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         unsigned queue_mode;
  94
  95         /*
  96          * We must use a mempool of dm_mpath_io structs so that we
  97          * can resubmit bios on error.
  98          */
  99         mempool_t *mpio_pool;
 100
 101         struct mutex work_mutex;
 102         struct work_struct trigger_event;
 103
 104         struct work_struct process_queued_bios;
 105         struct bio_list queued_bios;
 106 };
 107
 108 /*
 109  * Context information attached to each io we process.
 110  */
 111 struct dm_mpath_io {
 112         struct pgpath *pgpath;
 113         size_t nr_bytes;
 114 };
 115
 116 typedef int (*action_fn) (struct pgpath *pgpath);
 117
 118 static struct kmem_cache *_mpio_cache;
 119
 120 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 121 static void trigger_event(struct work_struct *work);
 122 static void activate_path(struct work_struct *work);
 123 static void process_queued_bios(struct work_struct *work);
 124
 125 /*-----------------------------------------------
 126  * Multipath state flags.
 127  *-----------------------------------------------*/
 128
 129 #define MPATHF_QUEUE_IO 0                       /* Must we queue all I/O? */
 130 #define MPATHF_QUEUE_IF_NO_PATH 1               /* Queue I/O if last path fails? */
 131 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2         /* Saved state during suspension */
 132 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3     /* If there's already a hw_handler present, don't change it. */
 133 #define MPATHF_PG_INIT_DISABLED 4               /* pg_init is not currently allowed */
 134 #define MPATHF_PG_INIT_REQUIRED 5               /* pg_init needs calling? */
 135 #define MPATHF_PG_INIT_DELAY_RETRY 6            /* Delay pg_init retry? */
 136
 137 /*-----------------------------------------------
 138  * Allocation routines
 139  *-----------------------------------------------*/
 140
 141 static struct pgpath *alloc_pgpath(void)
 142 {
 143         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 144
 145         if (pgpath) {
 146                 pgpath->is_active = true;
 147                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 148         }
 149
 150         return pgpath;
 151 }
 152
 153 static void free_pgpath(struct pgpath *pgpath)
 154 {
 155         kfree(pgpath);
 156 }
 157
 158 static struct priority_group *alloc_priority_group(void)
 159 {
 160         struct priority_group *pg;
 161
 162         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 163
 164         if (pg)
 165                 INIT_LIST_HEAD(&pg->pgpaths);
 166
 167         return pg;
 168 }
 169
 170 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 171 {
 172         struct pgpath *pgpath, *tmp;
 173
 174         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 175                 list_del(&pgpath->list);
 176                 dm_put_device(ti, pgpath->path.dev);
 177                 free_pgpath(pgpath);
 178         }
 179 }
 180
 181 static void free_priority_group(struct priority_group *pg,
 182                                 struct dm_target *ti)
 183 {
 184         struct path_selector *ps = &pg->ps;
 185
 186         if (ps->type) {
 187                 ps->type->destroy(ps);
 188                 dm_put_path_selector(ps->type);
 189         }
 190
 191         free_pgpaths(&pg->pgpaths, ti);
 192         kfree(pg);
 193 }
 194
 195 static struct multipath *alloc_multipath(struct dm_target *ti)
 196 {
 197         struct multipath *m;
 198
 199         m = kzalloc(sizeof(*m), GFP_KERNEL);
 200         if (m) {
 201                 INIT_LIST_HEAD(&m->priority_groups);
 202                 spin_lock_init(&m->lock);
 203                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 204                 atomic_set(&m->nr_valid_paths, 0);
 205                 atomic_set(&m->pg_init_in_progress, 0);
 206                 atomic_set(&m->pg_init_count, 0);
 207                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 208                 INIT_WORK(&m->trigger_event, trigger_event);
 209                 init_waitqueue_head(&m->pg_init_wait);
 210                 mutex_init(&m->work_mutex);
 211
 212                 m->mpio_pool = NULL;
 213                 m->queue_mode = DM_TYPE_NONE;
 214
 215                 m->ti = ti;
 216                 ti->private = m;
 217         }
 218
 219         return m;
 220 }
 221
 222 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
 223 {
 224         if (m->queue_mode == DM_TYPE_NONE) {
 225                 /*
 226                  * Default to request-based.
 227                  */
 228                 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
 229                         m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
 230                 else
 231                         m->queue_mode = DM_TYPE_REQUEST_BASED;
 232         }
 233
 234         if (m->queue_mode == DM_TYPE_REQUEST_BASED) {
 235                 unsigned min_ios = dm_get_reserved_rq_based_ios();
 236
 237                 m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
 238                 if (!m->mpio_pool)
 239                         return -ENOMEM;
 240         }
 241         else if (m->queue_mode == DM_TYPE_BIO_BASED) {
 242                 INIT_WORK(&m->process_queued_bios, process_queued_bios);
 243                 /*
 244                  * bio-based doesn't support any direct scsi_dh management;
 245                  * it just discovers if a scsi_dh is attached.
 246                  */
 247                 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 248         }
 249
 250         dm_table_set_type(ti->table, m->queue_mode);
 251
 252         return 0;
 253 }
 254
 255 static void free_multipath(struct multipath *m)
 256 {
 257         struct priority_group *pg, *tmp;
 258
 259         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 260                 list_del(&pg->list);
 261                 free_priority_group(pg, m->ti);
 262         }
 263
 264         kfree(m->hw_handler_name);
 265         kfree(m->hw_handler_params);
 266         mempool_destroy(m->mpio_pool);
 267         kfree(m);
 268 }
 269
 270 static struct dm_mpath_io *get_mpio(union map_info *info)
 271 {
 272         return info->ptr;
 273 }
 274
 275 static struct dm_mpath_io *set_mpio(struct multipath *m, union map_info *info)
 276 {
 277         struct dm_mpath_io *mpio;
 278
 279         if (!m->mpio_pool) {
 280                 /* Use blk-mq pdu memory requested via per_io_data_size */
 281                 mpio = get_mpio(info);
 282                 memset(mpio, 0, sizeof(*mpio));
 283                 return mpio;
 284         }
 285
 286         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 287         if (!mpio)
 288                 return NULL;
 289
 290         memset(mpio, 0, sizeof(*mpio));
 291         info->ptr = mpio;
 292
 293         return mpio;
 294 }
 295
 296 static void clear_request_fn_mpio(struct multipath *m, union map_info *info)
 297 {
 298         /* Only needed for non blk-mq (.request_fn) multipath */
 299         if (m->mpio_pool) {
 300                 struct dm_mpath_io *mpio = info->ptr;
 301
 302                 info->ptr = NULL;
 303                 mempool_free(mpio, m->mpio_pool);
 304         }
 305 }
 306
 307 static size_t multipath_per_bio_data_size(void)
 308 {
 309         return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
 310 }
 311
 312 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
 313 {
 314         return dm_per_bio_data(bio, multipath_per_bio_data_size());
 315 }
 316
 317 static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
 318 {
 319         /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
 320         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 321         void *bio_details = mpio + 1;
 322
 323         return bio_details;
 324 }
 325
 326 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p,
 327                                         struct dm_bio_details **bio_details_p)
 328 {
 329         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 330         struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
 331
 332         memset(mpio, 0, sizeof(*mpio));
 333         memset(bio_details, 0, sizeof(*bio_details));
 334         dm_bio_record(bio_details, bio);
 335
 336         if (mpio_p)
 337                 *mpio_p = mpio;
 338         if (bio_details_p)
 339                 *bio_details_p = bio_details;
 340 }
 341
 342 /*-----------------------------------------------
 343  * Path selection
 344  *-----------------------------------------------*/
 345
 346 static int __pg_init_all_paths(struct multipath *m)
 347 {
 348         struct pgpath *pgpath;
 349         unsigned long pg_init_delay = 0;
 350
 351         if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
 352                 return 0;
 353
 354         atomic_inc(&m->pg_init_count);
 355         clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 356
 357         /* Check here to reset pg_init_required */
 358         if (!m->current_pg)
 359                 return 0;
 360
 361         if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
 362                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 363                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 364         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 365                 /* Skip failed paths */
 366                 if (!pgpath->is_active)
 367                         continue;
 368                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 369                                        pg_init_delay))
 370                         atomic_inc(&m->pg_init_in_progress);
 371         }
 372         return atomic_read(&m->pg_init_in_progress);
 373 }
 374
 375 static int pg_init_all_paths(struct multipath *m)
 376 {
 377         int r;
 378         unsigned long flags;
 379
 380         spin_lock_irqsave(&m->lock, flags);
 381         r = __pg_init_all_paths(m);
 382         spin_unlock_irqrestore(&m->lock, flags);
 383
 384         return r;
 385 }
 386
 387 static void __switch_pg(struct multipath *m, struct priority_group *pg)
 388 {
 389         m->current_pg = pg;
 390
 391         /* Must we initialise the PG first, and queue I/O till it's ready? */
 392         if (m->hw_handler_name) {
 393                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 394                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 395         } else {
 396                 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 397                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 398         }
 399
 400         atomic_set(&m->pg_init_count, 0);
 401 }
 402
 403 static struct pgpath *choose_path_in_pg(struct multipath *m,
 404                                         struct priority_group *pg,
 405                                         size_t nr_bytes)
 406 {
 407         unsigned long flags;
 408         struct dm_path *path;
 409         struct pgpath *pgpath;
 410
 411         path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 412         if (!path)
 413                 return ERR_PTR(-ENXIO);
 414
 415         pgpath = path_to_pgpath(path);
 416
 417         if (unlikely(lockless_dereference(m->current_pg) != pg)) {
 418                 /* Only update current_pgpath if pg changed */
 419                 spin_lock_irqsave(&m->lock, flags);
 420                 m->current_pgpath = pgpath;
 421                 __switch_pg(m, pg);
 422                 spin_unlock_irqrestore(&m->lock, flags);
 423         }
 424
 425         return pgpath;
 426 }
 427
 428 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
 429 {
 430         unsigned long flags;
 431         struct priority_group *pg;
 432         struct pgpath *pgpath;
 433         bool bypassed = true;
 434
 435         if (!atomic_read(&m->nr_valid_paths)) {
 436                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 437                 goto failed;
 438         }
 439
 440         /* Were we instructed to switch PG? */
 441         if (lockless_dereference(m->next_pg)) {
 442                 spin_lock_irqsave(&m->lock, flags);
 443                 pg = m->next_pg;
 444                 if (!pg) {
 445                         spin_unlock_irqrestore(&m->lock, flags);
 446                         goto check_current_pg;
 447                 }
 448                 m->next_pg = NULL;
 449                 spin_unlock_irqrestore(&m->lock, flags);
 450                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 451                 if (!IS_ERR_OR_NULL(pgpath))
 452                         return pgpath;
 453         }
 454
 455         /* Don't change PG until it has no remaining paths */
 456 check_current_pg:
 457         pg = lockless_dereference(m->current_pg);
 458         if (pg) {
 459                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 460                 if (!IS_ERR_OR_NULL(pgpath))
 461                         return pgpath;
 462         }
 463
 464         /*
 465          * Loop through priority groups until we find a valid path.
 466          * First time we skip PGs marked 'bypassed'.
 467          * Second time we only try the ones we skipped, but set
 468          * pg_init_delay_retry so we do not hammer controllers.
 469          */
 470         do {
 471                 list_for_each_entry(pg, &m->priority_groups, list) {
 472                         if (pg->bypassed == bypassed)
 473                                 continue;
 474                         pgpath = choose_path_in_pg(m, pg, nr_bytes);
 475                         if (!IS_ERR_OR_NULL(pgpath)) {
 476                                 if (!bypassed)
 477                                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
 478                                 return pgpath;
 479                         }
 480                 }
 481         } while (bypassed--);
 482
 483 failed:
 484         spin_lock_irqsave(&m->lock, flags);
 485         m->current_pgpath = NULL;
 486         m->current_pg = NULL;
 487         spin_unlock_irqrestore(&m->lock, flags);
 488
 489         return NULL;
 490 }
 491
 492 /*
 493  * Check whether bios must be queued in the device-mapper core rather
 494  * than here in the target.
 495  *
 496  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 497  * same value then we are not between multipath_presuspend()
 498  * and multipath_resume() calls and we have no need to check
 499  * for the DMF_NOFLUSH_SUSPENDING flag.
 500  */
 501 static bool __must_push_back(struct multipath *m)
 502 {
 503         return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) !=
 504                  test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) &&
 505                 dm_noflush_suspending(m->ti));
 506 }
 507
 508 static bool must_push_back_rq(struct multipath *m)
 509 {
 510         return (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) ||
 511                 __must_push_back(m));
 512 }
 513
 514 static bool must_push_back_bio(struct multipath *m)
 515 {
 516         return __must_push_back(m);
 517 }
 518
 519 /*
 520  * Map cloned requests (request-based multipath)
 521  */
 522 static int __multipath_map(struct dm_target *ti, struct request *clone,
 523                            union map_info *map_context,
 524                            struct request *rq, struct request **__clone)
 525 {
 526         struct multipath *m = ti->private;
 527         int r = DM_MAPIO_REQUEUE;
 528         size_t nr_bytes = clone ? blk_rq_bytes(clone) : blk_rq_bytes(rq);
 529         struct pgpath *pgpath;
 530         struct block_device *bdev;
 531         struct dm_mpath_io *mpio;
 532
 533         /* Do we need to select a new pgpath? */
 534         pgpath = lockless_dereference(m->current_pgpath);
 535         if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
 536                 pgpath = choose_pgpath(m, nr_bytes);
 537
 538         if (!pgpath) {
 539                 if (!must_push_back_rq(m))
 540                         r = -EIO;       /* Failed */
 541                 return r;
 542         } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
 543                    test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
 544                 pg_init_all_paths(m);
 545                 return r;
 546         }
 547
 548         mpio = set_mpio(m, map_context);
 549         if (!mpio)
 550                 /* ENOMEM, requeue */
 551                 return r;
 552
 553         mpio->pgpath = pgpath;
 554         mpio->nr_bytes = nr_bytes;
 555
 556         bdev = pgpath->path.dev->bdev;
 557
 558         if (clone) {
 559                 /*
 560                  * Old request-based interface: allocated clone is passed in.
 561                  * Used by: .request_fn stacked on .request_fn path(s).
 562                  */
 563                 clone->q = bdev_get_queue(bdev);
 564                 clone->rq_disk = bdev->bd_disk;
 565                 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 566         } else {
 567                 /*
 568                  * blk-mq request-based interface; used by both:
 569                  * .request_fn stacked on blk-mq path(s) and
 570                  * blk-mq stacked on blk-mq path(s).
 571                  */
 572                 *__clone = blk_mq_alloc_request(bdev_get_queue(bdev),
 573                                                 rq_data_dir(rq), BLK_MQ_REQ_NOWAIT);
 574                 if (IS_ERR(*__clone)) {
 575                         /* ENOMEM, requeue */
 576                         clear_request_fn_mpio(m, map_context);
 577                         return r;
 578                 }
 579                 (*__clone)->bio = (*__clone)->biotail = NULL;
 580                 (*__clone)->rq_disk = bdev->bd_disk;
 581                 (*__clone)->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 582         }
 583
 584         if (pgpath->pg->ps.type->start_io)
 585                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 586                                               &pgpath->path,
 587                                               nr_bytes);
 588         return DM_MAPIO_REMAPPED;
 589 }
 590
 591 static int multipath_map(struct dm_target *ti, struct request *clone,
 592                          union map_info *map_context)
 593 {
 594         return __multipath_map(ti, clone, map_context, NULL, NULL);
 595 }
 596
 597 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 598                                    union map_info *map_context,
 599                                    struct request **clone)
 600 {
 601         return __multipath_map(ti, NULL, map_context, rq, clone);
 602 }
 603
 604 static void multipath_release_clone(struct request *clone)
 605 {
 606         blk_mq_free_request(clone);
 607 }
 608
 609 /*
 610  * Map cloned bios (bio-based multipath)
 611  */
 612 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
 613 {
 614         size_t nr_bytes = bio->bi_iter.bi_size;
 615         struct pgpath *pgpath;
 616         unsigned long flags;
 617         bool queue_io;
 618
 619         /* Do we need to select a new pgpath? */
 620         pgpath = lockless_dereference(m->current_pgpath);
 621         queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
 622         if (!pgpath || !queue_io)
 623                 pgpath = choose_pgpath(m, nr_bytes);
 624
 625         if ((pgpath && queue_io) ||
 626             (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
 627                 /* Queue for the daemon to resubmit */
 628                 spin_lock_irqsave(&m->lock, flags);
 629                 bio_list_add(&m->queued_bios, bio);
 630                 spin_unlock_irqrestore(&m->lock, flags);
 631                 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
 632                 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
 633                         pg_init_all_paths(m);
 634                 else if (!queue_io)
 635                         queue_work(kmultipathd, &m->process_queued_bios);
 636                 return DM_MAPIO_SUBMITTED;
 637         }
 638
 639         if (!pgpath) {
 640                 if (!must_push_back_bio(m))
 641                         return -EIO;
 642                 return DM_MAPIO_REQUEUE;
 643         }
 644
 645         mpio->pgpath = pgpath;
 646         mpio->nr_bytes = nr_bytes;
 647
 648         bio->bi_error = 0;
 649         bio->bi_bdev = pgpath->path.dev->bdev;
 650         bio->bi_rw |= REQ_FAILFAST_TRANSPORT;
 651
 652         if (pgpath->pg->ps.type->start_io)
 653                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 654                                               &pgpath->path,
 655                                               nr_bytes);
 656         return DM_MAPIO_REMAPPED;
 657 }
 658
 659 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
 660 {
 661         struct multipath *m = ti->private;
 662         struct dm_mpath_io *mpio = NULL;
 663
 664         multipath_init_per_bio_data(bio, &mpio, NULL);
 665
 666         return __multipath_map_bio(m, bio, mpio);
 667 }
 668
 669 static void process_queued_bios_list(struct multipath *m)
 670 {
 671         if (m->queue_mode == DM_TYPE_BIO_BASED)
 672                 queue_work(kmultipathd, &m->process_queued_bios);
 673 }
 674
 675 static void process_queued_bios(struct work_struct *work)
 676 {
 677         int r;
 678         unsigned long flags;
 679         struct bio *bio;
 680         struct bio_list bios;
 681         struct blk_plug plug;
 682         struct multipath *m =
 683                 container_of(work, struct multipath, process_queued_bios);
 684
 685         bio_list_init(&bios);
 686
 687         spin_lock_irqsave(&m->lock, flags);
 688
 689         if (bio_list_empty(&m->queued_bios)) {
 690                 spin_unlock_irqrestore(&m->lock, flags);
 691                 return;
 692         }
 693
 694         bio_list_merge(&bios, &m->queued_bios);
 695         bio_list_init(&m->queued_bios);
 696
 697         spin_unlock_irqrestore(&m->lock, flags);
 698
 699         blk_start_plug(&plug);
 700         while ((bio = bio_list_pop(&bios))) {
 701                 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
 702                 if (r < 0 || r == DM_MAPIO_REQUEUE) {
 703                         bio->bi_error = r;
 704                         bio_endio(bio);
 705                 } else if (r == DM_MAPIO_REMAPPED)
 706                         generic_make_request(bio);
 707         }
 708         blk_finish_plug(&plug);
 709 }
 710
 711 /*
 712  * If we run out of usable paths, should we queue I/O or error it?
 713  */
 714 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
 715                             bool save_old_value)
 716 {
 717         unsigned long flags;
 718
 719         spin_lock_irqsave(&m->lock, flags);
 720
 721         if (save_old_value) {
 722                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
 723                         set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 724                 else
 725                         clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 726         } else {
 727                 if (queue_if_no_path)
 728                         set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 729                 else
 730                         clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 731         }
 732         if (queue_if_no_path)
 733                 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 734         else
 735                 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 736
 737         spin_unlock_irqrestore(&m->lock, flags);
 738
 739         if (!queue_if_no_path) {
 740                 dm_table_run_md_queue_async(m->ti->table);
 741                 process_queued_bios_list(m);
 742         }
 743
 744         return 0;
 745 }
 746
 747 /*
 748  * An event is triggered whenever a path is taken out of use.
 749  * Includes path failure and PG bypass.
 750  */
 751 static void trigger_event(struct work_struct *work)
 752 {
 753         struct multipath *m =
 754                 container_of(work, struct multipath, trigger_event);
 755
 756         dm_table_event(m->ti->table);
 757 }
 758
 759 /*-----------------------------------------------------------------
 760  * Constructor/argument parsing:
 761  * <#multipath feature args> [<arg>]*
 762  * <#hw_handler args> [hw_handler [<arg>]*]
 763  * <#priority groups>
 764  * <initial priority group>
 765  *     [<selector> <#selector args> [<arg>]*
 766  *      <#paths> <#per-path selector args>
 767  *         [<path> [<arg>]* ]+ ]+
 768  *---------------------------------------------------------------*/
 769 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 770                                struct dm_target *ti)
 771 {
 772         int r;
 773         struct path_selector_type *pst;
 774         unsigned ps_argc;
 775
 776         static struct dm_arg _args[] = {
 777                 {0, 1024, "invalid number of path selector args"},
 778         };
 779
 780         pst = dm_get_path_selector(dm_shift_arg(as));
 781         if (!pst) {
 782                 ti->error = "unknown path selector type";
 783                 return -EINVAL;
 784         }
 785
 786         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 787         if (r) {
 788                 dm_put_path_selector(pst);
 789                 return -EINVAL;
 790         }
 791
 792         r = pst->create(&pg->ps, ps_argc, as->argv);
 793         if (r) {
 794                 dm_put_path_selector(pst);
 795                 ti->error = "path selector constructor failed";
 796                 return r;
 797         }
 798
 799         pg->ps.type = pst;
 800         dm_consume_args(as, ps_argc);
 801
 802         return 0;
 803 }
 804
 805 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 806                                struct dm_target *ti)
 807 {
 808         int r;
 809         struct pgpath *p;
 810         struct multipath *m = ti->private;
 811         struct request_queue *q = NULL;
 812         const char *attached_handler_name;
 813
 814         /* we need at least a path arg */
 815         if (as->argc < 1) {
 816                 ti->error = "no device given";
 817                 return ERR_PTR(-EINVAL);
 818         }
 819
 820         p = alloc_pgpath();
 821         if (!p)
 822                 return ERR_PTR(-ENOMEM);
 823
 824         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 825                           &p->path.dev);
 826         if (r) {
 827                 ti->error = "error getting device";
 828                 goto bad;
 829         }
 830
 831         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
 832                 q = bdev_get_queue(p->path.dev->bdev);
 833
 834         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
 835 retain:
 836                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 837                 if (attached_handler_name) {
 838                         /*
 839                          * Reset hw_handler_name to match the attached handler
 840                          * and clear any hw_handler_params associated with the
 841                          * ignored handler.
 842                          *
 843                          * NB. This modifies the table line to show the actual
 844                          * handler instead of the original table passed in.
 845                          */
 846                         kfree(m->hw_handler_name);
 847                         m->hw_handler_name = attached_handler_name;
 848
 849                         kfree(m->hw_handler_params);
 850                         m->hw_handler_params = NULL;
 851                 }
 852         }
 853
 854         if (m->hw_handler_name) {
 855                 r = scsi_dh_attach(q, m->hw_handler_name);
 856                 if (r == -EBUSY) {
 857                         char b[BDEVNAME_SIZE];
 858
 859                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
 860                                 bdevname(p->path.dev->bdev, b));
 861                         goto retain;
 862                 }
 863                 if (r < 0) {
 864                         ti->error = "error attaching hardware handler";
 865                         dm_put_device(ti, p->path.dev);
 866                         goto bad;
 867                 }
 868
 869                 if (m->hw_handler_params) {
 870                         r = scsi_dh_set_params(q, m->hw_handler_params);
 871                         if (r < 0) {
 872                                 ti->error = "unable to set hardware "
 873                                                         "handler parameters";
 874                                 dm_put_device(ti, p->path.dev);
 875                                 goto bad;
 876                         }
 877                 }
 878         }
 879
 880         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 881         if (r) {
 882                 dm_put_device(ti, p->path.dev);
 883                 goto bad;
 884         }
 885
 886         return p;
 887
 888  bad:
 889         free_pgpath(p);
 890         return ERR_PTR(r);
 891 }
 892
 893 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 894                                                    struct multipath *m)
 895 {
 896         static struct dm_arg _args[] = {
 897                 {1, 1024, "invalid number of paths"},
 898                 {0, 1024, "invalid number of selector args"}
 899         };
 900
 901         int r;
 902         unsigned i, nr_selector_args, nr_args;
 903         struct priority_group *pg;
 904         struct dm_target *ti = m->ti;
 905
 906         if (as->argc < 2) {
 907                 as->argc = 0;
 908                 ti->error = "not enough priority group arguments";
 909                 return ERR_PTR(-EINVAL);
 910         }
 911
 912         pg = alloc_priority_group();
 913         if (!pg) {
 914                 ti->error = "couldn't allocate priority group";
 915                 return ERR_PTR(-ENOMEM);
 916         }
 917         pg->m = m;
 918
 919         r = parse_path_selector(as, pg, ti);
 920         if (r)
 921                 goto bad;
 922
 923         /*
 924          * read the paths
 925          */
 926         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 927         if (r)
 928                 goto bad;
 929
 930         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 931         if (r)
 932                 goto bad;
 933
 934         nr_args = 1 + nr_selector_args;
 935         for (i = 0; i < pg->nr_pgpaths; i++) {
 936                 struct pgpath *pgpath;
 937                 struct dm_arg_set path_args;
 938
 939                 if (as->argc < nr_args) {
 940                         ti->error = "not enough path parameters";
 941                         r = -EINVAL;
 942                         goto bad;
 943                 }
 944
 945                 path_args.argc = nr_args;
 946                 path_args.argv = as->argv;
 947
 948                 pgpath = parse_path(&path_args, &pg->ps, ti);
 949                 if (IS_ERR(pgpath)) {
 950                         r = PTR_ERR(pgpath);
 951                         goto bad;
 952                 }
 953
 954                 pgpath->pg = pg;
 955                 list_add_tail(&pgpath->list, &pg->pgpaths);
 956                 dm_consume_args(as, nr_args);
 957         }
 958
 959         return pg;
 960
 961  bad:
 962         free_priority_group(pg, ti);
 963         return ERR_PTR(r);
 964 }
 965
 966 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 967 {
 968         unsigned hw_argc;
 969         int ret;
 970         struct dm_target *ti = m->ti;
 971
 972         static struct dm_arg _args[] = {
 973                 {0, 1024, "invalid number of hardware handler args"},
 974         };
 975
 976         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 977                 return -EINVAL;
 978
 979         if (!hw_argc)
 980                 return 0;
 981
 982         if (m->queue_mode == DM_TYPE_BIO_BASED) {
 983                 dm_consume_args(as, hw_argc);
 984                 DMERR("bio-based multipath doesn't allow hardware handler args");
 985                 return 0;
 986         }
 987
 988         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 989
 990         if (hw_argc > 1) {
 991                 char *p;
 992                 int i, j, len = 4;
 993
 994                 for (i = 0; i <= hw_argc - 2; i++)
 995                         len += strlen(as->argv[i]) + 1;
 996                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 997                 if (!p) {
 998                         ti->error = "memory allocation failed";
 999                         ret = -ENOMEM;
1000                         goto fail;
1001                 }
1002                 j = sprintf(p, "%d", hw_argc - 1);
1003                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1004                         j = sprintf(p, "%s", as->argv[i]);
1005         }
1006         dm_consume_args(as, hw_argc - 1);
1007
1008         return 0;
1009 fail:
1010         kfree(m->hw_handler_name);
1011         m->hw_handler_name = NULL;
1012         return ret;
1013 }
1014
1015 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1016 {
1017         int r;
1018         unsigned argc;
1019         struct dm_target *ti = m->ti;
1020         const char *arg_name;
1021
1022         static struct dm_arg _args[] = {
1023                 {0, 8, "invalid number of feature args"},
1024                 {1, 50, "pg_init_retries must be between 1 and 50"},
1025                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1026         };
1027
1028         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1029         if (r)
1030                 return -EINVAL;
1031
1032         if (!argc)
1033                 return 0;
1034
1035         do {
1036                 arg_name = dm_shift_arg(as);
1037                 argc--;
1038
1039                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1040                         r = queue_if_no_path(m, true, false);
1041                         continue;
1042                 }
1043
1044                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1045                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1046                         continue;
1047                 }
1048
1049                 if (!strcasecmp(arg_name, "pg_init_retries") &&
1050                     (argc >= 1)) {
1051                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1052                         argc--;
1053                         continue;
1054                 }
1055
1056                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1057                     (argc >= 1)) {
1058                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1059                         argc--;
1060                         continue;
1061                 }
1062
1063                 if (!strcasecmp(arg_name, "queue_mode") &&
1064                     (argc >= 1)) {
1065                         const char *queue_mode_name = dm_shift_arg(as);
1066
1067                         if (!strcasecmp(queue_mode_name, "bio"))
1068                                 m->queue_mode = DM_TYPE_BIO_BASED;
1069                         else if (!strcasecmp(queue_mode_name, "rq"))
1070                                 m->queue_mode = DM_TYPE_REQUEST_BASED;
1071                         else if (!strcasecmp(queue_mode_name, "mq"))
1072                                 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1073                         else {
1074                                 ti->error = "Unknown 'queue_mode' requested";
1075                                 r = -EINVAL;
1076                         }
1077                         argc--;
1078                         continue;
1079                 }
1080
1081                 ti->error = "Unrecognised multipath feature request";
1082                 r = -EINVAL;
1083         } while (argc && !r);
1084
1085         return r;
1086 }
1087
1088 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1089 {
1090         /* target arguments */
1091         static struct dm_arg _args[] = {
1092                 {0, 1024, "invalid number of priority groups"},
1093                 {0, 1024, "invalid initial priority group number"},
1094         };
1095
1096         int r;
1097         struct multipath *m;
1098         struct dm_arg_set as;
1099         unsigned pg_count = 0;
1100         unsigned next_pg_num;
1101
1102         as.argc = argc;
1103         as.argv = argv;
1104
1105         m = alloc_multipath(ti);
1106         if (!m) {
1107                 ti->error = "can't allocate multipath";
1108                 return -EINVAL;
1109         }
1110
1111         r = parse_features(&as, m);
1112         if (r)
1113                 goto bad;
1114
1115         r = alloc_multipath_stage2(ti, m);
1116         if (r)
1117                 goto bad;
1118
1119         r = parse_hw_handler(&as, m);
1120         if (r)
1121                 goto bad;
1122
1123         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1124         if (r)
1125                 goto bad;
1126
1127         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1128         if (r)
1129                 goto bad;
1130
1131         if ((!m->nr_priority_groups && next_pg_num) ||
1132             (m->nr_priority_groups && !next_pg_num)) {
1133                 ti->error = "invalid initial priority group";
1134                 r = -EINVAL;
1135                 goto bad;
1136         }
1137
1138         /* parse the priority groups */
1139         while (as.argc) {
1140                 struct priority_group *pg;
1141                 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1142
1143                 pg = parse_priority_group(&as, m);
1144                 if (IS_ERR(pg)) {
1145                         r = PTR_ERR(pg);
1146                         goto bad;
1147                 }
1148
1149                 nr_valid_paths += pg->nr_pgpaths;
1150                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1151
1152                 list_add_tail(&pg->list, &m->priority_groups);
1153                 pg_count++;
1154                 pg->pg_num = pg_count;
1155                 if (!--next_pg_num)
1156                         m->next_pg = pg;
1157         }
1158
1159         if (pg_count != m->nr_priority_groups) {
1160                 ti->error = "priority group count mismatch";
1161                 r = -EINVAL;
1162                 goto bad;
1163         }
1164
1165         ti->num_flush_bios = 1;
1166         ti->num_discard_bios = 1;
1167         ti->num_write_same_bios = 1;
1168         if (m->queue_mode == DM_TYPE_BIO_BASED)
1169                 ti->per_io_data_size = multipath_per_bio_data_size();
1170         else if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
1171                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1172
1173         return 0;
1174
1175  bad:
1176         free_multipath(m);
1177         return r;
1178 }
1179
1180 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1181 {
1182         DECLARE_WAITQUEUE(wait, current);
1183
1184         add_wait_queue(&m->pg_init_wait, &wait);
1185
1186         while (1) {
1187                 set_current_state(TASK_UNINTERRUPTIBLE);
1188
1189                 if (!atomic_read(&m->pg_init_in_progress))
1190                         break;
1191
1192                 io_schedule();
1193         }
1194         set_current_state(TASK_RUNNING);
1195
1196         remove_wait_queue(&m->pg_init_wait, &wait);
1197 }
1198
1199 static void flush_multipath_work(struct multipath *m)
1200 {
1201         set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1202         smp_mb__after_atomic();
1203
1204         flush_workqueue(kmpath_handlerd);
1205         multipath_wait_for_pg_init_completion(m);
1206         flush_workqueue(kmultipathd);
1207         flush_work(&m->trigger_event);
1208
1209         clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1210         smp_mb__after_atomic();
1211 }
1212
1213 static void multipath_dtr(struct dm_target *ti)
1214 {
1215         struct multipath *m = ti->private;
1216
1217         flush_multipath_work(m);
1218         free_multipath(m);
1219 }
1220
1221 /*
1222  * Take a path out of use.
1223  */
1224 static int fail_path(struct pgpath *pgpath)
1225 {
1226         unsigned long flags;
1227         struct multipath *m = pgpath->pg->m;
1228
1229         spin_lock_irqsave(&m->lock, flags);
1230
1231         if (!pgpath->is_active)
1232                 goto out;
1233
1234         DMWARN("Failing path %s.", pgpath->path.dev->name);
1235
1236         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1237         pgpath->is_active = false;
1238         pgpath->fail_count++;
1239
1240         atomic_dec(&m->nr_valid_paths);
1241
1242         if (pgpath == m->current_pgpath)
1243                 m->current_pgpath = NULL;
1244
1245         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1246                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1247
1248         schedule_work(&m->trigger_event);
1249
1250 out:
1251         spin_unlock_irqrestore(&m->lock, flags);
1252
1253         return 0;
1254 }
1255
1256 /*
1257  * Reinstate a previously-failed path
1258  */
1259 static int reinstate_path(struct pgpath *pgpath)
1260 {
1261         int r = 0, run_queue = 0;
1262         unsigned long flags;
1263         struct multipath *m = pgpath->pg->m;
1264         unsigned nr_valid_paths;
1265
1266         spin_lock_irqsave(&m->lock, flags);
1267
1268         if (pgpath->is_active)
1269                 goto out;
1270
1271         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1272
1273         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1274         if (r)
1275                 goto out;
1276
1277         pgpath->is_active = true;
1278
1279         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1280         if (nr_valid_paths == 1) {
1281                 m->current_pgpath = NULL;
1282                 run_queue = 1;
1283         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1284                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1285                         atomic_inc(&m->pg_init_in_progress);
1286         }
1287
1288         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1289                        pgpath->path.dev->name, nr_valid_paths);
1290
1291         schedule_work(&m->trigger_event);
1292
1293 out:
1294         spin_unlock_irqrestore(&m->lock, flags);
1295         if (run_queue) {
1296                 dm_table_run_md_queue_async(m->ti->table);
1297                 process_queued_bios_list(m);
1298         }
1299
1300         return r;
1301 }
1302
1303 /*
1304  * Fail or reinstate all paths that match the provided struct dm_dev.
1305  */
1306 static int action_dev(struct multipath *m, struct dm_dev *dev,
1307                       action_fn action)
1308 {
1309         int r = -EINVAL;
1310         struct pgpath *pgpath;
1311         struct priority_group *pg;
1312
1313         list_for_each_entry(pg, &m->priority_groups, list) {
1314                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1315                         if (pgpath->path.dev == dev)
1316                                 r = action(pgpath);
1317                 }
1318         }
1319
1320         return r;
1321 }
1322
1323 /*
1324  * Temporarily try to avoid having to use the specified PG
1325  */
1326 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1327                       bool bypassed)
1328 {
1329         unsigned long flags;
1330
1331         spin_lock_irqsave(&m->lock, flags);
1332
1333         pg->bypassed = bypassed;
1334         m->current_pgpath = NULL;
1335         m->current_pg = NULL;
1336
1337         spin_unlock_irqrestore(&m->lock, flags);
1338
1339         schedule_work(&m->trigger_event);
1340 }
1341
1342 /*
1343  * Switch to using the specified PG from the next I/O that gets mapped
1344  */
1345 static int switch_pg_num(struct multipath *m, const char *pgstr)
1346 {
1347         struct priority_group *pg;
1348         unsigned pgnum;
1349         unsigned long flags;
1350         char dummy;
1351
1352         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1353             (pgnum > m->nr_priority_groups)) {
1354                 DMWARN("invalid PG number supplied to switch_pg_num");
1355                 return -EINVAL;
1356         }
1357
1358         spin_lock_irqsave(&m->lock, flags);
1359         list_for_each_entry(pg, &m->priority_groups, list) {
1360                 pg->bypassed = false;
1361                 if (--pgnum)
1362                         continue;
1363
1364                 m->current_pgpath = NULL;
1365                 m->current_pg = NULL;
1366                 m->next_pg = pg;
1367         }
1368         spin_unlock_irqrestore(&m->lock, flags);
1369
1370         schedule_work(&m->trigger_event);
1371         return 0;
1372 }
1373
1374 /*
1375  * Set/clear bypassed status of a PG.
1376  * PGs are numbered upwards from 1 in the order they were declared.
1377  */
1378 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1379 {
1380         struct priority_group *pg;
1381         unsigned pgnum;
1382         char dummy;
1383
1384         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1385             (pgnum > m->nr_priority_groups)) {
1386                 DMWARN("invalid PG number supplied to bypass_pg");
1387                 return -EINVAL;
1388         }
1389
1390         list_for_each_entry(pg, &m->priority_groups, list) {
1391                 if (!--pgnum)
1392                         break;
1393         }
1394
1395         bypass_pg(m, pg, bypassed);
1396         return 0;
1397 }
1398
1399 /*
1400  * Should we retry pg_init immediately?
1401  */
1402 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1403 {
1404         unsigned long flags;
1405         bool limit_reached = false;
1406
1407         spin_lock_irqsave(&m->lock, flags);
1408
1409         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1410             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1411                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1412         else
1413                 limit_reached = true;
1414
1415         spin_unlock_irqrestore(&m->lock, flags);
1416
1417         return limit_reached;
1418 }
1419
1420 static void pg_init_done(void *data, int errors)
1421 {
1422         struct pgpath *pgpath = data;
1423         struct priority_group *pg = pgpath->pg;
1424         struct multipath *m = pg->m;
1425         unsigned long flags;
1426         bool delay_retry = false;
1427
1428         /* device or driver problems */
1429         switch (errors) {
1430         case SCSI_DH_OK:
1431                 break;
1432         case SCSI_DH_NOSYS:
1433                 if (!m->hw_handler_name) {
1434                         errors = 0;
1435                         break;
1436                 }
1437                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1438                       "Error %d.", m->hw_handler_name, errors);
1439                 /*
1440                  * Fail path for now, so we do not ping pong
1441                  */
1442                 fail_path(pgpath);
1443                 break;
1444         case SCSI_DH_DEV_TEMP_BUSY:
1445                 /*
1446                  * Probably doing something like FW upgrade on the
1447                  * controller so try the other pg.
1448                  */
1449                 bypass_pg(m, pg, true);
1450                 break;
1451         case SCSI_DH_RETRY:
1452                 /* Wait before retrying. */
1453                 delay_retry = 1;
1454         case SCSI_DH_IMM_RETRY:
1455         case SCSI_DH_RES_TEMP_UNAVAIL:
1456                 if (pg_init_limit_reached(m, pgpath))
1457                         fail_path(pgpath);
1458                 errors = 0;
1459                 break;
1460         case SCSI_DH_DEV_OFFLINED:
1461         default:
1462                 /*
1463                  * We probably do not want to fail the path for a device
1464                  * error, but this is what the old dm did. In future
1465                  * patches we can do more advanced handling.
1466                  */
1467                 fail_path(pgpath);
1468         }
1469
1470         spin_lock_irqsave(&m->lock, flags);
1471         if (errors) {
1472                 if (pgpath == m->current_pgpath) {
1473                         DMERR("Could not failover device. Error %d.", errors);
1474                         m->current_pgpath = NULL;
1475                         m->current_pg = NULL;
1476                 }
1477         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1478                 pg->bypassed = false;
1479
1480         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1481                 /* Activations of other paths are still on going */
1482                 goto out;
1483
1484         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1485                 if (delay_retry)
1486                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1487                 else
1488                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1489
1490                 if (__pg_init_all_paths(m))
1491                         goto out;
1492         }
1493         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1494
1495         process_queued_bios_list(m);
1496
1497         /*
1498          * Wake up any thread waiting to suspend.
1499          */
1500         wake_up(&m->pg_init_wait);
1501
1502 out:
1503         spin_unlock_irqrestore(&m->lock, flags);
1504 }
1505
1506 static void activate_path(struct work_struct *work)
1507 {
1508         struct pgpath *pgpath =
1509                 container_of(work, struct pgpath, activate_path.work);
1510
1511         if (pgpath->is_active)
1512                 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1513                                  pg_init_done, pgpath);
1514         else
1515                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1516 }
1517
1518 static int noretry_error(int error)
1519 {
1520         switch (error) {
1521         case -EOPNOTSUPP:
1522         case -EREMOTEIO:
1523         case -EILSEQ:
1524         case -ENODATA:
1525         case -ENOSPC:
1526                 return 1;
1527         }
1528
1529         /* Anything else could be a path failure, so should be retried */
1530         return 0;
1531 }
1532
1533 /*
1534  * end_io handling
1535  */
1536 static int do_end_io(struct multipath *m, struct request *clone,
1537                      int error, struct dm_mpath_io *mpio)
1538 {
1539         /*
1540          * We don't queue any clone request inside the multipath target
1541          * during end I/O handling, since those clone requests don't have
1542          * bio clones.  If we queue them inside the multipath target,
1543          * we need to make bio clones, that requires memory allocation.
1544          * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1545          *  don't have bio clones.)
1546          * Instead of queueing the clone request here, we queue the original
1547          * request into dm core, which will remake a clone request and
1548          * clone bios for it and resubmit it later.
1549          */
1550         int r = DM_ENDIO_REQUEUE;
1551
1552         if (!error && !clone->errors)
1553                 return 0;       /* I/O complete */
1554
1555         if (noretry_error(error))
1556                 return error;
1557
1558         if (mpio->pgpath)
1559                 fail_path(mpio->pgpath);
1560
1561         if (!atomic_read(&m->nr_valid_paths)) {
1562                 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1563                         if (!must_push_back_rq(m))
1564                                 r = -EIO;
1565                 } else {
1566                         if (error == -EBADE)
1567                                 r = error;
1568                 }
1569         }
1570
1571         return r;
1572 }
1573
1574 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1575                             int error, union map_info *map_context)
1576 {
1577         struct multipath *m = ti->private;
1578         struct dm_mpath_io *mpio = get_mpio(map_context);
1579         struct pgpath *pgpath;
1580         struct path_selector *ps;
1581         int r;
1582
1583         BUG_ON(!mpio);
1584
1585         r = do_end_io(m, clone, error, mpio);
1586         pgpath = mpio->pgpath;
1587         if (pgpath) {
1588                 ps = &pgpath->pg->ps;
1589                 if (ps->type->end_io)
1590                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1591         }
1592         clear_request_fn_mpio(m, map_context);
1593
1594         return r;
1595 }
1596
1597 static int do_end_io_bio(struct multipath *m, struct bio *clone,
1598                          int error, struct dm_mpath_io *mpio)
1599 {
1600         unsigned long flags;
1601
1602         if (!error)
1603                 return 0;       /* I/O complete */
1604
1605         if (noretry_error(error))
1606                 return error;
1607
1608         if (mpio->pgpath)
1609                 fail_path(mpio->pgpath);
1610
1611         if (!atomic_read(&m->nr_valid_paths)) {
1612                 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1613                         if (!must_push_back_bio(m))
1614                                 return -EIO;
1615                         return DM_ENDIO_REQUEUE;
1616                 } else {
1617                         if (error == -EBADE)
1618                                 return error;
1619                 }
1620         }
1621
1622         /* Queue for the daemon to resubmit */
1623         dm_bio_restore(get_bio_details_from_bio(clone), clone);
1624
1625         spin_lock_irqsave(&m->lock, flags);
1626         bio_list_add(&m->queued_bios, clone);
1627         spin_unlock_irqrestore(&m->lock, flags);
1628         if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1629                 queue_work(kmultipathd, &m->process_queued_bios);
1630
1631         return DM_ENDIO_INCOMPLETE;
1632 }
1633
1634 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone, int error)
1635 {
1636         struct multipath *m = ti->private;
1637         struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1638         struct pgpath *pgpath;
1639         struct path_selector *ps;
1640         int r;
1641
1642         BUG_ON(!mpio);
1643
1644         r = do_end_io_bio(m, clone, error, mpio);
1645         pgpath = mpio->pgpath;
1646         if (pgpath) {
1647                 ps = &pgpath->pg->ps;
1648                 if (ps->type->end_io)
1649                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1650         }
1651
1652         return r;
1653 }
1654
1655 /*
1656  * Suspend can't complete until all the I/O is processed so if
1657  * the last path fails we must error any remaining I/O.
1658  * Note that if the freeze_bdev fails while suspending, the
1659  * queue_if_no_path state is lost - userspace should reset it.
1660  */
1661 static void multipath_presuspend(struct dm_target *ti)
1662 {
1663         struct multipath *m = ti->private;
1664
1665         queue_if_no_path(m, false, true);
1666 }
1667
1668 static void multipath_postsuspend(struct dm_target *ti)
1669 {
1670         struct multipath *m = ti->private;
1671
1672         mutex_lock(&m->work_mutex);
1673         flush_multipath_work(m);
1674         mutex_unlock(&m->work_mutex);
1675 }
1676
1677 /*
1678  * Restore the queue_if_no_path setting.
1679  */
1680 static void multipath_resume(struct dm_target *ti)
1681 {
1682         struct multipath *m = ti->private;
1683
1684         if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags))
1685                 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1686         else
1687                 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1688         smp_mb__after_atomic();
1689 }
1690
1691 /*
1692  * Info output has the following format:
1693  * num_multipath_feature_args [multipath_feature_args]*
1694  * num_handler_status_args [handler_status_args]*
1695  * num_groups init_group_number
1696  *            [A|D|E num_ps_status_args [ps_status_args]*
1697  *             num_paths num_selector_args
1698  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1699  *
1700  * Table output has the following format (identical to the constructor string):
1701  * num_feature_args [features_args]*
1702  * num_handler_args hw_handler [hw_handler_args]*
1703  * num_groups init_group_number
1704  *     [priority selector-name num_ps_args [ps_args]*
1705  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1706  */
1707 static void multipath_status(struct dm_target *ti, status_type_t type,
1708                              unsigned status_flags, char *result, unsigned maxlen)
1709 {
1710         int sz = 0;
1711         unsigned long flags;
1712         struct multipath *m = ti->private;
1713         struct priority_group *pg;
1714         struct pgpath *p;
1715         unsigned pg_num;
1716         char state;
1717
1718         spin_lock_irqsave(&m->lock, flags);
1719
1720         /* Features */
1721         if (type == STATUSTYPE_INFO)
1722                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1723                        atomic_read(&m->pg_init_count));
1724         else {
1725                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1726                               (m->pg_init_retries > 0) * 2 +
1727                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1728                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1729                               (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1730
1731                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1732                         DMEMIT("queue_if_no_path ");
1733                 if (m->pg_init_retries)
1734                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1735                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1736                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1737                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1738                         DMEMIT("retain_attached_hw_handler ");
1739                 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1740                         switch(m->queue_mode) {
1741                         case DM_TYPE_BIO_BASED:
1742                                 DMEMIT("queue_mode bio ");
1743                                 break;
1744                         case DM_TYPE_MQ_REQUEST_BASED:
1745                                 DMEMIT("queue_mode mq ");
1746                                 break;
1747                         }
1748                 }
1749         }
1750
1751         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1752                 DMEMIT("0 ");
1753         else
1754                 DMEMIT("1 %s ", m->hw_handler_name);
1755
1756         DMEMIT("%u ", m->nr_priority_groups);
1757
1758         if (m->next_pg)
1759                 pg_num = m->next_pg->pg_num;
1760         else if (m->current_pg)
1761                 pg_num = m->current_pg->pg_num;
1762         else
1763                 pg_num = (m->nr_priority_groups ? 1 : 0);
1764
1765         DMEMIT("%u ", pg_num);
1766
1767         switch (type) {
1768         case STATUSTYPE_INFO:
1769                 list_for_each_entry(pg, &m->priority_groups, list) {
1770                         if (pg->bypassed)
1771                                 state = 'D';    /* Disabled */
1772                         else if (pg == m->current_pg)
1773                                 state = 'A';    /* Currently Active */
1774                         else
1775                                 state = 'E';    /* Enabled */
1776
1777                         DMEMIT("%c ", state);
1778
1779                         if (pg->ps.type->status)
1780                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1781                                                           result + sz,
1782                                                           maxlen - sz);
1783                         else
1784                                 DMEMIT("0 ");
1785
1786                         DMEMIT("%u %u ", pg->nr_pgpaths,
1787                                pg->ps.type->info_args);
1788
1789                         list_for_each_entry(p, &pg->pgpaths, list) {
1790                                 DMEMIT("%s %s %u ", p->path.dev->name,
1791                                        p->is_active ? "A" : "F",
1792                                        p->fail_count);
1793                                 if (pg->ps.type->status)
1794                                         sz += pg->ps.type->status(&pg->ps,
1795                                               &p->path, type, result + sz,
1796                                               maxlen - sz);
1797                         }
1798                 }
1799                 break;
1800
1801         case STATUSTYPE_TABLE:
1802                 list_for_each_entry(pg, &m->priority_groups, list) {
1803                         DMEMIT("%s ", pg->ps.type->name);
1804
1805                         if (pg->ps.type->status)
1806                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1807                                                           result + sz,
1808                                                           maxlen - sz);
1809                         else
1810                                 DMEMIT("0 ");
1811
1812                         DMEMIT("%u %u ", pg->nr_pgpaths,
1813                                pg->ps.type->table_args);
1814
1815                         list_for_each_entry(p, &pg->pgpaths, list) {
1816                                 DMEMIT("%s ", p->path.dev->name);
1817                                 if (pg->ps.type->status)
1818                                         sz += pg->ps.type->status(&pg->ps,
1819                                               &p->path, type, result + sz,
1820                                               maxlen - sz);
1821                         }
1822                 }
1823                 break;
1824         }
1825
1826         spin_unlock_irqrestore(&m->lock, flags);
1827 }
1828
1829 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1830 {
1831         int r = -EINVAL;
1832         struct dm_dev *dev;
1833         struct multipath *m = ti->private;
1834         action_fn action;
1835
1836         mutex_lock(&m->work_mutex);
1837
1838         if (dm_suspended(ti)) {
1839                 r = -EBUSY;
1840                 goto out;
1841         }
1842
1843         if (argc == 1) {
1844                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1845                         r = queue_if_no_path(m, true, false);
1846                         goto out;
1847                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1848                         r = queue_if_no_path(m, false, false);
1849                         goto out;
1850                 }
1851         }
1852
1853         if (argc != 2) {
1854                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1855                 goto out;
1856         }
1857
1858         if (!strcasecmp(argv[0], "disable_group")) {
1859                 r = bypass_pg_num(m, argv[1], true);
1860                 goto out;
1861         } else if (!strcasecmp(argv[0], "enable_group")) {
1862                 r = bypass_pg_num(m, argv[1], false);
1863                 goto out;
1864         } else if (!strcasecmp(argv[0], "switch_group")) {
1865                 r = switch_pg_num(m, argv[1]);
1866                 goto out;
1867         } else if (!strcasecmp(argv[0], "reinstate_path"))
1868                 action = reinstate_path;
1869         else if (!strcasecmp(argv[0], "fail_path"))
1870                 action = fail_path;
1871         else {
1872                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1873                 goto out;
1874         }
1875
1876         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1877         if (r) {
1878                 DMWARN("message: error getting device %s",
1879                        argv[1]);
1880                 goto out;
1881         }
1882
1883         r = action_dev(m, dev, action);
1884
1885         dm_put_device(ti, dev);
1886
1887 out:
1888         mutex_unlock(&m->work_mutex);
1889         return r;
1890 }
1891
1892 static int multipath_prepare_ioctl(struct dm_target *ti,
1893                 struct block_device **bdev, fmode_t *mode)
1894 {
1895         struct multipath *m = ti->private;
1896         struct pgpath *current_pgpath;
1897         int r;
1898
1899         current_pgpath = lockless_dereference(m->current_pgpath);
1900         if (!current_pgpath)
1901                 current_pgpath = choose_pgpath(m, 0);
1902
1903         if (current_pgpath) {
1904                 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1905                         *bdev = current_pgpath->path.dev->bdev;
1906                         *mode = current_pgpath->path.dev->mode;
1907                         r = 0;
1908                 } else {
1909                         /* pg_init has not started or completed */
1910                         r = -ENOTCONN;
1911                 }
1912         } else {
1913                 /* No path is available */
1914                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1915                         r = -ENOTCONN;
1916                 else
1917                         r = -EIO;
1918         }
1919
1920         if (r == -ENOTCONN) {
1921                 if (!lockless_dereference(m->current_pg)) {
1922                         /* Path status changed, redo selection */
1923                         (void) choose_pgpath(m, 0);
1924                 }
1925                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1926                         pg_init_all_paths(m);
1927                 dm_table_run_md_queue_async(m->ti->table);
1928                 process_queued_bios_list(m);
1929         }
1930
1931         /*
1932          * Only pass ioctls through if the device sizes match exactly.
1933          */
1934         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1935                 return 1;
1936         return r;
1937 }
1938
1939 static int multipath_iterate_devices(struct dm_target *ti,
1940                                      iterate_devices_callout_fn fn, void *data)
1941 {
1942         struct multipath *m = ti->private;
1943         struct priority_group *pg;
1944         struct pgpath *p;
1945         int ret = 0;
1946
1947         list_for_each_entry(pg, &m->priority_groups, list) {
1948                 list_for_each_entry(p, &pg->pgpaths, list) {
1949                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1950                         if (ret)
1951                                 goto out;
1952                 }
1953         }
1954
1955 out:
1956         return ret;
1957 }
1958
1959 static int pgpath_busy(struct pgpath *pgpath)
1960 {
1961         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1962
1963         return blk_lld_busy(q);
1964 }
1965
1966 /*
1967  * We return "busy", only when we can map I/Os but underlying devices
1968  * are busy (so even if we map I/Os now, the I/Os will wait on
1969  * the underlying queue).
1970  * In other words, if we want to kill I/Os or queue them inside us
1971  * due to map unavailability, we don't return "busy".  Otherwise,
1972  * dm core won't give us the I/Os and we can't do what we want.
1973  */
1974 static int multipath_busy(struct dm_target *ti)
1975 {
1976         bool busy = false, has_active = false;
1977         struct multipath *m = ti->private;
1978         struct priority_group *pg, *next_pg;
1979         struct pgpath *pgpath;
1980
1981         /* pg_init in progress or no paths available */
1982         if (atomic_read(&m->pg_init_in_progress) ||
1983             (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)))
1984                 return true;
1985
1986         /* Guess which priority_group will be used at next mapping time */
1987         pg = lockless_dereference(m->current_pg);
1988         next_pg = lockless_dereference(m->next_pg);
1989         if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
1990                 pg = next_pg;
1991
1992         if (!pg) {
1993                 /*
1994                  * We don't know which pg will be used at next mapping time.
1995                  * We don't call choose_pgpath() here to avoid to trigger
1996                  * pg_init just by busy checking.
1997                  * So we don't know whether underlying devices we will be using
1998                  * at next mapping time are busy or not. Just try mapping.
1999                  */
2000                 return busy;
2001         }
2002
2003         /*
2004          * If there is one non-busy active path at least, the path selector
2005          * will be able to select it. So we consider such a pg as not busy.
2006          */
2007         busy = true;
2008         list_for_each_entry(pgpath, &pg->pgpaths, list) {
2009                 if (pgpath->is_active) {
2010                         has_active = true;
2011                         if (!pgpath_busy(pgpath)) {
2012                                 busy = false;
2013                                 break;
2014                         }
2015                 }
2016         }
2017
2018         if (!has_active) {
2019                 /*
2020                  * No active path in this pg, so this pg won't be used and
2021                  * the current_pg will be changed at next mapping time.
2022                  * We need to try mapping to determine it.
2023                  */
2024                 busy = false;
2025         }
2026
2027         return busy;
2028 }
2029
2030 /*-----------------------------------------------------------------
2031  * Module setup
2032  *---------------------------------------------------------------*/
2033 static struct target_type multipath_target = {
2034         .name = "multipath",
2035         .version = {1, 12, 0},
2036         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
2037         .module = THIS_MODULE,
2038         .ctr = multipath_ctr,
2039         .dtr = multipath_dtr,
2040         .map_rq = multipath_map,
2041         .clone_and_map_rq = multipath_clone_and_map,
2042         .release_clone_rq = multipath_release_clone,
2043         .rq_end_io = multipath_end_io,
2044         .map = multipath_map_bio,
2045         .end_io = multipath_end_io_bio,
2046         .presuspend = multipath_presuspend,
2047         .postsuspend = multipath_postsuspend,
2048         .resume = multipath_resume,
2049         .status = multipath_status,
2050         .message = multipath_message,
2051         .prepare_ioctl = multipath_prepare_ioctl,
2052         .iterate_devices = multipath_iterate_devices,
2053         .busy = multipath_busy,
2054 };
2055
2056 static int __init dm_multipath_init(void)
2057 {
2058         int r;
2059
2060         /* allocate a slab for the dm_mpath_ios */
2061         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
2062         if (!_mpio_cache)
2063                 return -ENOMEM;
2064
2065         r = dm_register_target(&multipath_target);
2066         if (r < 0) {
2067                 DMERR("request-based register failed %d", r);
2068                 r = -EINVAL;
2069                 goto bad_register_target;
2070         }
2071
2072         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2073         if (!kmultipathd) {
2074                 DMERR("failed to create workqueue kmpathd");
2075                 r = -ENOMEM;
2076                 goto bad_alloc_kmultipathd;
2077         }
2078
2079         /*
2080          * A separate workqueue is used to handle the device handlers
2081          * to avoid overloading existing workqueue. Overloading the
2082          * old workqueue would also create a bottleneck in the
2083          * path of the storage hardware device activation.
2084          */
2085         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2086                                                   WQ_MEM_RECLAIM);
2087         if (!kmpath_handlerd) {
2088                 DMERR("failed to create workqueue kmpath_handlerd");
2089                 r = -ENOMEM;
2090                 goto bad_alloc_kmpath_handlerd;
2091         }
2092
2093         return 0;
2094
2095 bad_alloc_kmpath_handlerd:
2096         destroy_workqueue(kmultipathd);
2097 bad_alloc_kmultipathd:
2098         dm_unregister_target(&multipath_target);
2099 bad_register_target:
2100         kmem_cache_destroy(_mpio_cache);
2101
2102         return r;
2103 }
2104
2105 static void __exit dm_multipath_exit(void)
2106 {
2107         destroy_workqueue(kmpath_handlerd);
2108         destroy_workqueue(kmultipathd);
2109
2110         dm_unregister_target(&multipath_target);
2111         kmem_cache_destroy(_mpio_cache);
2112 }
2113
2114 module_init(dm_multipath_init);
2115 module_exit(dm_multipath_exit);
2116
2117 MODULE_DESCRIPTION(DM_NAME " multipath target");
2118 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2119 MODULE_LICENSE("GPL");