than here in the target.
481 */
482static bool __must_push_back(struct multipath *m)
483{
484 return dm_noflush_suspending(m->ti);
485}
486
487static bool must_push_back_rq(struct multipath *m)
488{
489 unsigned long flags;
490 bool ret;
491
492 spin_lock_irqsave(&m->lock, flags);
493 ret = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) || __must_push_back(m));
494 spin_unlock_irqrestore(&m->lock, flags);
495
496 return ret;
497}
498
499/*
500 * Map cloned requests (request-based multipath)
501 */
502static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
503 union map_info *map_context,
504 struct request **__clone)
505{
506 struct multipath *m = ti->private;
507 size_t nr_bytes = blk_rq_bytes(rq);
508 struct pgpath *pgpath;
509 struct block_device *bdev;
510 struct dm_mpath_io *mpio = get_mpio(map_context);
511 struct request_queue *q;
512 struct request *clone;
513
514 /* Do we need to select a new pgpath? */
515 pgpath = READ_ONCE(m->current_pgpath);
516 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
517 pgpath = choose_pgpath(m, nr_bytes);
518
519 if (!pgpath) {
520 if (must_push_back_rq(m))
521 return DM_MAPIO_DELAY_REQUEUE;
522 dm_report_EIO(m); /* Failed */
523 return DM_MAPIO_KILL;
524 } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
525 mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
526 pg_init_all_paths(m);
527 return DM_MAPIO_DELAY_REQUEUE;
528 }
529
530 mpio->pgpath = pgpath;
531 mpio->nr_bytes = nr_bytes;
532
533 bdev = pgpath->path.dev->bdev;
534 q = bdev_get_queue(bdev);
535 clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE,
536 BLK_MQ_REQ_NOWAIT);
537 if (IS_ERR(clone)) {
538 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
539 if (blk_queue_dying(q)) {
540 atomic_inc(&m->pg_init_in_progress);
541 activate_or_offline_path(pgpath);
542 return DM_MAPIO_DELAY_REQUEUE;
543 }
544
545 /*
546 * blk-mq's SCHED_RESTART can cover this requeue, so we
547 * needn't deal with it by DELAY_REQUEUE. More importantly,
548 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
549 * get the queue busy feedback (via BLK_STS_RESOURCE),
550 * otherwise I/O merging can suffer.
551 */
552 return DM_MAPIO_REQUEUE;
553 }
554 clone->bio = clone->biotail = NULL;
555 clone->rq_disk = bdev->bd_disk;
556 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
557 *__clone = clone;
558
559 if (pgpath->pg->ps.type->start_io)
560 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
561 &pgpath->path,
562 nr_bytes);
563 return DM_MAPIO_REMAPPED;
564}
565
566static void multipath_release_clone(struct request *clone,
567 union map_info *map_context)
568{
569 if (unlikely(map_context)) {
570 /*
571 * non-NULL map_context means caller is still map
572 * method; must undo multipath_clone_and_map()
573 */
574 struct dm_mpath_io *mpio = get_mpio(map_context);
575 struct pgpath *pgpath = mpio->pgpath;
576
577 if (pgpath && pgpath->pg->ps.type->end_io)
578 pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
579 &pgpath->path,
580 mpio->nr_bytes,
581 clone->io_start_time_ns);
582 }
583
584 blk_put_request(clone);
585}
586
587/*
588 * Map cloned bios (bio-based multipath)
589 */
590
591static void __multipath_queue_bio(struct multipath *m, struct bio *bio)
592{
593 /* Queue for the daemon to resubmit */
594 bio_list_add(&m->queued_bios, bio);
595 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
596 queue_work(kmultipathd, &m->process_queued_bios);
597}
598
599static void multipath_queue_bio(struct multipath *m, struct bio *bio)
600{
601 unsigned long flags;
602
603 spin_lock_irqsave(&m->lock, flags);
604 __multipath_queue_bio(m, bio);
605 spin_unlock_irqrestore(&m->lock, flags);
606}
607
608static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
609{
610 struct pgpath *pgpath;
611 unsigned long flags;
612
613 /* Do we need to select a new pgpath? */
614 pgpath = READ_ONCE(m->current_pgpath);
615 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
616 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
617
618 if (!pgpath) {
619 spin_lock_irqsave(&m->lock, flags);
620 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
621 __multipath_queue_bio(m, bio);
622 pgpath = ERR_PTR(-EAGAIN);
623 }
624 spin_unlock_irqrestore(&m->lock, flags);
625
626 } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
627 mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
628 multipath_queue_bio(m, bio);
629 pg_init_all_paths(m);
630 return ERR_PTR(-EAGAIN);
631 }
632
633 return pgpath;
634}
635
636static int __multipath_map_bio(struct multipath *m, struct bio *bio,
637 struct dm_mpath_io *mpio)
638{
639 struct pgpath *pgpath = __map_bio(m, bio);
640
641 if (IS_ERR(pgpath))
642 return DM_MAPIO_SUBMITTED;
643
644 if (!pgpath) {
645 if (__must_push_back(m))
646 return DM_MAPIO_REQUEUE;
647 dm_report_EIO(m);
648 return DM_MAPIO_KILL;
649 }
650
651 mpio->pgpath = pgpath;
652
653 bio->bi_status = 0;
654 bio_set_dev(bio, pgpath->path.dev->bdev);
655 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
656
657 if (pgpath->pg->ps.type->start_io)
658 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
659 &pgpath->path,
660 mpio->nr_bytes);
661 return DM_MAPIO_REMAPPED;
662}
663
664static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
665{
666 struct multipath *m = ti->private;
667 struct dm_mpath_io *mpio = NULL;
668
669 multipath_init_per_bio_data(bio, &mpio);
670 return __multipath_map_bio(m, bio, mpio);
671}
672
673static void process_queued_io_list(struct multipath *m)
674{
675 if (m->queue_mode == DM_TYPE_REQUEST_BASED)
676 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
677 else if (m->queue_mode == DM_TYPE_BIO_BASED)
678 queue_work(kmultipathd, &m->process_queued_bios);
679}
680
681static void process_queued_bios(struct work_struct *work)
682{
683 int r;
684 unsigned long flags;
685 struct bio *bio;
686 struct bio_list bios;
687 struct blk_plug plug;
688 struct multipath *m =
689 container_of(work, struct multipath, process_queued_bios);
690
691 bio_list_init(&bios);
692
693 spin_lock_irqsave(&m->lock, flags);
694
695 if (bio_list_empty(&m->queued_bios)) {
696 spin_unlock_irqrestore(&m->lock, flags);
697 return;
698 }
699
700 bio_list_merge(&bios, &m->queued_bios);
701 bio_list_init(&m->queued_bios);
702
703 spin_unlock_irqrestore(&m->lock, flags);
704
705 blk_start_plug(&plug);
706 while ((bio = bio_list_pop(&bios))) {
707 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
708 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
709 r = __multipath_map_bio(m, bio, mpio);
710 switch (r) {
711 case DM_MAPIO_KILL:
712 bio->bi_status = BLK_STS_IOERR;
713 bio_endio(bio);
714 break;
715 case DM_MAPIO_REQUEUE:
716 bio->bi_status = BLK_STS_DM_REQUEUE;
717 bio_endio(bio);
718 break;
719 case DM_MAPIO_REMAPPED:
720 submit_bio_noacct(bio);
721 break;
722 case DM_MAPIO_SUBMITTED:
723 break;
724 default:
725 WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
726 }
727 }
728 blk_finish_plug(&plug);
729}
730
731/*
732 * If we run out of usable paths, should we queue I/O or error it?
733 */
734static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
735 bool save_old_value, const char *caller)
736{
737 unsigned long flags;
738 bool queue_if_no_path_bit, saved_queue_if_no_path_bit;
739 const char *dm_dev_name = dm_device_name(dm_table_get_md(m->ti->table));
740
741 DMDEBUG("%s: %s caller=%s queue_if_no_path=%d save_old_value=%d",
742 dm_dev_name, __func__, caller, queue_if_no_path, save_old_value);
743
744 spin_lock_irqsave(&m->lock, flags);
745
746 queue_if_no_path_bit = test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
747 saved_queue_if_no_path_bit = test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
748
749 if (save_old_value) {
750 if (unlikely(!queue_if_no_path_bit && saved_queue_if_no_path_bit)) {
751 DMERR("%s: QIFNP disabled but saved as enabled, saving again loses state, not saving!",
752 dm_dev_name);
753 } else
754 assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path_bit);
755 } else if (!queue_if_no_path && saved_queue_if_no_path_bit) {
756 /* due to "fail_if_no_path" message, need to honor it. */
757 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
758 }
759 assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
760
761 DMDEBUG("%s: after %s changes; QIFNP = %d; SQIFNP = %d; DNFS = %d",
762 dm_dev_name, __func__,
763 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
764 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
765 dm_noflush_suspending(m->ti));
766
767 spin_unlock_irqrestore(&m->lock, flags);
768
769 if (!queue_if_no_path) {
770 dm_table_run_md_queue_async(m->ti->table);
771 process_queued_io_list(m);
772 }
773
774 return 0;
775}
776
777/*
778 * If the queue_if_no_path timeout fires, turn off queue_if_no_path and
779 * process any queued I/O.
780 */
781static void queue_if_no_path_timeout_work(struct timer_list *t)
782{
783 struct multipath *m = from_timer(m, t, nopath_timer);
784 struct mapped_device *md = dm_table_get_md(m->ti->table);
785
786 DMWARN("queue_if_no_path timeout on %s, failing queued IO", dm_device_name(md));
787 queue_if_no_path(m, false, false, __func__);
788}
789
790/*
791 * Enable the queue_if_no_path timeout if necessary.
792 * Called with m->lock held.
793 */
794static void enable_nopath_timeout(struct multipath *m)
795{
796 unsigned long queue_if_no_path_timeout =
797 READ_ONCE(queue_if_no_path_timeout_secs) * HZ;
798
799 lockdep_assert_held(&m->lock);
800
801 if (queue_if_no_path_timeout > 0 &&
802 atomic_read(&m->nr_valid_paths) == 0 &&
803 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
804 mod_timer(&m->nopath_timer,
805 jiffies + queue_if_no_path_timeout);
806 }
807}
808
809static void disable_nopath_timeout(struct multipath *m)
810{
811 del_timer_sync(&m->nopath_timer);
812}
813
814/*
815 * An event is triggered whenever a path is taken out of use.
816 * Includes path failure and PG bypass.
817 */
818static void trigger_event(struct work_struct *work)
819{
820 struct multipath *m =
821 container_of(work, struct multipath, trigger_event);
822
823 dm_table_event(m->ti->table);
824}
825
826/*-----------------------------------------------------------------
827 * Constructor/argument parsing:
828 * <#multipath feature args> [<arg>]*
829 * <#hw_handler args> [hw_handler [<arg>]*]
830 * <#priority groups>
831 * <initial priority group>
832 * [<selector> <#selector args> [<arg>]*
833 * <#paths> <#per-path selector args>
834 * [<path> [<arg>]* ]+ ]+
835 *---------------------------------------------------------------*/
836static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
837 struct dm_target *ti)
838{
839 int r;
840 struct path_selector_type *pst;
841 unsigned ps_argc;
842
843 static const struct dm_arg _args[] = {
844 {0, 1024, "invalid number of path selector args"},
845 };
846
847 pst = dm_get_path_selector(dm_shift_arg(as));
848 if (!pst) {
849 ti->error = "unknown path selector type";
850 return -EINVAL;
851 }
852
853 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
854 if (r) {
855 dm_put_path_selector(pst);
856 return -EINVAL;
857 }
858
859 r = pst->create(&pg->ps, ps_argc, as->argv);
860 if (r) {
861 dm_put_path_selector(pst);
862 ti->error = "path selector constructor failed";
863 return r;
864 }
865
866 pg->ps.type = pst;
867 dm_consume_args(as, ps_argc);
868
869 return 0;
870}
871
872static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
873 const char **attached_handler_name, char **error)
874{
875 struct request_queue *q = bdev_get_queue(bdev);
876 int r;
877
878 if (mpath_double_check_test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, m)) {
879retain:
880 if (*attached_handler_name) {
881 /*
882 * Clear any hw_handler_params associated with a
883 * handler that isn't already attached.
884 */
885 if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
886 kfree(m->hw_handler_params);
887 m->hw_handler_params = NULL;
888 }
889
890 /*
891 * Reset hw_handler_name to match the attached handler
892 *
893 * NB. This modifies the table line to show the actual
894 * handler instead of the original table passed in.
895 */
896 kfree(m->hw_handler_name);
897 m->hw_handler_name = *attached_handler_name;
898 *attached_handler_name = NULL;
899 }
900 }
901
902 if (m->hw_handler_name) {
903 r = scsi_dh_attach(q, m->hw_handler_name);
904 if (r == -EBUSY) {
905 char b[BDEVNAME_SIZE];
906
907 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
908 bdevname(bdev, b));
909 goto retain;
910 }
911 if (r < 0) {
912 *error = "error attaching hardware handler";
913 return r;
914 }
915
916 if (m->hw_handler_params) {
917 r = scsi_dh_set_params(q, m->hw_handler_params);
918 if (r < 0) {
919 *error = "unable to set hardware handler parameters";
920 return r;
921 }
922 }
923 }
924
925 return 0;
926}
927
928static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
929 struct dm_target *ti)
930{
931 int r;
932 struct pgpath *p;
933 struct multipath *m = ti->private;
934 struct request_queue *q;
935 const char *attached_handler_name = NULL;
936
937 /* we need at least a path arg */
938 if (as->argc < 1) {
939 ti->error = "no device given";
940 return ERR_PTR(-EINVAL);
941 }
942
943 p = alloc_pgpath();
944 if (!p)
945 return ERR_PTR(-ENOMEM);
946
947 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
948 &p->path.dev);
949 if (r) {
950 ti->error = "error getting device";
951 goto bad;
952 }
953
954 q = bdev_get_queue(p->path.dev->bdev);
955 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
956 if (attached_handler_name || m->hw_handler_name) {
957 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
958 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
959 kfree(attached_handler_name);
960 if (r) {
961 dm_put_device(ti, p->path.dev);
962 goto bad;
963 }
964 }
965
966 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
967 if (r) {
968 dm_put_device(ti, p->path.dev);
969 goto bad;
970 }
971
972 return p;
973 bad:
974 free_pgpath(p);
975 return ERR_PTR(r);
976}
977
978static struct priority_group *parse_priority_group(struct dm_arg_set *as,
979 struct multipath *m)
980{
981 static const struct dm_arg _args[] = {
982 {1, 1024, "invalid number of paths"},
983 {0, 1024, "invalid number of selector args"}
984 };
985
986 int r;
987 unsigned i, nr_selector_args, nr_args;
988 struct priority_group *pg;
989 struct dm_target *ti = m->ti;
990
991 if (as->argc < 2) {
992 as->argc = 0;
993 ti->error = "not enough priority group arguments";
994 return ERR_PTR(-EINVAL);
995 }
996
997 pg = alloc_priority_group();
998 if (!pg) {
999 ti->error = "couldn't allocate priority group";
1000 return ERR_PTR(-ENOMEM);
1001 }
1002 pg->m = m;
1003
1004 r = parse_path_selector(as, pg, ti);
1005 if (r)
1006 goto bad;
1007
1008 /*
1009 * read the paths
1010 */
1011 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1012 if (r)
1013 goto bad;
1014
1015 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1016 if (r)
1017 goto bad;
1018
1019 nr_args = 1 + nr_selector_args;
1020 for (i = 0; i < pg->nr_pgpaths; i++) {
1021 struct pgpath *pgpath;
1022 struct dm_arg_set path_args;
1023
1024 if (as->argc < nr_args) {
1025 ti->error = "not enough path parameters";
1026 r = -EINVAL;
1027 goto bad;
1028 }
1029
1030 path_args.argc = nr_args;
1031 path_args.argv = as->argv;
1032
1033 pgpath = parse_path(&path_args, &pg->ps, ti);
1034 if (IS_ERR(pgpath)) {
1035 r = PTR_ERR(pgpath);
1036 goto bad;
1037 }
1038
1039 pgpath->pg = pg;
1040 list_add_tail(&pgpath->list, &pg->pgpaths);
1041 dm_consume_args(as, nr_args);
1042 }
1043
1044 return pg;
1045
1046 bad:
1047 free_priority_group(pg, ti);
1048 return ERR_PTR(r);
1049}
1050
1051static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1052{
1053 unsigned hw_argc;
1054 int ret;
1055 struct dm_target *ti = m->ti;
1056
1057 static const struct dm_arg _args[] = {
1058 {0, 1024, "invalid number of hardware handler args"},
1059 };
1060
1061 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1062 return -EINVAL;
1063
1064 if (!hw_argc)
1065 return 0;
1066
1067 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1068 dm_consume_args(as, hw_argc);
1069 DMERR("bio-based multipath doesn't allow hardware handler args");
1070 return 0;
1071 }
1072
1073 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1074 if (!m->hw_handler_name)
1075 return -EINVAL;
1076
1077 if (hw_argc > 1) {
1078 char *p;
1079 int i, j, len = 4;
1080
1081 for (i = 0; i <= hw_argc - 2; i++)
1082 len += strlen(as->argv[i]) + 1;
1083 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1084 if (!p) {
1085 ti->error = "memory allocation failed";
1086 ret = -ENOMEM;
1087 goto fail;
1088 }
1089 j = sprintf(p, "%d", hw_argc - 1);
1090 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1091 j = sprintf(p, "%s", as->argv[i]);
1092 }
1093 dm_consume_args(as, hw_argc - 1);
1094
1095 return 0;
1096fail:
1097 kfree(m->hw_handler_name);
1098 m->hw_handler_name = NULL;
1099 return ret;
1100}
1101
1102static int parse_features(struct dm_arg_set *as, struct multipath *m)
1103{
1104 int r;
1105 unsigned argc;
1106 struct dm_target *ti = m->ti;
1107 const char *arg_name;
1108
1109 static const struct dm_arg _args[] = {
1110 {0, 8, "invalid number of feature args"},
1111 {1, 50, "pg_init_retries must be between 1 and 50"},
1112 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1113 };
1114
1115 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1116 if (r)
1117 return -EINVAL;
1118
1119 if (!argc)
1120 return 0;
1121
1122 do {
1123 arg_name = dm_shift_arg(as);
1124 argc--;
1125
1126 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1127 r = queue_if_no_path(m, true, false, __func__);
1128 continue;
1129 }
1130
1131 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1132 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1133 continue;
1134 }
1135
1136 if (!strcasecmp(arg_name, "pg_init_retries") &&
1137 (argc >= 1)) {
1138 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1139 argc--;
1140 continue;
1141 }
1142
1143 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1144 (argc >= 1)) {
1145 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1146 argc--;
1147 continue;
1148 }
1149
1150 if (!strcasecmp(arg_name, "queue_mode") &&
1151 (argc >= 1)) {
1152 const char *queue_mode_name = dm_shift_arg(as);
1153
1154 if (!strcasecmp(queue_mode_name, "bio"))
1155 m->queue_mode = DM_TYPE_BIO_BASED;
1156 else if (!strcasecmp(queue_mode_name, "rq") ||
1157 !strcasecmp(queue_mode_name, "mq"))
1158 m->queue_mode = DM_TYPE_REQUEST_BASED;
1159 else {
1160 ti->error = "Unknown 'queue_mode' requested";
1161 r = -EINVAL;
1162 }
1163 argc--;
1164 continue;
1165 }
1166
1167 ti->error = "Unrecognised multipath feature request";
1168 r = -EINVAL;
1169 } while (argc && !r);
1170
1171 return r;
1172}
1173
1174static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1175{
1176 /* target arguments */
1177 static const struct dm_arg _args[] = {
1178 {0, 1024, "invalid number of priority groups"},
1179 {0, 1024, "invalid initial priority group number"},
1180 };
1181
1182 int r;
1183 struct multipath *m;
1184 struct dm_arg_set as;
1185 unsigned pg_count = 0;
1186 unsigned next_pg_num;
1187 unsigned long flags;
1188
1189 as.argc = argc;
1190 as.argv = argv;
1191
1192 m = alloc_multipath(ti);
1193 if (!m) {
1194 ti->error = "can't allocate multipath";
1195 return -EINVAL;
1196 }
1197
1198 r = parse_features(&as, m);
1199 if (r)
1200 goto bad;
1201
1202 r = alloc_multipath_stage2(ti, m);
1203 if (r)
1204 goto bad;
1205
1206 r = parse_hw_handler(&as, m);
1207 if (r)
1208 goto bad;
1209
1210 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1211 if (r)
1212 goto bad;
1213
1214 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1215 if (r)
1216 goto bad;
1217
1218 if ((!m->nr_priority_groups && next_pg_num) ||
1219 (m->nr_priority_groups && !next_pg_num)) {
1220 ti->error = "invalid initial priority group";
1221 r = -EINVAL;
1222 goto bad;
1223 }
1224
1225 /* parse the priority groups */
1226 while (as.argc) {
1227 struct priority_group *pg;
1228 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1229
1230 pg = parse_priority_group(&as, m);
1231 if (IS_ERR(pg)) {
1232 r = PTR_ERR(pg);
1233 goto bad;
1234 }
1235
1236 nr_valid_paths += pg->nr_pgpaths;
1237 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1238
1239 list_add_tail(&pg->list, &m->priority_groups);
1240 pg_count++;
1241 pg->pg_num = pg_count;
1242 if (!--next_pg_num)
1243 m->next_pg = pg;
1244 }
1245
1246 if (pg_count != m->nr_priority_groups) {
1247 ti->error = "priority group count mismatch";
1248 r = -EINVAL;
1249 goto bad;
1250 }
1251
1252 spin_lock_irqsave(&m->lock, flags);
1253 enable_nopath_timeout(m);
1254 spin_unlock_irqrestore(&m->lock, flags);
1255
1256 ti->num_flush_bios = 1;
1257 ti->num_discard_bios = 1;
1258 ti->num_write_same_bios = 1;
1259 ti->num_write_zeroes_bios = 1;
1260 if (m->queue_mode == DM_TYPE_BIO_BASED)
1261 ti->per_io_data_size = multipath_per_bio_data_size();
1262 else
1263 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1264
1265 return 0;
1266
1267 bad:
1268 free_multipath(m);
1269 return r;
1270}
1271
1272static void multipath_wait_for_pg_init_completion(struct multipath *m)
1273{
1274 DEFINE_WAIT(wait);
1275
1276 while (1) {
1277 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1278
1279 if (!atomic_read(&m->pg_init_in_progress))
1280 break;
1281
1282 io_schedule();
1283 }
1284 finish_wait(&m->pg_init_wait, &wait);
1285}
1286
1287static void flush_multipath_work(struct multipath *m)
1288{
1289 if (m->hw_handler_name) {
1290 unsigned long flags;
1291
1292 if (!atomic_read(&m->pg_init_in_progress))
1293 goto skip;
1294
1295 spin_lock_irqsave(&m->lock, flags);
1296 if (atomic_read(&m->pg_init_in_progress) &&
1297 !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1298 spin_unlock_irqrestore(&m->lock, flags);
1299
1300 flush_workqueue(kmpath_handlerd);
1301 multipath_wait_for_pg_init_completion(m);
1302
1303 spin_lock_irqsave(&m->lock, flags);
1304 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1305 }
1306 spin_unlock_irqrestore(&m->lock, flags);
1307 }
1308skip:
1309 if (m->queue_mode == DM_TYPE_BIO_BASED)
1310 flush_work(&m->process_queued_bios);
1311 flush_work(&m->trigger_event);
1312}
1313
1314static void multipath_dtr(struct dm_target *ti)
1315{
1316 struct multipath *m = ti->private;
1317
1318 disable_nopath_timeout(m);
1319 flush_multipath_work(m);
1320 free_multipath(m);
1321}
1322
1323/*
1324 * Take a path out of use.
1325 */
1326static int fail_path(struct pgpath *pgpath)
1327{
1328 unsigned long flags;
1329 struct multipath *m = pgpath->pg->m;
1330
1331 spin_lock_irqsave(&m->lock, flags);
1332
1333 if (!pgpath->is_active)
1334 goto out;
1335
1336 DMWARN("%s: Failing path %s.",
1337 dm_device_name(dm_table_get_md(m->ti->table)),
1338 pgpath->path.dev->name);
1339
1340 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1341 pgpath->is_active = false;
1342 pgpath->fail_count++;
1343
1344 atomic_dec(&m->nr_valid_paths);
1345
1346 if (pgpath == m->current_pgpath)
1347 m->current_pgpath = NULL;
1348
1349 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1350 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1351
1352 schedule_work(&m->trigger_event);
1353
1354 enable_nopath_timeout(m);
1355
1356out:
1357 spin_unlock_irqrestore(&m->lock, flags);
1358
1359 return 0;
1360}
1361
1362/*
1363 * Reinstate a previously-failed path
1364 */
1365static int reinstate_path(struct pgpath *pgpath)
1366{
1367 int r = 0, run_queue = 0;
1368 unsigned long flags;
1369 struct multipath *m = pgpath->pg->m;
1370 unsigned nr_valid_paths;
1371
1372 spin_lock_irqsave(&m->lock, flags);
1373
1374 if (pgpath->is_active)
1375 goto out;
1376
1377 DMWARN("%s: Reinstating path %s.",
1378 dm_device_name(dm_table_get_md(m->ti->table)),
1379 pgpath->path.dev->name);
1380
1381 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1382 if (r)
1383 goto out;
1384
1385 pgpath->is_active = true;
1386
1387 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1388 if (nr_valid_paths == 1) {
1389 m->current_pgpath = NULL;
1390 run_queue = 1;
1391 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1392 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1393 atomic_inc(&m->pg_init_in_progress);
1394 }
1395
1396 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1397 pgpath->path.dev->name, nr_valid_paths);
1398
1399 schedule_work(&m->trigger_event);
1400
1401out:
1402 spin_unlock_irqrestore(&m->lock, flags);
1403 if (run_queue) {
1404 dm_table_run_md_queue_async(m->ti->table);
1405 process_queued_io_list(m);
1406 }
1407
1408 if (pgpath->is_active)
1409 disable_nopath_timeout(m);
1410
1411 return r;
1412}
1413
1414/*
1415 * Fail or reinstate all paths that match the provided struct dm_dev.
1416 */
1417static int action_dev(struct multipath *m, struct dm_dev *dev,
1418 action_fn action)
1419{
1420 int r = -EINVAL;
1421 struct pgpath *pgpath;
1422 struct priority_group *pg;
1423
1424 list_for_each_entry(pg, &m->priority_groups, list) {
1425 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1426 if (pgpath->path.dev == dev)
1427 r = action(pgpath);
1428 }
1429 }
1430
1431 return r;
1432}
1433
1434/*
1435 * Temporarily try to avoid having to use the specified PG
1436 */
1437static void bypass_pg(struct multipath *m, struct priority_group *pg,
1438 bool bypassed)
1439{
1440 unsigned long flags;
1441
1442 spin_lock_irqsave(&m->lock, flags);
1443
1444 pg->bypassed = bypassed;
1445 m->current_pgpath = NULL;
1446 m->current_pg = NULL;
1447
1448 spin_unlock_irqrestore(&m->lock, flags);
1449
1450 schedule_work(&m->trigger_event);
1451}
1452
1453/*
1454 * Switch to using the specified PG from the next I/O that gets mapped
1455 */
1456static int switch_pg_num(struct multipath *m, const char *pgstr)
1457{
1458 struct priority_group *pg;
1459 unsigned pgnum;
1460 unsigned long flags;
1461 char dummy;
1462
1463 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1464 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1465 DMWARN("invalid PG number supplied to switch_pg_num");
1466 return -EINVAL;
1467 }
1468
1469 spin_lock_irqsave(&m->lock, flags);
1470 list_for_each_entry(pg, &m->priority_groups, list) {
1471 pg->bypassed = false;
1472 if (--pgnum)
1473 continue;
1474
1475 m->current_pgpath = NULL;
1476 m->current_pg = NULL;
1477 m->next_pg = pg;
1478 }
1479 spin_unlock_irqrestore(&m->lock, flags);
1480
1481 schedule_work(&m->trigger_event);
1482 return 0;
1483}
1484
1485/*
1486 * Set/clear bypassed status of a PG.
1487 * PGs are numbered upwards from 1 in the order they were declared.
1488 */
1489static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1490{
1491 struct priority_group *pg;
1492 unsigned pgnum;
1493 char dummy;
1494
1495 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1496 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1497 DMWARN("invalid PG number supplied to bypass_pg");
1498 return -EINVAL;
1499 }
1500
1501 list_for_each_entry(pg, &m->priority_groups, list) {
1502 if (!--pgnum)
1503 break;
1504 }
1505
1506 bypass_pg(m, pg, bypassed);
1507 return 0;
1508}
1509
1510/*
1511 * Should we retry pg_init immediately?
1512 */
1513static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1514{
1515 unsigned long flags;
1516 bool limit_reached = false;
1517
1518 spin_lock_irqsave(&m->lock, flags);
1519
1520 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1521 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1522 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1523 else
1524 limit_reached = true;
1525
1526 spin_unlock_irqrestore(&m->lock, flags);
1527
1528 return limit_reached;
1529}
1530
1531static void pg_init_done(void *data, int errors)
1532{
1533 struct pgpath *pgpath = data;
1534 struct priority_group *pg = pgpath->pg;
1535 struct multipath *m = pg->m;
1536 unsigned long flags;
1537 bool delay_retry = false;
1538
1539 /* device or driver problems */
1540 switch (errors) {
1541 case SCSI_DH_OK:
1542 break;
1543 case SCSI_DH_NOSYS:
1544 if (!m->hw_handler_name) {
1545 errors = 0;
1546 break;
1547 }
1548 DMERR("Could not failover the device: Handler scsi_dh_%s "
1549 "Error %d.", m->hw_handler_name, errors);
1550 /*
1551 * Fail path for now, so we do not ping pong
1552 */
1553 fail_path(pgpath);
1554 break;
1555 case SCSI_DH_DEV_TEMP_BUSY:
1556 /*
1557 * Probably doing something like FW upgrade on the
1558 * controller so try the other pg.
1559 */
1560 bypass_pg(m, pg, true);
1561 break;
1562 case SCSI_DH_RETRY:
1563 /* Wait before retrying. */
1564 delay_retry = true;
1565 fallthrough;
1566 case SCSI_DH_IMM_RETRY:
1567 case SCSI_DH_RES_TEMP_UNAVAIL:
1568 if (pg_init_limit_reached(m, pgpath))
1569 fail_path(pgpath);
1570 errors = 0;
1571 break;
1572 case SCSI_DH_DEV_OFFLINED:
1573 default:
1574 /*
1575 * We probably do not want to fail the path for a device
1576 * error, but this is what the old dm did. In future
1577 * patches we can do more advanced handling.
1578 */
1579 fail_path(pgpath);
1580 }
1581
1582 spin_lock_irqsave(&m->lock, flags);
1583 if (errors) {
1584 if (pgpath == m->current_pgpath) {
1585 DMERR("Could not failover device. Error %d.", errors);
1586 m->current_pgpath = NULL;
1587 m->current_pg = NULL;
1588 }
1589 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1590 pg->bypassed = false;
1591
1592 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1593 /* Activations of other paths are still on going */
1594 goto out;
1595
1596 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1597 if (delay_retry)
1598 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1599 else
1600 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1601
1602 if (__pg_init_all_paths(m))
1603 goto out;
1604 }
1605 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1606
1607 process_queued_io_list(m);
1608
1609 /*
1610 * Wake up any thread waiting to suspend.
1611 */
1612 wake_up(&m->pg_init_wait);
1613
1614out:
1615 spin_unlock_irqrestore(&m->lock, flags);
1616}
1617
1618static void activate_or_offline_path(struct pgpath *pgpath)
1619{
1620 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1621
1622 if (pgpath->is_active && !blk_queue_dying(q))
1623 scsi_dh_activate(q, pg_init_done, pgpath);
1624 else
1625 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1626}
1627
1628static void activate_path_work(struct work_struct *work)
1629{
1630 struct pgpath *pgpath =
1631 container_of(work, struct pgpath, activate_path.work);
1632
1633 activate_or_offline_path(pgpath);
1634}
1635
1636static int multipath_end_io(struct dm_target *ti, struct request *clone,
1637 blk_status_t error, union map_info *map_context)
1638{
1639 struct dm_mpath_io *mpio = get_mpio(map_context);
1640 struct pgpath *pgpath = mpio->pgpath;
1641 int r = DM_ENDIO_DONE;
1642
1643 /*
1644 * We don't queue any clone request inside the multipath target
1645 * during end I/O handling, since those clone requests don't have
1646 * bio clones. If we queue them inside the multipath target,
1647 * we need to make bio clones, that requires memory allocation.
1648 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1649 * don't have bio clones.)
1650 * Instead of queueing the clone request here, we queue the original
1651 * request into dm core, which will remake a clone request and
1652 * clone bios for it and resubmit it later.
1653 */
1654 if (error && blk_path_error(error)) {
1655 struct multipath *m = ti->private;
1656
1657 if (error == BLK_STS_RESOURCE)
1658 r = DM_ENDIO_DELAY_REQUEUE;
1659 else
1660 r = DM_ENDIO_REQUEUE;
1661
1662 if (pgpath)
1663 fail_path(pgpath);
1664
1665 if (!atomic_read(&m->nr_valid_paths) &&
1666 !must_push_back_rq(m)) {
1667 if (error == BLK_STS_IOERR)
1668 dm_report_EIO(m);
1669 /* complete with the original error */
1670 r = DM_ENDIO_DONE;
1671 }
1672 }
1673
1674 if (pgpath) {
1675 struct path_selector *ps = &pgpath->pg->ps;
1676
1677 if (ps->type->end_io)
1678 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1679 clone->io_start_time_ns);
1680 }
1681
1682 return r;
1683}
1684
1685static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1686 blk_status_t *error)
1687{
1688 struct multipath *m = ti->private;
1689 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1690 struct pgpath *pgpath = mpio->pgpath;
1691 unsigned long flags;
1692 int r = DM_ENDIO_DONE;
1693
1694 if (!*error || !blk_path_error(*error))
1695 goto done;
1696
1697 if (pgpath)
1698 fail_path(pgpath);
1699
1700 if (!atomic_read(&m->nr_valid_paths)) {
1701 spin_lock_irqsave(&m->lock, flags);
1702 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1703 if (__must_push_back(m)) {
1704 r = DM_ENDIO_REQUEUE;
1705 } else {
1706 dm_report_EIO(m);
1707 *error = BLK_STS_IOERR;
1708 }
1709 spin_unlock_irqrestore(&m->lock, flags);
1710 goto done;
1711 }
1712 spin_unlock_irqrestore(&m->lock, flags);
1713 }
1714
1715 multipath_queue_bio(m, clone);
1716 r = DM_ENDIO_INCOMPLETE;
1717done:
1718 if (pgpath) {
1719 struct path_selector *ps = &pgpath->pg->ps;
1720
1721 if (ps->type->end_io)
1722 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1723 dm_start_time_ns_from_clone(clone));
1724 }
1725
1726 return r;
1727}
1728
1729/*
1730 * Suspend with flush can't complete until all the I/O is processed
1731 * so if the last path fails we must error any remaining I/O.
1732 * - Note that if the freeze_bdev fails while suspending, the
1733 * queue_if_no_path state is lost - userspace should reset it.
1734 * Otherwise, during noflush suspend, queue_if_no_path will not change.
1735 */
1736static void multipath_presuspend(struct dm_target *ti)
1737{
1738 struct multipath *m = ti->private;
1739
1740 /* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1741 if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1742 queue_if_no_path(m, false, true, __func__);
1743}
1744
1745static void multipath_postsuspend(struct dm_target *ti)
1746{
1747 struct multipath *m = ti->private;
1748
1749 mutex_lock(&m->work_mutex);
1750 flush_multipath_work(m);
1751 mutex_unlock(&m->work_mutex);
1752}
1753
1754/*
1755 * Restore the queue_if_no_path setting.
1756 */
1757static void multipath_resume(struct dm_target *ti)
1758{
1759 struct multipath *m = ti->private;
1760 unsigned long flags;
1761
1762 spin_lock_irqsave(&m->lock, flags);
1763 if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1764 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1765 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1766 }
1767
1768 DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1769 dm_device_name(dm_table_get_md(m->ti->table)), __func__,
1770 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1771 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1772
1773 spin_unlock_irqrestore(&m->lock, flags);
1774}
1775
1776/*
1777 * Info output has the following format:
1778 * num_multipath_feature_args [multipath_feature_args]*
1779 * num_handler_status_args [handler_status_args]*
1780 * num_groups init_group_number
1781 * [A|D|E num_ps_status_args [ps_status_args]*
1782 * num_paths num_selector_args
1783 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1784 *
1785 * Table output has the following format (identical to the constructor string):
1786 * num_feature_args [features_args]*
1787 * num_handler_args hw_handler [hw_handler_args]*
1788 * num_groups init_group_number
1789 * [priority selector-name num_ps_args [ps_args]*
1790 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1791 */
1792static void multipath_status(struct dm_target *ti, status_type_t type,
1793 unsigned status_flags, char *result, unsigned maxlen)
1794{
1795 int sz = 0;
1796 unsigned long flags;
1797 struct multipath *m = ti->private;
1798 struct priority_group *pg;
1799 struct pgpath *p;
1800 unsigned pg_num;
1801 char state;
1802
1803 spin_lock_irqsave(&m->lock, flags);
1804
1805 /* Features */
1806 if (type == STATUSTYPE_INFO)
1807 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1808 atomic_read(&m->pg_init_count));
1809 else {
1810 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1811 (m->pg_init_retries > 0) * 2 +
1812 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1813 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1814 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1815
1816 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1817 DMEMIT("queue_if_no_path ");
1818 if (m->pg_init_retries)
1819 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1820 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1821 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1822 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1823 DMEMIT("retain_attached_hw_handler ");
1824 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1825 switch(m->queue_mode) {
1826 case DM_TYPE_BIO_BASED:
1827 DMEMIT("queue_mode bio ");
1828 break;
1829 default:
1830 WARN_ON_ONCE(true);
1831 break;
1832 }
1833 }
1834 }
1835
1836 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1837 DMEMIT("0 ");
1838 else
1839 DMEMIT("1 %s ", m->hw_handler_name);
1840
1841 DMEMIT("%u ", m->nr_priority_groups);
1842
1843 if (m->next_pg)
1844 pg_num = m->next_pg->pg_num;
1845 else if (m->current_pg)
1846 pg_num = m->current_pg->pg_num;
1847 else
1848 pg_num = (m->nr_priority_groups ? 1 : 0);
1849
1850 DMEMIT("%u ", pg_num);
1851
1852 switch (type) {
1853 case STATUSTYPE_INFO:
1854 list_for_each_entry(pg, &m->priority_groups, list) {
1855 if (pg->bypassed)
1856 state = 'D'; /* Disabled */
1857 else if (pg == m->current_pg)
1858 state = 'A'; /* Currently Active */
1859 else
1860 state = 'E'; /* Enabled */
1861
1862 DMEMIT("%c ", state);
1863
1864 if (pg->ps.type->status)
1865 sz += pg->ps.type->status(&pg->ps, NULL, type,
1866 result + sz,
1867 maxlen - sz);
1868 else
1869 DMEMIT("0 ");
1870
1871 DMEMIT("%u %u ", pg->nr_pgpaths,
1872 pg->ps.type->info_args);
1873
1874 list_for_each_entry(p, &pg->pgpaths, list) {
1875 DMEMIT("%s %s %u ", p->path.dev->name,
1876 p->is_active ? "A" : "F",
1877 p->fail_count);
1878 if (pg->ps.type->status)
1879 sz += pg->ps.type->status(&pg->ps,
1880 &p->path, type, result + sz,
1881 maxlen - sz);
1882 }
1883 }
1884 break;
1885
1886 case STATUSTYPE_TABLE:
1887 list_for_each_entry(pg, &m->priority_groups, list) {
1888 DMEMIT("%s ", pg->ps.type->name);
1889
1890 if (pg->ps.type->status)
1891 sz += pg->ps.type->status(&pg->ps, NULL, type,
1892 result + sz,
1893 maxlen - sz);
1894 else
1895 DMEMIT("0 ");
1896
1897 DMEMIT("%u %u ", pg->nr_pgpaths,
1898 pg->ps.type->table_args);
1899
1900 list_for_each_entry(p, &pg->pgpaths, list) {
1901 DMEMIT("%s ", p->path.dev->name);
1902 if (pg->ps.type->status)
1903 sz += pg->ps.type->status(&pg->ps,
1904 &p->path, type, result + sz,
1905 maxlen - sz);
1906 }
1907 }
1908 break;
1909 }
1910
1911 spin_unlock_irqrestore(&m->lock, flags);
1912}
1913
1914static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1915 char *result, unsigned maxlen)
1916{
1917 int r = -EINVAL;
1918 struct dm_dev *dev;
1919 struct multipath *m = ti->private;
1920 action_fn action;
1921 unsigned long flags;
1922
1923 mutex_lock(&m->work_mutex);
1924
1925 if (dm_suspended(ti)) {
1926 r = -EBUSY;
1927 goto out;
1928 }
1929
1930 if (argc == 1) {
1931 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1932 r = queue_if_no_path(m, true, false, __func__);
1933 spin_lock_irqsave(&m->lock, flags);
1934 enable_nopath_timeout(m);
1935 spin_unlock_irqrestore(&m->lock, flags);
1936 goto out;
1937 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1938 r = queue_if_no_path(m, false, false, __func__);
1939 disable_nopath_timeout(m);
1940 goto out;
1941 }
1942 }
1943
1944 if (argc != 2) {
1945 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1946 goto out;
1947 }
1948
1949 if (!strcasecmp(argv[0], "disable_group")) {
1950 r = bypass_pg_num(m, argv[1], true);
1951 goto out;
1952 } else if (!strcasecmp(argv[0], "enable_group")) {
1953 r = bypass_pg_num(m, argv[1], false);
1954 goto out;
1955 } else if (!strcasecmp(argv[0], "switch_group")) {
1956 r = switch_pg_num(m, argv[1]);
1957 goto out;
1958 } else if (!strcasecmp(argv[0], "reinstate_path"))
1959 action = reinstate_path;
1960 else if (!strcasecmp(argv[0], "fail_path"))
1961 action = fail_path;
1962 else {
1963 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1964 goto out;
1965 }
1966
1967 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1968 if (r) {
1969 DMWARN("message: error getting device %s",
1970 argv[1]);
1971 goto out;
1972 }
1973
1974 r = action_dev(m, dev, action);
1975
1976 dm_put_device(ti, dev);
1977
1978out:
1979 mutex_unlock(&m->work_mutex);
1980 return r;
1981}
1982
1983static int multipath_prepare_ioctl(struct dm_target *ti,
1984 struct block_device **bdev)
1985{
1986 struct multipath *m = ti->private;
1987 struct pgpath *pgpath;
1988 unsigned long flags;
1989 int r;
1990
1991 pgpath = READ_ONCE(m->current_pgpath);
1992 if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
1993 pgpath = choose_pgpath(m, 0);
1994
1995 if (pgpath) {
1996 if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
1997 *bdev = pgpath->path.dev->bdev;
1998 r = 0;
1999 } else {
2000 /* pg_init has not started or completed */
2001 r = -ENOTCONN;
2002 }
2003 } else {
2004 /* No path is available */
2005 r = -EIO;
2006 spin_lock_irqsave(&m->lock, flags);
2007 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2008 r = -ENOTCONN;
2009 spin_unlock_irqrestore(&m->lock, flags);
2010 }
2011
2012 if (r == -ENOTCONN) {
2013 if (!READ_ONCE(m->current_pg)) {
2014 /* Path status changed, redo selection */
2015 (void) choose_pgpath(m, 0);
2016 }
2017 spin_lock_irqsave(&m->lock, flags);
2018 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2019 (void) __pg_init_all_paths(m);
2020 spin_unlock_irqrestore(&m->lock, flags);
2021 dm_table_run_md_queue_async(m->ti->table);
2022 process_queued_io_list(m);
2023 }
2024
2025 /*
2026 * Only pass ioctls through if the device sizes match exactly.
2027 */
2028 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
2029 return 1;
2030 return r;
2031}
2032
2033static int multipath_iterate_devices(struct dm_target *ti,
2034 iterate_devices_callout_fn fn, void *data)
2035{
2036 struct multipath *m = ti->private;
2037 struct priority_group *pg;
2038 struct pgpath *p;
2039 int ret = 0;
2040
2041 list_for_each_entry(pg, &m->priority_groups, list) {
2042 list_for_each_entry(p, &pg->pgpaths, list) {
2043 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2044 if (ret)
2045 goto out;
2046 }
2047 }
2048
2049out:
2050 return ret;
2051}
2052
2053static int pgpath_busy(struct pgpath *pgpath)
2054{
2055 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2056
2057 return blk_lld_busy(q);
2058}
2059
2060/*
2061 * We return "busy", only when we can map I/Os but underlying devices
2062 * are busy (so even if we map I/Os now, the I/Os will wait on
2063 * the underlying queue).
2064 * In other words, if we want to kill I/Os or queue them inside us
2065 * due to map unavailability, we don't return "busy". Otherwise,
2066 * dm core won't give us the I/Os and we can't do what we want.
2067 */
2068static int multipath_busy(struct dm_target *ti)
2069{
2070 bool busy = false, has_active = false;
2071 struct multipath *m = ti->private;
2072 struct priority_group *pg, *next_pg;
2073 struct pgpath *pgpath;
2074
2075 /* pg_init in progress */
2076 if (atomic_read(&m->pg_init_in_progress))
2077 return true;
2078
2079 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2080 if (!atomic_read(&m->nr_valid_paths)) {
2081 unsigned long flags;
2082 spin_lock_irqsave(&m->lock, flags);
2083 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2084 spin_unlock_irqrestore(&m->lock, flags);
2085 return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2086 }
2087 spin_unlock_irqrestore(&m->lock, flags);
2088 }
2089
2090 /* Guess which priority_group will be used at next mapping time */
2091 pg = READ_ONCE(m->current_pg);
2092 next_pg = READ_ONCE(m->next_pg);
2093 if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2094 pg = next_pg;
2095
2096 if (!pg) {
2097 /*
2098 * We don't know which pg will be used at next mapping time.
2099 * We don't call choose_pgpath() here to avoid to trigger
2100 * pg_init just by busy checking.
2101 * So we don't know whether underlying devices we will be using
2102 * at next mapping time are busy or not. Just try mapping.
2103 */
2104 return busy;
2105 }
2106
2107 /*
2108 * If there is one non-busy active path at least, the path selector
2109 * will be able to select it. So we consider such a pg as not busy.
2110 */
2111 busy = true;
2112 list_for_each_entry(pgpath, &pg->pgpaths, list) {
2113 if (pgpath->is_active) {
2114 has_active = true;
2115 if (!pgpath_busy(pgpath)) {
2116 busy = false;
2117 break;
2118 }
2119 }
2120 }
2121
2122 if (!has_active) {
2123 /*
2124 * No active path in this pg, so this pg won't be used and
2125 * the current_pg will be changed at next mapping time.
2126 * We need to try mapping to determine it.
2127 */
2128 busy = false;
2129 }
2130
2131 return busy;
2132}
2133
2134/*-----------------------------------------------------------------
2135 * Module setup
2136 *---------------------------------------------------------------*/
2137static struct target_type multipath_target = {
2138 .name = "multipath",
2139 .version = {1, 14, 0},
2140 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2141 DM_TARGET_PASSES_INTEGRITY,
2142 .module = THIS_MODULE,
2143 .ctr = multipath_ctr,
2144 .dtr = multipath_dtr,
2145 .clone_and_map_rq = multipath_clone_and_map,
2146 .release_clone_rq = multipath_release_clone,
2147 .rq_end_io = multipath_end_io,
2148 .map = multipath_map_bio,
2149 .end_io = multipath_end_io_bio,
2150 .presuspend = multipath_presuspend,
2151 .postsuspend = multipath_postsuspend,
2152 .resume = multipath_resume,
2153 .status = multipath_status,
2154 .message = multipath_message,
2155 .prepare_ioctl = multipath_prepare_ioctl,
2156 .iterate_devices = multipath_iterate_devices,
2157 .busy = multipath_busy,
2158};
2159
2160static int __init dm_multipath_init(void)
2161{
2162 int r;
2163
2164 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2165 if (!kmultipathd) {
2166 DMERR("failed to create workqueue kmpathd");
2167 r = -ENOMEM;
2168 goto bad_alloc_kmultipathd;
2169 }
2170
2171 /*
2172 * A separate workqueue is used to handle the device handlers
2173 * to avoid overloading existing workqueue. Overloading the
2174 * old workqueue would also create a bottleneck in the
2175 * path of the storage hardware device activation.
2176 */
2177 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2178 WQ_MEM_RECLAIM);
2179 if (!kmpath_handlerd) {
2180 DMERR("failed to create workqueue kmpath_handlerd");
2181 r = -ENOMEM;
2182 goto bad_alloc_kmpath_handlerd;
2183 }
2184
2185 r = dm_register_target(&multipath_target);
2186 if (r < 0) {
2187 DMERR("request-based register failed %d", r);
2188 r = -EINVAL;
2189 goto bad_register_target;
2190 }
2191
2192 return 0;
2193
2194bad_register_target:
2195 destroy_workqueue(kmpath_handlerd);
2196bad_alloc_kmpath_handlerd:
2197 destroy_workqueue(kmultipathd);
2198bad_alloc_kmultipathd:
2199 return r;
2200}
2201
2202static void __exit dm_multipath_exit(void)
2203{
2204 destroy_workqueue(kmpath_handlerd);
2205 destroy_workqueue(kmultipathd);
2206
2207 dm_unregister_target(&multipath_target);
2208}
2209
2210module_init(dm_multipath_init);
2211module_exit(dm_multipath_exit);
2212
2213module_param_named(queue_if_no_path_timeout_secs,
2214 queue_if_no_path_timeout_secs, ulong, S_IRUGO | S_IWUSR);
2215MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2216
2217MODULE_DESCRIPTION(DM_NAME " multipath target");
2218MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2219MODULE_LICENSE("GPL");