Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dledford/rdma
[linux-2.6-microblaze.git] / drivers / md / dm-mpath.c
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 */
34 struct 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  */
52 struct 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 */
66 struct 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         enum dm_queue_mode queue_mode;
94
95         struct mutex work_mutex;
96         struct work_struct trigger_event;
97
98         struct work_struct process_queued_bios;
99         struct bio_list queued_bios;
100 };
101
102 /*
103  * Context information attached to each io we process.
104  */
105 struct dm_mpath_io {
106         struct pgpath *pgpath;
107         size_t nr_bytes;
108 };
109
110 typedef int (*action_fn) (struct pgpath *pgpath);
111
112 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
113 static void trigger_event(struct work_struct *work);
114 static void activate_or_offline_path(struct pgpath *pgpath);
115 static void activate_path_work(struct work_struct *work);
116 static void process_queued_bios(struct work_struct *work);
117
118 /*-----------------------------------------------
119  * Multipath state flags.
120  *-----------------------------------------------*/
121
122 #define MPATHF_QUEUE_IO 0                       /* Must we queue all I/O? */
123 #define MPATHF_QUEUE_IF_NO_PATH 1               /* Queue I/O if last path fails? */
124 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2         /* Saved state during suspension */
125 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3     /* If there's already a hw_handler present, don't change it. */
126 #define MPATHF_PG_INIT_DISABLED 4               /* pg_init is not currently allowed */
127 #define MPATHF_PG_INIT_REQUIRED 5               /* pg_init needs calling? */
128 #define MPATHF_PG_INIT_DELAY_RETRY 6            /* Delay pg_init retry? */
129
130 /*-----------------------------------------------
131  * Allocation routines
132  *-----------------------------------------------*/
133
134 static struct pgpath *alloc_pgpath(void)
135 {
136         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
137
138         if (pgpath) {
139                 pgpath->is_active = true;
140                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path_work);
141         }
142
143         return pgpath;
144 }
145
146 static void free_pgpath(struct pgpath *pgpath)
147 {
148         kfree(pgpath);
149 }
150
151 static struct priority_group *alloc_priority_group(void)
152 {
153         struct priority_group *pg;
154
155         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
156
157         if (pg)
158                 INIT_LIST_HEAD(&pg->pgpaths);
159
160         return pg;
161 }
162
163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
164 {
165         struct pgpath *pgpath, *tmp;
166
167         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
168                 list_del(&pgpath->list);
169                 dm_put_device(ti, pgpath->path.dev);
170                 free_pgpath(pgpath);
171         }
172 }
173
174 static void free_priority_group(struct priority_group *pg,
175                                 struct dm_target *ti)
176 {
177         struct path_selector *ps = &pg->ps;
178
179         if (ps->type) {
180                 ps->type->destroy(ps);
181                 dm_put_path_selector(ps->type);
182         }
183
184         free_pgpaths(&pg->pgpaths, ti);
185         kfree(pg);
186 }
187
188 static struct multipath *alloc_multipath(struct dm_target *ti)
189 {
190         struct multipath *m;
191
192         m = kzalloc(sizeof(*m), GFP_KERNEL);
193         if (m) {
194                 INIT_LIST_HEAD(&m->priority_groups);
195                 spin_lock_init(&m->lock);
196                 set_bit(MPATHF_QUEUE_IO, &m->flags);
197                 atomic_set(&m->nr_valid_paths, 0);
198                 atomic_set(&m->pg_init_in_progress, 0);
199                 atomic_set(&m->pg_init_count, 0);
200                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
201                 INIT_WORK(&m->trigger_event, trigger_event);
202                 init_waitqueue_head(&m->pg_init_wait);
203                 mutex_init(&m->work_mutex);
204
205                 m->queue_mode = DM_TYPE_NONE;
206
207                 m->ti = ti;
208                 ti->private = m;
209         }
210
211         return m;
212 }
213
214 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
215 {
216         if (m->queue_mode == DM_TYPE_NONE) {
217                 /*
218                  * Default to request-based.
219                  */
220                 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
221                         m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
222                 else
223                         m->queue_mode = DM_TYPE_REQUEST_BASED;
224         } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
225                 INIT_WORK(&m->process_queued_bios, process_queued_bios);
226                 /*
227                  * bio-based doesn't support any direct scsi_dh management;
228                  * it just discovers if a scsi_dh is attached.
229                  */
230                 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
231         }
232
233         dm_table_set_type(ti->table, m->queue_mode);
234
235         return 0;
236 }
237
238 static void free_multipath(struct multipath *m)
239 {
240         struct priority_group *pg, *tmp;
241
242         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
243                 list_del(&pg->list);
244                 free_priority_group(pg, m->ti);
245         }
246
247         kfree(m->hw_handler_name);
248         kfree(m->hw_handler_params);
249         kfree(m);
250 }
251
252 static struct dm_mpath_io *get_mpio(union map_info *info)
253 {
254         return info->ptr;
255 }
256
257 static size_t multipath_per_bio_data_size(void)
258 {
259         return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
260 }
261
262 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
263 {
264         return dm_per_bio_data(bio, multipath_per_bio_data_size());
265 }
266
267 static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
268 {
269         /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
270         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
271         void *bio_details = mpio + 1;
272
273         return bio_details;
274 }
275
276 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p,
277                                         struct dm_bio_details **bio_details_p)
278 {
279         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
280         struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
281
282         memset(mpio, 0, sizeof(*mpio));
283         memset(bio_details, 0, sizeof(*bio_details));
284         dm_bio_record(bio_details, bio);
285
286         if (mpio_p)
287                 *mpio_p = mpio;
288         if (bio_details_p)
289                 *bio_details_p = bio_details;
290 }
291
292 /*-----------------------------------------------
293  * Path selection
294  *-----------------------------------------------*/
295
296 static int __pg_init_all_paths(struct multipath *m)
297 {
298         struct pgpath *pgpath;
299         unsigned long pg_init_delay = 0;
300
301         lockdep_assert_held(&m->lock);
302
303         if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
304                 return 0;
305
306         atomic_inc(&m->pg_init_count);
307         clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
308
309         /* Check here to reset pg_init_required */
310         if (!m->current_pg)
311                 return 0;
312
313         if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
314                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
315                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
316         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
317                 /* Skip failed paths */
318                 if (!pgpath->is_active)
319                         continue;
320                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
321                                        pg_init_delay))
322                         atomic_inc(&m->pg_init_in_progress);
323         }
324         return atomic_read(&m->pg_init_in_progress);
325 }
326
327 static int pg_init_all_paths(struct multipath *m)
328 {
329         int ret;
330         unsigned long flags;
331
332         spin_lock_irqsave(&m->lock, flags);
333         ret = __pg_init_all_paths(m);
334         spin_unlock_irqrestore(&m->lock, flags);
335
336         return ret;
337 }
338
339 static void __switch_pg(struct multipath *m, struct priority_group *pg)
340 {
341         m->current_pg = pg;
342
343         /* Must we initialise the PG first, and queue I/O till it's ready? */
344         if (m->hw_handler_name) {
345                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
346                 set_bit(MPATHF_QUEUE_IO, &m->flags);
347         } else {
348                 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
349                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
350         }
351
352         atomic_set(&m->pg_init_count, 0);
353 }
354
355 static struct pgpath *choose_path_in_pg(struct multipath *m,
356                                         struct priority_group *pg,
357                                         size_t nr_bytes)
358 {
359         unsigned long flags;
360         struct dm_path *path;
361         struct pgpath *pgpath;
362
363         path = pg->ps.type->select_path(&pg->ps, nr_bytes);
364         if (!path)
365                 return ERR_PTR(-ENXIO);
366
367         pgpath = path_to_pgpath(path);
368
369         if (unlikely(READ_ONCE(m->current_pg) != pg)) {
370                 /* Only update current_pgpath if pg changed */
371                 spin_lock_irqsave(&m->lock, flags);
372                 m->current_pgpath = pgpath;
373                 __switch_pg(m, pg);
374                 spin_unlock_irqrestore(&m->lock, flags);
375         }
376
377         return pgpath;
378 }
379
380 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
381 {
382         unsigned long flags;
383         struct priority_group *pg;
384         struct pgpath *pgpath;
385         unsigned bypassed = 1;
386
387         if (!atomic_read(&m->nr_valid_paths)) {
388                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
389                 goto failed;
390         }
391
392         /* Were we instructed to switch PG? */
393         if (READ_ONCE(m->next_pg)) {
394                 spin_lock_irqsave(&m->lock, flags);
395                 pg = m->next_pg;
396                 if (!pg) {
397                         spin_unlock_irqrestore(&m->lock, flags);
398                         goto check_current_pg;
399                 }
400                 m->next_pg = NULL;
401                 spin_unlock_irqrestore(&m->lock, flags);
402                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
403                 if (!IS_ERR_OR_NULL(pgpath))
404                         return pgpath;
405         }
406
407         /* Don't change PG until it has no remaining paths */
408 check_current_pg:
409         pg = READ_ONCE(m->current_pg);
410         if (pg) {
411                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
412                 if (!IS_ERR_OR_NULL(pgpath))
413                         return pgpath;
414         }
415
416         /*
417          * Loop through priority groups until we find a valid path.
418          * First time we skip PGs marked 'bypassed'.
419          * Second time we only try the ones we skipped, but set
420          * pg_init_delay_retry so we do not hammer controllers.
421          */
422         do {
423                 list_for_each_entry(pg, &m->priority_groups, list) {
424                         if (pg->bypassed == !!bypassed)
425                                 continue;
426                         pgpath = choose_path_in_pg(m, pg, nr_bytes);
427                         if (!IS_ERR_OR_NULL(pgpath)) {
428                                 if (!bypassed)
429                                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
430                                 return pgpath;
431                         }
432                 }
433         } while (bypassed--);
434
435 failed:
436         spin_lock_irqsave(&m->lock, flags);
437         m->current_pgpath = NULL;
438         m->current_pg = NULL;
439         spin_unlock_irqrestore(&m->lock, flags);
440
441         return NULL;
442 }
443
444 /*
445  * dm_report_EIO() is a macro instead of a function to make pr_debug()
446  * report the function name and line number of the function from which
447  * it has been invoked.
448  */
449 #define dm_report_EIO(m)                                                \
450 do {                                                                    \
451         struct mapped_device *md = dm_table_get_md((m)->ti->table);     \
452                                                                         \
453         pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
454                  dm_device_name(md),                                    \
455                  test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags),        \
456                  test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags),  \
457                  dm_noflush_suspending((m)->ti));                       \
458 } while (0)
459
460 /*
461  * Map cloned requests (request-based multipath)
462  */
463 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
464                                    union map_info *map_context,
465                                    struct request **__clone)
466 {
467         struct multipath *m = ti->private;
468         size_t nr_bytes = blk_rq_bytes(rq);
469         struct pgpath *pgpath;
470         struct block_device *bdev;
471         struct dm_mpath_io *mpio = get_mpio(map_context);
472         struct request_queue *q;
473         struct request *clone;
474
475         /* Do we need to select a new pgpath? */
476         pgpath = READ_ONCE(m->current_pgpath);
477         if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
478                 pgpath = choose_pgpath(m, nr_bytes);
479
480         if (!pgpath) {
481                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
482                         return DM_MAPIO_DELAY_REQUEUE;
483                 dm_report_EIO(m);       /* Failed */
484                 return DM_MAPIO_KILL;
485         } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
486                    test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
487                 if (pg_init_all_paths(m))
488                         return DM_MAPIO_DELAY_REQUEUE;
489                 return DM_MAPIO_REQUEUE;
490         }
491
492         memset(mpio, 0, sizeof(*mpio));
493         mpio->pgpath = pgpath;
494         mpio->nr_bytes = nr_bytes;
495
496         bdev = pgpath->path.dev->bdev;
497         q = bdev_get_queue(bdev);
498         clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
499         if (IS_ERR(clone)) {
500                 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
501                 bool queue_dying = blk_queue_dying(q);
502                 DMERR_LIMIT("blk_get_request() returned %ld%s - requeuing",
503                             PTR_ERR(clone), queue_dying ? " (path offline)" : "");
504                 if (queue_dying) {
505                         atomic_inc(&m->pg_init_in_progress);
506                         activate_or_offline_path(pgpath);
507                 }
508                 return DM_MAPIO_DELAY_REQUEUE;
509         }
510         clone->bio = clone->biotail = NULL;
511         clone->rq_disk = bdev->bd_disk;
512         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
513         *__clone = clone;
514
515         if (pgpath->pg->ps.type->start_io)
516                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
517                                               &pgpath->path,
518                                               nr_bytes);
519         return DM_MAPIO_REMAPPED;
520 }
521
522 static void multipath_release_clone(struct request *clone)
523 {
524         blk_put_request(clone);
525 }
526
527 /*
528  * Map cloned bios (bio-based multipath)
529  */
530 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
531 {
532         size_t nr_bytes = bio->bi_iter.bi_size;
533         struct pgpath *pgpath;
534         unsigned long flags;
535         bool queue_io;
536
537         /* Do we need to select a new pgpath? */
538         pgpath = READ_ONCE(m->current_pgpath);
539         queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
540         if (!pgpath || !queue_io)
541                 pgpath = choose_pgpath(m, nr_bytes);
542
543         if ((pgpath && queue_io) ||
544             (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
545                 /* Queue for the daemon to resubmit */
546                 spin_lock_irqsave(&m->lock, flags);
547                 bio_list_add(&m->queued_bios, bio);
548                 spin_unlock_irqrestore(&m->lock, flags);
549                 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
550                 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
551                         pg_init_all_paths(m);
552                 else if (!queue_io)
553                         queue_work(kmultipathd, &m->process_queued_bios);
554                 return DM_MAPIO_SUBMITTED;
555         }
556
557         if (!pgpath) {
558                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
559                         return DM_MAPIO_REQUEUE;
560                 dm_report_EIO(m);
561                 return DM_MAPIO_KILL;
562         }
563
564         mpio->pgpath = pgpath;
565         mpio->nr_bytes = nr_bytes;
566
567         bio->bi_status = 0;
568         bio_set_dev(bio, pgpath->path.dev->bdev);
569         bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
570
571         if (pgpath->pg->ps.type->start_io)
572                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
573                                               &pgpath->path,
574                                               nr_bytes);
575         return DM_MAPIO_REMAPPED;
576 }
577
578 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
579 {
580         struct multipath *m = ti->private;
581         struct dm_mpath_io *mpio = NULL;
582
583         multipath_init_per_bio_data(bio, &mpio, NULL);
584
585         return __multipath_map_bio(m, bio, mpio);
586 }
587
588 static void process_queued_io_list(struct multipath *m)
589 {
590         if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
591                 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
592         else if (m->queue_mode == DM_TYPE_BIO_BASED)
593                 queue_work(kmultipathd, &m->process_queued_bios);
594 }
595
596 static void process_queued_bios(struct work_struct *work)
597 {
598         int r;
599         unsigned long flags;
600         struct bio *bio;
601         struct bio_list bios;
602         struct blk_plug plug;
603         struct multipath *m =
604                 container_of(work, struct multipath, process_queued_bios);
605
606         bio_list_init(&bios);
607
608         spin_lock_irqsave(&m->lock, flags);
609
610         if (bio_list_empty(&m->queued_bios)) {
611                 spin_unlock_irqrestore(&m->lock, flags);
612                 return;
613         }
614
615         bio_list_merge(&bios, &m->queued_bios);
616         bio_list_init(&m->queued_bios);
617
618         spin_unlock_irqrestore(&m->lock, flags);
619
620         blk_start_plug(&plug);
621         while ((bio = bio_list_pop(&bios))) {
622                 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
623                 switch (r) {
624                 case DM_MAPIO_KILL:
625                         bio->bi_status = BLK_STS_IOERR;
626                         bio_endio(bio);
627                         break;
628                 case DM_MAPIO_REQUEUE:
629                         bio->bi_status = BLK_STS_DM_REQUEUE;
630                         bio_endio(bio);
631                         break;
632                 case DM_MAPIO_REMAPPED:
633                         generic_make_request(bio);
634                         break;
635                 case 0:
636                         break;
637                 default:
638                         WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
639                 }
640         }
641         blk_finish_plug(&plug);
642 }
643
644 /*
645  * If we run out of usable paths, should we queue I/O or error it?
646  */
647 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
648                             bool save_old_value)
649 {
650         unsigned long flags;
651
652         spin_lock_irqsave(&m->lock, flags);
653         assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
654                    (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
655                    (!save_old_value && queue_if_no_path));
656         assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
657                    queue_if_no_path || dm_noflush_suspending(m->ti));
658         spin_unlock_irqrestore(&m->lock, flags);
659
660         if (!queue_if_no_path) {
661                 dm_table_run_md_queue_async(m->ti->table);
662                 process_queued_io_list(m);
663         }
664
665         return 0;
666 }
667
668 /*
669  * An event is triggered whenever a path is taken out of use.
670  * Includes path failure and PG bypass.
671  */
672 static void trigger_event(struct work_struct *work)
673 {
674         struct multipath *m =
675                 container_of(work, struct multipath, trigger_event);
676
677         dm_table_event(m->ti->table);
678 }
679
680 /*-----------------------------------------------------------------
681  * Constructor/argument parsing:
682  * <#multipath feature args> [<arg>]*
683  * <#hw_handler args> [hw_handler [<arg>]*]
684  * <#priority groups>
685  * <initial priority group>
686  *     [<selector> <#selector args> [<arg>]*
687  *      <#paths> <#per-path selector args>
688  *         [<path> [<arg>]* ]+ ]+
689  *---------------------------------------------------------------*/
690 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
691                                struct dm_target *ti)
692 {
693         int r;
694         struct path_selector_type *pst;
695         unsigned ps_argc;
696
697         static const struct dm_arg _args[] = {
698                 {0, 1024, "invalid number of path selector args"},
699         };
700
701         pst = dm_get_path_selector(dm_shift_arg(as));
702         if (!pst) {
703                 ti->error = "unknown path selector type";
704                 return -EINVAL;
705         }
706
707         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
708         if (r) {
709                 dm_put_path_selector(pst);
710                 return -EINVAL;
711         }
712
713         r = pst->create(&pg->ps, ps_argc, as->argv);
714         if (r) {
715                 dm_put_path_selector(pst);
716                 ti->error = "path selector constructor failed";
717                 return r;
718         }
719
720         pg->ps.type = pst;
721         dm_consume_args(as, ps_argc);
722
723         return 0;
724 }
725
726 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
727                                struct dm_target *ti)
728 {
729         int r;
730         struct pgpath *p;
731         struct multipath *m = ti->private;
732         struct request_queue *q = NULL;
733         const char *attached_handler_name;
734
735         /* we need at least a path arg */
736         if (as->argc < 1) {
737                 ti->error = "no device given";
738                 return ERR_PTR(-EINVAL);
739         }
740
741         p = alloc_pgpath();
742         if (!p)
743                 return ERR_PTR(-ENOMEM);
744
745         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
746                           &p->path.dev);
747         if (r) {
748                 ti->error = "error getting device";
749                 goto bad;
750         }
751
752         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
753                 q = bdev_get_queue(p->path.dev->bdev);
754
755         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
756 retain:
757                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
758                 if (attached_handler_name) {
759                         /*
760                          * Clear any hw_handler_params associated with a
761                          * handler that isn't already attached.
762                          */
763                         if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
764                                 kfree(m->hw_handler_params);
765                                 m->hw_handler_params = NULL;
766                         }
767
768                         /*
769                          * Reset hw_handler_name to match the attached handler
770                          *
771                          * NB. This modifies the table line to show the actual
772                          * handler instead of the original table passed in.
773                          */
774                         kfree(m->hw_handler_name);
775                         m->hw_handler_name = attached_handler_name;
776                 }
777         }
778
779         if (m->hw_handler_name) {
780                 r = scsi_dh_attach(q, m->hw_handler_name);
781                 if (r == -EBUSY) {
782                         char b[BDEVNAME_SIZE];
783
784                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
785                                 bdevname(p->path.dev->bdev, b));
786                         goto retain;
787                 }
788                 if (r < 0) {
789                         ti->error = "error attaching hardware handler";
790                         dm_put_device(ti, p->path.dev);
791                         goto bad;
792                 }
793
794                 if (m->hw_handler_params) {
795                         r = scsi_dh_set_params(q, m->hw_handler_params);
796                         if (r < 0) {
797                                 ti->error = "unable to set hardware "
798                                                         "handler parameters";
799                                 dm_put_device(ti, p->path.dev);
800                                 goto bad;
801                         }
802                 }
803         }
804
805         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
806         if (r) {
807                 dm_put_device(ti, p->path.dev);
808                 goto bad;
809         }
810
811         return p;
812
813  bad:
814         free_pgpath(p);
815         return ERR_PTR(r);
816 }
817
818 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
819                                                    struct multipath *m)
820 {
821         static const struct dm_arg _args[] = {
822                 {1, 1024, "invalid number of paths"},
823                 {0, 1024, "invalid number of selector args"}
824         };
825
826         int r;
827         unsigned i, nr_selector_args, nr_args;
828         struct priority_group *pg;
829         struct dm_target *ti = m->ti;
830
831         if (as->argc < 2) {
832                 as->argc = 0;
833                 ti->error = "not enough priority group arguments";
834                 return ERR_PTR(-EINVAL);
835         }
836
837         pg = alloc_priority_group();
838         if (!pg) {
839                 ti->error = "couldn't allocate priority group";
840                 return ERR_PTR(-ENOMEM);
841         }
842         pg->m = m;
843
844         r = parse_path_selector(as, pg, ti);
845         if (r)
846                 goto bad;
847
848         /*
849          * read the paths
850          */
851         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
852         if (r)
853                 goto bad;
854
855         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
856         if (r)
857                 goto bad;
858
859         nr_args = 1 + nr_selector_args;
860         for (i = 0; i < pg->nr_pgpaths; i++) {
861                 struct pgpath *pgpath;
862                 struct dm_arg_set path_args;
863
864                 if (as->argc < nr_args) {
865                         ti->error = "not enough path parameters";
866                         r = -EINVAL;
867                         goto bad;
868                 }
869
870                 path_args.argc = nr_args;
871                 path_args.argv = as->argv;
872
873                 pgpath = parse_path(&path_args, &pg->ps, ti);
874                 if (IS_ERR(pgpath)) {
875                         r = PTR_ERR(pgpath);
876                         goto bad;
877                 }
878
879                 pgpath->pg = pg;
880                 list_add_tail(&pgpath->list, &pg->pgpaths);
881                 dm_consume_args(as, nr_args);
882         }
883
884         return pg;
885
886  bad:
887         free_priority_group(pg, ti);
888         return ERR_PTR(r);
889 }
890
891 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
892 {
893         unsigned hw_argc;
894         int ret;
895         struct dm_target *ti = m->ti;
896
897         static const struct dm_arg _args[] = {
898                 {0, 1024, "invalid number of hardware handler args"},
899         };
900
901         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
902                 return -EINVAL;
903
904         if (!hw_argc)
905                 return 0;
906
907         if (m->queue_mode == DM_TYPE_BIO_BASED) {
908                 dm_consume_args(as, hw_argc);
909                 DMERR("bio-based multipath doesn't allow hardware handler args");
910                 return 0;
911         }
912
913         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
914         if (!m->hw_handler_name)
915                 return -EINVAL;
916
917         if (hw_argc > 1) {
918                 char *p;
919                 int i, j, len = 4;
920
921                 for (i = 0; i <= hw_argc - 2; i++)
922                         len += strlen(as->argv[i]) + 1;
923                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
924                 if (!p) {
925                         ti->error = "memory allocation failed";
926                         ret = -ENOMEM;
927                         goto fail;
928                 }
929                 j = sprintf(p, "%d", hw_argc - 1);
930                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
931                         j = sprintf(p, "%s", as->argv[i]);
932         }
933         dm_consume_args(as, hw_argc - 1);
934
935         return 0;
936 fail:
937         kfree(m->hw_handler_name);
938         m->hw_handler_name = NULL;
939         return ret;
940 }
941
942 static int parse_features(struct dm_arg_set *as, struct multipath *m)
943 {
944         int r;
945         unsigned argc;
946         struct dm_target *ti = m->ti;
947         const char *arg_name;
948
949         static const struct dm_arg _args[] = {
950                 {0, 8, "invalid number of feature args"},
951                 {1, 50, "pg_init_retries must be between 1 and 50"},
952                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
953         };
954
955         r = dm_read_arg_group(_args, as, &argc, &ti->error);
956         if (r)
957                 return -EINVAL;
958
959         if (!argc)
960                 return 0;
961
962         do {
963                 arg_name = dm_shift_arg(as);
964                 argc--;
965
966                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
967                         r = queue_if_no_path(m, true, false);
968                         continue;
969                 }
970
971                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
972                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
973                         continue;
974                 }
975
976                 if (!strcasecmp(arg_name, "pg_init_retries") &&
977                     (argc >= 1)) {
978                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
979                         argc--;
980                         continue;
981                 }
982
983                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
984                     (argc >= 1)) {
985                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
986                         argc--;
987                         continue;
988                 }
989
990                 if (!strcasecmp(arg_name, "queue_mode") &&
991                     (argc >= 1)) {
992                         const char *queue_mode_name = dm_shift_arg(as);
993
994                         if (!strcasecmp(queue_mode_name, "bio"))
995                                 m->queue_mode = DM_TYPE_BIO_BASED;
996                         else if (!strcasecmp(queue_mode_name, "rq"))
997                                 m->queue_mode = DM_TYPE_REQUEST_BASED;
998                         else if (!strcasecmp(queue_mode_name, "mq"))
999                                 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1000                         else {
1001                                 ti->error = "Unknown 'queue_mode' requested";
1002                                 r = -EINVAL;
1003                         }
1004                         argc--;
1005                         continue;
1006                 }
1007
1008                 ti->error = "Unrecognised multipath feature request";
1009                 r = -EINVAL;
1010         } while (argc && !r);
1011
1012         return r;
1013 }
1014
1015 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1016 {
1017         /* target arguments */
1018         static const struct dm_arg _args[] = {
1019                 {0, 1024, "invalid number of priority groups"},
1020                 {0, 1024, "invalid initial priority group number"},
1021         };
1022
1023         int r;
1024         struct multipath *m;
1025         struct dm_arg_set as;
1026         unsigned pg_count = 0;
1027         unsigned next_pg_num;
1028
1029         as.argc = argc;
1030         as.argv = argv;
1031
1032         m = alloc_multipath(ti);
1033         if (!m) {
1034                 ti->error = "can't allocate multipath";
1035                 return -EINVAL;
1036         }
1037
1038         r = parse_features(&as, m);
1039         if (r)
1040                 goto bad;
1041
1042         r = alloc_multipath_stage2(ti, m);
1043         if (r)
1044                 goto bad;
1045
1046         r = parse_hw_handler(&as, m);
1047         if (r)
1048                 goto bad;
1049
1050         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1051         if (r)
1052                 goto bad;
1053
1054         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1055         if (r)
1056                 goto bad;
1057
1058         if ((!m->nr_priority_groups && next_pg_num) ||
1059             (m->nr_priority_groups && !next_pg_num)) {
1060                 ti->error = "invalid initial priority group";
1061                 r = -EINVAL;
1062                 goto bad;
1063         }
1064
1065         /* parse the priority groups */
1066         while (as.argc) {
1067                 struct priority_group *pg;
1068                 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1069
1070                 pg = parse_priority_group(&as, m);
1071                 if (IS_ERR(pg)) {
1072                         r = PTR_ERR(pg);
1073                         goto bad;
1074                 }
1075
1076                 nr_valid_paths += pg->nr_pgpaths;
1077                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1078
1079                 list_add_tail(&pg->list, &m->priority_groups);
1080                 pg_count++;
1081                 pg->pg_num = pg_count;
1082                 if (!--next_pg_num)
1083                         m->next_pg = pg;
1084         }
1085
1086         if (pg_count != m->nr_priority_groups) {
1087                 ti->error = "priority group count mismatch";
1088                 r = -EINVAL;
1089                 goto bad;
1090         }
1091
1092         ti->num_flush_bios = 1;
1093         ti->num_discard_bios = 1;
1094         ti->num_write_same_bios = 1;
1095         ti->num_write_zeroes_bios = 1;
1096         if (m->queue_mode == DM_TYPE_BIO_BASED)
1097                 ti->per_io_data_size = multipath_per_bio_data_size();
1098         else
1099                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1100
1101         return 0;
1102
1103  bad:
1104         free_multipath(m);
1105         return r;
1106 }
1107
1108 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1109 {
1110         DEFINE_WAIT(wait);
1111
1112         while (1) {
1113                 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1114
1115                 if (!atomic_read(&m->pg_init_in_progress))
1116                         break;
1117
1118                 io_schedule();
1119         }
1120         finish_wait(&m->pg_init_wait, &wait);
1121 }
1122
1123 static void flush_multipath_work(struct multipath *m)
1124 {
1125         set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1126         smp_mb__after_atomic();
1127
1128         flush_workqueue(kmpath_handlerd);
1129         multipath_wait_for_pg_init_completion(m);
1130         flush_workqueue(kmultipathd);
1131         flush_work(&m->trigger_event);
1132
1133         clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1134         smp_mb__after_atomic();
1135 }
1136
1137 static void multipath_dtr(struct dm_target *ti)
1138 {
1139         struct multipath *m = ti->private;
1140
1141         flush_multipath_work(m);
1142         free_multipath(m);
1143 }
1144
1145 /*
1146  * Take a path out of use.
1147  */
1148 static int fail_path(struct pgpath *pgpath)
1149 {
1150         unsigned long flags;
1151         struct multipath *m = pgpath->pg->m;
1152
1153         spin_lock_irqsave(&m->lock, flags);
1154
1155         if (!pgpath->is_active)
1156                 goto out;
1157
1158         DMWARN("Failing path %s.", pgpath->path.dev->name);
1159
1160         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1161         pgpath->is_active = false;
1162         pgpath->fail_count++;
1163
1164         atomic_dec(&m->nr_valid_paths);
1165
1166         if (pgpath == m->current_pgpath)
1167                 m->current_pgpath = NULL;
1168
1169         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1170                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1171
1172         schedule_work(&m->trigger_event);
1173
1174 out:
1175         spin_unlock_irqrestore(&m->lock, flags);
1176
1177         return 0;
1178 }
1179
1180 /*
1181  * Reinstate a previously-failed path
1182  */
1183 static int reinstate_path(struct pgpath *pgpath)
1184 {
1185         int r = 0, run_queue = 0;
1186         unsigned long flags;
1187         struct multipath *m = pgpath->pg->m;
1188         unsigned nr_valid_paths;
1189
1190         spin_lock_irqsave(&m->lock, flags);
1191
1192         if (pgpath->is_active)
1193                 goto out;
1194
1195         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1196
1197         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1198         if (r)
1199                 goto out;
1200
1201         pgpath->is_active = true;
1202
1203         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1204         if (nr_valid_paths == 1) {
1205                 m->current_pgpath = NULL;
1206                 run_queue = 1;
1207         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1208                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1209                         atomic_inc(&m->pg_init_in_progress);
1210         }
1211
1212         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1213                        pgpath->path.dev->name, nr_valid_paths);
1214
1215         schedule_work(&m->trigger_event);
1216
1217 out:
1218         spin_unlock_irqrestore(&m->lock, flags);
1219         if (run_queue) {
1220                 dm_table_run_md_queue_async(m->ti->table);
1221                 process_queued_io_list(m);
1222         }
1223
1224         return r;
1225 }
1226
1227 /*
1228  * Fail or reinstate all paths that match the provided struct dm_dev.
1229  */
1230 static int action_dev(struct multipath *m, struct dm_dev *dev,
1231                       action_fn action)
1232 {
1233         int r = -EINVAL;
1234         struct pgpath *pgpath;
1235         struct priority_group *pg;
1236
1237         list_for_each_entry(pg, &m->priority_groups, list) {
1238                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1239                         if (pgpath->path.dev == dev)
1240                                 r = action(pgpath);
1241                 }
1242         }
1243
1244         return r;
1245 }
1246
1247 /*
1248  * Temporarily try to avoid having to use the specified PG
1249  */
1250 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1251                       bool bypassed)
1252 {
1253         unsigned long flags;
1254
1255         spin_lock_irqsave(&m->lock, flags);
1256
1257         pg->bypassed = bypassed;
1258         m->current_pgpath = NULL;
1259         m->current_pg = NULL;
1260
1261         spin_unlock_irqrestore(&m->lock, flags);
1262
1263         schedule_work(&m->trigger_event);
1264 }
1265
1266 /*
1267  * Switch to using the specified PG from the next I/O that gets mapped
1268  */
1269 static int switch_pg_num(struct multipath *m, const char *pgstr)
1270 {
1271         struct priority_group *pg;
1272         unsigned pgnum;
1273         unsigned long flags;
1274         char dummy;
1275
1276         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1277             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1278                 DMWARN("invalid PG number supplied to switch_pg_num");
1279                 return -EINVAL;
1280         }
1281
1282         spin_lock_irqsave(&m->lock, flags);
1283         list_for_each_entry(pg, &m->priority_groups, list) {
1284                 pg->bypassed = false;
1285                 if (--pgnum)
1286                         continue;
1287
1288                 m->current_pgpath = NULL;
1289                 m->current_pg = NULL;
1290                 m->next_pg = pg;
1291         }
1292         spin_unlock_irqrestore(&m->lock, flags);
1293
1294         schedule_work(&m->trigger_event);
1295         return 0;
1296 }
1297
1298 /*
1299  * Set/clear bypassed status of a PG.
1300  * PGs are numbered upwards from 1 in the order they were declared.
1301  */
1302 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1303 {
1304         struct priority_group *pg;
1305         unsigned pgnum;
1306         char dummy;
1307
1308         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1309             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1310                 DMWARN("invalid PG number supplied to bypass_pg");
1311                 return -EINVAL;
1312         }
1313
1314         list_for_each_entry(pg, &m->priority_groups, list) {
1315                 if (!--pgnum)
1316                         break;
1317         }
1318
1319         bypass_pg(m, pg, bypassed);
1320         return 0;
1321 }
1322
1323 /*
1324  * Should we retry pg_init immediately?
1325  */
1326 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1327 {
1328         unsigned long flags;
1329         bool limit_reached = false;
1330
1331         spin_lock_irqsave(&m->lock, flags);
1332
1333         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1334             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1335                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1336         else
1337                 limit_reached = true;
1338
1339         spin_unlock_irqrestore(&m->lock, flags);
1340
1341         return limit_reached;
1342 }
1343
1344 static void pg_init_done(void *data, int errors)
1345 {
1346         struct pgpath *pgpath = data;
1347         struct priority_group *pg = pgpath->pg;
1348         struct multipath *m = pg->m;
1349         unsigned long flags;
1350         bool delay_retry = false;
1351
1352         /* device or driver problems */
1353         switch (errors) {
1354         case SCSI_DH_OK:
1355                 break;
1356         case SCSI_DH_NOSYS:
1357                 if (!m->hw_handler_name) {
1358                         errors = 0;
1359                         break;
1360                 }
1361                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1362                       "Error %d.", m->hw_handler_name, errors);
1363                 /*
1364                  * Fail path for now, so we do not ping pong
1365                  */
1366                 fail_path(pgpath);
1367                 break;
1368         case SCSI_DH_DEV_TEMP_BUSY:
1369                 /*
1370                  * Probably doing something like FW upgrade on the
1371                  * controller so try the other pg.
1372                  */
1373                 bypass_pg(m, pg, true);
1374                 break;
1375         case SCSI_DH_RETRY:
1376                 /* Wait before retrying. */
1377                 delay_retry = 1;
1378                 /* fall through */
1379         case SCSI_DH_IMM_RETRY:
1380         case SCSI_DH_RES_TEMP_UNAVAIL:
1381                 if (pg_init_limit_reached(m, pgpath))
1382                         fail_path(pgpath);
1383                 errors = 0;
1384                 break;
1385         case SCSI_DH_DEV_OFFLINED:
1386         default:
1387                 /*
1388                  * We probably do not want to fail the path for a device
1389                  * error, but this is what the old dm did. In future
1390                  * patches we can do more advanced handling.
1391                  */
1392                 fail_path(pgpath);
1393         }
1394
1395         spin_lock_irqsave(&m->lock, flags);
1396         if (errors) {
1397                 if (pgpath == m->current_pgpath) {
1398                         DMERR("Could not failover device. Error %d.", errors);
1399                         m->current_pgpath = NULL;
1400                         m->current_pg = NULL;
1401                 }
1402         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1403                 pg->bypassed = false;
1404
1405         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1406                 /* Activations of other paths are still on going */
1407                 goto out;
1408
1409         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1410                 if (delay_retry)
1411                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1412                 else
1413                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1414
1415                 if (__pg_init_all_paths(m))
1416                         goto out;
1417         }
1418         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1419
1420         process_queued_io_list(m);
1421
1422         /*
1423          * Wake up any thread waiting to suspend.
1424          */
1425         wake_up(&m->pg_init_wait);
1426
1427 out:
1428         spin_unlock_irqrestore(&m->lock, flags);
1429 }
1430
1431 static void activate_or_offline_path(struct pgpath *pgpath)
1432 {
1433         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1434
1435         if (pgpath->is_active && !blk_queue_dying(q))
1436                 scsi_dh_activate(q, pg_init_done, pgpath);
1437         else
1438                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1439 }
1440
1441 static void activate_path_work(struct work_struct *work)
1442 {
1443         struct pgpath *pgpath =
1444                 container_of(work, struct pgpath, activate_path.work);
1445
1446         activate_or_offline_path(pgpath);
1447 }
1448
1449 static int noretry_error(blk_status_t error)
1450 {
1451         switch (error) {
1452         case BLK_STS_NOTSUPP:
1453         case BLK_STS_NOSPC:
1454         case BLK_STS_TARGET:
1455         case BLK_STS_NEXUS:
1456         case BLK_STS_MEDIUM:
1457                 return 1;
1458         }
1459
1460         /* Anything else could be a path failure, so should be retried */
1461         return 0;
1462 }
1463
1464 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1465                             blk_status_t error, union map_info *map_context)
1466 {
1467         struct dm_mpath_io *mpio = get_mpio(map_context);
1468         struct pgpath *pgpath = mpio->pgpath;
1469         int r = DM_ENDIO_DONE;
1470
1471         /*
1472          * We don't queue any clone request inside the multipath target
1473          * during end I/O handling, since those clone requests don't have
1474          * bio clones.  If we queue them inside the multipath target,
1475          * we need to make bio clones, that requires memory allocation.
1476          * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1477          *  don't have bio clones.)
1478          * Instead of queueing the clone request here, we queue the original
1479          * request into dm core, which will remake a clone request and
1480          * clone bios for it and resubmit it later.
1481          */
1482         if (error && !noretry_error(error)) {
1483                 struct multipath *m = ti->private;
1484
1485                 r = DM_ENDIO_REQUEUE;
1486
1487                 if (pgpath)
1488                         fail_path(pgpath);
1489
1490                 if (atomic_read(&m->nr_valid_paths) == 0 &&
1491                     !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1492                         if (error == BLK_STS_IOERR)
1493                                 dm_report_EIO(m);
1494                         /* complete with the original error */
1495                         r = DM_ENDIO_DONE;
1496                 }
1497         }
1498
1499         if (pgpath) {
1500                 struct path_selector *ps = &pgpath->pg->ps;
1501
1502                 if (ps->type->end_io)
1503                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1504         }
1505
1506         return r;
1507 }
1508
1509 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1510                 blk_status_t *error)
1511 {
1512         struct multipath *m = ti->private;
1513         struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1514         struct pgpath *pgpath = mpio->pgpath;
1515         unsigned long flags;
1516         int r = DM_ENDIO_DONE;
1517
1518         if (!*error || noretry_error(*error))
1519                 goto done;
1520
1521         if (pgpath)
1522                 fail_path(pgpath);
1523
1524         if (atomic_read(&m->nr_valid_paths) == 0 &&
1525             !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1526                 dm_report_EIO(m);
1527                 *error = BLK_STS_IOERR;
1528                 goto done;
1529         }
1530
1531         /* Queue for the daemon to resubmit */
1532         dm_bio_restore(get_bio_details_from_bio(clone), clone);
1533
1534         spin_lock_irqsave(&m->lock, flags);
1535         bio_list_add(&m->queued_bios, clone);
1536         spin_unlock_irqrestore(&m->lock, flags);
1537         if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1538                 queue_work(kmultipathd, &m->process_queued_bios);
1539
1540         r = DM_ENDIO_INCOMPLETE;
1541 done:
1542         if (pgpath) {
1543                 struct path_selector *ps = &pgpath->pg->ps;
1544
1545                 if (ps->type->end_io)
1546                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1547         }
1548
1549         return r;
1550 }
1551
1552 /*
1553  * Suspend can't complete until all the I/O is processed so if
1554  * the last path fails we must error any remaining I/O.
1555  * Note that if the freeze_bdev fails while suspending, the
1556  * queue_if_no_path state is lost - userspace should reset it.
1557  */
1558 static void multipath_presuspend(struct dm_target *ti)
1559 {
1560         struct multipath *m = ti->private;
1561
1562         queue_if_no_path(m, false, true);
1563 }
1564
1565 static void multipath_postsuspend(struct dm_target *ti)
1566 {
1567         struct multipath *m = ti->private;
1568
1569         mutex_lock(&m->work_mutex);
1570         flush_multipath_work(m);
1571         mutex_unlock(&m->work_mutex);
1572 }
1573
1574 /*
1575  * Restore the queue_if_no_path setting.
1576  */
1577 static void multipath_resume(struct dm_target *ti)
1578 {
1579         struct multipath *m = ti->private;
1580         unsigned long flags;
1581
1582         spin_lock_irqsave(&m->lock, flags);
1583         assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1584                    test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1585         spin_unlock_irqrestore(&m->lock, flags);
1586 }
1587
1588 /*
1589  * Info output has the following format:
1590  * num_multipath_feature_args [multipath_feature_args]*
1591  * num_handler_status_args [handler_status_args]*
1592  * num_groups init_group_number
1593  *            [A|D|E num_ps_status_args [ps_status_args]*
1594  *             num_paths num_selector_args
1595  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1596  *
1597  * Table output has the following format (identical to the constructor string):
1598  * num_feature_args [features_args]*
1599  * num_handler_args hw_handler [hw_handler_args]*
1600  * num_groups init_group_number
1601  *     [priority selector-name num_ps_args [ps_args]*
1602  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1603  */
1604 static void multipath_status(struct dm_target *ti, status_type_t type,
1605                              unsigned status_flags, char *result, unsigned maxlen)
1606 {
1607         int sz = 0;
1608         unsigned long flags;
1609         struct multipath *m = ti->private;
1610         struct priority_group *pg;
1611         struct pgpath *p;
1612         unsigned pg_num;
1613         char state;
1614
1615         spin_lock_irqsave(&m->lock, flags);
1616
1617         /* Features */
1618         if (type == STATUSTYPE_INFO)
1619                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1620                        atomic_read(&m->pg_init_count));
1621         else {
1622                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1623                               (m->pg_init_retries > 0) * 2 +
1624                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1625                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1626                               (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1627
1628                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1629                         DMEMIT("queue_if_no_path ");
1630                 if (m->pg_init_retries)
1631                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1632                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1633                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1634                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1635                         DMEMIT("retain_attached_hw_handler ");
1636                 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1637                         switch(m->queue_mode) {
1638                         case DM_TYPE_BIO_BASED:
1639                                 DMEMIT("queue_mode bio ");
1640                                 break;
1641                         case DM_TYPE_MQ_REQUEST_BASED:
1642                                 DMEMIT("queue_mode mq ");
1643                                 break;
1644                         default:
1645                                 WARN_ON_ONCE(true);
1646                                 break;
1647                         }
1648                 }
1649         }
1650
1651         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1652                 DMEMIT("0 ");
1653         else
1654                 DMEMIT("1 %s ", m->hw_handler_name);
1655
1656         DMEMIT("%u ", m->nr_priority_groups);
1657
1658         if (m->next_pg)
1659                 pg_num = m->next_pg->pg_num;
1660         else if (m->current_pg)
1661                 pg_num = m->current_pg->pg_num;
1662         else
1663                 pg_num = (m->nr_priority_groups ? 1 : 0);
1664
1665         DMEMIT("%u ", pg_num);
1666
1667         switch (type) {
1668         case STATUSTYPE_INFO:
1669                 list_for_each_entry(pg, &m->priority_groups, list) {
1670                         if (pg->bypassed)
1671                                 state = 'D';    /* Disabled */
1672                         else if (pg == m->current_pg)
1673                                 state = 'A';    /* Currently Active */
1674                         else
1675                                 state = 'E';    /* Enabled */
1676
1677                         DMEMIT("%c ", state);
1678
1679                         if (pg->ps.type->status)
1680                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1681                                                           result + sz,
1682                                                           maxlen - sz);
1683                         else
1684                                 DMEMIT("0 ");
1685
1686                         DMEMIT("%u %u ", pg->nr_pgpaths,
1687                                pg->ps.type->info_args);
1688
1689                         list_for_each_entry(p, &pg->pgpaths, list) {
1690                                 DMEMIT("%s %s %u ", p->path.dev->name,
1691                                        p->is_active ? "A" : "F",
1692                                        p->fail_count);
1693                                 if (pg->ps.type->status)
1694                                         sz += pg->ps.type->status(&pg->ps,
1695                                               &p->path, type, result + sz,
1696                                               maxlen - sz);
1697                         }
1698                 }
1699                 break;
1700
1701         case STATUSTYPE_TABLE:
1702                 list_for_each_entry(pg, &m->priority_groups, list) {
1703                         DMEMIT("%s ", pg->ps.type->name);
1704
1705                         if (pg->ps.type->status)
1706                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1707                                                           result + sz,
1708                                                           maxlen - sz);
1709                         else
1710                                 DMEMIT("0 ");
1711
1712                         DMEMIT("%u %u ", pg->nr_pgpaths,
1713                                pg->ps.type->table_args);
1714
1715                         list_for_each_entry(p, &pg->pgpaths, list) {
1716                                 DMEMIT("%s ", p->path.dev->name);
1717                                 if (pg->ps.type->status)
1718                                         sz += pg->ps.type->status(&pg->ps,
1719                                               &p->path, type, result + sz,
1720                                               maxlen - sz);
1721                         }
1722                 }
1723                 break;
1724         }
1725
1726         spin_unlock_irqrestore(&m->lock, flags);
1727 }
1728
1729 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1730 {
1731         int r = -EINVAL;
1732         struct dm_dev *dev;
1733         struct multipath *m = ti->private;
1734         action_fn action;
1735
1736         mutex_lock(&m->work_mutex);
1737
1738         if (dm_suspended(ti)) {
1739                 r = -EBUSY;
1740                 goto out;
1741         }
1742
1743         if (argc == 1) {
1744                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1745                         r = queue_if_no_path(m, true, false);
1746                         goto out;
1747                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1748                         r = queue_if_no_path(m, false, false);
1749                         goto out;
1750                 }
1751         }
1752
1753         if (argc != 2) {
1754                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1755                 goto out;
1756         }
1757
1758         if (!strcasecmp(argv[0], "disable_group")) {
1759                 r = bypass_pg_num(m, argv[1], true);
1760                 goto out;
1761         } else if (!strcasecmp(argv[0], "enable_group")) {
1762                 r = bypass_pg_num(m, argv[1], false);
1763                 goto out;
1764         } else if (!strcasecmp(argv[0], "switch_group")) {
1765                 r = switch_pg_num(m, argv[1]);
1766                 goto out;
1767         } else if (!strcasecmp(argv[0], "reinstate_path"))
1768                 action = reinstate_path;
1769         else if (!strcasecmp(argv[0], "fail_path"))
1770                 action = fail_path;
1771         else {
1772                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1773                 goto out;
1774         }
1775
1776         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1777         if (r) {
1778                 DMWARN("message: error getting device %s",
1779                        argv[1]);
1780                 goto out;
1781         }
1782
1783         r = action_dev(m, dev, action);
1784
1785         dm_put_device(ti, dev);
1786
1787 out:
1788         mutex_unlock(&m->work_mutex);
1789         return r;
1790 }
1791
1792 static int multipath_prepare_ioctl(struct dm_target *ti,
1793                 struct block_device **bdev, fmode_t *mode)
1794 {
1795         struct multipath *m = ti->private;
1796         struct pgpath *current_pgpath;
1797         int r;
1798
1799         current_pgpath = READ_ONCE(m->current_pgpath);
1800         if (!current_pgpath)
1801                 current_pgpath = choose_pgpath(m, 0);
1802
1803         if (current_pgpath) {
1804                 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1805                         *bdev = current_pgpath->path.dev->bdev;
1806                         *mode = current_pgpath->path.dev->mode;
1807                         r = 0;
1808                 } else {
1809                         /* pg_init has not started or completed */
1810                         r = -ENOTCONN;
1811                 }
1812         } else {
1813                 /* No path is available */
1814                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1815                         r = -ENOTCONN;
1816                 else
1817                         r = -EIO;
1818         }
1819
1820         if (r == -ENOTCONN) {
1821                 if (!READ_ONCE(m->current_pg)) {
1822                         /* Path status changed, redo selection */
1823                         (void) choose_pgpath(m, 0);
1824                 }
1825                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1826                         pg_init_all_paths(m);
1827                 dm_table_run_md_queue_async(m->ti->table);
1828                 process_queued_io_list(m);
1829         }
1830
1831         /*
1832          * Only pass ioctls through if the device sizes match exactly.
1833          */
1834         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1835                 return 1;
1836         return r;
1837 }
1838
1839 static int multipath_iterate_devices(struct dm_target *ti,
1840                                      iterate_devices_callout_fn fn, void *data)
1841 {
1842         struct multipath *m = ti->private;
1843         struct priority_group *pg;
1844         struct pgpath *p;
1845         int ret = 0;
1846
1847         list_for_each_entry(pg, &m->priority_groups, list) {
1848                 list_for_each_entry(p, &pg->pgpaths, list) {
1849                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1850                         if (ret)
1851                                 goto out;
1852                 }
1853         }
1854
1855 out:
1856         return ret;
1857 }
1858
1859 static int pgpath_busy(struct pgpath *pgpath)
1860 {
1861         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1862
1863         return blk_lld_busy(q);
1864 }
1865
1866 /*
1867  * We return "busy", only when we can map I/Os but underlying devices
1868  * are busy (so even if we map I/Os now, the I/Os will wait on
1869  * the underlying queue).
1870  * In other words, if we want to kill I/Os or queue them inside us
1871  * due to map unavailability, we don't return "busy".  Otherwise,
1872  * dm core won't give us the I/Os and we can't do what we want.
1873  */
1874 static int multipath_busy(struct dm_target *ti)
1875 {
1876         bool busy = false, has_active = false;
1877         struct multipath *m = ti->private;
1878         struct priority_group *pg, *next_pg;
1879         struct pgpath *pgpath;
1880
1881         /* pg_init in progress */
1882         if (atomic_read(&m->pg_init_in_progress))
1883                 return true;
1884
1885         /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1886         if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1887                 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1888
1889         /* Guess which priority_group will be used at next mapping time */
1890         pg = READ_ONCE(m->current_pg);
1891         next_pg = READ_ONCE(m->next_pg);
1892         if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1893                 pg = next_pg;
1894
1895         if (!pg) {
1896                 /*
1897                  * We don't know which pg will be used at next mapping time.
1898                  * We don't call choose_pgpath() here to avoid to trigger
1899                  * pg_init just by busy checking.
1900                  * So we don't know whether underlying devices we will be using
1901                  * at next mapping time are busy or not. Just try mapping.
1902                  */
1903                 return busy;
1904         }
1905
1906         /*
1907          * If there is one non-busy active path at least, the path selector
1908          * will be able to select it. So we consider such a pg as not busy.
1909          */
1910         busy = true;
1911         list_for_each_entry(pgpath, &pg->pgpaths, list) {
1912                 if (pgpath->is_active) {
1913                         has_active = true;
1914                         if (!pgpath_busy(pgpath)) {
1915                                 busy = false;
1916                                 break;
1917                         }
1918                 }
1919         }
1920
1921         if (!has_active) {
1922                 /*
1923                  * No active path in this pg, so this pg won't be used and
1924                  * the current_pg will be changed at next mapping time.
1925                  * We need to try mapping to determine it.
1926                  */
1927                 busy = false;
1928         }
1929
1930         return busy;
1931 }
1932
1933 /*-----------------------------------------------------------------
1934  * Module setup
1935  *---------------------------------------------------------------*/
1936 static struct target_type multipath_target = {
1937         .name = "multipath",
1938         .version = {1, 12, 0},
1939         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1940         .module = THIS_MODULE,
1941         .ctr = multipath_ctr,
1942         .dtr = multipath_dtr,
1943         .clone_and_map_rq = multipath_clone_and_map,
1944         .release_clone_rq = multipath_release_clone,
1945         .rq_end_io = multipath_end_io,
1946         .map = multipath_map_bio,
1947         .end_io = multipath_end_io_bio,
1948         .presuspend = multipath_presuspend,
1949         .postsuspend = multipath_postsuspend,
1950         .resume = multipath_resume,
1951         .status = multipath_status,
1952         .message = multipath_message,
1953         .prepare_ioctl = multipath_prepare_ioctl,
1954         .iterate_devices = multipath_iterate_devices,
1955         .busy = multipath_busy,
1956 };
1957
1958 static int __init dm_multipath_init(void)
1959 {
1960         int r;
1961
1962         r = dm_register_target(&multipath_target);
1963         if (r < 0) {
1964                 DMERR("request-based register failed %d", r);
1965                 r = -EINVAL;
1966                 goto bad_register_target;
1967         }
1968
1969         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1970         if (!kmultipathd) {
1971                 DMERR("failed to create workqueue kmpathd");
1972                 r = -ENOMEM;
1973                 goto bad_alloc_kmultipathd;
1974         }
1975
1976         /*
1977          * A separate workqueue is used to handle the device handlers
1978          * to avoid overloading existing workqueue. Overloading the
1979          * old workqueue would also create a bottleneck in the
1980          * path of the storage hardware device activation.
1981          */
1982         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1983                                                   WQ_MEM_RECLAIM);
1984         if (!kmpath_handlerd) {
1985                 DMERR("failed to create workqueue kmpath_handlerd");
1986                 r = -ENOMEM;
1987                 goto bad_alloc_kmpath_handlerd;
1988         }
1989
1990         return 0;
1991
1992 bad_alloc_kmpath_handlerd:
1993         destroy_workqueue(kmultipathd);
1994 bad_alloc_kmultipathd:
1995         dm_unregister_target(&multipath_target);
1996 bad_register_target:
1997         return r;
1998 }
1999
2000 static void __exit dm_multipath_exit(void)
2001 {
2002         destroy_workqueue(kmpath_handlerd);
2003         destroy_workqueue(kmultipathd);
2004
2005         dm_unregister_target(&multipath_target);
2006 }
2007
2008 module_init(dm_multipath_init);
2009 module_exit(dm_multipath_exit);
2010
2011 MODULE_DESCRIPTION(DM_NAME " multipath target");
2012 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2013 MODULE_LICENSE("GPL");