drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9
  10 #include "dm-rq.h"
  11 #include "dm-bio-record.h"
  12 #include "dm-path-selector.h"
  13 #include "dm-uevent.h"
  14
  15 #include <linux/blkdev.h>
  16 #include <linux/ctype.h>
  17 #include <linux/init.h>
  18 #include <linux/mempool.h>
  19 #include <linux/module.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/slab.h>
  22 #include <linux/time.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/delay.h>
  25 #include <scsi/scsi_dh.h>
  26 #include <linux/atomic.h>
  27 #include <linux/blk-mq.h>
  28
  29 #define DM_MSG_PREFIX "multipath"
  30 #define DM_PG_INIT_DELAY_MSECS 2000
  31 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  32
  33 /* Path properties */
  34 struct pgpath {
  35         struct list_head list;
  36
  37         struct priority_group *pg;      /* Owning PG */
  38         unsigned fail_count;            /* Cumulative failure count */
  39
  40         struct dm_path path;
  41         struct delayed_work activate_path;
  42
  43         bool is_active:1;               /* Path status */
  44 };
  45
  46 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  47
  48 /*
  49  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  50  * Each has a path selector which controls which path gets used.
  51  */
  52 struct priority_group {
  53         struct list_head list;
  54
  55         struct multipath *m;            /* Owning multipath instance */
  56         struct path_selector ps;
  57
  58         unsigned pg_num;                /* Reference number */
  59         unsigned nr_pgpaths;            /* Number of paths in PG */
  60         struct list_head pgpaths;
  61
  62         bool bypassed:1;                /* Temporarily bypass this PG? */
  63 };
  64
  65 /* Multipath context */
  66 struct multipath {
  67         struct list_head list;
  68         struct dm_target *ti;
  69
  70         const char *hw_handler_name;
  71         char *hw_handler_params;
  72
  73         spinlock_t lock;
  74
  75         unsigned nr_priority_groups;
  76         struct list_head priority_groups;
  77
  78         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  79
  80         struct pgpath *current_pgpath;
  81         struct priority_group *current_pg;
  82         struct priority_group *next_pg; /* Switch to this PG if set */
  83
  84         unsigned long flags;            /* Multipath state flags */
  85
  86         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  87         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  88
  89         atomic_t nr_valid_paths;        /* Total number of usable paths */
  90         atomic_t pg_init_in_progress;   /* Only one pg_init allowed at once */
  91         atomic_t pg_init_count;         /* Number of times pg_init called */
  92
  93         enum dm_queue_mode queue_mode;
  94
  95         struct mutex work_mutex;
  96         struct work_struct trigger_event;
  97
  98         struct work_struct process_queued_bios;
  99         struct bio_list queued_bios;
 100 };
 101
 102 /*
 103  * Context information attached to each io we process.
 104  */
 105 struct dm_mpath_io {
 106         struct pgpath *pgpath;
 107         size_t nr_bytes;
 108 };
 109
 110 typedef int (*action_fn) (struct pgpath *pgpath);
 111
 112 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 113 static void trigger_event(struct work_struct *work);
 114 static void activate_or_offline_path(struct pgpath *pgpath);
 115 static void activate_path_work(struct work_struct *work);
 116 static void process_queued_bios(struct work_struct *work);
 117
 118 /*-----------------------------------------------
 119  * Multipath state flags.
 120  *-----------------------------------------------*/
 121
 122 #define MPATHF_QUEUE_IO 0                       /* Must we queue all I/O? */
 123 #define MPATHF_QUEUE_IF_NO_PATH 1               /* Queue I/O if last path fails? */
 124 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2         /* Saved state during suspension */
 125 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3     /* If there's already a hw_handler present, don't change it. */
 126 #define MPATHF_PG_INIT_DISABLED 4               /* pg_init is not currently allowed */
 127 #define MPATHF_PG_INIT_REQUIRED 5               /* pg_init needs calling? */
 128 #define MPATHF_PG_INIT_DELAY_RETRY 6            /* Delay pg_init retry? */
 129
 130 /*-----------------------------------------------
 131  * Allocation routines
 132  *-----------------------------------------------*/
 133
 134 static struct pgpath *alloc_pgpath(void)
 135 {
 136         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 137
 138         if (pgpath) {
 139                 pgpath->is_active = true;
 140                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path_work);
 141         }
 142
 143         return pgpath;
 144 }
 145
 146 static void free_pgpath(struct pgpath *pgpath)
 147 {
 148         kfree(pgpath);
 149 }
 150
 151 static struct priority_group *alloc_priority_group(void)
 152 {
 153         struct priority_group *pg;
 154
 155         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 156
 157         if (pg)
 158                 INIT_LIST_HEAD(&pg->pgpaths);
 159
 160         return pg;
 161 }
 162
 163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 164 {
 165         struct pgpath *pgpath, *tmp;
 166
 167         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 168                 list_del(&pgpath->list);
 169                 dm_put_device(ti, pgpath->path.dev);
 170                 free_pgpath(pgpath);
 171         }
 172 }
 173
 174 static void free_priority_group(struct priority_group *pg,
 175                                 struct dm_target *ti)
 176 {
 177         struct path_selector *ps = &pg->ps;
 178
 179         if (ps->type) {
 180                 ps->type->destroy(ps);
 181                 dm_put_path_selector(ps->type);
 182         }
 183
 184         free_pgpaths(&pg->pgpaths, ti);
 185         kfree(pg);
 186 }
 187
 188 static struct multipath *alloc_multipath(struct dm_target *ti)
 189 {
 190         struct multipath *m;
 191
 192         m = kzalloc(sizeof(*m), GFP_KERNEL);
 193         if (m) {
 194                 INIT_LIST_HEAD(&m->priority_groups);
 195                 spin_lock_init(&m->lock);
 196                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 197                 atomic_set(&m->nr_valid_paths, 0);
 198                 atomic_set(&m->pg_init_in_progress, 0);
 199                 atomic_set(&m->pg_init_count, 0);
 200                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 201                 INIT_WORK(&m->trigger_event, trigger_event);
 202                 init_waitqueue_head(&m->pg_init_wait);
 203                 mutex_init(&m->work_mutex);
 204
 205                 m->queue_mode = DM_TYPE_NONE;
 206
 207                 m->ti = ti;
 208                 ti->private = m;
 209         }
 210
 211         return m;
 212 }
 213
 214 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
 215 {
 216         if (m->queue_mode == DM_TYPE_NONE) {
 217                 /*
 218                  * Default to request-based.
 219                  */
 220                 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
 221                         m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
 222                 else
 223                         m->queue_mode = DM_TYPE_REQUEST_BASED;
 224         } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
 225                 INIT_WORK(&m->process_queued_bios, process_queued_bios);
 226                 /*
 227                  * bio-based doesn't support any direct scsi_dh management;
 228                  * it just discovers if a scsi_dh is attached.
 229                  */
 230                 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 231         }
 232
 233         dm_table_set_type(ti->table, m->queue_mode);
 234
 235         return 0;
 236 }
 237
 238 static void free_multipath(struct multipath *m)
 239 {
 240         struct priority_group *pg, *tmp;
 241
 242         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 243                 list_del(&pg->list);
 244                 free_priority_group(pg, m->ti);
 245         }
 246
 247         kfree(m->hw_handler_name);
 248         kfree(m->hw_handler_params);
 249         kfree(m);
 250 }
 251
 252 static struct dm_mpath_io *get_mpio(union map_info *info)
 253 {
 254         return info->ptr;
 255 }
 256
 257 static size_t multipath_per_bio_data_size(void)
 258 {
 259         return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
 260 }
 261
 262 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
 263 {
 264         return dm_per_bio_data(bio, multipath_per_bio_data_size());
 265 }
 266
 267 static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
 268 {
 269         /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
 270         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 271         void *bio_details = mpio + 1;
 272
 273         return bio_details;
 274 }
 275
 276 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p,
 277                                         struct dm_bio_details **bio_details_p)
 278 {
 279         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 280         struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
 281
 282         memset(mpio, 0, sizeof(*mpio));
 283         memset(bio_details, 0, sizeof(*bio_details));
 284         dm_bio_record(bio_details, bio);
 285
 286         if (mpio_p)
 287                 *mpio_p = mpio;
 288         if (bio_details_p)
 289                 *bio_details_p = bio_details;
 290 }
 291
 292 /*-----------------------------------------------
 293  * Path selection
 294  *-----------------------------------------------*/
 295
 296 static int __pg_init_all_paths(struct multipath *m)
 297 {
 298         struct pgpath *pgpath;
 299         unsigned long pg_init_delay = 0;
 300
 301         lockdep_assert_held(&m->lock);
 302
 303         if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
 304                 return 0;
 305
 306         atomic_inc(&m->pg_init_count);
 307         clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 308
 309         /* Check here to reset pg_init_required */
 310         if (!m->current_pg)
 311                 return 0;
 312
 313         if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
 314                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 315                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 316         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 317                 /* Skip failed paths */
 318                 if (!pgpath->is_active)
 319                         continue;
 320                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 321                                        pg_init_delay))
 322                         atomic_inc(&m->pg_init_in_progress);
 323         }
 324         return atomic_read(&m->pg_init_in_progress);
 325 }
 326
 327 static int pg_init_all_paths(struct multipath *m)
 328 {
 329         int ret;
 330         unsigned long flags;
 331
 332         spin_lock_irqsave(&m->lock, flags);
 333         ret = __pg_init_all_paths(m);
 334         spin_unlock_irqrestore(&m->lock, flags);
 335
 336         return ret;
 337 }
 338
 339 static void __switch_pg(struct multipath *m, struct priority_group *pg)
 340 {
 341         m->current_pg = pg;
 342
 343         /* Must we initialise the PG first, and queue I/O till it's ready? */
 344         if (m->hw_handler_name) {
 345                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 346                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 347         } else {
 348                 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 349                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 350         }
 351
 352         atomic_set(&m->pg_init_count, 0);
 353 }
 354
 355 static struct pgpath *choose_path_in_pg(struct multipath *m,
 356                                         struct priority_group *pg,
 357                                         size_t nr_bytes)
 358 {
 359         unsigned long flags;
 360         struct dm_path *path;
 361         struct pgpath *pgpath;
 362
 363         path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 364         if (!path)
 365                 return ERR_PTR(-ENXIO);
 366
 367         pgpath = path_to_pgpath(path);
 368
 369         if (unlikely(lockless_dereference(m->current_pg) != pg)) {
 370                 /* Only update current_pgpath if pg changed */
 371                 spin_lock_irqsave(&m->lock, flags);
 372                 m->current_pgpath = pgpath;
 373                 __switch_pg(m, pg);
 374                 spin_unlock_irqrestore(&m->lock, flags);
 375         }
 376
 377         return pgpath;
 378 }
 379
 380 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
 381 {
 382         unsigned long flags;
 383         struct priority_group *pg;
 384         struct pgpath *pgpath;
 385         unsigned bypassed = 1;
 386
 387         if (!atomic_read(&m->nr_valid_paths)) {
 388                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 389                 goto failed;
 390         }
 391
 392         /* Were we instructed to switch PG? */
 393         if (lockless_dereference(m->next_pg)) {
 394                 spin_lock_irqsave(&m->lock, flags);
 395                 pg = m->next_pg;
 396                 if (!pg) {
 397                         spin_unlock_irqrestore(&m->lock, flags);
 398                         goto check_current_pg;
 399                 }
 400                 m->next_pg = NULL;
 401                 spin_unlock_irqrestore(&m->lock, flags);
 402                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 403                 if (!IS_ERR_OR_NULL(pgpath))
 404                         return pgpath;
 405         }
 406
 407         /* Don't change PG until it has no remaining paths */
 408 check_current_pg:
 409         pg = lockless_dereference(m->current_pg);
 410         if (pg) {
 411                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 412                 if (!IS_ERR_OR_NULL(pgpath))
 413                         return pgpath;
 414         }
 415
 416         /*
 417          * Loop through priority groups until we find a valid path.
 418          * First time we skip PGs marked 'bypassed'.
 419          * Second time we only try the ones we skipped, but set
 420          * pg_init_delay_retry so we do not hammer controllers.
 421          */
 422         do {
 423                 list_for_each_entry(pg, &m->priority_groups, list) {
 424                         if (pg->bypassed == !!bypassed)
 425                                 continue;
 426                         pgpath = choose_path_in_pg(m, pg, nr_bytes);
 427                         if (!IS_ERR_OR_NULL(pgpath)) {
 428                                 if (!bypassed)
 429                                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
 430                                 return pgpath;
 431                         }
 432                 }
 433         } while (bypassed--);
 434
 435 failed:
 436         spin_lock_irqsave(&m->lock, flags);
 437         m->current_pgpath = NULL;
 438         m->current_pg = NULL;
 439         spin_unlock_irqrestore(&m->lock, flags);
 440
 441         return NULL;
 442 }
 443
 444 /*
 445  * dm_report_EIO() is a macro instead of a function to make pr_debug()
 446  * report the function name and line number of the function from which
 447  * it has been invoked.
 448  */
 449 #define dm_report_EIO(m)                                                \
 450 ({                                                                      \
 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         -EIO;                                                           \
 459 })
 460
 461 /*
 462  * Map cloned requests (request-based multipath)
 463  */
 464 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 465                                    union map_info *map_context,
 466                                    struct request **__clone)
 467 {
 468         struct multipath *m = ti->private;
 469         size_t nr_bytes = blk_rq_bytes(rq);
 470         struct pgpath *pgpath;
 471         struct block_device *bdev;
 472         struct dm_mpath_io *mpio = get_mpio(map_context);
 473         struct request_queue *q;
 474         struct request *clone;
 475
 476         /* Do we need to select a new pgpath? */
 477         pgpath = lockless_dereference(m->current_pgpath);
 478         if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
 479                 pgpath = choose_pgpath(m, nr_bytes);
 480
 481         if (!pgpath) {
 482                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
 483                         return DM_MAPIO_DELAY_REQUEUE;
 484                 return dm_report_EIO(m);        /* Failed */
 485         } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
 486                    test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
 487                 if (pg_init_all_paths(m))
 488                         return DM_MAPIO_DELAY_REQUEUE;
 489                 return DM_MAPIO_REQUEUE;
 490         }
 491
 492         memset(mpio, 0, sizeof(*mpio));
 493         mpio->pgpath = pgpath;
 494         mpio->nr_bytes = nr_bytes;
 495
 496         bdev = pgpath->path.dev->bdev;
 497         q = bdev_get_queue(bdev);
 498         clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
 499         if (IS_ERR(clone)) {
 500                 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
 501                 bool queue_dying = blk_queue_dying(q);
 502                 DMERR_LIMIT("blk_get_request() returned %ld%s - requeuing",
 503                             PTR_ERR(clone), queue_dying ? " (path offline)" : "");
 504                 if (queue_dying) {
 505                         atomic_inc(&m->pg_init_in_progress);
 506                         activate_or_offline_path(pgpath);
 507                         return DM_MAPIO_REQUEUE;
 508                 }
 509                 return DM_MAPIO_DELAY_REQUEUE;
 510         }
 511         clone->bio = clone->biotail = NULL;
 512         clone->rq_disk = bdev->bd_disk;
 513         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 514         *__clone = clone;
 515
 516         if (pgpath->pg->ps.type->start_io)
 517                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 518                                               &pgpath->path,
 519                                               nr_bytes);
 520         return DM_MAPIO_REMAPPED;
 521 }
 522
 523 static void multipath_release_clone(struct request *clone)
 524 {
 525         blk_put_request(clone);
 526 }
 527
 528 /*
 529  * Map cloned bios (bio-based multipath)
 530  */
 531 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
 532 {
 533         size_t nr_bytes = bio->bi_iter.bi_size;
 534         struct pgpath *pgpath;
 535         unsigned long flags;
 536         bool queue_io;
 537
 538         /* Do we need to select a new pgpath? */
 539         pgpath = lockless_dereference(m->current_pgpath);
 540         queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
 541         if (!pgpath || !queue_io)
 542                 pgpath = choose_pgpath(m, nr_bytes);
 543
 544         if ((pgpath && queue_io) ||
 545             (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
 546                 /* Queue for the daemon to resubmit */
 547                 spin_lock_irqsave(&m->lock, flags);
 548                 bio_list_add(&m->queued_bios, bio);
 549                 spin_unlock_irqrestore(&m->lock, flags);
 550                 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
 551                 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
 552                         pg_init_all_paths(m);
 553                 else if (!queue_io)
 554                         queue_work(kmultipathd, &m->process_queued_bios);
 555                 return DM_MAPIO_SUBMITTED;
 556         }
 557
 558         if (!pgpath) {
 559                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
 560                         return DM_MAPIO_REQUEUE;
 561                 return dm_report_EIO(m);
 562         }
 563
 564         mpio->pgpath = pgpath;
 565         mpio->nr_bytes = nr_bytes;
 566
 567         bio->bi_error = 0;
 568         bio->bi_bdev = pgpath->path.dev->bdev;
 569         bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
 570
 571         if (pgpath->pg->ps.type->start_io)
 572                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 573                                               &pgpath->path,
 574                                               nr_bytes);
 575         return DM_MAPIO_REMAPPED;
 576 }
 577
 578 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
 579 {
 580         struct multipath *m = ti->private;
 581         struct dm_mpath_io *mpio = NULL;
 582
 583         multipath_init_per_bio_data(bio, &mpio, NULL);
 584
 585         return __multipath_map_bio(m, bio, mpio);
 586 }
 587
 588 static void process_queued_io_list(struct multipath *m)
 589 {
 590         if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
 591                 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
 592         else if (m->queue_mode == DM_TYPE_BIO_BASED)
 593                 queue_work(kmultipathd, &m->process_queued_bios);
 594 }
 595
 596 static void process_queued_bios(struct work_struct *work)
 597 {
 598         int r;
 599         unsigned long flags;
 600         struct bio *bio;
 601         struct bio_list bios;
 602         struct blk_plug plug;
 603         struct multipath *m =
 604                 container_of(work, struct multipath, process_queued_bios);
 605
 606         bio_list_init(&bios);
 607
 608         spin_lock_irqsave(&m->lock, flags);
 609
 610         if (bio_list_empty(&m->queued_bios)) {
 611                 spin_unlock_irqrestore(&m->lock, flags);
 612                 return;
 613         }
 614
 615         bio_list_merge(&bios, &m->queued_bios);
 616         bio_list_init(&m->queued_bios);
 617
 618         spin_unlock_irqrestore(&m->lock, flags);
 619
 620         blk_start_plug(&plug);
 621         while ((bio = bio_list_pop(&bios))) {
 622                 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
 623                 if (r < 0 || r == DM_MAPIO_REQUEUE) {
 624                         bio->bi_error = r;
 625                         bio_endio(bio);
 626                 } else if (r == DM_MAPIO_REMAPPED)
 627                         generic_make_request(bio);
 628         }
 629         blk_finish_plug(&plug);
 630 }
 631
 632 static void assign_bit(bool value, long nr, unsigned long *addr)
 633 {
 634         if (value)
 635                 set_bit(nr, addr);
 636         else
 637                 clear_bit(nr, addr);
 638 }
 639
 640 /*
 641  * If we run out of usable paths, should we queue I/O or error it?
 642  */
 643 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
 644                             bool save_old_value)
 645 {
 646         unsigned long flags;
 647
 648         spin_lock_irqsave(&m->lock, flags);
 649         assign_bit((save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
 650                    (!save_old_value && queue_if_no_path),
 651                    MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 652         assign_bit(queue_if_no_path || dm_noflush_suspending(m->ti),
 653                    MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 654         spin_unlock_irqrestore(&m->lock, flags);
 655
 656         if (!queue_if_no_path) {
 657                 dm_table_run_md_queue_async(m->ti->table);
 658                 process_queued_io_list(m);
 659         }
 660
 661         return 0;
 662 }
 663
 664 /*
 665  * An event is triggered whenever a path is taken out of use.
 666  * Includes path failure and PG bypass.
 667  */
 668 static void trigger_event(struct work_struct *work)
 669 {
 670         struct multipath *m =
 671                 container_of(work, struct multipath, trigger_event);
 672
 673         dm_table_event(m->ti->table);
 674 }
 675
 676 /*-----------------------------------------------------------------
 677  * Constructor/argument parsing:
 678  * <#multipath feature args> [<arg>]*
 679  * <#hw_handler args> [hw_handler [<arg>]*]
 680  * <#priority groups>
 681  * <initial priority group>
 682  *     [<selector> <#selector args> [<arg>]*
 683  *      <#paths> <#per-path selector args>
 684  *         [<path> [<arg>]* ]+ ]+
 685  *---------------------------------------------------------------*/
 686 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 687                                struct dm_target *ti)
 688 {
 689         int r;
 690         struct path_selector_type *pst;
 691         unsigned ps_argc;
 692
 693         static struct dm_arg _args[] = {
 694                 {0, 1024, "invalid number of path selector args"},
 695         };
 696
 697         pst = dm_get_path_selector(dm_shift_arg(as));
 698         if (!pst) {
 699                 ti->error = "unknown path selector type";
 700                 return -EINVAL;
 701         }
 702
 703         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 704         if (r) {
 705                 dm_put_path_selector(pst);
 706                 return -EINVAL;
 707         }
 708
 709         r = pst->create(&pg->ps, ps_argc, as->argv);
 710         if (r) {
 711                 dm_put_path_selector(pst);
 712                 ti->error = "path selector constructor failed";
 713                 return r;
 714         }
 715
 716         pg->ps.type = pst;
 717         dm_consume_args(as, ps_argc);
 718
 719         return 0;
 720 }
 721
 722 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 723                                struct dm_target *ti)
 724 {
 725         int r;
 726         struct pgpath *p;
 727         struct multipath *m = ti->private;
 728         struct request_queue *q = NULL;
 729         const char *attached_handler_name;
 730
 731         /* we need at least a path arg */
 732         if (as->argc < 1) {
 733                 ti->error = "no device given";
 734                 return ERR_PTR(-EINVAL);
 735         }
 736
 737         p = alloc_pgpath();
 738         if (!p)
 739                 return ERR_PTR(-ENOMEM);
 740
 741         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 742                           &p->path.dev);
 743         if (r) {
 744                 ti->error = "error getting device";
 745                 goto bad;
 746         }
 747
 748         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
 749                 q = bdev_get_queue(p->path.dev->bdev);
 750
 751         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
 752 retain:
 753                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 754                 if (attached_handler_name) {
 755                         /*
 756                          * Clear any hw_handler_params associated with a
 757                          * handler that isn't already attached.
 758                          */
 759                         if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
 760                                 kfree(m->hw_handler_params);
 761                                 m->hw_handler_params = NULL;
 762                         }
 763
 764                         /*
 765                          * Reset hw_handler_name to match the attached handler
 766                          *
 767                          * NB. This modifies the table line to show the actual
 768                          * handler instead of the original table passed in.
 769                          */
 770                         kfree(m->hw_handler_name);
 771                         m->hw_handler_name = attached_handler_name;
 772                 }
 773         }
 774
 775         if (m->hw_handler_name) {
 776                 r = scsi_dh_attach(q, m->hw_handler_name);
 777                 if (r == -EBUSY) {
 778                         char b[BDEVNAME_SIZE];
 779
 780                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
 781                                 bdevname(p->path.dev->bdev, b));
 782                         goto retain;
 783                 }
 784                 if (r < 0) {
 785                         ti->error = "error attaching hardware handler";
 786                         dm_put_device(ti, p->path.dev);
 787                         goto bad;
 788                 }
 789
 790                 if (m->hw_handler_params) {
 791                         r = scsi_dh_set_params(q, m->hw_handler_params);
 792                         if (r < 0) {
 793                                 ti->error = "unable to set hardware "
 794                                                         "handler parameters";
 795                                 dm_put_device(ti, p->path.dev);
 796                                 goto bad;
 797                         }
 798                 }
 799         }
 800
 801         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 802         if (r) {
 803                 dm_put_device(ti, p->path.dev);
 804                 goto bad;
 805         }
 806
 807         return p;
 808
 809  bad:
 810         free_pgpath(p);
 811         return ERR_PTR(r);
 812 }
 813
 814 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 815                                                    struct multipath *m)
 816 {
 817         static struct dm_arg _args[] = {
 818                 {1, 1024, "invalid number of paths"},
 819                 {0, 1024, "invalid number of selector args"}
 820         };
 821
 822         int r;
 823         unsigned i, nr_selector_args, nr_args;
 824         struct priority_group *pg;
 825         struct dm_target *ti = m->ti;
 826
 827         if (as->argc < 2) {
 828                 as->argc = 0;
 829                 ti->error = "not enough priority group arguments";
 830                 return ERR_PTR(-EINVAL);
 831         }
 832
 833         pg = alloc_priority_group();
 834         if (!pg) {
 835                 ti->error = "couldn't allocate priority group";
 836                 return ERR_PTR(-ENOMEM);
 837         }
 838         pg->m = m;
 839
 840         r = parse_path_selector(as, pg, ti);
 841         if (r)
 842                 goto bad;
 843
 844         /*
 845          * read the paths
 846          */
 847         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 848         if (r)
 849                 goto bad;
 850
 851         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 852         if (r)
 853                 goto bad;
 854
 855         nr_args = 1 + nr_selector_args;
 856         for (i = 0; i < pg->nr_pgpaths; i++) {
 857                 struct pgpath *pgpath;
 858                 struct dm_arg_set path_args;
 859
 860                 if (as->argc < nr_args) {
 861                         ti->error = "not enough path parameters";
 862                         r = -EINVAL;
 863                         goto bad;
 864                 }
 865
 866                 path_args.argc = nr_args;
 867                 path_args.argv = as->argv;
 868
 869                 pgpath = parse_path(&path_args, &pg->ps, ti);
 870                 if (IS_ERR(pgpath)) {
 871                         r = PTR_ERR(pgpath);
 872                         goto bad;
 873                 }
 874
 875                 pgpath->pg = pg;
 876                 list_add_tail(&pgpath->list, &pg->pgpaths);
 877                 dm_consume_args(as, nr_args);
 878         }
 879
 880         return pg;
 881
 882  bad:
 883         free_priority_group(pg, ti);
 884         return ERR_PTR(r);
 885 }
 886
 887 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 888 {
 889         unsigned hw_argc;
 890         int ret;
 891         struct dm_target *ti = m->ti;
 892
 893         static struct dm_arg _args[] = {
 894                 {0, 1024, "invalid number of hardware handler args"},
 895         };
 896
 897         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 898                 return -EINVAL;
 899
 900         if (!hw_argc)
 901                 return 0;
 902
 903         if (m->queue_mode == DM_TYPE_BIO_BASED) {
 904                 dm_consume_args(as, hw_argc);
 905                 DMERR("bio-based multipath doesn't allow hardware handler args");
 906                 return 0;
 907         }
 908
 909         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 910         if (!m->hw_handler_name)
 911                 return -EINVAL;
 912
 913         if (hw_argc > 1) {
 914                 char *p;
 915                 int i, j, len = 4;
 916
 917                 for (i = 0; i <= hw_argc - 2; i++)
 918                         len += strlen(as->argv[i]) + 1;
 919                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 920                 if (!p) {
 921                         ti->error = "memory allocation failed";
 922                         ret = -ENOMEM;
 923                         goto fail;
 924                 }
 925                 j = sprintf(p, "%d", hw_argc - 1);
 926                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 927                         j = sprintf(p, "%s", as->argv[i]);
 928         }
 929         dm_consume_args(as, hw_argc - 1);
 930
 931         return 0;
 932 fail:
 933         kfree(m->hw_handler_name);
 934         m->hw_handler_name = NULL;
 935         return ret;
 936 }
 937
 938 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 939 {
 940         int r;
 941         unsigned argc;
 942         struct dm_target *ti = m->ti;
 943         const char *arg_name;
 944
 945         static struct dm_arg _args[] = {
 946                 {0, 8, "invalid number of feature args"},
 947                 {1, 50, "pg_init_retries must be between 1 and 50"},
 948                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 949         };
 950
 951         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 952         if (r)
 953                 return -EINVAL;
 954
 955         if (!argc)
 956                 return 0;
 957
 958         do {
 959                 arg_name = dm_shift_arg(as);
 960                 argc--;
 961
 962                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 963                         r = queue_if_no_path(m, true, false);
 964                         continue;
 965                 }
 966
 967                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
 968                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 969                         continue;
 970                 }
 971
 972                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 973                     (argc >= 1)) {
 974                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 975                         argc--;
 976                         continue;
 977                 }
 978
 979                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 980                     (argc >= 1)) {
 981                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 982                         argc--;
 983                         continue;
 984                 }
 985
 986                 if (!strcasecmp(arg_name, "queue_mode") &&
 987                     (argc >= 1)) {
 988                         const char *queue_mode_name = dm_shift_arg(as);
 989
 990                         if (!strcasecmp(queue_mode_name, "bio"))
 991                                 m->queue_mode = DM_TYPE_BIO_BASED;
 992                         else if (!strcasecmp(queue_mode_name, "rq"))
 993                                 m->queue_mode = DM_TYPE_REQUEST_BASED;
 994                         else if (!strcasecmp(queue_mode_name, "mq"))
 995                                 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
 996                         else {
 997                                 ti->error = "Unknown 'queue_mode' requested";
 998                                 r = -EINVAL;
 999                         }
1000                         argc--;
1001                         continue;
1002                 }
1003
1004                 ti->error = "Unrecognised multipath feature request";
1005                 r = -EINVAL;
1006         } while (argc && !r);
1007
1008         return r;
1009 }
1010
1011 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1012 {
1013         /* target arguments */
1014         static struct dm_arg _args[] = {
1015                 {0, 1024, "invalid number of priority groups"},
1016                 {0, 1024, "invalid initial priority group number"},
1017         };
1018
1019         int r;
1020         struct multipath *m;
1021         struct dm_arg_set as;
1022         unsigned pg_count = 0;
1023         unsigned next_pg_num;
1024
1025         as.argc = argc;
1026         as.argv = argv;
1027
1028         m = alloc_multipath(ti);
1029         if (!m) {
1030                 ti->error = "can't allocate multipath";
1031                 return -EINVAL;
1032         }
1033
1034         r = parse_features(&as, m);
1035         if (r)
1036                 goto bad;
1037
1038         r = alloc_multipath_stage2(ti, m);
1039         if (r)
1040                 goto bad;
1041
1042         r = parse_hw_handler(&as, m);
1043         if (r)
1044                 goto bad;
1045
1046         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1047         if (r)
1048                 goto bad;
1049
1050         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1051         if (r)
1052                 goto bad;
1053
1054         if ((!m->nr_priority_groups && next_pg_num) ||
1055             (m->nr_priority_groups && !next_pg_num)) {
1056                 ti->error = "invalid initial priority group";
1057                 r = -EINVAL;
1058                 goto bad;
1059         }
1060
1061         /* parse the priority groups */
1062         while (as.argc) {
1063                 struct priority_group *pg;
1064                 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1065
1066                 pg = parse_priority_group(&as, m);
1067                 if (IS_ERR(pg)) {
1068                         r = PTR_ERR(pg);
1069                         goto bad;
1070                 }
1071
1072                 nr_valid_paths += pg->nr_pgpaths;
1073                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1074
1075                 list_add_tail(&pg->list, &m->priority_groups);
1076                 pg_count++;
1077                 pg->pg_num = pg_count;
1078                 if (!--next_pg_num)
1079                         m->next_pg = pg;
1080         }
1081
1082         if (pg_count != m->nr_priority_groups) {
1083                 ti->error = "priority group count mismatch";
1084                 r = -EINVAL;
1085                 goto bad;
1086         }
1087
1088         ti->num_flush_bios = 1;
1089         ti->num_discard_bios = 1;
1090         ti->num_write_same_bios = 1;
1091         ti->num_write_zeroes_bios = 1;
1092         if (m->queue_mode == DM_TYPE_BIO_BASED)
1093                 ti->per_io_data_size = multipath_per_bio_data_size();
1094         else
1095                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1096
1097         return 0;
1098
1099  bad:
1100         free_multipath(m);
1101         return r;
1102 }
1103
1104 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1105 {
1106         DEFINE_WAIT(wait);
1107
1108         while (1) {
1109                 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1110
1111                 if (!atomic_read(&m->pg_init_in_progress))
1112                         break;
1113
1114                 io_schedule();
1115         }
1116         finish_wait(&m->pg_init_wait, &wait);
1117 }
1118
1119 static void flush_multipath_work(struct multipath *m)
1120 {
1121         set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1122         smp_mb__after_atomic();
1123
1124         flush_workqueue(kmpath_handlerd);
1125         multipath_wait_for_pg_init_completion(m);
1126         flush_workqueue(kmultipathd);
1127         flush_work(&m->trigger_event);
1128
1129         clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1130         smp_mb__after_atomic();
1131 }
1132
1133 static void multipath_dtr(struct dm_target *ti)
1134 {
1135         struct multipath *m = ti->private;
1136
1137         flush_multipath_work(m);
1138         free_multipath(m);
1139 }
1140
1141 /*
1142  * Take a path out of use.
1143  */
1144 static int fail_path(struct pgpath *pgpath)
1145 {
1146         unsigned long flags;
1147         struct multipath *m = pgpath->pg->m;
1148
1149         spin_lock_irqsave(&m->lock, flags);
1150
1151         if (!pgpath->is_active)
1152                 goto out;
1153
1154         DMWARN("Failing path %s.", pgpath->path.dev->name);
1155
1156         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1157         pgpath->is_active = false;
1158         pgpath->fail_count++;
1159
1160         atomic_dec(&m->nr_valid_paths);
1161
1162         if (pgpath == m->current_pgpath)
1163                 m->current_pgpath = NULL;
1164
1165         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1166                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1167
1168         schedule_work(&m->trigger_event);
1169
1170 out:
1171         spin_unlock_irqrestore(&m->lock, flags);
1172
1173         return 0;
1174 }
1175
1176 /*
1177  * Reinstate a previously-failed path
1178  */
1179 static int reinstate_path(struct pgpath *pgpath)
1180 {
1181         int r = 0, run_queue = 0;
1182         unsigned long flags;
1183         struct multipath *m = pgpath->pg->m;
1184         unsigned nr_valid_paths;
1185
1186         spin_lock_irqsave(&m->lock, flags);
1187
1188         if (pgpath->is_active)
1189                 goto out;
1190
1191         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1192
1193         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1194         if (r)
1195                 goto out;
1196
1197         pgpath->is_active = true;
1198
1199         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1200         if (nr_valid_paths == 1) {
1201                 m->current_pgpath = NULL;
1202                 run_queue = 1;
1203         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1204                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1205                         atomic_inc(&m->pg_init_in_progress);
1206         }
1207
1208         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1209                        pgpath->path.dev->name, nr_valid_paths);
1210
1211         schedule_work(&m->trigger_event);
1212
1213 out:
1214         spin_unlock_irqrestore(&m->lock, flags);
1215         if (run_queue) {
1216                 dm_table_run_md_queue_async(m->ti->table);
1217                 process_queued_io_list(m);
1218         }
1219
1220         return r;
1221 }
1222
1223 /*
1224  * Fail or reinstate all paths that match the provided struct dm_dev.
1225  */
1226 static int action_dev(struct multipath *m, struct dm_dev *dev,
1227                       action_fn action)
1228 {
1229         int r = -EINVAL;
1230         struct pgpath *pgpath;
1231         struct priority_group *pg;
1232
1233         list_for_each_entry(pg, &m->priority_groups, list) {
1234                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1235                         if (pgpath->path.dev == dev)
1236                                 r = action(pgpath);
1237                 }
1238         }
1239
1240         return r;
1241 }
1242
1243 /*
1244  * Temporarily try to avoid having to use the specified PG
1245  */
1246 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1247                       bool bypassed)
1248 {
1249         unsigned long flags;
1250
1251         spin_lock_irqsave(&m->lock, flags);
1252
1253         pg->bypassed = bypassed;
1254         m->current_pgpath = NULL;
1255         m->current_pg = NULL;
1256
1257         spin_unlock_irqrestore(&m->lock, flags);
1258
1259         schedule_work(&m->trigger_event);
1260 }
1261
1262 /*
1263  * Switch to using the specified PG from the next I/O that gets mapped
1264  */
1265 static int switch_pg_num(struct multipath *m, const char *pgstr)
1266 {
1267         struct priority_group *pg;
1268         unsigned pgnum;
1269         unsigned long flags;
1270         char dummy;
1271
1272         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1273             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1274                 DMWARN("invalid PG number supplied to switch_pg_num");
1275                 return -EINVAL;
1276         }
1277
1278         spin_lock_irqsave(&m->lock, flags);
1279         list_for_each_entry(pg, &m->priority_groups, list) {
1280                 pg->bypassed = false;
1281                 if (--pgnum)
1282                         continue;
1283
1284                 m->current_pgpath = NULL;
1285                 m->current_pg = NULL;
1286                 m->next_pg = pg;
1287         }
1288         spin_unlock_irqrestore(&m->lock, flags);
1289
1290         schedule_work(&m->trigger_event);
1291         return 0;
1292 }
1293
1294 /*
1295  * Set/clear bypassed status of a PG.
1296  * PGs are numbered upwards from 1 in the order they were declared.
1297  */
1298 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1299 {
1300         struct priority_group *pg;
1301         unsigned pgnum;
1302         char dummy;
1303
1304         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1305             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1306                 DMWARN("invalid PG number supplied to bypass_pg");
1307                 return -EINVAL;
1308         }
1309
1310         list_for_each_entry(pg, &m->priority_groups, list) {
1311                 if (!--pgnum)
1312                         break;
1313         }
1314
1315         bypass_pg(m, pg, bypassed);
1316         return 0;
1317 }
1318
1319 /*
1320  * Should we retry pg_init immediately?
1321  */
1322 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1323 {
1324         unsigned long flags;
1325         bool limit_reached = false;
1326
1327         spin_lock_irqsave(&m->lock, flags);
1328
1329         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1330             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1331                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1332         else
1333                 limit_reached = true;
1334
1335         spin_unlock_irqrestore(&m->lock, flags);
1336
1337         return limit_reached;
1338 }
1339
1340 static void pg_init_done(void *data, int errors)
1341 {
1342         struct pgpath *pgpath = data;
1343         struct priority_group *pg = pgpath->pg;
1344         struct multipath *m = pg->m;
1345         unsigned long flags;
1346         bool delay_retry = false;
1347
1348         /* device or driver problems */
1349         switch (errors) {
1350         case SCSI_DH_OK:
1351                 break;
1352         case SCSI_DH_NOSYS:
1353                 if (!m->hw_handler_name) {
1354                         errors = 0;
1355                         break;
1356                 }
1357                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1358                       "Error %d.", m->hw_handler_name, errors);
1359                 /*
1360                  * Fail path for now, so we do not ping pong
1361                  */
1362                 fail_path(pgpath);
1363                 break;
1364         case SCSI_DH_DEV_TEMP_BUSY:
1365                 /*
1366                  * Probably doing something like FW upgrade on the
1367                  * controller so try the other pg.
1368                  */
1369                 bypass_pg(m, pg, true);
1370                 break;
1371         case SCSI_DH_RETRY:
1372                 /* Wait before retrying. */
1373                 delay_retry = 1;
1374         case SCSI_DH_IMM_RETRY:
1375         case SCSI_DH_RES_TEMP_UNAVAIL:
1376                 if (pg_init_limit_reached(m, pgpath))
1377                         fail_path(pgpath);
1378                 errors = 0;
1379                 break;
1380         case SCSI_DH_DEV_OFFLINED:
1381         default:
1382                 /*
1383                  * We probably do not want to fail the path for a device
1384                  * error, but this is what the old dm did. In future
1385                  * patches we can do more advanced handling.
1386                  */
1387                 fail_path(pgpath);
1388         }
1389
1390         spin_lock_irqsave(&m->lock, flags);
1391         if (errors) {
1392                 if (pgpath == m->current_pgpath) {
1393                         DMERR("Could not failover device. Error %d.", errors);
1394                         m->current_pgpath = NULL;
1395                         m->current_pg = NULL;
1396                 }
1397         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1398                 pg->bypassed = false;
1399
1400         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1401                 /* Activations of other paths are still on going */
1402                 goto out;
1403
1404         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1405                 if (delay_retry)
1406                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1407                 else
1408                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1409
1410                 if (__pg_init_all_paths(m))
1411                         goto out;
1412         }
1413         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1414
1415         process_queued_io_list(m);
1416
1417         /*
1418          * Wake up any thread waiting to suspend.
1419          */
1420         wake_up(&m->pg_init_wait);
1421
1422 out:
1423         spin_unlock_irqrestore(&m->lock, flags);
1424 }
1425
1426 static void activate_or_offline_path(struct pgpath *pgpath)
1427 {
1428         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1429
1430         if (pgpath->is_active && !blk_queue_dying(q))
1431                 scsi_dh_activate(q, pg_init_done, pgpath);
1432         else
1433                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1434 }
1435
1436 static void activate_path_work(struct work_struct *work)
1437 {
1438         struct pgpath *pgpath =
1439                 container_of(work, struct pgpath, activate_path.work);
1440
1441         activate_or_offline_path(pgpath);
1442 }
1443
1444 static int noretry_error(int error)
1445 {
1446         switch (error) {
1447         case -EBADE:
1448                 /*
1449                  * EBADE signals an reservation conflict.
1450                  * We shouldn't fail the path here as we can communicate with
1451                  * the target.  We should failover to the next path, but in
1452                  * doing so we might be causing a ping-pong between paths.
1453                  * So just return the reservation conflict error.
1454                  */
1455         case -EOPNOTSUPP:
1456         case -EREMOTEIO:
1457         case -EILSEQ:
1458         case -ENODATA:
1459         case -ENOSPC:
1460                 return 1;
1461         }
1462
1463         /* Anything else could be a path failure, so should be retried */
1464         return 0;
1465 }
1466
1467 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1468                             int error, union map_info *map_context)
1469 {
1470         struct dm_mpath_io *mpio = get_mpio(map_context);
1471         struct pgpath *pgpath = mpio->pgpath;
1472         int r = DM_ENDIO_DONE;
1473
1474         /*
1475          * We don't queue any clone request inside the multipath target
1476          * during end I/O handling, since those clone requests don't have
1477          * bio clones.  If we queue them inside the multipath target,
1478          * we need to make bio clones, that requires memory allocation.
1479          * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1480          *  don't have bio clones.)
1481          * Instead of queueing the clone request here, we queue the original
1482          * request into dm core, which will remake a clone request and
1483          * clone bios for it and resubmit it later.
1484          */
1485         if (error && !noretry_error(error)) {
1486                 struct multipath *m = ti->private;
1487
1488                 r = DM_ENDIO_REQUEUE;
1489
1490                 if (pgpath)
1491                         fail_path(pgpath);
1492
1493                 if (atomic_read(&m->nr_valid_paths) == 0 &&
1494                     !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1495                         if (error == -EIO)
1496                                 error = dm_report_EIO(m);
1497                         /* complete with the original error */
1498                         r = DM_ENDIO_DONE;
1499                 }
1500         }
1501
1502         if (pgpath) {
1503                 struct path_selector *ps = &pgpath->pg->ps;
1504
1505                 if (ps->type->end_io)
1506                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1507         }
1508
1509         return r;
1510 }
1511
1512 static int do_end_io_bio(struct multipath *m, struct bio *clone,
1513                          int error, struct dm_mpath_io *mpio)
1514 {
1515         unsigned long flags;
1516
1517         if (!error)
1518                 return 0;       /* I/O complete */
1519
1520         if (noretry_error(error))
1521                 return error;
1522
1523         if (mpio->pgpath)
1524                 fail_path(mpio->pgpath);
1525
1526         if (atomic_read(&m->nr_valid_paths) == 0 &&
1527             !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1528                 return dm_report_EIO(m);
1529
1530         /* Queue for the daemon to resubmit */
1531         dm_bio_restore(get_bio_details_from_bio(clone), clone);
1532
1533         spin_lock_irqsave(&m->lock, flags);
1534         bio_list_add(&m->queued_bios, clone);
1535         spin_unlock_irqrestore(&m->lock, flags);
1536         if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1537                 queue_work(kmultipathd, &m->process_queued_bios);
1538
1539         return DM_ENDIO_INCOMPLETE;
1540 }
1541
1542 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone, int error)
1543 {
1544         struct multipath *m = ti->private;
1545         struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1546         struct pgpath *pgpath;
1547         struct path_selector *ps;
1548         int r;
1549
1550         BUG_ON(!mpio);
1551
1552         r = do_end_io_bio(m, clone, error, mpio);
1553         pgpath = mpio->pgpath;
1554         if (pgpath) {
1555                 ps = &pgpath->pg->ps;
1556                 if (ps->type->end_io)
1557                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1558         }
1559
1560         return r;
1561 }
1562
1563 /*
1564  * Suspend can't complete until all the I/O is processed so if
1565  * the last path fails we must error any remaining I/O.
1566  * Note that if the freeze_bdev fails while suspending, the
1567  * queue_if_no_path state is lost - userspace should reset it.
1568  */
1569 static void multipath_presuspend(struct dm_target *ti)
1570 {
1571         struct multipath *m = ti->private;
1572
1573         queue_if_no_path(m, false, true);
1574 }
1575
1576 static void multipath_postsuspend(struct dm_target *ti)
1577 {
1578         struct multipath *m = ti->private;
1579
1580         mutex_lock(&m->work_mutex);
1581         flush_multipath_work(m);
1582         mutex_unlock(&m->work_mutex);
1583 }
1584
1585 /*
1586  * Restore the queue_if_no_path setting.
1587  */
1588 static void multipath_resume(struct dm_target *ti)
1589 {
1590         struct multipath *m = ti->private;
1591         unsigned long flags;
1592
1593         spin_lock_irqsave(&m->lock, flags);
1594         assign_bit(test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
1595                    MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1596         spin_unlock_irqrestore(&m->lock, flags);
1597 }
1598
1599 /*
1600  * Info output has the following format:
1601  * num_multipath_feature_args [multipath_feature_args]*
1602  * num_handler_status_args [handler_status_args]*
1603  * num_groups init_group_number
1604  *            [A|D|E num_ps_status_args [ps_status_args]*
1605  *             num_paths num_selector_args
1606  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1607  *
1608  * Table output has the following format (identical to the constructor string):
1609  * num_feature_args [features_args]*
1610  * num_handler_args hw_handler [hw_handler_args]*
1611  * num_groups init_group_number
1612  *     [priority selector-name num_ps_args [ps_args]*
1613  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1614  */
1615 static void multipath_status(struct dm_target *ti, status_type_t type,
1616                              unsigned status_flags, char *result, unsigned maxlen)
1617 {
1618         int sz = 0;
1619         unsigned long flags;
1620         struct multipath *m = ti->private;
1621         struct priority_group *pg;
1622         struct pgpath *p;
1623         unsigned pg_num;
1624         char state;
1625
1626         spin_lock_irqsave(&m->lock, flags);
1627
1628         /* Features */
1629         if (type == STATUSTYPE_INFO)
1630                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1631                        atomic_read(&m->pg_init_count));
1632         else {
1633                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1634                               (m->pg_init_retries > 0) * 2 +
1635                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1636                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1637                               (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1638
1639                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1640                         DMEMIT("queue_if_no_path ");
1641                 if (m->pg_init_retries)
1642                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1643                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1644                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1645                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1646                         DMEMIT("retain_attached_hw_handler ");
1647                 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1648                         switch(m->queue_mode) {
1649                         case DM_TYPE_BIO_BASED:
1650                                 DMEMIT("queue_mode bio ");
1651                                 break;
1652                         case DM_TYPE_MQ_REQUEST_BASED:
1653                                 DMEMIT("queue_mode mq ");
1654                                 break;
1655                         default:
1656                                 WARN_ON_ONCE(true);
1657                                 break;
1658                         }
1659                 }
1660         }
1661
1662         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1663                 DMEMIT("0 ");
1664         else
1665                 DMEMIT("1 %s ", m->hw_handler_name);
1666
1667         DMEMIT("%u ", m->nr_priority_groups);
1668
1669         if (m->next_pg)
1670                 pg_num = m->next_pg->pg_num;
1671         else if (m->current_pg)
1672                 pg_num = m->current_pg->pg_num;
1673         else
1674                 pg_num = (m->nr_priority_groups ? 1 : 0);
1675
1676         DMEMIT("%u ", pg_num);
1677
1678         switch (type) {
1679         case STATUSTYPE_INFO:
1680                 list_for_each_entry(pg, &m->priority_groups, list) {
1681                         if (pg->bypassed)
1682                                 state = 'D';    /* Disabled */
1683                         else if (pg == m->current_pg)
1684                                 state = 'A';    /* Currently Active */
1685                         else
1686                                 state = 'E';    /* Enabled */
1687
1688                         DMEMIT("%c ", state);
1689
1690                         if (pg->ps.type->status)
1691                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1692                                                           result + sz,
1693                                                           maxlen - sz);
1694                         else
1695                                 DMEMIT("0 ");
1696
1697                         DMEMIT("%u %u ", pg->nr_pgpaths,
1698                                pg->ps.type->info_args);
1699
1700                         list_for_each_entry(p, &pg->pgpaths, list) {
1701                                 DMEMIT("%s %s %u ", p->path.dev->name,
1702                                        p->is_active ? "A" : "F",
1703                                        p->fail_count);
1704                                 if (pg->ps.type->status)
1705                                         sz += pg->ps.type->status(&pg->ps,
1706                                               &p->path, type, result + sz,
1707                                               maxlen - sz);
1708                         }
1709                 }
1710                 break;
1711
1712         case STATUSTYPE_TABLE:
1713                 list_for_each_entry(pg, &m->priority_groups, list) {
1714                         DMEMIT("%s ", pg->ps.type->name);
1715
1716                         if (pg->ps.type->status)
1717                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1718                                                           result + sz,
1719                                                           maxlen - sz);
1720                         else
1721                                 DMEMIT("0 ");
1722
1723                         DMEMIT("%u %u ", pg->nr_pgpaths,
1724                                pg->ps.type->table_args);
1725
1726                         list_for_each_entry(p, &pg->pgpaths, list) {
1727                                 DMEMIT("%s ", p->path.dev->name);
1728                                 if (pg->ps.type->status)
1729                                         sz += pg->ps.type->status(&pg->ps,
1730                                               &p->path, type, result + sz,
1731                                               maxlen - sz);
1732                         }
1733                 }
1734                 break;
1735         }
1736
1737         spin_unlock_irqrestore(&m->lock, flags);
1738 }
1739
1740 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1741 {
1742         int r = -EINVAL;
1743         struct dm_dev *dev;
1744         struct multipath *m = ti->private;
1745         action_fn action;
1746
1747         mutex_lock(&m->work_mutex);
1748
1749         if (dm_suspended(ti)) {
1750                 r = -EBUSY;
1751                 goto out;
1752         }
1753
1754         if (argc == 1) {
1755                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1756                         r = queue_if_no_path(m, true, false);
1757                         goto out;
1758                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1759                         r = queue_if_no_path(m, false, false);
1760                         goto out;
1761                 }
1762         }
1763
1764         if (argc != 2) {
1765                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1766                 goto out;
1767         }
1768
1769         if (!strcasecmp(argv[0], "disable_group")) {
1770                 r = bypass_pg_num(m, argv[1], true);
1771                 goto out;
1772         } else if (!strcasecmp(argv[0], "enable_group")) {
1773                 r = bypass_pg_num(m, argv[1], false);
1774                 goto out;
1775         } else if (!strcasecmp(argv[0], "switch_group")) {
1776                 r = switch_pg_num(m, argv[1]);
1777                 goto out;
1778         } else if (!strcasecmp(argv[0], "reinstate_path"))
1779                 action = reinstate_path;
1780         else if (!strcasecmp(argv[0], "fail_path"))
1781                 action = fail_path;
1782         else {
1783                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1784                 goto out;
1785         }
1786
1787         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1788         if (r) {
1789                 DMWARN("message: error getting device %s",
1790                        argv[1]);
1791                 goto out;
1792         }
1793
1794         r = action_dev(m, dev, action);
1795
1796         dm_put_device(ti, dev);
1797
1798 out:
1799         mutex_unlock(&m->work_mutex);
1800         return r;
1801 }
1802
1803 static int multipath_prepare_ioctl(struct dm_target *ti,
1804                 struct block_device **bdev, fmode_t *mode)
1805 {
1806         struct multipath *m = ti->private;
1807         struct pgpath *current_pgpath;
1808         int r;
1809
1810         current_pgpath = lockless_dereference(m->current_pgpath);
1811         if (!current_pgpath)
1812                 current_pgpath = choose_pgpath(m, 0);
1813
1814         if (current_pgpath) {
1815                 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1816                         *bdev = current_pgpath->path.dev->bdev;
1817                         *mode = current_pgpath->path.dev->mode;
1818                         r = 0;
1819                 } else {
1820                         /* pg_init has not started or completed */
1821                         r = -ENOTCONN;
1822                 }
1823         } else {
1824                 /* No path is available */
1825                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1826                         r = -ENOTCONN;
1827                 else
1828                         r = -EIO;
1829         }
1830
1831         if (r == -ENOTCONN) {
1832                 if (!lockless_dereference(m->current_pg)) {
1833                         /* Path status changed, redo selection */
1834                         (void) choose_pgpath(m, 0);
1835                 }
1836                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1837                         pg_init_all_paths(m);
1838                 dm_table_run_md_queue_async(m->ti->table);
1839                 process_queued_io_list(m);
1840         }
1841
1842         /*
1843          * Only pass ioctls through if the device sizes match exactly.
1844          */
1845         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1846                 return 1;
1847         return r;
1848 }
1849
1850 static int multipath_iterate_devices(struct dm_target *ti,
1851                                      iterate_devices_callout_fn fn, void *data)
1852 {
1853         struct multipath *m = ti->private;
1854         struct priority_group *pg;
1855         struct pgpath *p;
1856         int ret = 0;
1857
1858         list_for_each_entry(pg, &m->priority_groups, list) {
1859                 list_for_each_entry(p, &pg->pgpaths, list) {
1860                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1861                         if (ret)
1862                                 goto out;
1863                 }
1864         }
1865
1866 out:
1867         return ret;
1868 }
1869
1870 static int pgpath_busy(struct pgpath *pgpath)
1871 {
1872         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1873
1874         return blk_lld_busy(q);
1875 }
1876
1877 /*
1878  * We return "busy", only when we can map I/Os but underlying devices
1879  * are busy (so even if we map I/Os now, the I/Os will wait on
1880  * the underlying queue).
1881  * In other words, if we want to kill I/Os or queue them inside us
1882  * due to map unavailability, we don't return "busy".  Otherwise,
1883  * dm core won't give us the I/Os and we can't do what we want.
1884  */
1885 static int multipath_busy(struct dm_target *ti)
1886 {
1887         bool busy = false, has_active = false;
1888         struct multipath *m = ti->private;
1889         struct priority_group *pg, *next_pg;
1890         struct pgpath *pgpath;
1891
1892         /* pg_init in progress */
1893         if (atomic_read(&m->pg_init_in_progress))
1894                 return true;
1895
1896         /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1897         if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1898                 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1899
1900         /* Guess which priority_group will be used at next mapping time */
1901         pg = lockless_dereference(m->current_pg);
1902         next_pg = lockless_dereference(m->next_pg);
1903         if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
1904                 pg = next_pg;
1905
1906         if (!pg) {
1907                 /*
1908                  * We don't know which pg will be used at next mapping time.
1909                  * We don't call choose_pgpath() here to avoid to trigger
1910                  * pg_init just by busy checking.
1911                  * So we don't know whether underlying devices we will be using
1912                  * at next mapping time are busy or not. Just try mapping.
1913                  */
1914                 return busy;
1915         }
1916
1917         /*
1918          * If there is one non-busy active path at least, the path selector
1919          * will be able to select it. So we consider such a pg as not busy.
1920          */
1921         busy = true;
1922         list_for_each_entry(pgpath, &pg->pgpaths, list) {
1923                 if (pgpath->is_active) {
1924                         has_active = true;
1925                         if (!pgpath_busy(pgpath)) {
1926                                 busy = false;
1927                                 break;
1928                         }
1929                 }
1930         }
1931
1932         if (!has_active) {
1933                 /*
1934                  * No active path in this pg, so this pg won't be used and
1935                  * the current_pg will be changed at next mapping time.
1936                  * We need to try mapping to determine it.
1937                  */
1938                 busy = false;
1939         }
1940
1941         return busy;
1942 }
1943
1944 /*-----------------------------------------------------------------
1945  * Module setup
1946  *---------------------------------------------------------------*/
1947 static struct target_type multipath_target = {
1948         .name = "multipath",
1949         .version = {1, 12, 0},
1950         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1951         .module = THIS_MODULE,
1952         .ctr = multipath_ctr,
1953         .dtr = multipath_dtr,
1954         .clone_and_map_rq = multipath_clone_and_map,
1955         .release_clone_rq = multipath_release_clone,
1956         .rq_end_io = multipath_end_io,
1957         .map = multipath_map_bio,
1958         .end_io = multipath_end_io_bio,
1959         .presuspend = multipath_presuspend,
1960         .postsuspend = multipath_postsuspend,
1961         .resume = multipath_resume,
1962         .status = multipath_status,
1963         .message = multipath_message,
1964         .prepare_ioctl = multipath_prepare_ioctl,
1965         .iterate_devices = multipath_iterate_devices,
1966         .busy = multipath_busy,
1967 };
1968
1969 static int __init dm_multipath_init(void)
1970 {
1971         int r;
1972
1973         r = dm_register_target(&multipath_target);
1974         if (r < 0) {
1975                 DMERR("request-based register failed %d", r);
1976                 r = -EINVAL;
1977                 goto bad_register_target;
1978         }
1979
1980         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1981         if (!kmultipathd) {
1982                 DMERR("failed to create workqueue kmpathd");
1983                 r = -ENOMEM;
1984                 goto bad_alloc_kmultipathd;
1985         }
1986
1987         /*
1988          * A separate workqueue is used to handle the device handlers
1989          * to avoid overloading existing workqueue. Overloading the
1990          * old workqueue would also create a bottleneck in the
1991          * path of the storage hardware device activation.
1992          */
1993         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1994                                                   WQ_MEM_RECLAIM);
1995         if (!kmpath_handlerd) {
1996                 DMERR("failed to create workqueue kmpath_handlerd");
1997                 r = -ENOMEM;
1998                 goto bad_alloc_kmpath_handlerd;
1999         }
2000
2001         return 0;
2002
2003 bad_alloc_kmpath_handlerd:
2004         destroy_workqueue(kmultipathd);
2005 bad_alloc_kmultipathd:
2006         dm_unregister_target(&multipath_target);
2007 bad_register_target:
2008         return r;
2009 }
2010
2011 static void __exit dm_multipath_exit(void)
2012 {
2013         destroy_workqueue(kmpath_handlerd);
2014         destroy_workqueue(kmultipathd);
2015
2016         dm_unregister_target(&multipath_target);
2017 }
2018
2019 module_init(dm_multipath_init);
2020 module_exit(dm_multipath_exit);
2021
2022 MODULE_DESCRIPTION(DM_NAME " multipath target");
2023 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2024 MODULE_LICENSE("GPL");