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.h"
11#include "dm-path-selector.h"
12#include "dm-uevent.h"
13
14#include <linux/blkdev.h>
15#include <linux/ctype.h>
16#include <linux/init.h>
17#include <linux/mempool.h>
18#include <linux/module.h>
19#include <linux/pagemap.h>
20#include <linux/slab.h>
21#include <linux/time.h>
22#include <linux/workqueue.h>
23#include <linux/delay.h>
24#include <scsi/scsi_dh.h>
25#include <linux/atomic.h>
26#include <linux/blk-mq.h>
27
28#define DM_MSG_PREFIX "multipath"
29#define DM_PG_INIT_DELAY_MSECS 2000
30#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
31
32/* Path properties */
33struct pgpath {
34 struct list_head list;
35
36 struct priority_group *pg; /* Owning PG */
37 unsigned fail_count; /* Cumulative failure count */
38
39 struct dm_path path;
40 struct delayed_work activate_path;
41
42 bool is_active:1; /* Path status */
43};
44
45#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
46
47/*
48 * Paths are grouped into Priority Groups and numbered from 1 upwards.
49 * Each has a path selector which controls which path gets used.
50 */
51struct priority_group {
52 struct list_head list;
53
54 struct multipath *m; /* Owning multipath instance */
55 struct path_selector ps;
56
57 unsigned pg_num; /* Reference number */
58 unsigned nr_pgpaths; /* Number of paths in PG */
59 struct list_head pgpaths;
60
61 bool bypassed:1; /* Temporarily bypass this PG? */
62};
63
64/* Multipath context */
65struct multipath {
66 struct list_head list;
67 struct dm_target *ti;
68
69 const char *hw_handler_name;
70 char *hw_handler_params;
71
72 spinlock_t lock;
73
74 unsigned nr_priority_groups;
75 struct list_head priority_groups;
76
77 wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
78
79 unsigned pg_init_in_progress; /* Only one pg_init allowed at once */
80
81 unsigned nr_valid_paths; /* Total number of usable paths */
82 struct pgpath *current_pgpath;
83 struct priority_group *current_pg;
84 struct priority_group *next_pg; /* Switch to this PG if set */
85
86 bool queue_io:1; /* Must we queue all I/O? */
87 bool queue_if_no_path:1; /* Queue I/O if last path fails? */
88 bool saved_queue_if_no_path:1; /* Saved state during suspension */
89 bool retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
90 bool pg_init_disabled:1; /* pg_init is not currently allowed */
91 bool pg_init_required:1; /* pg_init needs calling? */
92 bool pg_init_delay_retry:1; /* Delay pg_init retry? */
93
94 unsigned pg_init_retries; /* Number of times to retry pg_init */
95 unsigned pg_init_count; /* Number of times pg_init called */
96 unsigned pg_init_delay_msecs; /* Number of msecs before pg_init retry */
97
98 struct work_struct trigger_event;
99
100 /*
101 * We must use a mempool of dm_mpath_io structs so that we
102 * can resubmit bios on error.
103 */
104 mempool_t *mpio_pool;
105
106 struct mutex work_mutex;
107};
108
109/*
110 * Context information attached to each bio we process.
111 */
112struct dm_mpath_io {
113 struct pgpath *pgpath;
114 size_t nr_bytes;
115};
116
117typedef int (*action_fn) (struct pgpath *pgpath);
118
119static struct kmem_cache *_mpio_cache;
120
121static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
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 = true;
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
163 list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
164 list_del(&pgpath->list);
165 dm_put_device(ti, pgpath->path.dev);
166 free_pgpath(pgpath);
167 }
168}
169
170static void free_priority_group(struct priority_group *pg,
171 struct dm_target *ti)
172{
173 struct path_selector *ps = &pg->ps;
174
175 if (ps->type) {
176 ps->type->destroy(ps);
177 dm_put_path_selector(ps->type);
178 }
179
180 free_pgpaths(&pg->pgpaths, ti);
181 kfree(pg);
182}
183
184static struct multipath *alloc_multipath(struct dm_target *ti, bool use_blk_mq)
185{
186 struct multipath *m;
187
188 m = kzalloc(sizeof(*m), GFP_KERNEL);
189 if (m) {
190 INIT_LIST_HEAD(&m->priority_groups);
191 spin_lock_init(&m->lock);
192 m->queue_io = true;
193 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
194 INIT_WORK(&m->trigger_event, trigger_event);
195 init_waitqueue_head(&m->pg_init_wait);
196 mutex_init(&m->work_mutex);
197
198 m->mpio_pool = NULL;
199 if (!use_blk_mq) {
200 unsigned min_ios = dm_get_reserved_rq_based_ios();
201
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 }
208
209 m->ti = ti;
210 ti->private = m;
211 }
212
213 return m;
214}
215
216static void free_multipath(struct multipath *m)
217{
218 struct priority_group *pg, *tmp;
219
220 list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
221 list_del(&pg->list);
222 free_priority_group(pg, m->ti);
223 }
224
225 kfree(m->hw_handler_name);
226 kfree(m->hw_handler_params);
227 mempool_destroy(m->mpio_pool);
228 kfree(m);
229}
230
231static struct dm_mpath_io *get_mpio(union map_info *info)
232{
233 return info->ptr;
234}
235
236static struct dm_mpath_io *set_mpio(struct multipath *m, union map_info *info)
237{
238 struct dm_mpath_io *mpio;
239
240 if (!m->mpio_pool) {
241 /* Use blk-mq pdu memory requested via per_io_data_size */
242 mpio = get_mpio(info);
243 memset(mpio, 0, sizeof(*mpio));
244 return mpio;
245 }
246
247 mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
248 if (!mpio)
249 return NULL;
250
251 memset(mpio, 0, sizeof(*mpio));
252 info->ptr = mpio;
253
254 return mpio;
255}
256
257static void clear_request_fn_mpio(struct multipath *m, union map_info *info)
258{
259 /* Only needed for non blk-mq (.request_fn) multipath */
260 if (m->mpio_pool) {
261 struct dm_mpath_io *mpio = info->ptr;
262
263 info->ptr = NULL;
264 mempool_free(mpio, m->mpio_pool);
265 }
266}
267
268/*-----------------------------------------------
269 * Path selection
270 *-----------------------------------------------*/
271
272static int __pg_init_all_paths(struct multipath *m)
273{
274 struct pgpath *pgpath;
275 unsigned long pg_init_delay = 0;
276
277 if (m->pg_init_in_progress || m->pg_init_disabled)
278 return 0;
279
280 m->pg_init_count++;
281 m->pg_init_required = false;
282
283 /* Check here to reset pg_init_required */
284 if (!m->current_pg)
285 return 0;
286
287 if (m->pg_init_delay_retry)
288 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
289 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
290 list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
291 /* Skip failed paths */
292 if (!pgpath->is_active)
293 continue;
294 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
295 pg_init_delay))
296 m->pg_init_in_progress++;
297 }
298 return m->pg_init_in_progress;
299}
300
301static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
302{
303 m->current_pg = pgpath->pg;
304
305 /* Must we initialise the PG first, and queue I/O till it's ready? */
306 if (m->hw_handler_name) {
307 m->pg_init_required = true;
308 m->queue_io = true;
309 } else {
310 m->pg_init_required = false;
311 m->queue_io = false;
312 }
313
314 m->pg_init_count = 0;
315}
316
317static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
318 size_t nr_bytes)
319{
320 struct dm_path *path;
321
322 path = pg->ps.type->select_path(&pg->ps, nr_bytes);
323 if (!path)
324 return -ENXIO;
325
326 m->current_pgpath = path_to_pgpath(path);
327
328 if (m->current_pg != pg)
329 __switch_pg(m, m->current_pgpath);
330
331 return 0;
332}
333
334static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
335{
336 struct priority_group *pg;
337 bool bypassed = true;
338
339 if (!m->nr_valid_paths) {
340 m->queue_io = false;
341 goto failed;
342 }
343
344 /* Were we instructed to switch PG? */
345 if (m->next_pg) {
346 pg = m->next_pg;
347 m->next_pg = NULL;
348 if (!__choose_path_in_pg(m, pg, nr_bytes))
349 return;
350 }
351
352 /* Don't change PG until it has no remaining paths */
353 if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
354 return;
355
356 /*
357 * Loop through priority groups until we find a valid path.
358 * First time we skip PGs marked 'bypassed'.
359 * Second time we only try the ones we skipped, but set
360 * pg_init_delay_retry so we do not hammer controllers.
361 */
362 do {
363 list_for_each_entry(pg, &m->priority_groups, list) {
364 if (pg->bypassed == bypassed)
365 continue;
366 if (!__choose_path_in_pg(m, pg, nr_bytes)) {
367 if (!bypassed)
368 m->pg_init_delay_retry = true;
369 return;
370 }
371 }
372 } while (bypassed--);
373
374failed:
375 m->current_pgpath = NULL;
376 m->current_pg = NULL;
377}
378
379/*
380 * Check whether bios must be queued in the device-mapper core rather
381 * than here in the target.
382 *
383 * m->lock must be held on entry.
384 *
385 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
386 * same value then we are not between multipath_presuspend()
387 * and multipath_resume() calls and we have no need to check
388 * for the DMF_NOFLUSH_SUSPENDING flag.
389 */
390static int __must_push_back(struct multipath *m)
391{
392 return (m->queue_if_no_path ||
393 (m->queue_if_no_path != m->saved_queue_if_no_path &&
394 dm_noflush_suspending(m->ti)));
395}
396
397/*
398 * Map cloned requests
399 */
400static int __multipath_map(struct dm_target *ti, struct request *clone,
401 union map_info *map_context,
402 struct request *rq, struct request **__clone)
403{
404 struct multipath *m = ti->private;
405 int r = DM_MAPIO_REQUEUE;
406 size_t nr_bytes = clone ? blk_rq_bytes(clone) : blk_rq_bytes(rq);
407 struct pgpath *pgpath;
408 struct block_device *bdev;
409 struct dm_mpath_io *mpio;
410
411 spin_lock_irq(&m->lock);
412
413 /* Do we need to select a new pgpath? */
414 if (!m->current_pgpath || !m->queue_io)
415 __choose_pgpath(m, nr_bytes);
416
417 pgpath = m->current_pgpath;
418
419 if (!pgpath) {
420 if (!__must_push_back(m))
421 r = -EIO; /* Failed */
422 goto out_unlock;
423 } else if (m->queue_io || m->pg_init_required) {
424 __pg_init_all_paths(m);
425 goto out_unlock;
426 }
427
428 mpio = set_mpio(m, map_context);
429 if (!mpio)
430 /* ENOMEM, requeue */
431 goto out_unlock;
432
433 mpio->pgpath = pgpath;
434 mpio->nr_bytes = nr_bytes;
435
436 bdev = pgpath->path.dev->bdev;
437
438 spin_unlock_irq(&m->lock);
439
440 if (clone) {
441 /*
442 * Old request-based interface: allocated clone is passed in.
443 * Used by: .request_fn stacked on .request_fn path(s).
444 */
445 clone->q = bdev_get_queue(bdev);
446 clone->rq_disk = bdev->bd_disk;
447 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
448 } else {
449 /*
450 * blk-mq request-based interface; used by both:
451 * .request_fn stacked on blk-mq path(s) and
452 * blk-mq stacked on blk-mq path(s).
453 */
454 *__clone = blk_mq_alloc_request(bdev_get_queue(bdev),
455 rq_data_dir(rq), BLK_MQ_REQ_NOWAIT);
456 if (IS_ERR(*__clone)) {
457 /* ENOMEM, requeue */
458 clear_request_fn_mpio(m, map_context);
459 return r;
460 }
461 (*__clone)->bio = (*__clone)->biotail = NULL;
462 (*__clone)->rq_disk = bdev->bd_disk;
463 (*__clone)->cmd_flags |= REQ_FAILFAST_TRANSPORT;
464 }
465
466 if (pgpath->pg->ps.type->start_io)
467 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
468 &pgpath->path,
469 nr_bytes);
470 return DM_MAPIO_REMAPPED;
471
472out_unlock:
473 spin_unlock_irq(&m->lock);
474
475 return r;
476}
477
478static int multipath_map(struct dm_target *ti, struct request *clone,
479 union map_info *map_context)
480{
481 return __multipath_map(ti, clone, map_context, NULL, NULL);
482}
483
484static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
485 union map_info *map_context,
486 struct request **clone)
487{
488 return __multipath_map(ti, NULL, map_context, rq, clone);
489}
490
491static void multipath_release_clone(struct request *clone)
492{
493 blk_mq_free_request(clone);
494}
495
496/*
497 * If we run out of usable paths, should we queue I/O or error it?
498 */
499static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
500 bool save_old_value)
501{
502 unsigned long flags;
503
504 spin_lock_irqsave(&m->lock, flags);
505
506 if (save_old_value)
507 m->saved_queue_if_no_path = m->queue_if_no_path;
508 else
509 m->saved_queue_if_no_path = queue_if_no_path;
510 m->queue_if_no_path = queue_if_no_path;
511 spin_unlock_irqrestore(&m->lock, flags);
512
513 if (!queue_if_no_path)
514 dm_table_run_md_queue_async(m->ti->table);
515
516 return 0;
517}
518
519/*
520 * An event is triggered whenever a path is taken out of use.
521 * Includes path failure and PG bypass.
522 */
523static void trigger_event(struct work_struct *work)
524{
525 struct multipath *m =
526 container_of(work, struct multipath, trigger_event);
527
528 dm_table_event(m->ti->table);
529}
530
531/*-----------------------------------------------------------------
532 * Constructor/argument parsing:
533 * <#multipath feature args> [<arg>]*
534 * <#hw_handler args> [hw_handler [<arg>]*]
535 * <#priority groups>
536 * <initial priority group>
537 * [<selector> <#selector args> [<arg>]*
538 * <#paths> <#per-path selector args>
539 * [<path> [<arg>]* ]+ ]+
540 *---------------------------------------------------------------*/
541static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
542 struct dm_target *ti)
543{
544 int r;
545 struct path_selector_type *pst;
546 unsigned ps_argc;
547
548 static struct dm_arg _args[] = {
549 {0, 1024, "invalid number of path selector args"},
550 };
551
552 pst = dm_get_path_selector(dm_shift_arg(as));
553 if (!pst) {
554 ti->error = "unknown path selector type";
555 return -EINVAL;
556 }
557
558 r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
559 if (r) {
560 dm_put_path_selector(pst);
561 return -EINVAL;
562 }
563
564 r = pst->create(&pg->ps, ps_argc, as->argv);
565 if (r) {
566 dm_put_path_selector(pst);
567 ti->error = "path selector constructor failed";
568 return r;
569 }
570
571 pg->ps.type = pst;
572 dm_consume_args(as, ps_argc);
573
574 return 0;
575}
576
577static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
578 struct dm_target *ti)
579{
580 int r;
581 struct pgpath *p;
582 struct multipath *m = ti->private;
583 struct request_queue *q = NULL;
584 const char *attached_handler_name;
585
586 /* we need at least a path arg */
587 if (as->argc < 1) {
588 ti->error = "no device given";
589 return ERR_PTR(-EINVAL);
590 }
591
592 p = alloc_pgpath();
593 if (!p)
594 return ERR_PTR(-ENOMEM);
595
596 r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
597 &p->path.dev);
598 if (r) {
599 ti->error = "error getting device";
600 goto bad;
601 }
602
603 if (m->retain_attached_hw_handler || m->hw_handler_name)
604 q = bdev_get_queue(p->path.dev->bdev);
605
606 if (m->retain_attached_hw_handler) {
607retain:
608 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
609 if (attached_handler_name) {
610 /*
611 * Reset hw_handler_name to match the attached handler
612 * and clear any hw_handler_params associated with the
613 * ignored handler.
614 *
615 * NB. This modifies the table line to show the actual
616 * handler instead of the original table passed in.
617 */
618 kfree(m->hw_handler_name);
619 m->hw_handler_name = attached_handler_name;
620
621 kfree(m->hw_handler_params);
622 m->hw_handler_params = NULL;
623 }
624 }
625
626 if (m->hw_handler_name) {
627 r = scsi_dh_attach(q, m->hw_handler_name);
628 if (r == -EBUSY) {
629 char b[BDEVNAME_SIZE];
630
631 printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
632 bdevname(p->path.dev->bdev, b));
633 goto retain;
634 }
635 if (r < 0) {
636 ti->error = "error attaching hardware handler";
637 dm_put_device(ti, p->path.dev);
638 goto bad;
639 }
640
641 if (m->hw_handler_params) {
642 r = scsi_dh_set_params(q, m->hw_handler_params);
643 if (r < 0) {
644 ti->error = "unable to set hardware "
645 "handler parameters";
646 dm_put_device(ti, p->path.dev);
647 goto bad;
648 }
649 }
650 }
651
652 r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
653 if (r) {
654 dm_put_device(ti, p->path.dev);
655 goto bad;
656 }
657
658 return p;
659
660 bad:
661 free_pgpath(p);
662 return ERR_PTR(r);
663}
664
665static struct priority_group *parse_priority_group(struct dm_arg_set *as,
666 struct multipath *m)
667{
668 static struct dm_arg _args[] = {
669 {1, 1024, "invalid number of paths"},
670 {0, 1024, "invalid number of selector args"}
671 };
672
673 int r;
674 unsigned i, nr_selector_args, nr_args;
675 struct priority_group *pg;
676 struct dm_target *ti = m->ti;
677
678 if (as->argc < 2) {
679 as->argc = 0;
680 ti->error = "not enough priority group arguments";
681 return ERR_PTR(-EINVAL);
682 }
683
684 pg = alloc_priority_group();
685 if (!pg) {
686 ti->error = "couldn't allocate priority group";
687 return ERR_PTR(-ENOMEM);
688 }
689 pg->m = m;
690
691 r = parse_path_selector(as, pg, ti);
692 if (r)
693 goto bad;
694
695 /*
696 * read the paths
697 */
698 r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
699 if (r)
700 goto bad;
701
702 r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
703 if (r)
704 goto bad;
705
706 nr_args = 1 + nr_selector_args;
707 for (i = 0; i < pg->nr_pgpaths; i++) {
708 struct pgpath *pgpath;
709 struct dm_arg_set path_args;
710
711 if (as->argc < nr_args) {
712 ti->error = "not enough path parameters";
713 r = -EINVAL;
714 goto bad;
715 }
716
717 path_args.argc = nr_args;
718 path_args.argv = as->argv;
719
720 pgpath = parse_path(&path_args, &pg->ps, ti);
721 if (IS_ERR(pgpath)) {
722 r = PTR_ERR(pgpath);
723 goto bad;
724 }
725
726 pgpath->pg = pg;
727 list_add_tail(&pgpath->list, &pg->pgpaths);
728 dm_consume_args(as, nr_args);
729 }
730
731 return pg;
732
733 bad:
734 free_priority_group(pg, ti);
735 return ERR_PTR(r);
736}
737
738static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
739{
740 unsigned hw_argc;
741 int ret;
742 struct dm_target *ti = m->ti;
743
744 static struct dm_arg _args[] = {
745 {0, 1024, "invalid number of hardware handler args"},
746 };
747
748 if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
749 return -EINVAL;
750
751 if (!hw_argc)
752 return 0;
753
754 m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
755
756 if (hw_argc > 1) {
757 char *p;
758 int i, j, len = 4;
759
760 for (i = 0; i <= hw_argc - 2; i++)
761 len += strlen(as->argv[i]) + 1;
762 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
763 if (!p) {
764 ti->error = "memory allocation failed";
765 ret = -ENOMEM;
766 goto fail;
767 }
768 j = sprintf(p, "%d", hw_argc - 1);
769 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
770 j = sprintf(p, "%s", as->argv[i]);
771 }
772 dm_consume_args(as, hw_argc - 1);
773
774 return 0;
775fail:
776 kfree(m->hw_handler_name);
777 m->hw_handler_name = NULL;
778 return ret;
779}
780
781static int parse_features(struct dm_arg_set *as, struct multipath *m)
782{
783 int r;
784 unsigned argc;
785 struct dm_target *ti = m->ti;
786 const char *arg_name;
787
788 static struct dm_arg _args[] = {
789 {0, 6, "invalid number of feature args"},
790 {1, 50, "pg_init_retries must be between 1 and 50"},
791 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
792 };
793
794 r = dm_read_arg_group(_args, as, &argc, &ti->error);
795 if (r)
796 return -EINVAL;
797
798 if (!argc)
799 return 0;
800
801 do {
802 arg_name = dm_shift_arg(as);
803 argc--;
804
805 if (!strcasecmp(arg_name, "queue_if_no_path")) {
806 r = queue_if_no_path(m, true, false);
807 continue;
808 }
809
810 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
811 m->retain_attached_hw_handler = true;
812 continue;
813 }
814
815 if (!strcasecmp(arg_name, "pg_init_retries") &&
816 (argc >= 1)) {
817 r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
818 argc--;
819 continue;
820 }
821
822 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
823 (argc >= 1)) {
824 r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
825 argc--;
826 continue;
827 }
828
829 ti->error = "Unrecognised multipath feature request";
830 r = -EINVAL;
831 } while (argc && !r);
832
833 return r;
834}
835
836static int multipath_ctr(struct dm_target *ti, unsigned int argc,
837 char **argv)
838{
839 /* target arguments */
840 static struct dm_arg _args[] = {
841 {0, 1024, "invalid number of priority groups"},
842 {0, 1024, "invalid initial priority group number"},
843 };
844
845 int r;
846 struct multipath *m;
847 struct dm_arg_set as;
848 unsigned pg_count = 0;
849 unsigned next_pg_num;
850 bool use_blk_mq = dm_use_blk_mq(dm_table_get_md(ti->table));
851
852 as.argc = argc;
853 as.argv = argv;
854
855 m = alloc_multipath(ti, use_blk_mq);
856 if (!m) {
857 ti->error = "can't allocate multipath";
858 return -EINVAL;
859 }
860
861 r = parse_features(&as, m);
862 if (r)
863 goto bad;
864
865 r = parse_hw_handler(&as, m);
866 if (r)
867 goto bad;
868
869 r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
870 if (r)
871 goto bad;
872
873 r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
874 if (r)
875 goto bad;
876
877 if ((!m->nr_priority_groups && next_pg_num) ||
878 (m->nr_priority_groups && !next_pg_num)) {
879 ti->error = "invalid initial priority group";
880 r = -EINVAL;
881 goto bad;
882 }
883
884 /* parse the priority groups */
885 while (as.argc) {
886 struct priority_group *pg;
887
888 pg = parse_priority_group(&as, m);
889 if (IS_ERR(pg)) {
890 r = PTR_ERR(pg);
891 goto bad;
892 }
893
894 m->nr_valid_paths += pg->nr_pgpaths;
895 list_add_tail(&pg->list, &m->priority_groups);
896 pg_count++;
897 pg->pg_num = pg_count;
898 if (!--next_pg_num)
899 m->next_pg = pg;
900 }
901
902 if (pg_count != m->nr_priority_groups) {
903 ti->error = "priority group count mismatch";
904 r = -EINVAL;
905 goto bad;
906 }
907
908 ti->num_flush_bios = 1;
909 ti->num_discard_bios = 1;
910 ti->num_write_same_bios = 1;
911 if (use_blk_mq)
912 ti->per_io_data_size = sizeof(struct dm_mpath_io);
913
914 return 0;
915
916 bad:
917 free_multipath(m);
918 return r;
919}
920
921static void multipath_wait_for_pg_init_completion(struct multipath *m)
922{
923 DECLARE_WAITQUEUE(wait, current);
924 unsigned long flags;
925
926 add_wait_queue(&m->pg_init_wait, &wait);
927
928 while (1) {
929 set_current_state(TASK_UNINTERRUPTIBLE);
930
931 spin_lock_irqsave(&m->lock, flags);
932 if (!m->pg_init_in_progress) {
933 spin_unlock_irqrestore(&m->lock, flags);
934 break;
935 }
936 spin_unlock_irqrestore(&m->lock, flags);
937
938 io_schedule();
939 }
940 set_current_state(TASK_RUNNING);
941
942 remove_wait_queue(&m->pg_init_wait, &wait);
943}
944
945static void flush_multipath_work(struct multipath *m)
946{
947 unsigned long flags;
948
949 spin_lock_irqsave(&m->lock, flags);
950 m->pg_init_disabled = true;
951 spin_unlock_irqrestore(&m->lock, flags);
952
953 flush_workqueue(kmpath_handlerd);
954 multipath_wait_for_pg_init_completion(m);
955 flush_workqueue(kmultipathd);
956 flush_work(&m->trigger_event);
957
958 spin_lock_irqsave(&m->lock, flags);
959 m->pg_init_disabled = false;
960 spin_unlock_irqrestore(&m->lock, flags);
961}
962
963static void multipath_dtr(struct dm_target *ti)
964{
965 struct multipath *m = ti->private;
966
967 flush_multipath_work(m);
968 free_multipath(m);
969}
970
971/*
972 * Take a path out of use.
973 */
974static int fail_path(struct pgpath *pgpath)
975{
976 unsigned long flags;
977 struct multipath *m = pgpath->pg->m;
978
979 spin_lock_irqsave(&m->lock, flags);
980
981 if (!pgpath->is_active)
982 goto out;
983
984 DMWARN("Failing path %s.", pgpath->path.dev->name);
985
986 pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
987 pgpath->is_active = false;
988 pgpath->fail_count++;
989
990 m->nr_valid_paths--;
991
992 if (pgpath == m->current_pgpath)
993 m->current_pgpath = NULL;
994
995 dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
996 pgpath->path.dev->name, m->nr_valid_paths);
997
998 schedule_work(&m->trigger_event);
999
1000out:
1001 spin_unlock_irqrestore(&m->lock, flags);
1002
1003 return 0;
1004}
1005
1006/*
1007 * Reinstate a previously-failed path
1008 */
1009static int reinstate_path(struct pgpath *pgpath)
1010{
1011 int r = 0, run_queue = 0;
1012 unsigned long flags;
1013 struct multipath *m = pgpath->pg->m;
1014
1015 spin_lock_irqsave(&m->lock, flags);
1016
1017 if (pgpath->is_active)
1018 goto out;
1019
1020 DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1021
1022 r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1023 if (r)
1024 goto out;
1025
1026 pgpath->is_active = true;
1027
1028 if (!m->nr_valid_paths++) {
1029 m->current_pgpath = NULL;
1030 run_queue = 1;
1031 } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1032 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1033 m->pg_init_in_progress++;
1034 }
1035
1036 dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1037 pgpath->path.dev->name, m->nr_valid_paths);
1038
1039 schedule_work(&m->trigger_event);
1040
1041out:
1042 spin_unlock_irqrestore(&m->lock, flags);
1043 if (run_queue)
1044 dm_table_run_md_queue_async(m->ti->table);
1045
1046 return r;
1047}
1048
1049/*
1050 * Fail or reinstate all paths that match the provided struct dm_dev.
1051 */
1052static int action_dev(struct multipath *m, struct dm_dev *dev,
1053 action_fn action)
1054{
1055 int r = -EINVAL;
1056 struct pgpath *pgpath;
1057 struct priority_group *pg;
1058
1059 list_for_each_entry(pg, &m->priority_groups, list) {
1060 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1061 if (pgpath->path.dev == dev)
1062 r = action(pgpath);
1063 }
1064 }
1065
1066 return r;
1067}
1068
1069/*
1070 * Temporarily try to avoid having to use the specified PG
1071 */
1072static void bypass_pg(struct multipath *m, struct priority_group *pg,
1073 bool bypassed)
1074{
1075 unsigned long flags;
1076
1077 spin_lock_irqsave(&m->lock, flags);
1078
1079 pg->bypassed = bypassed;
1080 m->current_pgpath = NULL;
1081 m->current_pg = NULL;
1082
1083 spin_unlock_irqrestore(&m->lock, flags);
1084
1085 schedule_work(&m->trigger_event);
1086}
1087
1088/*
1089 * Switch to using the specified PG from the next I/O that gets mapped
1090 */
1091static int switch_pg_num(struct multipath *m, const char *pgstr)
1092{
1093 struct priority_group *pg;
1094 unsigned pgnum;
1095 unsigned long flags;
1096 char dummy;
1097
1098 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1099 (pgnum > m->nr_priority_groups)) {
1100 DMWARN("invalid PG number supplied to switch_pg_num");
1101 return -EINVAL;
1102 }
1103
1104 spin_lock_irqsave(&m->lock, flags);
1105 list_for_each_entry(pg, &m->priority_groups, list) {
1106 pg->bypassed = false;
1107 if (--pgnum)
1108 continue;
1109
1110 m->current_pgpath = NULL;
1111 m->current_pg = NULL;
1112 m->next_pg = pg;
1113 }
1114 spin_unlock_irqrestore(&m->lock, flags);
1115
1116 schedule_work(&m->trigger_event);
1117 return 0;
1118}
1119
1120/*
1121 * Set/clear bypassed status of a PG.
1122 * PGs are numbered upwards from 1 in the order they were declared.
1123 */
1124static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1125{
1126 struct priority_group *pg;
1127 unsigned pgnum;
1128 char dummy;
1129
1130 if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1131 (pgnum > m->nr_priority_groups)) {
1132 DMWARN("invalid PG number supplied to bypass_pg");
1133 return -EINVAL;
1134 }
1135
1136 list_for_each_entry(pg, &m->priority_groups, list) {
1137 if (!--pgnum)
1138 break;
1139 }
1140
1141 bypass_pg(m, pg, bypassed);
1142 return 0;
1143}
1144
1145/*
1146 * Should we retry pg_init immediately?
1147 */
1148static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1149{
1150 unsigned long flags;
1151 bool limit_reached = false;
1152
1153 spin_lock_irqsave(&m->lock, flags);
1154
1155 if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1156 m->pg_init_required = true;
1157 else
1158 limit_reached = true;
1159
1160 spin_unlock_irqrestore(&m->lock, flags);
1161
1162 return limit_reached;
1163}
1164
1165static void pg_init_done(void *data, int errors)
1166{
1167 struct pgpath *pgpath = data;
1168 struct priority_group *pg = pgpath->pg;
1169 struct multipath *m = pg->m;
1170 unsigned long flags;
1171 bool delay_retry = false;
1172
1173 /* device or driver problems */
1174 switch (errors) {
1175 case SCSI_DH_OK:
1176 break;
1177 case SCSI_DH_NOSYS:
1178 if (!m->hw_handler_name) {
1179 errors = 0;
1180 break;
1181 }
1182 DMERR("Could not failover the device: Handler scsi_dh_%s "
1183 "Error %d.", m->hw_handler_name, errors);
1184 /*
1185 * Fail path for now, so we do not ping pong
1186 */
1187 fail_path(pgpath);
1188 break;
1189 case SCSI_DH_DEV_TEMP_BUSY:
1190 /*
1191 * Probably doing something like FW upgrade on the
1192 * controller so try the other pg.
1193 */
1194 bypass_pg(m, pg, true);
1195 break;
1196 case SCSI_DH_RETRY:
1197 /* Wait before retrying. */
1198 delay_retry = 1;
1199 case SCSI_DH_IMM_RETRY:
1200 case SCSI_DH_RES_TEMP_UNAVAIL:
1201 if (pg_init_limit_reached(m, pgpath))
1202 fail_path(pgpath);
1203 errors = 0;
1204 break;
1205 case SCSI_DH_DEV_OFFLINED:
1206 default:
1207 /*
1208 * We probably do not want to fail the path for a device
1209 * error, but this is what the old dm did. In future
1210 * patches we can do more advanced handling.
1211 */
1212 fail_path(pgpath);
1213 }
1214
1215 spin_lock_irqsave(&m->lock, flags);
1216 if (errors) {
1217 if (pgpath == m->current_pgpath) {
1218 DMERR("Could not failover device. Error %d.", errors);
1219 m->current_pgpath = NULL;
1220 m->current_pg = NULL;
1221 }
1222 } else if (!m->pg_init_required)
1223 pg->bypassed = false;
1224
1225 if (--m->pg_init_in_progress)
1226 /* Activations of other paths are still on going */
1227 goto out;
1228
1229 if (m->pg_init_required) {
1230 m->pg_init_delay_retry = delay_retry;
1231 if (__pg_init_all_paths(m))
1232 goto out;
1233 }
1234 m->queue_io = false;
1235
1236 /*
1237 * Wake up any thread waiting to suspend.
1238 */
1239 wake_up(&m->pg_init_wait);
1240
1241out:
1242 spin_unlock_irqrestore(&m->lock, flags);
1243}
1244
1245static void activate_path(struct work_struct *work)
1246{
1247 struct pgpath *pgpath =
1248 container_of(work, struct pgpath, activate_path.work);
1249
1250 if (pgpath->is_active)
1251 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1252 pg_init_done, pgpath);
1253 else
1254 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1255}
1256
1257static int noretry_error(int error)
1258{
1259 switch (error) {
1260 case -EOPNOTSUPP:
1261 case -EREMOTEIO:
1262 case -EILSEQ:
1263 case -ENODATA:
1264 case -ENOSPC:
1265 return 1;
1266 }
1267
1268 /* Anything else could be a path failure, so should be retried */
1269 return 0;
1270}
1271
1272/*
1273 * end_io handling
1274 */
1275static int do_end_io(struct multipath *m, struct request *clone,
1276 int error, struct dm_mpath_io *mpio)
1277{
1278 /*
1279 * We don't queue any clone request inside the multipath target
1280 * during end I/O handling, since those clone requests don't have
1281 * bio clones. If we queue them inside the multipath target,
1282 * we need to make bio clones, that requires memory allocation.
1283 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1284 * don't have bio clones.)
1285 * Instead of queueing the clone request here, we queue the original
1286 * request into dm core, which will remake a clone request and
1287 * clone bios for it and resubmit it later.
1288 */
1289 int r = DM_ENDIO_REQUEUE;
1290 unsigned long flags;
1291
1292 if (!error && !clone->errors)
1293 return 0; /* I/O complete */
1294
1295 if (noretry_error(error))
1296 return error;
1297
1298 if (mpio->pgpath)
1299 fail_path(mpio->pgpath);
1300
1301 spin_lock_irqsave(&m->lock, flags);
1302 if (!m->nr_valid_paths) {
1303 if (!m->queue_if_no_path) {
1304 if (!__must_push_back(m))
1305 r = -EIO;
1306 } else {
1307 if (error == -EBADE)
1308 r = error;
1309 }
1310 }
1311 spin_unlock_irqrestore(&m->lock, flags);
1312
1313 return r;
1314}
1315
1316static int multipath_end_io(struct dm_target *ti, struct request *clone,
1317 int error, union map_info *map_context)
1318{
1319 struct multipath *m = ti->private;
1320 struct dm_mpath_io *mpio = get_mpio(map_context);
1321 struct pgpath *pgpath;
1322 struct path_selector *ps;
1323 int r;
1324
1325 BUG_ON(!mpio);
1326
1327 r = do_end_io(m, clone, error, mpio);
1328 pgpath = mpio->pgpath;
1329 if (pgpath) {
1330 ps = &pgpath->pg->ps;
1331 if (ps->type->end_io)
1332 ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1333 }
1334 clear_request_fn_mpio(m, map_context);
1335
1336 return r;
1337}
1338
1339/*
1340 * Suspend can't complete until all the I/O is processed so if
1341 * the last path fails we must error any remaining I/O.
1342 * Note that if the freeze_bdev fails while suspending, the
1343 * queue_if_no_path state is lost - userspace should reset it.
1344 */
1345static void multipath_presuspend(struct dm_target *ti)
1346{
1347 struct multipath *m = ti->private;
1348
1349 queue_if_no_path(m, false, true);
1350}
1351
1352static void multipath_postsuspend(struct dm_target *ti)
1353{
1354 struct multipath *m = ti->private;
1355
1356 mutex_lock(&m->work_mutex);
1357 flush_multipath_work(m);
1358 mutex_unlock(&m->work_mutex);
1359}
1360
1361/*
1362 * Restore the queue_if_no_path setting.
1363 */
1364static void multipath_resume(struct dm_target *ti)
1365{
1366 struct multipath *m = ti->private;
1367 unsigned long flags;
1368
1369 spin_lock_irqsave(&m->lock, flags);
1370 m->queue_if_no_path = m->saved_queue_if_no_path;
1371 spin_unlock_irqrestore(&m->lock, flags);
1372}
1373
1374/*
1375 * Info output has the following format:
1376 * num_multipath_feature_args [multipath_feature_args]*
1377 * num_handler_status_args [handler_status_args]*
1378 * num_groups init_group_number
1379 * [A|D|E num_ps_status_args [ps_status_args]*
1380 * num_paths num_selector_args
1381 * [path_dev A|F fail_count [selector_args]* ]+ ]+
1382 *
1383 * Table output has the following format (identical to the constructor string):
1384 * num_feature_args [features_args]*
1385 * num_handler_args hw_handler [hw_handler_args]*
1386 * num_groups init_group_number
1387 * [priority selector-name num_ps_args [ps_args]*
1388 * num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1389 */
1390static void multipath_status(struct dm_target *ti, status_type_t type,
1391 unsigned status_flags, char *result, unsigned maxlen)
1392{
1393 int sz = 0;
1394 unsigned long flags;
1395 struct multipath *m = ti->private;
1396 struct priority_group *pg;
1397 struct pgpath *p;
1398 unsigned pg_num;
1399 char state;
1400
1401 spin_lock_irqsave(&m->lock, flags);
1402
1403 /* Features */
1404 if (type == STATUSTYPE_INFO)
1405 DMEMIT("2 %u %u ", m->queue_io, m->pg_init_count);
1406 else {
1407 DMEMIT("%u ", m->queue_if_no_path +
1408 (m->pg_init_retries > 0) * 2 +
1409 (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1410 m->retain_attached_hw_handler);
1411 if (m->queue_if_no_path)
1412 DMEMIT("queue_if_no_path ");
1413 if (m->pg_init_retries)
1414 DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1415 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1416 DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1417 if (m->retain_attached_hw_handler)
1418 DMEMIT("retain_attached_hw_handler ");
1419 }
1420
1421 if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1422 DMEMIT("0 ");
1423 else
1424 DMEMIT("1 %s ", m->hw_handler_name);
1425
1426 DMEMIT("%u ", m->nr_priority_groups);
1427
1428 if (m->next_pg)
1429 pg_num = m->next_pg->pg_num;
1430 else if (m->current_pg)
1431 pg_num = m->current_pg->pg_num;
1432 else
1433 pg_num = (m->nr_priority_groups ? 1 : 0);
1434
1435 DMEMIT("%u ", pg_num);
1436
1437 switch (type) {
1438 case STATUSTYPE_INFO:
1439 list_for_each_entry(pg, &m->priority_groups, list) {
1440 if (pg->bypassed)
1441 state = 'D'; /* Disabled */
1442 else if (pg == m->current_pg)
1443 state = 'A'; /* Currently Active */
1444 else
1445 state = 'E'; /* Enabled */
1446
1447 DMEMIT("%c ", state);
1448
1449 if (pg->ps.type->status)
1450 sz += pg->ps.type->status(&pg->ps, NULL, type,
1451 result + sz,
1452 maxlen - sz);
1453 else
1454 DMEMIT("0 ");
1455
1456 DMEMIT("%u %u ", pg->nr_pgpaths,
1457 pg->ps.type->info_args);
1458
1459 list_for_each_entry(p, &pg->pgpaths, list) {
1460 DMEMIT("%s %s %u ", p->path.dev->name,
1461 p->is_active ? "A" : "F",
1462 p->fail_count);
1463 if (pg->ps.type->status)
1464 sz += pg->ps.type->status(&pg->ps,
1465 &p->path, type, result + sz,
1466 maxlen - sz);
1467 }
1468 }
1469 break;
1470
1471 case STATUSTYPE_TABLE:
1472 list_for_each_entry(pg, &m->priority_groups, list) {
1473 DMEMIT("%s ", pg->ps.type->name);
1474
1475 if (pg->ps.type->status)
1476 sz += pg->ps.type->status(&pg->ps, NULL, type,
1477 result + sz,
1478 maxlen - sz);
1479 else
1480 DMEMIT("0 ");
1481
1482 DMEMIT("%u %u ", pg->nr_pgpaths,
1483 pg->ps.type->table_args);
1484
1485 list_for_each_entry(p, &pg->pgpaths, list) {
1486 DMEMIT("%s ", p->path.dev->name);
1487 if (pg->ps.type->status)
1488 sz += pg->ps.type->status(&pg->ps,
1489 &p->path, type, result + sz,
1490 maxlen - sz);
1491 }
1492 }
1493 break;
1494 }
1495
1496 spin_unlock_irqrestore(&m->lock, flags);
1497}
1498
1499static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1500{
1501 int r = -EINVAL;
1502 struct dm_dev *dev;
1503 struct multipath *m = ti->private;
1504 action_fn action;
1505
1506 mutex_lock(&m->work_mutex);
1507
1508 if (dm_suspended(ti)) {
1509 r = -EBUSY;
1510 goto out;
1511 }
1512
1513 if (argc == 1) {
1514 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1515 r = queue_if_no_path(m, true, false);
1516 goto out;
1517 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1518 r = queue_if_no_path(m, false, false);
1519 goto out;
1520 }
1521 }
1522
1523 if (argc != 2) {
1524 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1525 goto out;
1526 }
1527
1528 if (!strcasecmp(argv[0], "disable_group")) {
1529 r = bypass_pg_num(m, argv[1], true);
1530 goto out;
1531 } else if (!strcasecmp(argv[0], "enable_group")) {
1532 r = bypass_pg_num(m, argv[1], false);
1533 goto out;
1534 } else if (!strcasecmp(argv[0], "switch_group")) {
1535 r = switch_pg_num(m, argv[1]);
1536 goto out;
1537 } else if (!strcasecmp(argv[0], "reinstate_path"))
1538 action = reinstate_path;
1539 else if (!strcasecmp(argv[0], "fail_path"))
1540 action = fail_path;
1541 else {
1542 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1543 goto out;
1544 }
1545
1546 r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1547 if (r) {
1548 DMWARN("message: error getting device %s",
1549 argv[1]);
1550 goto out;
1551 }
1552
1553 r = action_dev(m, dev, action);
1554
1555 dm_put_device(ti, dev);
1556
1557out:
1558 mutex_unlock(&m->work_mutex);
1559 return r;
1560}
1561
1562static int multipath_prepare_ioctl(struct dm_target *ti,
1563 struct block_device **bdev, fmode_t *mode)
1564{
1565 struct multipath *m = ti->private;
1566 unsigned long flags;
1567 int r;
1568
1569 spin_lock_irqsave(&m->lock, flags);
1570
1571 if (!m->current_pgpath)
1572 __choose_pgpath(m, 0);
1573
1574 if (m->current_pgpath) {
1575 if (!m->queue_io) {
1576 *bdev = m->current_pgpath->path.dev->bdev;
1577 *mode = m->current_pgpath->path.dev->mode;
1578 r = 0;
1579 } else {
1580 /* pg_init has not started or completed */
1581 r = -ENOTCONN;
1582 }
1583 } else {
1584 /* No path is available */
1585 if (m->queue_if_no_path)
1586 r = -ENOTCONN;
1587 else
1588 r = -EIO;
1589 }
1590
1591 spin_unlock_irqrestore(&m->lock, flags);
1592
1593 if (r == -ENOTCONN) {
1594 spin_lock_irqsave(&m->lock, flags);
1595 if (!m->current_pg) {
1596 /* Path status changed, redo selection */
1597 __choose_pgpath(m, 0);
1598 }
1599 if (m->pg_init_required)
1600 __pg_init_all_paths(m);
1601 spin_unlock_irqrestore(&m->lock, flags);
1602 dm_table_run_md_queue_async(m->ti->table);
1603 }
1604
1605 /*
1606 * Only pass ioctls through if the device sizes match exactly.
1607 */
1608 if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1609 return 1;
1610 return r;
1611}
1612
1613static int multipath_iterate_devices(struct dm_target *ti,
1614 iterate_devices_callout_fn fn, void *data)
1615{
1616 struct multipath *m = ti->private;
1617 struct priority_group *pg;
1618 struct pgpath *p;
1619 int ret = 0;
1620
1621 list_for_each_entry(pg, &m->priority_groups, list) {
1622 list_for_each_entry(p, &pg->pgpaths, list) {
1623 ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1624 if (ret)
1625 goto out;
1626 }
1627 }
1628
1629out:
1630 return ret;
1631}
1632
1633static int pgpath_busy(struct pgpath *pgpath)
1634{
1635 struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1636
1637 return blk_lld_busy(q);
1638}
1639
1640/*
1641 * We return "busy", only when we can map I/Os but underlying devices
1642 * are busy (so even if we map I/Os now, the I/Os will wait on
1643 * the underlying queue).
1644 * In other words, if we want to kill I/Os or queue them inside us
1645 * due to map unavailability, we don't return "busy". Otherwise,
1646 * dm core won't give us the I/Os and we can't do what we want.
1647 */
1648static int multipath_busy(struct dm_target *ti)
1649{
1650 bool busy = false, has_active = false;
1651 struct multipath *m = ti->private;
1652 struct priority_group *pg;
1653 struct pgpath *pgpath;
1654 unsigned long flags;
1655
1656 spin_lock_irqsave(&m->lock, flags);
1657
1658 /* pg_init in progress or no paths available */
1659 if (m->pg_init_in_progress ||
1660 (!m->nr_valid_paths && m->queue_if_no_path)) {
1661 busy = true;
1662 goto out;
1663 }
1664 /* Guess which priority_group will be used at next mapping time */
1665 if (unlikely(!m->current_pgpath && m->next_pg))
1666 pg = m->next_pg;
1667 else if (likely(m->current_pg))
1668 pg = m->current_pg;
1669 else
1670 /*
1671 * We don't know which pg will be used at next mapping time.
1672 * We don't call __choose_pgpath() here to avoid to trigger
1673 * pg_init just by busy checking.
1674 * So we don't know whether underlying devices we will be using
1675 * at next mapping time are busy or not. Just try mapping.
1676 */
1677 goto out;
1678
1679 /*
1680 * If there is one non-busy active path at least, the path selector
1681 * will be able to select it. So we consider such a pg as not busy.
1682 */
1683 busy = true;
1684 list_for_each_entry(pgpath, &pg->pgpaths, list)
1685 if (pgpath->is_active) {
1686 has_active = true;
1687 if (!pgpath_busy(pgpath)) {
1688 busy = false;
1689 break;
1690 }
1691 }
1692
1693 if (!has_active)
1694 /*
1695 * No active path in this pg, so this pg won't be used and
1696 * the current_pg will be changed at next mapping time.
1697 * We need to try mapping to determine it.
1698 */
1699 busy = false;
1700
1701out:
1702 spin_unlock_irqrestore(&m->lock, flags);
1703
1704 return busy;
1705}
1706
1707/*-----------------------------------------------------------------
1708 * Module setup
1709 *---------------------------------------------------------------*/
1710static struct target_type multipath_target = {
1711 .name = "multipath",
1712 .version = {1, 11, 0},
1713 .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1714 .module = THIS_MODULE,
1715 .ctr = multipath_ctr,
1716 .dtr = multipath_dtr,
1717 .map_rq = multipath_map,
1718 .clone_and_map_rq = multipath_clone_and_map,
1719 .release_clone_rq = multipath_release_clone,
1720 .rq_end_io = multipath_end_io,
1721 .presuspend = multipath_presuspend,
1722 .postsuspend = multipath_postsuspend,
1723 .resume = multipath_resume,
1724 .status = multipath_status,
1725 .message = multipath_message,
1726 .prepare_ioctl = multipath_prepare_ioctl,
1727 .iterate_devices = multipath_iterate_devices,
1728 .busy = multipath_busy,
1729};
1730
1731static int __init dm_multipath_init(void)
1732{
1733 int r;
1734
1735 /* allocate a slab for the dm_ios */
1736 _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1737 if (!_mpio_cache)
1738 return -ENOMEM;
1739
1740 r = dm_register_target(&multipath_target);
1741 if (r < 0) {
1742 DMERR("register failed %d", r);
1743 r = -EINVAL;
1744 goto bad_register_target;
1745 }
1746
1747 kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1748 if (!kmultipathd) {
1749 DMERR("failed to create workqueue kmpathd");
1750 r = -ENOMEM;
1751 goto bad_alloc_kmultipathd;
1752 }
1753
1754 /*
1755 * A separate workqueue is used to handle the device handlers
1756 * to avoid overloading existing workqueue. Overloading the
1757 * old workqueue would also create a bottleneck in the
1758 * path of the storage hardware device activation.
1759 */
1760 kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1761 WQ_MEM_RECLAIM);
1762 if (!kmpath_handlerd) {
1763 DMERR("failed to create workqueue kmpath_handlerd");
1764 r = -ENOMEM;
1765 goto bad_alloc_kmpath_handlerd;
1766 }
1767
1768 DMINFO("version %u.%u.%u loaded",
1769 multipath_target.version[0], multipath_target.version[1],
1770 multipath_target.version[2]);
1771
1772 return 0;
1773
1774bad_alloc_kmpath_handlerd:
1775 destroy_workqueue(kmultipathd);
1776bad_alloc_kmultipathd:
1777 dm_unregister_target(&multipath_target);
1778bad_register_target:
1779 kmem_cache_destroy(_mpio_cache);
1780
1781 return r;
1782}
1783
1784static void __exit dm_multipath_exit(void)
1785{
1786 destroy_workqueue(kmpath_handlerd);
1787 destroy_workqueue(kmultipathd);
1788
1789 dm_unregister_target(&multipath_target);
1790 kmem_cache_destroy(_mpio_cache);
1791}
1792
1793module_init(dm_multipath_init);
1794module_exit(dm_multipath_exit);
1795
1796MODULE_DESCRIPTION(DM_NAME " multipath target");
1797MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1798MODULE_LICENSE("GPL");