drivers/md/dm-mpath.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2003 Sistina Software Limited.
   4  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   5  *
   6  * This file is released under the GPL.
   7  */
   8
   9 #include <linux/device-mapper.h>
  10
  11 #include "dm-rq.h"
  12 #include "dm-bio-record.h"
  13 #include "dm-path-selector.h"
  14 #include "dm-uevent.h"
  15
  16 #include <linux/blkdev.h>
  17 #include <linux/ctype.h>
  18 #include <linux/init.h>
  19 #include <linux/mempool.h>
  20 #include <linux/module.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/slab.h>
  23 #include <linux/time.h>
  24 #include <linux/timer.h>
  25 #include <linux/workqueue.h>
  26 #include <linux/delay.h>
  27 #include <scsi/scsi_dh.h>
  28 #include <linux/atomic.h>
  29 #include <linux/blk-mq.h>
  30
  31 static struct workqueue_struct *dm_mpath_wq;
  32
  33 #define DM_MSG_PREFIX "multipath"
  34 #define DM_PG_INIT_DELAY_MSECS 2000
  35 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned int) -1)
  36 #define QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT 0
  37
  38 static unsigned long queue_if_no_path_timeout_secs = QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT;
  39
  40 /* Path properties */
  41 struct pgpath {
  42         struct list_head list;
  43
  44         struct priority_group *pg;      /* Owning PG */
  45         unsigned int fail_count;                /* Cumulative failure count */
  46
  47         struct dm_path path;
  48         struct delayed_work activate_path;
  49
  50         bool is_active:1;               /* Path status */
  51 };
  52
  53 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  54
  55 /*
  56  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  57  * Each has a path selector which controls which path gets used.
  58  */
  59 struct priority_group {
  60         struct list_head list;
  61
  62         struct multipath *m;            /* Owning multipath instance */
  63         struct path_selector ps;
  64
  65         unsigned int pg_num;            /* Reference number */
  66         unsigned int nr_pgpaths;                /* Number of paths in PG */
  67         struct list_head pgpaths;
  68
  69         bool bypassed:1;                /* Temporarily bypass this PG? */
  70 };
  71
  72 /* Multipath context */
  73 struct multipath {
  74         unsigned long flags;            /* Multipath state flags */
  75
  76         spinlock_t lock;
  77         enum dm_queue_mode queue_mode;
  78
  79         struct pgpath *current_pgpath;
  80         struct priority_group *current_pg;
  81         struct priority_group *next_pg; /* Switch to this PG if set */
  82
  83         atomic_t nr_valid_paths;        /* Total number of usable paths */
  84         unsigned int nr_priority_groups;
  85         struct list_head priority_groups;
  86
  87         const char *hw_handler_name;
  88         char *hw_handler_params;
  89         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  90         unsigned int pg_init_retries;   /* Number of times to retry pg_init */
  91         unsigned int pg_init_delay_msecs;       /* Number of msecs before pg_init retry */
  92         atomic_t pg_init_in_progress;   /* Only one pg_init allowed at once */
  93         atomic_t pg_init_count;         /* Number of times pg_init called */
  94
  95         struct mutex work_mutex;
  96         struct work_struct trigger_event;
  97         struct dm_target *ti;
  98
  99         struct work_struct process_queued_bios;
 100         struct bio_list queued_bios;
 101
 102         struct timer_list nopath_timer; /* Timeout for queue_if_no_path */
 103 };
 104
 105 /*
 106  * Context information attached to each io we process.
 107  */
 108 struct dm_mpath_io {
 109         struct pgpath *pgpath;
 110         size_t nr_bytes;
 111         u64 start_time_ns;
 112 };
 113
 114 typedef int (*action_fn) (struct pgpath *pgpath);
 115
 116 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 117 static void trigger_event(struct work_struct *work);
 118 static void activate_or_offline_path(struct pgpath *pgpath);
 119 static void activate_path_work(struct work_struct *work);
 120 static void process_queued_bios(struct work_struct *work);
 121 static void queue_if_no_path_timeout_work(struct timer_list *t);
 122
 123 /*
 124  *-----------------------------------------------
 125  * Multipath state flags.
 126  *-----------------------------------------------
 127  */
 128 #define MPATHF_QUEUE_IO 0                       /* Must we queue all I/O? */
 129 #define MPATHF_QUEUE_IF_NO_PATH 1               /* Queue I/O if last path fails? */
 130 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2         /* Saved state during suspension */
 131 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3     /* If there's already a hw_handler present, don't change it. */
 132 #define MPATHF_PG_INIT_DISABLED 4               /* pg_init is not currently allowed */
 133 #define MPATHF_PG_INIT_REQUIRED 5               /* pg_init needs calling? */
 134 #define MPATHF_PG_INIT_DELAY_RETRY 6            /* Delay pg_init retry? */
 135
 136 static bool mpath_double_check_test_bit(int MPATHF_bit, struct multipath *m)
 137 {
 138         bool r = test_bit(MPATHF_bit, &m->flags);
 139
 140         if (r) {
 141                 unsigned long flags;
 142
 143                 spin_lock_irqsave(&m->lock, flags);
 144                 r = test_bit(MPATHF_bit, &m->flags);
 145                 spin_unlock_irqrestore(&m->lock, flags);
 146         }
 147
 148         return r;
 149 }
 150
 151 /*
 152  *-----------------------------------------------
 153  * Allocation routines
 154  *-----------------------------------------------
 155  */
 156 static struct pgpath *alloc_pgpath(void)
 157 {
 158         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 159
 160         if (!pgpath)
 161                 return NULL;
 162
 163         pgpath->is_active = true;
 164
 165         return pgpath;
 166 }
 167
 168 static void free_pgpath(struct pgpath *pgpath)
 169 {
 170         kfree(pgpath);
 171 }
 172
 173 static struct priority_group *alloc_priority_group(void)
 174 {
 175         struct priority_group *pg;
 176
 177         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 178
 179         if (pg)
 180                 INIT_LIST_HEAD(&pg->pgpaths);
 181
 182         return pg;
 183 }
 184
 185 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 186 {
 187         struct pgpath *pgpath, *tmp;
 188
 189         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 190                 list_del(&pgpath->list);
 191                 dm_put_device(ti, pgpath->path.dev);
 192                 free_pgpath(pgpath);
 193         }
 194 }
 195
 196 static void free_priority_group(struct priority_group *pg,
 197                                 struct dm_target *ti)
 198 {
 199         struct path_selector *ps = &pg->ps;
 200
 201         if (ps->type) {
 202                 ps->type->destroy(ps);
 203                 dm_put_path_selector(ps->type);
 204         }
 205
 206         free_pgpaths(&pg->pgpaths, ti);
 207         kfree(pg);
 208 }
 209
 210 static struct multipath *alloc_multipath(struct dm_target *ti)
 211 {
 212         struct multipath *m;
 213
 214         m = kzalloc(sizeof(*m), GFP_KERNEL);
 215         if (m) {
 216                 INIT_LIST_HEAD(&m->priority_groups);
 217                 spin_lock_init(&m->lock);
 218                 atomic_set(&m->nr_valid_paths, 0);
 219                 INIT_WORK(&m->trigger_event, trigger_event);
 220                 mutex_init(&m->work_mutex);
 221
 222                 m->queue_mode = DM_TYPE_NONE;
 223
 224                 m->ti = ti;
 225                 ti->private = m;
 226
 227                 timer_setup(&m->nopath_timer, queue_if_no_path_timeout_work, 0);
 228         }
 229
 230         return m;
 231 }
 232
 233 static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
 234 {
 235         if (m->queue_mode == DM_TYPE_NONE) {
 236                 m->queue_mode = DM_TYPE_REQUEST_BASED;
 237         } else if (m->queue_mode == DM_TYPE_BIO_BASED) {
 238                 INIT_WORK(&m->process_queued_bios, process_queued_bios);
 239                 /*
 240                  * bio-based doesn't support any direct scsi_dh management;
 241                  * it just discovers if a scsi_dh is attached.
 242                  */
 243                 set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 244         }
 245
 246         dm_table_set_type(ti->table, m->queue_mode);
 247
 248         /*
 249          * Init fields that are only used when a scsi_dh is attached
 250          * - must do this unconditionally (really doesn't hurt non-SCSI uses)
 251          */
 252         set_bit(MPATHF_QUEUE_IO, &m->flags);
 253         atomic_set(&m->pg_init_in_progress, 0);
 254         atomic_set(&m->pg_init_count, 0);
 255         m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 256         init_waitqueue_head(&m->pg_init_wait);
 257
 258         return 0;
 259 }
 260
 261 static void free_multipath(struct multipath *m)
 262 {
 263         struct priority_group *pg, *tmp;
 264
 265         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 266                 list_del(&pg->list);
 267                 free_priority_group(pg, m->ti);
 268         }
 269
 270         kfree(m->hw_handler_name);
 271         kfree(m->hw_handler_params);
 272         mutex_destroy(&m->work_mutex);
 273         kfree(m);
 274 }
 275
 276 static struct dm_mpath_io *get_mpio(union map_info *info)
 277 {
 278         return info->ptr;
 279 }
 280
 281 static size_t multipath_per_bio_data_size(void)
 282 {
 283         return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
 284 }
 285
 286 static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
 287 {
 288         return dm_per_bio_data(bio, multipath_per_bio_data_size());
 289 }
 290
 291 static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
 292 {
 293         /* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
 294         void *bio_details = mpio + 1;
 295         return bio_details;
 296 }
 297
 298 static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
 299 {
 300         struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 301         struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
 302
 303         mpio->nr_bytes = bio->bi_iter.bi_size;
 304         mpio->pgpath = NULL;
 305         mpio->start_time_ns = 0;
 306         *mpio_p = mpio;
 307
 308         dm_bio_record(bio_details, bio);
 309 }
 310
 311 /*
 312  *-----------------------------------------------
 313  * Path selection
 314  *-----------------------------------------------
 315  */
 316 static int __pg_init_all_paths(struct multipath *m)
 317 {
 318         struct pgpath *pgpath;
 319         unsigned long pg_init_delay = 0;
 320
 321         lockdep_assert_held(&m->lock);
 322
 323         if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
 324                 return 0;
 325
 326         atomic_inc(&m->pg_init_count);
 327         clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 328
 329         /* Check here to reset pg_init_required */
 330         if (!m->current_pg)
 331                 return 0;
 332
 333         if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
 334                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 335                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 336         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 337                 /* Skip failed paths */
 338                 if (!pgpath->is_active)
 339                         continue;
 340                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 341                                        pg_init_delay))
 342                         atomic_inc(&m->pg_init_in_progress);
 343         }
 344         return atomic_read(&m->pg_init_in_progress);
 345 }
 346
 347 static int pg_init_all_paths(struct multipath *m)
 348 {
 349         int ret;
 350         unsigned long flags;
 351
 352         spin_lock_irqsave(&m->lock, flags);
 353         ret = __pg_init_all_paths(m);
 354         spin_unlock_irqrestore(&m->lock, flags);
 355
 356         return ret;
 357 }
 358
 359 static void __switch_pg(struct multipath *m, struct priority_group *pg)
 360 {
 361         lockdep_assert_held(&m->lock);
 362
 363         m->current_pg = pg;
 364
 365         /* Must we initialise the PG first, and queue I/O till it's ready? */
 366         if (m->hw_handler_name) {
 367                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 368                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 369         } else {
 370                 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 371                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 372         }
 373
 374         atomic_set(&m->pg_init_count, 0);
 375 }
 376
 377 static struct pgpath *choose_path_in_pg(struct multipath *m,
 378                                         struct priority_group *pg,
 379                                         size_t nr_bytes)
 380 {
 381         unsigned long flags;
 382         struct dm_path *path;
 383         struct pgpath *pgpath;
 384
 385         path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 386         if (!path)
 387                 return ERR_PTR(-ENXIO);
 388
 389         pgpath = path_to_pgpath(path);
 390
 391         if (unlikely(READ_ONCE(m->current_pg) != pg)) {
 392                 /* Only update current_pgpath if pg changed */
 393                 spin_lock_irqsave(&m->lock, flags);
 394                 m->current_pgpath = pgpath;
 395                 __switch_pg(m, pg);
 396                 spin_unlock_irqrestore(&m->lock, flags);
 397         }
 398
 399         return pgpath;
 400 }
 401
 402 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
 403 {
 404         unsigned long flags;
 405         struct priority_group *pg;
 406         struct pgpath *pgpath;
 407         unsigned int bypassed = 1;
 408
 409         if (!atomic_read(&m->nr_valid_paths)) {
 410                 spin_lock_irqsave(&m->lock, flags);
 411                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 412                 spin_unlock_irqrestore(&m->lock, flags);
 413                 goto failed;
 414         }
 415
 416         /* Were we instructed to switch PG? */
 417         if (READ_ONCE(m->next_pg)) {
 418                 spin_lock_irqsave(&m->lock, flags);
 419                 pg = m->next_pg;
 420                 if (!pg) {
 421                         spin_unlock_irqrestore(&m->lock, flags);
 422                         goto check_current_pg;
 423                 }
 424                 m->next_pg = NULL;
 425                 spin_unlock_irqrestore(&m->lock, flags);
 426                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 427                 if (!IS_ERR_OR_NULL(pgpath))
 428                         return pgpath;
 429         }
 430
 431         /* Don't change PG until it has no remaining paths */
 432 check_current_pg:
 433         pg = READ_ONCE(m->current_pg);
 434         if (pg) {
 435                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 436                 if (!IS_ERR_OR_NULL(pgpath))
 437                         return pgpath;
 438         }
 439
 440         /*
 441          * Loop through priority groups until we find a valid path.
 442          * First time we skip PGs marked 'bypassed'.
 443          * Second time we only try the ones we skipped, but set
 444          * pg_init_delay_retry so we do not hammer controllers.
 445          */
 446         do {
 447                 list_for_each_entry(pg, &m->priority_groups, list) {
 448                         if (pg->bypassed == !!bypassed)
 449                                 continue;
 450                         pgpath = choose_path_in_pg(m, pg, nr_bytes);
 451                         if (!IS_ERR_OR_NULL(pgpath)) {
 452                                 if (!bypassed) {
 453                                         spin_lock_irqsave(&m->lock, flags);
 454                                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
 455                                         spin_unlock_irqrestore(&m->lock, flags);
 456                                 }
 457                                 return pgpath;
 458                         }
 459                 }
 460         } while (bypassed--);
 461
 462 failed:
 463         spin_lock_irqsave(&m->lock, flags);
 464         m->current_pgpath = NULL;
 465         m->current_pg = NULL;
 466         spin_unlock_irqrestore(&m->lock, flags);
 467
 468         return NULL;
 469 }
 470
 471 /*
 472  * dm_report_EIO() is a macro instead of a function to make pr_debug_ratelimited()
 473  * report the function name and line number of the function from which
 474  * it has been invoked.
 475  */
 476 #define dm_report_EIO(m)                                                \
 477         DMDEBUG_LIMIT("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d", \
 478                       dm_table_device_name((m)->ti->table),             \
 479                       test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags),   \
 480                       test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
 481                       dm_noflush_suspending((m)->ti))
 482
 483 /*
 484  * Check whether bios must be queued in the device-mapper core rather
 485  * than here in the target.
 486  */
 487 static bool __must_push_back(struct multipath *m)
 488 {
 489         return dm_noflush_suspending(m->ti);
 490 }
 491
 492 static bool must_push_back_rq(struct multipath *m)
 493 {
 494         unsigned long flags;
 495         bool ret;
 496
 497         spin_lock_irqsave(&m->lock, flags);
 498         ret = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) || __must_push_back(m));
 499         spin_unlock_irqrestore(&m->lock, flags);
 500
 501         return ret;
 502 }
 503
 504 /*
 505  * Map cloned requests (request-based multipath)
 506  */
 507 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 508                                    union map_info *map_context,
 509                                    struct request **__clone)
 510 {
 511         struct multipath *m = ti->private;
 512         size_t nr_bytes = blk_rq_bytes(rq);
 513         struct pgpath *pgpath;
 514         struct block_device *bdev;
 515         struct dm_mpath_io *mpio = get_mpio(map_context);
 516         struct request_queue *q;
 517         struct request *clone;
 518
 519         /* Do we need to select a new pgpath? */
 520         pgpath = READ_ONCE(m->current_pgpath);
 521         if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
 522                 pgpath = choose_pgpath(m, nr_bytes);
 523
 524         if (!pgpath) {
 525                 if (must_push_back_rq(m))
 526                         return DM_MAPIO_DELAY_REQUEUE;
 527                 dm_report_EIO(m);       /* Failed */
 528                 return DM_MAPIO_KILL;
 529         } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
 530                    mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
 531                 pg_init_all_paths(m);
 532                 return DM_MAPIO_DELAY_REQUEUE;
 533         }
 534
 535         mpio->pgpath = pgpath;
 536         mpio->nr_bytes = nr_bytes;
 537
 538         bdev = pgpath->path.dev->bdev;
 539         q = bdev_get_queue(bdev);
 540         clone = blk_mq_alloc_request(q, rq->cmd_flags | REQ_NOMERGE,
 541                         BLK_MQ_REQ_NOWAIT);
 542         if (IS_ERR(clone)) {
 543                 /* EBUSY, ENODEV or EWOULDBLOCK: requeue */
 544                 if (blk_queue_dying(q)) {
 545                         atomic_inc(&m->pg_init_in_progress);
 546                         activate_or_offline_path(pgpath);
 547                         return DM_MAPIO_DELAY_REQUEUE;
 548                 }
 549
 550                 /*
 551                  * blk-mq's SCHED_RESTART can cover this requeue, so we
 552                  * needn't deal with it by DELAY_REQUEUE. More importantly,
 553                  * we have to return DM_MAPIO_REQUEUE so that blk-mq can
 554                  * get the queue busy feedback (via BLK_STS_RESOURCE),
 555                  * otherwise I/O merging can suffer.
 556                  */
 557                 return DM_MAPIO_REQUEUE;
 558         }
 559         clone->bio = clone->biotail = NULL;
 560         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 561         *__clone = clone;
 562
 563         if (pgpath->pg->ps.type->start_io)
 564                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 565                                               &pgpath->path,
 566                                               nr_bytes);
 567         return DM_MAPIO_REMAPPED;
 568 }
 569
 570 static void multipath_release_clone(struct request *clone,
 571                                     union map_info *map_context)
 572 {
 573         if (unlikely(map_context)) {
 574                 /*
 575                  * non-NULL map_context means caller is still map
 576                  * method; must undo multipath_clone_and_map()
 577                  */
 578                 struct dm_mpath_io *mpio = get_mpio(map_context);
 579                 struct pgpath *pgpath = mpio->pgpath;
 580
 581                 if (pgpath && pgpath->pg->ps.type->end_io)
 582                         pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
 583                                                     &pgpath->path,
 584                                                     mpio->nr_bytes,
 585                                                     clone->io_start_time_ns);
 586         }
 587
 588         blk_mq_free_request(clone);
 589 }
 590
 591 /*
 592  * Map cloned bios (bio-based multipath)
 593  */
 594
 595 static void __multipath_queue_bio(struct multipath *m, struct bio *bio)
 596 {
 597         /* Queue for the daemon to resubmit */
 598         bio_list_add(&m->queued_bios, bio);
 599         if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
 600                 queue_work(kmultipathd, &m->process_queued_bios);
 601 }
 602
 603 static void multipath_queue_bio(struct multipath *m, struct bio *bio)
 604 {
 605         unsigned long flags;
 606
 607         spin_lock_irqsave(&m->lock, flags);
 608         __multipath_queue_bio(m, bio);
 609         spin_unlock_irqrestore(&m->lock, flags);
 610 }
 611
 612 static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
 613 {
 614         struct pgpath *pgpath;
 615         unsigned long flags;
 616
 617         /* Do we need to select a new pgpath? */
 618         pgpath = READ_ONCE(m->current_pgpath);
 619         if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
 620                 pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
 621
 622         if (!pgpath) {
 623                 spin_lock_irqsave(&m->lock, flags);
 624                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
 625                         __multipath_queue_bio(m, bio);
 626                         pgpath = ERR_PTR(-EAGAIN);
 627                 }
 628                 spin_unlock_irqrestore(&m->lock, flags);
 629
 630         } else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
 631                    mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
 632                 multipath_queue_bio(m, bio);
 633                 pg_init_all_paths(m);
 634                 return ERR_PTR(-EAGAIN);
 635         }
 636
 637         return pgpath;
 638 }
 639
 640 static int __multipath_map_bio(struct multipath *m, struct bio *bio,
 641                                struct dm_mpath_io *mpio)
 642 {
 643         struct pgpath *pgpath = __map_bio(m, bio);
 644
 645         if (IS_ERR(pgpath))
 646                 return DM_MAPIO_SUBMITTED;
 647
 648         if (!pgpath) {
 649                 if (__must_push_back(m))
 650                         return DM_MAPIO_REQUEUE;
 651                 dm_report_EIO(m);
 652                 return DM_MAPIO_KILL;
 653         }
 654
 655         mpio->pgpath = pgpath;
 656
 657         if (dm_ps_use_hr_timer(pgpath->pg->ps.type))
 658                 mpio->start_time_ns = ktime_get_ns();
 659
 660         bio->bi_status = 0;
 661         bio_set_dev(bio, pgpath->path.dev->bdev);
 662         bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
 663
 664         if (pgpath->pg->ps.type->start_io)
 665                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 666                                               &pgpath->path,
 667                                               mpio->nr_bytes);
 668         return DM_MAPIO_REMAPPED;
 669 }
 670
 671 static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
 672 {
 673         struct multipath *m = ti->private;
 674         struct dm_mpath_io *mpio = NULL;
 675
 676         multipath_init_per_bio_data(bio, &mpio);
 677         return __multipath_map_bio(m, bio, mpio);
 678 }
 679
 680 static void process_queued_io_list(struct multipath *m)
 681 {
 682         if (m->queue_mode == DM_TYPE_REQUEST_BASED)
 683                 dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
 684         else if (m->queue_mode == DM_TYPE_BIO_BASED)
 685                 queue_work(kmultipathd, &m->process_queued_bios);
 686 }
 687
 688 static void process_queued_bios(struct work_struct *work)
 689 {
 690         int r;
 691         unsigned long flags;
 692         struct bio *bio;
 693         struct bio_list bios;
 694         struct blk_plug plug;
 695         struct multipath *m =
 696                 container_of(work, struct multipath, process_queued_bios);
 697
 698         bio_list_init(&bios);
 699
 700         spin_lock_irqsave(&m->lock, flags);
 701
 702         if (bio_list_empty(&m->queued_bios)) {
 703                 spin_unlock_irqrestore(&m->lock, flags);
 704                 return;
 705         }
 706
 707         bio_list_merge_init(&bios, &m->queued_bios);
 708
 709         spin_unlock_irqrestore(&m->lock, flags);
 710
 711         blk_start_plug(&plug);
 712         while ((bio = bio_list_pop(&bios))) {
 713                 struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 714
 715                 dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
 716                 r = __multipath_map_bio(m, bio, mpio);
 717                 switch (r) {
 718                 case DM_MAPIO_KILL:
 719                         bio->bi_status = BLK_STS_IOERR;
 720                         bio_endio(bio);
 721                         break;
 722                 case DM_MAPIO_REQUEUE:
 723                         bio->bi_status = BLK_STS_DM_REQUEUE;
 724                         bio_endio(bio);
 725                         break;
 726                 case DM_MAPIO_REMAPPED:
 727                         submit_bio_noacct(bio);
 728                         break;
 729                 case DM_MAPIO_SUBMITTED:
 730                         break;
 731                 default:
 732                         WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
 733                 }
 734         }
 735         blk_finish_plug(&plug);
 736 }
 737
 738 /*
 739  * If we run out of usable paths, should we queue I/O or error it?
 740  */
 741 static int queue_if_no_path(struct multipath *m, bool f_queue_if_no_path,
 742                             bool save_old_value, const char *caller)
 743 {
 744         unsigned long flags;
 745         bool queue_if_no_path_bit, saved_queue_if_no_path_bit;
 746         const char *dm_dev_name = dm_table_device_name(m->ti->table);
 747
 748         DMDEBUG("%s: %s caller=%s f_queue_if_no_path=%d save_old_value=%d",
 749                 dm_dev_name, __func__, caller, f_queue_if_no_path, save_old_value);
 750
 751         spin_lock_irqsave(&m->lock, flags);
 752
 753         queue_if_no_path_bit = test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 754         saved_queue_if_no_path_bit = test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 755
 756         if (save_old_value) {
 757                 if (unlikely(!queue_if_no_path_bit && saved_queue_if_no_path_bit)) {
 758                         DMERR("%s: QIFNP disabled but saved as enabled, saving again loses state, not saving!",
 759                               dm_dev_name);
 760                 } else
 761                         assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path_bit);
 762         } else if (!f_queue_if_no_path && saved_queue_if_no_path_bit) {
 763                 /* due to "fail_if_no_path" message, need to honor it. */
 764                 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 765         }
 766         assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, f_queue_if_no_path);
 767
 768         DMDEBUG("%s: after %s changes; QIFNP = %d; SQIFNP = %d; DNFS = %d",
 769                 dm_dev_name, __func__,
 770                 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
 771                 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
 772                 dm_noflush_suspending(m->ti));
 773
 774         spin_unlock_irqrestore(&m->lock, flags);
 775
 776         if (!f_queue_if_no_path) {
 777                 dm_table_run_md_queue_async(m->ti->table);
 778                 process_queued_io_list(m);
 779         }
 780
 781         return 0;
 782 }
 783
 784 /*
 785  * If the queue_if_no_path timeout fires, turn off queue_if_no_path and
 786  * process any queued I/O.
 787  */
 788 static void queue_if_no_path_timeout_work(struct timer_list *t)
 789 {
 790         struct multipath *m = from_timer(m, t, nopath_timer);
 791
 792         DMWARN("queue_if_no_path timeout on %s, failing queued IO",
 793                dm_table_device_name(m->ti->table));
 794         queue_if_no_path(m, false, false, __func__);
 795 }
 796
 797 /*
 798  * Enable the queue_if_no_path timeout if necessary.
 799  * Called with m->lock held.
 800  */
 801 static void enable_nopath_timeout(struct multipath *m)
 802 {
 803         unsigned long queue_if_no_path_timeout =
 804                 READ_ONCE(queue_if_no_path_timeout_secs) * HZ;
 805
 806         lockdep_assert_held(&m->lock);
 807
 808         if (queue_if_no_path_timeout > 0 &&
 809             atomic_read(&m->nr_valid_paths) == 0 &&
 810             test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
 811                 mod_timer(&m->nopath_timer,
 812                           jiffies + queue_if_no_path_timeout);
 813         }
 814 }
 815
 816 static void disable_nopath_timeout(struct multipath *m)
 817 {
 818         del_timer_sync(&m->nopath_timer);
 819 }
 820
 821 /*
 822  * An event is triggered whenever a path is taken out of use.
 823  * Includes path failure and PG bypass.
 824  */
 825 static void trigger_event(struct work_struct *work)
 826 {
 827         struct multipath *m =
 828                 container_of(work, struct multipath, trigger_event);
 829
 830         dm_table_event(m->ti->table);
 831 }
 832
 833 /*
 834  *---------------------------------------------------------------
 835  * Constructor/argument parsing:
 836  * <#multipath feature args> [<arg>]*
 837  * <#hw_handler args> [hw_handler [<arg>]*]
 838  * <#priority groups>
 839  * <initial priority group>
 840  *     [<selector> <#selector args> [<arg>]*
 841  *      <#paths> <#per-path selector args>
 842  *         [<path> [<arg>]* ]+ ]+
 843  *---------------------------------------------------------------
 844  */
 845 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 846                                struct dm_target *ti)
 847 {
 848         int r;
 849         struct path_selector_type *pst;
 850         unsigned int ps_argc;
 851
 852         static const struct dm_arg _args[] = {
 853                 {0, 1024, "invalid number of path selector args"},
 854         };
 855
 856         pst = dm_get_path_selector(dm_shift_arg(as));
 857         if (!pst) {
 858                 ti->error = "unknown path selector type";
 859                 return -EINVAL;
 860         }
 861
 862         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 863         if (r) {
 864                 dm_put_path_selector(pst);
 865                 return -EINVAL;
 866         }
 867
 868         r = pst->create(&pg->ps, ps_argc, as->argv);
 869         if (r) {
 870                 dm_put_path_selector(pst);
 871                 ti->error = "path selector constructor failed";
 872                 return r;
 873         }
 874
 875         pg->ps.type = pst;
 876         dm_consume_args(as, ps_argc);
 877
 878         return 0;
 879 }
 880
 881 static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
 882                          const char **attached_handler_name, char **error)
 883 {
 884         struct request_queue *q = bdev_get_queue(bdev);
 885         int r;
 886
 887         if (mpath_double_check_test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, m)) {
 888 retain:
 889                 if (*attached_handler_name) {
 890                         /*
 891                          * Clear any hw_handler_params associated with a
 892                          * handler that isn't already attached.
 893                          */
 894                         if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
 895                                 kfree(m->hw_handler_params);
 896                                 m->hw_handler_params = NULL;
 897                         }
 898
 899                         /*
 900                          * Reset hw_handler_name to match the attached handler
 901                          *
 902                          * NB. This modifies the table line to show the actual
 903                          * handler instead of the original table passed in.
 904                          */
 905                         kfree(m->hw_handler_name);
 906                         m->hw_handler_name = *attached_handler_name;
 907                         *attached_handler_name = NULL;
 908                 }
 909         }
 910
 911         if (m->hw_handler_name) {
 912                 r = scsi_dh_attach(q, m->hw_handler_name);
 913                 if (r == -EBUSY) {
 914                         DMINFO("retaining handler on device %pg", bdev);
 915                         goto retain;
 916                 }
 917                 if (r < 0) {
 918                         *error = "error attaching hardware handler";
 919                         return r;
 920                 }
 921
 922                 if (m->hw_handler_params) {
 923                         r = scsi_dh_set_params(q, m->hw_handler_params);
 924                         if (r < 0) {
 925                                 *error = "unable to set hardware handler parameters";
 926                                 return r;
 927                         }
 928                 }
 929         }
 930
 931         return 0;
 932 }
 933
 934 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 935                                  struct dm_target *ti)
 936 {
 937         int r;
 938         struct pgpath *p;
 939         struct multipath *m = ti->private;
 940         struct request_queue *q;
 941         const char *attached_handler_name = NULL;
 942
 943         /* we need at least a path arg */
 944         if (as->argc < 1) {
 945                 ti->error = "no device given";
 946                 return ERR_PTR(-EINVAL);
 947         }
 948
 949         p = alloc_pgpath();
 950         if (!p)
 951                 return ERR_PTR(-ENOMEM);
 952
 953         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 954                           &p->path.dev);
 955         if (r) {
 956                 ti->error = "error getting device";
 957                 goto bad;
 958         }
 959
 960         q = bdev_get_queue(p->path.dev->bdev);
 961         attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 962         if (attached_handler_name || m->hw_handler_name) {
 963                 INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
 964                 r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
 965                 kfree(attached_handler_name);
 966                 if (r) {
 967                         dm_put_device(ti, p->path.dev);
 968                         goto bad;
 969                 }
 970         }
 971
 972         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 973         if (r) {
 974                 dm_put_device(ti, p->path.dev);
 975                 goto bad;
 976         }
 977
 978         return p;
 979  bad:
 980         free_pgpath(p);
 981         return ERR_PTR(r);
 982 }
 983
 984 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 985                                                    struct multipath *m)
 986 {
 987         static const struct dm_arg _args[] = {
 988                 {1, 1024, "invalid number of paths"},
 989                 {0, 1024, "invalid number of selector args"}
 990         };
 991
 992         int r;
 993         unsigned int i, nr_selector_args, nr_args;
 994         struct priority_group *pg;
 995         struct dm_target *ti = m->ti;
 996
 997         if (as->argc < 2) {
 998                 as->argc = 0;
 999                 ti->error = "not enough priority group arguments";
1000                 return ERR_PTR(-EINVAL);
1001         }
1002
1003         pg = alloc_priority_group();
1004         if (!pg) {
1005                 ti->error = "couldn't allocate priority group";
1006                 return ERR_PTR(-ENOMEM);
1007         }
1008         pg->m = m;
1009
1010         r = parse_path_selector(as, pg, ti);
1011         if (r)
1012                 goto bad;
1013
1014         /*
1015          * read the paths
1016          */
1017         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1018         if (r)
1019                 goto bad;
1020
1021         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1022         if (r)
1023                 goto bad;
1024
1025         nr_args = 1 + nr_selector_args;
1026         for (i = 0; i < pg->nr_pgpaths; i++) {
1027                 struct pgpath *pgpath;
1028                 struct dm_arg_set path_args;
1029
1030                 if (as->argc < nr_args) {
1031                         ti->error = "not enough path parameters";
1032                         r = -EINVAL;
1033                         goto bad;
1034                 }
1035
1036                 path_args.argc = nr_args;
1037                 path_args.argv = as->argv;
1038
1039                 pgpath = parse_path(&path_args, &pg->ps, ti);
1040                 if (IS_ERR(pgpath)) {
1041                         r = PTR_ERR(pgpath);
1042                         goto bad;
1043                 }
1044
1045                 pgpath->pg = pg;
1046                 list_add_tail(&pgpath->list, &pg->pgpaths);
1047                 dm_consume_args(as, nr_args);
1048         }
1049
1050         return pg;
1051
1052  bad:
1053         free_priority_group(pg, ti);
1054         return ERR_PTR(r);
1055 }
1056
1057 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1058 {
1059         unsigned int hw_argc;
1060         int ret;
1061         struct dm_target *ti = m->ti;
1062
1063         static const struct dm_arg _args[] = {
1064                 {0, 1024, "invalid number of hardware handler args"},
1065         };
1066
1067         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1068                 return -EINVAL;
1069
1070         if (!hw_argc)
1071                 return 0;
1072
1073         if (m->queue_mode == DM_TYPE_BIO_BASED) {
1074                 dm_consume_args(as, hw_argc);
1075                 DMERR("bio-based multipath doesn't allow hardware handler args");
1076                 return 0;
1077         }
1078
1079         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1080         if (!m->hw_handler_name)
1081                 return -EINVAL;
1082
1083         if (hw_argc > 1) {
1084                 char *p;
1085                 int i, j, len = 4;
1086
1087                 for (i = 0; i <= hw_argc - 2; i++)
1088                         len += strlen(as->argv[i]) + 1;
1089                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1090                 if (!p) {
1091                         ti->error = "memory allocation failed";
1092                         ret = -ENOMEM;
1093                         goto fail;
1094                 }
1095                 j = sprintf(p, "%d", hw_argc - 1);
1096                 for (i = 0, p += j + 1; i <= hw_argc - 2; i++, p += j + 1)
1097                         j = sprintf(p, "%s", as->argv[i]);
1098         }
1099         dm_consume_args(as, hw_argc - 1);
1100
1101         return 0;
1102 fail:
1103         kfree(m->hw_handler_name);
1104         m->hw_handler_name = NULL;
1105         return ret;
1106 }
1107
1108 static int parse_features(struct dm_arg_set *as, struct multipath *m)
1109 {
1110         int r;
1111         unsigned int argc;
1112         struct dm_target *ti = m->ti;
1113         const char *arg_name;
1114
1115         static const struct dm_arg _args[] = {
1116                 {0, 8, "invalid number of feature args"},
1117                 {1, 50, "pg_init_retries must be between 1 and 50"},
1118                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1119         };
1120
1121         r = dm_read_arg_group(_args, as, &argc, &ti->error);
1122         if (r)
1123                 return -EINVAL;
1124
1125         if (!argc)
1126                 return 0;
1127
1128         do {
1129                 arg_name = dm_shift_arg(as);
1130                 argc--;
1131
1132                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
1133                         r = queue_if_no_path(m, true, false, __func__);
1134                         continue;
1135                 }
1136
1137                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1138                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1139                         continue;
1140                 }
1141
1142                 if (!strcasecmp(arg_name, "pg_init_retries") &&
1143                     (argc >= 1)) {
1144                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1145                         argc--;
1146                         continue;
1147                 }
1148
1149                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1150                     (argc >= 1)) {
1151                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1152                         argc--;
1153                         continue;
1154                 }
1155
1156                 if (!strcasecmp(arg_name, "queue_mode") &&
1157                     (argc >= 1)) {
1158                         const char *queue_mode_name = dm_shift_arg(as);
1159
1160                         if (!strcasecmp(queue_mode_name, "bio"))
1161                                 m->queue_mode = DM_TYPE_BIO_BASED;
1162                         else if (!strcasecmp(queue_mode_name, "rq") ||
1163                                  !strcasecmp(queue_mode_name, "mq"))
1164                                 m->queue_mode = DM_TYPE_REQUEST_BASED;
1165                         else {
1166                                 ti->error = "Unknown 'queue_mode' requested";
1167                                 r = -EINVAL;
1168                         }
1169                         argc--;
1170                         continue;
1171                 }
1172
1173                 ti->error = "Unrecognised multipath feature request";
1174                 r = -EINVAL;
1175         } while (argc && !r);
1176
1177         return r;
1178 }
1179
1180 static int multipath_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1181 {
1182         /* target arguments */
1183         static const struct dm_arg _args[] = {
1184                 {0, 1024, "invalid number of priority groups"},
1185                 {0, 1024, "invalid initial priority group number"},
1186         };
1187
1188         int r;
1189         struct multipath *m;
1190         struct dm_arg_set as;
1191         unsigned int pg_count = 0;
1192         unsigned int next_pg_num;
1193         unsigned long flags;
1194
1195         as.argc = argc;
1196         as.argv = argv;
1197
1198         m = alloc_multipath(ti);
1199         if (!m) {
1200                 ti->error = "can't allocate multipath";
1201                 return -EINVAL;
1202         }
1203
1204         r = parse_features(&as, m);
1205         if (r)
1206                 goto bad;
1207
1208         r = alloc_multipath_stage2(ti, m);
1209         if (r)
1210                 goto bad;
1211
1212         r = parse_hw_handler(&as, m);
1213         if (r)
1214                 goto bad;
1215
1216         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1217         if (r)
1218                 goto bad;
1219
1220         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1221         if (r)
1222                 goto bad;
1223
1224         if ((!m->nr_priority_groups && next_pg_num) ||
1225             (m->nr_priority_groups && !next_pg_num)) {
1226                 ti->error = "invalid initial priority group";
1227                 r = -EINVAL;
1228                 goto bad;
1229         }
1230
1231         /* parse the priority groups */
1232         while (as.argc) {
1233                 struct priority_group *pg;
1234                 unsigned int nr_valid_paths = atomic_read(&m->nr_valid_paths);
1235
1236                 pg = parse_priority_group(&as, m);
1237                 if (IS_ERR(pg)) {
1238                         r = PTR_ERR(pg);
1239                         goto bad;
1240                 }
1241
1242                 nr_valid_paths += pg->nr_pgpaths;
1243                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
1244
1245                 list_add_tail(&pg->list, &m->priority_groups);
1246                 pg_count++;
1247                 pg->pg_num = pg_count;
1248                 if (!--next_pg_num)
1249                         m->next_pg = pg;
1250         }
1251
1252         if (pg_count != m->nr_priority_groups) {
1253                 ti->error = "priority group count mismatch";
1254                 r = -EINVAL;
1255                 goto bad;
1256         }
1257
1258         spin_lock_irqsave(&m->lock, flags);
1259         enable_nopath_timeout(m);
1260         spin_unlock_irqrestore(&m->lock, flags);
1261
1262         ti->num_flush_bios = 1;
1263         ti->num_discard_bios = 1;
1264         ti->num_write_zeroes_bios = 1;
1265         if (m->queue_mode == DM_TYPE_BIO_BASED)
1266                 ti->per_io_data_size = multipath_per_bio_data_size();
1267         else
1268                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
1269
1270         return 0;
1271
1272  bad:
1273         free_multipath(m);
1274         return r;
1275 }
1276
1277 static void multipath_wait_for_pg_init_completion(struct multipath *m)
1278 {
1279         DEFINE_WAIT(wait);
1280
1281         while (1) {
1282                 prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1283
1284                 if (!atomic_read(&m->pg_init_in_progress))
1285                         break;
1286
1287                 io_schedule();
1288         }
1289         finish_wait(&m->pg_init_wait, &wait);
1290 }
1291
1292 static void flush_multipath_work(struct multipath *m)
1293 {
1294         if (m->hw_handler_name) {
1295                 unsigned long flags;
1296
1297                 if (!atomic_read(&m->pg_init_in_progress))
1298                         goto skip;
1299
1300                 spin_lock_irqsave(&m->lock, flags);
1301                 if (atomic_read(&m->pg_init_in_progress) &&
1302                     !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1303                         spin_unlock_irqrestore(&m->lock, flags);
1304
1305                         flush_workqueue(kmpath_handlerd);
1306                         multipath_wait_for_pg_init_completion(m);
1307
1308                         spin_lock_irqsave(&m->lock, flags);
1309                         clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1310                 }
1311                 spin_unlock_irqrestore(&m->lock, flags);
1312         }
1313 skip:
1314         if (m->queue_mode == DM_TYPE_BIO_BASED)
1315                 flush_work(&m->process_queued_bios);
1316         flush_work(&m->trigger_event);
1317 }
1318
1319 static void multipath_dtr(struct dm_target *ti)
1320 {
1321         struct multipath *m = ti->private;
1322
1323         disable_nopath_timeout(m);
1324         flush_multipath_work(m);
1325         free_multipath(m);
1326 }
1327
1328 /*
1329  * Take a path out of use.
1330  */
1331 static int fail_path(struct pgpath *pgpath)
1332 {
1333         unsigned long flags;
1334         struct multipath *m = pgpath->pg->m;
1335
1336         spin_lock_irqsave(&m->lock, flags);
1337
1338         if (!pgpath->is_active)
1339                 goto out;
1340
1341         DMWARN("%s: Failing path %s.",
1342                dm_table_device_name(m->ti->table),
1343                pgpath->path.dev->name);
1344
1345         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1346         pgpath->is_active = false;
1347         pgpath->fail_count++;
1348
1349         atomic_dec(&m->nr_valid_paths);
1350
1351         if (pgpath == m->current_pgpath)
1352                 m->current_pgpath = NULL;
1353
1354         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1355                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1356
1357         queue_work(dm_mpath_wq, &m->trigger_event);
1358
1359         enable_nopath_timeout(m);
1360
1361 out:
1362         spin_unlock_irqrestore(&m->lock, flags);
1363
1364         return 0;
1365 }
1366
1367 /*
1368  * Reinstate a previously-failed path
1369  */
1370 static int reinstate_path(struct pgpath *pgpath)
1371 {
1372         int r = 0, run_queue = 0;
1373         unsigned long flags;
1374         struct multipath *m = pgpath->pg->m;
1375         unsigned int nr_valid_paths;
1376
1377         spin_lock_irqsave(&m->lock, flags);
1378
1379         if (pgpath->is_active)
1380                 goto out;
1381
1382         DMWARN("%s: Reinstating path %s.",
1383                dm_table_device_name(m->ti->table),
1384                pgpath->path.dev->name);
1385
1386         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1387         if (r)
1388                 goto out;
1389
1390         pgpath->is_active = true;
1391
1392         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1393         if (nr_valid_paths == 1) {
1394                 m->current_pgpath = NULL;
1395                 run_queue = 1;
1396         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1397                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1398                         atomic_inc(&m->pg_init_in_progress);
1399         }
1400
1401         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1402                        pgpath->path.dev->name, nr_valid_paths);
1403
1404         schedule_work(&m->trigger_event);
1405
1406 out:
1407         spin_unlock_irqrestore(&m->lock, flags);
1408         if (run_queue) {
1409                 dm_table_run_md_queue_async(m->ti->table);
1410                 process_queued_io_list(m);
1411         }
1412
1413         if (pgpath->is_active)
1414                 disable_nopath_timeout(m);
1415
1416         return r;
1417 }
1418
1419 /*
1420  * Fail or reinstate all paths that match the provided struct dm_dev.
1421  */
1422 static int action_dev(struct multipath *m, dev_t dev, action_fn action)
1423 {
1424         int r = -EINVAL;
1425         struct pgpath *pgpath;
1426         struct priority_group *pg;
1427
1428         list_for_each_entry(pg, &m->priority_groups, list) {
1429                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1430                         if (pgpath->path.dev->bdev->bd_dev == dev)
1431                                 r = action(pgpath);
1432                 }
1433         }
1434
1435         return r;
1436 }
1437
1438 /*
1439  * Temporarily try to avoid having to use the specified PG
1440  */
1441 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1442                       bool bypassed)
1443 {
1444         unsigned long flags;
1445
1446         spin_lock_irqsave(&m->lock, flags);
1447
1448         pg->bypassed = bypassed;
1449         m->current_pgpath = NULL;
1450         m->current_pg = NULL;
1451
1452         spin_unlock_irqrestore(&m->lock, flags);
1453
1454         schedule_work(&m->trigger_event);
1455 }
1456
1457 /*
1458  * Switch to using the specified PG from the next I/O that gets mapped
1459  */
1460 static int switch_pg_num(struct multipath *m, const char *pgstr)
1461 {
1462         struct priority_group *pg;
1463         unsigned int pgnum;
1464         unsigned long flags;
1465         char dummy;
1466
1467         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1468             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1469                 DMWARN("invalid PG number supplied to %s", __func__);
1470                 return -EINVAL;
1471         }
1472
1473         spin_lock_irqsave(&m->lock, flags);
1474         list_for_each_entry(pg, &m->priority_groups, list) {
1475                 pg->bypassed = false;
1476                 if (--pgnum)
1477                         continue;
1478
1479                 m->current_pgpath = NULL;
1480                 m->current_pg = NULL;
1481                 m->next_pg = pg;
1482         }
1483         spin_unlock_irqrestore(&m->lock, flags);
1484
1485         schedule_work(&m->trigger_event);
1486         return 0;
1487 }
1488
1489 /*
1490  * Set/clear bypassed status of a PG.
1491  * PGs are numbered upwards from 1 in the order they were declared.
1492  */
1493 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1494 {
1495         struct priority_group *pg;
1496         unsigned int pgnum;
1497         char dummy;
1498
1499         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1500             !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1501                 DMWARN("invalid PG number supplied to bypass_pg");
1502                 return -EINVAL;
1503         }
1504
1505         list_for_each_entry(pg, &m->priority_groups, list) {
1506                 if (!--pgnum)
1507                         break;
1508         }
1509
1510         bypass_pg(m, pg, bypassed);
1511         return 0;
1512 }
1513
1514 /*
1515  * Should we retry pg_init immediately?
1516  */
1517 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1518 {
1519         unsigned long flags;
1520         bool limit_reached = false;
1521
1522         spin_lock_irqsave(&m->lock, flags);
1523
1524         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1525             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1526                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1527         else
1528                 limit_reached = true;
1529
1530         spin_unlock_irqrestore(&m->lock, flags);
1531
1532         return limit_reached;
1533 }
1534
1535 static void pg_init_done(void *data, int errors)
1536 {
1537         struct pgpath *pgpath = data;
1538         struct priority_group *pg = pgpath->pg;
1539         struct multipath *m = pg->m;
1540         unsigned long flags;
1541         bool delay_retry = false;
1542
1543         /* device or driver problems */
1544         switch (errors) {
1545         case SCSI_DH_OK:
1546                 break;
1547         case SCSI_DH_NOSYS:
1548                 if (!m->hw_handler_name) {
1549                         errors = 0;
1550                         break;
1551                 }
1552                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1553                       "Error %d.", m->hw_handler_name, errors);
1554                 /*
1555                  * Fail path for now, so we do not ping pong
1556                  */
1557                 fail_path(pgpath);
1558                 break;
1559         case SCSI_DH_DEV_TEMP_BUSY:
1560                 /*
1561                  * Probably doing something like FW upgrade on the
1562                  * controller so try the other pg.
1563                  */
1564                 bypass_pg(m, pg, true);
1565                 break;
1566         case SCSI_DH_RETRY:
1567                 /* Wait before retrying. */
1568                 delay_retry = true;
1569                 fallthrough;
1570         case SCSI_DH_IMM_RETRY:
1571         case SCSI_DH_RES_TEMP_UNAVAIL:
1572                 if (pg_init_limit_reached(m, pgpath))
1573                         fail_path(pgpath);
1574                 errors = 0;
1575                 break;
1576         case SCSI_DH_DEV_OFFLINED:
1577         default:
1578                 /*
1579                  * We probably do not want to fail the path for a device
1580                  * error, but this is what the old dm did. In future
1581                  * patches we can do more advanced handling.
1582                  */
1583                 fail_path(pgpath);
1584         }
1585
1586         spin_lock_irqsave(&m->lock, flags);
1587         if (errors) {
1588                 if (pgpath == m->current_pgpath) {
1589                         DMERR("Could not failover device. Error %d.", errors);
1590                         m->current_pgpath = NULL;
1591                         m->current_pg = NULL;
1592                 }
1593         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1594                 pg->bypassed = false;
1595
1596         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1597                 /* Activations of other paths are still on going */
1598                 goto out;
1599
1600         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1601                 if (delay_retry)
1602                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1603                 else
1604                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1605
1606                 if (__pg_init_all_paths(m))
1607                         goto out;
1608         }
1609         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1610
1611         process_queued_io_list(m);
1612
1613         /*
1614          * Wake up any thread waiting to suspend.
1615          */
1616         wake_up(&m->pg_init_wait);
1617
1618 out:
1619         spin_unlock_irqrestore(&m->lock, flags);
1620 }
1621
1622 static void activate_or_offline_path(struct pgpath *pgpath)
1623 {
1624         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1625
1626         if (pgpath->is_active && !blk_queue_dying(q))
1627                 scsi_dh_activate(q, pg_init_done, pgpath);
1628         else
1629                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1630 }
1631
1632 static void activate_path_work(struct work_struct *work)
1633 {
1634         struct pgpath *pgpath =
1635                 container_of(work, struct pgpath, activate_path.work);
1636
1637         activate_or_offline_path(pgpath);
1638 }
1639
1640 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1641                             blk_status_t error, union map_info *map_context)
1642 {
1643         struct dm_mpath_io *mpio = get_mpio(map_context);
1644         struct pgpath *pgpath = mpio->pgpath;
1645         int r = DM_ENDIO_DONE;
1646
1647         /*
1648          * We don't queue any clone request inside the multipath target
1649          * during end I/O handling, since those clone requests don't have
1650          * bio clones.  If we queue them inside the multipath target,
1651          * we need to make bio clones, that requires memory allocation.
1652          * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1653          *  don't have bio clones.)
1654          * Instead of queueing the clone request here, we queue the original
1655          * request into dm core, which will remake a clone request and
1656          * clone bios for it and resubmit it later.
1657          */
1658         if (error && blk_path_error(error)) {
1659                 struct multipath *m = ti->private;
1660
1661                 if (error == BLK_STS_RESOURCE)
1662                         r = DM_ENDIO_DELAY_REQUEUE;
1663                 else
1664                         r = DM_ENDIO_REQUEUE;
1665
1666                 if (pgpath)
1667                         fail_path(pgpath);
1668
1669                 if (!atomic_read(&m->nr_valid_paths) &&
1670                     !must_push_back_rq(m)) {
1671                         if (error == BLK_STS_IOERR)
1672                                 dm_report_EIO(m);
1673                         /* complete with the original error */
1674                         r = DM_ENDIO_DONE;
1675                 }
1676         }
1677
1678         if (pgpath) {
1679                 struct path_selector *ps = &pgpath->pg->ps;
1680
1681                 if (ps->type->end_io)
1682                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1683                                          clone->io_start_time_ns);
1684         }
1685
1686         return r;
1687 }
1688
1689 static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1690                                 blk_status_t *error)
1691 {
1692         struct multipath *m = ti->private;
1693         struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1694         struct pgpath *pgpath = mpio->pgpath;
1695         unsigned long flags;
1696         int r = DM_ENDIO_DONE;
1697
1698         if (!*error || !blk_path_error(*error))
1699                 goto done;
1700
1701         if (pgpath)
1702                 fail_path(pgpath);
1703
1704         if (!atomic_read(&m->nr_valid_paths)) {
1705                 spin_lock_irqsave(&m->lock, flags);
1706                 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1707                         if (__must_push_back(m)) {
1708                                 r = DM_ENDIO_REQUEUE;
1709                         } else {
1710                                 dm_report_EIO(m);
1711                                 *error = BLK_STS_IOERR;
1712                         }
1713                         spin_unlock_irqrestore(&m->lock, flags);
1714                         goto done;
1715                 }
1716                 spin_unlock_irqrestore(&m->lock, flags);
1717         }
1718
1719         multipath_queue_bio(m, clone);
1720         r = DM_ENDIO_INCOMPLETE;
1721 done:
1722         if (pgpath) {
1723                 struct path_selector *ps = &pgpath->pg->ps;
1724
1725                 if (ps->type->end_io)
1726                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1727                                          (mpio->start_time_ns ?:
1728                                           dm_start_time_ns_from_clone(clone)));
1729         }
1730
1731         return r;
1732 }
1733
1734 /*
1735  * Suspend with flush can't complete until all the I/O is processed
1736  * so if the last path fails we must error any remaining I/O.
1737  * - Note that if the freeze_bdev fails while suspending, the
1738  *   queue_if_no_path state is lost - userspace should reset it.
1739  * Otherwise, during noflush suspend, queue_if_no_path will not change.
1740  */
1741 static void multipath_presuspend(struct dm_target *ti)
1742 {
1743         struct multipath *m = ti->private;
1744
1745         /* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1746         if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1747                 queue_if_no_path(m, false, true, __func__);
1748 }
1749
1750 static void multipath_postsuspend(struct dm_target *ti)
1751 {
1752         struct multipath *m = ti->private;
1753
1754         mutex_lock(&m->work_mutex);
1755         flush_multipath_work(m);
1756         mutex_unlock(&m->work_mutex);
1757 }
1758
1759 /*
1760  * Restore the queue_if_no_path setting.
1761  */
1762 static void multipath_resume(struct dm_target *ti)
1763 {
1764         struct multipath *m = ti->private;
1765         unsigned long flags;
1766
1767         spin_lock_irqsave(&m->lock, flags);
1768         if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1769                 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1770                 clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1771         }
1772
1773         DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1774                 dm_table_device_name(m->ti->table), __func__,
1775                 test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1776                 test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1777
1778         spin_unlock_irqrestore(&m->lock, flags);
1779 }
1780
1781 /*
1782  * Info output has the following format:
1783  * num_multipath_feature_args [multipath_feature_args]*
1784  * num_handler_status_args [handler_status_args]*
1785  * num_groups init_group_number
1786  *            [A|D|E num_ps_status_args [ps_status_args]*
1787  *             num_paths num_selector_args
1788  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1789  *
1790  * Table output has the following format (identical to the constructor string):
1791  * num_feature_args [features_args]*
1792  * num_handler_args hw_handler [hw_handler_args]*
1793  * num_groups init_group_number
1794  *     [priority selector-name num_ps_args [ps_args]*
1795  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1796  */
1797 static void multipath_status(struct dm_target *ti, status_type_t type,
1798                              unsigned int status_flags, char *result, unsigned int maxlen)
1799 {
1800         int sz = 0, pg_counter, pgpath_counter;
1801         unsigned long flags;
1802         struct multipath *m = ti->private;
1803         struct priority_group *pg;
1804         struct pgpath *p;
1805         unsigned int pg_num;
1806         char state;
1807
1808         spin_lock_irqsave(&m->lock, flags);
1809
1810         /* Features */
1811         if (type == STATUSTYPE_INFO)
1812                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1813                        atomic_read(&m->pg_init_count));
1814         else {
1815                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1816                               (m->pg_init_retries > 0) * 2 +
1817                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1818                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1819                               (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1820
1821                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1822                         DMEMIT("queue_if_no_path ");
1823                 if (m->pg_init_retries)
1824                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1825                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1826                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1827                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1828                         DMEMIT("retain_attached_hw_handler ");
1829                 if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1830                         switch (m->queue_mode) {
1831                         case DM_TYPE_BIO_BASED:
1832                                 DMEMIT("queue_mode bio ");
1833                                 break;
1834                         default:
1835                                 WARN_ON_ONCE(true);
1836                                 break;
1837                         }
1838                 }
1839         }
1840
1841         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1842                 DMEMIT("0 ");
1843         else
1844                 DMEMIT("1 %s ", m->hw_handler_name);
1845
1846         DMEMIT("%u ", m->nr_priority_groups);
1847
1848         if (m->next_pg)
1849                 pg_num = m->next_pg->pg_num;
1850         else if (m->current_pg)
1851                 pg_num = m->current_pg->pg_num;
1852         else
1853                 pg_num = (m->nr_priority_groups ? 1 : 0);
1854
1855         DMEMIT("%u ", pg_num);
1856
1857         switch (type) {
1858         case STATUSTYPE_INFO:
1859                 list_for_each_entry(pg, &m->priority_groups, list) {
1860                         if (pg->bypassed)
1861                                 state = 'D';    /* Disabled */
1862                         else if (pg == m->current_pg)
1863                                 state = 'A';    /* Currently Active */
1864                         else
1865                                 state = 'E';    /* Enabled */
1866
1867                         DMEMIT("%c ", state);
1868
1869                         if (pg->ps.type->status)
1870                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1871                                                           result + sz,
1872                                                           maxlen - sz);
1873                         else
1874                                 DMEMIT("0 ");
1875
1876                         DMEMIT("%u %u ", pg->nr_pgpaths,
1877                                pg->ps.type->info_args);
1878
1879                         list_for_each_entry(p, &pg->pgpaths, list) {
1880                                 DMEMIT("%s %s %u ", p->path.dev->name,
1881                                        p->is_active ? "A" : "F",
1882                                        p->fail_count);
1883                                 if (pg->ps.type->status)
1884                                         sz += pg->ps.type->status(&pg->ps,
1885                                               &p->path, type, result + sz,
1886                                               maxlen - sz);
1887                         }
1888                 }
1889                 break;
1890
1891         case STATUSTYPE_TABLE:
1892                 list_for_each_entry(pg, &m->priority_groups, list) {
1893                         DMEMIT("%s ", pg->ps.type->name);
1894
1895                         if (pg->ps.type->status)
1896                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1897                                                           result + sz,
1898                                                           maxlen - sz);
1899                         else
1900                                 DMEMIT("0 ");
1901
1902                         DMEMIT("%u %u ", pg->nr_pgpaths,
1903                                pg->ps.type->table_args);
1904
1905                         list_for_each_entry(p, &pg->pgpaths, list) {
1906                                 DMEMIT("%s ", p->path.dev->name);
1907                                 if (pg->ps.type->status)
1908                                         sz += pg->ps.type->status(&pg->ps,
1909                                               &p->path, type, result + sz,
1910                                               maxlen - sz);
1911                         }
1912                 }
1913                 break;
1914
1915         case STATUSTYPE_IMA:
1916                 sz = 0; /*reset the result pointer*/
1917
1918                 DMEMIT_TARGET_NAME_VERSION(ti->type);
1919                 DMEMIT(",nr_priority_groups=%u", m->nr_priority_groups);
1920
1921                 pg_counter = 0;
1922                 list_for_each_entry(pg, &m->priority_groups, list) {
1923                         if (pg->bypassed)
1924                                 state = 'D';    /* Disabled */
1925                         else if (pg == m->current_pg)
1926                                 state = 'A';    /* Currently Active */
1927                         else
1928                                 state = 'E';    /* Enabled */
1929                         DMEMIT(",pg_state_%d=%c", pg_counter, state);
1930                         DMEMIT(",nr_pgpaths_%d=%u", pg_counter, pg->nr_pgpaths);
1931                         DMEMIT(",path_selector_name_%d=%s", pg_counter, pg->ps.type->name);
1932
1933                         pgpath_counter = 0;
1934                         list_for_each_entry(p, &pg->pgpaths, list) {
1935                                 DMEMIT(",path_name_%d_%d=%s,is_active_%d_%d=%c,fail_count_%d_%d=%u",
1936                                        pg_counter, pgpath_counter, p->path.dev->name,
1937                                        pg_counter, pgpath_counter, p->is_active ? 'A' : 'F',
1938                                        pg_counter, pgpath_counter, p->fail_count);
1939                                 if (pg->ps.type->status) {
1940                                         DMEMIT(",path_selector_status_%d_%d=",
1941                                                pg_counter, pgpath_counter);
1942                                         sz += pg->ps.type->status(&pg->ps, &p->path,
1943                                                                   type, result + sz,
1944                                                                   maxlen - sz);
1945                                 }
1946                                 pgpath_counter++;
1947                         }
1948                         pg_counter++;
1949                 }
1950                 DMEMIT(";");
1951                 break;
1952         }
1953
1954         spin_unlock_irqrestore(&m->lock, flags);
1955 }
1956
1957 static int multipath_message(struct dm_target *ti, unsigned int argc, char **argv,
1958                              char *result, unsigned int maxlen)
1959 {
1960         int r = -EINVAL;
1961         dev_t dev;
1962         struct multipath *m = ti->private;
1963         action_fn action;
1964         unsigned long flags;
1965
1966         mutex_lock(&m->work_mutex);
1967
1968         if (dm_suspended(ti)) {
1969                 r = -EBUSY;
1970                 goto out;
1971         }
1972
1973         if (argc == 1) {
1974                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1975                         r = queue_if_no_path(m, true, false, __func__);
1976                         spin_lock_irqsave(&m->lock, flags);
1977                         enable_nopath_timeout(m);
1978                         spin_unlock_irqrestore(&m->lock, flags);
1979                         goto out;
1980                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1981                         r = queue_if_no_path(m, false, false, __func__);
1982                         disable_nopath_timeout(m);
1983                         goto out;
1984                 }
1985         }
1986
1987         if (argc != 2) {
1988                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1989                 goto out;
1990         }
1991
1992         if (!strcasecmp(argv[0], "disable_group")) {
1993                 r = bypass_pg_num(m, argv[1], true);
1994                 goto out;
1995         } else if (!strcasecmp(argv[0], "enable_group")) {
1996                 r = bypass_pg_num(m, argv[1], false);
1997                 goto out;
1998         } else if (!strcasecmp(argv[0], "switch_group")) {
1999                 r = switch_pg_num(m, argv[1]);
2000                 goto out;
2001         } else if (!strcasecmp(argv[0], "reinstate_path"))
2002                 action = reinstate_path;
2003         else if (!strcasecmp(argv[0], "fail_path"))
2004                 action = fail_path;
2005         else {
2006                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
2007                 goto out;
2008         }
2009
2010         r = dm_devt_from_path(argv[1], &dev);
2011         if (r) {
2012                 DMWARN("message: error getting device %s",
2013                        argv[1]);
2014                 goto out;
2015         }
2016
2017         r = action_dev(m, dev, action);
2018
2019 out:
2020         mutex_unlock(&m->work_mutex);
2021         return r;
2022 }
2023
2024 static int multipath_prepare_ioctl(struct dm_target *ti,
2025                                    struct block_device **bdev)
2026 {
2027         struct multipath *m = ti->private;
2028         struct pgpath *pgpath;
2029         unsigned long flags;
2030         int r;
2031
2032         pgpath = READ_ONCE(m->current_pgpath);
2033         if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
2034                 pgpath = choose_pgpath(m, 0);
2035
2036         if (pgpath) {
2037                 if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
2038                         *bdev = pgpath->path.dev->bdev;
2039                         r = 0;
2040                 } else {
2041                         /* pg_init has not started or completed */
2042                         r = -ENOTCONN;
2043                 }
2044         } else {
2045                 /* No path is available */
2046                 r = -EIO;
2047                 spin_lock_irqsave(&m->lock, flags);
2048                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2049                         r = -ENOTCONN;
2050                 spin_unlock_irqrestore(&m->lock, flags);
2051         }
2052
2053         if (r == -ENOTCONN) {
2054                 if (!READ_ONCE(m->current_pg)) {
2055                         /* Path status changed, redo selection */
2056                         (void) choose_pgpath(m, 0);
2057                 }
2058                 spin_lock_irqsave(&m->lock, flags);
2059                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2060                         (void) __pg_init_all_paths(m);
2061                 spin_unlock_irqrestore(&m->lock, flags);
2062                 dm_table_run_md_queue_async(m->ti->table);
2063                 process_queued_io_list(m);
2064         }
2065
2066         /*
2067          * Only pass ioctls through if the device sizes match exactly.
2068          */
2069         if (!r && ti->len != bdev_nr_sectors((*bdev)))
2070                 return 1;
2071         return r;
2072 }
2073
2074 static int multipath_iterate_devices(struct dm_target *ti,
2075                                      iterate_devices_callout_fn fn, void *data)
2076 {
2077         struct multipath *m = ti->private;
2078         struct priority_group *pg;
2079         struct pgpath *p;
2080         int ret = 0;
2081
2082         list_for_each_entry(pg, &m->priority_groups, list) {
2083                 list_for_each_entry(p, &pg->pgpaths, list) {
2084                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2085                         if (ret)
2086                                 goto out;
2087                 }
2088         }
2089
2090 out:
2091         return ret;
2092 }
2093
2094 static int pgpath_busy(struct pgpath *pgpath)
2095 {
2096         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2097
2098         return blk_lld_busy(q);
2099 }
2100
2101 /*
2102  * We return "busy", only when we can map I/Os but underlying devices
2103  * are busy (so even if we map I/Os now, the I/Os will wait on
2104  * the underlying queue).
2105  * In other words, if we want to kill I/Os or queue them inside us
2106  * due to map unavailability, we don't return "busy".  Otherwise,
2107  * dm core won't give us the I/Os and we can't do what we want.
2108  */
2109 static int multipath_busy(struct dm_target *ti)
2110 {
2111         bool busy = false, has_active = false;
2112         struct multipath *m = ti->private;
2113         struct priority_group *pg, *next_pg;
2114         struct pgpath *pgpath;
2115
2116         /* pg_init in progress */
2117         if (atomic_read(&m->pg_init_in_progress))
2118                 return true;
2119
2120         /* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2121         if (!atomic_read(&m->nr_valid_paths)) {
2122                 unsigned long flags;
2123
2124                 spin_lock_irqsave(&m->lock, flags);
2125                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2126                         spin_unlock_irqrestore(&m->lock, flags);
2127                         return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2128                 }
2129                 spin_unlock_irqrestore(&m->lock, flags);
2130         }
2131
2132         /* Guess which priority_group will be used at next mapping time */
2133         pg = READ_ONCE(m->current_pg);
2134         next_pg = READ_ONCE(m->next_pg);
2135         if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2136                 pg = next_pg;
2137
2138         if (!pg) {
2139                 /*
2140                  * We don't know which pg will be used at next mapping time.
2141                  * We don't call choose_pgpath() here to avoid to trigger
2142                  * pg_init just by busy checking.
2143                  * So we don't know whether underlying devices we will be using
2144                  * at next mapping time are busy or not. Just try mapping.
2145                  */
2146                 return busy;
2147         }
2148
2149         /*
2150          * If there is one non-busy active path at least, the path selector
2151          * will be able to select it. So we consider such a pg as not busy.
2152          */
2153         busy = true;
2154         list_for_each_entry(pgpath, &pg->pgpaths, list) {
2155                 if (pgpath->is_active) {
2156                         has_active = true;
2157                         if (!pgpath_busy(pgpath)) {
2158                                 busy = false;
2159                                 break;
2160                         }
2161                 }
2162         }
2163
2164         if (!has_active) {
2165                 /*
2166                  * No active path in this pg, so this pg won't be used and
2167                  * the current_pg will be changed at next mapping time.
2168                  * We need to try mapping to determine it.
2169                  */
2170                 busy = false;
2171         }
2172
2173         return busy;
2174 }
2175
2176 /*
2177  *---------------------------------------------------------------
2178  * Module setup
2179  *---------------------------------------------------------------
2180  */
2181 static struct target_type multipath_target = {
2182         .name = "multipath",
2183         .version = {1, 14, 0},
2184         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2185                     DM_TARGET_PASSES_INTEGRITY,
2186         .module = THIS_MODULE,
2187         .ctr = multipath_ctr,
2188         .dtr = multipath_dtr,
2189         .clone_and_map_rq = multipath_clone_and_map,
2190         .release_clone_rq = multipath_release_clone,
2191         .rq_end_io = multipath_end_io,
2192         .map = multipath_map_bio,
2193         .end_io = multipath_end_io_bio,
2194         .presuspend = multipath_presuspend,
2195         .postsuspend = multipath_postsuspend,
2196         .resume = multipath_resume,
2197         .status = multipath_status,
2198         .message = multipath_message,
2199         .prepare_ioctl = multipath_prepare_ioctl,
2200         .iterate_devices = multipath_iterate_devices,
2201         .busy = multipath_busy,
2202 };
2203
2204 static int __init dm_multipath_init(void)
2205 {
2206         int r = -ENOMEM;
2207
2208         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2209         if (!kmultipathd) {
2210                 DMERR("failed to create workqueue kmpathd");
2211                 goto bad_alloc_kmultipathd;
2212         }
2213
2214         /*
2215          * A separate workqueue is used to handle the device handlers
2216          * to avoid overloading existing workqueue. Overloading the
2217          * old workqueue would also create a bottleneck in the
2218          * path of the storage hardware device activation.
2219          */
2220         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2221                                                   WQ_MEM_RECLAIM);
2222         if (!kmpath_handlerd) {
2223                 DMERR("failed to create workqueue kmpath_handlerd");
2224                 goto bad_alloc_kmpath_handlerd;
2225         }
2226
2227         dm_mpath_wq = alloc_workqueue("dm_mpath_wq", 0, 0);
2228         if (!dm_mpath_wq) {
2229                 DMERR("failed to create workqueue dm_mpath_wq");
2230                 goto bad_alloc_dm_mpath_wq;
2231         }
2232
2233         r = dm_register_target(&multipath_target);
2234         if (r < 0)
2235                 goto bad_register_target;
2236
2237         return 0;
2238
2239 bad_register_target:
2240         destroy_workqueue(dm_mpath_wq);
2241 bad_alloc_dm_mpath_wq:
2242         destroy_workqueue(kmpath_handlerd);
2243 bad_alloc_kmpath_handlerd:
2244         destroy_workqueue(kmultipathd);
2245 bad_alloc_kmultipathd:
2246         return r;
2247 }
2248
2249 static void __exit dm_multipath_exit(void)
2250 {
2251         destroy_workqueue(dm_mpath_wq);
2252         destroy_workqueue(kmpath_handlerd);
2253         destroy_workqueue(kmultipathd);
2254
2255         dm_unregister_target(&multipath_target);
2256 }
2257
2258 module_init(dm_multipath_init);
2259 module_exit(dm_multipath_exit);
2260
2261 module_param_named(queue_if_no_path_timeout_secs, queue_if_no_path_timeout_secs, ulong, 0644);
2262 MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2263
2264 MODULE_DESCRIPTION(DM_NAME " multipath target");
2265 MODULE_AUTHOR("Sistina Software <dm-devel@lists.linux.dev>");
2266 MODULE_LICENSE("GPL");