drivers/md/dm-mpath.c

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