]> git.kernelconcepts.de Git - karo-tx-linux.git/blob - drivers/md/dm-mpath.c
block: introduce new block status code type
[karo-tx-linux.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(lockless_dereference(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 (lockless_dereference(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 = lockless_dereference(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 = lockless_dereference(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                         return DM_MAPIO_REQUEUE;
508                 }
509                 return DM_MAPIO_DELAY_REQUEUE;
510         }
511         clone->bio = clone->biotail = NULL;
512         clone->rq_disk = bdev->bd_disk;
513         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
514         *__clone = clone;
515
516         if (pgpath->pg->ps.type->start_io)
517                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
518                                               &pgpath->path,
519                                               nr_bytes);
520         return DM_MAPIO_REMAPPED;
521 }
522
523 static void multipath_release_clone(struct request *clone)
524 {
525         blk_put_request(clone);
526 }
527
528 /*
529  * Map cloned bios (bio-based multipath)
530  */
531 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
532 {
533         size_t nr_bytes = bio->bi_iter.bi_size;
534         struct pgpath *pgpath;
535         unsigned long flags;
536         bool queue_io;
537
538         /* Do we need to select a new pgpath? */
539         pgpath = lockless_dereference(m->current_pgpath);
540         queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
541         if (!pgpath || !queue_io)
542                 pgpath = choose_pgpath(m, nr_bytes);
543
544         if ((pgpath && queue_io) ||
545             (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
546                 /* Queue for the daemon to resubmit */
547                 spin_lock_irqsave(&m->lock, flags);
548                 bio_list_add(&m->queued_bios, bio);
549                 spin_unlock_irqrestore(&m->lock, flags);
550                 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
551                 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
552                         pg_init_all_paths(m);
553                 else if (!queue_io)
554                         queue_work(kmultipathd, &m->process_queued_bios);
555                 return DM_MAPIO_SUBMITTED;
556         }
557
558         if (!pgpath) {
559                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
560                         return DM_MAPIO_REQUEUE;
561                 dm_report_EIO(m);
562                 return DM_MAPIO_KILL;
563         }
564
565         mpio->pgpath = pgpath;
566         mpio->nr_bytes = nr_bytes;
567
568         bio->bi_error = 0;
569         bio->bi_bdev = pgpath->path.dev->bdev;
570         bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
571
572         if (pgpath->pg->ps.type->start_io)
573                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
574                                               &pgpath->path,
575                                               nr_bytes);
576         return DM_MAPIO_REMAPPED;
577 }
578
579 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
580 {
581         struct multipath *m = ti->private;
582         struct dm_mpath_io *mpio = NULL;
583
584         multipath_init_per_bio_data(bio, &mpio, NULL);
585
586         return __multipath_map_bio(m, bio, mpio);
587 }
588
589 static void process_queued_io_list(struct multipath *m)
590 {
591         if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
592                 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
593         else if (m->queue_mode == DM_TYPE_BIO_BASED)
594                 queue_work(kmultipathd, &m->process_queued_bios);
595 }
596
597 static void process_queued_bios(struct work_struct *work)
598 {
599         int r;
600         unsigned long flags;
601         struct bio *bio;
602         struct bio_list bios;
603         struct blk_plug plug;
604         struct multipath *m =
605                 container_of(work, struct multipath, process_queued_bios);
606
607         bio_list_init(&bios);
608
609         spin_lock_irqsave(&m->lock, flags);
610
611         if (bio_list_empty(&m->queued_bios)) {
612                 spin_unlock_irqrestore(&m->lock, flags);
613                 return;
614         }
615
616         bio_list_merge(&bios, &m->queued_bios);
617         bio_list_init(&m->queued_bios);
618
619         spin_unlock_irqrestore(&m->lock, flags);
620
621         blk_start_plug(&plug);
622         while ((bio = bio_list_pop(&bios))) {
623                 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
624                 switch (r) {
625                 case DM_MAPIO_KILL:
626                         r = -EIO;
627                         /*FALLTHRU*/
628                 case DM_MAPIO_REQUEUE:
629                         bio->bi_error = r;
630                         bio_endio(bio);
631                         break;
632                 case DM_MAPIO_REMAPPED:
633                         generic_make_request(bio);
634                         break;
635                 }
636         }
637         blk_finish_plug(&plug);
638 }
639
640 static void assign_bit(bool value, long nr, unsigned long *addr)
641 {
642         if (value)
643                 set_bit(nr, addr);
644         else
645                 clear_bit(nr, addr);
646 }
647
648 /*
649  * If we run out of usable paths, should we queue I/O or error it?
650  */
651 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
652                             bool save_old_value)
653 {
654         unsigned long flags;
655
656         spin_lock_irqsave(&m->lock, flags);
657         assign_bit((save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
658                    (!save_old_value && queue_if_no_path),
659                    MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
660         assign_bit(queue_if_no_path || dm_noflush_suspending(m->ti),
661                    MPATHF_QUEUE_IF_NO_PATH, &m->flags);
662         spin_unlock_irqrestore(&m->lock, flags);
663
664         if (!queue_if_no_path) {
665                 dm_table_run_md_queue_async(m->ti->table);
666                 process_queued_io_list(m);
667         }
668
669         return 0;
670 }
671
672 /*
673  * An event is triggered whenever a path is taken out of use.
674  * Includes path failure and PG bypass.
675  */
676 static void trigger_event(struct work_struct *work)
677 {
678         struct multipath *m =
679                 container_of(work, struct multipath, trigger_event);
680
681         dm_table_event(m->ti->table);
682 }
683
684 /*-----------------------------------------------------------------
685  * Constructor/argument parsing:
686  * <#multipath feature args> [<arg>]*
687  * <#hw_handler args> [hw_handler [<arg>]*]
688  * <#priority groups>
689  * <initial priority group>
690  *     [<selector> <#selector args> [<arg>]*
691  *      <#paths> <#per-path selector args>
692  *         [<path> [<arg>]* ]+ ]+
693  *---------------------------------------------------------------*/
694 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
695                                struct dm_target *ti)
696 {
697         int r;
698         struct path_selector_type *pst;
699         unsigned ps_argc;
700
701         static struct dm_arg _args[] = {
702                 {0, 1024, "invalid number of path selector args"},
703         };
704
705         pst = dm_get_path_selector(dm_shift_arg(as));
706         if (!pst) {
707                 ti->error = "unknown path selector type";
708                 return -EINVAL;
709         }
710
711         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
712         if (r) {
713                 dm_put_path_selector(pst);
714                 return -EINVAL;
715         }
716
717         r = pst->create(&pg->ps, ps_argc, as->argv);
718         if (r) {
719                 dm_put_path_selector(pst);
720                 ti->error = "path selector constructor failed";
721                 return r;
722         }
723
724         pg->ps.type = pst;
725         dm_consume_args(as, ps_argc);
726
727         return 0;
728 }
729
730 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
731                                struct dm_target *ti)
732 {
733         int r;
734         struct pgpath *p;
735         struct multipath *m = ti->private;
736         struct request_queue *q = NULL;
737         const char *attached_handler_name;
738
739         /* we need at least a path arg */
740         if (as->argc < 1) {
741                 ti->error = "no device given";
742                 return ERR_PTR(-EINVAL);
743         }
744
745         p = alloc_pgpath();
746         if (!p)
747                 return ERR_PTR(-ENOMEM);
748
749         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
750                           &p->path.dev);
751         if (r) {
752                 ti->error = "error getting device";
753                 goto bad;
754         }
755
756         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
757                 q = bdev_get_queue(p->path.dev->bdev);
758
759         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
760 retain:
761                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
762                 if (attached_handler_name) {
763                         /*
764                          * Clear any hw_handler_params associated with a
765                          * handler that isn't already attached.
766                          */
767                         if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
768                                 kfree(m->hw_handler_params);
769                                 m->hw_handler_params = NULL;
770                         }
771
772                         /*
773                          * Reset hw_handler_name to match the attached handler
774                          *
775                          * NB. This modifies the table line to show the actual
776                          * handler instead of the original table passed in.
777                          */
778                         kfree(m->hw_handler_name);
779                         m->hw_handler_name = attached_handler_name;
780                 }
781         }
782
783         if (m->hw_handler_name) {
784                 r = scsi_dh_attach(q, m->hw_handler_name);
785                 if (r == -EBUSY) {
786                         char b[BDEVNAME_SIZE];
787
788                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
789                                 bdevname(p->path.dev->bdev, b));
790                         goto retain;
791                 }
792                 if (r < 0) {
793                         ti->error = "error attaching hardware handler";
794                         dm_put_device(ti, p->path.dev);
795                         goto bad;
796                 }
797
798                 if (m->hw_handler_params) {
799                         r = scsi_dh_set_params(q, m->hw_handler_params);
800                         if (r < 0) {
801                                 ti->error = "unable to set hardware "
802                                                         "handler parameters";
803                                 dm_put_device(ti, p->path.dev);
804                                 goto bad;
805                         }
806                 }
807         }
808
809         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
810         if (r) {
811                 dm_put_device(ti, p->path.dev);
812                 goto bad;
813         }
814
815         return p;
816
817  bad:
818         free_pgpath(p);
819         return ERR_PTR(r);
820 }
821
822 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
823                                                    struct multipath *m)
824 {
825         static struct dm_arg _args[] = {
826                 {1, 1024, "invalid number of paths"},
827                 {0, 1024, "invalid number of selector args"}
828         };
829
830         int r;
831         unsigned i, nr_selector_args, nr_args;
832         struct priority_group *pg;
833         struct dm_target *ti = m->ti;
834
835         if (as->argc < 2) {
836                 as->argc = 0;
837                 ti->error = "not enough priority group arguments";
838                 return ERR_PTR(-EINVAL);
839         }
840
841         pg = alloc_priority_group();
842         if (!pg) {
843                 ti->error = "couldn't allocate priority group";
844                 return ERR_PTR(-ENOMEM);
845         }
846         pg->m = m;
847
848         r = parse_path_selector(as, pg, ti);
849         if (r)
850                 goto bad;
851
852         /*
853          * read the paths
854          */
855         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
856         if (r)
857                 goto bad;
858
859         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
860         if (r)
861                 goto bad;
862
863         nr_args = 1 + nr_selector_args;
864         for (i = 0; i < pg->nr_pgpaths; i++) {
865                 struct pgpath *pgpath;
866                 struct dm_arg_set path_args;
867
868                 if (as->argc < nr_args) {
869                         ti->error = "not enough path parameters";
870                         r = -EINVAL;
871                         goto bad;
872                 }
873
874                 path_args.argc = nr_args;
875                 path_args.argv = as->argv;
876
877                 pgpath = parse_path(&path_args, &pg->ps, ti);
878                 if (IS_ERR(pgpath)) {
879                         r = PTR_ERR(pgpath);
880                         goto bad;
881                 }
882
883                 pgpath->pg = pg;
884                 list_add_tail(&pgpath->list, &pg->pgpaths);
885                 dm_consume_args(as, nr_args);
886         }
887
888         return pg;
889
890  bad:
891         free_priority_group(pg, ti);
892         return ERR_PTR(r);
893 }
894
895 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
896 {
897         unsigned hw_argc;
898         int ret;
899         struct dm_target *ti = m->ti;
900
901         static struct dm_arg _args[] = {
902                 {0, 1024, "invalid number of hardware handler args"},
903         };
904
905         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
906                 return -EINVAL;
907
908         if (!hw_argc)
909                 return 0;
910
911         if (m->queue_mode == DM_TYPE_BIO_BASED) {
912                 dm_consume_args(as, hw_argc);
913                 DMERR("bio-based multipath doesn't allow hardware handler args");
914                 return 0;
915         }
916
917         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
918         if (!m->hw_handler_name)
919                 return -EINVAL;
920
921         if (hw_argc > 1) {
922                 char *p;
923                 int i, j, len = 4;
924
925                 for (i = 0; i <= hw_argc - 2; i++)
926                         len += strlen(as->argv[i]) + 1;
927                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
928                 if (!p) {
929                         ti->error = "memory allocation failed";
930                         ret = -ENOMEM;
931                         goto fail;
932                 }
933                 j = sprintf(p, "%d", hw_argc - 1);
934                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
935                         j = sprintf(p, "%s", as->argv[i]);
936         }
937         dm_consume_args(as, hw_argc - 1);
938
939         return 0;
940 fail:
941         kfree(m->hw_handler_name);
942         m->hw_handler_name = NULL;
943         return ret;
944 }
945
946 static int parse_features(struct dm_arg_set *as, struct multipath *m)
947 {
948         int r;
949         unsigned argc;
950         struct dm_target *ti = m->ti;
951         const char *arg_name;
952
953         static struct dm_arg _args[] = {
954                 {0, 8, "invalid number of feature args"},
955                 {1, 50, "pg_init_retries must be between 1 and 50"},
956                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
957         };
958
959         r = dm_read_arg_group(_args, as, &argc, &ti->error);
960         if (r)
961                 return -EINVAL;
962
963         if (!argc)
964                 return 0;
965
966         do {
967                 arg_name = dm_shift_arg(as);
968                 argc--;
969
970                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
971                         r = queue_if_no_path(m, true, false);
972                         continue;
973                 }
974
975                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
976                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
977                         continue;
978                 }
979
980                 if (!strcasecmp(arg_name, "pg_init_retries") &&
981                     (argc >= 1)) {
982                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
983                         argc--;
984                         continue;
985                 }
986
987                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
988                     (argc >= 1)) {
989                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
990                         argc--;
991                         continue;
992                 }
993
994                 if (!strcasecmp(arg_name, "queue_mode") &&
995                     (argc >= 1)) {
996                         const char *queue_mode_name = dm_shift_arg(as);
997
998                         if (!strcasecmp(queue_mode_name, "bio"))
999                                 m->queue_mode = DM_TYPE_BIO_BASED;
1000                         else if (!strcasecmp(queue_mode_name, "rq"))
1001                                 m->queue_mode = DM_TYPE_REQUEST_BASED;
1002                         else if (!strcasecmp(queue_mode_name, "mq"))
1003                                 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1004                         else {
1005                                 ti->error = "Unknown 'queue_mode' requested";
1006                                 r = -EINVAL;
1007                         }
1008                         argc--;
1009                         continue;
1010                 }
1011
1012                 ti->error = "Unrecognised multipath feature request";
1013                 r = -EINVAL;
1014         } while (argc && !r);
1015
1016         return r;
1017 }
1018
1019 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1020 {
1021         /* target arguments */
1022         static struct dm_arg _args[] = {
1023                 {0, 1024, "invalid number of priority groups"},
1024                 {0, 1024, "invalid initial priority group number"},
1025         };
1026
1027         int r;
1028         struct multipath *m;
1029         struct dm_arg_set as;
1030         unsigned pg_count = 0;
1031         unsigned next_pg_num;
1032
1033         as.argc = argc;
1034         as.argv = argv;
1035
1036         m = alloc_multipath(ti);
1037         if (!m) {
1038                 ti->error = "can't allocate multipath";
1039                 return -EINVAL;
1040         }
1041
1042         r = parse_features(&as, m);
1043         if (r)
1044                 goto bad;
1045
1046         r = alloc_multipath_stage2(ti, m);
1047         if (r)
1048                 goto bad;
1049
1050         r = parse_hw_handler(&as, m);
1051         if (r)
1052                 goto bad;
1053
1054         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1055         if (r)
1056                 goto bad;
1057
1058         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1059         if (r)
1060                 goto bad;
1061
1062         if ((!m->nr_priority_groups && next_pg_num) ||
1063             (m->nr_priority_groups && !next_pg_num)) {
1064                 ti->error = "invalid initial priority group";
1065                 r = -EINVAL;
1066                 goto bad;
1067         }
1068
1069         /* parse the priority groups */
1070         while (as.argc) {
1071                 struct priority_group *pg;
1072                 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1073
1074                 pg = parse_priority_group(&as, m);
1075                 if (IS_ERR(pg)) {
1076                         r = PTR_ERR(pg);
1077                         goto bad;
1078                 }
1079
1080                 nr_valid_paths += pg->nr_pgpaths;
1081                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1082
1083                 list_add_tail(&pg->list, &m->priority_groups);
1084                 pg_count++;
1085                 pg->pg_num = pg_count;
1086                 if (!--next_pg_num)
1087                         m->next_pg = pg;
1088         }
1089
1090         if (pg_count != m->nr_priority_groups) {
1091                 ti->error = "priority group count mismatch";
1092                 r = -EINVAL;
1093                 goto bad;
1094         }
1095
1096         ti->num_flush_bios = 1;
1097         ti->num_discard_bios = 1;
1098         ti->num_write_same_bios = 1;
1099         ti->num_write_zeroes_bios = 1;
1100         if (m->queue_mode == DM_TYPE_BIO_BASED)
1101                 ti->per_io_data_size = multipath_per_bio_data_size();
1102         else
1103                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1104
1105         return 0;
1106
1107  bad:
1108         free_multipath(m);
1109         return r;
1110 }
1111
1112 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1113 {
1114         DEFINE_WAIT(wait);
1115
1116         while (1) {
1117                 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1118
1119                 if (!atomic_read(&m->pg_init_in_progress))
1120                         break;
1121
1122                 io_schedule();
1123         }
1124         finish_wait(&m->pg_init_wait, &wait);
1125 }
1126
1127 static void flush_multipath_work(struct multipath *m)
1128 {
1129         set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1130         smp_mb__after_atomic();
1131
1132         flush_workqueue(kmpath_handlerd);
1133         multipath_wait_for_pg_init_completion(m);
1134         flush_workqueue(kmultipathd);
1135         flush_work(&m->trigger_event);
1136
1137         clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1138         smp_mb__after_atomic();
1139 }
1140
1141 static void multipath_dtr(struct dm_target *ti)
1142 {
1143         struct multipath *m = ti->private;
1144
1145         flush_multipath_work(m);
1146         free_multipath(m);
1147 }
1148
1149 /*
1150  * Take a path out of use.
1151  */
1152 static int fail_path(struct pgpath *pgpath)
1153 {
1154         unsigned long flags;
1155         struct multipath *m = pgpath->pg->m;
1156
1157         spin_lock_irqsave(&m->lock, flags);
1158
1159         if (!pgpath->is_active)
1160                 goto out;
1161
1162         DMWARN("Failing path %s.", pgpath->path.dev->name);
1163
1164         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1165         pgpath->is_active = false;
1166         pgpath->fail_count++;
1167
1168         atomic_dec(&m->nr_valid_paths);
1169
1170         if (pgpath == m->current_pgpath)
1171                 m->current_pgpath = NULL;
1172
1173         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1174                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1175
1176         schedule_work(&m->trigger_event);
1177
1178 out:
1179         spin_unlock_irqrestore(&m->lock, flags);
1180
1181         return 0;
1182 }
1183
1184 /*
1185  * Reinstate a previously-failed path
1186  */
1187 static int reinstate_path(struct pgpath *pgpath)
1188 {
1189         int r = 0, run_queue = 0;
1190         unsigned long flags;
1191         struct multipath *m = pgpath->pg->m;
1192         unsigned nr_valid_paths;
1193
1194         spin_lock_irqsave(&m->lock, flags);
1195
1196         if (pgpath->is_active)
1197                 goto out;
1198
1199         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1200
1201         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1202         if (r)
1203                 goto out;
1204
1205         pgpath->is_active = true;
1206
1207         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1208         if (nr_valid_paths == 1) {
1209                 m->current_pgpath = NULL;
1210                 run_queue = 1;
1211         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1212                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1213                         atomic_inc(&m->pg_init_in_progress);
1214         }
1215
1216         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1217                        pgpath->path.dev->name, nr_valid_paths);
1218
1219         schedule_work(&m->trigger_event);
1220
1221 out:
1222         spin_unlock_irqrestore(&m->lock, flags);
1223         if (run_queue) {
1224                 dm_table_run_md_queue_async(m->ti->table);
1225                 process_queued_io_list(m);
1226         }
1227
1228         return r;
1229 }
1230
1231 /*
1232  * Fail or reinstate all paths that match the provided struct dm_dev.
1233  */
1234 static int action_dev(struct multipath *m, struct dm_dev *dev,
1235                       action_fn action)
1236 {
1237         int r = -EINVAL;
1238         struct pgpath *pgpath;
1239         struct priority_group *pg;
1240
1241         list_for_each_entry(pg, &m->priority_groups, list) {
1242                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1243                         if (pgpath->path.dev == dev)
1244                                 r = action(pgpath);
1245                 }
1246         }
1247
1248         return r;
1249 }
1250
1251 /*
1252  * Temporarily try to avoid having to use the specified PG
1253  */
1254 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1255                       bool bypassed)
1256 {
1257         unsigned long flags;
1258
1259         spin_lock_irqsave(&m->lock, flags);
1260
1261         pg->bypassed = bypassed;
1262         m->current_pgpath = NULL;
1263         m->current_pg = NULL;
1264
1265         spin_unlock_irqrestore(&m->lock, flags);
1266
1267         schedule_work(&m->trigger_event);
1268 }
1269
1270 /*
1271  * Switch to using the specified PG from the next I/O that gets mapped
1272  */
1273 static int switch_pg_num(struct multipath *m, const char *pgstr)
1274 {
1275         struct priority_group *pg;
1276         unsigned pgnum;
1277         unsigned long flags;
1278         char dummy;
1279
1280         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1281             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1282                 DMWARN("invalid PG number supplied to switch_pg_num");
1283                 return -EINVAL;
1284         }
1285
1286         spin_lock_irqsave(&m->lock, flags);
1287         list_for_each_entry(pg, &m->priority_groups, list) {
1288                 pg->bypassed = false;
1289                 if (--pgnum)
1290                         continue;
1291
1292                 m->current_pgpath = NULL;
1293                 m->current_pg = NULL;
1294                 m->next_pg = pg;
1295         }
1296         spin_unlock_irqrestore(&m->lock, flags);
1297
1298         schedule_work(&m->trigger_event);
1299         return 0;
1300 }
1301
1302 /*
1303  * Set/clear bypassed status of a PG.
1304  * PGs are numbered upwards from 1 in the order they were declared.
1305  */
1306 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1307 {
1308         struct priority_group *pg;
1309         unsigned pgnum;
1310         char dummy;
1311
1312         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1313             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1314                 DMWARN("invalid PG number supplied to bypass_pg");
1315                 return -EINVAL;
1316         }
1317
1318         list_for_each_entry(pg, &m->priority_groups, list) {
1319                 if (!--pgnum)
1320                         break;
1321         }
1322
1323         bypass_pg(m, pg, bypassed);
1324         return 0;
1325 }
1326
1327 /*
1328  * Should we retry pg_init immediately?
1329  */
1330 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1331 {
1332         unsigned long flags;
1333         bool limit_reached = false;
1334
1335         spin_lock_irqsave(&m->lock, flags);
1336
1337         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1338             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1339                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1340         else
1341                 limit_reached = true;
1342
1343         spin_unlock_irqrestore(&m->lock, flags);
1344
1345         return limit_reached;
1346 }
1347
1348 static void pg_init_done(void *data, int errors)
1349 {
1350         struct pgpath *pgpath = data;
1351         struct priority_group *pg = pgpath->pg;
1352         struct multipath *m = pg->m;
1353         unsigned long flags;
1354         bool delay_retry = false;
1355
1356         /* device or driver problems */
1357         switch (errors) {
1358         case SCSI_DH_OK:
1359                 break;
1360         case SCSI_DH_NOSYS:
1361                 if (!m->hw_handler_name) {
1362                         errors = 0;
1363                         break;
1364                 }
1365                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1366                       "Error %d.", m->hw_handler_name, errors);
1367                 /*
1368                  * Fail path for now, so we do not ping pong
1369                  */
1370                 fail_path(pgpath);
1371                 break;
1372         case SCSI_DH_DEV_TEMP_BUSY:
1373                 /*
1374                  * Probably doing something like FW upgrade on the
1375                  * controller so try the other pg.
1376                  */
1377                 bypass_pg(m, pg, true);
1378                 break;
1379         case SCSI_DH_RETRY:
1380                 /* Wait before retrying. */
1381                 delay_retry = 1;
1382         case SCSI_DH_IMM_RETRY:
1383         case SCSI_DH_RES_TEMP_UNAVAIL:
1384                 if (pg_init_limit_reached(m, pgpath))
1385                         fail_path(pgpath);
1386                 errors = 0;
1387                 break;
1388         case SCSI_DH_DEV_OFFLINED:
1389         default:
1390                 /*
1391                  * We probably do not want to fail the path for a device
1392                  * error, but this is what the old dm did. In future
1393                  * patches we can do more advanced handling.
1394                  */
1395                 fail_path(pgpath);
1396         }
1397
1398         spin_lock_irqsave(&m->lock, flags);
1399         if (errors) {
1400                 if (pgpath == m->current_pgpath) {
1401                         DMERR("Could not failover device. Error %d.", errors);
1402                         m->current_pgpath = NULL;
1403                         m->current_pg = NULL;
1404                 }
1405         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1406                 pg->bypassed = false;
1407
1408         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1409                 /* Activations of other paths are still on going */
1410                 goto out;
1411
1412         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1413                 if (delay_retry)
1414                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1415                 else
1416                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1417
1418                 if (__pg_init_all_paths(m))
1419                         goto out;
1420         }
1421         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1422
1423         process_queued_io_list(m);
1424
1425         /*
1426          * Wake up any thread waiting to suspend.
1427          */
1428         wake_up(&m->pg_init_wait);
1429
1430 out:
1431         spin_unlock_irqrestore(&m->lock, flags);
1432 }
1433
1434 static void activate_or_offline_path(struct pgpath *pgpath)
1435 {
1436         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1437
1438         if (pgpath->is_active && !blk_queue_dying(q))
1439                 scsi_dh_activate(q, pg_init_done, pgpath);
1440         else
1441                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1442 }
1443
1444 static void activate_path_work(struct work_struct *work)
1445 {
1446         struct pgpath *pgpath =
1447                 container_of(work, struct pgpath, activate_path.work);
1448
1449         activate_or_offline_path(pgpath);
1450 }
1451
1452 static int noretry_error(blk_status_t error)
1453 {
1454         switch (error) {
1455         case BLK_STS_NOTSUPP:
1456         case BLK_STS_NOSPC:
1457         case BLK_STS_TARGET:
1458         case BLK_STS_NEXUS:
1459         case BLK_STS_MEDIUM:
1460         case BLK_STS_RESOURCE:
1461                 return 1;
1462         }
1463
1464         /* Anything else could be a path failure, so should be retried */
1465         return 0;
1466 }
1467
1468 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1469                             blk_status_t error, union map_info *map_context)
1470 {
1471         struct dm_mpath_io *mpio = get_mpio(map_context);
1472         struct pgpath *pgpath = mpio->pgpath;
1473         int r = DM_ENDIO_DONE;
1474
1475         /*
1476          * We don't queue any clone request inside the multipath target
1477          * during end I/O handling, since those clone requests don't have
1478          * bio clones.  If we queue them inside the multipath target,
1479          * we need to make bio clones, that requires memory allocation.
1480          * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1481          *  don't have bio clones.)
1482          * Instead of queueing the clone request here, we queue the original
1483          * request into dm core, which will remake a clone request and
1484          * clone bios for it and resubmit it later.
1485          */
1486         if (error && !noretry_error(error)) {
1487                 struct multipath *m = ti->private;
1488
1489                 r = DM_ENDIO_REQUEUE;
1490
1491                 if (pgpath)
1492                         fail_path(pgpath);
1493
1494                 if (atomic_read(&m->nr_valid_paths) == 0 &&
1495                     !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1496                         if (error == BLK_STS_IOERR)
1497                                 dm_report_EIO(m);
1498                         /* complete with the original error */
1499                         r = DM_ENDIO_DONE;
1500                 }
1501         }
1502
1503         if (pgpath) {
1504                 struct path_selector *ps = &pgpath->pg->ps;
1505
1506                 if (ps->type->end_io)
1507                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1508         }
1509
1510         return r;
1511 }
1512
1513 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone, int *error)
1514 {
1515         struct multipath *m = ti->private;
1516         struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1517         struct pgpath *pgpath = mpio->pgpath;
1518         unsigned long flags;
1519         int r = DM_ENDIO_DONE;
1520
1521         if (!*error || noretry_error(errno_to_blk_status(*error)))
1522                 goto done;
1523
1524         if (pgpath)
1525                 fail_path(pgpath);
1526
1527         if (atomic_read(&m->nr_valid_paths) == 0 &&
1528             !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1529                 dm_report_EIO(m);
1530                 *error = -EIO;
1531                 goto done;
1532         }
1533
1534         /* Queue for the daemon to resubmit */
1535         dm_bio_restore(get_bio_details_from_bio(clone), clone);
1536
1537         spin_lock_irqsave(&m->lock, flags);
1538         bio_list_add(&m->queued_bios, clone);
1539         spin_unlock_irqrestore(&m->lock, flags);
1540         if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1541                 queue_work(kmultipathd, &m->process_queued_bios);
1542
1543         r = DM_ENDIO_INCOMPLETE;
1544 done:
1545         if (pgpath) {
1546                 struct path_selector *ps = &pgpath->pg->ps;
1547
1548                 if (ps->type->end_io)
1549                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1550         }
1551
1552         return r;
1553 }
1554
1555 /*
1556  * Suspend can't complete until all the I/O is processed so if
1557  * the last path fails we must error any remaining I/O.
1558  * Note that if the freeze_bdev fails while suspending, the
1559  * queue_if_no_path state is lost - userspace should reset it.
1560  */
1561 static void multipath_presuspend(struct dm_target *ti)
1562 {
1563         struct multipath *m = ti->private;
1564
1565         queue_if_no_path(m, false, true);
1566 }
1567
1568 static void multipath_postsuspend(struct dm_target *ti)
1569 {
1570         struct multipath *m = ti->private;
1571
1572         mutex_lock(&m->work_mutex);
1573         flush_multipath_work(m);
1574         mutex_unlock(&m->work_mutex);
1575 }
1576
1577 /*
1578  * Restore the queue_if_no_path setting.
1579  */
1580 static void multipath_resume(struct dm_target *ti)
1581 {
1582         struct multipath *m = ti->private;
1583         unsigned long flags;
1584
1585         spin_lock_irqsave(&m->lock, flags);
1586         assign_bit(test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
1587                    MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1588         spin_unlock_irqrestore(&m->lock, flags);
1589 }
1590
1591 /*
1592  * Info output has the following format:
1593  * num_multipath_feature_args [multipath_feature_args]*
1594  * num_handler_status_args [handler_status_args]*
1595  * num_groups init_group_number
1596  *            [A|D|E num_ps_status_args [ps_status_args]*
1597  *             num_paths num_selector_args
1598  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1599  *
1600  * Table output has the following format (identical to the constructor string):
1601  * num_feature_args [features_args]*
1602  * num_handler_args hw_handler [hw_handler_args]*
1603  * num_groups init_group_number
1604  *     [priority selector-name num_ps_args [ps_args]*
1605  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1606  */
1607 static void multipath_status(struct dm_target *ti, status_type_t type,
1608                              unsigned status_flags, char *result, unsigned maxlen)
1609 {
1610         int sz = 0;
1611         unsigned long flags;
1612         struct multipath *m = ti->private;
1613         struct priority_group *pg;
1614         struct pgpath *p;
1615         unsigned pg_num;
1616         char state;
1617
1618         spin_lock_irqsave(&m->lock, flags);
1619
1620         /* Features */
1621         if (type == STATUSTYPE_INFO)
1622                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1623                        atomic_read(&m->pg_init_count));
1624         else {
1625                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1626                               (m->pg_init_retries > 0) * 2 +
1627                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1628                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1629                               (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1630
1631                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1632                         DMEMIT("queue_if_no_path ");
1633                 if (m->pg_init_retries)
1634                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1635                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1636                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1637                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1638                         DMEMIT("retain_attached_hw_handler ");
1639                 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1640                         switch(m->queue_mode) {
1641                         case DM_TYPE_BIO_BASED:
1642                                 DMEMIT("queue_mode bio ");
1643                                 break;
1644                         case DM_TYPE_MQ_REQUEST_BASED:
1645                                 DMEMIT("queue_mode mq ");
1646                                 break;
1647                         default:
1648                                 WARN_ON_ONCE(true);
1649                                 break;
1650                         }
1651                 }
1652         }
1653
1654         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1655                 DMEMIT("0 ");
1656         else
1657                 DMEMIT("1 %s ", m->hw_handler_name);
1658
1659         DMEMIT("%u ", m->nr_priority_groups);
1660
1661         if (m->next_pg)
1662                 pg_num = m->next_pg->pg_num;
1663         else if (m->current_pg)
1664                 pg_num = m->current_pg->pg_num;
1665         else
1666                 pg_num = (m->nr_priority_groups ? 1 : 0);
1667
1668         DMEMIT("%u ", pg_num);
1669
1670         switch (type) {
1671         case STATUSTYPE_INFO:
1672                 list_for_each_entry(pg, &m->priority_groups, list) {
1673                         if (pg->bypassed)
1674                                 state = 'D';    /* Disabled */
1675                         else if (pg == m->current_pg)
1676                                 state = 'A';    /* Currently Active */
1677                         else
1678                                 state = 'E';    /* Enabled */
1679
1680                         DMEMIT("%c ", state);
1681
1682                         if (pg->ps.type->status)
1683                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1684                                                           result + sz,
1685                                                           maxlen - sz);
1686                         else
1687                                 DMEMIT("0 ");
1688
1689                         DMEMIT("%u %u ", pg->nr_pgpaths,
1690                                pg->ps.type->info_args);
1691
1692                         list_for_each_entry(p, &pg->pgpaths, list) {
1693                                 DMEMIT("%s %s %u ", p->path.dev->name,
1694                                        p->is_active ? "A" : "F",
1695                                        p->fail_count);
1696                                 if (pg->ps.type->status)
1697                                         sz += pg->ps.type->status(&pg->ps,
1698                                               &p->path, type, result + sz,
1699                                               maxlen - sz);
1700                         }
1701                 }
1702                 break;
1703
1704         case STATUSTYPE_TABLE:
1705                 list_for_each_entry(pg, &m->priority_groups, list) {
1706                         DMEMIT("%s ", pg->ps.type->name);
1707
1708                         if (pg->ps.type->status)
1709                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1710                                                           result + sz,
1711                                                           maxlen - sz);
1712                         else
1713                                 DMEMIT("0 ");
1714
1715                         DMEMIT("%u %u ", pg->nr_pgpaths,
1716                                pg->ps.type->table_args);
1717
1718                         list_for_each_entry(p, &pg->pgpaths, list) {
1719                                 DMEMIT("%s ", p->path.dev->name);
1720                                 if (pg->ps.type->status)
1721                                         sz += pg->ps.type->status(&pg->ps,
1722                                               &p->path, type, result + sz,
1723                                               maxlen - sz);
1724                         }
1725                 }
1726                 break;
1727         }
1728
1729         spin_unlock_irqrestore(&m->lock, flags);
1730 }
1731
1732 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1733 {
1734         int r = -EINVAL;
1735         struct dm_dev *dev;
1736         struct multipath *m = ti->private;
1737         action_fn action;
1738
1739         mutex_lock(&m->work_mutex);
1740
1741         if (dm_suspended(ti)) {
1742                 r = -EBUSY;
1743                 goto out;
1744         }
1745
1746         if (argc == 1) {
1747                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1748                         r = queue_if_no_path(m, true, false);
1749                         goto out;
1750                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1751                         r = queue_if_no_path(m, false, false);
1752                         goto out;
1753                 }
1754         }
1755
1756         if (argc != 2) {
1757                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1758                 goto out;
1759         }
1760
1761         if (!strcasecmp(argv[0], "disable_group")) {
1762                 r = bypass_pg_num(m, argv[1], true);
1763                 goto out;
1764         } else if (!strcasecmp(argv[0], "enable_group")) {
1765                 r = bypass_pg_num(m, argv[1], false);
1766                 goto out;
1767         } else if (!strcasecmp(argv[0], "switch_group")) {
1768                 r = switch_pg_num(m, argv[1]);
1769                 goto out;
1770         } else if (!strcasecmp(argv[0], "reinstate_path"))
1771                 action = reinstate_path;
1772         else if (!strcasecmp(argv[0], "fail_path"))
1773                 action = fail_path;
1774         else {
1775                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1776                 goto out;
1777         }
1778
1779         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1780         if (r) {
1781                 DMWARN("message: error getting device %s",
1782                        argv[1]);
1783                 goto out;
1784         }
1785
1786         r = action_dev(m, dev, action);
1787
1788         dm_put_device(ti, dev);
1789
1790 out:
1791         mutex_unlock(&m->work_mutex);
1792         return r;
1793 }
1794
1795 static int multipath_prepare_ioctl(struct dm_target *ti,
1796                 struct block_device **bdev, fmode_t *mode)
1797 {
1798         struct multipath *m = ti->private;
1799         struct pgpath *current_pgpath;
1800         int r;
1801
1802         current_pgpath = lockless_dereference(m->current_pgpath);
1803         if (!current_pgpath)
1804                 current_pgpath = choose_pgpath(m, 0);
1805
1806         if (current_pgpath) {
1807                 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1808                         *bdev = current_pgpath->path.dev->bdev;
1809                         *mode = current_pgpath->path.dev->mode;
1810                         r = 0;
1811                 } else {
1812                         /* pg_init has not started or completed */
1813                         r = -ENOTCONN;
1814                 }
1815         } else {
1816                 /* No path is available */
1817                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1818                         r = -ENOTCONN;
1819                 else
1820                         r = -EIO;
1821         }
1822
1823         if (r == -ENOTCONN) {
1824                 if (!lockless_dereference(m->current_pg)) {
1825                         /* Path status changed, redo selection */
1826                         (void) choose_pgpath(m, 0);
1827                 }
1828                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1829                         pg_init_all_paths(m);
1830                 dm_table_run_md_queue_async(m->ti->table);
1831                 process_queued_io_list(m);
1832         }
1833
1834         /*
1835          * Only pass ioctls through if the device sizes match exactly.
1836          */
1837         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1838                 return 1;
1839         return r;
1840 }
1841
1842 static int multipath_iterate_devices(struct dm_target *ti,
1843                                      iterate_devices_callout_fn fn, void *data)
1844 {
1845         struct multipath *m = ti->private;
1846         struct priority_group *pg;
1847         struct pgpath *p;
1848         int ret = 0;
1849
1850         list_for_each_entry(pg, &m->priority_groups, list) {
1851                 list_for_each_entry(p, &pg->pgpaths, list) {
1852                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1853                         if (ret)
1854                                 goto out;
1855                 }
1856         }
1857
1858 out:
1859         return ret;
1860 }
1861
1862 static int pgpath_busy(struct pgpath *pgpath)
1863 {
1864         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1865
1866         return blk_lld_busy(q);
1867 }
1868
1869 /*
1870  * We return "busy", only when we can map I/Os but underlying devices
1871  * are busy (so even if we map I/Os now, the I/Os will wait on
1872  * the underlying queue).
1873  * In other words, if we want to kill I/Os or queue them inside us
1874  * due to map unavailability, we don't return "busy".  Otherwise,
1875  * dm core won't give us the I/Os and we can't do what we want.
1876  */
1877 static int multipath_busy(struct dm_target *ti)
1878 {
1879         bool busy = false, has_active = false;
1880         struct multipath *m = ti->private;
1881         struct priority_group *pg, *next_pg;
1882         struct pgpath *pgpath;
1883
1884         /* pg_init in progress */
1885         if (atomic_read(&m->pg_init_in_progress))
1886                 return true;
1887
1888         /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1889         if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1890                 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1891
1892         /* Guess which priority_group will be used at next mapping time */
1893         pg = lockless_dereference(m->current_pg);
1894         next_pg = lockless_dereference(m->next_pg);
1895         if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
1896                 pg = next_pg;
1897
1898         if (!pg) {
1899                 /*
1900                  * We don't know which pg will be used at next mapping time.
1901                  * We don't call choose_pgpath() here to avoid to trigger
1902                  * pg_init just by busy checking.
1903                  * So we don't know whether underlying devices we will be using
1904                  * at next mapping time are busy or not. Just try mapping.
1905                  */
1906                 return busy;
1907         }
1908
1909         /*
1910          * If there is one non-busy active path at least, the path selector
1911          * will be able to select it. So we consider such a pg as not busy.
1912          */
1913         busy = true;
1914         list_for_each_entry(pgpath, &pg->pgpaths, list) {
1915                 if (pgpath->is_active) {
1916                         has_active = true;
1917                         if (!pgpath_busy(pgpath)) {
1918                                 busy = false;
1919                                 break;
1920                         }
1921                 }
1922         }
1923
1924         if (!has_active) {
1925                 /*
1926                  * No active path in this pg, so this pg won't be used and
1927                  * the current_pg will be changed at next mapping time.
1928                  * We need to try mapping to determine it.
1929                  */
1930                 busy = false;
1931         }
1932
1933         return busy;
1934 }
1935
1936 /*-----------------------------------------------------------------
1937  * Module setup
1938  *---------------------------------------------------------------*/
1939 static struct target_type multipath_target = {
1940         .name = "multipath",
1941         .version = {1, 12, 0},
1942         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1943         .module = THIS_MODULE,
1944         .ctr = multipath_ctr,
1945         .dtr = multipath_dtr,
1946         .clone_and_map_rq = multipath_clone_and_map,
1947         .release_clone_rq = multipath_release_clone,
1948         .rq_end_io = multipath_end_io,
1949         .map = multipath_map_bio,
1950         .end_io = multipath_end_io_bio,
1951         .presuspend = multipath_presuspend,
1952         .postsuspend = multipath_postsuspend,
1953         .resume = multipath_resume,
1954         .status = multipath_status,
1955         .message = multipath_message,
1956         .prepare_ioctl = multipath_prepare_ioctl,
1957         .iterate_devices = multipath_iterate_devices,
1958         .busy = multipath_busy,
1959 };
1960
1961 static int __init dm_multipath_init(void)
1962 {
1963         int r;
1964
1965         r = dm_register_target(&multipath_target);
1966         if (r < 0) {
1967                 DMERR("request-based register failed %d", r);
1968                 r = -EINVAL;
1969                 goto bad_register_target;
1970         }
1971
1972         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1973         if (!kmultipathd) {
1974                 DMERR("failed to create workqueue kmpathd");
1975                 r = -ENOMEM;
1976                 goto bad_alloc_kmultipathd;
1977         }
1978
1979         /*
1980          * A separate workqueue is used to handle the device handlers
1981          * to avoid overloading existing workqueue. Overloading the
1982          * old workqueue would also create a bottleneck in the
1983          * path of the storage hardware device activation.
1984          */
1985         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1986                                                   WQ_MEM_RECLAIM);
1987         if (!kmpath_handlerd) {
1988                 DMERR("failed to create workqueue kmpath_handlerd");
1989                 r = -ENOMEM;
1990                 goto bad_alloc_kmpath_handlerd;
1991         }
1992
1993         return 0;
1994
1995 bad_alloc_kmpath_handlerd:
1996         destroy_workqueue(kmultipathd);
1997 bad_alloc_kmultipathd:
1998         dm_unregister_target(&multipath_target);
1999 bad_register_target:
2000         return r;
2001 }
2002
2003 static void __exit dm_multipath_exit(void)
2004 {
2005         destroy_workqueue(kmpath_handlerd);
2006         destroy_workqueue(kmultipathd);
2007
2008         dm_unregister_target(&multipath_target);
2009 }
2010
2011 module_init(dm_multipath_init);
2012 module_exit(dm_multipath_exit);
2013
2014 MODULE_DESCRIPTION(DM_NAME " multipath target");
2015 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2016 MODULE_LICENSE("GPL");