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/workqueue.h>
24#include <linux/delay.h>
25#include <scsi/scsi_dh.h>
26#include <linux/atomic.h>
27#include <linux/blk-mq.h>
28
29#define DM_MSG_PREFIX "multipath"
30#define DM_PG_INIT_DELAY_MSECS 2000
31#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
32
33/* Path properties */
34struct pgpath {
35 struct list_head list;
36
37 struct priority_group *pg; /* Owning PG */
38 unsigned fail_count; /* Cumulative failure count */
39
40 struct dm_path path;
41 struct delayed_work activate_path;
42
43 bool is_active:1; /* Path status */
44};
45
46#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
47
48/*
49 * Paths are grouped into Priority Groups and numbered from 1 upwards.
50 * Each has a path selector which controls which path gets used.
51 */
52struct priority_group {
53 struct list_head list;
54
55 struct multipath *m; /* Owning multipath instance */
56 struct path_selector ps;
57
58 unsigned pg_num; /* Reference number */
59 unsigned nr_pgpaths; /* Number of paths in PG */
60 struct list_head pgpaths;
61
62 bool bypassed:1; /* Temporarily bypass this PG? */
63};
64
65/* Multipath context */
66struct multipath {
67 struct list_head list;
68 struct dm_target *ti;
69
70 const char *hw_handler_name;
71 char *hw_handler_params;
72
73 spinlock_t lock;
74
75 unsigned nr_priority_groups;
76 struct list_head priority_groups;
77
78 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
79
80 struct pgpath *current_pgpath;
81 struct priority_group *current_pg;
82 struct priority_group *next_pg; /* Switch to this PG if set */
83
84 unsigned long flags; /* Multipath state flags */
85
86 unsigned pg_init_retries; /* Number of times to retry pg_init */
87 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
88
89 atomic_t nr_valid_paths; /* Total number of usable paths */
90 atomic_t pg_init_in_progress; /* Only one pg_init allowed at once */
91 atomic_t pg_init_count; /* Number of times pg_init called */
92
93 unsigned queue_mode;
94
95 /*
96 * We must use a mempool of dm_mpath_io structs so that we
97 * can resubmit bios on error.
98 */
99 mempool_t *mpio_pool;
100
101 struct mutex work_mutex;
102 struct work_struct trigger_event;
103
104 struct work_struct process_queued_bios;
105 struct bio_list queued_bios;
106};
107
108/*
109 * Context information attached to each io we process.
110 */
111struct dm_mpath_io {
112 struct pgpath *pgpath;
113 size_t nr_bytes;
114};
115
116typedef int (*action_fn) (struct pgpath *pgpath);
117
118static struct kmem_cache *_mpio_cache;
119
120static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
121static void trigger_event(struct work_struct *work);
122static void activate_path(struct work_struct *work);
123static void process_queued_bios(struct work_struct *work);
124
125/*-----------------------------------------------
126 * Multipath state flags.
127 *-----------------------------------------------*/
128
129#define MPATHF_QUEUE_IO 0 /* Must we queue all I/O? */
130#define MPATHF_QUEUE_IF_NO_PATH 1 /* Queue I/O if last path fails? */
131#define MPATHF_SAVED_QUEUE_IF_NO_PATH 2 /* Saved state during suspension */
132#define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3 /* If there's already a hw_handler present, don't change it. */
133#define MPATHF_PG_INIT_DISABLED 4 /* pg_init is not currently allowed */
134#define MPATHF_PG_INIT_REQUIRED 5 /* pg_init needs calling? */
135#define MPATHF_PG_INIT_DELAY_RETRY 6 /* Delay pg_init retry? */
136
137/*-----------------------------------------------
138 * Allocation routines
139 *-----------------------------------------------*/
140
141static struct pgpath *alloc_pgpath(void)
142{
143 struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
144
145 if (pgpath) {
146 pgpath->is_active = true;
147 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
148 }
149
150 return pgpath;
151}
152
153static void free_pgpath(struct pgpath *pgpath)
154{
155 kfree(pgpath);
156}
157
158static struct priority_group *alloc_priority_group(void)
159{
160 struct priority_group *pg;
161
162 pg = kzalloc(sizeof(*pg), GFP_KERNEL);
163
164 if (pg)
165 INIT_LIST_HEAD(&pg->pgpaths);
166
167 return pg;
168}
169
170static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
171{
172 struct pgpath *pgpath, *tmp;
173
174 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
175 list_del(&pgpath->list);
176 dm_put_device(ti, pgpath->path.dev);
177 free_pgpath(pgpath);
178 }
179}
180
181static void free_priority_group(struct priority_group *pg,
182 struct dm_target *ti)
183{
184 struct path_selector *ps = &pg->ps;
185
186 if (ps->type) {
187 ps->type->destroy(ps);
188 dm_put_path_selector(ps->type);
189 }
190
191 free_pgpaths(&pg->pgpaths, ti);
192 kfree(pg);
193}
194
195static struct multipath *alloc_multipath(struct dm_target *ti)
196{
197 struct multipath *m;
198
199 m = kzalloc(sizeof(*m), GFP_KERNEL);
200 if (m) {
201 INIT_LIST_HEAD(&m->priority_groups);
202 spin_lock_init(&m->lock);
203 set_bit(MPATHF_QUEUE_IO, &m->flags);
204 atomic_set(&m->nr_valid_paths, 0);
205 atomic_set(&m->pg_init_in_progress, 0);
206 atomic_set(&m->pg_init_count, 0);
207 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
208 INIT_WORK(&m->trigger_event, trigger_event);
209 init_waitqueue_head(&m->pg_init_wait);
210 mutex_init(&m->work_mutex);
211
212 m->mpio_pool = NULL;
213 m->queue_mode = DM_TYPE_NONE;
214
215 m->ti = ti;
216 ti->private = m;
217 }
218
219 return m;
220}
221
222static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
223{
224 if (m->queue_mode == DM_TYPE_NONE) {
225 /*
226 * Default to request-based.
227 */
228 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
229 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
230 else
231 m->queue_mode = DM_TYPE_REQUEST_BASED;
232 }
233
234 if (m->queue_mode == DM_TYPE_REQUEST_BASED) {
235 unsigned min_ios = dm_get_reserved_rq_based_ios();
236
237 m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
238 if (!m->mpio_pool)
239 return -ENOMEM;
240 }
241 else if (m->queue_mode == DM_TYPE_BIO_BASED) {
242 INIT_WORK(&m->process_queued_bios, process_queued_bios);
243 /*
244 * bio-based doesn't support any direct scsi_dh management;
245 * it just discovers if a scsi_dh is attached.
246 */
247 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
248 }
249
250 dm_table_set_type(ti->table, m->queue_mode);
251
252 return 0;
253}
254
255static void free_multipath(struct multipath *m)
256{
257 struct priority_group *pg, *tmp;
258
259 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
260 list_del(&pg->list);
261 free_priority_group(pg, m->ti);
262 }
263
264 kfree(m->hw_handler_name);
265 kfree(m->hw_handler_params);
266 mempool_destroy(m->mpio_pool);
267 kfree(m);
268}
269
270static struct dm_mpath_io *get_mpio(union map_info *info)
271{
272 return info->ptr;
273}
274
275static struct dm_mpath_io *set_mpio(struct multipath *m, union map_info *info)
276{
277 struct dm_mpath_io *mpio;
278
279 if (!m->mpio_pool) {
280 /* Use blk-mq pdu memory requested via per_io_data_size */
281 mpio = get_mpio(info);
282 memset(mpio, 0, sizeof(*mpio));
283 return mpio;
284 }
285
286 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
287 if (!mpio)
288 return NULL;
289
290 memset(mpio, 0, sizeof(*mpio));
291 info->ptr = mpio;
292
293 return mpio;
294}
295
296static void clear_request_fn_mpio(struct multipath *m, union map_info *info)
297{
298 /* Only needed for non blk-mq (.request_fn) multipath */
299 if (m->mpio_pool) {
300 struct dm_mpath_io *mpio = info->ptr;
301
302 info->ptr = NULL;
303 mempool_free(mpio, m->mpio_pool);
304 }
305}
306
307static size_t multipath_per_bio_data_size(void)
308{
309 return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
310}
311
312static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
313{
314 return dm_per_bio_data(bio, multipath_per_bio_data_size());
315}
316
317static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
318{
319 /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
320 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
321 void *bio_details = mpio + 1;
322
323 return bio_details;
324}
325
326static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p,
327 struct dm_bio_details **bio_details_p)
328{
329 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
330 struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
331
332 memset(mpio, 0, sizeof(*mpio));
333 memset(bio_details, 0, sizeof(*bio_details));
334 dm_bio_record(bio_details, bio);
335
336 if (mpio_p)
337 *mpio_p = mpio;
338 if (bio_details_p)
339 *bio_details_p = bio_details;
340}
341
342/*-----------------------------------------------
343 * Path selection
344 *-----------------------------------------------*/
345
346static int __pg_init_all_paths(struct multipath *m)
347{
348 struct pgpath *pgpath;
349 unsigned long pg_init_delay = 0;
350
351 if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
352 return 0;
353
354 atomic_inc(&m->pg_init_count);
355 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
356
357 /* Check here to reset pg_init_required */
358 if (!m->current_pg)
359 return 0;
360
361 if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
362 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
363 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
364 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
365 /* Skip failed paths */
366 if (!pgpath->is_active)
367 continue;
368 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
369 pg_init_delay))
370 atomic_inc(&m->pg_init_in_progress);
371 }
372 return atomic_read(&m->pg_init_in_progress);
373}
374
375static void pg_init_all_paths(struct multipath *m)
376{
377 unsigned long flags;
378
379 spin_lock_irqsave(&m->lock, flags);
380 __pg_init_all_paths(m);
381 spin_unlock_irqrestore(&m->lock, flags);
382}
383
384static void __switch_pg(struct multipath *m, struct priority_group *pg)
385{
386 m->current_pg = pg;
387
388 /* Must we initialise the PG first, and queue I/O till it's ready? */
389 if (m->hw_handler_name) {
390 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
391 set_bit(MPATHF_QUEUE_IO, &m->flags);
392 } else {
393 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
394 clear_bit(MPATHF_QUEUE_IO, &m->flags);
395 }
396
397 atomic_set(&m->pg_init_count, 0);
398}
399
400static struct pgpath *choose_path_in_pg(struct multipath *m,
401 struct priority_group *pg,
402 size_t nr_bytes)
403{
404 unsigned long flags;
405 struct dm_path *path;
406 struct pgpath *pgpath;
407
408 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
409 if (!path)
410 return ERR_PTR(-ENXIO);
411
412 pgpath = path_to_pgpath(path);
413
414 if (unlikely(lockless_dereference(m->current_pg) != pg)) {
415 /* Only update current_pgpath if pg changed */
416 spin_lock_irqsave(&m->lock, flags);
417 m->current_pgpath = pgpath;
418 __switch_pg(m, pg);
419 spin_unlock_irqrestore(&m->lock, flags);
420 }
421
422 return pgpath;
423}
424
425static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
426{
427 unsigned long flags;
428 struct priority_group *pg;
429 struct pgpath *pgpath;
430 unsigned bypassed = 1;
431
432 if (!atomic_read(&m->nr_valid_paths)) {
433 clear_bit(MPATHF_QUEUE_IO, &m->flags);
434 goto failed;
435 }
436
437 /* Were we instructed to switch PG? */
438 if (lockless_dereference(m->next_pg)) {
439 spin_lock_irqsave(&m->lock, flags);
440 pg = m->next_pg;
441 if (!pg) {
442 spin_unlock_irqrestore(&m->lock, flags);
443 goto check_current_pg;
444 }
445 m->next_pg = NULL;
446 spin_unlock_irqrestore(&m->lock, flags);
447 pgpath = choose_path_in_pg(m, pg, nr_bytes);
448 if (!IS_ERR_OR_NULL(pgpath))
449 return pgpath;
450 }
451
452 /* Don't change PG until it has no remaining paths */
453check_current_pg:
454 pg = lockless_dereference(m->current_pg);
455 if (pg) {
456 pgpath = choose_path_in_pg(m, pg, nr_bytes);
457 if (!IS_ERR_OR_NULL(pgpath))
458 return pgpath;
459 }
460
461 /*
462 * Loop through priority groups until we find a valid path.
463 * First time we skip PGs marked 'bypassed'.
464 * Second time we only try the ones we skipped, but set
465 * pg_init_delay_retry so we do not hammer controllers.
466 */
467 do {
468 list_for_each_entry(pg, &m->priority_groups, list) {
469 if (pg->bypassed == !!bypassed)
470 continue;
471 pgpath = choose_path_in_pg(m, pg, nr_bytes);
472 if (!IS_ERR_OR_NULL(pgpath)) {
473 if (!bypassed)
474 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
475 return pgpath;
476 }
477 }
478 } while (bypassed--);
479
480failed:
481 spin_lock_irqsave(&m->lock, flags);
482 m->current_pgpath = NULL;
483 m->current_pg = NULL;
484 spin_unlock_irqrestore(&m->lock, flags);
485
486 return NULL;
487}
488
489/*
490 * Check whether bios must be queued in the device-mapper core rather
491 * than here in the target.
492 *
493 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
494 * same value then we are not between multipath_presuspend()
495 * and multipath_resume() calls and we have no need to check
496 * for the DMF_NOFLUSH_SUSPENDING flag.
497 */
498static bool __must_push_back(struct multipath *m)
499{
500 return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) !=
501 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) &&
502 dm_noflush_suspending(m->ti));
503}
504
505static bool must_push_back_rq(struct multipath *m)
506{
507 bool r;
508 unsigned long flags;
509
510 spin_lock_irqsave(&m->lock, flags);
511 r = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) ||
512 __must_push_back(m));
513 spin_unlock_irqrestore(&m->lock, flags);
514
515 return r;
516}
517
518static bool must_push_back_bio(struct multipath *m)
519{
520 bool r;
521 unsigned long flags;
522
523 spin_lock_irqsave(&m->lock, flags);
524 r = __must_push_back(m);
525 spin_unlock_irqrestore(&m->lock, flags);
526
527 return r;
528}
529
530/*
531 * Map cloned requests (request-based multipath)
532 */
533static int __multipath_map(struct dm_target *ti, struct request *clone,
534 union map_info *map_context,
535 struct request *rq, struct request **__clone)
536{
537 struct multipath *m = ti->private;
538 int r = DM_MAPIO_REQUEUE;
539 size_t nr_bytes = clone ? blk_rq_bytes(clone) : blk_rq_bytes(rq);
540 struct pgpath *pgpath;
541 struct block_device *bdev;
542 struct dm_mpath_io *mpio;
543
544 /* Do we need to select a new pgpath? */
545 pgpath = lockless_dereference(m->current_pgpath);
546 if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
547 pgpath = choose_pgpath(m, nr_bytes);
548
549 if (!pgpath) {
550 if (must_push_back_rq(m))
551 return DM_MAPIO_DELAY_REQUEUE;
552 return -EIO; /* Failed */
553 } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
554 test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
555 pg_init_all_paths(m);
556 return r;
557 }
558
559 mpio = set_mpio(m, map_context);
560 if (!mpio)
561 /* ENOMEM, requeue */
562 return r;
563
564 mpio->pgpath = pgpath;
565 mpio->nr_bytes = nr_bytes;
566
567 bdev = pgpath->path.dev->bdev;
568
569 if (clone) {
570 /*
571 * Old request-based interface: allocated clone is passed in.
572 * Used by: .request_fn stacked on .request_fn path(s).
573 */
574 clone->q = bdev_get_queue(bdev);
575 clone->rq_disk = bdev->bd_disk;
576 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
577 } else {
578 /*
579 * blk-mq request-based interface; used by both:
580 * .request_fn stacked on blk-mq path(s) and
581 * blk-mq stacked on blk-mq path(s).
582 */
583 clone = blk_mq_alloc_request(bdev_get_queue(bdev),
584 rq_data_dir(rq), BLK_MQ_REQ_NOWAIT);
585 if (IS_ERR(clone)) {
586 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
587 clear_request_fn_mpio(m, map_context);
588 return r;
589 }
590 clone->bio = clone->biotail = NULL;
591 clone->rq_disk = bdev->bd_disk;
592 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
593 *__clone = clone;
594 }
595
596 if (pgpath->pg->ps.type->start_io)
597 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
598 &pgpath->path,
599 nr_bytes);
600 return DM_MAPIO_REMAPPED;
601}
602
603static int multipath_map(struct dm_target *ti, struct request *clone,
604 union map_info *map_context)
605{
606 return __multipath_map(ti, clone, map_context, NULL, NULL);
607}
608
609static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
610 union map_info *map_context,
611 struct request **clone)
612{
613 return __multipath_map(ti, NULL, map_context, rq, clone);
614}
615
616static void multipath_release_clone(struct request *clone)
617{
618 blk_mq_free_request(clone);
619}
620
621/*
622 * Map cloned bios (bio-based multipath)
623 */
624static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
625{
626 size_t nr_bytes = bio->bi_iter.bi_size;
627 struct pgpath *pgpath;
628 unsigned long flags;
629 bool queue_io;
630
631 /* Do we need to select a new pgpath? */
632 pgpath = lockless_dereference(m->current_pgpath);
633 queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
634 if (!pgpath || !queue_io)
635 pgpath = choose_pgpath(m, nr_bytes);
636
637 if ((pgpath && queue_io) ||
638 (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
639 /* Queue for the daemon to resubmit */
640 spin_lock_irqsave(&m->lock, flags);
641 bio_list_add(&m->queued_bios, bio);
642 spin_unlock_irqrestore(&m->lock, flags);
643 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
644 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
645 pg_init_all_paths(m);
646 else if (!queue_io)
647 queue_work(kmultipathd, &m->process_queued_bios);
648 return DM_MAPIO_SUBMITTED;
649 }
650
651 if (!pgpath) {
652 if (!must_push_back_bio(m))
653 return -EIO;
654 return DM_MAPIO_REQUEUE;
655 }
656
657 mpio->pgpath = pgpath;
658 mpio->nr_bytes = nr_bytes;
659
660 bio->bi_error = 0;
661 bio->bi_bdev = pgpath->path.dev->bdev;
662 bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
663
664 if (pgpath->pg->ps.type->start_io)
665 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
666 &pgpath->path,
667 nr_bytes);
668 return DM_MAPIO_REMAPPED;
669}
670
671static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
672{
673 struct multipath *m = ti->private;
674 struct dm_mpath_io *mpio = NULL;
675
676 multipath_init_per_bio_data(bio, &mpio, NULL);
677
678 return __multipath_map_bio(m, bio, mpio);
679}
680
681static void process_queued_io_list(struct multipath *m)
682{
683 if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
684 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
685 else if (m->queue_mode == DM_TYPE_BIO_BASED)
686 queue_work(kmultipathd, &m->process_queued_bios);
687}
688
689static void process_queued_bios(struct work_struct *work)
690{
691 int r;
692 unsigned long flags;
693 struct bio *bio;
694 struct bio_list bios;
695 struct blk_plug plug;
696 struct multipath *m =
697 container_of(work, struct multipath, process_queued_bios);
698
699 bio_list_init(&bios);
700
701 spin_lock_irqsave(&m->lock, flags);
702
703 if (bio_list_empty(&m->queued_bios)) {
704 spin_unlock_irqrestore(&m->lock, flags);
705 return;
706 }
707
708 bio_list_merge(&bios, &m->queued_bios);
709 bio_list_init(&m->queued_bios);
710
711 spin_unlock_irqrestore(&m->lock, flags);
712
713 blk_start_plug(&plug);
714 while ((bio = bio_list_pop(&bios))) {
715 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
716 if (r < 0 || r == DM_MAPIO_REQUEUE) {
717 bio->bi_error = r;
718 bio_endio(bio);
719 } else if (r == DM_MAPIO_REMAPPED)
720 generic_make_request(bio);
721 }
722 blk_finish_plug(&plug);
723}
724
725/*
726 * If we run out of usable paths, should we queue I/O or error it?
727 */
728static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
729 bool save_old_value)
730{
731 unsigned long flags;
732
733 spin_lock_irqsave(&m->lock, flags);
734
735 if (save_old_value) {
736 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
737 set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
738 else
739 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
740 } else {
741 if (queue_if_no_path)
742 set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
743 else
744 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
745 }
746 if (queue_if_no_path)
747 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
748 else
749 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
750
751 spin_unlock_irqrestore(&m->lock, flags);
752
753 if (!queue_if_no_path) {
754 dm_table_run_md_queue_async(m->ti->table);
755 process_queued_io_list(m);
756 }
757
758 return 0;
759}
760
761/*
762 * An event is triggered whenever a path is taken out of use.
763 * Includes path failure and PG bypass.
764 */
765static void trigger_event(struct work_struct *work)
766{
767 struct multipath *m =
768 container_of(work, struct multipath, trigger_event);
769
770 dm_table_event(m->ti->table);
771}
772
773/*-----------------------------------------------------------------
774 * Constructor/argument parsing:
775 * <#multipath feature args> [<arg>]*
776 * <#hw_handler args> [hw_handler [<arg>]*]
777 * <#priority groups>
778 * <initial priority group>
779 * [<selector> <#selector args> [<arg>]*
780 * <#paths> <#per-path selector args>
781 * [<path> [<arg>]* ]+ ]+
782 *---------------------------------------------------------------*/
783static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
784 struct dm_target *ti)
785{
786 int r;
787 struct path_selector_type *pst;
788 unsigned ps_argc;
789
790 static struct dm_arg _args[] = {
791 {0, 1024, "invalid number of path selector args"},
792 };
793
794 pst = dm_get_path_selector(dm_shift_arg(as));
795 if (!pst) {
796 ti->error = "unknown path selector type";
797 return -EINVAL;
798 }
799
800 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
801 if (r) {
802 dm_put_path_selector(pst);
803 return -EINVAL;
804 }
805
806 r = pst->create(&pg->ps, ps_argc, as->argv);
807 if (r) {
808 dm_put_path_selector(pst);
809 ti->error = "path selector constructor failed";
810 return r;
811 }
812
813 pg->ps.type = pst;
814 dm_consume_args(as, ps_argc);
815
816 return 0;
817}
818
819static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
820 struct dm_target *ti)
821{
822 int r;
823 struct pgpath *p;
824 struct multipath *m = ti->private;
825 struct request_queue *q = NULL;
826 const char *attached_handler_name;
827
828 /* we need at least a path arg */
829 if (as->argc < 1) {
830 ti->error = "no device given";
831 return ERR_PTR(-EINVAL);
832 }
833
834 p = alloc_pgpath();
835 if (!p)
836 return ERR_PTR(-ENOMEM);
837
838 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
839 &p->path.dev);
840 if (r) {
841 ti->error = "error getting device";
842 goto bad;
843 }
844
845 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
846 q = bdev_get_queue(p->path.dev->bdev);
847
848 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
849retain:
850 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
851 if (attached_handler_name) {
852 /*
853 * Clear any hw_handler_params associated with a
854 * handler that isn't already attached.
855 */
856 if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
857 kfree(m->hw_handler_params);
858 m->hw_handler_params = NULL;
859 }
860
861 /*
862 * Reset hw_handler_name to match the attached handler
863 *
864 * NB. This modifies the table line to show the actual
865 * handler instead of the original table passed in.
866 */
867 kfree(m->hw_handler_name);
868 m->hw_handler_name = attached_handler_name;
869 }
870 }
871
872 if (m->hw_handler_name) {
873 r = scsi_dh_attach(q, m->hw_handler_name);
874 if (r == -EBUSY) {
875 char b[BDEVNAME_SIZE];
876
877 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
878 bdevname(p->path.dev->bdev, b));
879 goto retain;
880 }
881 if (r < 0) {
882 ti->error = "error attaching hardware handler";
883 dm_put_device(ti, p->path.dev);
884 goto bad;
885 }
886
887 if (m->hw_handler_params) {
888 r = scsi_dh_set_params(q, m->hw_handler_params);
889 if (r < 0) {
890 ti->error = "unable to set hardware "
891 "handler parameters";
892 dm_put_device(ti, p->path.dev);
893 goto bad;
894 }
895 }
896 }
897
898 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
899 if (r) {
900 dm_put_device(ti, p->path.dev);
901 goto bad;
902 }
903
904 return p;
905
906 bad:
907 free_pgpath(p);
908 return ERR_PTR(r);
909}
910
911static struct priority_group *parse_priority_group(struct dm_arg_set *as,
912 struct multipath *m)
913{
914 static struct dm_arg _args[] = {
915 {1, 1024, "invalid number of paths"},
916 {0, 1024, "invalid number of selector args"}
917 };
918
919 int r;
920 unsigned i, nr_selector_args, nr_args;
921 struct priority_group *pg;
922 struct dm_target *ti = m->ti;
923
924 if (as->argc < 2) {
925 as->argc = 0;
926 ti->error = "not enough priority group arguments";
927 return ERR_PTR(-EINVAL);
928 }
929
930 pg = alloc_priority_group();
931 if (!pg) {
932 ti->error = "couldn't allocate priority group";
933 return ERR_PTR(-ENOMEM);
934 }
935 pg->m = m;
936
937 r = parse_path_selector(as, pg, ti);
938 if (r)
939 goto bad;
940
941 /*
942 * read the paths
943 */
944 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
945 if (r)
946 goto bad;
947
948 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
949 if (r)
950 goto bad;
951
952 nr_args = 1 + nr_selector_args;
953 for (i = 0; i < pg->nr_pgpaths; i++) {
954 struct pgpath *pgpath;
955 struct dm_arg_set path_args;
956
957 if (as->argc < nr_args) {
958 ti->error = "not enough path parameters";
959 r = -EINVAL;
960 goto bad;
961 }
962
963 path_args.argc = nr_args;
964 path_args.argv = as->argv;
965
966 pgpath = parse_path(&path_args, &pg->ps, ti);
967 if (IS_ERR(pgpath)) {
968 r = PTR_ERR(pgpath);
969 goto bad;
970 }
971
972 pgpath->pg = pg;
973 list_add_tail(&pgpath->list, &pg->pgpaths);
974 dm_consume_args(as, nr_args);
975 }
976
977 return pg;
978
979 bad:
980 free_priority_group(pg, ti);
981 return ERR_PTR(r);
982}
983
984static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
985{
986 unsigned hw_argc;
987 int ret;
988 struct dm_target *ti = m->ti;
989
990 static struct dm_arg _args[] = {
991 {0, 1024, "invalid number of hardware handler args"},
992 };
993
994 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
995 return -EINVAL;
996
997 if (!hw_argc)
998 return 0;
999
1000 if (m->queue_mode == DM_TYPE_BIO_BASED) {
1001 dm_consume_args(as, hw_argc);
1002 DMERR("bio-based multipath doesn't allow hardware handler args");
1003 return 0;
1004 }
1005
1006 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1007 if (!m->hw_handler_name)
1008 return -EINVAL;
1009
1010 if (hw_argc > 1) {
1011 char *p;
1012 int i, j, len = 4;
1013
1014 for (i = 0; i <= hw_argc - 2; i++)
1015 len += strlen(as->argv[i]) + 1;
1016 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1017 if (!p) {
1018 ti->error = "memory allocation failed";
1019 ret = -ENOMEM;
1020 goto fail;
1021 }
1022 j = sprintf(p, "%d", hw_argc - 1);
1023 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1024 j = sprintf(p, "%s", as->argv[i]);
1025 }
1026 dm_consume_args(as, hw_argc - 1);
1027
1028 return 0;
1029fail:
1030 kfree(m->hw_handler_name);
1031 m->hw_handler_name = NULL;
1032 return ret;
1033}
1034
1035static int parse_features(struct dm_arg_set *as, struct multipath *m)
1036{
1037 int r;
1038 unsigned argc;
1039 struct dm_target *ti = m->ti;
1040 const char *arg_name;
1041
1042 static struct dm_arg _args[] = {
1043 {0, 8, "invalid number of feature args"},
1044 {1, 50, "pg_init_retries must be between 1 and 50"},
1045 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1046 };
1047
1048 r = dm_read_arg_group(_args, as, &argc, &ti->error);
1049 if (r)
1050 return -EINVAL;
1051
1052 if (!argc)
1053 return 0;
1054
1055 do {
1056 arg_name = dm_shift_arg(as);
1057 argc--;
1058
1059 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1060 r = queue_if_no_path(m, true, false);
1061 continue;
1062 }
1063
1064 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1065 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1066 continue;
1067 }
1068
1069 if (!strcasecmp(arg_name, "pg_init_retries") &&
1070 (argc >= 1)) {
1071 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1072 argc--;
1073 continue;
1074 }
1075
1076 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1077 (argc >= 1)) {
1078 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1079 argc--;
1080 continue;
1081 }
1082
1083 if (!strcasecmp(arg_name, "queue_mode") &&
1084 (argc >= 1)) {
1085 const char *queue_mode_name = dm_shift_arg(as);
1086
1087 if (!strcasecmp(queue_mode_name, "bio"))
1088 m->queue_mode = DM_TYPE_BIO_BASED;
1089 else if (!strcasecmp(queue_mode_name, "rq"))
1090 m->queue_mode = DM_TYPE_REQUEST_BASED;
1091 else if (!strcasecmp(queue_mode_name, "mq"))
1092 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1093 else {
1094 ti->error = "Unknown 'queue_mode' requested";
1095 r = -EINVAL;
1096 }
1097 argc--;
1098 continue;
1099 }
1100
1101 ti->error = "Unrecognised multipath feature request";
1102 r = -EINVAL;
1103 } while (argc && !r);
1104
1105 return r;
1106}
1107
1108static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1109{
1110 /* target arguments */
1111 static struct dm_arg _args[] = {
1112 {0, 1024, "invalid number of priority groups"},
1113 {0, 1024, "invalid initial priority group number"},
1114 };
1115
1116 int r;
1117 struct multipath *m;
1118 struct dm_arg_set as;
1119 unsigned pg_count = 0;
1120 unsigned next_pg_num;
1121
1122 as.argc = argc;
1123 as.argv = argv;
1124
1125 m = alloc_multipath(ti);
1126 if (!m) {
1127 ti->error = "can't allocate multipath";
1128 return -EINVAL;
1129 }
1130
1131 r = parse_features(&as, m);
1132 if (r)
1133 goto bad;
1134
1135 r = alloc_multipath_stage2(ti, m);
1136 if (r)
1137 goto bad;
1138
1139 r = parse_hw_handler(&as, m);
1140 if (r)
1141 goto bad;
1142
1143 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1144 if (r)
1145 goto bad;
1146
1147 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1148 if (r)
1149 goto bad;
1150
1151 if ((!m->nr_priority_groups && next_pg_num) ||
1152 (m->nr_priority_groups && !next_pg_num)) {
1153 ti->error = "invalid initial priority group";
1154 r = -EINVAL;
1155 goto bad;
1156 }
1157
1158 /* parse the priority groups */
1159 while (as.argc) {
1160 struct priority_group *pg;
1161 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1162
1163 pg = parse_priority_group(&as, m);
1164 if (IS_ERR(pg)) {
1165 r = PTR_ERR(pg);
1166 goto bad;
1167 }
1168
1169 nr_valid_paths += pg->nr_pgpaths;
1170 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1171
1172 list_add_tail(&pg->list, &m->priority_groups);
1173 pg_count++;
1174 pg->pg_num = pg_count;
1175 if (!--next_pg_num)
1176 m->next_pg = pg;
1177 }
1178
1179 if (pg_count != m->nr_priority_groups) {
1180 ti->error = "priority group count mismatch";
1181 r = -EINVAL;
1182 goto bad;
1183 }
1184
1185 ti->num_flush_bios = 1;
1186 ti->num_discard_bios = 1;
1187 ti->num_write_same_bios = 1;
1188 if (m->queue_mode == DM_TYPE_BIO_BASED)
1189 ti->per_io_data_size = multipath_per_bio_data_size();
1190 else if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
1191 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1192
1193 return 0;
1194
1195 bad:
1196 free_multipath(m);
1197 return r;
1198}
1199
1200static void multipath_wait_for_pg_init_completion(struct multipath *m)
1201{
1202 DEFINE_WAIT(wait);
1203
1204 while (1) {
1205 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1206
1207 if (!atomic_read(&m->pg_init_in_progress))
1208 break;
1209
1210 io_schedule();
1211 }
1212 finish_wait(&m->pg_init_wait, &wait);
1213}
1214
1215static void flush_multipath_work(struct multipath *m)
1216{
1217 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1218 smp_mb__after_atomic();
1219
1220 flush_workqueue(kmpath_handlerd);
1221 multipath_wait_for_pg_init_completion(m);
1222 flush_workqueue(kmultipathd);
1223 flush_work(&m->trigger_event);
1224
1225 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1226 smp_mb__after_atomic();
1227}
1228
1229static void multipath_dtr(struct dm_target *ti)
1230{
1231 struct multipath *m = ti->private;
1232
1233 flush_multipath_work(m);
1234 free_multipath(m);
1235}
1236
1237/*
1238 * Take a path out of use.
1239 */
1240static int fail_path(struct pgpath *pgpath)
1241{
1242 unsigned long flags;
1243 struct multipath *m = pgpath->pg->m;
1244
1245 spin_lock_irqsave(&m->lock, flags);
1246
1247 if (!pgpath->is_active)
1248 goto out;
1249
1250 DMWARN("Failing path %s.", pgpath->path.dev->name);
1251
1252 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1253 pgpath->is_active = false;
1254 pgpath->fail_count++;
1255
1256 atomic_dec(&m->nr_valid_paths);
1257
1258 if (pgpath == m->current_pgpath)
1259 m->current_pgpath = NULL;
1260
1261 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1262 pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1263
1264 schedule_work(&m->trigger_event);
1265
1266out:
1267 spin_unlock_irqrestore(&m->lock, flags);
1268
1269 return 0;
1270}
1271
1272/*
1273 * Reinstate a previously-failed path
1274 */
1275static int reinstate_path(struct pgpath *pgpath)
1276{
1277 int r = 0, run_queue = 0;
1278 unsigned long flags;
1279 struct multipath *m = pgpath->pg->m;
1280 unsigned nr_valid_paths;
1281
1282 spin_lock_irqsave(&m->lock, flags);
1283
1284 if (pgpath->is_active)
1285 goto out;
1286
1287 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1288
1289 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1290 if (r)
1291 goto out;
1292
1293 pgpath->is_active = true;
1294
1295 nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1296 if (nr_valid_paths == 1) {
1297 m->current_pgpath = NULL;
1298 run_queue = 1;
1299 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1300 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1301 atomic_inc(&m->pg_init_in_progress);
1302 }
1303
1304 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1305 pgpath->path.dev->name, nr_valid_paths);
1306
1307 schedule_work(&m->trigger_event);
1308
1309out:
1310 spin_unlock_irqrestore(&m->lock, flags);
1311 if (run_queue) {
1312 dm_table_run_md_queue_async(m->ti->table);
1313 process_queued_io_list(m);
1314 }
1315
1316 return r;
1317}
1318
1319/*
1320 * Fail or reinstate all paths that match the provided struct dm_dev.
1321 */
1322static int action_dev(struct multipath *m, struct dm_dev *dev,
1323 action_fn action)
1324{
1325 int r = -EINVAL;
1326 struct pgpath *pgpath;
1327 struct priority_group *pg;
1328
1329 list_for_each_entry(pg, &m->priority_groups, list) {
1330 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1331 if (pgpath->path.dev == dev)
1332 r = action(pgpath);
1333 }
1334 }
1335
1336 return r;
1337}
1338
1339/*
1340 * Temporarily try to avoid having to use the specified PG
1341 */
1342static void bypass_pg(struct multipath *m, struct priority_group *pg,
1343 bool bypassed)
1344{
1345 unsigned long flags;
1346
1347 spin_lock_irqsave(&m->lock, flags);
1348
1349 pg->bypassed = bypassed;
1350 m->current_pgpath = NULL;
1351 m->current_pg = NULL;
1352
1353 spin_unlock_irqrestore(&m->lock, flags);
1354
1355 schedule_work(&m->trigger_event);
1356}
1357
1358/*
1359 * Switch to using the specified PG from the next I/O that gets mapped
1360 */
1361static int switch_pg_num(struct multipath *m, const char *pgstr)
1362{
1363 struct priority_group *pg;
1364 unsigned pgnum;
1365 unsigned long flags;
1366 char dummy;
1367
1368 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1369 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1370 DMWARN("invalid PG number supplied to switch_pg_num");
1371 return -EINVAL;
1372 }
1373
1374 spin_lock_irqsave(&m->lock, flags);
1375 list_for_each_entry(pg, &m->priority_groups, list) {
1376 pg->bypassed = false;
1377 if (--pgnum)
1378 continue;
1379
1380 m->current_pgpath = NULL;
1381 m->current_pg = NULL;
1382 m->next_pg = pg;
1383 }
1384 spin_unlock_irqrestore(&m->lock, flags);
1385
1386 schedule_work(&m->trigger_event);
1387 return 0;
1388}
1389
1390/*
1391 * Set/clear bypassed status of a PG.
1392 * PGs are numbered upwards from 1 in the order they were declared.
1393 */
1394static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1395{
1396 struct priority_group *pg;
1397 unsigned pgnum;
1398 char dummy;
1399
1400 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1401 !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1402 DMWARN("invalid PG number supplied to bypass_pg");
1403 return -EINVAL;
1404 }
1405
1406 list_for_each_entry(pg, &m->priority_groups, list) {
1407 if (!--pgnum)
1408 break;
1409 }
1410
1411 bypass_pg(m, pg, bypassed);
1412 return 0;
1413}
1414
1415/*
1416 * Should we retry pg_init immediately?
1417 */
1418static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1419{
1420 unsigned long flags;
1421 bool limit_reached = false;
1422
1423 spin_lock_irqsave(&m->lock, flags);
1424
1425 if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1426 !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1427 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1428 else
1429 limit_reached = true;
1430
1431 spin_unlock_irqrestore(&m->lock, flags);
1432
1433 return limit_reached;
1434}
1435
1436static void pg_init_done(void *data, int errors)
1437{
1438 struct pgpath *pgpath = data;
1439 struct priority_group *pg = pgpath->pg;
1440 struct multipath *m = pg->m;
1441 unsigned long flags;
1442 bool delay_retry = false;
1443
1444 /* device or driver problems */
1445 switch (errors) {
1446 case SCSI_DH_OK:
1447 break;
1448 case SCSI_DH_NOSYS:
1449 if (!m->hw_handler_name) {
1450 errors = 0;
1451 break;
1452 }
1453 DMERR("Could not failover the device: Handler scsi_dh_%s "
1454 "Error %d.", m->hw_handler_name, errors);
1455 /*
1456 * Fail path for now, so we do not ping pong
1457 */
1458 fail_path(pgpath);
1459 break;
1460 case SCSI_DH_DEV_TEMP_BUSY:
1461 /*
1462 * Probably doing something like FW upgrade on the
1463 * controller so try the other pg.
1464 */
1465 bypass_pg(m, pg, true);
1466 break;
1467 case SCSI_DH_RETRY:
1468 /* Wait before retrying. */
1469 delay_retry = 1;
1470 case SCSI_DH_IMM_RETRY:
1471 case SCSI_DH_RES_TEMP_UNAVAIL:
1472 if (pg_init_limit_reached(m, pgpath))
1473 fail_path(pgpath);
1474 errors = 0;
1475 break;
1476 case SCSI_DH_DEV_OFFLINED:
1477 default:
1478 /*
1479 * We probably do not want to fail the path for a device
1480 * error, but this is what the old dm did. In future
1481 * patches we can do more advanced handling.
1482 */
1483 fail_path(pgpath);
1484 }
1485
1486 spin_lock_irqsave(&m->lock, flags);
1487 if (errors) {
1488 if (pgpath == m->current_pgpath) {
1489 DMERR("Could not failover device. Error %d.", errors);
1490 m->current_pgpath = NULL;
1491 m->current_pg = NULL;
1492 }
1493 } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1494 pg->bypassed = false;
1495
1496 if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1497 /* Activations of other paths are still on going */
1498 goto out;
1499
1500 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1501 if (delay_retry)
1502 set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1503 else
1504 clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1505
1506 if (__pg_init_all_paths(m))
1507 goto out;
1508 }
1509 clear_bit(MPATHF_QUEUE_IO, &m->flags);
1510
1511 process_queued_io_list(m);
1512
1513 /*
1514 * Wake up any thread waiting to suspend.
1515 */
1516 wake_up(&m->pg_init_wait);
1517
1518out:
1519 spin_unlock_irqrestore(&m->lock, flags);
1520}
1521
1522static void activate_path(struct work_struct *work)
1523{
1524 struct pgpath *pgpath =
1525 container_of(work, struct pgpath, activate_path.work);
1526 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1527
1528 if (pgpath->is_active && !blk_queue_dying(q))
1529 scsi_dh_activate(q, pg_init_done, pgpath);
1530 else
1531 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1532}
1533
1534static int noretry_error(int error)
1535{
1536 switch (error) {
1537 case -EBADE:
1538 /*
1539 * EBADE signals an reservation conflict.
1540 * We shouldn't fail the path here as we can communicate with
1541 * the target. We should failover to the next path, but in
1542 * doing so we might be causing a ping-pong between paths.
1543 * So just return the reservation conflict error.
1544 */
1545 case -EOPNOTSUPP:
1546 case -EREMOTEIO:
1547 case -EILSEQ:
1548 case -ENODATA:
1549 case -ENOSPC:
1550 return 1;
1551 }
1552
1553 /* Anything else could be a path failure, so should be retried */
1554 return 0;
1555}
1556
1557/*
1558 * end_io handling
1559 */
1560static int do_end_io(struct multipath *m, struct request *clone,
1561 int error, struct dm_mpath_io *mpio)
1562{
1563 /*
1564 * We don't queue any clone request inside the multipath target
1565 * during end I/O handling, since those clone requests don't have
1566 * bio clones. If we queue them inside the multipath target,
1567 * we need to make bio clones, that requires memory allocation.
1568 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1569 * don't have bio clones.)
1570 * Instead of queueing the clone request here, we queue the original
1571 * request into dm core, which will remake a clone request and
1572 * clone bios for it and resubmit it later.
1573 */
1574 int r = DM_ENDIO_REQUEUE;
1575
1576 if (!error && !clone->errors)
1577 return 0; /* I/O complete */
1578
1579 if (noretry_error(error))
1580 return error;
1581
1582 if (mpio->pgpath)
1583 fail_path(mpio->pgpath);
1584
1585 if (!atomic_read(&m->nr_valid_paths)) {
1586 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1587 if (!must_push_back_rq(m))
1588 r = -EIO;
1589 }
1590 }
1591
1592 return r;
1593}
1594
1595static int multipath_end_io(struct dm_target *ti, struct request *clone,
1596 int error, union map_info *map_context)
1597{
1598 struct multipath *m = ti->private;
1599 struct dm_mpath_io *mpio = get_mpio(map_context);
1600 struct pgpath *pgpath;
1601 struct path_selector *ps;
1602 int r;
1603
1604 BUG_ON(!mpio);
1605
1606 r = do_end_io(m, clone, error, mpio);
1607 pgpath = mpio->pgpath;
1608 if (pgpath) {
1609 ps = &pgpath->pg->ps;
1610 if (ps->type->end_io)
1611 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1612 }
1613 clear_request_fn_mpio(m, map_context);
1614
1615 return r;
1616}
1617
1618static int do_end_io_bio(struct multipath *m, struct bio *clone,
1619 int error, struct dm_mpath_io *mpio)
1620{
1621 unsigned long flags;
1622
1623 if (!error)
1624 return 0; /* I/O complete */
1625
1626 if (noretry_error(error))
1627 return error;
1628
1629 if (mpio->pgpath)
1630 fail_path(mpio->pgpath);
1631
1632 if (!atomic_read(&m->nr_valid_paths)) {
1633 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1634 if (!must_push_back_bio(m))
1635 return -EIO;
1636 return DM_ENDIO_REQUEUE;
1637 }
1638 }
1639
1640 /* Queue for the daemon to resubmit */
1641 dm_bio_restore(get_bio_details_from_bio(clone), clone);
1642
1643 spin_lock_irqsave(&m->lock, flags);
1644 bio_list_add(&m->queued_bios, clone);
1645 spin_unlock_irqrestore(&m->lock, flags);
1646 if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1647 queue_work(kmultipathd, &m->process_queued_bios);
1648
1649 return DM_ENDIO_INCOMPLETE;
1650}
1651
1652static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone, int error)
1653{
1654 struct multipath *m = ti->private;
1655 struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1656 struct pgpath *pgpath;
1657 struct path_selector *ps;
1658 int r;
1659
1660 BUG_ON(!mpio);
1661
1662 r = do_end_io_bio(m, clone, error, mpio);
1663 pgpath = mpio->pgpath;
1664 if (pgpath) {
1665 ps = &pgpath->pg->ps;
1666 if (ps->type->end_io)
1667 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1668 }
1669
1670 return r;
1671}
1672
1673/*
1674 * Suspend can't complete until all the I/O is processed so if
1675 * the last path fails we must error any remaining I/O.
1676 * Note that if the freeze_bdev fails while suspending, the
1677 * queue_if_no_path state is lost - userspace should reset it.
1678 */
1679static void multipath_presuspend(struct dm_target *ti)
1680{
1681 struct multipath *m = ti->private;
1682
1683 queue_if_no_path(m, false, true);
1684}
1685
1686static void multipath_postsuspend(struct dm_target *ti)
1687{
1688 struct multipath *m = ti->private;
1689
1690 mutex_lock(&m->work_mutex);
1691 flush_multipath_work(m);
1692 mutex_unlock(&m->work_mutex);
1693}
1694
1695/*
1696 * Restore the queue_if_no_path setting.
1697 */
1698static void multipath_resume(struct dm_target *ti)
1699{
1700 struct multipath *m = ti->private;
1701 unsigned long flags;
1702
1703 spin_lock_irqsave(&m->lock, flags);
1704 if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags))
1705 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1706 else
1707 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1708 spin_unlock_irqrestore(&m->lock, flags);
1709}
1710
1711/*
1712 * Info output has the following format:
1713 * num_multipath_feature_args [multipath_feature_args]*
1714 * num_handler_status_args [handler_status_args]*
1715 * num_groups init_group_number
1716 * [A|D|E num_ps_status_args [ps_status_args]*
1717 * num_paths num_selector_args
1718 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1719 *
1720 * Table output has the following format (identical to the constructor string):
1721 * num_feature_args [features_args]*
1722 * num_handler_args hw_handler [hw_handler_args]*
1723 * num_groups init_group_number
1724 * [priority selector-name num_ps_args [ps_args]*
1725 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1726 */
1727static void multipath_status(struct dm_target *ti, status_type_t type,
1728 unsigned status_flags, char *result, unsigned maxlen)
1729{
1730 int sz = 0;
1731 unsigned long flags;
1732 struct multipath *m = ti->private;
1733 struct priority_group *pg;
1734 struct pgpath *p;
1735 unsigned pg_num;
1736 char state;
1737
1738 spin_lock_irqsave(&m->lock, flags);
1739
1740 /* Features */
1741 if (type == STATUSTYPE_INFO)
1742 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1743 atomic_read(&m->pg_init_count));
1744 else {
1745 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1746 (m->pg_init_retries > 0) * 2 +
1747 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1748 test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1749 (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1750
1751 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1752 DMEMIT("queue_if_no_path ");
1753 if (m->pg_init_retries)
1754 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1755 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1756 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1757 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1758 DMEMIT("retain_attached_hw_handler ");
1759 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1760 switch(m->queue_mode) {
1761 case DM_TYPE_BIO_BASED:
1762 DMEMIT("queue_mode bio ");
1763 break;
1764 case DM_TYPE_MQ_REQUEST_BASED:
1765 DMEMIT("queue_mode mq ");
1766 break;
1767 }
1768 }
1769 }
1770
1771 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1772 DMEMIT("0 ");
1773 else
1774 DMEMIT("1 %s ", m->hw_handler_name);
1775
1776 DMEMIT("%u ", m->nr_priority_groups);
1777
1778 if (m->next_pg)
1779 pg_num = m->next_pg->pg_num;
1780 else if (m->current_pg)
1781 pg_num = m->current_pg->pg_num;
1782 else
1783 pg_num = (m->nr_priority_groups ? 1 : 0);
1784
1785 DMEMIT("%u ", pg_num);
1786
1787 switch (type) {
1788 case STATUSTYPE_INFO:
1789 list_for_each_entry(pg, &m->priority_groups, list) {
1790 if (pg->bypassed)
1791 state = 'D'; /* Disabled */
1792 else if (pg == m->current_pg)
1793 state = 'A'; /* Currently Active */
1794 else
1795 state = 'E'; /* Enabled */
1796
1797 DMEMIT("%c ", state);
1798
1799 if (pg->ps.type->status)
1800 sz += pg->ps.type->status(&pg->ps, NULL, type,
1801 result + sz,
1802 maxlen - sz);
1803 else
1804 DMEMIT("0 ");
1805
1806 DMEMIT("%u %u ", pg->nr_pgpaths,
1807 pg->ps.type->info_args);
1808
1809 list_for_each_entry(p, &pg->pgpaths, list) {
1810 DMEMIT("%s %s %u ", p->path.dev->name,
1811 p->is_active ? "A" : "F",
1812 p->fail_count);
1813 if (pg->ps.type->status)
1814 sz += pg->ps.type->status(&pg->ps,
1815 &p->path, type, result + sz,
1816 maxlen - sz);
1817 }
1818 }
1819 break;
1820
1821 case STATUSTYPE_TABLE:
1822 list_for_each_entry(pg, &m->priority_groups, list) {
1823 DMEMIT("%s ", pg->ps.type->name);
1824
1825 if (pg->ps.type->status)
1826 sz += pg->ps.type->status(&pg->ps, NULL, type,
1827 result + sz,
1828 maxlen - sz);
1829 else
1830 DMEMIT("0 ");
1831
1832 DMEMIT("%u %u ", pg->nr_pgpaths,
1833 pg->ps.type->table_args);
1834
1835 list_for_each_entry(p, &pg->pgpaths, list) {
1836 DMEMIT("%s ", p->path.dev->name);
1837 if (pg->ps.type->status)
1838 sz += pg->ps.type->status(&pg->ps,
1839 &p->path, type, result + sz,
1840 maxlen - sz);
1841 }
1842 }
1843 break;
1844 }
1845
1846 spin_unlock_irqrestore(&m->lock, flags);
1847}
1848
1849static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1850{
1851 int r = -EINVAL;
1852 struct dm_dev *dev;
1853 struct multipath *m = ti->private;
1854 action_fn action;
1855
1856 mutex_lock(&m->work_mutex);
1857
1858 if (dm_suspended(ti)) {
1859 r = -EBUSY;
1860 goto out;
1861 }
1862
1863 if (argc == 1) {
1864 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1865 r = queue_if_no_path(m, true, false);
1866 goto out;
1867 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1868 r = queue_if_no_path(m, false, false);
1869 goto out;
1870 }
1871 }
1872
1873 if (argc != 2) {
1874 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1875 goto out;
1876 }
1877
1878 if (!strcasecmp(argv[0], "disable_group")) {
1879 r = bypass_pg_num(m, argv[1], true);
1880 goto out;
1881 } else if (!strcasecmp(argv[0], "enable_group")) {
1882 r = bypass_pg_num(m, argv[1], false);
1883 goto out;
1884 } else if (!strcasecmp(argv[0], "switch_group")) {
1885 r = switch_pg_num(m, argv[1]);
1886 goto out;
1887 } else if (!strcasecmp(argv[0], "reinstate_path"))
1888 action = reinstate_path;
1889 else if (!strcasecmp(argv[0], "fail_path"))
1890 action = fail_path;
1891 else {
1892 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1893 goto out;
1894 }
1895
1896 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1897 if (r) {
1898 DMWARN("message: error getting device %s",
1899 argv[1]);
1900 goto out;
1901 }
1902
1903 r = action_dev(m, dev, action);
1904
1905 dm_put_device(ti, dev);
1906
1907out:
1908 mutex_unlock(&m->work_mutex);
1909 return r;
1910}
1911
1912static int multipath_prepare_ioctl(struct dm_target *ti,
1913 struct block_device **bdev, fmode_t *mode)
1914{
1915 struct multipath *m = ti->private;
1916 struct pgpath *current_pgpath;
1917 int r;
1918
1919 current_pgpath = lockless_dereference(m->current_pgpath);
1920 if (!current_pgpath)
1921 current_pgpath = choose_pgpath(m, 0);
1922
1923 if (current_pgpath) {
1924 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1925 *bdev = current_pgpath->path.dev->bdev;
1926 *mode = current_pgpath->path.dev->mode;
1927 r = 0;
1928 } else {
1929 /* pg_init has not started or completed */
1930 r = -ENOTCONN;
1931 }
1932 } else {
1933 /* No path is available */
1934 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1935 r = -ENOTCONN;
1936 else
1937 r = -EIO;
1938 }
1939
1940 if (r == -ENOTCONN) {
1941 if (!lockless_dereference(m->current_pg)) {
1942 /* Path status changed, redo selection */
1943 (void) choose_pgpath(m, 0);
1944 }
1945 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1946 pg_init_all_paths(m);
1947 dm_table_run_md_queue_async(m->ti->table);
1948 process_queued_io_list(m);
1949 }
1950
1951 /*
1952 * Only pass ioctls through if the device sizes match exactly.
1953 */
1954 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1955 return 1;
1956 return r;
1957}
1958
1959static int multipath_iterate_devices(struct dm_target *ti,
1960 iterate_devices_callout_fn fn, void *data)
1961{
1962 struct multipath *m = ti->private;
1963 struct priority_group *pg;
1964 struct pgpath *p;
1965 int ret = 0;
1966
1967 list_for_each_entry(pg, &m->priority_groups, list) {
1968 list_for_each_entry(p, &pg->pgpaths, list) {
1969 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1970 if (ret)
1971 goto out;
1972 }
1973 }
1974
1975out:
1976 return ret;
1977}
1978
1979static int pgpath_busy(struct pgpath *pgpath)
1980{
1981 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1982
1983 return blk_lld_busy(q);
1984}
1985
1986/*
1987 * We return "busy", only when we can map I/Os but underlying devices
1988 * are busy (so even if we map I/Os now, the I/Os will wait on
1989 * the underlying queue).
1990 * In other words, if we want to kill I/Os or queue them inside us
1991 * due to map unavailability, we don't return "busy". Otherwise,
1992 * dm core won't give us the I/Os and we can't do what we want.
1993 */
1994static int multipath_busy(struct dm_target *ti)
1995{
1996 bool busy = false, has_active = false;
1997 struct multipath *m = ti->private;
1998 struct priority_group *pg, *next_pg;
1999 struct pgpath *pgpath;
2000
2001 /* pg_init in progress */
2002 if (atomic_read(&m->pg_init_in_progress))
2003 return true;
2004
2005 /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2006 if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2007 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
2008
2009 /* Guess which priority_group will be used at next mapping time */
2010 pg = lockless_dereference(m->current_pg);
2011 next_pg = lockless_dereference(m->next_pg);
2012 if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
2013 pg = next_pg;
2014
2015 if (!pg) {
2016 /*
2017 * We don't know which pg will be used at next mapping time.
2018 * We don't call choose_pgpath() here to avoid to trigger
2019 * pg_init just by busy checking.
2020 * So we don't know whether underlying devices we will be using
2021 * at next mapping time are busy or not. Just try mapping.
2022 */
2023 return busy;
2024 }
2025
2026 /*
2027 * If there is one non-busy active path at least, the path selector
2028 * will be able to select it. So we consider such a pg as not busy.
2029 */
2030 busy = true;
2031 list_for_each_entry(pgpath, &pg->pgpaths, list) {
2032 if (pgpath->is_active) {
2033 has_active = true;
2034 if (!pgpath_busy(pgpath)) {
2035 busy = false;
2036 break;
2037 }
2038 }
2039 }
2040
2041 if (!has_active) {
2042 /*
2043 * No active path in this pg, so this pg won't be used and
2044 * the current_pg will be changed at next mapping time.
2045 * We need to try mapping to determine it.
2046 */
2047 busy = false;
2048 }
2049
2050 return busy;
2051}
2052
2053/*-----------------------------------------------------------------
2054 * Module setup
2055 *---------------------------------------------------------------*/
2056static struct target_type multipath_target = {
2057 .name = "multipath",
2058 .version = {1, 12, 0},
2059 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
2060 .module = THIS_MODULE,
2061 .ctr = multipath_ctr,
2062 .dtr = multipath_dtr,
2063 .map_rq = multipath_map,
2064 .clone_and_map_rq = multipath_clone_and_map,
2065 .release_clone_rq = multipath_release_clone,
2066 .rq_end_io = multipath_end_io,
2067 .map = multipath_map_bio,
2068 .end_io = multipath_end_io_bio,
2069 .presuspend = multipath_presuspend,
2070 .postsuspend = multipath_postsuspend,
2071 .resume = multipath_resume,
2072 .status = multipath_status,
2073 .message = multipath_message,
2074 .prepare_ioctl = multipath_prepare_ioctl,
2075 .iterate_devices = multipath_iterate_devices,
2076 .busy = multipath_busy,
2077};
2078
2079static int __init dm_multipath_init(void)
2080{
2081 int r;
2082
2083 /* allocate a slab for the dm_mpath_ios */
2084 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
2085 if (!_mpio_cache)
2086 return -ENOMEM;
2087
2088 r = dm_register_target(&multipath_target);
2089 if (r < 0) {
2090 DMERR("request-based register failed %d", r);
2091 r = -EINVAL;
2092 goto bad_register_target;
2093 }
2094
2095 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2096 if (!kmultipathd) {
2097 DMERR("failed to create workqueue kmpathd");
2098 r = -ENOMEM;
2099 goto bad_alloc_kmultipathd;
2100 }
2101
2102 /*
2103 * A separate workqueue is used to handle the device handlers
2104 * to avoid overloading existing workqueue. Overloading the
2105 * old workqueue would also create a bottleneck in the
2106 * path of the storage hardware device activation.
2107 */
2108 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2109 WQ_MEM_RECLAIM);
2110 if (!kmpath_handlerd) {
2111 DMERR("failed to create workqueue kmpath_handlerd");
2112 r = -ENOMEM;
2113 goto bad_alloc_kmpath_handlerd;
2114 }
2115
2116 return 0;
2117
2118bad_alloc_kmpath_handlerd:
2119 destroy_workqueue(kmultipathd);
2120bad_alloc_kmultipathd:
2121 dm_unregister_target(&multipath_target);
2122bad_register_target:
2123 kmem_cache_destroy(_mpio_cache);
2124
2125 return r;
2126}
2127
2128static void __exit dm_multipath_exit(void)
2129{
2130 destroy_workqueue(kmpath_handlerd);
2131 destroy_workqueue(kmultipathd);
2132
2133 dm_unregister_target(&multipath_target);
2134 kmem_cache_destroy(_mpio_cache);
2135}
2136
2137module_init(dm_multipath_init);
2138module_exit(dm_multipath_exit);
2139
2140MODULE_DESCRIPTION(DM_NAME " multipath target");
2141MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2142MODULE_LICENSE("GPL");