Merge tag 'for-4.15/dm-changes-2' of git://git.kernel.org/pub/scm/linux/kernel/git...
[muen/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(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 /*
643  * If we run out of usable paths, should we queue I/O or error it?
644  */
645 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
646                             bool save_old_value)
647 {
648         unsigned long flags;
649
650         spin_lock_irqsave(&m->lock, flags);
651         assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
652                    (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
653                    (!save_old_value && queue_if_no_path));
654         assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
655                    queue_if_no_path || dm_noflush_suspending(m->ti));
656         spin_unlock_irqrestore(&m->lock, flags);
657
658         if (!queue_if_no_path) {
659                 dm_table_run_md_queue_async(m->ti->table);
660                 process_queued_io_list(m);
661         }
662
663         return 0;
664 }
665
666 /*
667  * An event is triggered whenever a path is taken out of use.
668  * Includes path failure and PG bypass.
669  */
670 static void trigger_event(struct work_struct *work)
671 {
672         struct multipath *m =
673                 container_of(work, struct multipath, trigger_event);
674
675         dm_table_event(m->ti->table);
676 }
677
678 /*-----------------------------------------------------------------
679  * Constructor/argument parsing:
680  * <#multipath feature args> [<arg>]*
681  * <#hw_handler args> [hw_handler [<arg>]*]
682  * <#priority groups>
683  * <initial priority group>
684  *     [<selector> <#selector args> [<arg>]*
685  *      <#paths> <#per-path selector args>
686  *         [<path> [<arg>]* ]+ ]+
687  *---------------------------------------------------------------*/
688 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
689                                struct dm_target *ti)
690 {
691         int r;
692         struct path_selector_type *pst;
693         unsigned ps_argc;
694
695         static const struct dm_arg _args[] = {
696                 {0, 1024, "invalid number of path selector args"},
697         };
698
699         pst = dm_get_path_selector(dm_shift_arg(as));
700         if (!pst) {
701                 ti->error = "unknown path selector type";
702                 return -EINVAL;
703         }
704
705         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
706         if (r) {
707                 dm_put_path_selector(pst);
708                 return -EINVAL;
709         }
710
711         r = pst->create(&pg->ps, ps_argc, as->argv);
712         if (r) {
713                 dm_put_path_selector(pst);
714                 ti->error = "path selector constructor failed";
715                 return r;
716         }
717
718         pg->ps.type = pst;
719         dm_consume_args(as, ps_argc);
720
721         return 0;
722 }
723
724 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
725                                struct dm_target *ti)
726 {
727         int r;
728         struct pgpath *p;
729         struct multipath *m = ti->private;
730         struct request_queue *q = NULL;
731         const char *attached_handler_name;
732
733         /* we need at least a path arg */
734         if (as->argc < 1) {
735                 ti->error = "no device given";
736                 return ERR_PTR(-EINVAL);
737         }
738
739         p = alloc_pgpath();
740         if (!p)
741                 return ERR_PTR(-ENOMEM);
742
743         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
744                           &p->path.dev);
745         if (r) {
746                 ti->error = "error getting device";
747                 goto bad;
748         }
749
750         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
751                 q = bdev_get_queue(p->path.dev->bdev);
752
753         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
754 retain:
755                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
756                 if (attached_handler_name) {
757                         /*
758                          * Clear any hw_handler_params associated with a
759                          * handler that isn't already attached.
760                          */
761                         if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
762                                 kfree(m->hw_handler_params);
763                                 m->hw_handler_params = NULL;
764                         }
765
766                         /*
767                          * Reset hw_handler_name to match the attached handler
768                          *
769                          * NB. This modifies the table line to show the actual
770                          * handler instead of the original table passed in.
771                          */
772                         kfree(m->hw_handler_name);
773                         m->hw_handler_name = attached_handler_name;
774                 }
775         }
776
777         if (m->hw_handler_name) {
778                 r = scsi_dh_attach(q, m->hw_handler_name);
779                 if (r == -EBUSY) {
780                         char b[BDEVNAME_SIZE];
781
782                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
783                                 bdevname(p->path.dev->bdev, b));
784                         goto retain;
785                 }
786                 if (r < 0) {
787                         ti->error = "error attaching hardware handler";
788                         dm_put_device(ti, p->path.dev);
789                         goto bad;
790                 }
791
792                 if (m->hw_handler_params) {
793                         r = scsi_dh_set_params(q, m->hw_handler_params);
794                         if (r < 0) {
795                                 ti->error = "unable to set hardware "
796                                                         "handler parameters";
797                                 dm_put_device(ti, p->path.dev);
798                                 goto bad;
799                         }
800                 }
801         }
802
803         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
804         if (r) {
805                 dm_put_device(ti, p->path.dev);
806                 goto bad;
807         }
808
809         return p;
810
811  bad:
812         free_pgpath(p);
813         return ERR_PTR(r);
814 }
815
816 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
817                                                    struct multipath *m)
818 {
819         static const struct dm_arg _args[] = {
820                 {1, 1024, "invalid number of paths"},
821                 {0, 1024, "invalid number of selector args"}
822         };
823
824         int r;
825         unsigned i, nr_selector_args, nr_args;
826         struct priority_group *pg;
827         struct dm_target *ti = m->ti;
828
829         if (as->argc < 2) {
830                 as->argc = 0;
831                 ti->error = "not enough priority group arguments";
832                 return ERR_PTR(-EINVAL);
833         }
834
835         pg = alloc_priority_group();
836         if (!pg) {
837                 ti->error = "couldn't allocate priority group";
838                 return ERR_PTR(-ENOMEM);
839         }
840         pg->m = m;
841
842         r = parse_path_selector(as, pg, ti);
843         if (r)
844                 goto bad;
845
846         /*
847          * read the paths
848          */
849         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
850         if (r)
851                 goto bad;
852
853         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
854         if (r)
855                 goto bad;
856
857         nr_args = 1 + nr_selector_args;
858         for (i = 0; i < pg->nr_pgpaths; i++) {
859                 struct pgpath *pgpath;
860                 struct dm_arg_set path_args;
861
862                 if (as->argc < nr_args) {
863                         ti->error = "not enough path parameters";
864                         r = -EINVAL;
865                         goto bad;
866                 }
867
868                 path_args.argc = nr_args;
869                 path_args.argv = as->argv;
870
871                 pgpath = parse_path(&path_args, &pg->ps, ti);
872                 if (IS_ERR(pgpath)) {
873                         r = PTR_ERR(pgpath);
874                         goto bad;
875                 }
876
877                 pgpath->pg = pg;
878                 list_add_tail(&pgpath->list, &pg->pgpaths);
879                 dm_consume_args(as, nr_args);
880         }
881
882         return pg;
883
884  bad:
885         free_priority_group(pg, ti);
886         return ERR_PTR(r);
887 }
888
889 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
890 {
891         unsigned hw_argc;
892         int ret;
893         struct dm_target *ti = m->ti;
894
895         static const struct dm_arg _args[] = {
896                 {0, 1024, "invalid number of hardware handler args"},
897         };
898
899         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
900                 return -EINVAL;
901
902         if (!hw_argc)
903                 return 0;
904
905         if (m->queue_mode == DM_TYPE_BIO_BASED) {
906                 dm_consume_args(as, hw_argc);
907                 DMERR("bio-based multipath doesn't allow hardware handler args");
908                 return 0;
909         }
910
911         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
912         if (!m->hw_handler_name)
913                 return -EINVAL;
914
915         if (hw_argc > 1) {
916                 char *p;
917                 int i, j, len = 4;
918
919                 for (i = 0; i <= hw_argc - 2; i++)
920                         len += strlen(as->argv[i]) + 1;
921                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
922                 if (!p) {
923                         ti->error = "memory allocation failed";
924                         ret = -ENOMEM;
925                         goto fail;
926                 }
927                 j = sprintf(p, "%d", hw_argc - 1);
928                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
929                         j = sprintf(p, "%s", as->argv[i]);
930         }
931         dm_consume_args(as, hw_argc - 1);
932
933         return 0;
934 fail:
935         kfree(m->hw_handler_name);
936         m->hw_handler_name = NULL;
937         return ret;
938 }
939
940 static int parse_features(struct dm_arg_set *as, struct multipath *m)
941 {
942         int r;
943         unsigned argc;
944         struct dm_target *ti = m->ti;
945         const char *arg_name;
946
947         static const struct dm_arg _args[] = {
948                 {0, 8, "invalid number of feature args"},
949                 {1, 50, "pg_init_retries must be between 1 and 50"},
950                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
951         };
952
953         r = dm_read_arg_group(_args, as, &argc, &ti->error);
954         if (r)
955                 return -EINVAL;
956
957         if (!argc)
958                 return 0;
959
960         do {
961                 arg_name = dm_shift_arg(as);
962                 argc--;
963
964                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
965                         r = queue_if_no_path(m, true, false);
966                         continue;
967                 }
968
969                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
970                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
971                         continue;
972                 }
973
974                 if (!strcasecmp(arg_name, "pg_init_retries") &&
975                     (argc >= 1)) {
976                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
977                         argc--;
978                         continue;
979                 }
980
981                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
982                     (argc >= 1)) {
983                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
984                         argc--;
985                         continue;
986                 }
987
988                 if (!strcasecmp(arg_name, "queue_mode") &&
989                     (argc >= 1)) {
990                         const char *queue_mode_name = dm_shift_arg(as);
991
992                         if (!strcasecmp(queue_mode_name, "bio"))
993                                 m->queue_mode = DM_TYPE_BIO_BASED;
994                         else if (!strcasecmp(queue_mode_name, "rq"))
995                                 m->queue_mode = DM_TYPE_REQUEST_BASED;
996                         else if (!strcasecmp(queue_mode_name, "mq"))
997                                 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
998                         else {
999                                 ti->error = "Unknown 'queue_mode' requested";
1000                                 r = -EINVAL;
1001                         }
1002                         argc--;
1003                         continue;
1004                 }
1005
1006                 ti->error = "Unrecognised multipath feature request";
1007                 r = -EINVAL;
1008         } while (argc && !r);
1009
1010         return r;
1011 }
1012
1013 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1014 {
1015         /* target arguments */
1016         static const struct dm_arg _args[] = {
1017                 {0, 1024, "invalid number of priority groups"},
1018                 {0, 1024, "invalid initial priority group number"},
1019         };
1020
1021         int r;
1022         struct multipath *m;
1023         struct dm_arg_set as;
1024         unsigned pg_count = 0;
1025         unsigned next_pg_num;
1026
1027         as.argc = argc;
1028         as.argv = argv;
1029
1030         m = alloc_multipath(ti);
1031         if (!m) {
1032                 ti->error = "can't allocate multipath";
1033                 return -EINVAL;
1034         }
1035
1036         r = parse_features(&as, m);
1037         if (r)
1038                 goto bad;
1039
1040         r = alloc_multipath_stage2(ti, m);
1041         if (r)
1042                 goto bad;
1043
1044         r = parse_hw_handler(&as, m);
1045         if (r)
1046                 goto bad;
1047
1048         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1049         if (r)
1050                 goto bad;
1051
1052         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1053         if (r)
1054                 goto bad;
1055
1056         if ((!m->nr_priority_groups && next_pg_num) ||
1057             (m->nr_priority_groups && !next_pg_num)) {
1058                 ti->error = "invalid initial priority group";
1059                 r = -EINVAL;
1060                 goto bad;
1061         }
1062
1063         /* parse the priority groups */
1064         while (as.argc) {
1065                 struct priority_group *pg;
1066                 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1067
1068                 pg = parse_priority_group(&as, m);
1069                 if (IS_ERR(pg)) {
1070                         r = PTR_ERR(pg);
1071                         goto bad;
1072                 }
1073
1074                 nr_valid_paths += pg->nr_pgpaths;
1075                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1076
1077                 list_add_tail(&pg->list, &m->priority_groups);
1078                 pg_count++;
1079                 pg->pg_num = pg_count;
1080                 if (!--next_pg_num)
1081                         m->next_pg = pg;
1082         }
1083
1084         if (pg_count != m->nr_priority_groups) {
1085                 ti->error = "priority group count mismatch";
1086                 r = -EINVAL;
1087                 goto bad;
1088         }
1089
1090         ti->num_flush_bios = 1;
1091         ti->num_discard_bios = 1;
1092         ti->num_write_same_bios = 1;
1093         ti->num_write_zeroes_bios = 1;
1094         if (m->queue_mode == DM_TYPE_BIO_BASED)
1095                 ti->per_io_data_size = multipath_per_bio_data_size();
1096         else
1097                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1098
1099         return 0;
1100
1101  bad:
1102         free_multipath(m);
1103         return r;
1104 }
1105
1106 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1107 {
1108         DEFINE_WAIT(wait);
1109
1110         while (1) {
1111                 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1112
1113                 if (!atomic_read(&m->pg_init_in_progress))
1114                         break;
1115
1116                 io_schedule();
1117         }
1118         finish_wait(&m->pg_init_wait, &wait);
1119 }
1120
1121 static void flush_multipath_work(struct multipath *m)
1122 {
1123         set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1124         smp_mb__after_atomic();
1125
1126         flush_workqueue(kmpath_handlerd);
1127         multipath_wait_for_pg_init_completion(m);
1128         flush_workqueue(kmultipathd);
1129         flush_work(&m->trigger_event);
1130
1131         clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1132         smp_mb__after_atomic();
1133 }
1134
1135 static void multipath_dtr(struct dm_target *ti)
1136 {
1137         struct multipath *m = ti->private;
1138
1139         flush_multipath_work(m);
1140         free_multipath(m);
1141 }
1142
1143 /*
1144  * Take a path out of use.
1145  */
1146 static int fail_path(struct pgpath *pgpath)
1147 {
1148         unsigned long flags;
1149         struct multipath *m = pgpath->pg->m;
1150
1151         spin_lock_irqsave(&m->lock, flags);
1152
1153         if (!pgpath->is_active)
1154                 goto out;
1155
1156         DMWARN("Failing path %s.", pgpath->path.dev->name);
1157
1158         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1159         pgpath->is_active = false;
1160         pgpath->fail_count++;
1161
1162         atomic_dec(&m->nr_valid_paths);
1163
1164         if (pgpath == m->current_pgpath)
1165                 m->current_pgpath = NULL;
1166
1167         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1168                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1169
1170         schedule_work(&m->trigger_event);
1171
1172 out:
1173         spin_unlock_irqrestore(&m->lock, flags);
1174
1175         return 0;
1176 }
1177
1178 /*
1179  * Reinstate a previously-failed path
1180  */
1181 static int reinstate_path(struct pgpath *pgpath)
1182 {
1183         int r = 0, run_queue = 0;
1184         unsigned long flags;
1185         struct multipath *m = pgpath->pg->m;
1186         unsigned nr_valid_paths;
1187
1188         spin_lock_irqsave(&m->lock, flags);
1189
1190         if (pgpath->is_active)
1191                 goto out;
1192
1193         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1194
1195         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1196         if (r)
1197                 goto out;
1198
1199         pgpath->is_active = true;
1200
1201         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1202         if (nr_valid_paths == 1) {
1203                 m->current_pgpath = NULL;
1204                 run_queue = 1;
1205         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1206                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1207                         atomic_inc(&m->pg_init_in_progress);
1208         }
1209
1210         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1211                        pgpath->path.dev->name, nr_valid_paths);
1212
1213         schedule_work(&m->trigger_event);
1214
1215 out:
1216         spin_unlock_irqrestore(&m->lock, flags);
1217         if (run_queue) {
1218                 dm_table_run_md_queue_async(m->ti->table);
1219                 process_queued_io_list(m);
1220         }
1221
1222         return r;
1223 }
1224
1225 /*
1226  * Fail or reinstate all paths that match the provided struct dm_dev.
1227  */
1228 static int action_dev(struct multipath *m, struct dm_dev *dev,
1229                       action_fn action)
1230 {
1231         int r = -EINVAL;
1232         struct pgpath *pgpath;
1233         struct priority_group *pg;
1234
1235         list_for_each_entry(pg, &m->priority_groups, list) {
1236                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1237                         if (pgpath->path.dev == dev)
1238                                 r = action(pgpath);
1239                 }
1240         }
1241
1242         return r;
1243 }
1244
1245 /*
1246  * Temporarily try to avoid having to use the specified PG
1247  */
1248 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1249                       bool bypassed)
1250 {
1251         unsigned long flags;
1252
1253         spin_lock_irqsave(&m->lock, flags);
1254
1255         pg->bypassed = bypassed;
1256         m->current_pgpath = NULL;
1257         m->current_pg = NULL;
1258
1259         spin_unlock_irqrestore(&m->lock, flags);
1260
1261         schedule_work(&m->trigger_event);
1262 }
1263
1264 /*
1265  * Switch to using the specified PG from the next I/O that gets mapped
1266  */
1267 static int switch_pg_num(struct multipath *m, const char *pgstr)
1268 {
1269         struct priority_group *pg;
1270         unsigned pgnum;
1271         unsigned long flags;
1272         char dummy;
1273
1274         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1275             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1276                 DMWARN("invalid PG number supplied to switch_pg_num");
1277                 return -EINVAL;
1278         }
1279
1280         spin_lock_irqsave(&m->lock, flags);
1281         list_for_each_entry(pg, &m->priority_groups, list) {
1282                 pg->bypassed = false;
1283                 if (--pgnum)
1284                         continue;
1285
1286                 m->current_pgpath = NULL;
1287                 m->current_pg = NULL;
1288                 m->next_pg = pg;
1289         }
1290         spin_unlock_irqrestore(&m->lock, flags);
1291
1292         schedule_work(&m->trigger_event);
1293         return 0;
1294 }
1295
1296 /*
1297  * Set/clear bypassed status of a PG.
1298  * PGs are numbered upwards from 1 in the order they were declared.
1299  */
1300 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1301 {
1302         struct priority_group *pg;
1303         unsigned pgnum;
1304         char dummy;
1305
1306         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1307             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1308                 DMWARN("invalid PG number supplied to bypass_pg");
1309                 return -EINVAL;
1310         }
1311
1312         list_for_each_entry(pg, &m->priority_groups, list) {
1313                 if (!--pgnum)
1314                         break;
1315         }
1316
1317         bypass_pg(m, pg, bypassed);
1318         return 0;
1319 }
1320
1321 /*
1322  * Should we retry pg_init immediately?
1323  */
1324 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1325 {
1326         unsigned long flags;
1327         bool limit_reached = false;
1328
1329         spin_lock_irqsave(&m->lock, flags);
1330
1331         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1332             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1333                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1334         else
1335                 limit_reached = true;
1336
1337         spin_unlock_irqrestore(&m->lock, flags);
1338
1339         return limit_reached;
1340 }
1341
1342 static void pg_init_done(void *data, int errors)
1343 {
1344         struct pgpath *pgpath = data;
1345         struct priority_group *pg = pgpath->pg;
1346         struct multipath *m = pg->m;
1347         unsigned long flags;
1348         bool delay_retry = false;
1349
1350         /* device or driver problems */
1351         switch (errors) {
1352         case SCSI_DH_OK:
1353                 break;
1354         case SCSI_DH_NOSYS:
1355                 if (!m->hw_handler_name) {
1356                         errors = 0;
1357                         break;
1358                 }
1359                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1360                       "Error %d.", m->hw_handler_name, errors);
1361                 /*
1362                  * Fail path for now, so we do not ping pong
1363                  */
1364                 fail_path(pgpath);
1365                 break;
1366         case SCSI_DH_DEV_TEMP_BUSY:
1367                 /*
1368                  * Probably doing something like FW upgrade on the
1369                  * controller so try the other pg.
1370                  */
1371                 bypass_pg(m, pg, true);
1372                 break;
1373         case SCSI_DH_RETRY:
1374                 /* Wait before retrying. */
1375                 delay_retry = 1;
1376                 /* fall through */
1377         case SCSI_DH_IMM_RETRY:
1378         case SCSI_DH_RES_TEMP_UNAVAIL:
1379                 if (pg_init_limit_reached(m, pgpath))
1380                         fail_path(pgpath);
1381                 errors = 0;
1382                 break;
1383         case SCSI_DH_DEV_OFFLINED:
1384         default:
1385                 /*
1386                  * We probably do not want to fail the path for a device
1387                  * error, but this is what the old dm did. In future
1388                  * patches we can do more advanced handling.
1389                  */
1390                 fail_path(pgpath);
1391         }
1392
1393         spin_lock_irqsave(&m->lock, flags);
1394         if (errors) {
1395                 if (pgpath == m->current_pgpath) {
1396                         DMERR("Could not failover device. Error %d.", errors);
1397                         m->current_pgpath = NULL;
1398                         m->current_pg = NULL;
1399                 }
1400         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1401                 pg->bypassed = false;
1402
1403         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1404                 /* Activations of other paths are still on going */
1405                 goto out;
1406
1407         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1408                 if (delay_retry)
1409                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1410                 else
1411                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1412
1413                 if (__pg_init_all_paths(m))
1414                         goto out;
1415         }
1416         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1417
1418         process_queued_io_list(m);
1419
1420         /*
1421          * Wake up any thread waiting to suspend.
1422          */
1423         wake_up(&m->pg_init_wait);
1424
1425 out:
1426         spin_unlock_irqrestore(&m->lock, flags);
1427 }
1428
1429 static void activate_or_offline_path(struct pgpath *pgpath)
1430 {
1431         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1432
1433         if (pgpath->is_active && !blk_queue_dying(q))
1434                 scsi_dh_activate(q, pg_init_done, pgpath);
1435         else
1436                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1437 }
1438
1439 static void activate_path_work(struct work_struct *work)
1440 {
1441         struct pgpath *pgpath =
1442                 container_of(work, struct pgpath, activate_path.work);
1443
1444         activate_or_offline_path(pgpath);
1445 }
1446
1447 static int noretry_error(blk_status_t error)
1448 {
1449         switch (error) {
1450         case BLK_STS_NOTSUPP:
1451         case BLK_STS_NOSPC:
1452         case BLK_STS_TARGET:
1453         case BLK_STS_NEXUS:
1454         case BLK_STS_MEDIUM:
1455                 return 1;
1456         }
1457
1458         /* Anything else could be a path failure, so should be retried */
1459         return 0;
1460 }
1461
1462 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1463                             blk_status_t error, union map_info *map_context)
1464 {
1465         struct dm_mpath_io *mpio = get_mpio(map_context);
1466         struct pgpath *pgpath = mpio->pgpath;
1467         int r = DM_ENDIO_DONE;
1468
1469         /*
1470          * We don't queue any clone request inside the multipath target
1471          * during end I/O handling, since those clone requests don't have
1472          * bio clones.  If we queue them inside the multipath target,
1473          * we need to make bio clones, that requires memory allocation.
1474          * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1475          *  don't have bio clones.)
1476          * Instead of queueing the clone request here, we queue the original
1477          * request into dm core, which will remake a clone request and
1478          * clone bios for it and resubmit it later.
1479          */
1480         if (error && !noretry_error(error)) {
1481                 struct multipath *m = ti->private;
1482
1483                 r = DM_ENDIO_REQUEUE;
1484
1485                 if (pgpath)
1486                         fail_path(pgpath);
1487
1488                 if (atomic_read(&m->nr_valid_paths) == 0 &&
1489                     !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1490                         if (error == BLK_STS_IOERR)
1491                                 dm_report_EIO(m);
1492                         /* complete with the original error */
1493                         r = DM_ENDIO_DONE;
1494                 }
1495         }
1496
1497         if (pgpath) {
1498                 struct path_selector *ps = &pgpath->pg->ps;
1499
1500                 if (ps->type->end_io)
1501                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1502         }
1503
1504         return r;
1505 }
1506
1507 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1508                 blk_status_t *error)
1509 {
1510         struct multipath *m = ti->private;
1511         struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1512         struct pgpath *pgpath = mpio->pgpath;
1513         unsigned long flags;
1514         int r = DM_ENDIO_DONE;
1515
1516         if (!*error || noretry_error(*error))
1517                 goto done;
1518
1519         if (pgpath)
1520                 fail_path(pgpath);
1521
1522         if (atomic_read(&m->nr_valid_paths) == 0 &&
1523             !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1524                 dm_report_EIO(m);
1525                 *error = BLK_STS_IOERR;
1526                 goto done;
1527         }
1528
1529         /* Queue for the daemon to resubmit */
1530         dm_bio_restore(get_bio_details_from_bio(clone), clone);
1531
1532         spin_lock_irqsave(&m->lock, flags);
1533         bio_list_add(&m->queued_bios, clone);
1534         spin_unlock_irqrestore(&m->lock, flags);
1535         if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1536                 queue_work(kmultipathd, &m->process_queued_bios);
1537
1538         r = DM_ENDIO_INCOMPLETE;
1539 done:
1540         if (pgpath) {
1541                 struct path_selector *ps = &pgpath->pg->ps;
1542
1543                 if (ps->type->end_io)
1544                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1545         }
1546
1547         return r;
1548 }
1549
1550 /*
1551  * Suspend can't complete until all the I/O is processed so if
1552  * the last path fails we must error any remaining I/O.
1553  * Note that if the freeze_bdev fails while suspending, the
1554  * queue_if_no_path state is lost - userspace should reset it.
1555  */
1556 static void multipath_presuspend(struct dm_target *ti)
1557 {
1558         struct multipath *m = ti->private;
1559
1560         queue_if_no_path(m, false, true);
1561 }
1562
1563 static void multipath_postsuspend(struct dm_target *ti)
1564 {
1565         struct multipath *m = ti->private;
1566
1567         mutex_lock(&m->work_mutex);
1568         flush_multipath_work(m);
1569         mutex_unlock(&m->work_mutex);
1570 }
1571
1572 /*
1573  * Restore the queue_if_no_path setting.
1574  */
1575 static void multipath_resume(struct dm_target *ti)
1576 {
1577         struct multipath *m = ti->private;
1578         unsigned long flags;
1579
1580         spin_lock_irqsave(&m->lock, flags);
1581         assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1582                    test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1583         spin_unlock_irqrestore(&m->lock, flags);
1584 }
1585
1586 /*
1587  * Info output has the following format:
1588  * num_multipath_feature_args [multipath_feature_args]*
1589  * num_handler_status_args [handler_status_args]*
1590  * num_groups init_group_number
1591  *            [A|D|E num_ps_status_args [ps_status_args]*
1592  *             num_paths num_selector_args
1593  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1594  *
1595  * Table output has the following format (identical to the constructor string):
1596  * num_feature_args [features_args]*
1597  * num_handler_args hw_handler [hw_handler_args]*
1598  * num_groups init_group_number
1599  *     [priority selector-name num_ps_args [ps_args]*
1600  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1601  */
1602 static void multipath_status(struct dm_target *ti, status_type_t type,
1603                              unsigned status_flags, char *result, unsigned maxlen)
1604 {
1605         int sz = 0;
1606         unsigned long flags;
1607         struct multipath *m = ti->private;
1608         struct priority_group *pg;
1609         struct pgpath *p;
1610         unsigned pg_num;
1611         char state;
1612
1613         spin_lock_irqsave(&m->lock, flags);
1614
1615         /* Features */
1616         if (type == STATUSTYPE_INFO)
1617                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1618                        atomic_read(&m->pg_init_count));
1619         else {
1620                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1621                               (m->pg_init_retries > 0) * 2 +
1622                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1623                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1624                               (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1625
1626                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1627                         DMEMIT("queue_if_no_path ");
1628                 if (m->pg_init_retries)
1629                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1630                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1631                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1632                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1633                         DMEMIT("retain_attached_hw_handler ");
1634                 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1635                         switch(m->queue_mode) {
1636                         case DM_TYPE_BIO_BASED:
1637                                 DMEMIT("queue_mode bio ");
1638                                 break;
1639                         case DM_TYPE_MQ_REQUEST_BASED:
1640                                 DMEMIT("queue_mode mq ");
1641                                 break;
1642                         default:
1643                                 WARN_ON_ONCE(true);
1644                                 break;
1645                         }
1646                 }
1647         }
1648
1649         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1650                 DMEMIT("0 ");
1651         else
1652                 DMEMIT("1 %s ", m->hw_handler_name);
1653
1654         DMEMIT("%u ", m->nr_priority_groups);
1655
1656         if (m->next_pg)
1657                 pg_num = m->next_pg->pg_num;
1658         else if (m->current_pg)
1659                 pg_num = m->current_pg->pg_num;
1660         else
1661                 pg_num = (m->nr_priority_groups ? 1 : 0);
1662
1663         DMEMIT("%u ", pg_num);
1664
1665         switch (type) {
1666         case STATUSTYPE_INFO:
1667                 list_for_each_entry(pg, &m->priority_groups, list) {
1668                         if (pg->bypassed)
1669                                 state = 'D';    /* Disabled */
1670                         else if (pg == m->current_pg)
1671                                 state = 'A';    /* Currently Active */
1672                         else
1673                                 state = 'E';    /* Enabled */
1674
1675                         DMEMIT("%c ", state);
1676
1677                         if (pg->ps.type->status)
1678                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1679                                                           result + sz,
1680                                                           maxlen - sz);
1681                         else
1682                                 DMEMIT("0 ");
1683
1684                         DMEMIT("%u %u ", pg->nr_pgpaths,
1685                                pg->ps.type->info_args);
1686
1687                         list_for_each_entry(p, &pg->pgpaths, list) {
1688                                 DMEMIT("%s %s %u ", p->path.dev->name,
1689                                        p->is_active ? "A" : "F",
1690                                        p->fail_count);
1691                                 if (pg->ps.type->status)
1692                                         sz += pg->ps.type->status(&pg->ps,
1693                                               &p->path, type, result + sz,
1694                                               maxlen - sz);
1695                         }
1696                 }
1697                 break;
1698
1699         case STATUSTYPE_TABLE:
1700                 list_for_each_entry(pg, &m->priority_groups, list) {
1701                         DMEMIT("%s ", pg->ps.type->name);
1702
1703                         if (pg->ps.type->status)
1704                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1705                                                           result + sz,
1706                                                           maxlen - sz);
1707                         else
1708                                 DMEMIT("0 ");
1709
1710                         DMEMIT("%u %u ", pg->nr_pgpaths,
1711                                pg->ps.type->table_args);
1712
1713                         list_for_each_entry(p, &pg->pgpaths, list) {
1714                                 DMEMIT("%s ", p->path.dev->name);
1715                                 if (pg->ps.type->status)
1716                                         sz += pg->ps.type->status(&pg->ps,
1717                                               &p->path, type, result + sz,
1718                                               maxlen - sz);
1719                         }
1720                 }
1721                 break;
1722         }
1723
1724         spin_unlock_irqrestore(&m->lock, flags);
1725 }
1726
1727 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1728 {
1729         int r = -EINVAL;
1730         struct dm_dev *dev;
1731         struct multipath *m = ti->private;
1732         action_fn action;
1733
1734         mutex_lock(&m->work_mutex);
1735
1736         if (dm_suspended(ti)) {
1737                 r = -EBUSY;
1738                 goto out;
1739         }
1740
1741         if (argc == 1) {
1742                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1743                         r = queue_if_no_path(m, true, false);
1744                         goto out;
1745                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1746                         r = queue_if_no_path(m, false, false);
1747                         goto out;
1748                 }
1749         }
1750
1751         if (argc != 2) {
1752                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1753                 goto out;
1754         }
1755
1756         if (!strcasecmp(argv[0], "disable_group")) {
1757                 r = bypass_pg_num(m, argv[1], true);
1758                 goto out;
1759         } else if (!strcasecmp(argv[0], "enable_group")) {
1760                 r = bypass_pg_num(m, argv[1], false);
1761                 goto out;
1762         } else if (!strcasecmp(argv[0], "switch_group")) {
1763                 r = switch_pg_num(m, argv[1]);
1764                 goto out;
1765         } else if (!strcasecmp(argv[0], "reinstate_path"))
1766                 action = reinstate_path;
1767         else if (!strcasecmp(argv[0], "fail_path"))
1768                 action = fail_path;
1769         else {
1770                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1771                 goto out;
1772         }
1773
1774         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1775         if (r) {
1776                 DMWARN("message: error getting device %s",
1777                        argv[1]);
1778                 goto out;
1779         }
1780
1781         r = action_dev(m, dev, action);
1782
1783         dm_put_device(ti, dev);
1784
1785 out:
1786         mutex_unlock(&m->work_mutex);
1787         return r;
1788 }
1789
1790 static int multipath_prepare_ioctl(struct dm_target *ti,
1791                 struct block_device **bdev, fmode_t *mode)
1792 {
1793         struct multipath *m = ti->private;
1794         struct pgpath *current_pgpath;
1795         int r;
1796
1797         current_pgpath = READ_ONCE(m->current_pgpath);
1798         if (!current_pgpath)
1799                 current_pgpath = choose_pgpath(m, 0);
1800
1801         if (current_pgpath) {
1802                 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1803                         *bdev = current_pgpath->path.dev->bdev;
1804                         *mode = current_pgpath->path.dev->mode;
1805                         r = 0;
1806                 } else {
1807                         /* pg_init has not started or completed */
1808                         r = -ENOTCONN;
1809                 }
1810         } else {
1811                 /* No path is available */
1812                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1813                         r = -ENOTCONN;
1814                 else
1815                         r = -EIO;
1816         }
1817
1818         if (r == -ENOTCONN) {
1819                 if (!READ_ONCE(m->current_pg)) {
1820                         /* Path status changed, redo selection */
1821                         (void) choose_pgpath(m, 0);
1822                 }
1823                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1824                         pg_init_all_paths(m);
1825                 dm_table_run_md_queue_async(m->ti->table);
1826                 process_queued_io_list(m);
1827         }
1828
1829         /*
1830          * Only pass ioctls through if the device sizes match exactly.
1831          */
1832         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1833                 return 1;
1834         return r;
1835 }
1836
1837 static int multipath_iterate_devices(struct dm_target *ti,
1838                                      iterate_devices_callout_fn fn, void *data)
1839 {
1840         struct multipath *m = ti->private;
1841         struct priority_group *pg;
1842         struct pgpath *p;
1843         int ret = 0;
1844
1845         list_for_each_entry(pg, &m->priority_groups, list) {
1846                 list_for_each_entry(p, &pg->pgpaths, list) {
1847                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1848                         if (ret)
1849                                 goto out;
1850                 }
1851         }
1852
1853 out:
1854         return ret;
1855 }
1856
1857 static int pgpath_busy(struct pgpath *pgpath)
1858 {
1859         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1860
1861         return blk_lld_busy(q);
1862 }
1863
1864 /*
1865  * We return "busy", only when we can map I/Os but underlying devices
1866  * are busy (so even if we map I/Os now, the I/Os will wait on
1867  * the underlying queue).
1868  * In other words, if we want to kill I/Os or queue them inside us
1869  * due to map unavailability, we don't return "busy".  Otherwise,
1870  * dm core won't give us the I/Os and we can't do what we want.
1871  */
1872 static int multipath_busy(struct dm_target *ti)
1873 {
1874         bool busy = false, has_active = false;
1875         struct multipath *m = ti->private;
1876         struct priority_group *pg, *next_pg;
1877         struct pgpath *pgpath;
1878
1879         /* pg_init in progress */
1880         if (atomic_read(&m->pg_init_in_progress))
1881                 return true;
1882
1883         /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1884         if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1885                 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1886
1887         /* Guess which priority_group will be used at next mapping time */
1888         pg = READ_ONCE(m->current_pg);
1889         next_pg = READ_ONCE(m->next_pg);
1890         if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1891                 pg = next_pg;
1892
1893         if (!pg) {
1894                 /*
1895                  * We don't know which pg will be used at next mapping time.
1896                  * We don't call choose_pgpath() here to avoid to trigger
1897                  * pg_init just by busy checking.
1898                  * So we don't know whether underlying devices we will be using
1899                  * at next mapping time are busy or not. Just try mapping.
1900                  */
1901                 return busy;
1902         }
1903
1904         /*
1905          * If there is one non-busy active path at least, the path selector
1906          * will be able to select it. So we consider such a pg as not busy.
1907          */
1908         busy = true;
1909         list_for_each_entry(pgpath, &pg->pgpaths, list) {
1910                 if (pgpath->is_active) {
1911                         has_active = true;
1912                         if (!pgpath_busy(pgpath)) {
1913                                 busy = false;
1914                                 break;
1915                         }
1916                 }
1917         }
1918
1919         if (!has_active) {
1920                 /*
1921                  * No active path in this pg, so this pg won't be used and
1922                  * the current_pg will be changed at next mapping time.
1923                  * We need to try mapping to determine it.
1924                  */
1925                 busy = false;
1926         }
1927
1928         return busy;
1929 }
1930
1931 /*-----------------------------------------------------------------
1932  * Module setup
1933  *---------------------------------------------------------------*/
1934 static struct target_type multipath_target = {
1935         .name = "multipath",
1936         .version = {1, 12, 0},
1937         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1938         .module = THIS_MODULE,
1939         .ctr = multipath_ctr,
1940         .dtr = multipath_dtr,
1941         .clone_and_map_rq = multipath_clone_and_map,
1942         .release_clone_rq = multipath_release_clone,
1943         .rq_end_io = multipath_end_io,
1944         .map = multipath_map_bio,
1945         .end_io = multipath_end_io_bio,
1946         .presuspend = multipath_presuspend,
1947         .postsuspend = multipath_postsuspend,
1948         .resume = multipath_resume,
1949         .status = multipath_status,
1950         .message = multipath_message,
1951         .prepare_ioctl = multipath_prepare_ioctl,
1952         .iterate_devices = multipath_iterate_devices,
1953         .busy = multipath_busy,
1954 };
1955
1956 static int __init dm_multipath_init(void)
1957 {
1958         int r;
1959
1960         r = dm_register_target(&multipath_target);
1961         if (r < 0) {
1962                 DMERR("request-based register failed %d", r);
1963                 r = -EINVAL;
1964                 goto bad_register_target;
1965         }
1966
1967         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1968         if (!kmultipathd) {
1969                 DMERR("failed to create workqueue kmpathd");
1970                 r = -ENOMEM;
1971                 goto bad_alloc_kmultipathd;
1972         }
1973
1974         /*
1975          * A separate workqueue is used to handle the device handlers
1976          * to avoid overloading existing workqueue. Overloading the
1977          * old workqueue would also create a bottleneck in the
1978          * path of the storage hardware device activation.
1979          */
1980         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1981                                                   WQ_MEM_RECLAIM);
1982         if (!kmpath_handlerd) {
1983                 DMERR("failed to create workqueue kmpath_handlerd");
1984                 r = -ENOMEM;
1985                 goto bad_alloc_kmpath_handlerd;
1986         }
1987
1988         return 0;
1989
1990 bad_alloc_kmpath_handlerd:
1991         destroy_workqueue(kmultipathd);
1992 bad_alloc_kmultipathd:
1993         dm_unregister_target(&multipath_target);
1994 bad_register_target:
1995         return r;
1996 }
1997
1998 static void __exit dm_multipath_exit(void)
1999 {
2000         destroy_workqueue(kmpath_handlerd);
2001         destroy_workqueue(kmultipathd);
2002
2003         dm_unregister_target(&multipath_target);
2004 }
2005
2006 module_init(dm_multipath_init);
2007 module_exit(dm_multipath_exit);
2008
2009 MODULE_DESCRIPTION(DM_NAME " multipath target");
2010 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2011 MODULE_LICENSE("GPL");