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-path-selector.h"
  11 #include "dm-bio-record.h"
  12 #include "dm-uevent.h"
  13
  14 #include <linux/ctype.h>
  15 #include <linux/init.h>
  16 #include <linux/mempool.h>
  17 #include <linux/module.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/slab.h>
  20 #include <linux/time.h>
  21 #include <linux/workqueue.h>
  22 #include <scsi/scsi_dh.h>
  23 #include <asm/atomic.h>
  24
  25 #define DM_MSG_PREFIX "multipath"
  26 #define MESG_STR(x) x, sizeof(x)
  27
  28 /* Path properties */
  29 struct pgpath {
  30         struct list_head list;
  31
  32         struct priority_group *pg;      /* Owning PG */
  33         unsigned is_active;             /* Path status */
  34         unsigned fail_count;            /* Cumulative failure count */
  35
  36         struct dm_path path;
  37         struct work_struct deactivate_path;
  38 };
  39
  40 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  41
  42 /*
  43  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  44  * Each has a path selector which controls which path gets used.
  45  */
  46 struct priority_group {
  47         struct list_head list;
  48
  49         struct multipath *m;            /* Owning multipath instance */
  50         struct path_selector ps;
  51
  52         unsigned pg_num;                /* Reference number */
  53         unsigned bypassed;              /* Temporarily bypass this PG? */
  54
  55         unsigned nr_pgpaths;            /* Number of paths in PG */
  56         struct list_head pgpaths;
  57 };
  58
  59 /* Multipath context */
  60 struct multipath {
  61         struct list_head list;
  62         struct dm_target *ti;
  63
  64         spinlock_t lock;
  65
  66         const char *hw_handler_name;
  67         struct work_struct activate_path;
  68         struct pgpath *pgpath_to_activate;
  69         unsigned nr_priority_groups;
  70         struct list_head priority_groups;
  71         unsigned pg_init_required;      /* pg_init needs calling? */
  72         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  73
  74         unsigned nr_valid_paths;        /* Total number of usable paths */
  75         struct pgpath *current_pgpath;
  76         struct priority_group *current_pg;
  77         struct priority_group *next_pg; /* Switch to this PG if set */
  78         unsigned repeat_count;          /* I/Os left before calling PS again */
  79
  80         unsigned queue_io;              /* Must we queue all I/O? */
  81         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  82         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  83         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  84         unsigned pg_init_count;         /* Number of times pg_init called */
  85
  86         struct work_struct process_queued_ios;
  87         struct bio_list queued_ios;
  88         unsigned queue_size;
  89
  90         struct work_struct trigger_event;
  91
  92         /*
  93          * We must use a mempool of dm_mpath_io structs so that we
  94          * can resubmit bios on error.
  95          */
  96         mempool_t *mpio_pool;
  97 };
  98
  99 /*
 100  * Context information attached to each bio we process.
 101  */
 102 struct dm_mpath_io {
 103         struct pgpath *pgpath;
 104         struct dm_bio_details details;
 105 };
 106
 107 typedef int (*action_fn) (struct pgpath *pgpath);
 108
 109 #define MIN_IOS 256     /* Mempool size */
 110
 111 static struct kmem_cache *_mpio_cache;
 112
 113 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 114 static void process_queued_ios(struct work_struct *work);
 115 static void trigger_event(struct work_struct *work);
 116 static void activate_path(struct work_struct *work);
 117 static void deactivate_path(struct work_struct *work);
 118
 119
 120 /*-----------------------------------------------
 121  * Allocation routines
 122  *-----------------------------------------------*/
 123
 124 static struct pgpath *alloc_pgpath(void)
 125 {
 126         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 127
 128         if (pgpath) {
 129                 pgpath->is_active = 1;
 130                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
 131         }
 132
 133         return pgpath;
 134 }
 135
 136 static void free_pgpath(struct pgpath *pgpath)
 137 {
 138         kfree(pgpath);
 139 }
 140
 141 static void deactivate_path(struct work_struct *work)
 142 {
 143         struct pgpath *pgpath =
 144                 container_of(work, struct pgpath, deactivate_path);
 145
 146         blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
 147 }
 148
 149 static struct priority_group *alloc_priority_group(void)
 150 {
 151         struct priority_group *pg;
 152
 153         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 154
 155         if (pg)
 156                 INIT_LIST_HEAD(&pg->pgpaths);
 157
 158         return pg;
 159 }
 160
 161 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 162 {
 163         unsigned long flags;
 164         struct pgpath *pgpath, *tmp;
 165         struct multipath *m = ti->private;
 166
 167         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 168                 list_del(&pgpath->list);
 169                 if (m->hw_handler_name)
 170                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 171                 dm_put_device(ti, pgpath->path.dev);
 172                 spin_lock_irqsave(&m->lock, flags);
 173                 if (m->pgpath_to_activate == pgpath)
 174                         m->pgpath_to_activate = NULL;
 175                 spin_unlock_irqrestore(&m->lock, flags);
 176                 free_pgpath(pgpath);
 177         }
 178 }
 179
 180 static void free_priority_group(struct priority_group *pg,
 181                                 struct dm_target *ti)
 182 {
 183         struct path_selector *ps = &pg->ps;
 184
 185         if (ps->type) {
 186                 ps->type->destroy(ps);
 187                 dm_put_path_selector(ps->type);
 188         }
 189
 190         free_pgpaths(&pg->pgpaths, ti);
 191         kfree(pg);
 192 }
 193
 194 static struct multipath *alloc_multipath(struct dm_target *ti)
 195 {
 196         struct multipath *m;
 197
 198         m = kzalloc(sizeof(*m), GFP_KERNEL);
 199         if (m) {
 200                 INIT_LIST_HEAD(&m->priority_groups);
 201                 spin_lock_init(&m->lock);
 202                 m->queue_io = 1;
 203                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 204                 INIT_WORK(&m->trigger_event, trigger_event);
 205                 INIT_WORK(&m->activate_path, activate_path);
 206                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 207                 if (!m->mpio_pool) {
 208                         kfree(m);
 209                         return NULL;
 210                 }
 211                 m->ti = ti;
 212                 ti->private = m;
 213         }
 214
 215         return m;
 216 }
 217
 218 static void free_multipath(struct multipath *m)
 219 {
 220         struct priority_group *pg, *tmp;
 221
 222         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 223                 list_del(&pg->list);
 224                 free_priority_group(pg, m->ti);
 225         }
 226
 227         kfree(m->hw_handler_name);
 228         mempool_destroy(m->mpio_pool);
 229         kfree(m);
 230 }
 231
 232
 233 /*-----------------------------------------------
 234  * Path selection
 235  *-----------------------------------------------*/
 236
 237 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 238 {
 239         m->current_pg = pgpath->pg;
 240
 241         /* Must we initialise the PG first, and queue I/O till it's ready? */
 242         if (m->hw_handler_name) {
 243                 m->pg_init_required = 1;
 244                 m->queue_io = 1;
 245         } else {
 246                 m->pg_init_required = 0;
 247                 m->queue_io = 0;
 248         }
 249
 250         m->pg_init_count = 0;
 251 }
 252
 253 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
 254 {
 255         struct dm_path *path;
 256
 257         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
 258         if (!path)
 259                 return -ENXIO;
 260
 261         m->current_pgpath = path_to_pgpath(path);
 262
 263         if (m->current_pg != pg)
 264                 __switch_pg(m, m->current_pgpath);
 265
 266         return 0;
 267 }
 268
 269 static void __choose_pgpath(struct multipath *m)
 270 {
 271         struct priority_group *pg;
 272         unsigned bypassed = 1;
 273
 274         if (!m->nr_valid_paths)
 275                 goto failed;
 276
 277         /* Were we instructed to switch PG? */
 278         if (m->next_pg) {
 279                 pg = m->next_pg;
 280                 m->next_pg = NULL;
 281                 if (!__choose_path_in_pg(m, pg))
 282                         return;
 283         }
 284
 285         /* Don't change PG until it has no remaining paths */
 286         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
 287                 return;
 288
 289         /*
 290          * Loop through priority groups until we find a valid path.
 291          * First time we skip PGs marked 'bypassed'.
 292          * Second time we only try the ones we skipped.
 293          */
 294         do {
 295                 list_for_each_entry(pg, &m->priority_groups, list) {
 296                         if (pg->bypassed == bypassed)
 297                                 continue;
 298                         if (!__choose_path_in_pg(m, pg))
 299                                 return;
 300                 }
 301         } while (bypassed--);
 302
 303 failed:
 304         m->current_pgpath = NULL;
 305         m->current_pg = NULL;
 306 }
 307
 308 /*
 309  * Check whether bios must be queued in the device-mapper core rather
 310  * than here in the target.
 311  *
 312  * m->lock must be held on entry.
 313  *
 314  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 315  * same value then we are not between multipath_presuspend()
 316  * and multipath_resume() calls and we have no need to check
 317  * for the DMF_NOFLUSH_SUSPENDING flag.
 318  */
 319 static int __must_push_back(struct multipath *m)
 320 {
 321         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 322                 dm_noflush_suspending(m->ti));
 323 }
 324
 325 static int map_io(struct multipath *m, struct bio *bio,
 326                   struct dm_mpath_io *mpio, unsigned was_queued)
 327 {
 328         int r = DM_MAPIO_REMAPPED;
 329         unsigned long flags;
 330         struct pgpath *pgpath;
 331
 332         spin_lock_irqsave(&m->lock, flags);
 333
 334         /* Do we need to select a new pgpath? */
 335         if (!m->current_pgpath ||
 336             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 337                 __choose_pgpath(m);
 338
 339         pgpath = m->current_pgpath;
 340
 341         if (was_queued)
 342                 m->queue_size--;
 343
 344         if ((pgpath && m->queue_io) ||
 345             (!pgpath && m->queue_if_no_path)) {
 346                 /* Queue for the daemon to resubmit */
 347                 bio_list_add(&m->queued_ios, bio);
 348                 m->queue_size++;
 349                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 350                     !m->queue_io)
 351                         queue_work(kmultipathd, &m->process_queued_ios);
 352                 pgpath = NULL;
 353                 r = DM_MAPIO_SUBMITTED;
 354         } else if (pgpath)
 355                 bio->bi_bdev = pgpath->path.dev->bdev;
 356         else if (__must_push_back(m))
 357                 r = DM_MAPIO_REQUEUE;
 358         else
 359                 r = -EIO;       /* Failed */
 360
 361         mpio->pgpath = pgpath;
 362
 363         spin_unlock_irqrestore(&m->lock, flags);
 364
 365         return r;
 366 }
 367
 368 /*
 369  * If we run out of usable paths, should we queue I/O or error it?
 370  */
 371 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 372                             unsigned save_old_value)
 373 {
 374         unsigned long flags;
 375
 376         spin_lock_irqsave(&m->lock, flags);
 377
 378         if (save_old_value)
 379                 m->saved_queue_if_no_path = m->queue_if_no_path;
 380         else
 381                 m->saved_queue_if_no_path = queue_if_no_path;
 382         m->queue_if_no_path = queue_if_no_path;
 383         if (!m->queue_if_no_path && m->queue_size)
 384                 queue_work(kmultipathd, &m->process_queued_ios);
 385
 386         spin_unlock_irqrestore(&m->lock, flags);
 387
 388         return 0;
 389 }
 390
 391 /*-----------------------------------------------------------------
 392  * The multipath daemon is responsible for resubmitting queued ios.
 393  *---------------------------------------------------------------*/
 394
 395 static void dispatch_queued_ios(struct multipath *m)
 396 {
 397         int r;
 398         unsigned long flags;
 399         struct bio *bio = NULL, *next;
 400         struct dm_mpath_io *mpio;
 401         union map_info *info;
 402
 403         spin_lock_irqsave(&m->lock, flags);
 404         bio = bio_list_get(&m->queued_ios);
 405         spin_unlock_irqrestore(&m->lock, flags);
 406
 407         while (bio) {
 408                 next = bio->bi_next;
 409                 bio->bi_next = NULL;
 410
 411                 info = dm_get_mapinfo(bio);
 412                 mpio = info->ptr;
 413
 414                 r = map_io(m, bio, mpio, 1);
 415                 if (r < 0)
 416                         bio_endio(bio, r);
 417                 else if (r == DM_MAPIO_REMAPPED)
 418                         generic_make_request(bio);
 419                 else if (r == DM_MAPIO_REQUEUE)
 420                         bio_endio(bio, -EIO);
 421
 422                 bio = next;
 423         }
 424 }
 425
 426 static void process_queued_ios(struct work_struct *work)
 427 {
 428         struct multipath *m =
 429                 container_of(work, struct multipath, process_queued_ios);
 430         struct pgpath *pgpath = NULL;
 431         unsigned init_required = 0, must_queue = 1;
 432         unsigned long flags;
 433
 434         spin_lock_irqsave(&m->lock, flags);
 435
 436         if (!m->queue_size)
 437                 goto out;
 438
 439         if (!m->current_pgpath)
 440                 __choose_pgpath(m);
 441
 442         pgpath = m->current_pgpath;
 443
 444         if ((pgpath && !m->queue_io) ||
 445             (!pgpath && !m->queue_if_no_path))
 446                 must_queue = 0;
 447
 448         if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
 449                 m->pgpath_to_activate = pgpath;
 450                 m->pg_init_count++;
 451                 m->pg_init_required = 0;
 452                 m->pg_init_in_progress = 1;
 453                 init_required = 1;
 454         }
 455
 456 out:
 457         spin_unlock_irqrestore(&m->lock, flags);
 458
 459         if (init_required)
 460                 queue_work(kmpath_handlerd, &m->activate_path);
 461
 462         if (!must_queue)
 463                 dispatch_queued_ios(m);
 464 }
 465
 466 /*
 467  * An event is triggered whenever a path is taken out of use.
 468  * Includes path failure and PG bypass.
 469  */
 470 static void trigger_event(struct work_struct *work)
 471 {
 472         struct multipath *m =
 473                 container_of(work, struct multipath, trigger_event);
 474
 475         dm_table_event(m->ti->table);
 476 }
 477
 478 /*-----------------------------------------------------------------
 479  * Constructor/argument parsing:
 480  * <#multipath feature args> [<arg>]*
 481  * <#hw_handler args> [hw_handler [<arg>]*]
 482  * <#priority groups>
 483  * <initial priority group>
 484  *     [<selector> <#selector args> [<arg>]*
 485  *      <#paths> <#per-path selector args>
 486  *         [<path> [<arg>]* ]+ ]+
 487  *---------------------------------------------------------------*/
 488 struct param {
 489         unsigned min;
 490         unsigned max;
 491         char *error;
 492 };
 493
 494 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 495 {
 496         if (!str ||
 497             (sscanf(str, "%u", v) != 1) ||
 498             (*v < param->min) ||
 499             (*v > param->max)) {
 500                 *error = param->error;
 501                 return -EINVAL;
 502         }
 503
 504         return 0;
 505 }
 506
 507 struct arg_set {
 508         unsigned argc;
 509         char **argv;
 510 };
 511
 512 static char *shift(struct arg_set *as)
 513 {
 514         char *r;
 515
 516         if (as->argc) {
 517                 as->argc--;
 518                 r = *as->argv;
 519                 as->argv++;
 520                 return r;
 521         }
 522
 523         return NULL;
 524 }
 525
 526 static void consume(struct arg_set *as, unsigned n)
 527 {
 528         BUG_ON (as->argc < n);
 529         as->argc -= n;
 530         as->argv += n;
 531 }
 532
 533 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 534                                struct dm_target *ti)
 535 {
 536         int r;
 537         struct path_selector_type *pst;
 538         unsigned ps_argc;
 539
 540         static struct param _params[] = {
 541                 {0, 1024, "invalid number of path selector args"},
 542         };
 543
 544         pst = dm_get_path_selector(shift(as));
 545         if (!pst) {
 546                 ti->error = "unknown path selector type";
 547                 return -EINVAL;
 548         }
 549
 550         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 551         if (r) {
 552                 dm_put_path_selector(pst);
 553                 return -EINVAL;
 554         }
 555
 556         if (ps_argc > as->argc) {
 557                 dm_put_path_selector(pst);
 558                 ti->error = "not enough arguments for path selector";
 559                 return -EINVAL;
 560         }
 561
 562         r = pst->create(&pg->ps, ps_argc, as->argv);
 563         if (r) {
 564                 dm_put_path_selector(pst);
 565                 ti->error = "path selector constructor failed";
 566                 return r;
 567         }
 568
 569         pg->ps.type = pst;
 570         consume(as, ps_argc);
 571
 572         return 0;
 573 }
 574
 575 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 576                                struct dm_target *ti)
 577 {
 578         int r;
 579         struct pgpath *p;
 580         struct multipath *m = ti->private;
 581
 582         /* we need at least a path arg */
 583         if (as->argc < 1) {
 584                 ti->error = "no device given";
 585                 return ERR_PTR(-EINVAL);
 586         }
 587
 588         p = alloc_pgpath();
 589         if (!p)
 590                 return ERR_PTR(-ENOMEM);
 591
 592         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 593                           dm_table_get_mode(ti->table), &p->path.dev);
 594         if (r) {
 595                 ti->error = "error getting device";
 596                 goto bad;
 597         }
 598
 599         if (m->hw_handler_name) {
 600                 r = scsi_dh_attach(bdev_get_queue(p->path.dev->bdev),
 601                                    m->hw_handler_name);
 602                 if (r < 0) {
 603                         dm_put_device(ti, p->path.dev);
 604                         goto bad;
 605                 }
 606         }
 607
 608         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 609         if (r) {
 610                 dm_put_device(ti, p->path.dev);
 611                 goto bad;
 612         }
 613
 614         return p;
 615
 616  bad:
 617         free_pgpath(p);
 618         return ERR_PTR(r);
 619 }
 620
 621 static struct priority_group *parse_priority_group(struct arg_set *as,
 622                                                    struct multipath *m)
 623 {
 624         static struct param _params[] = {
 625                 {1, 1024, "invalid number of paths"},
 626                 {0, 1024, "invalid number of selector args"}
 627         };
 628
 629         int r;
 630         unsigned i, nr_selector_args, nr_params;
 631         struct priority_group *pg;
 632         struct dm_target *ti = m->ti;
 633
 634         if (as->argc < 2) {
 635                 as->argc = 0;
 636                 ti->error = "not enough priority group arguments";
 637                 return ERR_PTR(-EINVAL);
 638         }
 639
 640         pg = alloc_priority_group();
 641         if (!pg) {
 642                 ti->error = "couldn't allocate priority group";
 643                 return ERR_PTR(-ENOMEM);
 644         }
 645         pg->m = m;
 646
 647         r = parse_path_selector(as, pg, ti);
 648         if (r)
 649                 goto bad;
 650
 651         /*
 652          * read the paths
 653          */
 654         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 655         if (r)
 656                 goto bad;
 657
 658         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 659         if (r)
 660                 goto bad;
 661
 662         nr_params = 1 + nr_selector_args;
 663         for (i = 0; i < pg->nr_pgpaths; i++) {
 664                 struct pgpath *pgpath;
 665                 struct arg_set path_args;
 666
 667                 if (as->argc < nr_params) {
 668                         ti->error = "not enough path parameters";
 669                         goto bad;
 670                 }
 671
 672                 path_args.argc = nr_params;
 673                 path_args.argv = as->argv;
 674
 675                 pgpath = parse_path(&path_args, &pg->ps, ti);
 676                 if (IS_ERR(pgpath)) {
 677                         r = PTR_ERR(pgpath);
 678                         goto bad;
 679                 }
 680
 681                 pgpath->pg = pg;
 682                 list_add_tail(&pgpath->list, &pg->pgpaths);
 683                 consume(as, nr_params);
 684         }
 685
 686         return pg;
 687
 688  bad:
 689         free_priority_group(pg, ti);
 690         return ERR_PTR(r);
 691 }
 692
 693 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 694 {
 695         unsigned hw_argc;
 696         struct dm_target *ti = m->ti;
 697
 698         static struct param _params[] = {
 699                 {0, 1024, "invalid number of hardware handler args"},
 700         };
 701
 702         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 703                 return -EINVAL;
 704
 705         if (!hw_argc)
 706                 return 0;
 707
 708         if (hw_argc > as->argc) {
 709                 ti->error = "not enough arguments for hardware handler";
 710                 return -EINVAL;
 711         }
 712
 713         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 714         request_module("scsi_dh_%s", m->hw_handler_name);
 715         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 716                 ti->error = "unknown hardware handler type";
 717                 kfree(m->hw_handler_name);
 718                 m->hw_handler_name = NULL;
 719                 return -EINVAL;
 720         }
 721
 722         if (hw_argc > 1)
 723                 DMWARN("Ignoring user-specified arguments for "
 724                        "hardware handler \"%s\"", m->hw_handler_name);
 725         consume(as, hw_argc - 1);
 726
 727         return 0;
 728 }
 729
 730 static int parse_features(struct arg_set *as, struct multipath *m)
 731 {
 732         int r;
 733         unsigned argc;
 734         struct dm_target *ti = m->ti;
 735         const char *param_name;
 736
 737         static struct param _params[] = {
 738                 {0, 3, "invalid number of feature args"},
 739                 {1, 50, "pg_init_retries must be between 1 and 50"},
 740         };
 741
 742         r = read_param(_params, shift(as), &argc, &ti->error);
 743         if (r)
 744                 return -EINVAL;
 745
 746         if (!argc)
 747                 return 0;
 748
 749         do {
 750                 param_name = shift(as);
 751                 argc--;
 752
 753                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 754                         r = queue_if_no_path(m, 1, 0);
 755                         continue;
 756                 }
 757
 758                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 759                     (argc >= 1)) {
 760                         r = read_param(_params + 1, shift(as),
 761                                        &m->pg_init_retries, &ti->error);
 762                         argc--;
 763                         continue;
 764                 }
 765
 766                 ti->error = "Unrecognised multipath feature request";
 767                 r = -EINVAL;
 768         } while (argc && !r);
 769
 770         return r;
 771 }
 772
 773 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 774                          char **argv)
 775 {
 776         /* target parameters */
 777         static struct param _params[] = {
 778                 {1, 1024, "invalid number of priority groups"},
 779                 {1, 1024, "invalid initial priority group number"},
 780         };
 781
 782         int r;
 783         struct multipath *m;
 784         struct arg_set as;
 785         unsigned pg_count = 0;
 786         unsigned next_pg_num;
 787
 788         as.argc = argc;
 789         as.argv = argv;
 790
 791         m = alloc_multipath(ti);
 792         if (!m) {
 793                 ti->error = "can't allocate multipath";
 794                 return -EINVAL;
 795         }
 796
 797         r = parse_features(&as, m);
 798         if (r)
 799                 goto bad;
 800
 801         r = parse_hw_handler(&as, m);
 802         if (r)
 803                 goto bad;
 804
 805         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 806         if (r)
 807                 goto bad;
 808
 809         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 810         if (r)
 811                 goto bad;
 812
 813         /* parse the priority groups */
 814         while (as.argc) {
 815                 struct priority_group *pg;
 816
 817                 pg = parse_priority_group(&as, m);
 818                 if (IS_ERR(pg)) {
 819                         r = PTR_ERR(pg);
 820                         goto bad;
 821                 }
 822
 823                 m->nr_valid_paths += pg->nr_pgpaths;
 824                 list_add_tail(&pg->list, &m->priority_groups);
 825                 pg_count++;
 826                 pg->pg_num = pg_count;
 827                 if (!--next_pg_num)
 828                         m->next_pg = pg;
 829         }
 830
 831         if (pg_count != m->nr_priority_groups) {
 832                 ti->error = "priority group count mismatch";
 833                 r = -EINVAL;
 834                 goto bad;
 835         }
 836
 837         return 0;
 838
 839  bad:
 840         free_multipath(m);
 841         return r;
 842 }
 843
 844 static void multipath_dtr(struct dm_target *ti)
 845 {
 846         struct multipath *m = (struct multipath *) ti->private;
 847
 848         flush_workqueue(kmpath_handlerd);
 849         flush_workqueue(kmultipathd);
 850         free_multipath(m);
 851 }
 852
 853 /*
 854  * Map bios, recording original fields for later in case we have to resubmit
 855  */
 856 static int multipath_map(struct dm_target *ti, struct bio *bio,
 857                          union map_info *map_context)
 858 {
 859         int r;
 860         struct dm_mpath_io *mpio;
 861         struct multipath *m = (struct multipath *) ti->private;
 862
 863         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 864         dm_bio_record(&mpio->details, bio);
 865
 866         map_context->ptr = mpio;
 867         bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
 868         r = map_io(m, bio, mpio, 0);
 869         if (r < 0 || r == DM_MAPIO_REQUEUE)
 870                 mempool_free(mpio, m->mpio_pool);
 871
 872         return r;
 873 }
 874
 875 /*
 876  * Take a path out of use.
 877  */
 878 static int fail_path(struct pgpath *pgpath)
 879 {
 880         unsigned long flags;
 881         struct multipath *m = pgpath->pg->m;
 882
 883         spin_lock_irqsave(&m->lock, flags);
 884
 885         if (!pgpath->is_active)
 886                 goto out;
 887
 888         DMWARN("Failing path %s.", pgpath->path.dev->name);
 889
 890         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 891         pgpath->is_active = 0;
 892         pgpath->fail_count++;
 893
 894         m->nr_valid_paths--;
 895
 896         if (pgpath == m->current_pgpath)
 897                 m->current_pgpath = NULL;
 898
 899         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 900                       pgpath->path.dev->name, m->nr_valid_paths);
 901
 902         schedule_work(&m->trigger_event);
 903         queue_work(kmultipathd, &pgpath->deactivate_path);
 904
 905 out:
 906         spin_unlock_irqrestore(&m->lock, flags);
 907
 908         return 0;
 909 }
 910
 911 /*
 912  * Reinstate a previously-failed path
 913  */
 914 static int reinstate_path(struct pgpath *pgpath)
 915 {
 916         int r = 0;
 917         unsigned long flags;
 918         struct multipath *m = pgpath->pg->m;
 919
 920         spin_lock_irqsave(&m->lock, flags);
 921
 922         if (pgpath->is_active)
 923                 goto out;
 924
 925         if (!pgpath->pg->ps.type->reinstate_path) {
 926                 DMWARN("Reinstate path not supported by path selector %s",
 927                        pgpath->pg->ps.type->name);
 928                 r = -EINVAL;
 929                 goto out;
 930         }
 931
 932         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 933         if (r)
 934                 goto out;
 935
 936         pgpath->is_active = 1;
 937
 938         m->current_pgpath = NULL;
 939         if (!m->nr_valid_paths++ && m->queue_size)
 940                 queue_work(kmultipathd, &m->process_queued_ios);
 941
 942         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 943                       pgpath->path.dev->name, m->nr_valid_paths);
 944
 945         schedule_work(&m->trigger_event);
 946
 947 out:
 948         spin_unlock_irqrestore(&m->lock, flags);
 949
 950         return r;
 951 }
 952
 953 /*
 954  * Fail or reinstate all paths that match the provided struct dm_dev.
 955  */
 956 static int action_dev(struct multipath *m, struct dm_dev *dev,
 957                       action_fn action)
 958 {
 959         int r = 0;
 960         struct pgpath *pgpath;
 961         struct priority_group *pg;
 962
 963         list_for_each_entry(pg, &m->priority_groups, list) {
 964                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 965                         if (pgpath->path.dev == dev)
 966                                 r = action(pgpath);
 967                 }
 968         }
 969
 970         return r;
 971 }
 972
 973 /*
 974  * Temporarily try to avoid having to use the specified PG
 975  */
 976 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 977                       int bypassed)
 978 {
 979         unsigned long flags;
 980
 981         spin_lock_irqsave(&m->lock, flags);
 982
 983         pg->bypassed = bypassed;
 984         m->current_pgpath = NULL;
 985         m->current_pg = NULL;
 986
 987         spin_unlock_irqrestore(&m->lock, flags);
 988
 989         schedule_work(&m->trigger_event);
 990 }
 991
 992 /*
 993  * Switch to using the specified PG from the next I/O that gets mapped
 994  */
 995 static int switch_pg_num(struct multipath *m, const char *pgstr)
 996 {
 997         struct priority_group *pg;
 998         unsigned pgnum;
 999         unsigned long flags;
1000
1001         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1002             (pgnum > m->nr_priority_groups)) {
1003                 DMWARN("invalid PG number supplied to switch_pg_num");
1004                 return -EINVAL;
1005         }
1006
1007         spin_lock_irqsave(&m->lock, flags);
1008         list_for_each_entry(pg, &m->priority_groups, list) {
1009                 pg->bypassed = 0;
1010                 if (--pgnum)
1011                         continue;
1012
1013                 m->current_pgpath = NULL;
1014                 m->current_pg = NULL;
1015                 m->next_pg = pg;
1016         }
1017         spin_unlock_irqrestore(&m->lock, flags);
1018
1019         schedule_work(&m->trigger_event);
1020         return 0;
1021 }
1022
1023 /*
1024  * Set/clear bypassed status of a PG.
1025  * PGs are numbered upwards from 1 in the order they were declared.
1026  */
1027 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1028 {
1029         struct priority_group *pg;
1030         unsigned pgnum;
1031
1032         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1033             (pgnum > m->nr_priority_groups)) {
1034                 DMWARN("invalid PG number supplied to bypass_pg");
1035                 return -EINVAL;
1036         }
1037
1038         list_for_each_entry(pg, &m->priority_groups, list) {
1039                 if (!--pgnum)
1040                         break;
1041         }
1042
1043         bypass_pg(m, pg, bypassed);
1044         return 0;
1045 }
1046
1047 /*
1048  * Should we retry pg_init immediately?
1049  */
1050 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1051 {
1052         unsigned long flags;
1053         int limit_reached = 0;
1054
1055         spin_lock_irqsave(&m->lock, flags);
1056
1057         if (m->pg_init_count <= m->pg_init_retries)
1058                 m->pg_init_required = 1;
1059         else
1060                 limit_reached = 1;
1061
1062         spin_unlock_irqrestore(&m->lock, flags);
1063
1064         return limit_reached;
1065 }
1066
1067 static void pg_init_done(struct dm_path *path, int errors)
1068 {
1069         struct pgpath *pgpath = path_to_pgpath(path);
1070         struct priority_group *pg = pgpath->pg;
1071         struct multipath *m = pg->m;
1072         unsigned long flags;
1073
1074         /* device or driver problems */
1075         switch (errors) {
1076         case SCSI_DH_OK:
1077                 break;
1078         case SCSI_DH_NOSYS:
1079                 if (!m->hw_handler_name) {
1080                         errors = 0;
1081                         break;
1082                 }
1083                 DMERR("Cannot failover device because scsi_dh_%s was not "
1084                       "loaded.", m->hw_handler_name);
1085                 /*
1086                  * Fail path for now, so we do not ping pong
1087                  */
1088                 fail_path(pgpath);
1089                 break;
1090         case SCSI_DH_DEV_TEMP_BUSY:
1091                 /*
1092                  * Probably doing something like FW upgrade on the
1093                  * controller so try the other pg.
1094                  */
1095                 bypass_pg(m, pg, 1);
1096                 break;
1097         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1098         case SCSI_DH_RETRY:
1099         case SCSI_DH_IMM_RETRY:
1100         case SCSI_DH_RES_TEMP_UNAVAIL:
1101                 if (pg_init_limit_reached(m, pgpath))
1102                         fail_path(pgpath);
1103                 errors = 0;
1104                 break;
1105         default:
1106                 /*
1107                  * We probably do not want to fail the path for a device
1108                  * error, but this is what the old dm did. In future
1109                  * patches we can do more advanced handling.
1110                  */
1111                 fail_path(pgpath);
1112         }
1113
1114         spin_lock_irqsave(&m->lock, flags);
1115         if (errors) {
1116                 DMERR("Could not failover device. Error %d.", errors);
1117                 m->current_pgpath = NULL;
1118                 m->current_pg = NULL;
1119         } else if (!m->pg_init_required) {
1120                 m->queue_io = 0;
1121                 pg->bypassed = 0;
1122         }
1123
1124         m->pg_init_in_progress = 0;
1125         queue_work(kmultipathd, &m->process_queued_ios);
1126         spin_unlock_irqrestore(&m->lock, flags);
1127 }
1128
1129 static void activate_path(struct work_struct *work)
1130 {
1131         int ret;
1132         struct multipath *m =
1133                 container_of(work, struct multipath, activate_path);
1134         struct dm_path *path;
1135         unsigned long flags;
1136
1137         spin_lock_irqsave(&m->lock, flags);
1138         path = &m->pgpath_to_activate->path;
1139         m->pgpath_to_activate = NULL;
1140         spin_unlock_irqrestore(&m->lock, flags);
1141         if (!path)
1142                 return;
1143         ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
1144         pg_init_done(path, ret);
1145 }
1146
1147 /*
1148  * end_io handling
1149  */
1150 static int do_end_io(struct multipath *m, struct bio *bio,
1151                      int error, struct dm_mpath_io *mpio)
1152 {
1153         unsigned long flags;
1154
1155         if (!error)
1156                 return 0;       /* I/O complete */
1157
1158         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1159                 return error;
1160
1161         if (error == -EOPNOTSUPP)
1162                 return error;
1163
1164         spin_lock_irqsave(&m->lock, flags);
1165         if (!m->nr_valid_paths) {
1166                 if (__must_push_back(m)) {
1167                         spin_unlock_irqrestore(&m->lock, flags);
1168                         return DM_ENDIO_REQUEUE;
1169                 } else if (!m->queue_if_no_path) {
1170                         spin_unlock_irqrestore(&m->lock, flags);
1171                         return -EIO;
1172                 } else {
1173                         spin_unlock_irqrestore(&m->lock, flags);
1174                         goto requeue;
1175                 }
1176         }
1177         spin_unlock_irqrestore(&m->lock, flags);
1178
1179         if (mpio->pgpath)
1180                 fail_path(mpio->pgpath);
1181
1182       requeue:
1183         dm_bio_restore(&mpio->details, bio);
1184
1185         /* queue for the daemon to resubmit or fail */
1186         spin_lock_irqsave(&m->lock, flags);
1187         bio_list_add(&m->queued_ios, bio);
1188         m->queue_size++;
1189         if (!m->queue_io)
1190                 queue_work(kmultipathd, &m->process_queued_ios);
1191         spin_unlock_irqrestore(&m->lock, flags);
1192
1193         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1194 }
1195
1196 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1197                             int error, union map_info *map_context)
1198 {
1199         struct multipath *m = ti->private;
1200         struct dm_mpath_io *mpio = map_context->ptr;
1201         struct pgpath *pgpath = mpio->pgpath;
1202         struct path_selector *ps;
1203         int r;
1204
1205         r  = do_end_io(m, bio, error, mpio);
1206         if (pgpath) {
1207                 ps = &pgpath->pg->ps;
1208                 if (ps->type->end_io)
1209                         ps->type->end_io(ps, &pgpath->path);
1210         }
1211         if (r != DM_ENDIO_INCOMPLETE)
1212                 mempool_free(mpio, m->mpio_pool);
1213
1214         return r;
1215 }
1216
1217 /*
1218  * Suspend can't complete until all the I/O is processed so if
1219  * the last path fails we must error any remaining I/O.
1220  * Note that if the freeze_bdev fails while suspending, the
1221  * queue_if_no_path state is lost - userspace should reset it.
1222  */
1223 static void multipath_presuspend(struct dm_target *ti)
1224 {
1225         struct multipath *m = (struct multipath *) ti->private;
1226
1227         queue_if_no_path(m, 0, 1);
1228 }
1229
1230 /*
1231  * Restore the queue_if_no_path setting.
1232  */
1233 static void multipath_resume(struct dm_target *ti)
1234 {
1235         struct multipath *m = (struct multipath *) ti->private;
1236         unsigned long flags;
1237
1238         spin_lock_irqsave(&m->lock, flags);
1239         m->queue_if_no_path = m->saved_queue_if_no_path;
1240         spin_unlock_irqrestore(&m->lock, flags);
1241 }
1242
1243 /*
1244  * Info output has the following format:
1245  * num_multipath_feature_args [multipath_feature_args]*
1246  * num_handler_status_args [handler_status_args]*
1247  * num_groups init_group_number
1248  *            [A|D|E num_ps_status_args [ps_status_args]*
1249  *             num_paths num_selector_args
1250  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1251  *
1252  * Table output has the following format (identical to the constructor string):
1253  * num_feature_args [features_args]*
1254  * num_handler_args hw_handler [hw_handler_args]*
1255  * num_groups init_group_number
1256  *     [priority selector-name num_ps_args [ps_args]*
1257  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1258  */
1259 static int multipath_status(struct dm_target *ti, status_type_t type,
1260                             char *result, unsigned int maxlen)
1261 {
1262         int sz = 0;
1263         unsigned long flags;
1264         struct multipath *m = (struct multipath *) ti->private;
1265         struct priority_group *pg;
1266         struct pgpath *p;
1267         unsigned pg_num;
1268         char state;
1269
1270         spin_lock_irqsave(&m->lock, flags);
1271
1272         /* Features */
1273         if (type == STATUSTYPE_INFO)
1274                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1275         else {
1276                 DMEMIT("%u ", m->queue_if_no_path +
1277                               (m->pg_init_retries > 0) * 2);
1278                 if (m->queue_if_no_path)
1279                         DMEMIT("queue_if_no_path ");
1280                 if (m->pg_init_retries)
1281                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1282         }
1283
1284         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1285                 DMEMIT("0 ");
1286         else
1287                 DMEMIT("1 %s ", m->hw_handler_name);
1288
1289         DMEMIT("%u ", m->nr_priority_groups);
1290
1291         if (m->next_pg)
1292                 pg_num = m->next_pg->pg_num;
1293         else if (m->current_pg)
1294                 pg_num = m->current_pg->pg_num;
1295         else
1296                         pg_num = 1;
1297
1298         DMEMIT("%u ", pg_num);
1299
1300         switch (type) {
1301         case STATUSTYPE_INFO:
1302                 list_for_each_entry(pg, &m->priority_groups, list) {
1303                         if (pg->bypassed)
1304                                 state = 'D';    /* Disabled */
1305                         else if (pg == m->current_pg)
1306                                 state = 'A';    /* Currently Active */
1307                         else
1308                                 state = 'E';    /* Enabled */
1309
1310                         DMEMIT("%c ", state);
1311
1312                         if (pg->ps.type->status)
1313                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1314                                                           result + sz,
1315                                                           maxlen - sz);
1316                         else
1317                                 DMEMIT("0 ");
1318
1319                         DMEMIT("%u %u ", pg->nr_pgpaths,
1320                                pg->ps.type->info_args);
1321
1322                         list_for_each_entry(p, &pg->pgpaths, list) {
1323                                 DMEMIT("%s %s %u ", p->path.dev->name,
1324                                        p->is_active ? "A" : "F",
1325                                        p->fail_count);
1326                                 if (pg->ps.type->status)
1327                                         sz += pg->ps.type->status(&pg->ps,
1328                                               &p->path, type, result + sz,
1329                                               maxlen - sz);
1330                         }
1331                 }
1332                 break;
1333
1334         case STATUSTYPE_TABLE:
1335                 list_for_each_entry(pg, &m->priority_groups, list) {
1336                         DMEMIT("%s ", pg->ps.type->name);
1337
1338                         if (pg->ps.type->status)
1339                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1340                                                           result + sz,
1341                                                           maxlen - sz);
1342                         else
1343                                 DMEMIT("0 ");
1344
1345                         DMEMIT("%u %u ", pg->nr_pgpaths,
1346                                pg->ps.type->table_args);
1347
1348                         list_for_each_entry(p, &pg->pgpaths, list) {
1349                                 DMEMIT("%s ", p->path.dev->name);
1350                                 if (pg->ps.type->status)
1351                                         sz += pg->ps.type->status(&pg->ps,
1352                                               &p->path, type, result + sz,
1353                                               maxlen - sz);
1354                         }
1355                 }
1356                 break;
1357         }
1358
1359         spin_unlock_irqrestore(&m->lock, flags);
1360
1361         return 0;
1362 }
1363
1364 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1365 {
1366         int r;
1367         struct dm_dev *dev;
1368         struct multipath *m = (struct multipath *) ti->private;
1369         action_fn action;
1370
1371         if (argc == 1) {
1372                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1373                         return queue_if_no_path(m, 1, 0);
1374                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1375                         return queue_if_no_path(m, 0, 0);
1376         }
1377
1378         if (argc != 2)
1379                 goto error;
1380
1381         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1382                 return bypass_pg_num(m, argv[1], 1);
1383         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1384                 return bypass_pg_num(m, argv[1], 0);
1385         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1386                 return switch_pg_num(m, argv[1]);
1387         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1388                 action = reinstate_path;
1389         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1390                 action = fail_path;
1391         else
1392                 goto error;
1393
1394         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1395                           dm_table_get_mode(ti->table), &dev);
1396         if (r) {
1397                 DMWARN("message: error getting device %s",
1398                        argv[1]);
1399                 return -EINVAL;
1400         }
1401
1402         r = action_dev(m, dev, action);
1403
1404         dm_put_device(ti, dev);
1405
1406         return r;
1407
1408 error:
1409         DMWARN("Unrecognised multipath message received.");
1410         return -EINVAL;
1411 }
1412
1413 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1414                            unsigned long arg)
1415 {
1416         struct multipath *m = (struct multipath *) ti->private;
1417         struct block_device *bdev = NULL;
1418         fmode_t mode = 0;
1419         unsigned long flags;
1420         int r = 0;
1421
1422         spin_lock_irqsave(&m->lock, flags);
1423
1424         if (!m->current_pgpath)
1425                 __choose_pgpath(m);
1426
1427         if (m->current_pgpath) {
1428                 bdev = m->current_pgpath->path.dev->bdev;
1429                 mode = m->current_pgpath->path.dev->mode;
1430         }
1431
1432         if (m->queue_io)
1433                 r = -EAGAIN;
1434         else if (!bdev)
1435                 r = -EIO;
1436
1437         spin_unlock_irqrestore(&m->lock, flags);
1438
1439         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1440 }
1441
1442 /*-----------------------------------------------------------------
1443  * Module setup
1444  *---------------------------------------------------------------*/
1445 static struct target_type multipath_target = {
1446         .name = "multipath",
1447         .version = {1, 0, 5},
1448         .module = THIS_MODULE,
1449         .ctr = multipath_ctr,
1450         .dtr = multipath_dtr,
1451         .map = multipath_map,
1452         .end_io = multipath_end_io,
1453         .presuspend = multipath_presuspend,
1454         .resume = multipath_resume,
1455         .status = multipath_status,
1456         .message = multipath_message,
1457         .ioctl  = multipath_ioctl,
1458 };
1459
1460 static int __init dm_multipath_init(void)
1461 {
1462         int r;
1463
1464         /* allocate a slab for the dm_ios */
1465         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1466         if (!_mpio_cache)
1467                 return -ENOMEM;
1468
1469         r = dm_register_target(&multipath_target);
1470         if (r < 0) {
1471                 DMERR("register failed %d", r);
1472                 kmem_cache_destroy(_mpio_cache);
1473                 return -EINVAL;
1474         }
1475
1476         kmultipathd = create_workqueue("kmpathd");
1477         if (!kmultipathd) {
1478                 DMERR("failed to create workqueue kmpathd");
1479                 dm_unregister_target(&multipath_target);
1480                 kmem_cache_destroy(_mpio_cache);
1481                 return -ENOMEM;
1482         }
1483
1484         /*
1485          * A separate workqueue is used to handle the device handlers
1486          * to avoid overloading existing workqueue. Overloading the
1487          * old workqueue would also create a bottleneck in the
1488          * path of the storage hardware device activation.
1489          */
1490         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1491         if (!kmpath_handlerd) {
1492                 DMERR("failed to create workqueue kmpath_handlerd");
1493                 destroy_workqueue(kmultipathd);
1494                 dm_unregister_target(&multipath_target);
1495                 kmem_cache_destroy(_mpio_cache);
1496                 return -ENOMEM;
1497         }
1498
1499         DMINFO("version %u.%u.%u loaded",
1500                multipath_target.version[0], multipath_target.version[1],
1501                multipath_target.version[2]);
1502
1503         return r;
1504 }
1505
1506 static void __exit dm_multipath_exit(void)
1507 {
1508         destroy_workqueue(kmpath_handlerd);
1509         destroy_workqueue(kmultipathd);
1510
1511         dm_unregister_target(&multipath_target);
1512         kmem_cache_destroy(_mpio_cache);
1513 }
1514
1515 module_init(dm_multipath_init);
1516 module_exit(dm_multipath_exit);
1517
1518 MODULE_DESCRIPTION(DM_NAME " multipath target");
1519 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1520 MODULE_LICENSE("GPL");