drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9
  10 #include "dm-rq.h"
  11 #include "dm-bio-record.h"
  12 #include "dm-path-selector.h"
  13 #include "dm-uevent.h"
  14
  15 #include <linux/blkdev.h>
  16 #include <linux/ctype.h>
  17 #include <linux/init.h>
  18 #include <linux/mempool.h>
  19 #include <linux/module.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/slab.h>
  22 #include <linux/time.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/delay.h>
  25 #include <scsi/scsi_dh.h>
  26 #include <linux/atomic.h>
  27 #include <linux/blk-mq.h>
  28
  29 #define DM_MSG_PREFIX "multipath"
  30 #define DM_PG_INIT_DELAY_MSECS 2000
  31 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  32
  33 /* Path properties */
  34 struct pgpath {
  35         struct list_head list;
  36
  37         struct priority_group *pg;      /* Owning PG */
  38         unsigned fail_count;            /* Cumulative failure count */
  39
  40         struct dm_path path;
  41         struct delayed_work activate_path;
  42
  43         bool is_active:1;               /* Path status */
  44 };
  45
  46 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  47
  48 /*
  49  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  50  * Each has a path selector which controls which path gets used.
  51  */
  52 struct priority_group {
  53         struct list_head list;
  54
  55         struct multipath *m;            /* Owning multipath instance */
  56         struct path_selector ps;
  57
  58         unsigned pg_num;                /* Reference number */
  59         unsigned nr_pgpaths;            /* Number of paths in PG */
  60         struct list_head pgpaths;
  61
  62         bool bypassed:1;                /* Temporarily bypass this PG? */
  63 };
  64
  65 /* Multipath context */
  66 struct multipath {
  67         struct list_head list;
  68         struct dm_target *ti;
  69
  70         const char *hw_handler_name;
  71         char *hw_handler_params;
  72
  73         spinlock_t lock;
  74
  75         unsigned nr_priority_groups;
  76         struct list_head priority_groups;
  77
  78         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  79
  80         struct pgpath *current_pgpath;
  81         struct priority_group *current_pg;
  82         struct priority_group *next_pg; /* Switch to this PG if set */
  83
  84         unsigned long flags;            /* Multipath state flags */
  85
  86         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  87         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  88
  89         atomic_t nr_valid_paths;        /* Total number of usable paths */
  90         atomic_t pg_init_in_progress;   /* Only one pg_init allowed at once */
  91         atomic_t pg_init_count;         /* Number of times pg_init called */
  92
  93         unsigned queue_mode;
  94
  95         /*
  96          * We must use a mempool of dm_mpath_io structs so that we
  97          * can resubmit bios on error.
  98          */
  99         mempool_t *mpio_pool;
 100
 101         struct mutex work_mutex;
 102         struct work_struct trigger_event;
 103
 104         struct work_struct process_queued_bios;
 105         struct bio_list queued_bios;
 106 };
 107
 108 /*
 109  * Context information attached to each io we process.
 110  */
 111 struct dm_mpath_io {
 112         struct pgpath *pgpath;
 113         size_t nr_bytes;
 114 };
 115
 116 typedef int (*action_fn) (struct pgpath *pgpath);
 117
 118 static struct kmem_cache *_mpio_cache;
 119
 120 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 121 static void trigger_event(struct work_struct *work);
 122 static void activate_or_offline_path(struct pgpath *pgpath);
 123 static void activate_path_work(struct work_struct *work);
 124 static void process_queued_bios(struct work_struct *work);
 125
 126 /*-----------------------------------------------
 127  * Multipath state flags.
 128  *-----------------------------------------------*/
 129
 130 #define MPATHF_QUEUE_IO 0                       /* Must we queue all I/O? */
 131 #define MPATHF_QUEUE_IF_NO_PATH 1               /* Queue I/O if last path fails? */
 132 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2         /* Saved state during suspension */
 133 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3     /* If there's already a hw_handler present, don't change it. */
 134 #define MPATHF_PG_INIT_DISABLED 4               /* pg_init is not currently allowed */
 135 #define MPATHF_PG_INIT_REQUIRED 5               /* pg_init needs calling? */
 136 #define MPATHF_PG_INIT_DELAY_RETRY 6            /* Delay pg_init retry? */
 137
 138 /*-----------------------------------------------
 139  * Allocation routines
 140  *-----------------------------------------------*/
 141
 142 static struct pgpath *alloc_pgpath(void)
 143 {
 144         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 145
 146         if (pgpath) {
 147                 pgpath->is_active = true;
 148                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path_work);
 149         }
 150
 151         return pgpath;
 152 }
 153
 154 static void free_pgpath(struct pgpath *pgpath)
 155 {
 156         kfree(pgpath);
 157 }
 158
 159 static struct priority_group *alloc_priority_group(void)
 160 {
 161         struct priority_group *pg;
 162
 163         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 164
 165         if (pg)
 166                 INIT_LIST_HEAD(&pg->pgpaths);
 167
 168         return pg;
 169 }
 170
 171 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 172 {
 173         struct pgpath *pgpath, *tmp;
 174
 175         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 176                 list_del(&pgpath->list);
 177                 dm_put_device(ti, pgpath->path.dev);
 178                 free_pgpath(pgpath);
 179         }
 180 }
 181
 182 static void free_priority_group(struct priority_group *pg,
 183                                 struct dm_target *ti)
 184 {
 185         struct path_selector *ps = &pg->ps;
 186
 187         if (ps->type) {
 188                 ps->type->destroy(ps);
 189                 dm_put_path_selector(ps->type);
 190         }
 191
 192         free_pgpaths(&pg->pgpaths, ti);
 193         kfree(pg);
 194 }
 195
 196 static struct multipath *alloc_multipath(struct dm_target *ti)
 197 {
 198         struct multipath *m;
 199
 200         m = kzalloc(sizeof(*m), GFP_KERNEL);
 201         if (m) {
 202                 INIT_LIST_HEAD(&m->priority_groups);
 203                 spin_lock_init(&m->lock);
 204                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 205                 atomic_set(&m->nr_valid_paths, 0);
 206                 atomic_set(&m->pg_init_in_progress, 0);
 207                 atomic_set(&m->pg_init_count, 0);
 208                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 209                 INIT_WORK(&m->trigger_event, trigger_event);
 210                 init_waitqueue_head(&m->pg_init_wait);
 211                 mutex_init(&m->work_mutex);
 212
 213                 m->mpio_pool = NULL;
 214                 m->queue_mode = DM_TYPE_NONE;
 215
 216                 m->ti = ti;
 217                 ti->private = m;
 218         }
 219
 220         return m;
 221 }
 222
 223 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
 224 {
 225         if (m->queue_mode == DM_TYPE_NONE) {
 226                 /*
 227                  * Default to request-based.
 228                  */
 229                 if (dm_use_blk_mq(dm_table_get_md(ti->table)))
 230                         m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
 231                 else
 232                         m->queue_mode = DM_TYPE_REQUEST_BASED;
 233         }
 234
 235         if (m->queue_mode == DM_TYPE_REQUEST_BASED) {
 236                 unsigned min_ios = dm_get_reserved_rq_based_ios();
 237
 238                 m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
 239                 if (!m->mpio_pool)
 240                         return -ENOMEM;
 241         }
 242         else if (m->queue_mode == DM_TYPE_BIO_BASED) {
 243                 INIT_WORK(&m->process_queued_bios, process_queued_bios);
 244                 /*
 245                  * bio-based doesn't support any direct scsi_dh management;
 246                  * it just discovers if a scsi_dh is attached.
 247                  */
 248                 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 249         }
 250
 251         dm_table_set_type(ti->table, m->queue_mode);
 252
 253         return 0;
 254 }
 255
 256 static void free_multipath(struct multipath *m)
 257 {
 258         struct priority_group *pg, *tmp;
 259
 260         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 261                 list_del(&pg->list);
 262                 free_priority_group(pg, m->ti);
 263         }
 264
 265         kfree(m->hw_handler_name);
 266         kfree(m->hw_handler_params);
 267         mempool_destroy(m->mpio_pool);
 268         kfree(m);
 269 }
 270
 271 static struct dm_mpath_io *get_mpio(union map_info *info)
 272 {
 273         return info->ptr;
 274 }
 275
 276 static struct dm_mpath_io *set_mpio(struct multipath *m, union map_info *info)
 277 {
 278         struct dm_mpath_io *mpio;
 279
 280         if (!m->mpio_pool) {
 281                 /* Use blk-mq pdu memory requested via per_io_data_size */
 282                 mpio = get_mpio(info);
 283                 memset(mpio, 0, sizeof(*mpio));
 284                 return mpio;
 285         }
 286
 287         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 288         if (!mpio)
 289                 return NULL;
 290
 291         memset(mpio, 0, sizeof(*mpio));
 292         info->ptr = mpio;
 293
 294         return mpio;
 295 }
 296
 297 static void clear_request_fn_mpio(struct multipath *m, union map_info *info)
 298 {
 299         /* Only needed for non blk-mq (.request_fn) multipath */
 300         if (m->mpio_pool) {
 301                 struct dm_mpath_io *mpio = info->ptr;
 302
 303                 info->ptr = NULL;
 304                 mempool_free(mpio, m->mpio_pool);
 305         }
 306 }
 307
 308 static size_t multipath_per_bio_data_size(void)
 309 {
 310         return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
 311 }
 312
 313 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
 314 {
 315         return dm_per_bio_data(bio, multipath_per_bio_data_size());
 316 }
 317
 318 static struct dm_bio_details *get_bio_details_from_bio(struct bio *bio)
 319 {
 320         /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
 321         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 322         void *bio_details = mpio + 1;
 323
 324         return bio_details;
 325 }
 326
 327 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p,
 328                                         struct dm_bio_details **bio_details_p)
 329 {
 330         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 331         struct dm_bio_details *bio_details = get_bio_details_from_bio(bio);
 332
 333         memset(mpio, 0, sizeof(*mpio));
 334         memset(bio_details, 0, sizeof(*bio_details));
 335         dm_bio_record(bio_details, bio);
 336
 337         if (mpio_p)
 338                 *mpio_p = mpio;
 339         if (bio_details_p)
 340                 *bio_details_p = bio_details;
 341 }
 342
 343 /*-----------------------------------------------
 344  * Path selection
 345  *-----------------------------------------------*/
 346
 347 static int __pg_init_all_paths(struct multipath *m)
 348 {
 349         struct pgpath *pgpath;
 350         unsigned long pg_init_delay = 0;
 351
 352         if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
 353                 return 0;
 354
 355         atomic_inc(&m->pg_init_count);
 356         clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 357
 358         /* Check here to reset pg_init_required */
 359         if (!m->current_pg)
 360                 return 0;
 361
 362         if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
 363                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 364                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 365         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 366                 /* Skip failed paths */
 367                 if (!pgpath->is_active)
 368                         continue;
 369                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 370                                        pg_init_delay))
 371                         atomic_inc(&m->pg_init_in_progress);
 372         }
 373         return atomic_read(&m->pg_init_in_progress);
 374 }
 375
 376 static int pg_init_all_paths(struct multipath *m)
 377 {
 378         int r;
 379         unsigned long flags;
 380
 381         spin_lock_irqsave(&m->lock, flags);
 382         r = __pg_init_all_paths(m);
 383         spin_unlock_irqrestore(&m->lock, flags);
 384
 385         return r;
 386 }
 387
 388 static void __switch_pg(struct multipath *m, struct priority_group *pg)
 389 {
 390         m->current_pg = pg;
 391
 392         /* Must we initialise the PG first, and queue I/O till it's ready? */
 393         if (m->hw_handler_name) {
 394                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 395                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 396         } else {
 397                 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 398                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 399         }
 400
 401         atomic_set(&m->pg_init_count, 0);
 402 }
 403
 404 static struct pgpath *choose_path_in_pg(struct multipath *m,
 405                                         struct priority_group *pg,
 406                                         size_t nr_bytes)
 407 {
 408         unsigned long flags;
 409         struct dm_path *path;
 410         struct pgpath *pgpath;
 411
 412         path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 413         if (!path)
 414                 return ERR_PTR(-ENXIO);
 415
 416         pgpath = path_to_pgpath(path);
 417
 418         if (unlikely(lockless_dereference(m->current_pg) != pg)) {
 419                 /* Only update current_pgpath if pg changed */
 420                 spin_lock_irqsave(&m->lock, flags);
 421                 m->current_pgpath = pgpath;
 422                 __switch_pg(m, pg);
 423                 spin_unlock_irqrestore(&m->lock, flags);
 424         }
 425
 426         return pgpath;
 427 }
 428
 429 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
 430 {
 431         unsigned long flags;
 432         struct priority_group *pg;
 433         struct pgpath *pgpath;
 434         unsigned bypassed = 1;
 435
 436         if (!atomic_read(&m->nr_valid_paths)) {
 437                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 438                 goto failed;
 439         }
 440
 441         /* Were we instructed to switch PG? */
 442         if (lockless_dereference(m->next_pg)) {
 443                 spin_lock_irqsave(&m->lock, flags);
 444                 pg = m->next_pg;
 445                 if (!pg) {
 446                         spin_unlock_irqrestore(&m->lock, flags);
 447                         goto check_current_pg;
 448                 }
 449                 m->next_pg = NULL;
 450                 spin_unlock_irqrestore(&m->lock, flags);
 451                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 452                 if (!IS_ERR_OR_NULL(pgpath))
 453                         return pgpath;
 454         }
 455
 456         /* Don't change PG until it has no remaining paths */
 457 check_current_pg:
 458         pg = lockless_dereference(m->current_pg);
 459         if (pg) {
 460                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 461                 if (!IS_ERR_OR_NULL(pgpath))
 462                         return pgpath;
 463         }
 464
 465         /*
 466          * Loop through priority groups until we find a valid path.
 467          * First time we skip PGs marked 'bypassed'.
 468          * Second time we only try the ones we skipped, but set
 469          * pg_init_delay_retry so we do not hammer controllers.
 470          */
 471         do {
 472                 list_for_each_entry(pg, &m->priority_groups, list) {
 473                         if (pg->bypassed == !!bypassed)
 474                                 continue;
 475                         pgpath = choose_path_in_pg(m, pg, nr_bytes);
 476                         if (!IS_ERR_OR_NULL(pgpath)) {
 477                                 if (!bypassed)
 478                                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
 479                                 return pgpath;
 480                         }
 481                 }
 482         } while (bypassed--);
 483
 484 failed:
 485         spin_lock_irqsave(&m->lock, flags);
 486         m->current_pgpath = NULL;
 487         m->current_pg = NULL;
 488         spin_unlock_irqrestore(&m->lock, flags);
 489
 490         return NULL;
 491 }
 492
 493 /*
 494  * Check whether bios must be queued in the device-mapper core rather
 495  * than here in the target.
 496  *
 497  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 498  * same value then we are not between multipath_presuspend()
 499  * and multipath_resume() calls and we have no need to check
 500  * for the DMF_NOFLUSH_SUSPENDING flag.
 501  */
 502 static bool __must_push_back(struct multipath *m)
 503 {
 504         return ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) !=
 505                  test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) &&
 506                 dm_noflush_suspending(m->ti));
 507 }
 508
 509 static bool must_push_back_rq(struct multipath *m)
 510 {
 511         bool r;
 512         unsigned long flags;
 513
 514         spin_lock_irqsave(&m->lock, flags);
 515         r = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) ||
 516              __must_push_back(m));
 517         spin_unlock_irqrestore(&m->lock, flags);
 518
 519         return r;
 520 }
 521
 522 static bool must_push_back_bio(struct multipath *m)
 523 {
 524         bool r;
 525         unsigned long flags;
 526
 527         spin_lock_irqsave(&m->lock, flags);
 528         r = __must_push_back(m);
 529         spin_unlock_irqrestore(&m->lock, flags);
 530
 531         return r;
 532 }
 533
 534 /*
 535  * Map cloned requests (request-based multipath)
 536  */
 537 static int __multipath_map(struct dm_target *ti, struct request *clone,
 538                            union map_info *map_context,
 539                            struct request *rq, struct request **__clone)
 540 {
 541         struct multipath *m = ti->private;
 542         int r = DM_MAPIO_REQUEUE;
 543         size_t nr_bytes = clone ? blk_rq_bytes(clone) : blk_rq_bytes(rq);
 544         struct pgpath *pgpath;
 545         struct block_device *bdev;
 546         struct dm_mpath_io *mpio;
 547
 548         /* Do we need to select a new pgpath? */
 549         pgpath = lockless_dereference(m->current_pgpath);
 550         if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
 551                 pgpath = choose_pgpath(m, nr_bytes);
 552
 553         if (!pgpath) {
 554                 if (must_push_back_rq(m))
 555                         return DM_MAPIO_DELAY_REQUEUE;
 556                 return -EIO;    /* Failed */
 557         } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
 558                    test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
 559                 pg_init_all_paths(m);
 560                 return r;
 561         }
 562
 563         mpio = set_mpio(m, map_context);
 564         if (!mpio)
 565                 /* ENOMEM, requeue */
 566                 return r;
 567
 568         mpio->pgpath = pgpath;
 569         mpio->nr_bytes = nr_bytes;
 570
 571         bdev = pgpath->path.dev->bdev;
 572
 573         if (clone) {
 574                 /*
 575                  * Old request-based interface: allocated clone is passed in.
 576                  * Used by: .request_fn stacked on .request_fn path(s).
 577                  */
 578                 clone->q = bdev_get_queue(bdev);
 579                 clone->rq_disk = bdev->bd_disk;
 580                 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 581         } else {
 582                 /*
 583                  * blk-mq request-based interface; used by both:
 584                  * .request_fn stacked on blk-mq path(s) and
 585                  * blk-mq stacked on blk-mq path(s).
 586                  */
 587                 *__clone = blk_mq_alloc_request(bdev_get_queue(bdev),
 588                                                 rq_data_dir(rq), BLK_MQ_REQ_NOWAIT);
 589                 if (IS_ERR(*__clone)) {
 590                         /* ENOMEM, requeue */
 591                         clear_request_fn_mpio(m, map_context);
 592                         return r;
 593                 }
 594                 (*__clone)->bio = (*__clone)->biotail = NULL;
 595                 (*__clone)->rq_disk = bdev->bd_disk;
 596                 (*__clone)->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 597         }
 598
 599         if (pgpath->pg->ps.type->start_io)
 600                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 601                                               &pgpath->path,
 602                                               nr_bytes);
 603         return DM_MAPIO_REMAPPED;
 604 }
 605
 606 static int multipath_map(struct dm_target *ti, struct request *clone,
 607                          union map_info *map_context)
 608 {
 609         return __multipath_map(ti, clone, map_context, NULL, NULL);
 610 }
 611
 612 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 613                                    union map_info *map_context,
 614                                    struct request **clone)
 615 {
 616         return __multipath_map(ti, NULL, map_context, rq, clone);
 617 }
 618
 619 static void multipath_release_clone(struct request *clone)
 620 {
 621         blk_mq_free_request(clone);
 622 }
 623
 624 /*
 625  * Map cloned bios (bio-based multipath)
 626  */
 627 static int __multipath_map_bio(struct multipath *m, struct bio *bio, struct dm_mpath_io *mpio)
 628 {
 629         size_t nr_bytes = bio->bi_iter.bi_size;
 630         struct pgpath *pgpath;
 631         unsigned long flags;
 632         bool queue_io;
 633
 634         /* Do we need to select a new pgpath? */
 635         pgpath = lockless_dereference(m->current_pgpath);
 636         queue_io = test_bit(MPATHF_QUEUE_IO, &m->flags);
 637         if (!pgpath || !queue_io)
 638                 pgpath = choose_pgpath(m, nr_bytes);
 639
 640         if ((pgpath && queue_io) ||
 641             (!pgpath && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))) {
 642                 /* Queue for the daemon to resubmit */
 643                 spin_lock_irqsave(&m->lock, flags);
 644                 bio_list_add(&m->queued_bios, bio);
 645                 spin_unlock_irqrestore(&m->lock, flags);
 646                 /* PG_INIT_REQUIRED cannot be set without QUEUE_IO */
 647                 if (queue_io || test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
 648                         pg_init_all_paths(m);
 649                 else if (!queue_io)
 650                         queue_work(kmultipathd, &m->process_queued_bios);
 651                 return DM_MAPIO_SUBMITTED;
 652         }
 653
 654         if (!pgpath) {
 655                 if (!must_push_back_bio(m))
 656                         return -EIO;
 657                 return DM_MAPIO_REQUEUE;
 658         }
 659
 660         mpio->pgpath = pgpath;
 661         mpio->nr_bytes = nr_bytes;
 662
 663         bio->bi_error = 0;
 664         bio->bi_bdev = pgpath->path.dev->bdev;
 665         bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
 666
 667         if (pgpath->pg->ps.type->start_io)
 668                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 669                                               &pgpath->path,
 670                                               nr_bytes);
 671         return DM_MAPIO_REMAPPED;
 672 }
 673
 674 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
 675 {
 676         struct multipath *m = ti->private;
 677         struct dm_mpath_io *mpio = NULL;
 678
 679         multipath_init_per_bio_data(bio, &mpio, NULL);
 680
 681         return __multipath_map_bio(m, bio, mpio);
 682 }
 683
 684 static void process_queued_io_list(struct multipath *m)
 685 {
 686         if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
 687                 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
 688         else if (m->queue_mode == DM_TYPE_BIO_BASED)
 689                 queue_work(kmultipathd, &m->process_queued_bios);
 690 }
 691
 692 static void process_queued_bios(struct work_struct *work)
 693 {
 694         int r;
 695         unsigned long flags;
 696         struct bio *bio;
 697         struct bio_list bios;
 698         struct blk_plug plug;
 699         struct multipath *m =
 700                 container_of(work, struct multipath, process_queued_bios);
 701
 702         bio_list_init(&bios);
 703
 704         spin_lock_irqsave(&m->lock, flags);
 705
 706         if (bio_list_empty(&m->queued_bios)) {
 707                 spin_unlock_irqrestore(&m->lock, flags);
 708                 return;
 709         }
 710
 711         bio_list_merge(&bios, &m->queued_bios);
 712         bio_list_init(&m->queued_bios);
 713
 714         spin_unlock_irqrestore(&m->lock, flags);
 715
 716         blk_start_plug(&plug);
 717         while ((bio = bio_list_pop(&bios))) {
 718                 r = __multipath_map_bio(m, bio, get_mpio_from_bio(bio));
 719                 if (r < 0 || r == DM_MAPIO_REQUEUE) {
 720                         bio->bi_error = r;
 721                         bio_endio(bio);
 722                 } else if (r == DM_MAPIO_REMAPPED)
 723                         generic_make_request(bio);
 724         }
 725         blk_finish_plug(&plug);
 726 }
 727
 728 /*
 729  * If we run out of usable paths, should we queue I/O or error it?
 730  */
 731 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
 732                             bool save_old_value)
 733 {
 734         unsigned long flags;
 735
 736         spin_lock_irqsave(&m->lock, flags);
 737
 738         if (save_old_value) {
 739                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
 740                         set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 741                 else
 742                         clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 743         } else {
 744                 if (queue_if_no_path)
 745                         set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 746                 else
 747                         clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 748         }
 749         if (queue_if_no_path)
 750                 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 751         else
 752                 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 753
 754         spin_unlock_irqrestore(&m->lock, flags);
 755
 756         if (!queue_if_no_path) {
 757                 dm_table_run_md_queue_async(m->ti->table);
 758                 process_queued_io_list(m);
 759         }
 760
 761         return 0;
 762 }
 763
 764 /*
 765  * An event is triggered whenever a path is taken out of use.
 766  * Includes path failure and PG bypass.
 767  */
 768 static void trigger_event(struct work_struct *work)
 769 {
 770         struct multipath *m =
 771                 container_of(work, struct multipath, trigger_event);
 772
 773         dm_table_event(m->ti->table);
 774 }
 775
 776 /*-----------------------------------------------------------------
 777  * Constructor/argument parsing:
 778  * <#multipath feature args> [<arg>]*
 779  * <#hw_handler args> [hw_handler [<arg>]*]
 780  * <#priority groups>
 781  * <initial priority group>
 782  *     [<selector> <#selector args> [<arg>]*
 783  *      <#paths> <#per-path selector args>
 784  *         [<path> [<arg>]* ]+ ]+
 785  *---------------------------------------------------------------*/
 786 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 787                                struct dm_target *ti)
 788 {
 789         int r;
 790         struct path_selector_type *pst;
 791         unsigned ps_argc;
 792
 793         static struct dm_arg _args[] = {
 794                 {0, 1024, "invalid number of path selector args"},
 795         };
 796
 797         pst = dm_get_path_selector(dm_shift_arg(as));
 798         if (!pst) {
 799                 ti->error = "unknown path selector type";
 800                 return -EINVAL;
 801         }
 802
 803         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 804         if (r) {
 805                 dm_put_path_selector(pst);
 806                 return -EINVAL;
 807         }
 808
 809         r = pst->create(&pg->ps, ps_argc, as->argv);
 810         if (r) {
 811                 dm_put_path_selector(pst);
 812                 ti->error = "path selector constructor failed";
 813                 return r;
 814         }
 815
 816         pg->ps.type = pst;
 817         dm_consume_args(as, ps_argc);
 818
 819         return 0;
 820 }
 821
 822 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 823                                struct dm_target *ti)
 824 {
 825         int r;
 826         struct pgpath *p;
 827         struct multipath *m = ti->private;
 828         struct request_queue *q = NULL;
 829         const char *attached_handler_name;
 830
 831         /* we need at least a path arg */
 832         if (as->argc < 1) {
 833                 ti->error = "no device given";
 834                 return ERR_PTR(-EINVAL);
 835         }
 836
 837         p = alloc_pgpath();
 838         if (!p)
 839                 return ERR_PTR(-ENOMEM);
 840
 841         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 842                           &p->path.dev);
 843         if (r) {
 844                 ti->error = "error getting device";
 845                 goto bad;
 846         }
 847
 848         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
 849                 q = bdev_get_queue(p->path.dev->bdev);
 850
 851         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
 852 retain:
 853                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 854                 if (attached_handler_name) {
 855                         /*
 856                          * Reset hw_handler_name to match the attached handler
 857                          * and clear any hw_handler_params associated with the
 858                          * ignored handler.
 859                          *
 860                          * NB. This modifies the table line to show the actual
 861                          * handler instead of the original table passed in.
 862                          */
 863                         kfree(m->hw_handler_name);
 864                         m->hw_handler_name = attached_handler_name;
 865
 866                         kfree(m->hw_handler_params);
 867                         m->hw_handler_params = NULL;
 868                 }
 869         }
 870
 871         if (m->hw_handler_name) {
 872                 r = scsi_dh_attach(q, m->hw_handler_name);
 873                 if (r == -EBUSY) {
 874                         char b[BDEVNAME_SIZE];
 875
 876                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
 877                                 bdevname(p->path.dev->bdev, b));
 878                         goto retain;
 879                 }
 880                 if (r < 0) {
 881                         ti->error = "error attaching hardware handler";
 882                         dm_put_device(ti, p->path.dev);
 883                         goto bad;
 884                 }
 885
 886                 if (m->hw_handler_params) {
 887                         r = scsi_dh_set_params(q, m->hw_handler_params);
 888                         if (r < 0) {
 889                                 ti->error = "unable to set hardware "
 890                                                         "handler parameters";
 891                                 dm_put_device(ti, p->path.dev);
 892                                 goto bad;
 893                         }
 894                 }
 895         }
 896
 897         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 898         if (r) {
 899                 dm_put_device(ti, p->path.dev);
 900                 goto bad;
 901         }
 902
 903         return p;
 904
 905  bad:
 906         free_pgpath(p);
 907         return ERR_PTR(r);
 908 }
 909
 910 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 911                                                    struct multipath *m)
 912 {
 913         static struct dm_arg _args[] = {
 914                 {1, 1024, "invalid number of paths"},
 915                 {0, 1024, "invalid number of selector args"}
 916         };
 917
 918         int r;
 919         unsigned i, nr_selector_args, nr_args;
 920         struct priority_group *pg;
 921         struct dm_target *ti = m->ti;
 922
 923         if (as->argc < 2) {
 924                 as->argc = 0;
 925                 ti->error = "not enough priority group arguments";
 926                 return ERR_PTR(-EINVAL);
 927         }
 928
 929         pg = alloc_priority_group();
 930         if (!pg) {
 931                 ti->error = "couldn't allocate priority group";
 932                 return ERR_PTR(-ENOMEM);
 933         }
 934         pg->m = m;
 935
 936         r = parse_path_selector(as, pg, ti);
 937         if (r)
 938                 goto bad;
 939
 940         /*
 941          * read the paths
 942          */
 943         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 944         if (r)
 945                 goto bad;
 946
 947         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 948         if (r)
 949                 goto bad;
 950
 951         nr_args = 1 + nr_selector_args;
 952         for (i = 0; i < pg->nr_pgpaths; i++) {
 953                 struct pgpath *pgpath;
 954                 struct dm_arg_set path_args;
 955
 956                 if (as->argc < nr_args) {
 957                         ti->error = "not enough path parameters";
 958                         r = -EINVAL;
 959                         goto bad;
 960                 }
 961
 962                 path_args.argc = nr_args;
 963                 path_args.argv = as->argv;
 964
 965                 pgpath = parse_path(&path_args, &pg->ps, ti);
 966                 if (IS_ERR(pgpath)) {
 967                         r = PTR_ERR(pgpath);
 968                         goto bad;
 969                 }
 970
 971                 pgpath->pg = pg;
 972                 list_add_tail(&pgpath->list, &pg->pgpaths);
 973                 dm_consume_args(as, nr_args);
 974         }
 975
 976         return pg;
 977
 978  bad:
 979         free_priority_group(pg, ti);
 980         return ERR_PTR(r);
 981 }
 982
 983 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 984 {
 985         unsigned hw_argc;
 986         int ret;
 987         struct dm_target *ti = m->ti;
 988
 989         static struct dm_arg _args[] = {
 990                 {0, 1024, "invalid number of hardware handler args"},
 991         };
 992
 993         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 994                 return -EINVAL;
 995
 996         if (!hw_argc)
 997                 return 0;
 998
 999         if (m->queue_mode == DM_TYPE_BIO_BASED) {
1000                 dm_consume_args(as, hw_argc);
1001                 DMERR("bio-based multipath doesn't allow hardware handler args");
1002                 return 0;
1003         }
1004
1005         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1006
1007         if (hw_argc > 1) {
1008                 char *p;
1009                 int i, j, len = 4;
1010
1011                 for (i = 0; i <= hw_argc - 2; i++)
1012                         len += strlen(as->argv[i]) + 1;
1013                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1014                 if (!p) {
1015                         ti->error = "memory allocation failed";
1016                         ret = -ENOMEM;
1017                         goto fail;
1018                 }
1019                 j = sprintf(p, "%d", hw_argc - 1);
1020                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1021                         j = sprintf(p, "%s", as->argv[i]);
1022         }
1023         dm_consume_args(as, hw_argc - 1);
1024
1025         return 0;
1026 fail:
1027         kfree(m->hw_handler_name);
1028         m->hw_handler_name = NULL;
1029         return ret;
1030 }
1031
1032 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1033 {
1034         int r;
1035         unsigned argc;
1036         struct dm_target *ti = m->ti;
1037         const char *arg_name;
1038
1039         static struct dm_arg _args[] = {
1040                 {0, 8, "invalid number of feature args"},
1041                 {1, 50, "pg_init_retries must be between 1 and 50"},
1042                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1043         };
1044
1045         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1046         if (r)
1047                 return -EINVAL;
1048
1049         if (!argc)
1050                 return 0;
1051
1052         do {
1053                 arg_name = dm_shift_arg(as);
1054                 argc--;
1055
1056                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1057                         r = queue_if_no_path(m, true, false);
1058                         continue;
1059                 }
1060
1061                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1062                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1063                         continue;
1064                 }
1065
1066                 if (!strcasecmp(arg_name, "pg_init_retries") &&
1067                     (argc >= 1)) {
1068                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1069                         argc--;
1070                         continue;
1071                 }
1072
1073                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1074                     (argc >= 1)) {
1075                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1076                         argc--;
1077                         continue;
1078                 }
1079
1080                 if (!strcasecmp(arg_name, "queue_mode") &&
1081                     (argc >= 1)) {
1082                         const char *queue_mode_name = dm_shift_arg(as);
1083
1084                         if (!strcasecmp(queue_mode_name, "bio"))
1085                                 m->queue_mode = DM_TYPE_BIO_BASED;
1086                         else if (!strcasecmp(queue_mode_name, "rq"))
1087                                 m->queue_mode = DM_TYPE_REQUEST_BASED;
1088                         else if (!strcasecmp(queue_mode_name, "mq"))
1089                                 m->queue_mode = DM_TYPE_MQ_REQUEST_BASED;
1090                         else {
1091                                 ti->error = "Unknown 'queue_mode' requested";
1092                                 r = -EINVAL;
1093                         }
1094                         argc--;
1095                         continue;
1096                 }
1097
1098                 ti->error = "Unrecognised multipath feature request";
1099                 r = -EINVAL;
1100         } while (argc && !r);
1101
1102         return r;
1103 }
1104
1105 static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
1106 {
1107         /* target arguments */
1108         static struct dm_arg _args[] = {
1109                 {0, 1024, "invalid number of priority groups"},
1110                 {0, 1024, "invalid initial priority group number"},
1111         };
1112
1113         int r;
1114         struct multipath *m;
1115         struct dm_arg_set as;
1116         unsigned pg_count = 0;
1117         unsigned next_pg_num;
1118
1119         as.argc = argc;
1120         as.argv = argv;
1121
1122         m = alloc_multipath(ti);
1123         if (!m) {
1124                 ti->error = "can't allocate multipath";
1125                 return -EINVAL;
1126         }
1127
1128         r = parse_features(&as, m);
1129         if (r)
1130                 goto bad;
1131
1132         r = alloc_multipath_stage2(ti, m);
1133         if (r)
1134                 goto bad;
1135
1136         r = parse_hw_handler(&as, m);
1137         if (r)
1138                 goto bad;
1139
1140         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1141         if (r)
1142                 goto bad;
1143
1144         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1145         if (r)
1146                 goto bad;
1147
1148         if ((!m->nr_priority_groups && next_pg_num) ||
1149             (m->nr_priority_groups && !next_pg_num)) {
1150                 ti->error = "invalid initial priority group";
1151                 r = -EINVAL;
1152                 goto bad;
1153         }
1154
1155         /* parse the priority groups */
1156         while (as.argc) {
1157                 struct priority_group *pg;
1158                 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1159
1160                 pg = parse_priority_group(&as, m);
1161                 if (IS_ERR(pg)) {
1162                         r = PTR_ERR(pg);
1163                         goto bad;
1164                 }
1165
1166                 nr_valid_paths += pg->nr_pgpaths;
1167                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1168
1169                 list_add_tail(&pg->list, &m->priority_groups);
1170                 pg_count++;
1171                 pg->pg_num = pg_count;
1172                 if (!--next_pg_num)
1173                         m->next_pg = pg;
1174         }
1175
1176         if (pg_count != m->nr_priority_groups) {
1177                 ti->error = "priority group count mismatch";
1178                 r = -EINVAL;
1179                 goto bad;
1180         }
1181
1182         ti->num_flush_bios = 1;
1183         ti->num_discard_bios = 1;
1184         ti->num_write_same_bios = 1;
1185         if (m->queue_mode == DM_TYPE_BIO_BASED)
1186                 ti->per_io_data_size = multipath_per_bio_data_size();
1187         else if (m->queue_mode == DM_TYPE_MQ_REQUEST_BASED)
1188                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1189
1190         return 0;
1191
1192  bad:
1193         free_multipath(m);
1194         return r;
1195 }
1196
1197 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1198 {
1199         DEFINE_WAIT(wait);
1200
1201         while (1) {
1202                 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1203
1204                 if (!atomic_read(&m->pg_init_in_progress))
1205                         break;
1206
1207                 io_schedule();
1208         }
1209         finish_wait(&m->pg_init_wait, &wait);
1210 }
1211
1212 static void flush_multipath_work(struct multipath *m)
1213 {
1214         set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1215         smp_mb__after_atomic();
1216
1217         flush_workqueue(kmpath_handlerd);
1218         multipath_wait_for_pg_init_completion(m);
1219         flush_workqueue(kmultipathd);
1220         flush_work(&m->trigger_event);
1221
1222         clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1223         smp_mb__after_atomic();
1224 }
1225
1226 static void multipath_dtr(struct dm_target *ti)
1227 {
1228         struct multipath *m = ti->private;
1229
1230         flush_multipath_work(m);
1231         free_multipath(m);
1232 }
1233
1234 /*
1235  * Take a path out of use.
1236  */
1237 static int fail_path(struct pgpath *pgpath)
1238 {
1239         unsigned long flags;
1240         struct multipath *m = pgpath->pg->m;
1241
1242         spin_lock_irqsave(&m->lock, flags);
1243
1244         if (!pgpath->is_active)
1245                 goto out;
1246
1247         DMWARN("Failing path %s.", pgpath->path.dev->name);
1248
1249         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1250         pgpath->is_active = false;
1251         pgpath->fail_count++;
1252
1253         atomic_dec(&m->nr_valid_paths);
1254
1255         if (pgpath == m->current_pgpath)
1256                 m->current_pgpath = NULL;
1257
1258         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1259                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1260
1261         schedule_work(&m->trigger_event);
1262
1263 out:
1264         spin_unlock_irqrestore(&m->lock, flags);
1265
1266         return 0;
1267 }
1268
1269 /*
1270  * Reinstate a previously-failed path
1271  */
1272 static int reinstate_path(struct pgpath *pgpath)
1273 {
1274         int r = 0, run_queue = 0;
1275         unsigned long flags;
1276         struct multipath *m = pgpath->pg->m;
1277         unsigned nr_valid_paths;
1278
1279         spin_lock_irqsave(&m->lock, flags);
1280
1281         if (pgpath->is_active)
1282                 goto out;
1283
1284         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1285
1286         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1287         if (r)
1288                 goto out;
1289
1290         pgpath->is_active = true;
1291
1292         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1293         if (nr_valid_paths == 1) {
1294                 m->current_pgpath = NULL;
1295                 run_queue = 1;
1296         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1297                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1298                         atomic_inc(&m->pg_init_in_progress);
1299         }
1300
1301         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1302                        pgpath->path.dev->name, nr_valid_paths);
1303
1304         schedule_work(&m->trigger_event);
1305
1306 out:
1307         spin_unlock_irqrestore(&m->lock, flags);
1308         if (run_queue) {
1309                 dm_table_run_md_queue_async(m->ti->table);
1310                 process_queued_io_list(m);
1311         }
1312
1313         return r;
1314 }
1315
1316 /*
1317  * Fail or reinstate all paths that match the provided struct dm_dev.
1318  */
1319 static int action_dev(struct multipath *m, struct dm_dev *dev,
1320                       action_fn action)
1321 {
1322         int r = -EINVAL;
1323         struct pgpath *pgpath;
1324         struct priority_group *pg;
1325
1326         list_for_each_entry(pg, &m->priority_groups, list) {
1327                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1328                         if (pgpath->path.dev == dev)
1329                                 r = action(pgpath);
1330                 }
1331         }
1332
1333         return r;
1334 }
1335
1336 /*
1337  * Temporarily try to avoid having to use the specified PG
1338  */
1339 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1340                       bool bypassed)
1341 {
1342         unsigned long flags;
1343
1344         spin_lock_irqsave(&m->lock, flags);
1345
1346         pg->bypassed = bypassed;
1347         m->current_pgpath = NULL;
1348         m->current_pg = NULL;
1349
1350         spin_unlock_irqrestore(&m->lock, flags);
1351
1352         schedule_work(&m->trigger_event);
1353 }
1354
1355 /*
1356  * Switch to using the specified PG from the next I/O that gets mapped
1357  */
1358 static int switch_pg_num(struct multipath *m, const char *pgstr)
1359 {
1360         struct priority_group *pg;
1361         unsigned pgnum;
1362         unsigned long flags;
1363         char dummy;
1364
1365         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1366             (pgnum > m->nr_priority_groups)) {
1367                 DMWARN("invalid PG number supplied to switch_pg_num");
1368                 return -EINVAL;
1369         }
1370
1371         spin_lock_irqsave(&m->lock, flags);
1372         list_for_each_entry(pg, &m->priority_groups, list) {
1373                 pg->bypassed = false;
1374                 if (--pgnum)
1375                         continue;
1376
1377                 m->current_pgpath = NULL;
1378                 m->current_pg = NULL;
1379                 m->next_pg = pg;
1380         }
1381         spin_unlock_irqrestore(&m->lock, flags);
1382
1383         schedule_work(&m->trigger_event);
1384         return 0;
1385 }
1386
1387 /*
1388  * Set/clear bypassed status of a PG.
1389  * PGs are numbered upwards from 1 in the order they were declared.
1390  */
1391 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1392 {
1393         struct priority_group *pg;
1394         unsigned pgnum;
1395         char dummy;
1396
1397         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1398             (pgnum > m->nr_priority_groups)) {
1399                 DMWARN("invalid PG number supplied to bypass_pg");
1400                 return -EINVAL;
1401         }
1402
1403         list_for_each_entry(pg, &m->priority_groups, list) {
1404                 if (!--pgnum)
1405                         break;
1406         }
1407
1408         bypass_pg(m, pg, bypassed);
1409         return 0;
1410 }
1411
1412 /*
1413  * Should we retry pg_init immediately?
1414  */
1415 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1416 {
1417         unsigned long flags;
1418         bool limit_reached = false;
1419
1420         spin_lock_irqsave(&m->lock, flags);
1421
1422         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1423             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1424                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1425         else
1426                 limit_reached = true;
1427
1428         spin_unlock_irqrestore(&m->lock, flags);
1429
1430         return limit_reached;
1431 }
1432
1433 static void pg_init_done(void *data, int errors)
1434 {
1435         struct pgpath *pgpath = data;
1436         struct priority_group *pg = pgpath->pg;
1437         struct multipath *m = pg->m;
1438         unsigned long flags;
1439         bool delay_retry = false;
1440
1441         /* device or driver problems */
1442         switch (errors) {
1443         case SCSI_DH_OK:
1444                 break;
1445         case SCSI_DH_NOSYS:
1446                 if (!m->hw_handler_name) {
1447                         errors = 0;
1448                         break;
1449                 }
1450                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1451                       "Error %d.", m->hw_handler_name, errors);
1452                 /*
1453                  * Fail path for now, so we do not ping pong
1454                  */
1455                 fail_path(pgpath);
1456                 break;
1457         case SCSI_DH_DEV_TEMP_BUSY:
1458                 /*
1459                  * Probably doing something like FW upgrade on the
1460                  * controller so try the other pg.
1461                  */
1462                 bypass_pg(m, pg, true);
1463                 break;
1464         case SCSI_DH_RETRY:
1465                 /* Wait before retrying. */
1466                 delay_retry = 1;
1467         case SCSI_DH_IMM_RETRY:
1468         case SCSI_DH_RES_TEMP_UNAVAIL:
1469                 if (pg_init_limit_reached(m, pgpath))
1470                         fail_path(pgpath);
1471                 errors = 0;
1472                 break;
1473         case SCSI_DH_DEV_OFFLINED:
1474         default:
1475                 /*
1476                  * We probably do not want to fail the path for a device
1477                  * error, but this is what the old dm did. In future
1478                  * patches we can do more advanced handling.
1479                  */
1480                 fail_path(pgpath);
1481         }
1482
1483         spin_lock_irqsave(&m->lock, flags);
1484         if (errors) {
1485                 if (pgpath == m->current_pgpath) {
1486                         DMERR("Could not failover device. Error %d.", errors);
1487                         m->current_pgpath = NULL;
1488                         m->current_pg = NULL;
1489                 }
1490         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1491                 pg->bypassed = false;
1492
1493         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1494                 /* Activations of other paths are still on going */
1495                 goto out;
1496
1497         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1498                 if (delay_retry)
1499                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1500                 else
1501                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1502
1503                 if (__pg_init_all_paths(m))
1504                         goto out;
1505         }
1506         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1507
1508         process_queued_io_list(m);
1509
1510         /*
1511          * Wake up any thread waiting to suspend.
1512          */
1513         wake_up(&m->pg_init_wait);
1514
1515 out:
1516         spin_unlock_irqrestore(&m->lock, flags);
1517 }
1518
1519 static void activate_or_offline_path(struct pgpath *pgpath)
1520 {
1521         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1522
1523         if (pgpath->is_active && !blk_queue_dying(q))
1524                 scsi_dh_activate(q, pg_init_done, pgpath);
1525         else
1526                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1527 }
1528
1529 static void activate_path_work(struct work_struct *work)
1530 {
1531         struct pgpath *pgpath =
1532                 container_of(work, struct pgpath, activate_path.work);
1533
1534         activate_or_offline_path(pgpath);
1535 }
1536
1537 static int noretry_error(int error)
1538 {
1539         switch (error) {
1540         case -EBADE:
1541                 /*
1542                  * EBADE signals an reservation conflict.
1543                  * We shouldn't fail the path here as we can communicate with
1544                  * the target.  We should failover to the next path, but in
1545                  * doing so we might be causing a ping-pong between paths.
1546                  * So just return the reservation conflict error.
1547                  */
1548         case -EOPNOTSUPP:
1549         case -EREMOTEIO:
1550         case -EILSEQ:
1551         case -ENODATA:
1552         case -ENOSPC:
1553                 return 1;
1554         }
1555
1556         /* Anything else could be a path failure, so should be retried */
1557         return 0;
1558 }
1559
1560 /*
1561  * end_io handling
1562  */
1563 static int do_end_io(struct multipath *m, struct request *clone,
1564                      int error, struct dm_mpath_io *mpio)
1565 {
1566         /*
1567          * We don't queue any clone request inside the multipath target
1568          * during end I/O handling, since those clone requests don't have
1569          * bio clones.  If we queue them inside the multipath target,
1570          * we need to make bio clones, that requires memory allocation.
1571          * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1572          *  don't have bio clones.)
1573          * Instead of queueing the clone request here, we queue the original
1574          * request into dm core, which will remake a clone request and
1575          * clone bios for it and resubmit it later.
1576          */
1577         int r = DM_ENDIO_REQUEUE;
1578
1579         if (!error && !clone->errors)
1580                 return 0;       /* I/O complete */
1581
1582         if (noretry_error(error))
1583                 return error;
1584
1585         if (mpio->pgpath)
1586                 fail_path(mpio->pgpath);
1587
1588         if (!atomic_read(&m->nr_valid_paths)) {
1589                 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1590                         if (!must_push_back_rq(m))
1591                                 r = -EIO;
1592                 }
1593         }
1594
1595         return r;
1596 }
1597
1598 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1599                             int error, union map_info *map_context)
1600 {
1601         struct multipath *m = ti->private;
1602         struct dm_mpath_io *mpio = get_mpio(map_context);
1603         struct pgpath *pgpath;
1604         struct path_selector *ps;
1605         int r;
1606
1607         BUG_ON(!mpio);
1608
1609         r = do_end_io(m, clone, error, mpio);
1610         pgpath = mpio->pgpath;
1611         if (pgpath) {
1612                 ps = &pgpath->pg->ps;
1613                 if (ps->type->end_io)
1614                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1615         }
1616         clear_request_fn_mpio(m, map_context);
1617
1618         return r;
1619 }
1620
1621 static int do_end_io_bio(struct multipath *m, struct bio *clone,
1622                          int error, struct dm_mpath_io *mpio)
1623 {
1624         unsigned long flags;
1625
1626         if (!error)
1627                 return 0;       /* I/O complete */
1628
1629         if (noretry_error(error))
1630                 return error;
1631
1632         if (mpio->pgpath)
1633                 fail_path(mpio->pgpath);
1634
1635         if (!atomic_read(&m->nr_valid_paths)) {
1636                 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1637                         if (!must_push_back_bio(m))
1638                                 return -EIO;
1639                         return DM_ENDIO_REQUEUE;
1640                 }
1641         }
1642
1643         /* Queue for the daemon to resubmit */
1644         dm_bio_restore(get_bio_details_from_bio(clone), clone);
1645
1646         spin_lock_irqsave(&m->lock, flags);
1647         bio_list_add(&m->queued_bios, clone);
1648         spin_unlock_irqrestore(&m->lock, flags);
1649         if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
1650                 queue_work(kmultipathd, &m->process_queued_bios);
1651
1652         return DM_ENDIO_INCOMPLETE;
1653 }
1654
1655 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone, int error)
1656 {
1657         struct multipath *m = ti->private;
1658         struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1659         struct pgpath *pgpath;
1660         struct path_selector *ps;
1661         int r;
1662
1663         BUG_ON(!mpio);
1664
1665         r = do_end_io_bio(m, clone, error, mpio);
1666         pgpath = mpio->pgpath;
1667         if (pgpath) {
1668                 ps = &pgpath->pg->ps;
1669                 if (ps->type->end_io)
1670                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1671         }
1672
1673         return r;
1674 }
1675
1676 /*
1677  * Suspend can't complete until all the I/O is processed so if
1678  * the last path fails we must error any remaining I/O.
1679  * Note that if the freeze_bdev fails while suspending, the
1680  * queue_if_no_path state is lost - userspace should reset it.
1681  */
1682 static void multipath_presuspend(struct dm_target *ti)
1683 {
1684         struct multipath *m = ti->private;
1685
1686         queue_if_no_path(m, false, true);
1687 }
1688
1689 static void multipath_postsuspend(struct dm_target *ti)
1690 {
1691         struct multipath *m = ti->private;
1692
1693         mutex_lock(&m->work_mutex);
1694         flush_multipath_work(m);
1695         mutex_unlock(&m->work_mutex);
1696 }
1697
1698 /*
1699  * Restore the queue_if_no_path setting.
1700  */
1701 static void multipath_resume(struct dm_target *ti)
1702 {
1703         struct multipath *m = ti->private;
1704         unsigned long flags;
1705
1706         spin_lock_irqsave(&m->lock, flags);
1707         if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags))
1708                 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1709         else
1710                 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1711         spin_unlock_irqrestore(&m->lock, flags);
1712 }
1713
1714 /*
1715  * Info output has the following format:
1716  * num_multipath_feature_args [multipath_feature_args]*
1717  * num_handler_status_args [handler_status_args]*
1718  * num_groups init_group_number
1719  *            [A|D|E num_ps_status_args [ps_status_args]*
1720  *             num_paths num_selector_args
1721  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1722  *
1723  * Table output has the following format (identical to the constructor string):
1724  * num_feature_args [features_args]*
1725  * num_handler_args hw_handler [hw_handler_args]*
1726  * num_groups init_group_number
1727  *     [priority selector-name num_ps_args [ps_args]*
1728  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1729  */
1730 static void multipath_status(struct dm_target *ti, status_type_t type,
1731                              unsigned status_flags, char *result, unsigned maxlen)
1732 {
1733         int sz = 0;
1734         unsigned long flags;
1735         struct multipath *m = ti->private;
1736         struct priority_group *pg;
1737         struct pgpath *p;
1738         unsigned pg_num;
1739         char state;
1740
1741         spin_lock_irqsave(&m->lock, flags);
1742
1743         /* Features */
1744         if (type == STATUSTYPE_INFO)
1745                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1746                        atomic_read(&m->pg_init_count));
1747         else {
1748                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1749                               (m->pg_init_retries > 0) * 2 +
1750                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1751                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1752                               (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1753
1754                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1755                         DMEMIT("queue_if_no_path ");
1756                 if (m->pg_init_retries)
1757                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1758                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1759                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1760                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1761                         DMEMIT("retain_attached_hw_handler ");
1762                 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1763                         switch(m->queue_mode) {
1764                         case DM_TYPE_BIO_BASED:
1765                                 DMEMIT("queue_mode bio ");
1766                                 break;
1767                         case DM_TYPE_MQ_REQUEST_BASED:
1768                                 DMEMIT("queue_mode mq ");
1769                                 break;
1770                         }
1771                 }
1772         }
1773
1774         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1775                 DMEMIT("0 ");
1776         else
1777                 DMEMIT("1 %s ", m->hw_handler_name);
1778
1779         DMEMIT("%u ", m->nr_priority_groups);
1780
1781         if (m->next_pg)
1782                 pg_num = m->next_pg->pg_num;
1783         else if (m->current_pg)
1784                 pg_num = m->current_pg->pg_num;
1785         else
1786                 pg_num = (m->nr_priority_groups ? 1 : 0);
1787
1788         DMEMIT("%u ", pg_num);
1789
1790         switch (type) {
1791         case STATUSTYPE_INFO:
1792                 list_for_each_entry(pg, &m->priority_groups, list) {
1793                         if (pg->bypassed)
1794                                 state = 'D';    /* Disabled */
1795                         else if (pg == m->current_pg)
1796                                 state = 'A';    /* Currently Active */
1797                         else
1798                                 state = 'E';    /* Enabled */
1799
1800                         DMEMIT("%c ", state);
1801
1802                         if (pg->ps.type->status)
1803                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1804                                                           result + sz,
1805                                                           maxlen - sz);
1806                         else
1807                                 DMEMIT("0 ");
1808
1809                         DMEMIT("%u %u ", pg->nr_pgpaths,
1810                                pg->ps.type->info_args);
1811
1812                         list_for_each_entry(p, &pg->pgpaths, list) {
1813                                 DMEMIT("%s %s %u ", p->path.dev->name,
1814                                        p->is_active ? "A" : "F",
1815                                        p->fail_count);
1816                                 if (pg->ps.type->status)
1817                                         sz += pg->ps.type->status(&pg->ps,
1818                                               &p->path, type, result + sz,
1819                                               maxlen - sz);
1820                         }
1821                 }
1822                 break;
1823
1824         case STATUSTYPE_TABLE:
1825                 list_for_each_entry(pg, &m->priority_groups, list) {
1826                         DMEMIT("%s ", pg->ps.type->name);
1827
1828                         if (pg->ps.type->status)
1829                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1830                                                           result + sz,
1831                                                           maxlen - sz);
1832                         else
1833                                 DMEMIT("0 ");
1834
1835                         DMEMIT("%u %u ", pg->nr_pgpaths,
1836                                pg->ps.type->table_args);
1837
1838                         list_for_each_entry(p, &pg->pgpaths, list) {
1839                                 DMEMIT("%s ", p->path.dev->name);
1840                                 if (pg->ps.type->status)
1841                                         sz += pg->ps.type->status(&pg->ps,
1842                                               &p->path, type, result + sz,
1843                                               maxlen - sz);
1844                         }
1845                 }
1846                 break;
1847         }
1848
1849         spin_unlock_irqrestore(&m->lock, flags);
1850 }
1851
1852 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1853 {
1854         int r = -EINVAL;
1855         struct dm_dev *dev;
1856         struct multipath *m = ti->private;
1857         action_fn action;
1858
1859         mutex_lock(&m->work_mutex);
1860
1861         if (dm_suspended(ti)) {
1862                 r = -EBUSY;
1863                 goto out;
1864         }
1865
1866         if (argc == 1) {
1867                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1868                         r = queue_if_no_path(m, true, false);
1869                         goto out;
1870                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1871                         r = queue_if_no_path(m, false, false);
1872                         goto out;
1873                 }
1874         }
1875
1876         if (argc != 2) {
1877                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1878                 goto out;
1879         }
1880
1881         if (!strcasecmp(argv[0], "disable_group")) {
1882                 r = bypass_pg_num(m, argv[1], true);
1883                 goto out;
1884         } else if (!strcasecmp(argv[0], "enable_group")) {
1885                 r = bypass_pg_num(m, argv[1], false);
1886                 goto out;
1887         } else if (!strcasecmp(argv[0], "switch_group")) {
1888                 r = switch_pg_num(m, argv[1]);
1889                 goto out;
1890         } else if (!strcasecmp(argv[0], "reinstate_path"))
1891                 action = reinstate_path;
1892         else if (!strcasecmp(argv[0], "fail_path"))
1893                 action = fail_path;
1894         else {
1895                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1896                 goto out;
1897         }
1898
1899         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1900         if (r) {
1901                 DMWARN("message: error getting device %s",
1902                        argv[1]);
1903                 goto out;
1904         }
1905
1906         r = action_dev(m, dev, action);
1907
1908         dm_put_device(ti, dev);
1909
1910 out:
1911         mutex_unlock(&m->work_mutex);
1912         return r;
1913 }
1914
1915 static int multipath_prepare_ioctl(struct dm_target *ti,
1916                 struct block_device **bdev, fmode_t *mode)
1917 {
1918         struct multipath *m = ti->private;
1919         struct pgpath *current_pgpath;
1920         int r;
1921
1922         current_pgpath = lockless_dereference(m->current_pgpath);
1923         if (!current_pgpath)
1924                 current_pgpath = choose_pgpath(m, 0);
1925
1926         if (current_pgpath) {
1927                 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1928                         *bdev = current_pgpath->path.dev->bdev;
1929                         *mode = current_pgpath->path.dev->mode;
1930                         r = 0;
1931                 } else {
1932                         /* pg_init has not started or completed */
1933                         r = -ENOTCONN;
1934                 }
1935         } else {
1936                 /* No path is available */
1937                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1938                         r = -ENOTCONN;
1939                 else
1940                         r = -EIO;
1941         }
1942
1943         if (r == -ENOTCONN) {
1944                 if (!lockless_dereference(m->current_pg)) {
1945                         /* Path status changed, redo selection */
1946                         (void) choose_pgpath(m, 0);
1947                 }
1948                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1949                         pg_init_all_paths(m);
1950                 dm_table_run_md_queue_async(m->ti->table);
1951                 process_queued_io_list(m);
1952         }
1953
1954         /*
1955          * Only pass ioctls through if the device sizes match exactly.
1956          */
1957         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1958                 return 1;
1959         return r;
1960 }
1961
1962 static int multipath_iterate_devices(struct dm_target *ti,
1963                                      iterate_devices_callout_fn fn, void *data)
1964 {
1965         struct multipath *m = ti->private;
1966         struct priority_group *pg;
1967         struct pgpath *p;
1968         int ret = 0;
1969
1970         list_for_each_entry(pg, &m->priority_groups, list) {
1971                 list_for_each_entry(p, &pg->pgpaths, list) {
1972                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1973                         if (ret)
1974                                 goto out;
1975                 }
1976         }
1977
1978 out:
1979         return ret;
1980 }
1981
1982 static int pgpath_busy(struct pgpath *pgpath)
1983 {
1984         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1985
1986         return blk_lld_busy(q);
1987 }
1988
1989 /*
1990  * We return "busy", only when we can map I/Os but underlying devices
1991  * are busy (so even if we map I/Os now, the I/Os will wait on
1992  * the underlying queue).
1993  * In other words, if we want to kill I/Os or queue them inside us
1994  * due to map unavailability, we don't return "busy".  Otherwise,
1995  * dm core won't give us the I/Os and we can't do what we want.
1996  */
1997 static int multipath_busy(struct dm_target *ti)
1998 {
1999         bool busy = false, has_active = false;
2000         struct multipath *m = ti->private;
2001         struct priority_group *pg, *next_pg;
2002         struct pgpath *pgpath;
2003
2004         /* pg_init in progress */
2005         if (atomic_read(&m->pg_init_in_progress))
2006                 return true;
2007
2008         /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2009         if (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2010                 return (m->queue_mode != DM_TYPE_MQ_REQUEST_BASED);
2011
2012         /* Guess which priority_group will be used at next mapping time */
2013         pg = lockless_dereference(m->current_pg);
2014         next_pg = lockless_dereference(m->next_pg);
2015         if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
2016                 pg = next_pg;
2017
2018         if (!pg) {
2019                 /*
2020                  * We don't know which pg will be used at next mapping time.
2021                  * We don't call choose_pgpath() here to avoid to trigger
2022                  * pg_init just by busy checking.
2023                  * So we don't know whether underlying devices we will be using
2024                  * at next mapping time are busy or not. Just try mapping.
2025                  */
2026                 return busy;
2027         }
2028
2029         /*
2030          * If there is one non-busy active path at least, the path selector
2031          * will be able to select it. So we consider such a pg as not busy.
2032          */
2033         busy = true;
2034         list_for_each_entry(pgpath, &pg->pgpaths, list) {
2035                 if (pgpath->is_active) {
2036                         has_active = true;
2037                         if (!pgpath_busy(pgpath)) {
2038                                 busy = false;
2039                                 break;
2040                         }
2041                 }
2042         }
2043
2044         if (!has_active) {
2045                 /*
2046                  * No active path in this pg, so this pg won't be used and
2047                  * the current_pg will be changed at next mapping time.
2048                  * We need to try mapping to determine it.
2049                  */
2050                 busy = false;
2051         }
2052
2053         return busy;
2054 }
2055
2056 /*-----------------------------------------------------------------
2057  * Module setup
2058  *---------------------------------------------------------------*/
2059 static struct target_type multipath_target = {
2060         .name = "multipath",
2061         .version = {1, 12, 0},
2062         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
2063         .module = THIS_MODULE,
2064         .ctr = multipath_ctr,
2065         .dtr = multipath_dtr,
2066         .map_rq = multipath_map,
2067         .clone_and_map_rq = multipath_clone_and_map,
2068         .release_clone_rq = multipath_release_clone,
2069         .rq_end_io = multipath_end_io,
2070         .map = multipath_map_bio,
2071         .end_io = multipath_end_io_bio,
2072         .presuspend = multipath_presuspend,
2073         .postsuspend = multipath_postsuspend,
2074         .resume = multipath_resume,
2075         .status = multipath_status,
2076         .message = multipath_message,
2077         .prepare_ioctl = multipath_prepare_ioctl,
2078         .iterate_devices = multipath_iterate_devices,
2079         .busy = multipath_busy,
2080 };
2081
2082 static int __init dm_multipath_init(void)
2083 {
2084         int r;
2085
2086         /* allocate a slab for the dm_mpath_ios */
2087         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
2088         if (!_mpio_cache)
2089                 return -ENOMEM;
2090
2091         r = dm_register_target(&multipath_target);
2092         if (r < 0) {
2093                 DMERR("request-based register failed %d", r);
2094                 r = -EINVAL;
2095                 goto bad_register_target;
2096         }
2097
2098         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2099         if (!kmultipathd) {
2100                 DMERR("failed to create workqueue kmpathd");
2101                 r = -ENOMEM;
2102                 goto bad_alloc_kmultipathd;
2103         }
2104
2105         /*
2106          * A separate workqueue is used to handle the device handlers
2107          * to avoid overloading existing workqueue. Overloading the
2108          * old workqueue would also create a bottleneck in the
2109          * path of the storage hardware device activation.
2110          */
2111         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2112                                                   WQ_MEM_RECLAIM);
2113         if (!kmpath_handlerd) {
2114                 DMERR("failed to create workqueue kmpath_handlerd");
2115                 r = -ENOMEM;
2116                 goto bad_alloc_kmpath_handlerd;
2117         }
2118
2119         return 0;
2120
2121 bad_alloc_kmpath_handlerd:
2122         destroy_workqueue(kmultipathd);
2123 bad_alloc_kmultipathd:
2124         dm_unregister_target(&multipath_target);
2125 bad_register_target:
2126         kmem_cache_destroy(_mpio_cache);
2127
2128         return r;
2129 }
2130
2131 static void __exit dm_multipath_exit(void)
2132 {
2133         destroy_workqueue(kmpath_handlerd);
2134         destroy_workqueue(kmultipathd);
2135
2136         dm_unregister_target(&multipath_target);
2137         kmem_cache_destroy(_mpio_cache);
2138 }
2139
2140 module_init(dm_multipath_init);
2141 module_exit(dm_multipath_exit);
2142
2143 MODULE_DESCRIPTION(DM_NAME " multipath target");
2144 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2145 MODULE_LICENSE("GPL");