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         struct work_struct activate_path;
  39 };
  40
  41 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  42
  43 /*
  44  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  45  * Each has a path selector which controls which path gets used.
  46  */
  47 struct priority_group {
  48         struct list_head list;
  49
  50         struct multipath *m;            /* Owning multipath instance */
  51         struct path_selector ps;
  52
  53         unsigned pg_num;                /* Reference number */
  54         unsigned bypassed;              /* Temporarily bypass this PG? */
  55
  56         unsigned nr_pgpaths;            /* Number of paths in PG */
  57         struct list_head pgpaths;
  58 };
  59
  60 /* Multipath context */
  61 struct multipath {
  62         struct list_head list;
  63         struct dm_target *ti;
  64
  65         spinlock_t lock;
  66
  67         const char *hw_handler_name;
  68         unsigned nr_priority_groups;
  69         struct list_head priority_groups;
  70         unsigned pg_init_required;      /* pg_init needs calling? */
  71         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  72
  73         unsigned nr_valid_paths;        /* Total number of usable paths */
  74         struct pgpath *current_pgpath;
  75         struct priority_group *current_pg;
  76         struct priority_group *next_pg; /* Switch to this PG if set */
  77         unsigned repeat_count;          /* I/Os left before calling PS again */
  78
  79         unsigned queue_io;              /* Must we queue all I/O? */
  80         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  81         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  82         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  83         unsigned pg_init_count;         /* Number of times pg_init called */
  84
  85         struct work_struct process_queued_ios;
  86         struct bio_list queued_ios;
  87         unsigned queue_size;
  88
  89         struct work_struct trigger_event;
  90
  91         /*
  92          * We must use a mempool of dm_mpath_io structs so that we
  93          * can resubmit bios on error.
  94          */
  95         mempool_t *mpio_pool;
  96 };
  97
  98 /*
  99  * Context information attached to each bio we process.
 100  */
 101 struct dm_mpath_io {
 102         struct pgpath *pgpath;
 103         struct dm_bio_details details;
 104         size_t nr_bytes;
 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                 INIT_WORK(&pgpath->activate_path, activate_path);
 132         }
 133
 134         return pgpath;
 135 }
 136
 137 static void free_pgpath(struct pgpath *pgpath)
 138 {
 139         kfree(pgpath);
 140 }
 141
 142 static void deactivate_path(struct work_struct *work)
 143 {
 144         struct pgpath *pgpath =
 145                 container_of(work, struct pgpath, deactivate_path);
 146
 147         blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
 148 }
 149
 150 static struct priority_group *alloc_priority_group(void)
 151 {
 152         struct priority_group *pg;
 153
 154         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 155
 156         if (pg)
 157                 INIT_LIST_HEAD(&pg->pgpaths);
 158
 159         return pg;
 160 }
 161
 162 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 163 {
 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                 free_pgpath(pgpath);
 173         }
 174 }
 175
 176 static void free_priority_group(struct priority_group *pg,
 177                                 struct dm_target *ti)
 178 {
 179         struct path_selector *ps = &pg->ps;
 180
 181         if (ps->type) {
 182                 ps->type->destroy(ps);
 183                 dm_put_path_selector(ps->type);
 184         }
 185
 186         free_pgpaths(&pg->pgpaths, ti);
 187         kfree(pg);
 188 }
 189
 190 static struct multipath *alloc_multipath(struct dm_target *ti)
 191 {
 192         struct multipath *m;
 193
 194         m = kzalloc(sizeof(*m), GFP_KERNEL);
 195         if (m) {
 196                 INIT_LIST_HEAD(&m->priority_groups);
 197                 spin_lock_init(&m->lock);
 198                 m->queue_io = 1;
 199                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 200                 INIT_WORK(&m->trigger_event, trigger_event);
 201                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 202                 if (!m->mpio_pool) {
 203                         kfree(m);
 204                         return NULL;
 205                 }
 206                 m->ti = ti;
 207                 ti->private = m;
 208         }
 209
 210         return m;
 211 }
 212
 213 static void free_multipath(struct multipath *m)
 214 {
 215         struct priority_group *pg, *tmp;
 216
 217         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 218                 list_del(&pg->list);
 219                 free_priority_group(pg, m->ti);
 220         }
 221
 222         kfree(m->hw_handler_name);
 223         mempool_destroy(m->mpio_pool);
 224         kfree(m);
 225 }
 226
 227
 228 /*-----------------------------------------------
 229  * Path selection
 230  *-----------------------------------------------*/
 231
 232 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 233 {
 234         m->current_pg = pgpath->pg;
 235
 236         /* Must we initialise the PG first, and queue I/O till it's ready? */
 237         if (m->hw_handler_name) {
 238                 m->pg_init_required = 1;
 239                 m->queue_io = 1;
 240         } else {
 241                 m->pg_init_required = 0;
 242                 m->queue_io = 0;
 243         }
 244
 245         m->pg_init_count = 0;
 246 }
 247
 248 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 249                                size_t nr_bytes)
 250 {
 251         struct dm_path *path;
 252
 253         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 254         if (!path)
 255                 return -ENXIO;
 256
 257         m->current_pgpath = path_to_pgpath(path);
 258
 259         if (m->current_pg != pg)
 260                 __switch_pg(m, m->current_pgpath);
 261
 262         return 0;
 263 }
 264
 265 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 266 {
 267         struct priority_group *pg;
 268         unsigned bypassed = 1;
 269
 270         if (!m->nr_valid_paths)
 271                 goto failed;
 272
 273         /* Were we instructed to switch PG? */
 274         if (m->next_pg) {
 275                 pg = m->next_pg;
 276                 m->next_pg = NULL;
 277                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 278                         return;
 279         }
 280
 281         /* Don't change PG until it has no remaining paths */
 282         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 283                 return;
 284
 285         /*
 286          * Loop through priority groups until we find a valid path.
 287          * First time we skip PGs marked 'bypassed'.
 288          * Second time we only try the ones we skipped.
 289          */
 290         do {
 291                 list_for_each_entry(pg, &m->priority_groups, list) {
 292                         if (pg->bypassed == bypassed)
 293                                 continue;
 294                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 295                                 return;
 296                 }
 297         } while (bypassed--);
 298
 299 failed:
 300         m->current_pgpath = NULL;
 301         m->current_pg = NULL;
 302 }
 303
 304 /*
 305  * Check whether bios must be queued in the device-mapper core rather
 306  * than here in the target.
 307  *
 308  * m->lock must be held on entry.
 309  *
 310  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 311  * same value then we are not between multipath_presuspend()
 312  * and multipath_resume() calls and we have no need to check
 313  * for the DMF_NOFLUSH_SUSPENDING flag.
 314  */
 315 static int __must_push_back(struct multipath *m)
 316 {
 317         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 318                 dm_noflush_suspending(m->ti));
 319 }
 320
 321 static int map_io(struct multipath *m, struct bio *bio,
 322                   struct dm_mpath_io *mpio, unsigned was_queued)
 323 {
 324         int r = DM_MAPIO_REMAPPED;
 325         size_t nr_bytes = bio->bi_size;
 326         unsigned long flags;
 327         struct pgpath *pgpath;
 328
 329         spin_lock_irqsave(&m->lock, flags);
 330
 331         /* Do we need to select a new pgpath? */
 332         if (!m->current_pgpath ||
 333             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 334                 __choose_pgpath(m, nr_bytes);
 335
 336         pgpath = m->current_pgpath;
 337
 338         if (was_queued)
 339                 m->queue_size--;
 340
 341         if ((pgpath && m->queue_io) ||
 342             (!pgpath && m->queue_if_no_path)) {
 343                 /* Queue for the daemon to resubmit */
 344                 bio_list_add(&m->queued_ios, bio);
 345                 m->queue_size++;
 346                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 347                     !m->queue_io)
 348                         queue_work(kmultipathd, &m->process_queued_ios);
 349                 pgpath = NULL;
 350                 r = DM_MAPIO_SUBMITTED;
 351         } else if (pgpath)
 352                 bio->bi_bdev = pgpath->path.dev->bdev;
 353         else if (__must_push_back(m))
 354                 r = DM_MAPIO_REQUEUE;
 355         else
 356                 r = -EIO;       /* Failed */
 357
 358         mpio->pgpath = pgpath;
 359         mpio->nr_bytes = nr_bytes;
 360
 361         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 362                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 363                                               nr_bytes);
 364
 365         spin_unlock_irqrestore(&m->lock, flags);
 366
 367         return r;
 368 }
 369
 370 /*
 371  * If we run out of usable paths, should we queue I/O or error it?
 372  */
 373 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 374                             unsigned save_old_value)
 375 {
 376         unsigned long flags;
 377
 378         spin_lock_irqsave(&m->lock, flags);
 379
 380         if (save_old_value)
 381                 m->saved_queue_if_no_path = m->queue_if_no_path;
 382         else
 383                 m->saved_queue_if_no_path = queue_if_no_path;
 384         m->queue_if_no_path = queue_if_no_path;
 385         if (!m->queue_if_no_path && m->queue_size)
 386                 queue_work(kmultipathd, &m->process_queued_ios);
 387
 388         spin_unlock_irqrestore(&m->lock, flags);
 389
 390         return 0;
 391 }
 392
 393 /*-----------------------------------------------------------------
 394  * The multipath daemon is responsible for resubmitting queued ios.
 395  *---------------------------------------------------------------*/
 396
 397 static void dispatch_queued_ios(struct multipath *m)
 398 {
 399         int r;
 400         unsigned long flags;
 401         struct bio *bio = NULL, *next;
 402         struct dm_mpath_io *mpio;
 403         union map_info *info;
 404
 405         spin_lock_irqsave(&m->lock, flags);
 406         bio = bio_list_get(&m->queued_ios);
 407         spin_unlock_irqrestore(&m->lock, flags);
 408
 409         while (bio) {
 410                 next = bio->bi_next;
 411                 bio->bi_next = NULL;
 412
 413                 info = dm_get_mapinfo(bio);
 414                 mpio = info->ptr;
 415
 416                 r = map_io(m, bio, mpio, 1);
 417                 if (r < 0)
 418                         bio_endio(bio, r);
 419                 else if (r == DM_MAPIO_REMAPPED)
 420                         generic_make_request(bio);
 421                 else if (r == DM_MAPIO_REQUEUE)
 422                         bio_endio(bio, -EIO);
 423
 424                 bio = next;
 425         }
 426 }
 427
 428 static void process_queued_ios(struct work_struct *work)
 429 {
 430         struct multipath *m =
 431                 container_of(work, struct multipath, process_queued_ios);
 432         struct pgpath *pgpath = NULL, *tmp;
 433         unsigned must_queue = 1;
 434         unsigned long flags;
 435
 436         spin_lock_irqsave(&m->lock, flags);
 437
 438         if (!m->queue_size)
 439                 goto out;
 440
 441         if (!m->current_pgpath)
 442                 __choose_pgpath(m, 0);
 443
 444         pgpath = m->current_pgpath;
 445
 446         if ((pgpath && !m->queue_io) ||
 447             (!pgpath && !m->queue_if_no_path))
 448                 must_queue = 0;
 449
 450         if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
 451                 m->pg_init_count++;
 452                 m->pg_init_required = 0;
 453                 list_for_each_entry(tmp, &pgpath->pg->pgpaths, list) {
 454                         if (queue_work(kmpath_handlerd, &tmp->activate_path))
 455                                 m->pg_init_in_progress++;
 456                 }
 457         }
 458 out:
 459         spin_unlock_irqrestore(&m->lock, flags);
 460         if (!must_queue)
 461                 dispatch_queued_ios(m);
 462 }
 463
 464 /*
 465  * An event is triggered whenever a path is taken out of use.
 466  * Includes path failure and PG bypass.
 467  */
 468 static void trigger_event(struct work_struct *work)
 469 {
 470         struct multipath *m =
 471                 container_of(work, struct multipath, trigger_event);
 472
 473         dm_table_event(m->ti->table);
 474 }
 475
 476 /*-----------------------------------------------------------------
 477  * Constructor/argument parsing:
 478  * <#multipath feature args> [<arg>]*
 479  * <#hw_handler args> [hw_handler [<arg>]*]
 480  * <#priority groups>
 481  * <initial priority group>
 482  *     [<selector> <#selector args> [<arg>]*
 483  *      <#paths> <#per-path selector args>
 484  *         [<path> [<arg>]* ]+ ]+
 485  *---------------------------------------------------------------*/
 486 struct param {
 487         unsigned min;
 488         unsigned max;
 489         char *error;
 490 };
 491
 492 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 493 {
 494         if (!str ||
 495             (sscanf(str, "%u", v) != 1) ||
 496             (*v < param->min) ||
 497             (*v > param->max)) {
 498                 *error = param->error;
 499                 return -EINVAL;
 500         }
 501
 502         return 0;
 503 }
 504
 505 struct arg_set {
 506         unsigned argc;
 507         char **argv;
 508 };
 509
 510 static char *shift(struct arg_set *as)
 511 {
 512         char *r;
 513
 514         if (as->argc) {
 515                 as->argc--;
 516                 r = *as->argv;
 517                 as->argv++;
 518                 return r;
 519         }
 520
 521         return NULL;
 522 }
 523
 524 static void consume(struct arg_set *as, unsigned n)
 525 {
 526         BUG_ON (as->argc < n);
 527         as->argc -= n;
 528         as->argv += n;
 529 }
 530
 531 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 532                                struct dm_target *ti)
 533 {
 534         int r;
 535         struct path_selector_type *pst;
 536         unsigned ps_argc;
 537
 538         static struct param _params[] = {
 539                 {0, 1024, "invalid number of path selector args"},
 540         };
 541
 542         pst = dm_get_path_selector(shift(as));
 543         if (!pst) {
 544                 ti->error = "unknown path selector type";
 545                 return -EINVAL;
 546         }
 547
 548         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 549         if (r) {
 550                 dm_put_path_selector(pst);
 551                 return -EINVAL;
 552         }
 553
 554         if (ps_argc > as->argc) {
 555                 dm_put_path_selector(pst);
 556                 ti->error = "not enough arguments for path selector";
 557                 return -EINVAL;
 558         }
 559
 560         r = pst->create(&pg->ps, ps_argc, as->argv);
 561         if (r) {
 562                 dm_put_path_selector(pst);
 563                 ti->error = "path selector constructor failed";
 564                 return r;
 565         }
 566
 567         pg->ps.type = pst;
 568         consume(as, ps_argc);
 569
 570         return 0;
 571 }
 572
 573 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 574                                struct dm_target *ti)
 575 {
 576         int r;
 577         struct pgpath *p;
 578         struct multipath *m = ti->private;
 579
 580         /* we need at least a path arg */
 581         if (as->argc < 1) {
 582                 ti->error = "no device given";
 583                 return ERR_PTR(-EINVAL);
 584         }
 585
 586         p = alloc_pgpath();
 587         if (!p)
 588                 return ERR_PTR(-ENOMEM);
 589
 590         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 591                           dm_table_get_mode(ti->table), &p->path.dev);
 592         if (r) {
 593                 ti->error = "error getting device";
 594                 goto bad;
 595         }
 596
 597         if (m->hw_handler_name) {
 598                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 599
 600                 r = scsi_dh_attach(q, m->hw_handler_name);
 601                 if (r == -EBUSY) {
 602                         /*
 603                          * Already attached to different hw_handler,
 604                          * try to reattach with correct one.
 605                          */
 606                         scsi_dh_detach(q);
 607                         r = scsi_dh_attach(q, m->hw_handler_name);
 608                 }
 609
 610                 if (r < 0) {
 611                         ti->error = "error attaching hardware handler";
 612                         dm_put_device(ti, p->path.dev);
 613                         goto bad;
 614                 }
 615         }
 616
 617         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 618         if (r) {
 619                 dm_put_device(ti, p->path.dev);
 620                 goto bad;
 621         }
 622
 623         return p;
 624
 625  bad:
 626         free_pgpath(p);
 627         return ERR_PTR(r);
 628 }
 629
 630 static struct priority_group *parse_priority_group(struct arg_set *as,
 631                                                    struct multipath *m)
 632 {
 633         static struct param _params[] = {
 634                 {1, 1024, "invalid number of paths"},
 635                 {0, 1024, "invalid number of selector args"}
 636         };
 637
 638         int r;
 639         unsigned i, nr_selector_args, nr_params;
 640         struct priority_group *pg;
 641         struct dm_target *ti = m->ti;
 642
 643         if (as->argc < 2) {
 644                 as->argc = 0;
 645                 ti->error = "not enough priority group arguments";
 646                 return ERR_PTR(-EINVAL);
 647         }
 648
 649         pg = alloc_priority_group();
 650         if (!pg) {
 651                 ti->error = "couldn't allocate priority group";
 652                 return ERR_PTR(-ENOMEM);
 653         }
 654         pg->m = m;
 655
 656         r = parse_path_selector(as, pg, ti);
 657         if (r)
 658                 goto bad;
 659
 660         /*
 661          * read the paths
 662          */
 663         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 664         if (r)
 665                 goto bad;
 666
 667         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 668         if (r)
 669                 goto bad;
 670
 671         nr_params = 1 + nr_selector_args;
 672         for (i = 0; i < pg->nr_pgpaths; i++) {
 673                 struct pgpath *pgpath;
 674                 struct arg_set path_args;
 675
 676                 if (as->argc < nr_params) {
 677                         ti->error = "not enough path parameters";
 678                         goto bad;
 679                 }
 680
 681                 path_args.argc = nr_params;
 682                 path_args.argv = as->argv;
 683
 684                 pgpath = parse_path(&path_args, &pg->ps, ti);
 685                 if (IS_ERR(pgpath)) {
 686                         r = PTR_ERR(pgpath);
 687                         goto bad;
 688                 }
 689
 690                 pgpath->pg = pg;
 691                 list_add_tail(&pgpath->list, &pg->pgpaths);
 692                 consume(as, nr_params);
 693         }
 694
 695         return pg;
 696
 697  bad:
 698         free_priority_group(pg, ti);
 699         return ERR_PTR(r);
 700 }
 701
 702 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 703 {
 704         unsigned hw_argc;
 705         struct dm_target *ti = m->ti;
 706
 707         static struct param _params[] = {
 708                 {0, 1024, "invalid number of hardware handler args"},
 709         };
 710
 711         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 712                 return -EINVAL;
 713
 714         if (!hw_argc)
 715                 return 0;
 716
 717         if (hw_argc > as->argc) {
 718                 ti->error = "not enough arguments for hardware handler";
 719                 return -EINVAL;
 720         }
 721
 722         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 723         request_module("scsi_dh_%s", m->hw_handler_name);
 724         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 725                 ti->error = "unknown hardware handler type";
 726                 kfree(m->hw_handler_name);
 727                 m->hw_handler_name = NULL;
 728                 return -EINVAL;
 729         }
 730
 731         if (hw_argc > 1)
 732                 DMWARN("Ignoring user-specified arguments for "
 733                        "hardware handler \"%s\"", m->hw_handler_name);
 734         consume(as, hw_argc - 1);
 735
 736         return 0;
 737 }
 738
 739 static int parse_features(struct arg_set *as, struct multipath *m)
 740 {
 741         int r;
 742         unsigned argc;
 743         struct dm_target *ti = m->ti;
 744         const char *param_name;
 745
 746         static struct param _params[] = {
 747                 {0, 3, "invalid number of feature args"},
 748                 {1, 50, "pg_init_retries must be between 1 and 50"},
 749         };
 750
 751         r = read_param(_params, shift(as), &argc, &ti->error);
 752         if (r)
 753                 return -EINVAL;
 754
 755         if (!argc)
 756                 return 0;
 757
 758         do {
 759                 param_name = shift(as);
 760                 argc--;
 761
 762                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 763                         r = queue_if_no_path(m, 1, 0);
 764                         continue;
 765                 }
 766
 767                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 768                     (argc >= 1)) {
 769                         r = read_param(_params + 1, shift(as),
 770                                        &m->pg_init_retries, &ti->error);
 771                         argc--;
 772                         continue;
 773                 }
 774
 775                 ti->error = "Unrecognised multipath feature request";
 776                 r = -EINVAL;
 777         } while (argc && !r);
 778
 779         return r;
 780 }
 781
 782 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 783                          char **argv)
 784 {
 785         /* target parameters */
 786         static struct param _params[] = {
 787                 {1, 1024, "invalid number of priority groups"},
 788                 {1, 1024, "invalid initial priority group number"},
 789         };
 790
 791         int r;
 792         struct multipath *m;
 793         struct arg_set as;
 794         unsigned pg_count = 0;
 795         unsigned next_pg_num;
 796
 797         as.argc = argc;
 798         as.argv = argv;
 799
 800         m = alloc_multipath(ti);
 801         if (!m) {
 802                 ti->error = "can't allocate multipath";
 803                 return -EINVAL;
 804         }
 805
 806         r = parse_features(&as, m);
 807         if (r)
 808                 goto bad;
 809
 810         r = parse_hw_handler(&as, m);
 811         if (r)
 812                 goto bad;
 813
 814         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 815         if (r)
 816                 goto bad;
 817
 818         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 819         if (r)
 820                 goto bad;
 821
 822         /* parse the priority groups */
 823         while (as.argc) {
 824                 struct priority_group *pg;
 825
 826                 pg = parse_priority_group(&as, m);
 827                 if (IS_ERR(pg)) {
 828                         r = PTR_ERR(pg);
 829                         goto bad;
 830                 }
 831
 832                 m->nr_valid_paths += pg->nr_pgpaths;
 833                 list_add_tail(&pg->list, &m->priority_groups);
 834                 pg_count++;
 835                 pg->pg_num = pg_count;
 836                 if (!--next_pg_num)
 837                         m->next_pg = pg;
 838         }
 839
 840         if (pg_count != m->nr_priority_groups) {
 841                 ti->error = "priority group count mismatch";
 842                 r = -EINVAL;
 843                 goto bad;
 844         }
 845
 846         ti->num_flush_requests = 1;
 847
 848         return 0;
 849
 850  bad:
 851         free_multipath(m);
 852         return r;
 853 }
 854
 855 static void multipath_dtr(struct dm_target *ti)
 856 {
 857         struct multipath *m = (struct multipath *) ti->private;
 858
 859         flush_workqueue(kmpath_handlerd);
 860         flush_workqueue(kmultipathd);
 861         flush_scheduled_work();
 862         free_multipath(m);
 863 }
 864
 865 /*
 866  * Map bios, recording original fields for later in case we have to resubmit
 867  */
 868 static int multipath_map(struct dm_target *ti, struct bio *bio,
 869                          union map_info *map_context)
 870 {
 871         int r;
 872         struct dm_mpath_io *mpio;
 873         struct multipath *m = (struct multipath *) ti->private;
 874
 875         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 876         dm_bio_record(&mpio->details, bio);
 877
 878         map_context->ptr = mpio;
 879         bio->bi_rw |= (1 << BIO_RW_FAILFAST_TRANSPORT);
 880         r = map_io(m, bio, mpio, 0);
 881         if (r < 0 || r == DM_MAPIO_REQUEUE)
 882                 mempool_free(mpio, m->mpio_pool);
 883
 884         return r;
 885 }
 886
 887 /*
 888  * Take a path out of use.
 889  */
 890 static int fail_path(struct pgpath *pgpath)
 891 {
 892         unsigned long flags;
 893         struct multipath *m = pgpath->pg->m;
 894
 895         spin_lock_irqsave(&m->lock, flags);
 896
 897         if (!pgpath->is_active)
 898                 goto out;
 899
 900         DMWARN("Failing path %s.", pgpath->path.dev->name);
 901
 902         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 903         pgpath->is_active = 0;
 904         pgpath->fail_count++;
 905
 906         m->nr_valid_paths--;
 907
 908         if (pgpath == m->current_pgpath)
 909                 m->current_pgpath = NULL;
 910
 911         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 912                       pgpath->path.dev->name, m->nr_valid_paths);
 913
 914         schedule_work(&m->trigger_event);
 915         queue_work(kmultipathd, &pgpath->deactivate_path);
 916
 917 out:
 918         spin_unlock_irqrestore(&m->lock, flags);
 919
 920         return 0;
 921 }
 922
 923 /*
 924  * Reinstate a previously-failed path
 925  */
 926 static int reinstate_path(struct pgpath *pgpath)
 927 {
 928         int r = 0;
 929         unsigned long flags;
 930         struct multipath *m = pgpath->pg->m;
 931
 932         spin_lock_irqsave(&m->lock, flags);
 933
 934         if (pgpath->is_active)
 935                 goto out;
 936
 937         if (!pgpath->pg->ps.type->reinstate_path) {
 938                 DMWARN("Reinstate path not supported by path selector %s",
 939                        pgpath->pg->ps.type->name);
 940                 r = -EINVAL;
 941                 goto out;
 942         }
 943
 944         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 945         if (r)
 946                 goto out;
 947
 948         pgpath->is_active = 1;
 949
 950         if (!m->nr_valid_paths++ && m->queue_size) {
 951                 m->current_pgpath = NULL;
 952                 queue_work(kmultipathd, &m->process_queued_ios);
 953         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
 954                 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
 955                         m->pg_init_in_progress++;
 956         }
 957
 958         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 959                       pgpath->path.dev->name, m->nr_valid_paths);
 960
 961         schedule_work(&m->trigger_event);
 962
 963 out:
 964         spin_unlock_irqrestore(&m->lock, flags);
 965
 966         return r;
 967 }
 968
 969 /*
 970  * Fail or reinstate all paths that match the provided struct dm_dev.
 971  */
 972 static int action_dev(struct multipath *m, struct dm_dev *dev,
 973                       action_fn action)
 974 {
 975         int r = 0;
 976         struct pgpath *pgpath;
 977         struct priority_group *pg;
 978
 979         list_for_each_entry(pg, &m->priority_groups, list) {
 980                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 981                         if (pgpath->path.dev == dev)
 982                                 r = action(pgpath);
 983                 }
 984         }
 985
 986         return r;
 987 }
 988
 989 /*
 990  * Temporarily try to avoid having to use the specified PG
 991  */
 992 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 993                       int bypassed)
 994 {
 995         unsigned long flags;
 996
 997         spin_lock_irqsave(&m->lock, flags);
 998
 999         pg->bypassed = bypassed;
1000         m->current_pgpath = NULL;
1001         m->current_pg = NULL;
1002
1003         spin_unlock_irqrestore(&m->lock, flags);
1004
1005         schedule_work(&m->trigger_event);
1006 }
1007
1008 /*
1009  * Switch to using the specified PG from the next I/O that gets mapped
1010  */
1011 static int switch_pg_num(struct multipath *m, const char *pgstr)
1012 {
1013         struct priority_group *pg;
1014         unsigned pgnum;
1015         unsigned long flags;
1016
1017         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1018             (pgnum > m->nr_priority_groups)) {
1019                 DMWARN("invalid PG number supplied to switch_pg_num");
1020                 return -EINVAL;
1021         }
1022
1023         spin_lock_irqsave(&m->lock, flags);
1024         list_for_each_entry(pg, &m->priority_groups, list) {
1025                 pg->bypassed = 0;
1026                 if (--pgnum)
1027                         continue;
1028
1029                 m->current_pgpath = NULL;
1030                 m->current_pg = NULL;
1031                 m->next_pg = pg;
1032         }
1033         spin_unlock_irqrestore(&m->lock, flags);
1034
1035         schedule_work(&m->trigger_event);
1036         return 0;
1037 }
1038
1039 /*
1040  * Set/clear bypassed status of a PG.
1041  * PGs are numbered upwards from 1 in the order they were declared.
1042  */
1043 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1044 {
1045         struct priority_group *pg;
1046         unsigned pgnum;
1047
1048         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1049             (pgnum > m->nr_priority_groups)) {
1050                 DMWARN("invalid PG number supplied to bypass_pg");
1051                 return -EINVAL;
1052         }
1053
1054         list_for_each_entry(pg, &m->priority_groups, list) {
1055                 if (!--pgnum)
1056                         break;
1057         }
1058
1059         bypass_pg(m, pg, bypassed);
1060         return 0;
1061 }
1062
1063 /*
1064  * Should we retry pg_init immediately?
1065  */
1066 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1067 {
1068         unsigned long flags;
1069         int limit_reached = 0;
1070
1071         spin_lock_irqsave(&m->lock, flags);
1072
1073         if (m->pg_init_count <= m->pg_init_retries)
1074                 m->pg_init_required = 1;
1075         else
1076                 limit_reached = 1;
1077
1078         spin_unlock_irqrestore(&m->lock, flags);
1079
1080         return limit_reached;
1081 }
1082
1083 static void pg_init_done(struct dm_path *path, int errors)
1084 {
1085         struct pgpath *pgpath = path_to_pgpath(path);
1086         struct priority_group *pg = pgpath->pg;
1087         struct multipath *m = pg->m;
1088         unsigned long flags;
1089
1090         /* device or driver problems */
1091         switch (errors) {
1092         case SCSI_DH_OK:
1093                 break;
1094         case SCSI_DH_NOSYS:
1095                 if (!m->hw_handler_name) {
1096                         errors = 0;
1097                         break;
1098                 }
1099                 DMERR("Cannot failover device because scsi_dh_%s was not "
1100                       "loaded.", m->hw_handler_name);
1101                 /*
1102                  * Fail path for now, so we do not ping pong
1103                  */
1104                 fail_path(pgpath);
1105                 break;
1106         case SCSI_DH_DEV_TEMP_BUSY:
1107                 /*
1108                  * Probably doing something like FW upgrade on the
1109                  * controller so try the other pg.
1110                  */
1111                 bypass_pg(m, pg, 1);
1112                 break;
1113         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1114         case SCSI_DH_RETRY:
1115         case SCSI_DH_IMM_RETRY:
1116         case SCSI_DH_RES_TEMP_UNAVAIL:
1117                 if (pg_init_limit_reached(m, pgpath))
1118                         fail_path(pgpath);
1119                 errors = 0;
1120                 break;
1121         default:
1122                 /*
1123                  * We probably do not want to fail the path for a device
1124                  * error, but this is what the old dm did. In future
1125                  * patches we can do more advanced handling.
1126                  */
1127                 fail_path(pgpath);
1128         }
1129
1130         spin_lock_irqsave(&m->lock, flags);
1131         if (errors) {
1132                 if (pgpath == m->current_pgpath) {
1133                         DMERR("Could not failover device. Error %d.", errors);
1134                         m->current_pgpath = NULL;
1135                         m->current_pg = NULL;
1136                 }
1137         } else if (!m->pg_init_required) {
1138                 m->queue_io = 0;
1139                 pg->bypassed = 0;
1140         }
1141
1142         m->pg_init_in_progress--;
1143         if (!m->pg_init_in_progress)
1144                 queue_work(kmultipathd, &m->process_queued_ios);
1145         spin_unlock_irqrestore(&m->lock, flags);
1146 }
1147
1148 static void activate_path(struct work_struct *work)
1149 {
1150         int ret;
1151         struct pgpath *pgpath =
1152                 container_of(work, struct pgpath, activate_path);
1153
1154         ret = scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev));
1155         pg_init_done(&pgpath->path, ret);
1156 }
1157
1158 /*
1159  * end_io handling
1160  */
1161 static int do_end_io(struct multipath *m, struct bio *bio,
1162                      int error, struct dm_mpath_io *mpio)
1163 {
1164         unsigned long flags;
1165
1166         if (!error)
1167                 return 0;       /* I/O complete */
1168
1169         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1170                 return error;
1171
1172         if (error == -EOPNOTSUPP)
1173                 return error;
1174
1175         spin_lock_irqsave(&m->lock, flags);
1176         if (!m->nr_valid_paths) {
1177                 if (__must_push_back(m)) {
1178                         spin_unlock_irqrestore(&m->lock, flags);
1179                         return DM_ENDIO_REQUEUE;
1180                 } else if (!m->queue_if_no_path) {
1181                         spin_unlock_irqrestore(&m->lock, flags);
1182                         return -EIO;
1183                 } else {
1184                         spin_unlock_irqrestore(&m->lock, flags);
1185                         goto requeue;
1186                 }
1187         }
1188         spin_unlock_irqrestore(&m->lock, flags);
1189
1190         if (mpio->pgpath)
1191                 fail_path(mpio->pgpath);
1192
1193       requeue:
1194         dm_bio_restore(&mpio->details, bio);
1195
1196         /* queue for the daemon to resubmit or fail */
1197         spin_lock_irqsave(&m->lock, flags);
1198         bio_list_add(&m->queued_ios, bio);
1199         m->queue_size++;
1200         if (!m->queue_io)
1201                 queue_work(kmultipathd, &m->process_queued_ios);
1202         spin_unlock_irqrestore(&m->lock, flags);
1203
1204         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1205 }
1206
1207 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1208                             int error, union map_info *map_context)
1209 {
1210         struct multipath *m = ti->private;
1211         struct dm_mpath_io *mpio = map_context->ptr;
1212         struct pgpath *pgpath = mpio->pgpath;
1213         struct path_selector *ps;
1214         int r;
1215
1216         r  = do_end_io(m, bio, error, mpio);
1217         if (pgpath) {
1218                 ps = &pgpath->pg->ps;
1219                 if (ps->type->end_io)
1220                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1221         }
1222         if (r != DM_ENDIO_INCOMPLETE)
1223                 mempool_free(mpio, m->mpio_pool);
1224
1225         return r;
1226 }
1227
1228 /*
1229  * Suspend can't complete until all the I/O is processed so if
1230  * the last path fails we must error any remaining I/O.
1231  * Note that if the freeze_bdev fails while suspending, the
1232  * queue_if_no_path state is lost - userspace should reset it.
1233  */
1234 static void multipath_presuspend(struct dm_target *ti)
1235 {
1236         struct multipath *m = (struct multipath *) ti->private;
1237
1238         queue_if_no_path(m, 0, 1);
1239 }
1240
1241 /*
1242  * Restore the queue_if_no_path setting.
1243  */
1244 static void multipath_resume(struct dm_target *ti)
1245 {
1246         struct multipath *m = (struct multipath *) ti->private;
1247         unsigned long flags;
1248
1249         spin_lock_irqsave(&m->lock, flags);
1250         m->queue_if_no_path = m->saved_queue_if_no_path;
1251         spin_unlock_irqrestore(&m->lock, flags);
1252 }
1253
1254 /*
1255  * Info output has the following format:
1256  * num_multipath_feature_args [multipath_feature_args]*
1257  * num_handler_status_args [handler_status_args]*
1258  * num_groups init_group_number
1259  *            [A|D|E num_ps_status_args [ps_status_args]*
1260  *             num_paths num_selector_args
1261  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1262  *
1263  * Table output has the following format (identical to the constructor string):
1264  * num_feature_args [features_args]*
1265  * num_handler_args hw_handler [hw_handler_args]*
1266  * num_groups init_group_number
1267  *     [priority selector-name num_ps_args [ps_args]*
1268  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1269  */
1270 static int multipath_status(struct dm_target *ti, status_type_t type,
1271                             char *result, unsigned int maxlen)
1272 {
1273         int sz = 0;
1274         unsigned long flags;
1275         struct multipath *m = (struct multipath *) ti->private;
1276         struct priority_group *pg;
1277         struct pgpath *p;
1278         unsigned pg_num;
1279         char state;
1280
1281         spin_lock_irqsave(&m->lock, flags);
1282
1283         /* Features */
1284         if (type == STATUSTYPE_INFO)
1285                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1286         else {
1287                 DMEMIT("%u ", m->queue_if_no_path +
1288                               (m->pg_init_retries > 0) * 2);
1289                 if (m->queue_if_no_path)
1290                         DMEMIT("queue_if_no_path ");
1291                 if (m->pg_init_retries)
1292                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1293         }
1294
1295         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1296                 DMEMIT("0 ");
1297         else
1298                 DMEMIT("1 %s ", m->hw_handler_name);
1299
1300         DMEMIT("%u ", m->nr_priority_groups);
1301
1302         if (m->next_pg)
1303                 pg_num = m->next_pg->pg_num;
1304         else if (m->current_pg)
1305                 pg_num = m->current_pg->pg_num;
1306         else
1307                         pg_num = 1;
1308
1309         DMEMIT("%u ", pg_num);
1310
1311         switch (type) {
1312         case STATUSTYPE_INFO:
1313                 list_for_each_entry(pg, &m->priority_groups, list) {
1314                         if (pg->bypassed)
1315                                 state = 'D';    /* Disabled */
1316                         else if (pg == m->current_pg)
1317                                 state = 'A';    /* Currently Active */
1318                         else
1319                                 state = 'E';    /* Enabled */
1320
1321                         DMEMIT("%c ", state);
1322
1323                         if (pg->ps.type->status)
1324                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1325                                                           result + sz,
1326                                                           maxlen - sz);
1327                         else
1328                                 DMEMIT("0 ");
1329
1330                         DMEMIT("%u %u ", pg->nr_pgpaths,
1331                                pg->ps.type->info_args);
1332
1333                         list_for_each_entry(p, &pg->pgpaths, list) {
1334                                 DMEMIT("%s %s %u ", p->path.dev->name,
1335                                        p->is_active ? "A" : "F",
1336                                        p->fail_count);
1337                                 if (pg->ps.type->status)
1338                                         sz += pg->ps.type->status(&pg->ps,
1339                                               &p->path, type, result + sz,
1340                                               maxlen - sz);
1341                         }
1342                 }
1343                 break;
1344
1345         case STATUSTYPE_TABLE:
1346                 list_for_each_entry(pg, &m->priority_groups, list) {
1347                         DMEMIT("%s ", pg->ps.type->name);
1348
1349                         if (pg->ps.type->status)
1350                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1351                                                           result + sz,
1352                                                           maxlen - sz);
1353                         else
1354                                 DMEMIT("0 ");
1355
1356                         DMEMIT("%u %u ", pg->nr_pgpaths,
1357                                pg->ps.type->table_args);
1358
1359                         list_for_each_entry(p, &pg->pgpaths, list) {
1360                                 DMEMIT("%s ", p->path.dev->name);
1361                                 if (pg->ps.type->status)
1362                                         sz += pg->ps.type->status(&pg->ps,
1363                                               &p->path, type, result + sz,
1364                                               maxlen - sz);
1365                         }
1366                 }
1367                 break;
1368         }
1369
1370         spin_unlock_irqrestore(&m->lock, flags);
1371
1372         return 0;
1373 }
1374
1375 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1376 {
1377         int r;
1378         struct dm_dev *dev;
1379         struct multipath *m = (struct multipath *) ti->private;
1380         action_fn action;
1381
1382         if (argc == 1) {
1383                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1384                         return queue_if_no_path(m, 1, 0);
1385                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1386                         return queue_if_no_path(m, 0, 0);
1387         }
1388
1389         if (argc != 2)
1390                 goto error;
1391
1392         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1393                 return bypass_pg_num(m, argv[1], 1);
1394         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1395                 return bypass_pg_num(m, argv[1], 0);
1396         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1397                 return switch_pg_num(m, argv[1]);
1398         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1399                 action = reinstate_path;
1400         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1401                 action = fail_path;
1402         else
1403                 goto error;
1404
1405         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1406                           dm_table_get_mode(ti->table), &dev);
1407         if (r) {
1408                 DMWARN("message: error getting device %s",
1409                        argv[1]);
1410                 return -EINVAL;
1411         }
1412
1413         r = action_dev(m, dev, action);
1414
1415         dm_put_device(ti, dev);
1416
1417         return r;
1418
1419 error:
1420         DMWARN("Unrecognised multipath message received.");
1421         return -EINVAL;
1422 }
1423
1424 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1425                            unsigned long arg)
1426 {
1427         struct multipath *m = (struct multipath *) ti->private;
1428         struct block_device *bdev = NULL;
1429         fmode_t mode = 0;
1430         unsigned long flags;
1431         int r = 0;
1432
1433         spin_lock_irqsave(&m->lock, flags);
1434
1435         if (!m->current_pgpath)
1436                 __choose_pgpath(m, 0);
1437
1438         if (m->current_pgpath) {
1439                 bdev = m->current_pgpath->path.dev->bdev;
1440                 mode = m->current_pgpath->path.dev->mode;
1441         }
1442
1443         if (m->queue_io)
1444                 r = -EAGAIN;
1445         else if (!bdev)
1446                 r = -EIO;
1447
1448         spin_unlock_irqrestore(&m->lock, flags);
1449
1450         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1451 }
1452
1453 /*-----------------------------------------------------------------
1454  * Module setup
1455  *---------------------------------------------------------------*/
1456 static struct target_type multipath_target = {
1457         .name = "multipath",
1458         .version = {1, 0, 5},
1459         .module = THIS_MODULE,
1460         .ctr = multipath_ctr,
1461         .dtr = multipath_dtr,
1462         .map = multipath_map,
1463         .end_io = multipath_end_io,
1464         .presuspend = multipath_presuspend,
1465         .resume = multipath_resume,
1466         .status = multipath_status,
1467         .message = multipath_message,
1468         .ioctl  = multipath_ioctl,
1469 };
1470
1471 static int __init dm_multipath_init(void)
1472 {
1473         int r;
1474
1475         /* allocate a slab for the dm_ios */
1476         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1477         if (!_mpio_cache)
1478                 return -ENOMEM;
1479
1480         r = dm_register_target(&multipath_target);
1481         if (r < 0) {
1482                 DMERR("register failed %d", r);
1483                 kmem_cache_destroy(_mpio_cache);
1484                 return -EINVAL;
1485         }
1486
1487         kmultipathd = create_workqueue("kmpathd");
1488         if (!kmultipathd) {
1489                 DMERR("failed to create workqueue kmpathd");
1490                 dm_unregister_target(&multipath_target);
1491                 kmem_cache_destroy(_mpio_cache);
1492                 return -ENOMEM;
1493         }
1494
1495         /*
1496          * A separate workqueue is used to handle the device handlers
1497          * to avoid overloading existing workqueue. Overloading the
1498          * old workqueue would also create a bottleneck in the
1499          * path of the storage hardware device activation.
1500          */
1501         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1502         if (!kmpath_handlerd) {
1503                 DMERR("failed to create workqueue kmpath_handlerd");
1504                 destroy_workqueue(kmultipathd);
1505                 dm_unregister_target(&multipath_target);
1506                 kmem_cache_destroy(_mpio_cache);
1507                 return -ENOMEM;
1508         }
1509
1510         DMINFO("version %u.%u.%u loaded",
1511                multipath_target.version[0], multipath_target.version[1],
1512                multipath_target.version[2]);
1513
1514         return r;
1515 }
1516
1517 static void __exit dm_multipath_exit(void)
1518 {
1519         destroy_workqueue(kmpath_handlerd);
1520         destroy_workqueue(kmultipathd);
1521
1522         dm_unregister_target(&multipath_target);
1523         kmem_cache_destroy(_mpio_cache);
1524 }
1525
1526 module_init(dm_multipath_init);
1527 module_exit(dm_multipath_exit);
1528
1529 MODULE_DESCRIPTION(DM_NAME " multipath target");
1530 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1531 MODULE_LICENSE("GPL");