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