drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9
  10 #include "dm-rq.h"
  11 #include "dm-bio-record.h"
  12 #include "dm-path-selector.h"
  13 #include "dm-uevent.h"
  14
  15 #include <linux/blkdev.h>
  16 #include <linux/ctype.h>
  17 #include <linux/init.h>
  18 #include <linux/mempool.h>
  19 #include <linux/module.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/slab.h>
  22 #include <linux/time.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/delay.h>
  25 #include <scsi/scsi_dh.h>
  26 #include <linux/atomic.h>
  27 #include <linux/blk-mq.h>
  28
  29 #define DM_MSG_PREFIX "multipath"
  30 #define DM_PG_INIT_DELAY_MSECS 2000
  31 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  32
  33 /* Path properties */
  34 struct pgpath {
  35         struct list_head list;
  36
  37         struct priority_group *pg;      /* Owning PG */
  38         unsigned fail_count;            /* Cumulative failure count */
  39
  40         struct dm_path path;
  41         struct delayed_work activate_path;
  42
  43         bool is_active:1;               /* Path status */
  44 };
  45
  46 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  47
  48 /*
  49  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  50  * Each has a path selector which controls which path gets used.
  51  */
  52 struct priority_group {
  53         struct list_head list;
  54
  55         struct multipath *m;            /* Owning multipath instance */
  56         struct path_selector ps;
  57
  58         unsigned pg_num;                /* Reference number */
  59         unsigned nr_pgpaths;            /* Number of paths in PG */
  60         struct list_head pgpaths;
  61
  62         bool bypassed:1;                /* Temporarily bypass this PG? */
  63 };
  64
  65 /* Multipath context */
  66 struct multipath {
  67         struct list_head list;
  68         struct dm_target *ti;
  69
  70         const char *hw_handler_name;
  71         char *hw_handler_params;
  72
  73         spinlock_t lock;
  74
  75         unsigned nr_priority_groups;
  76         struct list_head priority_groups;
  77
  78         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  79
  80         struct pgpath *current_pgpath;
  81         struct priority_group *current_pg;
  82         struct priority_group *next_pg; /* Switch to this PG if set */
  83
  84         unsigned long flags;            /* Multipath state flags */
  85
  86         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  87         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  88
  89         atomic_t nr_valid_paths;        /* Total number of usable paths */
  90         atomic_t pg_init_in_progress;   /* Only one pg_init allowed at once */
  91         atomic_t pg_init_count;         /* Number of times pg_init called */
  92
  93         enum dm_queue_mode queue_mode;
  94
  95         struct mutex work_mutex;
  96         struct work_struct trigger_event;
  97
  98         struct work_struct process_queued_bios;
  99         struct bio_list queued_bios;
 100 };
 101
 102 /*
 103  * Context information attached to each io we process.
 104  */
 105 struct dm_mpath_io {
 106         struct pgpath *pgpath;
 107         size_t nr_bytes;
 108 };
 109
 110 typedef int (*action_fn) (struct pgpath *pgpath);
 111
 112 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 113 static void trigger_event(struct work_struct *work);
 114 static void activate_or_offline_path(struct pgpath *pgpath);
 115 static void activate_path_work(struct work_struct *work);
 116 static void process_queued_bios(struct work_struct *work);
 117
 118 /*-----------------------------------------------
 119  * Multipath state flags.
 120  *-----------------------------------------------*/
 121
 122 #define MPATHF_QUEUE_IO 0                       /* Must we queue all I/O? */
 123 #define MPATHF_QUEUE_IF_NO_PATH 1               /* Queue I/O if last path fails? */
 124 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2         /* Saved state during suspension */
 125 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3     /* If there's already a hw_handler present, don't change it. */
 126 #define MPATHF_PG_INIT_DISABLED 4               /* pg_init is not currently allowed */
 127 #define MPATHF_PG_INIT_REQUIRED 5               /* pg_init needs calling? */
 128 #define MPATHF_PG_INIT_DELAY_RETRY 6            /* Delay pg_init retry? */
 129
 130 /*-----------------------------------------------
 131  * Allocation routines
 132  *-----------------------------------------------*/
 133
 134 static struct pgpath *alloc_pgpath(void)
 135 {
 136         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 137
 138         if (pgpath) {
 139                 pgpath->is_active = true;
 140                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path_work);
 141         }
 142
 143         return pgpath;
 144 }
 145
 146 static void free_pgpath(struct pgpath *pgpath)
 147 {
 148         kfree(pgpath);
 149 }
 150
 151 static struct priority_group *alloc_priority_group(void)
 152 {
 153         struct priority_group *pg;
 154
 155         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 156
 157         if (pg)
 158                 INIT_LIST_HEAD(&pg->pgpaths);
 159
 160         return pg;
 161 }
 162
 163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 164 {
 165         struct pgpath *pgpath, *tmp;
 166
 167         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 168                 list_del(&pgpath->list);
 169                 dm_put_device(ti, pgpath->path.dev);
 170                 free_pgpath(pgpath);
 171         }
 172 }
 173
 174 static void free_priority_group(struct priority_group *pg,
 175                                 struct dm_target *ti)
 176 {
 177         struct path_selector *ps = &pg->ps;
 178
 179         if (ps->type) {
 180                 ps->type->destroy(ps);
 181                 dm_put_path_selector(ps->type);
 182         }
 183
 184         free_pgpaths(&pg->pgpaths, ti);
 185         kfree(pg);
 186 }
 187
 188 static struct multipath *alloc_multipath(struct dm_target *ti)
 189 {
 190         struct multipath *m;
 191
 192         m = kzalloc(sizeof(*m), GFP_KERNEL);
 193         if (m) {
 194                 INIT_LIST_HEAD(&m->priority_groups);
 195                 spin_lock_init(&m->lock);
 196                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 197                 atomic_set(&m->nr_valid_paths, 0);
 198                 atomic_set(&m->pg_init_in_progress, 0);
 199                 atomic_set(&m->pg_init_count, 0);
 200                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 201                 INIT_WORK(&m->trigger_event, trigger_event);
 202                 init_waitqueue_head(&m->pg_init_wait);
 203                 mutex_init(&m->work_mutex);
 204
 205                 m->queue_mode = DM_TYPE_NONE;
 206
 207                 m->ti = ti;
 208                 ti->private = m;
 209         }
 210
 211         return m;
 212 }
 213
 214 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
 215 {
 216         if (m->queue_mode == DM_TYPE_NONE) {
 217                 /*
 218                  * Default to request-based.
 219                  */
 220                 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
 221                         m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
 222                 else
 223                         m->queue_mode = DM_TYPE_REQUEST_BASED;
 224         } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
 225                 INIT_WORK(&m->process_queued_bios, process_queued_bios);
 226                 /*
 227                  * bio-based doesn't support any direct scsi_dh management;
 228                  * it just discovers if a scsi_dh is attached.
 229                  */
 230                 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 231         }
 232
 233         dm_table_set_type(ti->table, m->queue_mode);
 234
 235         return 0;
 236 }
 237
 238 static void free_multipath(struct multipath *m)
 239 {
 240         struct priority_group *pg, *tmp;
 241
 242         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 243                 list_del(&pg->list);
 244                 free_priority_group(pg, m->ti);
 245         }
 246
 247         kfree(m->hw_handler_name);
 248         kfree(m->hw_handler_params);
 249         kfree(m);
 250 }
 251
 252 static struct dm_mpath_io *get_mpio(union map_info *info)
 253 {
 254         return info->ptr;
 255 }
 256
 257 static size_t multipath_per_bio_data_size(void)
 258 {
 259         return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
 260 }
 261
 262 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
 263 {
 264         return dm_per_bio_data(bio, multipath_per_bio_data_size());
 265 }
 266
 267 static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
 268 {
 269         /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
 270         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 271         void *bio_details = mpio + 1;
 272
 273         return bio_details;
 274 }
 275
 276 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
 277 {
 278         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 279         struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
 280
 281         memset(mpio, 0, sizeof(*mpio));
 282         memset(bio_details, 0, sizeof(*bio_details));
 283         dm_bio_record(bio_details, bio);
 284
 285         if (mpio_p)
 286                 *mpio_p = mpio;
 287 }
 288
 289 /*-----------------------------------------------
 290  * Path selection
 291  *-----------------------------------------------*/
 292
 293 static int __pg_init_all_paths(struct multipath *m)
 294 {
 295         struct pgpath *pgpath;
 296         unsigned long pg_init_delay = 0;
 297
 298         lockdep_assert_held(&m->lock);
 299
 300         if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
 301                 return 0;
 302
 303         atomic_inc(&m->pg_init_count);
 304         clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 305
 306         /* Check here to reset pg_init_required */
 307         if (!m->current_pg)
 308                 return 0;
 309
 310         if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
 311                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 312                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 313         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 314                 /* Skip failed paths */
 315                 if (!pgpath->is_active)
 316                         continue;
 317                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 318                                        pg_init_delay))
 319                         atomic_inc(&m->pg_init_in_progress);
 320         }
 321         return atomic_read(&m->pg_init_in_progress);
 322 }
 323
 324 static int pg_init_all_paths(struct multipath *m)
 325 {
 326         int ret;
 327         unsigned long flags;
 328
 329         spin_lock_irqsave(&m->lock, flags);
 330         ret = __pg_init_all_paths(m);
 331         spin_unlock_irqrestore(&m->lock, flags);
 332
 333         return ret;
 334 }
 335
 336 static void __switch_pg(struct multipath *m, struct priority_group *pg)
 337 {
 338         m->current_pg = pg;
 339
 340         /* Must we initialise the PG first, and queue I/O till it's ready? */
 341         if (m->hw_handler_name) {
 342                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 343                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 344         } else {
 345                 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 346                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 347         }
 348
 349         atomic_set(&m->pg_init_count, 0);
 350 }
 351
 352 static struct pgpath *choose_path_in_pg(struct multipath *m,
 353                                         struct priority_group *pg,
 354                                         size_t nr_bytes)
 355 {
 356         unsigned long flags;
 357         struct dm_path *path;
 358         struct pgpath *pgpath;
 359
 360         path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 361         if (!path)
 362                 return ERR_PTR(-ENXIO);
 363
 364         pgpath = path_to_pgpath(path);
 365
 366         if (unlikely(READ_ONCE(m->current_pg) != pg)) {
 367                 /* Only update current_pgpath if pg changed */
 368                 spin_lock_irqsave(&m->lock, flags);
 369                 m->current_pgpath = pgpath;
 370                 __switch_pg(m, pg);
 371                 spin_unlock_irqrestore(&m->lock, flags);
 372         }
 373
 374         return pgpath;
 375 }
 376
 377 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
 378 {
 379         unsigned long flags;
 380         struct priority_group *pg;
 381         struct pgpath *pgpath;
 382         unsigned bypassed = 1;
 383
 384         if (!atomic_read(&m->nr_valid_paths)) {
 385                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 386                 goto failed;
 387         }
 388
 389         /* Were we instructed to switch PG? */
 390         if (READ_ONCE(m->next_pg)) {
 391                 spin_lock_irqsave(&m->lock, flags);
 392                 pg = m->next_pg;
 393                 if (!pg) {
 394                         spin_unlock_irqrestore(&m->lock, flags);
 395                         goto check_current_pg;
 396                 }
 397                 m->next_pg = NULL;
 398                 spin_unlock_irqrestore(&m->lock, flags);
 399                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 400                 if (!IS_ERR_OR_NULL(pgpath))
 401                         return pgpath;
 402         }
 403
 404         /* Don't change PG until it has no remaining paths */
 405 check_current_pg:
 406         pg = READ_ONCE(m->current_pg);
 407         if (pg) {
 408                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 409                 if (!IS_ERR_OR_NULL(pgpath))
 410                         return pgpath;
 411         }
 412
 413         /*
 414          * Loop through priority groups until we find a valid path.
 415          * First time we skip PGs marked 'bypassed'.
 416          * Second time we only try the ones we skipped, but set
 417          * pg_init_delay_retry so we do not hammer controllers.
 418          */
 419         do {
 420                 list_for_each_entry(pg, &m->priority_groups, list) {
 421                         if (pg->bypassed == !!bypassed)
 422                                 continue;
 423                         pgpath = choose_path_in_pg(m, pg, nr_bytes);
 424                         if (!IS_ERR_OR_NULL(pgpath)) {
 425                                 if (!bypassed)
 426                                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
 427                                 return pgpath;
 428                         }
 429                 }
 430         } while (bypassed--);
 431
 432 failed:
 433         spin_lock_irqsave(&m->lock, flags);
 434         m->current_pgpath = NULL;
 435         m->current_pg = NULL;
 436         spin_unlock_irqrestore(&m->lock, flags);
 437
 438         return NULL;
 439 }
 440
 441 /*
 442  * dm_report_EIO() is a macro instead of a function to make pr_debug()
 443  * report the function name and line number of the function from which
 444  * it has been invoked.
 445  */
 446 #define dm_report_EIO(m)                                                \
 447 do {                                                                    \
 448         struct mapped_device *md = dm_table_get_md((m)->ti->table);     \
 449                                                                         \
 450         pr_debug("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d\n", \
 451                  dm_device_name(md),                                    \
 452                  test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags),        \
 453                  test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags),  \
 454                  dm_noflush_suspending((m)->ti));                       \
 455 } while (0)
 456
 457 /*
 458  * Check whether bios must be queued in the device-mapper core rather
 459  * than here in the target.
 460  *
 461  * If MPATHF_QUEUE_IF_NO_PATH and MPATHF_SAVED_QUEUE_IF_NO_PATH hold
 462  * the same value then we are not between multipath_presuspend()
 463  * and multipath_resume() calls and we have no need to check
 464  * for the DMF_NOFLUSH_SUSPENDING flag.
 465  */
 466 static bool __must_push_back(struct multipath *m, unsigned long flags)
 467 {
 468         return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) !=
 469                  test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &flags)) &&
 470                 dm_noflush_suspending(m->ti));
 471 }
 472
 473 /*
 474  * Following functions use READ_ONCE to get atomic access to
 475  * all m->flags to avoid taking spinlock
 476  */
 477 static bool must_push_back_rq(struct multipath *m)
 478 {
 479         unsigned long flags = READ_ONCE(m->flags);
 480         return test_bit(MPATHF_QUEUE_IF_NO_PATH, &flags) || __must_push_back(m, flags);
 481 }
 482
 483 static bool must_push_back_bio(struct multipath *m)
 484 {
 485         unsigned long flags = READ_ONCE(m->flags);
 486         return __must_push_back(m, flags);
 487 }
 488
 489 /*
 490  * Map cloned requests (request-based multipath)
 491  */
 492 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 493                                    union map_info *map_context,
 494                                    struct request **__clone)
 495 {
 496         struct multipath *m = ti->private;
 497         size_t nr_bytes = blk_rq_bytes(rq);
 498         struct pgpath *pgpath;
 499         struct block_device *bdev;
 500         struct dm_mpath_io *mpio = get_mpio(map_context);
 501         struct request_queue *q;
 502         struct request *clone;
 503
 504         /* Do we need to select a new pgpath? */
 505         pgpath = READ_ONCE(m->current_pgpath);
 506         if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
 507                 pgpath = choose_pgpath(m, nr_bytes);
 508
 509         if (!pgpath) {
 510                 if (must_push_back_rq(m))
 511                         return DM_MAPIO_DELAY_REQUEUE;
 512                 dm_report_EIO(m);       /* Failed */
 513                 return DM_MAPIO_KILL;
 514         } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
 515                    test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
 516                 if (pg_init_all_paths(m))
 517                         return DM_MAPIO_DELAY_REQUEUE;
 518                 return DM_MAPIO_REQUEUE;
 519         }
 520
 521         memset(mpio, 0, sizeof(*mpio));
 522         mpio->pgpath = pgpath;
 523         mpio->nr_bytes = nr_bytes;
 524
 525         bdev = pgpath->path.dev->bdev;
 526         q = bdev_get_queue(bdev);
 527         clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE, GFP_ATOMIC);
 528         if (IS_ERR(clone)) {
 529                 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
 530                 bool queue_dying = blk_queue_dying(q);
 531                 if (queue_dying) {
 532                         atomic_inc(&m->pg_init_in_progress);
 533                         activate_or_offline_path(pgpath);
 534                 }
 535                 return DM_MAPIO_DELAY_REQUEUE;
 536         }
 537         clone->bio = clone->biotail = NULL;
 538         clone->rq_disk = bdev->bd_disk;
 539         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 540         *__clone = clone;
 541
 542         if (pgpath->pg->ps.type->start_io)
 543                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 544                                               &pgpath->path,
 545                                               nr_bytes);
 546         return DM_MAPIO_REMAPPED;
 547 }
 548
 549 static void multipath_release_clone(struct request *clone)
 550 {
 551         blk_put_request(clone);
 552 }
 553
 554 /*
 555  * Map cloned bios (bio-based multipath)
 556  */
 557 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
 558 {
 559         size_t nr_bytes = bio->bi_iter.bi_size;
 560         struct pgpath *pgpath;
 561         unsigned long flags;
 562         bool queue_io;
 563
 564         /* Do we need to select a new pgpath? */
 565         pgpath = READ_ONCE(m->current_pgpath);
 566         queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
 567         if (!pgpath || !queue_io)
 568                 pgpath = choose_pgpath(m, nr_bytes);
 569
 570         if ((pgpath && queue_io) ||
 571             (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
 572                 /* Queue for the daemon to resubmit */
 573                 spin_lock_irqsave(&m->lock, flags);
 574                 bio_list_add(&m->queued_bios, bio);
 575                 spin_unlock_irqrestore(&m->lock, flags);
 576                 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
 577                 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
 578                         pg_init_all_paths(m);
 579                 else if (!queue_io)
 580                         queue_work(kmultipathd, &m->process_queued_bios);
 581                 return DM_MAPIO_SUBMITTED;
 582         }
 583
 584         if (!pgpath) {
 585                 if (must_push_back_bio(m))
 586                         return DM_MAPIO_REQUEUE;
 587                 dm_report_EIO(m);
 588                 return DM_MAPIO_KILL;
 589         }
 590
 591         mpio->pgpath = pgpath;
 592         mpio->nr_bytes = nr_bytes;
 593
 594         bio->bi_status = 0;
 595         bio_set_dev(bio, pgpath->path.dev->bdev);
 596         bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
 597
 598         if (pgpath->pg->ps.type->start_io)
 599                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 600                                               &pgpath->path,
 601                                               nr_bytes);
 602         return DM_MAPIO_REMAPPED;
 603 }
 604
 605 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
 606 {
 607         struct multipath *m = ti->private;
 608         struct dm_mpath_io *mpio = NULL;
 609
 610         multipath_init_per_bio_data(bio, &mpio);
 611         return __multipath_map_bio(m, bio, mpio);
 612 }
 613
 614 static void process_queued_io_list(struct multipath *m)
 615 {
 616         if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
 617                 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
 618         else if (m->queue_mode == DM_TYPE_BIO_BASED)
 619                 queue_work(kmultipathd, &m->process_queued_bios);
 620 }
 621
 622 static void process_queued_bios(struct work_struct *work)
 623 {
 624         int r;
 625         unsigned long flags;
 626         struct bio *bio;
 627         struct bio_list bios;
 628         struct blk_plug plug;
 629         struct multipath *m =
 630                 container_of(work, struct multipath, process_queued_bios);
 631
 632         bio_list_init(&bios);
 633
 634         spin_lock_irqsave(&m->lock, flags);
 635
 636         if (bio_list_empty(&m->queued_bios)) {
 637                 spin_unlock_irqrestore(&m->lock, flags);
 638                 return;
 639         }
 640
 641         bio_list_merge(&bios, &m->queued_bios);
 642         bio_list_init(&m->queued_bios);
 643
 644         spin_unlock_irqrestore(&m->lock, flags);
 645
 646         blk_start_plug(&plug);
 647         while ((bio = bio_list_pop(&bios))) {
 648                 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
 649                 switch (r) {
 650                 case DM_MAPIO_KILL:
 651                         bio->bi_status = BLK_STS_IOERR;
 652                         bio_endio(bio);
 653                         break;
 654                 case DM_MAPIO_REQUEUE:
 655                         bio->bi_status = BLK_STS_DM_REQUEUE;
 656                         bio_endio(bio);
 657                         break;
 658                 case DM_MAPIO_REMAPPED:
 659                         generic_make_request(bio);
 660                         break;
 661                 case 0:
 662                         break;
 663                 default:
 664                         WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
 665                 }
 666         }
 667         blk_finish_plug(&plug);
 668 }
 669
 670 /*
 671  * If we run out of usable paths, should we queue I/O or error it?
 672  */
 673 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
 674                             bool save_old_value)
 675 {
 676         unsigned long flags;
 677
 678         spin_lock_irqsave(&m->lock, flags);
 679         assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags,
 680                    (save_old_value && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) ||
 681                    (!save_old_value && queue_if_no_path));
 682         assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
 683         spin_unlock_irqrestore(&m->lock, flags);
 684
 685         if (!queue_if_no_path) {
 686                 dm_table_run_md_queue_async(m->ti->table);
 687                 process_queued_io_list(m);
 688         }
 689
 690         return 0;
 691 }
 692
 693 /*
 694  * An event is triggered whenever a path is taken out of use.
 695  * Includes path failure and PG bypass.
 696  */
 697 static void trigger_event(struct work_struct *work)
 698 {
 699         struct multipath *m =
 700                 container_of(work, struct multipath, trigger_event);
 701
 702         dm_table_event(m->ti->table);
 703 }
 704
 705 /*-----------------------------------------------------------------
 706  * Constructor/argument parsing:
 707  * <#multipath feature args> [<arg>]*
 708  * <#hw_handler args> [hw_handler [<arg>]*]
 709  * <#priority groups>
 710  * <initial priority group>
 711  *     [<selector> <#selector args> [<arg>]*
 712  *      <#paths> <#per-path selector args>
 713  *         [<path> [<arg>]* ]+ ]+
 714  *---------------------------------------------------------------*/
 715 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 716                                struct dm_target *ti)
 717 {
 718         int r;
 719         struct path_selector_type *pst;
 720         unsigned ps_argc;
 721
 722         static const struct dm_arg _args[] = {
 723                 {0, 1024, "invalid number of path selector args"},
 724         };
 725
 726         pst = dm_get_path_selector(dm_shift_arg(as));
 727         if (!pst) {
 728                 ti->error = "unknown path selector type";
 729                 return -EINVAL;
 730         }
 731
 732         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 733         if (r) {
 734                 dm_put_path_selector(pst);
 735                 return -EINVAL;
 736         }
 737
 738         r = pst->create(&pg->ps, ps_argc, as->argv);
 739         if (r) {
 740                 dm_put_path_selector(pst);
 741                 ti->error = "path selector constructor failed";
 742                 return r;
 743         }
 744
 745         pg->ps.type = pst;
 746         dm_consume_args(as, ps_argc);
 747
 748         return 0;
 749 }
 750
 751 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 752                                struct dm_target *ti)
 753 {
 754         int r;
 755         struct pgpath *p;
 756         struct multipath *m = ti->private;
 757         struct request_queue *q = NULL;
 758         const char *attached_handler_name;
 759
 760         /* we need at least a path arg */
 761         if (as->argc < 1) {
 762                 ti->error = "no device given";
 763                 return ERR_PTR(-EINVAL);
 764         }
 765
 766         p = alloc_pgpath();
 767         if (!p)
 768                 return ERR_PTR(-ENOMEM);
 769
 770         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 771                           &p->path.dev);
 772         if (r) {
 773                 ti->error = "error getting device";
 774                 goto bad;
 775         }
 776
 777         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
 778                 q = bdev_get_queue(p->path.dev->bdev);
 779
 780         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
 781 retain:
 782                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 783                 if (attached_handler_name) {
 784                         /*
 785                          * Clear any hw_handler_params associated with a
 786                          * handler that isn't already attached.
 787                          */
 788                         if (m->hw_handler_name && strcmp(attached_handler_name, m->hw_handler_name)) {
 789                                 kfree(m->hw_handler_params);
 790                                 m->hw_handler_params = NULL;
 791                         }
 792
 793                         /*
 794                          * Reset hw_handler_name to match the attached handler
 795                          *
 796                          * NB. This modifies the table line to show the actual
 797                          * handler instead of the original table passed in.
 798                          */
 799                         kfree(m->hw_handler_name);
 800                         m->hw_handler_name = attached_handler_name;
 801                 }
 802         }
 803
 804         if (m->hw_handler_name) {
 805                 r = scsi_dh_attach(q, m->hw_handler_name);
 806                 if (r == -EBUSY) {
 807                         char b[BDEVNAME_SIZE];
 808
 809                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
 810                                 bdevname(p->path.dev->bdev, b));
 811                         goto retain;
 812                 }
 813                 if (r < 0) {
 814                         ti->error = "error attaching hardware handler";
 815                         dm_put_device(ti, p->path.dev);
 816                         goto bad;
 817                 }
 818
 819                 if (m->hw_handler_params) {
 820                         r = scsi_dh_set_params(q, m->hw_handler_params);
 821                         if (r < 0) {
 822                                 ti->error = "unable to set hardware "
 823                                                         "handler parameters";
 824                                 dm_put_device(ti, p->path.dev);
 825                                 goto bad;
 826                         }
 827                 }
 828         }
 829
 830         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 831         if (r) {
 832                 dm_put_device(ti, p->path.dev);
 833                 goto bad;
 834         }
 835
 836         return p;
 837
 838  bad:
 839         free_pgpath(p);
 840         return ERR_PTR(r);
 841 }
 842
 843 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 844                                                    struct multipath *m)
 845 {
 846         static const struct dm_arg _args[] = {
 847                 {1, 1024, "invalid number of paths"},
 848                 {0, 1024, "invalid number of selector args"}
 849         };
 850
 851         int r;
 852         unsigned i, nr_selector_args, nr_args;
 853         struct priority_group *pg;
 854         struct dm_target *ti = m->ti;
 855
 856         if (as->argc < 2) {
 857                 as->argc = 0;
 858                 ti->error = "not enough priority group arguments";
 859                 return ERR_PTR(-EINVAL);
 860         }
 861
 862         pg = alloc_priority_group();
 863         if (!pg) {
 864                 ti->error = "couldn't allocate priority group";
 865                 return ERR_PTR(-ENOMEM);
 866         }
 867         pg->m = m;
 868
 869         r = parse_path_selector(as, pg, ti);
 870         if (r)
 871                 goto bad;
 872
 873         /*
 874          * read the paths
 875          */
 876         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 877         if (r)
 878                 goto bad;
 879
 880         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 881         if (r)
 882                 goto bad;
 883
 884         nr_args = 1 + nr_selector_args;
 885         for (i = 0; i < pg->nr_pgpaths; i++) {
 886                 struct pgpath *pgpath;
 887                 struct dm_arg_set path_args;
 888
 889                 if (as->argc < nr_args) {
 890                         ti->error = "not enough path parameters";
 891                         r = -EINVAL;
 892                         goto bad;
 893                 }
 894
 895                 path_args.argc = nr_args;
 896                 path_args.argv = as->argv;
 897
 898                 pgpath = parse_path(&path_args, &pg->ps, ti);
 899                 if (IS_ERR(pgpath)) {
 900                         r = PTR_ERR(pgpath);
 901                         goto bad;
 902                 }
 903
 904                 pgpath->pg = pg;
 905                 list_add_tail(&pgpath->list, &pg->pgpaths);
 906                 dm_consume_args(as, nr_args);
 907         }
 908
 909         return pg;
 910
 911  bad:
 912         free_priority_group(pg, ti);
 913         return ERR_PTR(r);
 914 }
 915
 916 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 917 {
 918         unsigned hw_argc;
 919         int ret;
 920         struct dm_target *ti = m->ti;
 921
 922         static const struct dm_arg _args[] = {
 923                 {0, 1024, "invalid number of hardware handler args"},
 924         };
 925
 926         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 927                 return -EINVAL;
 928
 929         if (!hw_argc)
 930                 return 0;
 931
 932         if (m->queue_mode == DM_TYPE_BIO_BASED) {
 933                 dm_consume_args(as, hw_argc);
 934                 DMERR("bio-based multipath doesn't allow hardware handler args");
 935                 return 0;
 936         }
 937
 938         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 939         if (!m->hw_handler_name)
 940                 return -EINVAL;
 941
 942         if (hw_argc > 1) {
 943                 char *p;
 944                 int i, j, len = 4;
 945
 946                 for (i = 0; i <= hw_argc - 2; i++)
 947                         len += strlen(as->argv[i]) + 1;
 948                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 949                 if (!p) {
 950                         ti->error = "memory allocation failed";
 951                         ret = -ENOMEM;
 952                         goto fail;
 953                 }
 954                 j = sprintf(p, "%d", hw_argc - 1);
 955                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 956                         j = sprintf(p, "%s", as->argv[i]);
 957         }
 958         dm_consume_args(as, hw_argc - 1);
 959
 960         return 0;
 961 fail:
 962         kfree(m->hw_handler_name);
 963         m->hw_handler_name = NULL;
 964         return ret;
 965 }
 966
 967 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 968 {
 969         int r;
 970         unsigned argc;
 971         struct dm_target *ti = m->ti;
 972         const char *arg_name;
 973
 974         static const struct dm_arg _args[] = {
 975                 {0, 8, "invalid number of feature args"},
 976                 {1, 50, "pg_init_retries must be between 1 and 50"},
 977                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 978         };
 979
 980         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 981         if (r)
 982                 return -EINVAL;
 983
 984         if (!argc)
 985                 return 0;
 986
 987         do {
 988                 arg_name = dm_shift_arg(as);
 989                 argc--;
 990
 991                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 992                         r = queue_if_no_path(m, true, false);
 993                         continue;
 994                 }
 995
 996                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
 997                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 998                         continue;
 999                 }
1000
1001                 if (!strcasecmp(arg_name, "pg_init_retries") &&
1002                     (argc >= 1)) {
1003                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1004                         argc--;
1005                         continue;
1006                 }
1007
1008                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1009                     (argc >= 1)) {
1010                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1011                         argc--;
1012                         continue;
1013                 }
1014
1015                 if (!strcasecmp(arg_name, "queue_mode") &&
1016                     (argc >= 1)) {
1017                         const char *queue_mode_name = dm_shift_arg(as);
1018
1019                         if (!strcasecmp(queue_mode_name, "bio"))
1020                                 m->queue_mode = DM_TYPE_BIO_BASED;
1021                         else if (!strcasecmp(queue_mode_name, "rq"))
1022                                 m->queue_mode = DM_TYPE_REQUEST_BASED;
1023                         else if (!strcasecmp(queue_mode_name, "mq"))
1024                                 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1025                         else {
1026                                 ti->error = "Unknown 'queue_mode' requested";
1027                                 r = -EINVAL;
1028                         }
1029                         argc--;
1030                         continue;
1031                 }
1032
1033                 ti->error = "Unrecognised multipath feature request";
1034                 r = -EINVAL;
1035         } while (argc && !r);
1036
1037         return r;
1038 }
1039
1040 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1041 {
1042         /* target arguments */
1043         static const struct dm_arg _args[] = {
1044                 {0, 1024, "invalid number of priority groups"},
1045                 {0, 1024, "invalid initial priority group number"},
1046         };
1047
1048         int r;
1049         struct multipath *m;
1050         struct dm_arg_set as;
1051         unsigned pg_count = 0;
1052         unsigned next_pg_num;
1053
1054         as.argc = argc;
1055         as.argv = argv;
1056
1057         m = alloc_multipath(ti);
1058         if (!m) {
1059                 ti->error = "can't allocate multipath";
1060                 return -EINVAL;
1061         }
1062
1063         r = parse_features(&as, m);
1064         if (r)
1065                 goto bad;
1066
1067         r = alloc_multipath_stage2(ti, m);
1068         if (r)
1069                 goto bad;
1070
1071         r = parse_hw_handler(&as, m);
1072         if (r)
1073                 goto bad;
1074
1075         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1076         if (r)
1077                 goto bad;
1078
1079         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1080         if (r)
1081                 goto bad;
1082
1083         if ((!m->nr_priority_groups && next_pg_num) ||
1084             (m->nr_priority_groups && !next_pg_num)) {
1085                 ti->error = "invalid initial priority group";
1086                 r = -EINVAL;
1087                 goto bad;
1088         }
1089
1090         /* parse the priority groups */
1091         while (as.argc) {
1092                 struct priority_group *pg;
1093                 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1094
1095                 pg = parse_priority_group(&as, m);
1096                 if (IS_ERR(pg)) {
1097                         r = PTR_ERR(pg);
1098                         goto bad;
1099                 }
1100
1101                 nr_valid_paths += pg->nr_pgpaths;
1102                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1103
1104                 list_add_tail(&pg->list, &m->priority_groups);
1105                 pg_count++;
1106                 pg->pg_num = pg_count;
1107                 if (!--next_pg_num)
1108                         m->next_pg = pg;
1109         }
1110
1111         if (pg_count != m->nr_priority_groups) {
1112                 ti->error = "priority group count mismatch";
1113                 r = -EINVAL;
1114                 goto bad;
1115         }
1116
1117         ti->num_flush_bios = 1;
1118         ti->num_discard_bios = 1;
1119         ti->num_write_same_bios = 1;
1120         ti->num_write_zeroes_bios = 1;
1121         if (m->queue_mode == DM_TYPE_BIO_BASED)
1122                 ti->per_io_data_size = multipath_per_bio_data_size();
1123         else
1124                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1125
1126         return 0;
1127
1128  bad:
1129         free_multipath(m);
1130         return r;
1131 }
1132
1133 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1134 {
1135         DEFINE_WAIT(wait);
1136
1137         while (1) {
1138                 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1139
1140                 if (!atomic_read(&m->pg_init_in_progress))
1141                         break;
1142
1143                 io_schedule();
1144         }
1145         finish_wait(&m->pg_init_wait, &wait);
1146 }
1147
1148 static void flush_multipath_work(struct multipath *m)
1149 {
1150         set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1151         smp_mb__after_atomic();
1152
1153         flush_workqueue(kmpath_handlerd);
1154         multipath_wait_for_pg_init_completion(m);
1155         flush_workqueue(kmultipathd);
1156         flush_work(&m->trigger_event);
1157
1158         clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1159         smp_mb__after_atomic();
1160 }
1161
1162 static void multipath_dtr(struct dm_target *ti)
1163 {
1164         struct multipath *m = ti->private;
1165
1166         flush_multipath_work(m);
1167         free_multipath(m);
1168 }
1169
1170 /*
1171  * Take a path out of use.
1172  */
1173 static int fail_path(struct pgpath *pgpath)
1174 {
1175         unsigned long flags;
1176         struct multipath *m = pgpath->pg->m;
1177
1178         spin_lock_irqsave(&m->lock, flags);
1179
1180         if (!pgpath->is_active)
1181                 goto out;
1182
1183         DMWARN("Failing path %s.", pgpath->path.dev->name);
1184
1185         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1186         pgpath->is_active = false;
1187         pgpath->fail_count++;
1188
1189         atomic_dec(&m->nr_valid_paths);
1190
1191         if (pgpath == m->current_pgpath)
1192                 m->current_pgpath = NULL;
1193
1194         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1195                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1196
1197         schedule_work(&m->trigger_event);
1198
1199 out:
1200         spin_unlock_irqrestore(&m->lock, flags);
1201
1202         return 0;
1203 }
1204
1205 /*
1206  * Reinstate a previously-failed path
1207  */
1208 static int reinstate_path(struct pgpath *pgpath)
1209 {
1210         int r = 0, run_queue = 0;
1211         unsigned long flags;
1212         struct multipath *m = pgpath->pg->m;
1213         unsigned nr_valid_paths;
1214
1215         spin_lock_irqsave(&m->lock, flags);
1216
1217         if (pgpath->is_active)
1218                 goto out;
1219
1220         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1221
1222         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1223         if (r)
1224                 goto out;
1225
1226         pgpath->is_active = true;
1227
1228         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1229         if (nr_valid_paths == 1) {
1230                 m->current_pgpath = NULL;
1231                 run_queue = 1;
1232         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1233                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1234                         atomic_inc(&m->pg_init_in_progress);
1235         }
1236
1237         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1238                        pgpath->path.dev->name, nr_valid_paths);
1239
1240         schedule_work(&m->trigger_event);
1241
1242 out:
1243         spin_unlock_irqrestore(&m->lock, flags);
1244         if (run_queue) {
1245                 dm_table_run_md_queue_async(m->ti->table);
1246                 process_queued_io_list(m);
1247         }
1248
1249         return r;
1250 }
1251
1252 /*
1253  * Fail or reinstate all paths that match the provided struct dm_dev.
1254  */
1255 static int action_dev(struct multipath *m, struct dm_dev *dev,
1256                       action_fn action)
1257 {
1258         int r = -EINVAL;
1259         struct pgpath *pgpath;
1260         struct priority_group *pg;
1261
1262         list_for_each_entry(pg, &m->priority_groups, list) {
1263                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1264                         if (pgpath->path.dev == dev)
1265                                 r = action(pgpath);
1266                 }
1267         }
1268
1269         return r;
1270 }
1271
1272 /*
1273  * Temporarily try to avoid having to use the specified PG
1274  */
1275 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1276                       bool bypassed)
1277 {
1278         unsigned long flags;
1279
1280         spin_lock_irqsave(&m->lock, flags);
1281
1282         pg->bypassed = bypassed;
1283         m->current_pgpath = NULL;
1284         m->current_pg = NULL;
1285
1286         spin_unlock_irqrestore(&m->lock, flags);
1287
1288         schedule_work(&m->trigger_event);
1289 }
1290
1291 /*
1292  * Switch to using the specified PG from the next I/O that gets mapped
1293  */
1294 static int switch_pg_num(struct multipath *m, const char *pgstr)
1295 {
1296         struct priority_group *pg;
1297         unsigned pgnum;
1298         unsigned long flags;
1299         char dummy;
1300
1301         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1302             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1303                 DMWARN("invalid PG number supplied to switch_pg_num");
1304                 return -EINVAL;
1305         }
1306
1307         spin_lock_irqsave(&m->lock, flags);
1308         list_for_each_entry(pg, &m->priority_groups, list) {
1309                 pg->bypassed = false;
1310                 if (--pgnum)
1311                         continue;
1312
1313                 m->current_pgpath = NULL;
1314                 m->current_pg = NULL;
1315                 m->next_pg = pg;
1316         }
1317         spin_unlock_irqrestore(&m->lock, flags);
1318
1319         schedule_work(&m->trigger_event);
1320         return 0;
1321 }
1322
1323 /*
1324  * Set/clear bypassed status of a PG.
1325  * PGs are numbered upwards from 1 in the order they were declared.
1326  */
1327 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1328 {
1329         struct priority_group *pg;
1330         unsigned pgnum;
1331         char dummy;
1332
1333         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1334             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1335                 DMWARN("invalid PG number supplied to bypass_pg");
1336                 return -EINVAL;
1337         }
1338
1339         list_for_each_entry(pg, &m->priority_groups, list) {
1340                 if (!--pgnum)
1341                         break;
1342         }
1343
1344         bypass_pg(m, pg, bypassed);
1345         return 0;
1346 }
1347
1348 /*
1349  * Should we retry pg_init immediately?
1350  */
1351 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1352 {
1353         unsigned long flags;
1354         bool limit_reached = false;
1355
1356         spin_lock_irqsave(&m->lock, flags);
1357
1358         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1359             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1360                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1361         else
1362                 limit_reached = true;
1363
1364         spin_unlock_irqrestore(&m->lock, flags);
1365
1366         return limit_reached;
1367 }
1368
1369 static void pg_init_done(void *data, int errors)
1370 {
1371         struct pgpath *pgpath = data;
1372         struct priority_group *pg = pgpath->pg;
1373         struct multipath *m = pg->m;
1374         unsigned long flags;
1375         bool delay_retry = false;
1376
1377         /* device or driver problems */
1378         switch (errors) {
1379         case SCSI_DH_OK:
1380                 break;
1381         case SCSI_DH_NOSYS:
1382                 if (!m->hw_handler_name) {
1383                         errors = 0;
1384                         break;
1385                 }
1386                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1387                       "Error %d.", m->hw_handler_name, errors);
1388                 /*
1389                  * Fail path for now, so we do not ping pong
1390                  */
1391                 fail_path(pgpath);
1392                 break;
1393         case SCSI_DH_DEV_TEMP_BUSY:
1394                 /*
1395                  * Probably doing something like FW upgrade on the
1396                  * controller so try the other pg.
1397                  */
1398                 bypass_pg(m, pg, true);
1399                 break;
1400         case SCSI_DH_RETRY:
1401                 /* Wait before retrying. */
1402                 delay_retry = 1;
1403                 /* fall through */
1404         case SCSI_DH_IMM_RETRY:
1405         case SCSI_DH_RES_TEMP_UNAVAIL:
1406                 if (pg_init_limit_reached(m, pgpath))
1407                         fail_path(pgpath);
1408                 errors = 0;
1409                 break;
1410         case SCSI_DH_DEV_OFFLINED:
1411         default:
1412                 /*
1413                  * We probably do not want to fail the path for a device
1414                  * error, but this is what the old dm did. In future
1415                  * patches we can do more advanced handling.
1416                  */
1417                 fail_path(pgpath);
1418         }
1419
1420         spin_lock_irqsave(&m->lock, flags);
1421         if (errors) {
1422                 if (pgpath == m->current_pgpath) {
1423                         DMERR("Could not failover device. Error %d.", errors);
1424                         m->current_pgpath = NULL;
1425                         m->current_pg = NULL;
1426                 }
1427         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1428                 pg->bypassed = false;
1429
1430         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1431                 /* Activations of other paths are still on going */
1432                 goto out;
1433
1434         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1435                 if (delay_retry)
1436                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1437                 else
1438                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1439
1440                 if (__pg_init_all_paths(m))
1441                         goto out;
1442         }
1443         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1444
1445         process_queued_io_list(m);
1446
1447         /*
1448          * Wake up any thread waiting to suspend.
1449          */
1450         wake_up(&m->pg_init_wait);
1451
1452 out:
1453         spin_unlock_irqrestore(&m->lock, flags);
1454 }
1455
1456 static void activate_or_offline_path(struct pgpath *pgpath)
1457 {
1458         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1459
1460         if (pgpath->is_active && !blk_queue_dying(q))
1461                 scsi_dh_activate(q, pg_init_done, pgpath);
1462         else
1463                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1464 }
1465
1466 static void activate_path_work(struct work_struct *work)
1467 {
1468         struct pgpath *pgpath =
1469                 container_of(work, struct pgpath, activate_path.work);
1470
1471         activate_or_offline_path(pgpath);
1472 }
1473
1474 static int noretry_error(blk_status_t error)
1475 {
1476         switch (error) {
1477         case BLK_STS_NOTSUPP:
1478         case BLK_STS_NOSPC:
1479         case BLK_STS_TARGET:
1480         case BLK_STS_NEXUS:
1481         case BLK_STS_MEDIUM:
1482                 return 1;
1483         }
1484
1485         /* Anything else could be a path failure, so should be retried */
1486         return 0;
1487 }
1488
1489 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1490                             blk_status_t error, union map_info *map_context)
1491 {
1492         struct dm_mpath_io *mpio = get_mpio(map_context);
1493         struct pgpath *pgpath = mpio->pgpath;
1494         int r = DM_ENDIO_DONE;
1495
1496         /*
1497          * We don't queue any clone request inside the multipath target
1498          * during end I/O handling, since those clone requests don't have
1499          * bio clones.  If we queue them inside the multipath target,
1500          * we need to make bio clones, that requires memory allocation.
1501          * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1502          *  don't have bio clones.)
1503          * Instead of queueing the clone request here, we queue the original
1504          * request into dm core, which will remake a clone request and
1505          * clone bios for it and resubmit it later.
1506          */
1507         if (error && !noretry_error(error)) {
1508                 struct multipath *m = ti->private;
1509
1510                 r = DM_ENDIO_REQUEUE;
1511
1512                 if (pgpath)
1513                         fail_path(pgpath);
1514
1515                 if (atomic_read(&m->nr_valid_paths) == 0 &&
1516                     !must_push_back_rq(m)) {
1517                         if (error == BLK_STS_IOERR)
1518                                 dm_report_EIO(m);
1519                         /* complete with the original error */
1520                         r = DM_ENDIO_DONE;
1521                 }
1522         }
1523
1524         if (pgpath) {
1525                 struct path_selector *ps = &pgpath->pg->ps;
1526
1527                 if (ps->type->end_io)
1528                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1529         }
1530
1531         return r;
1532 }
1533
1534 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1535                 blk_status_t *error)
1536 {
1537         struct multipath *m = ti->private;
1538         struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1539         struct pgpath *pgpath = mpio->pgpath;
1540         unsigned long flags;
1541         int r = DM_ENDIO_DONE;
1542
1543         if (!*error || noretry_error(*error))
1544                 goto done;
1545
1546         if (pgpath)
1547                 fail_path(pgpath);
1548
1549         if (atomic_read(&m->nr_valid_paths) == 0 &&
1550             !test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1551                 if (must_push_back_bio(m)) {
1552                         r = DM_ENDIO_REQUEUE;
1553                 } else {
1554                         dm_report_EIO(m);
1555                         *error = BLK_STS_IOERR;
1556                 }
1557                 goto done;
1558         }
1559
1560         /* Queue for the daemon to resubmit */
1561         dm_bio_restore(get_bio_details_from_bio(clone), clone);
1562
1563         spin_lock_irqsave(&m->lock, flags);
1564         bio_list_add(&m->queued_bios, clone);
1565         spin_unlock_irqrestore(&m->lock, flags);
1566         if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1567                 queue_work(kmultipathd, &m->process_queued_bios);
1568
1569         r = DM_ENDIO_INCOMPLETE;
1570 done:
1571         if (pgpath) {
1572                 struct path_selector *ps = &pgpath->pg->ps;
1573
1574                 if (ps->type->end_io)
1575                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1576         }
1577
1578         return r;
1579 }
1580
1581 /*
1582  * Suspend can't complete until all the I/O is processed so if
1583  * the last path fails we must error any remaining I/O.
1584  * Note that if the freeze_bdev fails while suspending, the
1585  * queue_if_no_path state is lost - userspace should reset it.
1586  */
1587 static void multipath_presuspend(struct dm_target *ti)
1588 {
1589         struct multipath *m = ti->private;
1590
1591         queue_if_no_path(m, false, true);
1592 }
1593
1594 static void multipath_postsuspend(struct dm_target *ti)
1595 {
1596         struct multipath *m = ti->private;
1597
1598         mutex_lock(&m->work_mutex);
1599         flush_multipath_work(m);
1600         mutex_unlock(&m->work_mutex);
1601 }
1602
1603 /*
1604  * Restore the queue_if_no_path setting.
1605  */
1606 static void multipath_resume(struct dm_target *ti)
1607 {
1608         struct multipath *m = ti->private;
1609         unsigned long flags;
1610
1611         spin_lock_irqsave(&m->lock, flags);
1612         assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags,
1613                    test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1614         spin_unlock_irqrestore(&m->lock, flags);
1615 }
1616
1617 /*
1618  * Info output has the following format:
1619  * num_multipath_feature_args [multipath_feature_args]*
1620  * num_handler_status_args [handler_status_args]*
1621  * num_groups init_group_number
1622  *            [A|D|E num_ps_status_args [ps_status_args]*
1623  *             num_paths num_selector_args
1624  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1625  *
1626  * Table output has the following format (identical to the constructor string):
1627  * num_feature_args [features_args]*
1628  * num_handler_args hw_handler [hw_handler_args]*
1629  * num_groups init_group_number
1630  *     [priority selector-name num_ps_args [ps_args]*
1631  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1632  */
1633 static void multipath_status(struct dm_target *ti, status_type_t type,
1634                              unsigned status_flags, char *result, unsigned maxlen)
1635 {
1636         int sz = 0;
1637         unsigned long flags;
1638         struct multipath *m = ti->private;
1639         struct priority_group *pg;
1640         struct pgpath *p;
1641         unsigned pg_num;
1642         char state;
1643
1644         spin_lock_irqsave(&m->lock, flags);
1645
1646         /* Features */
1647         if (type == STATUSTYPE_INFO)
1648                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1649                        atomic_read(&m->pg_init_count));
1650         else {
1651                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1652                               (m->pg_init_retries > 0) * 2 +
1653                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1654                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1655                               (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1656
1657                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1658                         DMEMIT("queue_if_no_path ");
1659                 if (m->pg_init_retries)
1660                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1661                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1662                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1663                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1664                         DMEMIT("retain_attached_hw_handler ");
1665                 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1666                         switch(m->queue_mode) {
1667                         case DM_TYPE_BIO_BASED:
1668                                 DMEMIT("queue_mode bio ");
1669                                 break;
1670                         case DM_TYPE_MQ_REQUEST_BASED:
1671                                 DMEMIT("queue_mode mq ");
1672                                 break;
1673                         default:
1674                                 WARN_ON_ONCE(true);
1675                                 break;
1676                         }
1677                 }
1678         }
1679
1680         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1681                 DMEMIT("0 ");
1682         else
1683                 DMEMIT("1 %s ", m->hw_handler_name);
1684
1685         DMEMIT("%u ", m->nr_priority_groups);
1686
1687         if (m->next_pg)
1688                 pg_num = m->next_pg->pg_num;
1689         else if (m->current_pg)
1690                 pg_num = m->current_pg->pg_num;
1691         else
1692                 pg_num = (m->nr_priority_groups ? 1 : 0);
1693
1694         DMEMIT("%u ", pg_num);
1695
1696         switch (type) {
1697         case STATUSTYPE_INFO:
1698                 list_for_each_entry(pg, &m->priority_groups, list) {
1699                         if (pg->bypassed)
1700                                 state = 'D';    /* Disabled */
1701                         else if (pg == m->current_pg)
1702                                 state = 'A';    /* Currently Active */
1703                         else
1704                                 state = 'E';    /* Enabled */
1705
1706                         DMEMIT("%c ", state);
1707
1708                         if (pg->ps.type->status)
1709                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1710                                                           result + sz,
1711                                                           maxlen - sz);
1712                         else
1713                                 DMEMIT("0 ");
1714
1715                         DMEMIT("%u %u ", pg->nr_pgpaths,
1716                                pg->ps.type->info_args);
1717
1718                         list_for_each_entry(p, &pg->pgpaths, list) {
1719                                 DMEMIT("%s %s %u ", p->path.dev->name,
1720                                        p->is_active ? "A" : "F",
1721                                        p->fail_count);
1722                                 if (pg->ps.type->status)
1723                                         sz += pg->ps.type->status(&pg->ps,
1724                                               &p->path, type, result + sz,
1725                                               maxlen - sz);
1726                         }
1727                 }
1728                 break;
1729
1730         case STATUSTYPE_TABLE:
1731                 list_for_each_entry(pg, &m->priority_groups, list) {
1732                         DMEMIT("%s ", pg->ps.type->name);
1733
1734                         if (pg->ps.type->status)
1735                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1736                                                           result + sz,
1737                                                           maxlen - sz);
1738                         else
1739                                 DMEMIT("0 ");
1740
1741                         DMEMIT("%u %u ", pg->nr_pgpaths,
1742                                pg->ps.type->table_args);
1743
1744                         list_for_each_entry(p, &pg->pgpaths, list) {
1745                                 DMEMIT("%s ", p->path.dev->name);
1746                                 if (pg->ps.type->status)
1747                                         sz += pg->ps.type->status(&pg->ps,
1748                                               &p->path, type, result + sz,
1749                                               maxlen - sz);
1750                         }
1751                 }
1752                 break;
1753         }
1754
1755         spin_unlock_irqrestore(&m->lock, flags);
1756 }
1757
1758 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1759 {
1760         int r = -EINVAL;
1761         struct dm_dev *dev;
1762         struct multipath *m = ti->private;
1763         action_fn action;
1764
1765         mutex_lock(&m->work_mutex);
1766
1767         if (dm_suspended(ti)) {
1768                 r = -EBUSY;
1769                 goto out;
1770         }
1771
1772         if (argc == 1) {
1773                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1774                         r = queue_if_no_path(m, true, false);
1775                         goto out;
1776                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1777                         r = queue_if_no_path(m, false, false);
1778                         goto out;
1779                 }
1780         }
1781
1782         if (argc != 2) {
1783                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1784                 goto out;
1785         }
1786
1787         if (!strcasecmp(argv[0], "disable_group")) {
1788                 r = bypass_pg_num(m, argv[1], true);
1789                 goto out;
1790         } else if (!strcasecmp(argv[0], "enable_group")) {
1791                 r = bypass_pg_num(m, argv[1], false);
1792                 goto out;
1793         } else if (!strcasecmp(argv[0], "switch_group")) {
1794                 r = switch_pg_num(m, argv[1]);
1795                 goto out;
1796         } else if (!strcasecmp(argv[0], "reinstate_path"))
1797                 action = reinstate_path;
1798         else if (!strcasecmp(argv[0], "fail_path"))
1799                 action = fail_path;
1800         else {
1801                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1802                 goto out;
1803         }
1804
1805         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1806         if (r) {
1807                 DMWARN("message: error getting device %s",
1808                        argv[1]);
1809                 goto out;
1810         }
1811
1812         r = action_dev(m, dev, action);
1813
1814         dm_put_device(ti, dev);
1815
1816 out:
1817         mutex_unlock(&m->work_mutex);
1818         return r;
1819 }
1820
1821 static int multipath_prepare_ioctl(struct dm_target *ti,
1822                 struct block_device **bdev, fmode_t *mode)
1823 {
1824         struct multipath *m = ti->private;
1825         struct pgpath *current_pgpath;
1826         int r;
1827
1828         current_pgpath = READ_ONCE(m->current_pgpath);
1829         if (!current_pgpath)
1830                 current_pgpath = choose_pgpath(m, 0);
1831
1832         if (current_pgpath) {
1833                 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1834                         *bdev = current_pgpath->path.dev->bdev;
1835                         *mode = current_pgpath->path.dev->mode;
1836                         r = 0;
1837                 } else {
1838                         /* pg_init has not started or completed */
1839                         r = -ENOTCONN;
1840                 }
1841         } else {
1842                 /* No path is available */
1843                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1844                         r = -ENOTCONN;
1845                 else
1846                         r = -EIO;
1847         }
1848
1849         if (r == -ENOTCONN) {
1850                 if (!READ_ONCE(m->current_pg)) {
1851                         /* Path status changed, redo selection */
1852                         (void) choose_pgpath(m, 0);
1853                 }
1854                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1855                         pg_init_all_paths(m);
1856                 dm_table_run_md_queue_async(m->ti->table);
1857                 process_queued_io_list(m);
1858         }
1859
1860         /*
1861          * Only pass ioctls through if the device sizes match exactly.
1862          */
1863         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1864                 return 1;
1865         return r;
1866 }
1867
1868 static int multipath_iterate_devices(struct dm_target *ti,
1869                                      iterate_devices_callout_fn fn, void *data)
1870 {
1871         struct multipath *m = ti->private;
1872         struct priority_group *pg;
1873         struct pgpath *p;
1874         int ret = 0;
1875
1876         list_for_each_entry(pg, &m->priority_groups, list) {
1877                 list_for_each_entry(p, &pg->pgpaths, list) {
1878                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1879                         if (ret)
1880                                 goto out;
1881                 }
1882         }
1883
1884 out:
1885         return ret;
1886 }
1887
1888 static int pgpath_busy(struct pgpath *pgpath)
1889 {
1890         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1891
1892         return blk_lld_busy(q);
1893 }
1894
1895 /*
1896  * We return "busy", only when we can map I/Os but underlying devices
1897  * are busy (so even if we map I/Os now, the I/Os will wait on
1898  * the underlying queue).
1899  * In other words, if we want to kill I/Os or queue them inside us
1900  * due to map unavailability, we don't return "busy".  Otherwise,
1901  * dm core won't give us the I/Os and we can't do what we want.
1902  */
1903 static int multipath_busy(struct dm_target *ti)
1904 {
1905         bool busy = false, has_active = false;
1906         struct multipath *m = ti->private;
1907         struct priority_group *pg, *next_pg;
1908         struct pgpath *pgpath;
1909
1910         /* pg_init in progress */
1911         if (atomic_read(&m->pg_init_in_progress))
1912                 return true;
1913
1914         /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
1915         if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1916                 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
1917
1918         /* Guess which priority_group will be used at next mapping time */
1919         pg = READ_ONCE(m->current_pg);
1920         next_pg = READ_ONCE(m->next_pg);
1921         if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
1922                 pg = next_pg;
1923
1924         if (!pg) {
1925                 /*
1926                  * We don't know which pg will be used at next mapping time.
1927                  * We don't call choose_pgpath() here to avoid to trigger
1928                  * pg_init just by busy checking.
1929                  * So we don't know whether underlying devices we will be using
1930                  * at next mapping time are busy or not. Just try mapping.
1931                  */
1932                 return busy;
1933         }
1934
1935         /*
1936          * If there is one non-busy active path at least, the path selector
1937          * will be able to select it. So we consider such a pg as not busy.
1938          */
1939         busy = true;
1940         list_for_each_entry(pgpath, &pg->pgpaths, list) {
1941                 if (pgpath->is_active) {
1942                         has_active = true;
1943                         if (!pgpath_busy(pgpath)) {
1944                                 busy = false;
1945                                 break;
1946                         }
1947                 }
1948         }
1949
1950         if (!has_active) {
1951                 /*
1952                  * No active path in this pg, so this pg won't be used and
1953                  * the current_pg will be changed at next mapping time.
1954                  * We need to try mapping to determine it.
1955                  */
1956                 busy = false;
1957         }
1958
1959         return busy;
1960 }
1961
1962 /*-----------------------------------------------------------------
1963  * Module setup
1964  *---------------------------------------------------------------*/
1965 static struct target_type multipath_target = {
1966         .name = "multipath",
1967         .version = {1, 12, 0},
1968         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1969         .module = THIS_MODULE,
1970         .ctr = multipath_ctr,
1971         .dtr = multipath_dtr,
1972         .clone_and_map_rq = multipath_clone_and_map,
1973         .release_clone_rq = multipath_release_clone,
1974         .rq_end_io = multipath_end_io,
1975         .map = multipath_map_bio,
1976         .end_io = multipath_end_io_bio,
1977         .presuspend = multipath_presuspend,
1978         .postsuspend = multipath_postsuspend,
1979         .resume = multipath_resume,
1980         .status = multipath_status,
1981         .message = multipath_message,
1982         .prepare_ioctl = multipath_prepare_ioctl,
1983         .iterate_devices = multipath_iterate_devices,
1984         .busy = multipath_busy,
1985 };
1986
1987 static int __init dm_multipath_init(void)
1988 {
1989         int r;
1990
1991         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1992         if (!kmultipathd) {
1993                 DMERR("failed to create workqueue kmpathd");
1994                 r = -ENOMEM;
1995                 goto bad_alloc_kmultipathd;
1996         }
1997
1998         /*
1999          * A separate workqueue is used to handle the device handlers
2000          * to avoid overloading existing workqueue. Overloading the
2001          * old workqueue would also create a bottleneck in the
2002          * path of the storage hardware device activation.
2003          */
2004         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2005                                                   WQ_MEM_RECLAIM);
2006         if (!kmpath_handlerd) {
2007                 DMERR("failed to create workqueue kmpath_handlerd");
2008                 r = -ENOMEM;
2009                 goto bad_alloc_kmpath_handlerd;
2010         }
2011
2012         r = dm_register_target(&multipath_target);
2013         if (r < 0) {
2014                 DMERR("request-based register failed %d", r);
2015                 r = -EINVAL;
2016                 goto bad_register_target;
2017         }
2018
2019         return 0;
2020
2021 bad_register_target:
2022         destroy_workqueue(kmpath_handlerd);
2023 bad_alloc_kmpath_handlerd:
2024         destroy_workqueue(kmultipathd);
2025 bad_alloc_kmultipathd:
2026         return r;
2027 }
2028
2029 static void __exit dm_multipath_exit(void)
2030 {
2031         destroy_workqueue(kmpath_handlerd);
2032         destroy_workqueue(kmultipathd);
2033
2034         dm_unregister_target(&multipath_target);
2035 }
2036
2037 module_init(dm_multipath_init);
2038 module_exit(dm_multipath_exit);
2039
2040 MODULE_DESCRIPTION(DM_NAME " multipath target");
2041 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2042 MODULE_LICENSE("GPL");