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         unsigned long flags;            /* Multipath state flags */
  68
  69         spinlock_t lock;
  70         enum dm_queue_mode queue_mode;
  71
  72         struct pgpath *current_pgpath;
  73         struct priority_group *current_pg;
  74         struct priority_group *next_pg; /* Switch to this PG if set */
  75
  76         atomic_t nr_valid_paths;        /* Total number of usable paths */
  77         unsigned nr_priority_groups;
  78         struct list_head priority_groups;
  79
  80         const char *hw_handler_name;
  81         char *hw_handler_params;
  82         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  83         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  84         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  85         atomic_t pg_init_in_progress;   /* Only one pg_init allowed at once */
  86         atomic_t pg_init_count;         /* Number of times pg_init called */
  87
  88         struct mutex work_mutex;
  89         struct work_struct trigger_event;
  90         struct dm_target *ti;
  91
  92         struct work_struct process_queued_bios;
  93         struct bio_list queued_bios;
  94 };
  95
  96 /*
  97  * Context information attached to each io we process.
  98  */
  99 struct dm_mpath_io {
 100         struct pgpath *pgpath;
 101         size_t nr_bytes;
 102 };
 103
 104 typedef int (*action_fn) (struct pgpath *pgpath);
 105
 106 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 107 static void trigger_event(struct work_struct *work);
 108 static void activate_or_offline_path(struct pgpath *pgpath);
 109 static void activate_path_work(struct work_struct *work);
 110 static void process_queued_bios(struct work_struct *work);
 111
 112 /*-----------------------------------------------
 113  * Multipath state flags.
 114  *-----------------------------------------------*/
 115
 116 #define MPATHF_QUEUE_IO 0                       /* Must we queue all I/O? */
 117 #define MPATHF_QUEUE_IF_NO_PATH 1               /* Queue I/O if last path fails? */
 118 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2         /* Saved state during suspension */
 119 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3     /* If there's already a hw_handler present, don't change it. */
 120 #define MPATHF_PG_INIT_DISABLED 4               /* pg_init is not currently allowed */
 121 #define MPATHF_PG_INIT_REQUIRED 5               /* pg_init needs calling? */
 122 #define MPATHF_PG_INIT_DELAY_RETRY 6            /* Delay pg_init retry? */
 123
 124 /*-----------------------------------------------
 125  * Allocation routines
 126  *-----------------------------------------------*/
 127
 128 static struct pgpath *alloc_pgpath(void)
 129 {
 130         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 131
 132         if (!pgpath)
 133                 return NULL;
 134
 135         pgpath->is_active = true;
 136
 137         return pgpath;
 138 }
 139
 140 static void free_pgpath(struct pgpath *pgpath)
 141 {
 142         kfree(pgpath);
 143 }
 144
 145 static struct priority_group *alloc_priority_group(void)
 146 {
 147         struct priority_group *pg;
 148
 149         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 150
 151         if (pg)
 152                 INIT_LIST_HEAD(&pg->pgpaths);
 153
 154         return pg;
 155 }
 156
 157 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 158 {
 159         struct pgpath *pgpath, *tmp;
 160
 161         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 162                 list_del(&pgpath->list);
 163                 dm_put_device(ti, pgpath->path.dev);
 164                 free_pgpath(pgpath);
 165         }
 166 }
 167
 168 static void free_priority_group(struct priority_group *pg,
 169                                 struct dm_target *ti)
 170 {
 171         struct path_selector *ps = &pg->ps;
 172
 173         if (ps->type) {
 174                 ps->type->destroy(ps);
 175                 dm_put_path_selector(ps->type);
 176         }
 177
 178         free_pgpaths(&pg->pgpaths, ti);
 179         kfree(pg);
 180 }
 181
 182 static struct multipath *alloc_multipath(struct dm_target *ti)
 183 {
 184         struct multipath *m;
 185
 186         m = kzalloc(sizeof(*m), GFP_KERNEL);
 187         if (m) {
 188                 INIT_LIST_HEAD(&m->priority_groups);
 189                 spin_lock_init(&m->lock);
 190                 atomic_set(&m->nr_valid_paths, 0);
 191                 INIT_WORK(&m->trigger_event, trigger_event);
 192                 mutex_init(&m->work_mutex);
 193
 194                 m->queue_mode = DM_TYPE_NONE;
 195
 196                 m->ti = ti;
 197                 ti->private = m;
 198         }
 199
 200         return m;
 201 }
 202
 203 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
 204 {
 205         if (m->queue_mode == DM_TYPE_NONE) {
 206                 m->queue_mode = DM_TYPE_REQUEST_BASED;
 207         } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
 208                 INIT_WORK(&m->process_queued_bios, process_queued_bios);
 209                 /*
 210                  * bio-based doesn't support any direct scsi_dh management;
 211                  * it just discovers if a scsi_dh is attached.
 212                  */
 213                 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 214         }
 215
 216         dm_table_set_type(ti->table, m->queue_mode);
 217
 218         /*
 219          * Init fields that are only used when a scsi_dh is attached
 220          * - must do this unconditionally (really doesn't hurt non-SCSI uses)
 221          */
 222         set_bit(MPATHF_QUEUE_IO, &m->flags);
 223         atomic_set(&m->pg_init_in_progress, 0);
 224         atomic_set(&m->pg_init_count, 0);
 225         m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 226         init_waitqueue_head(&m->pg_init_wait);
 227
 228         return 0;
 229 }
 230
 231 static void free_multipath(struct multipath *m)
 232 {
 233         struct priority_group *pg, *tmp;
 234
 235         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 236                 list_del(&pg->list);
 237                 free_priority_group(pg, m->ti);
 238         }
 239
 240         kfree(m->hw_handler_name);
 241         kfree(m->hw_handler_params);
 242         mutex_destroy(&m->work_mutex);
 243         kfree(m);
 244 }
 245
 246 static struct dm_mpath_io *get_mpio(union map_info *info)
 247 {
 248         return info->ptr;
 249 }
 250
 251 static size_t multipath_per_bio_data_size(void)
 252 {
 253         return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
 254 }
 255
 256 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
 257 {
 258         return dm_per_bio_data(bio, multipath_per_bio_data_size());
 259 }
 260
 261 static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
 262 {
 263         /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
 264         void *bio_details = mpio + 1;
 265         return bio_details;
 266 }
 267
 268 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
 269 {
 270         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 271         struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
 272
 273         mpio->nr_bytes = bio->bi_iter.bi_size;
 274         mpio->pgpath = NULL;
 275         *mpio_p = mpio;
 276
 277         dm_bio_record(bio_details, bio);
 278 }
 279
 280 /*-----------------------------------------------
 281  * Path selection
 282  *-----------------------------------------------*/
 283
 284 static int __pg_init_all_paths(struct multipath *m)
 285 {
 286         struct pgpath *pgpath;
 287         unsigned long pg_init_delay = 0;
 288
 289         lockdep_assert_held(&m->lock);
 290
 291         if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
 292                 return 0;
 293
 294         atomic_inc(&m->pg_init_count);
 295         clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 296
 297         /* Check here to reset pg_init_required */
 298         if (!m->current_pg)
 299                 return 0;
 300
 301         if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
 302                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 303                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 304         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 305                 /* Skip failed paths */
 306                 if (!pgpath->is_active)
 307                         continue;
 308                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 309                                        pg_init_delay))
 310                         atomic_inc(&m->pg_init_in_progress);
 311         }
 312         return atomic_read(&m->pg_init_in_progress);
 313 }
 314
 315 static int pg_init_all_paths(struct multipath *m)
 316 {
 317         int ret;
 318         unsigned long flags;
 319
 320         spin_lock_irqsave(&m->lock, flags);
 321         ret = __pg_init_all_paths(m);
 322         spin_unlock_irqrestore(&m->lock, flags);
 323
 324         return ret;
 325 }
 326
 327 static void __switch_pg(struct multipath *m, struct priority_group *pg)
 328 {
 329         m->current_pg = pg;
 330
 331         /* Must we initialise the PG first, and queue I/O till it's ready? */
 332         if (m->hw_handler_name) {
 333                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 334                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 335         } else {
 336                 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 337                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 338         }
 339
 340         atomic_set(&m->pg_init_count, 0);
 341 }
 342
 343 static struct pgpath *choose_path_in_pg(struct multipath *m,
 344                                         struct priority_group *pg,
 345                                         size_t nr_bytes)
 346 {
 347         unsigned long flags;
 348         struct dm_path *path;
 349         struct pgpath *pgpath;
 350
 351         path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 352         if (!path)
 353                 return ERR_PTR(-ENXIO);
 354
 355         pgpath = path_to_pgpath(path);
 356
 357         if (unlikely(READ_ONCE(m->current_pg) != pg)) {
 358                 /* Only update current_pgpath if pg changed */
 359                 spin_lock_irqsave(&m->lock, flags);
 360                 m->current_pgpath = pgpath;
 361                 __switch_pg(m, pg);
 362                 spin_unlock_irqrestore(&m->lock, flags);
 363         }
 364
 365         return pgpath;
 366 }
 367
 368 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
 369 {
 370         unsigned long flags;
 371         struct priority_group *pg;
 372         struct pgpath *pgpath;
 373         unsigned bypassed = 1;
 374
 375         if (!atomic_read(&m->nr_valid_paths)) {
 376                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 377                 goto failed;
 378         }
 379
 380         /* Were we instructed to switch PG? */
 381         if (READ_ONCE(m->next_pg)) {
 382                 spin_lock_irqsave(&m->lock, flags);
 383                 pg = m->next_pg;
 384                 if (!pg) {
 385                         spin_unlock_irqrestore(&m->lock, flags);
 386                         goto check_current_pg;
 387                 }
 388                 m->next_pg = NULL;
 389                 spin_unlock_irqrestore(&m->lock, flags);
 390                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 391                 if (!IS_ERR_OR_NULL(pgpath))
 392                         return pgpath;
 393         }
 394
 395         /* Don't change PG until it has no remaining paths */
 396 check_current_pg:
 397         pg = READ_ONCE(m->current_pg);
 398         if (pg) {
 399                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 400                 if (!IS_ERR_OR_NULL(pgpath))
 401                         return pgpath;
 402         }
 403
 404         /*
 405          * Loop through priority groups until we find a valid path.
 406          * First time we skip PGs marked 'bypassed'.
 407          * Second time we only try the ones we skipped, but set
 408          * pg_init_delay_retry so we do not hammer controllers.
 409          */
 410         do {
 411                 list_for_each_entry(pg, &m->priority_groups, list) {
 412                         if (pg->bypassed == !!bypassed)
 413                                 continue;
 414                         pgpath = choose_path_in_pg(m, pg, nr_bytes);
 415                         if (!IS_ERR_OR_NULL(pgpath)) {
 416                                 if (!bypassed)
 417                                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
 418                                 return pgpath;
 419                         }
 420                 }
 421         } while (bypassed--);
 422
 423 failed:
 424         spin_lock_irqsave(&m->lock, flags);
 425         m->current_pgpath = NULL;
 426         m->current_pg = NULL;
 427         spin_unlock_irqrestore(&m->lock, flags);
 428
 429         return NULL;
 430 }
 431
 432 /*
 433  * dm_report_EIO() is a macro instead of a function to make pr_debug()
 434  * report the function name and line number of the function from which
 435  * it has been invoked.
 436  */
 437 #define dm_report_EIO(m)                                                \
 438 do {                                                                    \
 439         struct mapped_device *md = dm_table_get_md((m)->ti->table);     \
 440                                                                         \
 441         pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
 442                  dm_device_name(md),                                    \
 443                  test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags),        \
 444                  test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags),  \
 445                  dm_noflush_suspending((m)->ti));                       \
 446 } while (0)
 447
 448 /*
 449  * Check whether bios must be queued in the device-mapper core rather
 450  * than here in the target.
 451  *
 452  * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
 453  * the same value then we are not between multipath_presuspend()
 454  * and multipath_resume() calls and we have no need to check
 455  * for the DMF_NOFLUSH_SUSPENDING flag.
 456  */
 457 static bool __must_push_back(struct multipath *m, unsigned long flags)
 458 {
 459         return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
 460                  test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
 461                 dm_noflush_suspending(m->ti));
 462 }
 463
 464 /*
 465  * Following functions use READ_ONCE to get atomic access to
 466  * all m->flags to avoid taking spinlock
 467  */
 468 static bool must_push_back_rq(struct multipath *m)
 469 {
 470         unsigned long flags = READ_ONCE(m->flags);
 471         return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
 472 }
 473
 474 static bool must_push_back_bio(struct multipath *m)
 475 {
 476         unsigned long flags = READ_ONCE(m->flags);
 477         return __must_push_back(m, flags);
 478 }
 479
 480 /*
 481  * Map cloned requests (request-based multipath)
 482  */
 483 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 484                                    union map_info *map_context,
 485                                    struct request **__clone)
 486 {
 487         struct multipath *m = ti->private;
 488         size_t nr_bytes = blk_rq_bytes(rq);
 489         struct pgpath *pgpath;
 490         struct block_device *bdev;
 491         struct dm_mpath_io *mpio = get_mpio(map_context);
 492         struct request_queue *q;
 493         struct request *clone;
 494
 495         /* Do we need to select a new pgpath? */
 496         pgpath = READ_ONCE(m->current_pgpath);
 497         if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
 498                 pgpath = choose_pgpath(m, nr_bytes);
 499
 500         if (!pgpath) {
 501                 if (must_push_back_rq(m))
 502                         return DM_MAPIO_DELAY_REQUEUE;
 503                 dm_report_EIO(m);       /* Failed */
 504                 return DM_MAPIO_KILL;
 505         } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
 506                    test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
 507                 pg_init_all_paths(m);
 508                 return DM_MAPIO_DELAY_REQUEUE;
 509         }
 510
 511         mpio->pgpath = pgpath;
 512         mpio->nr_bytes = nr_bytes;
 513
 514         bdev = pgpath->path.dev->bdev;
 515         q = bdev_get_queue(bdev);
 516         clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE,
 517                         BLK_MQ_REQ_NOWAIT);
 518         if (IS_ERR(clone)) {
 519                 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
 520                 if (blk_queue_dying(q)) {
 521                         atomic_inc(&m->pg_init_in_progress);
 522                         activate_or_offline_path(pgpath);
 523                         return DM_MAPIO_DELAY_REQUEUE;
 524                 }
 525
 526                 /*
 527                  * blk-mq's SCHED_RESTART can cover this requeue, so we
 528                  * needn't deal with it by DELAY_REQUEUE. More importantly,
 529                  * we have to return DM_MAPIO_REQUEUE so that blk-mq can
 530                  * get the queue busy feedback (via BLK_STS_RESOURCE),
 531                  * otherwise I/O merging can suffer.
 532                  */
 533                 return DM_MAPIO_REQUEUE;
 534         }
 535         clone->bio = clone->biotail = NULL;
 536         clone->rq_disk = bdev->bd_disk;
 537         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 538         *__clone = clone;
 539
 540         if (pgpath->pg->ps.type->start_io)
 541                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 542                                               &pgpath->path,
 543                                               nr_bytes);
 544         return DM_MAPIO_REMAPPED;
 545 }
 546
 547 static void multipath_release_clone(struct request *clone,
 548                                     union map_info *map_context)
 549 {
 550         if (unlikely(map_context)) {
 551                 /*
 552                  * non-NULL map_context means caller is still map
 553                  * method; must undo multipath_clone_and_map()
 554                  */
 555                 struct dm_mpath_io *mpio = get_mpio(map_context);
 556                 struct pgpath *pgpath = mpio->pgpath;
 557
 558                 if (pgpath && pgpath->pg->ps.type->end_io)
 559                         pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
 560                                                     &pgpath->path,
 561                                                     mpio->nr_bytes);
 562         }
 563
 564         blk_put_request(clone);
 565 }
 566
 567 /*
 568  * Map cloned bios (bio-based multipath)
 569  */
 570
 571 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
 572 {
 573         struct pgpath *pgpath;
 574         unsigned long flags;
 575         bool queue_io;
 576
 577         /* Do we need to select a new pgpath? */
 578         pgpath = READ_ONCE(m->current_pgpath);
 579         queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
 580         if (!pgpath || !queue_io)
 581                 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
 582
 583         if ((pgpath && queue_io) ||
 584             (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
 585                 /* Queue for the daemon to resubmit */
 586                 spin_lock_irqsave(&m->lock, flags);
 587                 bio_list_add(&m->queued_bios, bio);
 588                 spin_unlock_irqrestore(&m->lock, flags);
 589
 590                 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
 591                 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
 592                         pg_init_all_paths(m);
 593                 else if (!queue_io)
 594                         queue_work(kmultipathd, &m->process_queued_bios);
 595
 596                 return ERR_PTR(-EAGAIN);
 597         }
 598
 599         return pgpath;
 600 }
 601
 602 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
 603                                struct dm_mpath_io *mpio)
 604 {
 605         struct pgpath *pgpath = __map_bio(m, bio);
 606
 607         if (IS_ERR(pgpath))
 608                 return DM_MAPIO_SUBMITTED;
 609
 610         if (!pgpath) {
 611                 if (must_push_back_bio(m))
 612                         return DM_MAPIO_REQUEUE;
 613                 dm_report_EIO(m);
 614                 return DM_MAPIO_KILL;
 615         }
 616
 617         mpio->pgpath = pgpath;
 618
 619         bio->bi_status = 0;
 620         bio_set_dev(bio, pgpath->path.dev->bdev);
 621         bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
 622
 623         if (pgpath->pg->ps.type->start_io)
 624                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 625                                               &pgpath->path,
 626                                               mpio->nr_bytes);
 627         return DM_MAPIO_REMAPPED;
 628 }
 629
 630 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
 631 {
 632         struct multipath *m = ti->private;
 633         struct dm_mpath_io *mpio = NULL;
 634
 635         multipath_init_per_bio_data(bio, &mpio);
 636         return __multipath_map_bio(m, bio, mpio);
 637 }
 638
 639 static void process_queued_io_list(struct multipath *m)
 640 {
 641         if (m->queue_mode == DM_TYPE_REQUEST_BASED)
 642                 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
 643         else if (m->queue_mode == DM_TYPE_BIO_BASED)
 644                 queue_work(kmultipathd, &m->process_queued_bios);
 645 }
 646
 647 static void process_queued_bios(struct work_struct *work)
 648 {
 649         int r;
 650         unsigned long flags;
 651         struct bio *bio;
 652         struct bio_list bios;
 653         struct blk_plug plug;
 654         struct multipath *m =
 655                 container_of(work, struct multipath, process_queued_bios);
 656
 657         bio_list_init(&bios);
 658
 659         spin_lock_irqsave(&m->lock, flags);
 660
 661         if (bio_list_empty(&m->queued_bios)) {
 662                 spin_unlock_irqrestore(&m->lock, flags);
 663                 return;
 664         }
 665
 666         bio_list_merge(&bios, &m->queued_bios);
 667         bio_list_init(&m->queued_bios);
 668
 669         spin_unlock_irqrestore(&m->lock, flags);
 670
 671         blk_start_plug(&plug);
 672         while ((bio = bio_list_pop(&bios))) {
 673                 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 674                 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
 675                 r = __multipath_map_bio(m, bio, mpio);
 676                 switch (r) {
 677                 case DM_MAPIO_KILL:
 678                         bio->bi_status = BLK_STS_IOERR;
 679                         bio_endio(bio);
 680                         break;
 681                 case DM_MAPIO_REQUEUE:
 682                         bio->bi_status = BLK_STS_DM_REQUEUE;
 683                         bio_endio(bio);
 684                         break;
 685                 case DM_MAPIO_REMAPPED:
 686                         generic_make_request(bio);
 687                         break;
 688                 case DM_MAPIO_SUBMITTED:
 689                         break;
 690                 default:
 691                         WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
 692                 }
 693         }
 694         blk_finish_plug(&plug);
 695 }
 696
 697 /*
 698  * If we run out of usable paths, should we queue I/O or error it?
 699  */
 700 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
 701                             bool save_old_value)
 702 {
 703         unsigned long flags;
 704
 705         spin_lock_irqsave(&m->lock, flags);
 706         assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
 707                    (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
 708                    (!save_old_value && queue_if_no_path));
 709         assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
 710         spin_unlock_irqrestore(&m->lock, flags);
 711
 712         if (!queue_if_no_path) {
 713                 dm_table_run_md_queue_async(m->ti->table);
 714                 process_queued_io_list(m);
 715         }
 716
 717         return 0;
 718 }
 719
 720 /*
 721  * An event is triggered whenever a path is taken out of use.
 722  * Includes path failure and PG bypass.
 723  */
 724 static void trigger_event(struct work_struct *work)
 725 {
 726         struct multipath *m =
 727                 container_of(work, struct multipath, trigger_event);
 728
 729         dm_table_event(m->ti->table);
 730 }
 731
 732 /*-----------------------------------------------------------------
 733  * Constructor/argument parsing:
 734  * <#multipath feature args> [<arg>]*
 735  * <#hw_handler args> [hw_handler [<arg>]*]
 736  * <#priority groups>
 737  * <initial priority group>
 738  *     [<selector> <#selector args> [<arg>]*
 739  *      <#paths> <#per-path selector args>
 740  *         [<path> [<arg>]* ]+ ]+
 741  *---------------------------------------------------------------*/
 742 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 743                                struct dm_target *ti)
 744 {
 745         int r;
 746         struct path_selector_type *pst;
 747         unsigned ps_argc;
 748
 749         static const struct dm_arg _args[] = {
 750                 {0, 1024, "invalid number of path selector args"},
 751         };
 752
 753         pst = dm_get_path_selector(dm_shift_arg(as));
 754         if (!pst) {
 755                 ti->error = "unknown path selector type";
 756                 return -EINVAL;
 757         }
 758
 759         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 760         if (r) {
 761                 dm_put_path_selector(pst);
 762                 return -EINVAL;
 763         }
 764
 765         r = pst->create(&pg->ps, ps_argc, as->argv);
 766         if (r) {
 767                 dm_put_path_selector(pst);
 768                 ti->error = "path selector constructor failed";
 769                 return r;
 770         }
 771
 772         pg->ps.type = pst;
 773         dm_consume_args(as, ps_argc);
 774
 775         return 0;
 776 }
 777
 778 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
 779                          const char **attached_handler_name, char **error)
 780 {
 781         struct request_queue *q = bdev_get_queue(bdev);
 782         int r;
 783
 784         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
 785 retain:
 786                 if (*attached_handler_name) {
 787                         /*
 788                          * Clear any hw_handler_params associated with a
 789                          * handler that isn't already attached.
 790                          */
 791                         if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
 792                                 kfree(m->hw_handler_params);
 793                                 m->hw_handler_params = NULL;
 794                         }
 795
 796                         /*
 797                          * Reset hw_handler_name to match the attached handler
 798                          *
 799                          * NB. This modifies the table line to show the actual
 800                          * handler instead of the original table passed in.
 801                          */
 802                         kfree(m->hw_handler_name);
 803                         m->hw_handler_name = *attached_handler_name;
 804                         *attached_handler_name = NULL;
 805                 }
 806         }
 807
 808         if (m->hw_handler_name) {
 809                 r = scsi_dh_attach(q, m->hw_handler_name);
 810                 if (r == -EBUSY) {
 811                         char b[BDEVNAME_SIZE];
 812
 813                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
 814                                bdevname(bdev, b));
 815                         goto retain;
 816                 }
 817                 if (r < 0) {
 818                         *error = "error attaching hardware handler";
 819                         return r;
 820                 }
 821
 822                 if (m->hw_handler_params) {
 823                         r = scsi_dh_set_params(q, m->hw_handler_params);
 824                         if (r < 0) {
 825                                 *error = "unable to set hardware handler parameters";
 826                                 return r;
 827                         }
 828                 }
 829         }
 830
 831         return 0;
 832 }
 833
 834 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 835                                  struct dm_target *ti)
 836 {
 837         int r;
 838         struct pgpath *p;
 839         struct multipath *m = ti->private;
 840         struct request_queue *q;
 841         const char *attached_handler_name = NULL;
 842
 843         /* we need at least a path arg */
 844         if (as->argc < 1) {
 845                 ti->error = "no device given";
 846                 return ERR_PTR(-EINVAL);
 847         }
 848
 849         p = alloc_pgpath();
 850         if (!p)
 851                 return ERR_PTR(-ENOMEM);
 852
 853         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 854                           &p->path.dev);
 855         if (r) {
 856                 ti->error = "error getting device";
 857                 goto bad;
 858         }
 859
 860         q = bdev_get_queue(p->path.dev->bdev);
 861         attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 862         if (attached_handler_name || m->hw_handler_name) {
 863                 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
 864                 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
 865                 kfree(attached_handler_name);
 866                 if (r) {
 867                         dm_put_device(ti, p->path.dev);
 868                         goto bad;
 869                 }
 870         }
 871
 872         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 873         if (r) {
 874                 dm_put_device(ti, p->path.dev);
 875                 goto bad;
 876         }
 877
 878         return p;
 879  bad:
 880         free_pgpath(p);
 881         return ERR_PTR(r);
 882 }
 883
 884 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 885                                                    struct multipath *m)
 886 {
 887         static const struct dm_arg _args[] = {
 888                 {1, 1024, "invalid number of paths"},
 889                 {0, 1024, "invalid number of selector args"}
 890         };
 891
 892         int r;
 893         unsigned i, nr_selector_args, nr_args;
 894         struct priority_group *pg;
 895         struct dm_target *ti = m->ti;
 896
 897         if (as->argc < 2) {
 898                 as->argc = 0;
 899                 ti->error = "not enough priority group arguments";
 900                 return ERR_PTR(-EINVAL);
 901         }
 902
 903         pg = alloc_priority_group();
 904         if (!pg) {
 905                 ti->error = "couldn't allocate priority group";
 906                 return ERR_PTR(-ENOMEM);
 907         }
 908         pg->m = m;
 909
 910         r = parse_path_selector(as, pg, ti);
 911         if (r)
 912                 goto bad;
 913
 914         /*
 915          * read the paths
 916          */
 917         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 918         if (r)
 919                 goto bad;
 920
 921         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 922         if (r)
 923                 goto bad;
 924
 925         nr_args = 1 + nr_selector_args;
 926         for (i = 0; i < pg->nr_pgpaths; i++) {
 927                 struct pgpath *pgpath;
 928                 struct dm_arg_set path_args;
 929
 930                 if (as->argc < nr_args) {
 931                         ti->error = "not enough path parameters";
 932                         r = -EINVAL;
 933                         goto bad;
 934                 }
 935
 936                 path_args.argc = nr_args;
 937                 path_args.argv = as->argv;
 938
 939                 pgpath = parse_path(&path_args, &pg->ps, ti);
 940                 if (IS_ERR(pgpath)) {
 941                         r = PTR_ERR(pgpath);
 942                         goto bad;
 943                 }
 944
 945                 pgpath->pg = pg;
 946                 list_add_tail(&pgpath->list, &pg->pgpaths);
 947                 dm_consume_args(as, nr_args);
 948         }
 949
 950         return pg;
 951
 952  bad:
 953         free_priority_group(pg, ti);
 954         return ERR_PTR(r);
 955 }
 956
 957 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 958 {
 959         unsigned hw_argc;
 960         int ret;
 961         struct dm_target *ti = m->ti;
 962
 963         static const struct dm_arg _args[] = {
 964                 {0, 1024, "invalid number of hardware handler args"},
 965         };
 966
 967         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 968                 return -EINVAL;
 969
 970         if (!hw_argc)
 971                 return 0;
 972
 973         if (m->queue_mode == DM_TYPE_BIO_BASED) {
 974                 dm_consume_args(as, hw_argc);
 975                 DMERR("bio-based multipath doesn't allow hardware handler args");
 976                 return 0;
 977         }
 978
 979         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 980         if (!m->hw_handler_name)
 981                 return -EINVAL;
 982
 983         if (hw_argc > 1) {
 984                 char *p;
 985                 int i, j, len = 4;
 986
 987                 for (i = 0; i <= hw_argc - 2; i++)
 988                         len += strlen(as->argv[i]) + 1;
 989                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 990                 if (!p) {
 991                         ti->error = "memory allocation failed";
 992                         ret = -ENOMEM;
 993                         goto fail;
 994                 }
 995                 j = sprintf(p, "%d", hw_argc - 1);
 996                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 997                         j = sprintf(p, "%s", as->argv[i]);
 998         }
 999         dm_consume_args(as, hw_argc - 1);
1000
1001         return 0;
1002 fail:
1003         kfree(m->hw_handler_name);
1004         m->hw_handler_name = NULL;
1005         return ret;
1006 }
1007
1008 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1009 {
1010         int r;
1011         unsigned argc;
1012         struct dm_target *ti = m->ti;
1013         const char *arg_name;
1014
1015         static const struct dm_arg _args[] = {
1016                 {0, 8, "invalid number of feature args"},
1017                 {1, 50, "pg_init_retries must be between 1 and 50"},
1018                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1019         };
1020
1021         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1022         if (r)
1023                 return -EINVAL;
1024
1025         if (!argc)
1026                 return 0;
1027
1028         do {
1029                 arg_name = dm_shift_arg(as);
1030                 argc--;
1031
1032                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1033                         r = queue_if_no_path(m, true, false);
1034                         continue;
1035                 }
1036
1037                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1038                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1039                         continue;
1040                 }
1041
1042                 if (!strcasecmp(arg_name, "pg_init_retries") &&
1043                     (argc >= 1)) {
1044                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1045                         argc--;
1046                         continue;
1047                 }
1048
1049                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1050                     (argc >= 1)) {
1051                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1052                         argc--;
1053                         continue;
1054                 }
1055
1056                 if (!strcasecmp(arg_name, "queue_mode") &&
1057                     (argc >= 1)) {
1058                         const char *queue_mode_name = dm_shift_arg(as);
1059
1060                         if (!strcasecmp(queue_mode_name, "bio"))
1061                                 m->queue_mode = DM_TYPE_BIO_BASED;
1062                         else if (!strcasecmp(queue_mode_name, "rq") ||
1063                                  !strcasecmp(queue_mode_name, "mq"))
1064                                 m->queue_mode = DM_TYPE_REQUEST_BASED;
1065                         else {
1066                                 ti->error = "Unknown 'queue_mode' requested";
1067                                 r = -EINVAL;
1068                         }
1069                         argc--;
1070                         continue;
1071                 }
1072
1073                 ti->error = "Unrecognised multipath feature request";
1074                 r = -EINVAL;
1075         } while (argc && !r);
1076
1077         return r;
1078 }
1079
1080 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1081 {
1082         /* target arguments */
1083         static const struct dm_arg _args[] = {
1084                 {0, 1024, "invalid number of priority groups"},
1085                 {0, 1024, "invalid initial priority group number"},
1086         };
1087
1088         int r;
1089         struct multipath *m;
1090         struct dm_arg_set as;
1091         unsigned pg_count = 0;
1092         unsigned next_pg_num;
1093
1094         as.argc = argc;
1095         as.argv = argv;
1096
1097         m = alloc_multipath(ti);
1098         if (!m) {
1099                 ti->error = "can't allocate multipath";
1100                 return -EINVAL;
1101         }
1102
1103         r = parse_features(&as, m);
1104         if (r)
1105                 goto bad;
1106
1107         r = alloc_multipath_stage2(ti, m);
1108         if (r)
1109                 goto bad;
1110
1111         r = parse_hw_handler(&as, m);
1112         if (r)
1113                 goto bad;
1114
1115         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1116         if (r)
1117                 goto bad;
1118
1119         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1120         if (r)
1121                 goto bad;
1122
1123         if ((!m->nr_priority_groups && next_pg_num) ||
1124             (m->nr_priority_groups && !next_pg_num)) {
1125                 ti->error = "invalid initial priority group";
1126                 r = -EINVAL;
1127                 goto bad;
1128         }
1129
1130         /* parse the priority groups */
1131         while (as.argc) {
1132                 struct priority_group *pg;
1133                 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1134
1135                 pg = parse_priority_group(&as, m);
1136                 if (IS_ERR(pg)) {
1137                         r = PTR_ERR(pg);
1138                         goto bad;
1139                 }
1140
1141                 nr_valid_paths += pg->nr_pgpaths;
1142                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1143
1144                 list_add_tail(&pg->list, &m->priority_groups);
1145                 pg_count++;
1146                 pg->pg_num = pg_count;
1147                 if (!--next_pg_num)
1148                         m->next_pg = pg;
1149         }
1150
1151         if (pg_count != m->nr_priority_groups) {
1152                 ti->error = "priority group count mismatch";
1153                 r = -EINVAL;
1154                 goto bad;
1155         }
1156
1157         ti->num_flush_bios = 1;
1158         ti->num_discard_bios = 1;
1159         ti->num_write_same_bios = 1;
1160         ti->num_write_zeroes_bios = 1;
1161         if (m->queue_mode == DM_TYPE_BIO_BASED)
1162                 ti->per_io_data_size = multipath_per_bio_data_size();
1163         else
1164                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1165
1166         return 0;
1167
1168  bad:
1169         free_multipath(m);
1170         return r;
1171 }
1172
1173 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1174 {
1175         DEFINE_WAIT(wait);
1176
1177         while (1) {
1178                 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1179
1180                 if (!atomic_read(&m->pg_init_in_progress))
1181                         break;
1182
1183                 io_schedule();
1184         }
1185         finish_wait(&m->pg_init_wait, &wait);
1186 }
1187
1188 static void flush_multipath_work(struct multipath *m)
1189 {
1190         if (m->hw_handler_name) {
1191                 set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1192                 smp_mb__after_atomic();
1193
1194                 if (atomic_read(&m->pg_init_in_progress))
1195                         flush_workqueue(kmpath_handlerd);
1196                 multipath_wait_for_pg_init_completion(m);
1197
1198                 clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1199                 smp_mb__after_atomic();
1200         }
1201
1202         if (m->queue_mode == DM_TYPE_BIO_BASED)
1203                 flush_work(&m->process_queued_bios);
1204         flush_work(&m->trigger_event);
1205 }
1206
1207 static void multipath_dtr(struct dm_target *ti)
1208 {
1209         struct multipath *m = ti->private;
1210
1211         flush_multipath_work(m);
1212         free_multipath(m);
1213 }
1214
1215 /*
1216  * Take a path out of use.
1217  */
1218 static int fail_path(struct pgpath *pgpath)
1219 {
1220         unsigned long flags;
1221         struct multipath *m = pgpath->pg->m;
1222
1223         spin_lock_irqsave(&m->lock, flags);
1224
1225         if (!pgpath->is_active)
1226                 goto out;
1227
1228         DMWARN("Failing path %s.", pgpath->path.dev->name);
1229
1230         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1231         pgpath->is_active = false;
1232         pgpath->fail_count++;
1233
1234         atomic_dec(&m->nr_valid_paths);
1235
1236         if (pgpath == m->current_pgpath)
1237                 m->current_pgpath = NULL;
1238
1239         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1240                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1241
1242         schedule_work(&m->trigger_event);
1243
1244 out:
1245         spin_unlock_irqrestore(&m->lock, flags);
1246
1247         return 0;
1248 }
1249
1250 /*
1251  * Reinstate a previously-failed path
1252  */
1253 static int reinstate_path(struct pgpath *pgpath)
1254 {
1255         int r = 0, run_queue = 0;
1256         unsigned long flags;
1257         struct multipath *m = pgpath->pg->m;
1258         unsigned nr_valid_paths;
1259
1260         spin_lock_irqsave(&m->lock, flags);
1261
1262         if (pgpath->is_active)
1263                 goto out;
1264
1265         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1266
1267         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1268         if (r)
1269                 goto out;
1270
1271         pgpath->is_active = true;
1272
1273         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1274         if (nr_valid_paths == 1) {
1275                 m->current_pgpath = NULL;
1276                 run_queue = 1;
1277         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1278                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1279                         atomic_inc(&m->pg_init_in_progress);
1280         }
1281
1282         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1283                        pgpath->path.dev->name, nr_valid_paths);
1284
1285         schedule_work(&m->trigger_event);
1286
1287 out:
1288         spin_unlock_irqrestore(&m->lock, flags);
1289         if (run_queue) {
1290                 dm_table_run_md_queue_async(m->ti->table);
1291                 process_queued_io_list(m);
1292         }
1293
1294         return r;
1295 }
1296
1297 /*
1298  * Fail or reinstate all paths that match the provided struct dm_dev.
1299  */
1300 static int action_dev(struct multipath *m, struct dm_dev *dev,
1301                       action_fn action)
1302 {
1303         int r = -EINVAL;
1304         struct pgpath *pgpath;
1305         struct priority_group *pg;
1306
1307         list_for_each_entry(pg, &m->priority_groups, list) {
1308                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1309                         if (pgpath->path.dev == dev)
1310                                 r = action(pgpath);
1311                 }
1312         }
1313
1314         return r;
1315 }
1316
1317 /*
1318  * Temporarily try to avoid having to use the specified PG
1319  */
1320 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1321                       bool bypassed)
1322 {
1323         unsigned long flags;
1324
1325         spin_lock_irqsave(&m->lock, flags);
1326
1327         pg->bypassed = bypassed;
1328         m->current_pgpath = NULL;
1329         m->current_pg = NULL;
1330
1331         spin_unlock_irqrestore(&m->lock, flags);
1332
1333         schedule_work(&m->trigger_event);
1334 }
1335
1336 /*
1337  * Switch to using the specified PG from the next I/O that gets mapped
1338  */
1339 static int switch_pg_num(struct multipath *m, const char *pgstr)
1340 {
1341         struct priority_group *pg;
1342         unsigned pgnum;
1343         unsigned long flags;
1344         char dummy;
1345
1346         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1347             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1348                 DMWARN("invalid PG number supplied to switch_pg_num");
1349                 return -EINVAL;
1350         }
1351
1352         spin_lock_irqsave(&m->lock, flags);
1353         list_for_each_entry(pg, &m->priority_groups, list) {
1354                 pg->bypassed = false;
1355                 if (--pgnum)
1356                         continue;
1357
1358                 m->current_pgpath = NULL;
1359                 m->current_pg = NULL;
1360                 m->next_pg = pg;
1361         }
1362         spin_unlock_irqrestore(&m->lock, flags);
1363
1364         schedule_work(&m->trigger_event);
1365         return 0;
1366 }
1367
1368 /*
1369  * Set/clear bypassed status of a PG.
1370  * PGs are numbered upwards from 1 in the order they were declared.
1371  */
1372 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1373 {
1374         struct priority_group *pg;
1375         unsigned pgnum;
1376         char dummy;
1377
1378         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1379             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1380                 DMWARN("invalid PG number supplied to bypass_pg");
1381                 return -EINVAL;
1382         }
1383
1384         list_for_each_entry(pg, &m->priority_groups, list) {
1385                 if (!--pgnum)
1386                         break;
1387         }
1388
1389         bypass_pg(m, pg, bypassed);
1390         return 0;
1391 }
1392
1393 /*
1394  * Should we retry pg_init immediately?
1395  */
1396 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1397 {
1398         unsigned long flags;
1399         bool limit_reached = false;
1400
1401         spin_lock_irqsave(&m->lock, flags);
1402
1403         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1404             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1405                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1406         else
1407                 limit_reached = true;
1408
1409         spin_unlock_irqrestore(&m->lock, flags);
1410
1411         return limit_reached;
1412 }
1413
1414 static void pg_init_done(void *data, int errors)
1415 {
1416         struct pgpath *pgpath = data;
1417         struct priority_group *pg = pgpath->pg;
1418         struct multipath *m = pg->m;
1419         unsigned long flags;
1420         bool delay_retry = false;
1421
1422         /* device or driver problems */
1423         switch (errors) {
1424         case SCSI_DH_OK:
1425                 break;
1426         case SCSI_DH_NOSYS:
1427                 if (!m->hw_handler_name) {
1428                         errors = 0;
1429                         break;
1430                 }
1431                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1432                       "Error %d.", m->hw_handler_name, errors);
1433                 /*
1434                  * Fail path for now, so we do not ping pong
1435                  */
1436                 fail_path(pgpath);
1437                 break;
1438         case SCSI_DH_DEV_TEMP_BUSY:
1439                 /*
1440                  * Probably doing something like FW upgrade on the
1441                  * controller so try the other pg.
1442                  */
1443                 bypass_pg(m, pg, true);
1444                 break;
1445         case SCSI_DH_RETRY:
1446                 /* Wait before retrying. */
1447                 delay_retry = 1;
1448                 /* fall through */
1449         case SCSI_DH_IMM_RETRY:
1450         case SCSI_DH_RES_TEMP_UNAVAIL:
1451                 if (pg_init_limit_reached(m, pgpath))
1452                         fail_path(pgpath);
1453                 errors = 0;
1454                 break;
1455         case SCSI_DH_DEV_OFFLINED:
1456         default:
1457                 /*
1458                  * We probably do not want to fail the path for a device
1459                  * error, but this is what the old dm did. In future
1460                  * patches we can do more advanced handling.
1461                  */
1462                 fail_path(pgpath);
1463         }
1464
1465         spin_lock_irqsave(&m->lock, flags);
1466         if (errors) {
1467                 if (pgpath == m->current_pgpath) {
1468                         DMERR("Could not failover device. Error %d.", errors);
1469                         m->current_pgpath = NULL;
1470                         m->current_pg = NULL;
1471                 }
1472         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1473                 pg->bypassed = false;
1474
1475         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1476                 /* Activations of other paths are still on going */
1477                 goto out;
1478
1479         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1480                 if (delay_retry)
1481                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1482                 else
1483                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1484
1485                 if (__pg_init_all_paths(m))
1486                         goto out;
1487         }
1488         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1489
1490         process_queued_io_list(m);
1491
1492         /*
1493          * Wake up any thread waiting to suspend.
1494          */
1495         wake_up(&m->pg_init_wait);
1496
1497 out:
1498         spin_unlock_irqrestore(&m->lock, flags);
1499 }
1500
1501 static void activate_or_offline_path(struct pgpath *pgpath)
1502 {
1503         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1504
1505         if (pgpath->is_active && !blk_queue_dying(q))
1506                 scsi_dh_activate(q, pg_init_done, pgpath);
1507         else
1508                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1509 }
1510
1511 static void activate_path_work(struct work_struct *work)
1512 {
1513         struct pgpath *pgpath =
1514                 container_of(work, struct pgpath, activate_path.work);
1515
1516         activate_or_offline_path(pgpath);
1517 }
1518
1519 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1520                             blk_status_t error, union map_info *map_context)
1521 {
1522         struct dm_mpath_io *mpio = get_mpio(map_context);
1523         struct pgpath *pgpath = mpio->pgpath;
1524         int r = DM_ENDIO_DONE;
1525
1526         /*
1527          * We don't queue any clone request inside the multipath target
1528          * during end I/O handling, since those clone requests don't have
1529          * bio clones.  If we queue them inside the multipath target,
1530          * we need to make bio clones, that requires memory allocation.
1531          * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1532          *  don't have bio clones.)
1533          * Instead of queueing the clone request here, we queue the original
1534          * request into dm core, which will remake a clone request and
1535          * clone bios for it and resubmit it later.
1536          */
1537         if (error && blk_path_error(error)) {
1538                 struct multipath *m = ti->private;
1539
1540                 if (error == BLK_STS_RESOURCE)
1541                         r = DM_ENDIO_DELAY_REQUEUE;
1542                 else
1543                         r = DM_ENDIO_REQUEUE;
1544
1545                 if (pgpath)
1546                         fail_path(pgpath);
1547
1548                 if (atomic_read(&m->nr_valid_paths) == 0 &&
1549                     !must_push_back_rq(m)) {
1550                         if (error == BLK_STS_IOERR)
1551                                 dm_report_EIO(m);
1552                         /* complete with the original error */
1553                         r = DM_ENDIO_DONE;
1554                 }
1555         }
1556
1557         if (pgpath) {
1558                 struct path_selector *ps = &pgpath->pg->ps;
1559
1560                 if (ps->type->end_io)
1561                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1562         }
1563
1564         return r;
1565 }
1566
1567 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1568                                 blk_status_t *error)
1569 {
1570         struct multipath *m = ti->private;
1571         struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1572         struct pgpath *pgpath = mpio->pgpath;
1573         unsigned long flags;
1574         int r = DM_ENDIO_DONE;
1575
1576         if (!*error || !blk_path_error(*error))
1577                 goto done;
1578
1579         if (pgpath)
1580                 fail_path(pgpath);
1581
1582         if (atomic_read(&m->nr_valid_paths) == 0 &&
1583             !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1584                 if (must_push_back_bio(m)) {
1585                         r = DM_ENDIO_REQUEUE;
1586                 } else {
1587                         dm_report_EIO(m);
1588                         *error = BLK_STS_IOERR;
1589                 }
1590                 goto done;
1591         }
1592
1593         spin_lock_irqsave(&m->lock, flags);
1594         bio_list_add(&m->queued_bios, clone);
1595         spin_unlock_irqrestore(&m->lock, flags);
1596         if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1597                 queue_work(kmultipathd, &m->process_queued_bios);
1598
1599         r = DM_ENDIO_INCOMPLETE;
1600 done:
1601         if (pgpath) {
1602                 struct path_selector *ps = &pgpath->pg->ps;
1603
1604                 if (ps->type->end_io)
1605                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1606         }
1607
1608         return r;
1609 }
1610
1611 /*
1612  * Suspend can't complete until all the I/O is processed so if
1613  * the last path fails we must error any remaining I/O.
1614  * Note that if the freeze_bdev fails while suspending, the
1615  * queue_if_no_path state is lost - userspace should reset it.
1616  */
1617 static void multipath_presuspend(struct dm_target *ti)
1618 {
1619         struct multipath *m = ti->private;
1620
1621         queue_if_no_path(m, false, true);
1622 }
1623
1624 static void multipath_postsuspend(struct dm_target *ti)
1625 {
1626         struct multipath *m = ti->private;
1627
1628         mutex_lock(&m->work_mutex);
1629         flush_multipath_work(m);
1630         mutex_unlock(&m->work_mutex);
1631 }
1632
1633 /*
1634  * Restore the queue_if_no_path setting.
1635  */
1636 static void multipath_resume(struct dm_target *ti)
1637 {
1638         struct multipath *m = ti->private;
1639         unsigned long flags;
1640
1641         spin_lock_irqsave(&m->lock, flags);
1642         assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1643                    test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1644         spin_unlock_irqrestore(&m->lock, flags);
1645 }
1646
1647 /*
1648  * Info output has the following format:
1649  * num_multipath_feature_args [multipath_feature_args]*
1650  * num_handler_status_args [handler_status_args]*
1651  * num_groups init_group_number
1652  *            [A|D|E num_ps_status_args [ps_status_args]*
1653  *             num_paths num_selector_args
1654  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1655  *
1656  * Table output has the following format (identical to the constructor string):
1657  * num_feature_args [features_args]*
1658  * num_handler_args hw_handler [hw_handler_args]*
1659  * num_groups init_group_number
1660  *     [priority selector-name num_ps_args [ps_args]*
1661  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1662  */
1663 static void multipath_status(struct dm_target *ti, status_type_t type,
1664                              unsigned status_flags, char *result, unsigned maxlen)
1665 {
1666         int sz = 0;
1667         unsigned long flags;
1668         struct multipath *m = ti->private;
1669         struct priority_group *pg;
1670         struct pgpath *p;
1671         unsigned pg_num;
1672         char state;
1673
1674         spin_lock_irqsave(&m->lock, flags);
1675
1676         /* Features */
1677         if (type == STATUSTYPE_INFO)
1678                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1679                        atomic_read(&m->pg_init_count));
1680         else {
1681                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1682                               (m->pg_init_retries > 0) * 2 +
1683                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1684                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1685                               (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1686
1687                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1688                         DMEMIT("queue_if_no_path ");
1689                 if (m->pg_init_retries)
1690                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1691                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1692                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1693                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1694                         DMEMIT("retain_attached_hw_handler ");
1695                 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1696                         switch(m->queue_mode) {
1697                         case DM_TYPE_BIO_BASED:
1698                                 DMEMIT("queue_mode bio ");
1699                                 break;
1700                         default:
1701                                 WARN_ON_ONCE(true);
1702                                 break;
1703                         }
1704                 }
1705         }
1706
1707         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1708                 DMEMIT("0 ");
1709         else
1710                 DMEMIT("1 %s ", m->hw_handler_name);
1711
1712         DMEMIT("%u ", m->nr_priority_groups);
1713
1714         if (m->next_pg)
1715                 pg_num = m->next_pg->pg_num;
1716         else if (m->current_pg)
1717                 pg_num = m->current_pg->pg_num;
1718         else
1719                 pg_num = (m->nr_priority_groups ? 1 : 0);
1720
1721         DMEMIT("%u ", pg_num);
1722
1723         switch (type) {
1724         case STATUSTYPE_INFO:
1725                 list_for_each_entry(pg, &m->priority_groups, list) {
1726                         if (pg->bypassed)
1727                                 state = 'D';    /* Disabled */
1728                         else if (pg == m->current_pg)
1729                                 state = 'A';    /* Currently Active */
1730                         else
1731                                 state = 'E';    /* Enabled */
1732
1733                         DMEMIT("%c ", state);
1734
1735                         if (pg->ps.type->status)
1736                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1737                                                           result + sz,
1738                                                           maxlen - sz);
1739                         else
1740                                 DMEMIT("0 ");
1741
1742                         DMEMIT("%u %u ", pg->nr_pgpaths,
1743                                pg->ps.type->info_args);
1744
1745                         list_for_each_entry(p, &pg->pgpaths, list) {
1746                                 DMEMIT("%s %s %u ", p->path.dev->name,
1747                                        p->is_active ? "A" : "F",
1748                                        p->fail_count);
1749                                 if (pg->ps.type->status)
1750                                         sz += pg->ps.type->status(&pg->ps,
1751                                               &p->path, type, result + sz,
1752                                               maxlen - sz);
1753                         }
1754                 }
1755                 break;
1756
1757         case STATUSTYPE_TABLE:
1758                 list_for_each_entry(pg, &m->priority_groups, list) {
1759                         DMEMIT("%s ", pg->ps.type->name);
1760
1761                         if (pg->ps.type->status)
1762                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1763                                                           result + sz,
1764                                                           maxlen - sz);
1765                         else
1766                                 DMEMIT("0 ");
1767
1768                         DMEMIT("%u %u ", pg->nr_pgpaths,
1769                                pg->ps.type->table_args);
1770
1771                         list_for_each_entry(p, &pg->pgpaths, list) {
1772                                 DMEMIT("%s ", p->path.dev->name);
1773                                 if (pg->ps.type->status)
1774                                         sz += pg->ps.type->status(&pg->ps,
1775                                               &p->path, type, result + sz,
1776                                               maxlen - sz);
1777                         }
1778                 }
1779                 break;
1780         }
1781
1782         spin_unlock_irqrestore(&m->lock, flags);
1783 }
1784
1785 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1786                              char *result, unsigned maxlen)
1787 {
1788         int r = -EINVAL;
1789         struct dm_dev *dev;
1790         struct multipath *m = ti->private;
1791         action_fn action;
1792
1793         mutex_lock(&m->work_mutex);
1794
1795         if (dm_suspended(ti)) {
1796                 r = -EBUSY;
1797                 goto out;
1798         }
1799
1800         if (argc == 1) {
1801                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1802                         r = queue_if_no_path(m, true, false);
1803                         goto out;
1804                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1805                         r = queue_if_no_path(m, false, false);
1806                         goto out;
1807                 }
1808         }
1809
1810         if (argc != 2) {
1811                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1812                 goto out;
1813         }
1814
1815         if (!strcasecmp(argv[0], "disable_group")) {
1816                 r = bypass_pg_num(m, argv[1], true);
1817                 goto out;
1818         } else if (!strcasecmp(argv[0], "enable_group")) {
1819                 r = bypass_pg_num(m, argv[1], false);
1820                 goto out;
1821         } else if (!strcasecmp(argv[0], "switch_group")) {
1822                 r = switch_pg_num(m, argv[1]);
1823                 goto out;
1824         } else if (!strcasecmp(argv[0], "reinstate_path"))
1825                 action = reinstate_path;
1826         else if (!strcasecmp(argv[0], "fail_path"))
1827                 action = fail_path;
1828         else {
1829                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1830                 goto out;
1831         }
1832
1833         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1834         if (r) {
1835                 DMWARN("message: error getting device %s",
1836                        argv[1]);
1837                 goto out;
1838         }
1839
1840         r = action_dev(m, dev, action);
1841
1842         dm_put_device(ti, dev);
1843
1844 out:
1845         mutex_unlock(&m->work_mutex);
1846         return r;
1847 }
1848
1849 static int multipath_prepare_ioctl(struct dm_target *ti,
1850                                    struct block_device **bdev)
1851 {
1852         struct multipath *m = ti->private;
1853         struct pgpath *current_pgpath;
1854         int r;
1855
1856         current_pgpath = READ_ONCE(m->current_pgpath);
1857         if (!current_pgpath)
1858                 current_pgpath = choose_pgpath(m, 0);
1859
1860         if (current_pgpath) {
1861                 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1862                         *bdev = current_pgpath->path.dev->bdev;
1863                         r = 0;
1864                 } else {
1865                         /* pg_init has not started or completed */
1866                         r = -ENOTCONN;
1867                 }
1868         } else {
1869                 /* No path is available */
1870                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1871                         r = -ENOTCONN;
1872                 else
1873                         r = -EIO;
1874         }
1875
1876         if (r == -ENOTCONN) {
1877                 if (!READ_ONCE(m->current_pg)) {
1878                         /* Path status changed, redo selection */
1879                         (void) choose_pgpath(m, 0);
1880                 }
1881                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1882                         pg_init_all_paths(m);
1883                 dm_table_run_md_queue_async(m->ti->table);
1884                 process_queued_io_list(m);
1885         }
1886
1887         /*
1888          * Only pass ioctls through if the device sizes match exactly.
1889          */
1890         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1891                 return 1;
1892         return r;
1893 }
1894
1895 static int multipath_iterate_devices(struct dm_target *ti,
1896                                      iterate_devices_callout_fn fn, void *data)
1897 {
1898         struct multipath *m = ti->private;
1899         struct priority_group *pg;
1900         struct pgpath *p;
1901         int ret = 0;
1902
1903         list_for_each_entry(pg, &m->priority_groups, list) {
1904                 list_for_each_entry(p, &pg->pgpaths, list) {
1905                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1906                         if (ret)
1907                                 goto out;
1908                 }
1909         }
1910
1911 out:
1912         return ret;
1913 }
1914
1915 static int pgpath_busy(struct pgpath *pgpath)
1916 {
1917         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1918
1919         return blk_lld_busy(q);
1920 }
1921
1922 /*
1923  * We return "busy", only when we can map I/Os but underlying devices
1924  * are busy (so even if we map I/Os now, the I/Os will wait on
1925  * the underlying queue).
1926  * In other words, if we want to kill I/Os or queue them inside us
1927  * due to map unavailability, we don't return "busy".  Otherwise,
1928  * dm core won't give us the I/Os and we can't do what we want.
1929  */
1930 static int multipath_busy(struct dm_target *ti)
1931 {
1932         bool busy = false, has_active = false;
1933         struct multipath *m = ti->private;
1934         struct priority_group *pg, *next_pg;
1935         struct pgpath *pgpath;
1936
1937         /* pg_init in progress */
1938         if (atomic_read(&m->pg_init_in_progress))
1939                 return true;
1940
1941         /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1942         if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1943                 return (m->queue_mode != DM_TYPE_REQUEST_BASED);
1944
1945         /* Guess which priority_group will be used at next mapping time */
1946         pg = READ_ONCE(m->current_pg);
1947         next_pg = READ_ONCE(m->next_pg);
1948         if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1949                 pg = next_pg;
1950
1951         if (!pg) {
1952                 /*
1953                  * We don't know which pg will be used at next mapping time.
1954                  * We don't call choose_pgpath() here to avoid to trigger
1955                  * pg_init just by busy checking.
1956                  * So we don't know whether underlying devices we will be using
1957                  * at next mapping time are busy or not. Just try mapping.
1958                  */
1959                 return busy;
1960         }
1961
1962         /*
1963          * If there is one non-busy active path at least, the path selector
1964          * will be able to select it. So we consider such a pg as not busy.
1965          */
1966         busy = true;
1967         list_for_each_entry(pgpath, &pg->pgpaths, list) {
1968                 if (pgpath->is_active) {
1969                         has_active = true;
1970                         if (!pgpath_busy(pgpath)) {
1971                                 busy = false;
1972                                 break;
1973                         }
1974                 }
1975         }
1976
1977         if (!has_active) {
1978                 /*
1979                  * No active path in this pg, so this pg won't be used and
1980                  * the current_pg will be changed at next mapping time.
1981                  * We need to try mapping to determine it.
1982                  */
1983                 busy = false;
1984         }
1985
1986         return busy;
1987 }
1988
1989 /*-----------------------------------------------------------------
1990  * Module setup
1991  *---------------------------------------------------------------*/
1992 static struct target_type multipath_target = {
1993         .name = "multipath",
1994         .version = {1, 13, 0},
1995         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
1996                     DM_TARGET_PASSES_INTEGRITY,
1997         .module = THIS_MODULE,
1998         .ctr = multipath_ctr,
1999         .dtr = multipath_dtr,
2000         .clone_and_map_rq = multipath_clone_and_map,
2001         .release_clone_rq = multipath_release_clone,
2002         .rq_end_io = multipath_end_io,
2003         .map = multipath_map_bio,
2004         .end_io = multipath_end_io_bio,
2005         .presuspend = multipath_presuspend,
2006         .postsuspend = multipath_postsuspend,
2007         .resume = multipath_resume,
2008         .status = multipath_status,
2009         .message = multipath_message,
2010         .prepare_ioctl = multipath_prepare_ioctl,
2011         .iterate_devices = multipath_iterate_devices,
2012         .busy = multipath_busy,
2013 };
2014
2015 static int __init dm_multipath_init(void)
2016 {
2017         int r;
2018
2019         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2020         if (!kmultipathd) {
2021                 DMERR("failed to create workqueue kmpathd");
2022                 r = -ENOMEM;
2023                 goto bad_alloc_kmultipathd;
2024         }
2025
2026         /*
2027          * A separate workqueue is used to handle the device handlers
2028          * to avoid overloading existing workqueue. Overloading the
2029          * old workqueue would also create a bottleneck in the
2030          * path of the storage hardware device activation.
2031          */
2032         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2033                                                   WQ_MEM_RECLAIM);
2034         if (!kmpath_handlerd) {
2035                 DMERR("failed to create workqueue kmpath_handlerd");
2036                 r = -ENOMEM;
2037                 goto bad_alloc_kmpath_handlerd;
2038         }
2039
2040         r = dm_register_target(&multipath_target);
2041         if (r < 0) {
2042                 DMERR("request-based register failed %d", r);
2043                 r = -EINVAL;
2044                 goto bad_register_target;
2045         }
2046
2047         return 0;
2048
2049 bad_register_target:
2050         destroy_workqueue(kmpath_handlerd);
2051 bad_alloc_kmpath_handlerd:
2052         destroy_workqueue(kmultipathd);
2053 bad_alloc_kmultipathd:
2054         return r;
2055 }
2056
2057 static void __exit dm_multipath_exit(void)
2058 {
2059         destroy_workqueue(kmpath_handlerd);
2060         destroy_workqueue(kmultipathd);
2061
2062         dm_unregister_target(&multipath_target);
2063 }
2064
2065 module_init(dm_multipath_init);
2066 module_exit(dm_multipath_exit);
2067
2068 MODULE_DESCRIPTION(DM_NAME " multipath target");
2069 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2070 MODULE_LICENSE("GPL");