ubifs: free the encrypted symlink target
[platform/kernel/linux-rpi.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                 if (queue_dying) {
503                         atomic_inc(&m->pg_init_in_progress);
504                         activate_or_offline_path(pgpath);
505                 }
506                 return DM_MAPIO_DELAY_REQUEUE;
507         }
508         clone->bio = clone->biotail = NULL;
509         clone->rq_disk = bdev->bd_disk;
510         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
511         *__clone = clone;
512
513         if (pgpath->pg->ps.type->start_io)
514                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
515                                               &pgpath->path,
516                                               nr_bytes);
517         return DM_MAPIO_REMAPPED;
518 }
519
520 static void multipath_release_clone(struct request *clone)
521 {
522         blk_put_request(clone);
523 }
524
525 /*
526  * Map cloned bios (bio-based multipath)
527  */
528 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
529 {
530         size_t nr_bytes = bio->bi_iter.bi_size;
531         struct pgpath *pgpath;
532         unsigned long flags;
533         bool queue_io;
534
535         /* Do we need to select a new pgpath? */
536         pgpath = READ_ONCE(m->current_pgpath);
537         queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
538         if (!pgpath || !queue_io)
539                 pgpath = choose_pgpath(m, nr_bytes);
540
541         if ((pgpath && queue_io) ||
542             (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
543                 /* Queue for the daemon to resubmit */
544                 spin_lock_irqsave(&m->lock, flags);
545                 bio_list_add(&m->queued_bios, bio);
546                 spin_unlock_irqrestore(&m->lock, flags);
547                 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
548                 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
549                         pg_init_all_paths(m);
550                 else if (!queue_io)
551                         queue_work(kmultipathd, &m->process_queued_bios);
552                 return DM_MAPIO_SUBMITTED;
553         }
554
555         if (!pgpath) {
556                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
557                         return DM_MAPIO_REQUEUE;
558                 dm_report_EIO(m);
559                 return DM_MAPIO_KILL;
560         }
561
562         mpio->pgpath = pgpath;
563         mpio->nr_bytes = nr_bytes;
564
565         bio->bi_status = 0;
566         bio_set_dev(bio, pgpath->path.dev->bdev);
567         bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
568
569         if (pgpath->pg->ps.type->start_io)
570                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
571                                               &pgpath->path,
572                                               nr_bytes);
573         return DM_MAPIO_REMAPPED;
574 }
575
576 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
577 {
578         struct multipath *m = ti->private;
579         struct dm_mpath_io *mpio = NULL;
580
581         multipath_init_per_bio_data(bio, &mpio, NULL);
582
583         return __multipath_map_bio(m, bio, mpio);
584 }
585
586 static void process_queued_io_list(struct multipath *m)
587 {
588         if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
589                 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
590         else if (m->queue_mode == DM_TYPE_BIO_BASED)
591                 queue_work(kmultipathd, &m->process_queued_bios);
592 }
593
594 static void process_queued_bios(struct work_struct *work)
595 {
596         int r;
597         unsigned long flags;
598         struct bio *bio;
599         struct bio_list bios;
600         struct blk_plug plug;
601         struct multipath *m =
602                 container_of(work, struct multipath, process_queued_bios);
603
604         bio_list_init(&bios);
605
606         spin_lock_irqsave(&m->lock, flags);
607
608         if (bio_list_empty(&m->queued_bios)) {
609                 spin_unlock_irqrestore(&m->lock, flags);
610                 return;
611         }
612
613         bio_list_merge(&bios, &m->queued_bios);
614         bio_list_init(&m->queued_bios);
615
616         spin_unlock_irqrestore(&m->lock, flags);
617
618         blk_start_plug(&plug);
619         while ((bio = bio_list_pop(&bios))) {
620                 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
621                 switch (r) {
622                 case DM_MAPIO_KILL:
623                         bio->bi_status = BLK_STS_IOERR;
624                         bio_endio(bio);
625                         break;
626                 case DM_MAPIO_REQUEUE:
627                         bio->bi_status = BLK_STS_DM_REQUEUE;
628                         bio_endio(bio);
629                         break;
630                 case DM_MAPIO_REMAPPED:
631                         generic_make_request(bio);
632                         break;
633                 case 0:
634                         break;
635                 default:
636                         WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
637                 }
638         }
639         blk_finish_plug(&plug);
640 }
641
642 static void assign_bit(bool value, long nr, unsigned long *addr)
643 {
644         if (value)
645                 set_bit(nr, addr);
646         else
647                 clear_bit(nr, addr);
648 }
649
650 /*
651  * If we run out of usable paths, should we queue I/O or error it?
652  */
653 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
654                             bool save_old_value)
655 {
656         unsigned long flags;
657
658         spin_lock_irqsave(&m->lock, flags);
659         assign_bit((save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
660                    (!save_old_value && queue_if_no_path),
661                    MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
662         assign_bit(queue_if_no_path || dm_noflush_suspending(m->ti),
663                    MPATHF_QUEUE_IF_NO_PATH, &m->flags);
664         spin_unlock_irqrestore(&m->lock, flags);
665
666         if (!queue_if_no_path) {
667                 dm_table_run_md_queue_async(m->ti->table);
668                 process_queued_io_list(m);
669         }
670
671         return 0;
672 }
673
674 /*
675  * An event is triggered whenever a path is taken out of use.
676  * Includes path failure and PG bypass.
677  */
678 static void trigger_event(struct work_struct *work)
679 {
680         struct multipath *m =
681                 container_of(work, struct multipath, trigger_event);
682
683         dm_table_event(m->ti->table);
684 }
685
686 /*-----------------------------------------------------------------
687  * Constructor/argument parsing:
688  * <#multipath feature args> [<arg>]*
689  * <#hw_handler args> [hw_handler [<arg>]*]
690  * <#priority groups>
691  * <initial priority group>
692  *     [<selector> <#selector args> [<arg>]*
693  *      <#paths> <#per-path selector args>
694  *         [<path> [<arg>]* ]+ ]+
695  *---------------------------------------------------------------*/
696 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
697                                struct dm_target *ti)
698 {
699         int r;
700         struct path_selector_type *pst;
701         unsigned ps_argc;
702
703         static const struct dm_arg _args[] = {
704                 {0, 1024, "invalid number of path selector args"},
705         };
706
707         pst = dm_get_path_selector(dm_shift_arg(as));
708         if (!pst) {
709                 ti->error = "unknown path selector type";
710                 return -EINVAL;
711         }
712
713         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
714         if (r) {
715                 dm_put_path_selector(pst);
716                 return -EINVAL;
717         }
718
719         r = pst->create(&pg->ps, ps_argc, as->argv);
720         if (r) {
721                 dm_put_path_selector(pst);
722                 ti->error = "path selector constructor failed";
723                 return r;
724         }
725
726         pg->ps.type = pst;
727         dm_consume_args(as, ps_argc);
728
729         return 0;
730 }
731
732 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
733                                struct dm_target *ti)
734 {
735         int r;
736         struct pgpath *p;
737         struct multipath *m = ti->private;
738         struct request_queue *q = NULL;
739         const char *attached_handler_name;
740
741         /* we need at least a path arg */
742         if (as->argc < 1) {
743                 ti->error = "no device given";
744                 return ERR_PTR(-EINVAL);
745         }
746
747         p = alloc_pgpath();
748         if (!p)
749                 return ERR_PTR(-ENOMEM);
750
751         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
752                           &p->path.dev);
753         if (r) {
754                 ti->error = "error getting device";
755                 goto bad;
756         }
757
758         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
759                 q = bdev_get_queue(p->path.dev->bdev);
760
761         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
762 retain:
763                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
764                 if (attached_handler_name) {
765                         /*
766                          * Clear any hw_handler_params associated with a
767                          * handler that isn't already attached.
768                          */
769                         if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
770                                 kfree(m->hw_handler_params);
771                                 m->hw_handler_params = NULL;
772                         }
773
774                         /*
775                          * Reset hw_handler_name to match the attached handler
776                          *
777                          * NB. This modifies the table line to show the actual
778                          * handler instead of the original table passed in.
779                          */
780                         kfree(m->hw_handler_name);
781                         m->hw_handler_name = attached_handler_name;
782                 }
783         }
784
785         if (m->hw_handler_name) {
786                 r = scsi_dh_attach(q, m->hw_handler_name);
787                 if (r == -EBUSY) {
788                         char b[BDEVNAME_SIZE];
789
790                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
791                                 bdevname(p->path.dev->bdev, b));
792                         goto retain;
793                 }
794                 if (r < 0) {
795                         ti->error = "error attaching hardware handler";
796                         dm_put_device(ti, p->path.dev);
797                         goto bad;
798                 }
799
800                 if (m->hw_handler_params) {
801                         r = scsi_dh_set_params(q, m->hw_handler_params);
802                         if (r < 0) {
803                                 ti->error = "unable to set hardware "
804                                                         "handler parameters";
805                                 dm_put_device(ti, p->path.dev);
806                                 goto bad;
807                         }
808                 }
809         }
810
811         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
812         if (r) {
813                 dm_put_device(ti, p->path.dev);
814                 goto bad;
815         }
816
817         return p;
818
819  bad:
820         free_pgpath(p);
821         return ERR_PTR(r);
822 }
823
824 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
825                                                    struct multipath *m)
826 {
827         static const struct dm_arg _args[] = {
828                 {1, 1024, "invalid number of paths"},
829                 {0, 1024, "invalid number of selector args"}
830         };
831
832         int r;
833         unsigned i, nr_selector_args, nr_args;
834         struct priority_group *pg;
835         struct dm_target *ti = m->ti;
836
837         if (as->argc < 2) {
838                 as->argc = 0;
839                 ti->error = "not enough priority group arguments";
840                 return ERR_PTR(-EINVAL);
841         }
842
843         pg = alloc_priority_group();
844         if (!pg) {
845                 ti->error = "couldn't allocate priority group";
846                 return ERR_PTR(-ENOMEM);
847         }
848         pg->m = m;
849
850         r = parse_path_selector(as, pg, ti);
851         if (r)
852                 goto bad;
853
854         /*
855          * read the paths
856          */
857         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
858         if (r)
859                 goto bad;
860
861         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
862         if (r)
863                 goto bad;
864
865         nr_args = 1 + nr_selector_args;
866         for (i = 0; i < pg->nr_pgpaths; i++) {
867                 struct pgpath *pgpath;
868                 struct dm_arg_set path_args;
869
870                 if (as->argc < nr_args) {
871                         ti->error = "not enough path parameters";
872                         r = -EINVAL;
873                         goto bad;
874                 }
875
876                 path_args.argc = nr_args;
877                 path_args.argv = as->argv;
878
879                 pgpath = parse_path(&path_args, &pg->ps, ti);
880                 if (IS_ERR(pgpath)) {
881                         r = PTR_ERR(pgpath);
882                         goto bad;
883                 }
884
885                 pgpath->pg = pg;
886                 list_add_tail(&pgpath->list, &pg->pgpaths);
887                 dm_consume_args(as, nr_args);
888         }
889
890         return pg;
891
892  bad:
893         free_priority_group(pg, ti);
894         return ERR_PTR(r);
895 }
896
897 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
898 {
899         unsigned hw_argc;
900         int ret;
901         struct dm_target *ti = m->ti;
902
903         static const struct dm_arg _args[] = {
904                 {0, 1024, "invalid number of hardware handler args"},
905         };
906
907         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
908                 return -EINVAL;
909
910         if (!hw_argc)
911                 return 0;
912
913         if (m->queue_mode == DM_TYPE_BIO_BASED) {
914                 dm_consume_args(as, hw_argc);
915                 DMERR("bio-based multipath doesn't allow hardware handler args");
916                 return 0;
917         }
918
919         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
920         if (!m->hw_handler_name)
921                 return -EINVAL;
922
923         if (hw_argc > 1) {
924                 char *p;
925                 int i, j, len = 4;
926
927                 for (i = 0; i <= hw_argc - 2; i++)
928                         len += strlen(as->argv[i]) + 1;
929                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
930                 if (!p) {
931                         ti->error = "memory allocation failed";
932                         ret = -ENOMEM;
933                         goto fail;
934                 }
935                 j = sprintf(p, "%d", hw_argc - 1);
936                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
937                         j = sprintf(p, "%s", as->argv[i]);
938         }
939         dm_consume_args(as, hw_argc - 1);
940
941         return 0;
942 fail:
943         kfree(m->hw_handler_name);
944         m->hw_handler_name = NULL;
945         return ret;
946 }
947
948 static int parse_features(struct dm_arg_set *as, struct multipath *m)
949 {
950         int r;
951         unsigned argc;
952         struct dm_target *ti = m->ti;
953         const char *arg_name;
954
955         static const struct dm_arg _args[] = {
956                 {0, 8, "invalid number of feature args"},
957                 {1, 50, "pg_init_retries must be between 1 and 50"},
958                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
959         };
960
961         r = dm_read_arg_group(_args, as, &argc, &ti->error);
962         if (r)
963                 return -EINVAL;
964
965         if (!argc)
966                 return 0;
967
968         do {
969                 arg_name = dm_shift_arg(as);
970                 argc--;
971
972                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
973                         r = queue_if_no_path(m, true, false);
974                         continue;
975                 }
976
977                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
978                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
979                         continue;
980                 }
981
982                 if (!strcasecmp(arg_name, "pg_init_retries") &&
983                     (argc >= 1)) {
984                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
985                         argc--;
986                         continue;
987                 }
988
989                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
990                     (argc >= 1)) {
991                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
992                         argc--;
993                         continue;
994                 }
995
996                 if (!strcasecmp(arg_name, "queue_mode") &&
997                     (argc >= 1)) {
998                         const char *queue_mode_name = dm_shift_arg(as);
999
1000                         if (!strcasecmp(queue_mode_name, "bio"))
1001                                 m->queue_mode = DM_TYPE_BIO_BASED;
1002                         else if (!strcasecmp(queue_mode_name, "rq"))
1003                                 m->queue_mode = DM_TYPE_REQUEST_BASED;
1004                         else if (!strcasecmp(queue_mode_name, "mq"))
1005                                 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1006                         else {
1007                                 ti->error = "Unknown 'queue_mode' requested";
1008                                 r = -EINVAL;
1009                         }
1010                         argc--;
1011                         continue;
1012                 }
1013
1014                 ti->error = "Unrecognised multipath feature request";
1015                 r = -EINVAL;
1016         } while (argc && !r);
1017
1018         return r;
1019 }
1020
1021 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1022 {
1023         /* target arguments */
1024         static const struct dm_arg _args[] = {
1025                 {0, 1024, "invalid number of priority groups"},
1026                 {0, 1024, "invalid initial priority group number"},
1027         };
1028
1029         int r;
1030         struct multipath *m;
1031         struct dm_arg_set as;
1032         unsigned pg_count = 0;
1033         unsigned next_pg_num;
1034
1035         as.argc = argc;
1036         as.argv = argv;
1037
1038         m = alloc_multipath(ti);
1039         if (!m) {
1040                 ti->error = "can't allocate multipath";
1041                 return -EINVAL;
1042         }
1043
1044         r = parse_features(&as, m);
1045         if (r)
1046                 goto bad;
1047
1048         r = alloc_multipath_stage2(ti, m);
1049         if (r)
1050                 goto bad;
1051
1052         r = parse_hw_handler(&as, m);
1053         if (r)
1054                 goto bad;
1055
1056         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1057         if (r)
1058                 goto bad;
1059
1060         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1061         if (r)
1062                 goto bad;
1063
1064         if ((!m->nr_priority_groups && next_pg_num) ||
1065             (m->nr_priority_groups && !next_pg_num)) {
1066                 ti->error = "invalid initial priority group";
1067                 r = -EINVAL;
1068                 goto bad;
1069         }
1070
1071         /* parse the priority groups */
1072         while (as.argc) {
1073                 struct priority_group *pg;
1074                 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1075
1076                 pg = parse_priority_group(&as, m);
1077                 if (IS_ERR(pg)) {
1078                         r = PTR_ERR(pg);
1079                         goto bad;
1080                 }
1081
1082                 nr_valid_paths += pg->nr_pgpaths;
1083                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1084
1085                 list_add_tail(&pg->list, &m->priority_groups);
1086                 pg_count++;
1087                 pg->pg_num = pg_count;
1088                 if (!--next_pg_num)
1089                         m->next_pg = pg;
1090         }
1091
1092         if (pg_count != m->nr_priority_groups) {
1093                 ti->error = "priority group count mismatch";
1094                 r = -EINVAL;
1095                 goto bad;
1096         }
1097
1098         ti->num_flush_bios = 1;
1099         ti->num_discard_bios = 1;
1100         ti->num_write_same_bios = 1;
1101         ti->num_write_zeroes_bios = 1;
1102         if (m->queue_mode == DM_TYPE_BIO_BASED)
1103                 ti->per_io_data_size = multipath_per_bio_data_size();
1104         else
1105                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1106
1107         return 0;
1108
1109  bad:
1110         free_multipath(m);
1111         return r;
1112 }
1113
1114 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1115 {
1116         DEFINE_WAIT(wait);
1117
1118         while (1) {
1119                 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1120
1121                 if (!atomic_read(&m->pg_init_in_progress))
1122                         break;
1123
1124                 io_schedule();
1125         }
1126         finish_wait(&m->pg_init_wait, &wait);
1127 }
1128
1129 static void flush_multipath_work(struct multipath *m)
1130 {
1131         set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1132         smp_mb__after_atomic();
1133
1134         flush_workqueue(kmpath_handlerd);
1135         multipath_wait_for_pg_init_completion(m);
1136         flush_workqueue(kmultipathd);
1137         flush_work(&m->trigger_event);
1138
1139         clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1140         smp_mb__after_atomic();
1141 }
1142
1143 static void multipath_dtr(struct dm_target *ti)
1144 {
1145         struct multipath *m = ti->private;
1146
1147         flush_multipath_work(m);
1148         free_multipath(m);
1149 }
1150
1151 /*
1152  * Take a path out of use.
1153  */
1154 static int fail_path(struct pgpath *pgpath)
1155 {
1156         unsigned long flags;
1157         struct multipath *m = pgpath->pg->m;
1158
1159         spin_lock_irqsave(&m->lock, flags);
1160
1161         if (!pgpath->is_active)
1162                 goto out;
1163
1164         DMWARN("Failing path %s.", pgpath->path.dev->name);
1165
1166         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1167         pgpath->is_active = false;
1168         pgpath->fail_count++;
1169
1170         atomic_dec(&m->nr_valid_paths);
1171
1172         if (pgpath == m->current_pgpath)
1173                 m->current_pgpath = NULL;
1174
1175         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1176                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1177
1178         schedule_work(&m->trigger_event);
1179
1180 out:
1181         spin_unlock_irqrestore(&m->lock, flags);
1182
1183         return 0;
1184 }
1185
1186 /*
1187  * Reinstate a previously-failed path
1188  */
1189 static int reinstate_path(struct pgpath *pgpath)
1190 {
1191         int r = 0, run_queue = 0;
1192         unsigned long flags;
1193         struct multipath *m = pgpath->pg->m;
1194         unsigned nr_valid_paths;
1195
1196         spin_lock_irqsave(&m->lock, flags);
1197
1198         if (pgpath->is_active)
1199                 goto out;
1200
1201         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1202
1203         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1204         if (r)
1205                 goto out;
1206
1207         pgpath->is_active = true;
1208
1209         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1210         if (nr_valid_paths == 1) {
1211                 m->current_pgpath = NULL;
1212                 run_queue = 1;
1213         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1214                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1215                         atomic_inc(&m->pg_init_in_progress);
1216         }
1217
1218         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1219                        pgpath->path.dev->name, nr_valid_paths);
1220
1221         schedule_work(&m->trigger_event);
1222
1223 out:
1224         spin_unlock_irqrestore(&m->lock, flags);
1225         if (run_queue) {
1226                 dm_table_run_md_queue_async(m->ti->table);
1227                 process_queued_io_list(m);
1228         }
1229
1230         return r;
1231 }
1232
1233 /*
1234  * Fail or reinstate all paths that match the provided struct dm_dev.
1235  */
1236 static int action_dev(struct multipath *m, struct dm_dev *dev,
1237                       action_fn action)
1238 {
1239         int r = -EINVAL;
1240         struct pgpath *pgpath;
1241         struct priority_group *pg;
1242
1243         list_for_each_entry(pg, &m->priority_groups, list) {
1244                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1245                         if (pgpath->path.dev == dev)
1246                                 r = action(pgpath);
1247                 }
1248         }
1249
1250         return r;
1251 }
1252
1253 /*
1254  * Temporarily try to avoid having to use the specified PG
1255  */
1256 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1257                       bool bypassed)
1258 {
1259         unsigned long flags;
1260
1261         spin_lock_irqsave(&m->lock, flags);
1262
1263         pg->bypassed = bypassed;
1264         m->current_pgpath = NULL;
1265         m->current_pg = NULL;
1266
1267         spin_unlock_irqrestore(&m->lock, flags);
1268
1269         schedule_work(&m->trigger_event);
1270 }
1271
1272 /*
1273  * Switch to using the specified PG from the next I/O that gets mapped
1274  */
1275 static int switch_pg_num(struct multipath *m, const char *pgstr)
1276 {
1277         struct priority_group *pg;
1278         unsigned pgnum;
1279         unsigned long flags;
1280         char dummy;
1281
1282         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1283             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1284                 DMWARN("invalid PG number supplied to switch_pg_num");
1285                 return -EINVAL;
1286         }
1287
1288         spin_lock_irqsave(&m->lock, flags);
1289         list_for_each_entry(pg, &m->priority_groups, list) {
1290                 pg->bypassed = false;
1291                 if (--pgnum)
1292                         continue;
1293
1294                 m->current_pgpath = NULL;
1295                 m->current_pg = NULL;
1296                 m->next_pg = pg;
1297         }
1298         spin_unlock_irqrestore(&m->lock, flags);
1299
1300         schedule_work(&m->trigger_event);
1301         return 0;
1302 }
1303
1304 /*
1305  * Set/clear bypassed status of a PG.
1306  * PGs are numbered upwards from 1 in the order they were declared.
1307  */
1308 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1309 {
1310         struct priority_group *pg;
1311         unsigned pgnum;
1312         char dummy;
1313
1314         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1315             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1316                 DMWARN("invalid PG number supplied to bypass_pg");
1317                 return -EINVAL;
1318         }
1319
1320         list_for_each_entry(pg, &m->priority_groups, list) {
1321                 if (!--pgnum)
1322                         break;
1323         }
1324
1325         bypass_pg(m, pg, bypassed);
1326         return 0;
1327 }
1328
1329 /*
1330  * Should we retry pg_init immediately?
1331  */
1332 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1333 {
1334         unsigned long flags;
1335         bool limit_reached = false;
1336
1337         spin_lock_irqsave(&m->lock, flags);
1338
1339         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1340             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1341                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1342         else
1343                 limit_reached = true;
1344
1345         spin_unlock_irqrestore(&m->lock, flags);
1346
1347         return limit_reached;
1348 }
1349
1350 static void pg_init_done(void *data, int errors)
1351 {
1352         struct pgpath *pgpath = data;
1353         struct priority_group *pg = pgpath->pg;
1354         struct multipath *m = pg->m;
1355         unsigned long flags;
1356         bool delay_retry = false;
1357
1358         /* device or driver problems */
1359         switch (errors) {
1360         case SCSI_DH_OK:
1361                 break;
1362         case SCSI_DH_NOSYS:
1363                 if (!m->hw_handler_name) {
1364                         errors = 0;
1365                         break;
1366                 }
1367                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1368                       "Error %d.", m->hw_handler_name, errors);
1369                 /*
1370                  * Fail path for now, so we do not ping pong
1371                  */
1372                 fail_path(pgpath);
1373                 break;
1374         case SCSI_DH_DEV_TEMP_BUSY:
1375                 /*
1376                  * Probably doing something like FW upgrade on the
1377                  * controller so try the other pg.
1378                  */
1379                 bypass_pg(m, pg, true);
1380                 break;
1381         case SCSI_DH_RETRY:
1382                 /* Wait before retrying. */
1383                 delay_retry = 1;
1384                 /* fall through */
1385         case SCSI_DH_IMM_RETRY:
1386         case SCSI_DH_RES_TEMP_UNAVAIL:
1387                 if (pg_init_limit_reached(m, pgpath))
1388                         fail_path(pgpath);
1389                 errors = 0;
1390                 break;
1391         case SCSI_DH_DEV_OFFLINED:
1392         default:
1393                 /*
1394                  * We probably do not want to fail the path for a device
1395                  * error, but this is what the old dm did. In future
1396                  * patches we can do more advanced handling.
1397                  */
1398                 fail_path(pgpath);
1399         }
1400
1401         spin_lock_irqsave(&m->lock, flags);
1402         if (errors) {
1403                 if (pgpath == m->current_pgpath) {
1404                         DMERR("Could not failover device. Error %d.", errors);
1405                         m->current_pgpath = NULL;
1406                         m->current_pg = NULL;
1407                 }
1408         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1409                 pg->bypassed = false;
1410
1411         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1412                 /* Activations of other paths are still on going */
1413                 goto out;
1414
1415         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1416                 if (delay_retry)
1417                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1418                 else
1419                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1420
1421                 if (__pg_init_all_paths(m))
1422                         goto out;
1423         }
1424         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1425
1426         process_queued_io_list(m);
1427
1428         /*
1429          * Wake up any thread waiting to suspend.
1430          */
1431         wake_up(&m->pg_init_wait);
1432
1433 out:
1434         spin_unlock_irqrestore(&m->lock, flags);
1435 }
1436
1437 static void activate_or_offline_path(struct pgpath *pgpath)
1438 {
1439         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1440
1441         if (pgpath->is_active && !blk_queue_dying(q))
1442                 scsi_dh_activate(q, pg_init_done, pgpath);
1443         else
1444                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1445 }
1446
1447 static void activate_path_work(struct work_struct *work)
1448 {
1449         struct pgpath *pgpath =
1450                 container_of(work, struct pgpath, activate_path.work);
1451
1452         activate_or_offline_path(pgpath);
1453 }
1454
1455 static int noretry_error(blk_status_t error)
1456 {
1457         switch (error) {
1458         case BLK_STS_NOTSUPP:
1459         case BLK_STS_NOSPC:
1460         case BLK_STS_TARGET:
1461         case BLK_STS_NEXUS:
1462         case BLK_STS_MEDIUM:
1463                 return 1;
1464         }
1465
1466         /* Anything else could be a path failure, so should be retried */
1467         return 0;
1468 }
1469
1470 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1471                             blk_status_t error, union map_info *map_context)
1472 {
1473         struct dm_mpath_io *mpio = get_mpio(map_context);
1474         struct pgpath *pgpath = mpio->pgpath;
1475         int r = DM_ENDIO_DONE;
1476
1477         /*
1478          * We don't queue any clone request inside the multipath target
1479          * during end I/O handling, since those clone requests don't have
1480          * bio clones.  If we queue them inside the multipath target,
1481          * we need to make bio clones, that requires memory allocation.
1482          * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1483          *  don't have bio clones.)
1484          * Instead of queueing the clone request here, we queue the original
1485          * request into dm core, which will remake a clone request and
1486          * clone bios for it and resubmit it later.
1487          */
1488         if (error && !noretry_error(error)) {
1489                 struct multipath *m = ti->private;
1490
1491                 r = DM_ENDIO_REQUEUE;
1492
1493                 if (pgpath)
1494                         fail_path(pgpath);
1495
1496                 if (atomic_read(&m->nr_valid_paths) == 0 &&
1497                     !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1498                         if (error == BLK_STS_IOERR)
1499                                 dm_report_EIO(m);
1500                         /* complete with the original error */
1501                         r = DM_ENDIO_DONE;
1502                 }
1503         }
1504
1505         if (pgpath) {
1506                 struct path_selector *ps = &pgpath->pg->ps;
1507
1508                 if (ps->type->end_io)
1509                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1510         }
1511
1512         return r;
1513 }
1514
1515 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1516                 blk_status_t *error)
1517 {
1518         struct multipath *m = ti->private;
1519         struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1520         struct pgpath *pgpath = mpio->pgpath;
1521         unsigned long flags;
1522         int r = DM_ENDIO_DONE;
1523
1524         if (!*error || noretry_error(*error))
1525                 goto done;
1526
1527         if (pgpath)
1528                 fail_path(pgpath);
1529
1530         if (atomic_read(&m->nr_valid_paths) == 0 &&
1531             !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1532                 dm_report_EIO(m);
1533                 *error = BLK_STS_IOERR;
1534                 goto done;
1535         }
1536
1537         /* Queue for the daemon to resubmit */
1538         dm_bio_restore(get_bio_details_from_bio(clone), clone);
1539
1540         spin_lock_irqsave(&m->lock, flags);
1541         bio_list_add(&m->queued_bios, clone);
1542         spin_unlock_irqrestore(&m->lock, flags);
1543         if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1544                 queue_work(kmultipathd, &m->process_queued_bios);
1545
1546         r = DM_ENDIO_INCOMPLETE;
1547 done:
1548         if (pgpath) {
1549                 struct path_selector *ps = &pgpath->pg->ps;
1550
1551                 if (ps->type->end_io)
1552                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1553         }
1554
1555         return r;
1556 }
1557
1558 /*
1559  * Suspend can't complete until all the I/O is processed so if
1560  * the last path fails we must error any remaining I/O.
1561  * Note that if the freeze_bdev fails while suspending, the
1562  * queue_if_no_path state is lost - userspace should reset it.
1563  */
1564 static void multipath_presuspend(struct dm_target *ti)
1565 {
1566         struct multipath *m = ti->private;
1567
1568         queue_if_no_path(m, false, true);
1569 }
1570
1571 static void multipath_postsuspend(struct dm_target *ti)
1572 {
1573         struct multipath *m = ti->private;
1574
1575         mutex_lock(&m->work_mutex);
1576         flush_multipath_work(m);
1577         mutex_unlock(&m->work_mutex);
1578 }
1579
1580 /*
1581  * Restore the queue_if_no_path setting.
1582  */
1583 static void multipath_resume(struct dm_target *ti)
1584 {
1585         struct multipath *m = ti->private;
1586         unsigned long flags;
1587
1588         spin_lock_irqsave(&m->lock, flags);
1589         assign_bit(test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
1590                    MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1591         spin_unlock_irqrestore(&m->lock, flags);
1592 }
1593
1594 /*
1595  * Info output has the following format:
1596  * num_multipath_feature_args [multipath_feature_args]*
1597  * num_handler_status_args [handler_status_args]*
1598  * num_groups init_group_number
1599  *            [A|D|E num_ps_status_args [ps_status_args]*
1600  *             num_paths num_selector_args
1601  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1602  *
1603  * Table output has the following format (identical to the constructor string):
1604  * num_feature_args [features_args]*
1605  * num_handler_args hw_handler [hw_handler_args]*
1606  * num_groups init_group_number
1607  *     [priority selector-name num_ps_args [ps_args]*
1608  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1609  */
1610 static void multipath_status(struct dm_target *ti, status_type_t type,
1611                              unsigned status_flags, char *result, unsigned maxlen)
1612 {
1613         int sz = 0;
1614         unsigned long flags;
1615         struct multipath *m = ti->private;
1616         struct priority_group *pg;
1617         struct pgpath *p;
1618         unsigned pg_num;
1619         char state;
1620
1621         spin_lock_irqsave(&m->lock, flags);
1622
1623         /* Features */
1624         if (type == STATUSTYPE_INFO)
1625                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1626                        atomic_read(&m->pg_init_count));
1627         else {
1628                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1629                               (m->pg_init_retries > 0) * 2 +
1630                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1631                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1632                               (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1633
1634                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1635                         DMEMIT("queue_if_no_path ");
1636                 if (m->pg_init_retries)
1637                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1638                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1639                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1640                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1641                         DMEMIT("retain_attached_hw_handler ");
1642                 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1643                         switch(m->queue_mode) {
1644                         case DM_TYPE_BIO_BASED:
1645                                 DMEMIT("queue_mode bio ");
1646                                 break;
1647                         case DM_TYPE_MQ_REQUEST_BASED:
1648                                 DMEMIT("queue_mode mq ");
1649                                 break;
1650                         default:
1651                                 WARN_ON_ONCE(true);
1652                                 break;
1653                         }
1654                 }
1655         }
1656
1657         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1658                 DMEMIT("0 ");
1659         else
1660                 DMEMIT("1 %s ", m->hw_handler_name);
1661
1662         DMEMIT("%u ", m->nr_priority_groups);
1663
1664         if (m->next_pg)
1665                 pg_num = m->next_pg->pg_num;
1666         else if (m->current_pg)
1667                 pg_num = m->current_pg->pg_num;
1668         else
1669                 pg_num = (m->nr_priority_groups ? 1 : 0);
1670
1671         DMEMIT("%u ", pg_num);
1672
1673         switch (type) {
1674         case STATUSTYPE_INFO:
1675                 list_for_each_entry(pg, &m->priority_groups, list) {
1676                         if (pg->bypassed)
1677                                 state = 'D';    /* Disabled */
1678                         else if (pg == m->current_pg)
1679                                 state = 'A';    /* Currently Active */
1680                         else
1681                                 state = 'E';    /* Enabled */
1682
1683                         DMEMIT("%c ", state);
1684
1685                         if (pg->ps.type->status)
1686                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1687                                                           result + sz,
1688                                                           maxlen - sz);
1689                         else
1690                                 DMEMIT("0 ");
1691
1692                         DMEMIT("%u %u ", pg->nr_pgpaths,
1693                                pg->ps.type->info_args);
1694
1695                         list_for_each_entry(p, &pg->pgpaths, list) {
1696                                 DMEMIT("%s %s %u ", p->path.dev->name,
1697                                        p->is_active ? "A" : "F",
1698                                        p->fail_count);
1699                                 if (pg->ps.type->status)
1700                                         sz += pg->ps.type->status(&pg->ps,
1701                                               &p->path, type, result + sz,
1702                                               maxlen - sz);
1703                         }
1704                 }
1705                 break;
1706
1707         case STATUSTYPE_TABLE:
1708                 list_for_each_entry(pg, &m->priority_groups, list) {
1709                         DMEMIT("%s ", pg->ps.type->name);
1710
1711                         if (pg->ps.type->status)
1712                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1713                                                           result + sz,
1714                                                           maxlen - sz);
1715                         else
1716                                 DMEMIT("0 ");
1717
1718                         DMEMIT("%u %u ", pg->nr_pgpaths,
1719                                pg->ps.type->table_args);
1720
1721                         list_for_each_entry(p, &pg->pgpaths, list) {
1722                                 DMEMIT("%s ", p->path.dev->name);
1723                                 if (pg->ps.type->status)
1724                                         sz += pg->ps.type->status(&pg->ps,
1725                                               &p->path, type, result + sz,
1726                                               maxlen - sz);
1727                         }
1728                 }
1729                 break;
1730         }
1731
1732         spin_unlock_irqrestore(&m->lock, flags);
1733 }
1734
1735 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1736 {
1737         int r = -EINVAL;
1738         struct dm_dev *dev;
1739         struct multipath *m = ti->private;
1740         action_fn action;
1741
1742         mutex_lock(&m->work_mutex);
1743
1744         if (dm_suspended(ti)) {
1745                 r = -EBUSY;
1746                 goto out;
1747         }
1748
1749         if (argc == 1) {
1750                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1751                         r = queue_if_no_path(m, true, false);
1752                         goto out;
1753                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1754                         r = queue_if_no_path(m, false, false);
1755                         goto out;
1756                 }
1757         }
1758
1759         if (argc != 2) {
1760                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1761                 goto out;
1762         }
1763
1764         if (!strcasecmp(argv[0], "disable_group")) {
1765                 r = bypass_pg_num(m, argv[1], true);
1766                 goto out;
1767         } else if (!strcasecmp(argv[0], "enable_group")) {
1768                 r = bypass_pg_num(m, argv[1], false);
1769                 goto out;
1770         } else if (!strcasecmp(argv[0], "switch_group")) {
1771                 r = switch_pg_num(m, argv[1]);
1772                 goto out;
1773         } else if (!strcasecmp(argv[0], "reinstate_path"))
1774                 action = reinstate_path;
1775         else if (!strcasecmp(argv[0], "fail_path"))
1776                 action = fail_path;
1777         else {
1778                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1779                 goto out;
1780         }
1781
1782         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1783         if (r) {
1784                 DMWARN("message: error getting device %s",
1785                        argv[1]);
1786                 goto out;
1787         }
1788
1789         r = action_dev(m, dev, action);
1790
1791         dm_put_device(ti, dev);
1792
1793 out:
1794         mutex_unlock(&m->work_mutex);
1795         return r;
1796 }
1797
1798 static int multipath_prepare_ioctl(struct dm_target *ti,
1799                 struct block_device **bdev, fmode_t *mode)
1800 {
1801         struct multipath *m = ti->private;
1802         struct pgpath *current_pgpath;
1803         int r;
1804
1805         current_pgpath = READ_ONCE(m->current_pgpath);
1806         if (!current_pgpath)
1807                 current_pgpath = choose_pgpath(m, 0);
1808
1809         if (current_pgpath) {
1810                 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1811                         *bdev = current_pgpath->path.dev->bdev;
1812                         *mode = current_pgpath->path.dev->mode;
1813                         r = 0;
1814                 } else {
1815                         /* pg_init has not started or completed */
1816                         r = -ENOTCONN;
1817                 }
1818         } else {
1819                 /* No path is available */
1820                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1821                         r = -ENOTCONN;
1822                 else
1823                         r = -EIO;
1824         }
1825
1826         if (r == -ENOTCONN) {
1827                 if (!READ_ONCE(m->current_pg)) {
1828                         /* Path status changed, redo selection */
1829                         (void) choose_pgpath(m, 0);
1830                 }
1831                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1832                         pg_init_all_paths(m);
1833                 dm_table_run_md_queue_async(m->ti->table);
1834                 process_queued_io_list(m);
1835         }
1836
1837         /*
1838          * Only pass ioctls through if the device sizes match exactly.
1839          */
1840         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1841                 return 1;
1842         return r;
1843 }
1844
1845 static int multipath_iterate_devices(struct dm_target *ti,
1846                                      iterate_devices_callout_fn fn, void *data)
1847 {
1848         struct multipath *m = ti->private;
1849         struct priority_group *pg;
1850         struct pgpath *p;
1851         int ret = 0;
1852
1853         list_for_each_entry(pg, &m->priority_groups, list) {
1854                 list_for_each_entry(p, &pg->pgpaths, list) {
1855                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1856                         if (ret)
1857                                 goto out;
1858                 }
1859         }
1860
1861 out:
1862         return ret;
1863 }
1864
1865 static int pgpath_busy(struct pgpath *pgpath)
1866 {
1867         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1868
1869         return blk_lld_busy(q);
1870 }
1871
1872 /*
1873  * We return "busy", only when we can map I/Os but underlying devices
1874  * are busy (so even if we map I/Os now, the I/Os will wait on
1875  * the underlying queue).
1876  * In other words, if we want to kill I/Os or queue them inside us
1877  * due to map unavailability, we don't return "busy".  Otherwise,
1878  * dm core won't give us the I/Os and we can't do what we want.
1879  */
1880 static int multipath_busy(struct dm_target *ti)
1881 {
1882         bool busy = false, has_active = false;
1883         struct multipath *m = ti->private;
1884         struct priority_group *pg, *next_pg;
1885         struct pgpath *pgpath;
1886
1887         /* pg_init in progress */
1888         if (atomic_read(&m->pg_init_in_progress))
1889                 return true;
1890
1891         /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1892         if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1893                 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1894
1895         /* Guess which priority_group will be used at next mapping time */
1896         pg = READ_ONCE(m->current_pg);
1897         next_pg = READ_ONCE(m->next_pg);
1898         if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1899                 pg = next_pg;
1900
1901         if (!pg) {
1902                 /*
1903                  * We don't know which pg will be used at next mapping time.
1904                  * We don't call choose_pgpath() here to avoid to trigger
1905                  * pg_init just by busy checking.
1906                  * So we don't know whether underlying devices we will be using
1907                  * at next mapping time are busy or not. Just try mapping.
1908                  */
1909                 return busy;
1910         }
1911
1912         /*
1913          * If there is one non-busy active path at least, the path selector
1914          * will be able to select it. So we consider such a pg as not busy.
1915          */
1916         busy = true;
1917         list_for_each_entry(pgpath, &pg->pgpaths, list) {
1918                 if (pgpath->is_active) {
1919                         has_active = true;
1920                         if (!pgpath_busy(pgpath)) {
1921                                 busy = false;
1922                                 break;
1923                         }
1924                 }
1925         }
1926
1927         if (!has_active) {
1928                 /*
1929                  * No active path in this pg, so this pg won't be used and
1930                  * the current_pg will be changed at next mapping time.
1931                  * We need to try mapping to determine it.
1932                  */
1933                 busy = false;
1934         }
1935
1936         return busy;
1937 }
1938
1939 /*-----------------------------------------------------------------
1940  * Module setup
1941  *---------------------------------------------------------------*/
1942 static struct target_type multipath_target = {
1943         .name = "multipath",
1944         .version = {1, 12, 0},
1945         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1946         .module = THIS_MODULE,
1947         .ctr = multipath_ctr,
1948         .dtr = multipath_dtr,
1949         .clone_and_map_rq = multipath_clone_and_map,
1950         .release_clone_rq = multipath_release_clone,
1951         .rq_end_io = multipath_end_io,
1952         .map = multipath_map_bio,
1953         .end_io = multipath_end_io_bio,
1954         .presuspend = multipath_presuspend,
1955         .postsuspend = multipath_postsuspend,
1956         .resume = multipath_resume,
1957         .status = multipath_status,
1958         .message = multipath_message,
1959         .prepare_ioctl = multipath_prepare_ioctl,
1960         .iterate_devices = multipath_iterate_devices,
1961         .busy = multipath_busy,
1962 };
1963
1964 static int __init dm_multipath_init(void)
1965 {
1966         int r;
1967
1968         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1969         if (!kmultipathd) {
1970                 DMERR("failed to create workqueue kmpathd");
1971                 r = -ENOMEM;
1972                 goto bad_alloc_kmultipathd;
1973         }
1974
1975         /*
1976          * A separate workqueue is used to handle the device handlers
1977          * to avoid overloading existing workqueue. Overloading the
1978          * old workqueue would also create a bottleneck in the
1979          * path of the storage hardware device activation.
1980          */
1981         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1982                                                   WQ_MEM_RECLAIM);
1983         if (!kmpath_handlerd) {
1984                 DMERR("failed to create workqueue kmpath_handlerd");
1985                 r = -ENOMEM;
1986                 goto bad_alloc_kmpath_handlerd;
1987         }
1988
1989         r = dm_register_target(&multipath_target);
1990         if (r < 0) {
1991                 DMERR("request-based register failed %d", r);
1992                 r = -EINVAL;
1993                 goto bad_register_target;
1994         }
1995
1996         return 0;
1997
1998 bad_register_target:
1999         destroy_workqueue(kmpath_handlerd);
2000 bad_alloc_kmpath_handlerd:
2001         destroy_workqueue(kmultipathd);
2002 bad_alloc_kmultipathd:
2003         return r;
2004 }
2005
2006 static void __exit dm_multipath_exit(void)
2007 {
2008         destroy_workqueue(kmpath_handlerd);
2009         destroy_workqueue(kmultipathd);
2010
2011         dm_unregister_target(&multipath_target);
2012 }
2013
2014 module_init(dm_multipath_init);
2015 module_exit(dm_multipath_exit);
2016
2017 MODULE_DESCRIPTION(DM_NAME " multipath target");
2018 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2019 MODULE_LICENSE("GPL");