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-uevent.h"
  12
  13 #include <linux/ctype.h>
  14 #include <linux/init.h>
  15 #include <linux/mempool.h>
  16 #include <linux/module.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/slab.h>
  19 #include <linux/time.h>
  20 #include <linux/workqueue.h>
  21 #include <scsi/scsi_dh.h>
  22 #include <asm/atomic.h>
  23
  24 #define DM_MSG_PREFIX "multipath"
  25 #define MESG_STR(x) x, sizeof(x)
  26
  27 /* Path properties */
  28 struct pgpath {
  29         struct list_head list;
  30
  31         struct priority_group *pg;      /* Owning PG */
  32         unsigned is_active;             /* Path status */
  33         unsigned fail_count;            /* Cumulative failure count */
  34
  35         struct dm_path path;
  36         struct work_struct deactivate_path;
  37         struct work_struct activate_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         char *hw_handler_params;
  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 list_head 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         struct mutex work_mutex;
  98
  99         unsigned suspended;     /* Don't create new I/O internally when set. */
 100 };
 101
 102 /*
 103  * Context information attached to each bio we process.
 104  */
 105 struct dm_mpath_io {
 106         struct pgpath *pgpath;
 107         size_t nr_bytes;
 108 };
 109
 110 typedef int (*action_fn) (struct pgpath *pgpath);
 111
 112 #define MIN_IOS 256     /* Mempool size */
 113
 114 static struct kmem_cache *_mpio_cache;
 115
 116 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 117 static void process_queued_ios(struct work_struct *work);
 118 static void trigger_event(struct work_struct *work);
 119 static void activate_path(struct work_struct *work);
 120 static void deactivate_path(struct work_struct *work);
 121
 122
 123 /*-----------------------------------------------
 124  * Allocation routines
 125  *-----------------------------------------------*/
 126
 127 static struct pgpath *alloc_pgpath(void)
 128 {
 129         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 130
 131         if (pgpath) {
 132                 pgpath->is_active = 1;
 133                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
 134                 INIT_WORK(&pgpath->activate_path, activate_path);
 135         }
 136
 137         return pgpath;
 138 }
 139
 140 static void free_pgpath(struct pgpath *pgpath)
 141 {
 142         kfree(pgpath);
 143 }
 144
 145 static void deactivate_path(struct work_struct *work)
 146 {
 147         struct pgpath *pgpath =
 148                 container_of(work, struct pgpath, deactivate_path);
 149
 150         blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
 151 }
 152
 153 static struct priority_group *alloc_priority_group(void)
 154 {
 155         struct priority_group *pg;
 156
 157         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 158
 159         if (pg)
 160                 INIT_LIST_HEAD(&pg->pgpaths);
 161
 162         return pg;
 163 }
 164
 165 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 166 {
 167         struct pgpath *pgpath, *tmp;
 168         struct multipath *m = ti->private;
 169
 170         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 171                 list_del(&pgpath->list);
 172                 if (m->hw_handler_name)
 173                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 174                 dm_put_device(ti, pgpath->path.dev);
 175                 free_pgpath(pgpath);
 176         }
 177 }
 178
 179 static void free_priority_group(struct priority_group *pg,
 180                                 struct dm_target *ti)
 181 {
 182         struct path_selector *ps = &pg->ps;
 183
 184         if (ps->type) {
 185                 ps->type->destroy(ps);
 186                 dm_put_path_selector(ps->type);
 187         }
 188
 189         free_pgpaths(&pg->pgpaths, ti);
 190         kfree(pg);
 191 }
 192
 193 static struct multipath *alloc_multipath(struct dm_target *ti)
 194 {
 195         struct multipath *m;
 196
 197         m = kzalloc(sizeof(*m), GFP_KERNEL);
 198         if (m) {
 199                 INIT_LIST_HEAD(&m->priority_groups);
 200                 INIT_LIST_HEAD(&m->queued_ios);
 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                 mutex_init(&m->work_mutex);
 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         kfree(m->hw_handler_params);
 229         mempool_destroy(m->mpio_pool);
 230         kfree(m);
 231 }
 232
 233
 234 /*-----------------------------------------------
 235  * Path selection
 236  *-----------------------------------------------*/
 237
 238 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 239 {
 240         m->current_pg = pgpath->pg;
 241
 242         /* Must we initialise the PG first, and queue I/O till it's ready? */
 243         if (m->hw_handler_name) {
 244                 m->pg_init_required = 1;
 245                 m->queue_io = 1;
 246         } else {
 247                 m->pg_init_required = 0;
 248                 m->queue_io = 0;
 249         }
 250
 251         m->pg_init_count = 0;
 252 }
 253
 254 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 255                                size_t nr_bytes)
 256 {
 257         struct dm_path *path;
 258
 259         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 260         if (!path)
 261                 return -ENXIO;
 262
 263         m->current_pgpath = path_to_pgpath(path);
 264
 265         if (m->current_pg != pg)
 266                 __switch_pg(m, m->current_pgpath);
 267
 268         return 0;
 269 }
 270
 271 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 272 {
 273         struct priority_group *pg;
 274         unsigned bypassed = 1;
 275
 276         if (!m->nr_valid_paths)
 277                 goto failed;
 278
 279         /* Were we instructed to switch PG? */
 280         if (m->next_pg) {
 281                 pg = m->next_pg;
 282                 m->next_pg = NULL;
 283                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 284                         return;
 285         }
 286
 287         /* Don't change PG until it has no remaining paths */
 288         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 289                 return;
 290
 291         /*
 292          * Loop through priority groups until we find a valid path.
 293          * First time we skip PGs marked 'bypassed'.
 294          * Second time we only try the ones we skipped.
 295          */
 296         do {
 297                 list_for_each_entry(pg, &m->priority_groups, list) {
 298                         if (pg->bypassed == bypassed)
 299                                 continue;
 300                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 301                                 return;
 302                 }
 303         } while (bypassed--);
 304
 305 failed:
 306         m->current_pgpath = NULL;
 307         m->current_pg = NULL;
 308 }
 309
 310 /*
 311  * Check whether bios must be queued in the device-mapper core rather
 312  * than here in the target.
 313  *
 314  * m->lock must be held on entry.
 315  *
 316  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 317  * same value then we are not between multipath_presuspend()
 318  * and multipath_resume() calls and we have no need to check
 319  * for the DMF_NOFLUSH_SUSPENDING flag.
 320  */
 321 static int __must_push_back(struct multipath *m)
 322 {
 323         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 324                 dm_noflush_suspending(m->ti));
 325 }
 326
 327 static int map_io(struct multipath *m, struct request *clone,
 328                   struct dm_mpath_io *mpio, unsigned was_queued)
 329 {
 330         int r = DM_MAPIO_REMAPPED;
 331         size_t nr_bytes = blk_rq_bytes(clone);
 332         unsigned long flags;
 333         struct pgpath *pgpath;
 334         struct block_device *bdev;
 335
 336         spin_lock_irqsave(&m->lock, flags);
 337
 338         /* Do we need to select a new pgpath? */
 339         if (!m->current_pgpath ||
 340             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 341                 __choose_pgpath(m, nr_bytes);
 342
 343         pgpath = m->current_pgpath;
 344
 345         if (was_queued)
 346                 m->queue_size--;
 347
 348         if ((pgpath && m->queue_io) ||
 349             (!pgpath && m->queue_if_no_path)) {
 350                 /* Queue for the daemon to resubmit */
 351                 list_add_tail(&clone->queuelist, &m->queued_ios);
 352                 m->queue_size++;
 353                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 354                     !m->queue_io)
 355                         queue_work(kmultipathd, &m->process_queued_ios);
 356                 pgpath = NULL;
 357                 r = DM_MAPIO_SUBMITTED;
 358         } else if (pgpath) {
 359                 bdev = pgpath->path.dev->bdev;
 360                 clone->q = bdev_get_queue(bdev);
 361                 clone->rq_disk = bdev->bd_disk;
 362         } else if (__must_push_back(m))
 363                 r = DM_MAPIO_REQUEUE;
 364         else
 365                 r = -EIO;       /* Failed */
 366
 367         mpio->pgpath = pgpath;
 368         mpio->nr_bytes = nr_bytes;
 369
 370         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 371                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 372                                               nr_bytes);
 373
 374         spin_unlock_irqrestore(&m->lock, flags);
 375
 376         return r;
 377 }
 378
 379 /*
 380  * If we run out of usable paths, should we queue I/O or error it?
 381  */
 382 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 383                             unsigned save_old_value)
 384 {
 385         unsigned long flags;
 386
 387         spin_lock_irqsave(&m->lock, flags);
 388
 389         if (save_old_value)
 390                 m->saved_queue_if_no_path = m->queue_if_no_path;
 391         else
 392                 m->saved_queue_if_no_path = queue_if_no_path;
 393         m->queue_if_no_path = queue_if_no_path;
 394         if (!m->queue_if_no_path && m->queue_size)
 395                 queue_work(kmultipathd, &m->process_queued_ios);
 396
 397         spin_unlock_irqrestore(&m->lock, flags);
 398
 399         return 0;
 400 }
 401
 402 /*-----------------------------------------------------------------
 403  * The multipath daemon is responsible for resubmitting queued ios.
 404  *---------------------------------------------------------------*/
 405
 406 static void dispatch_queued_ios(struct multipath *m)
 407 {
 408         int r;
 409         unsigned long flags;
 410         struct dm_mpath_io *mpio;
 411         union map_info *info;
 412         struct request *clone, *n;
 413         LIST_HEAD(cl);
 414
 415         spin_lock_irqsave(&m->lock, flags);
 416         list_splice_init(&m->queued_ios, &cl);
 417         spin_unlock_irqrestore(&m->lock, flags);
 418
 419         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 420                 list_del_init(&clone->queuelist);
 421
 422                 info = dm_get_rq_mapinfo(clone);
 423                 mpio = info->ptr;
 424
 425                 r = map_io(m, clone, mpio, 1);
 426                 if (r < 0) {
 427                         mempool_free(mpio, m->mpio_pool);
 428                         dm_kill_unmapped_request(clone, r);
 429                 } else if (r == DM_MAPIO_REMAPPED)
 430                         dm_dispatch_request(clone);
 431                 else if (r == DM_MAPIO_REQUEUE) {
 432                         mempool_free(mpio, m->mpio_pool);
 433                         dm_requeue_unmapped_request(clone);
 434                 }
 435         }
 436 }
 437
 438 static void process_queued_ios(struct work_struct *work)
 439 {
 440         struct multipath *m =
 441                 container_of(work, struct multipath, process_queued_ios);
 442         struct pgpath *pgpath = NULL, *tmp;
 443         unsigned must_queue = 1;
 444         unsigned long flags;
 445
 446         spin_lock_irqsave(&m->lock, flags);
 447
 448         if (!m->queue_size)
 449                 goto out;
 450
 451         if (!m->current_pgpath)
 452                 __choose_pgpath(m, 0);
 453
 454         pgpath = m->current_pgpath;
 455
 456         if ((pgpath && !m->queue_io) ||
 457             (!pgpath && !m->queue_if_no_path))
 458                 must_queue = 0;
 459
 460         if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
 461                 m->pg_init_count++;
 462                 m->pg_init_required = 0;
 463                 list_for_each_entry(tmp, &pgpath->pg->pgpaths, list) {
 464                         /* Skip failed paths */
 465                         if (!tmp->is_active)
 466                                 continue;
 467                         if (queue_work(kmpath_handlerd, &tmp->activate_path))
 468                                 m->pg_init_in_progress++;
 469                 }
 470         }
 471 out:
 472         spin_unlock_irqrestore(&m->lock, flags);
 473         if (!must_queue)
 474                 dispatch_queued_ios(m);
 475 }
 476
 477 /*
 478  * An event is triggered whenever a path is taken out of use.
 479  * Includes path failure and PG bypass.
 480  */
 481 static void trigger_event(struct work_struct *work)
 482 {
 483         struct multipath *m =
 484                 container_of(work, struct multipath, trigger_event);
 485
 486         dm_table_event(m->ti->table);
 487 }
 488
 489 /*-----------------------------------------------------------------
 490  * Constructor/argument parsing:
 491  * <#multipath feature args> [<arg>]*
 492  * <#hw_handler args> [hw_handler [<arg>]*]
 493  * <#priority groups>
 494  * <initial priority group>
 495  *     [<selector> <#selector args> [<arg>]*
 496  *      <#paths> <#per-path selector args>
 497  *         [<path> [<arg>]* ]+ ]+
 498  *---------------------------------------------------------------*/
 499 struct param {
 500         unsigned min;
 501         unsigned max;
 502         char *error;
 503 };
 504
 505 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 506 {
 507         if (!str ||
 508             (sscanf(str, "%u", v) != 1) ||
 509             (*v < param->min) ||
 510             (*v > param->max)) {
 511                 *error = param->error;
 512                 return -EINVAL;
 513         }
 514
 515         return 0;
 516 }
 517
 518 struct arg_set {
 519         unsigned argc;
 520         char **argv;
 521 };
 522
 523 static char *shift(struct arg_set *as)
 524 {
 525         char *r;
 526
 527         if (as->argc) {
 528                 as->argc--;
 529                 r = *as->argv;
 530                 as->argv++;
 531                 return r;
 532         }
 533
 534         return NULL;
 535 }
 536
 537 static void consume(struct arg_set *as, unsigned n)
 538 {
 539         BUG_ON (as->argc < n);
 540         as->argc -= n;
 541         as->argv += n;
 542 }
 543
 544 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 545                                struct dm_target *ti)
 546 {
 547         int r;
 548         struct path_selector_type *pst;
 549         unsigned ps_argc;
 550
 551         static struct param _params[] = {
 552                 {0, 1024, "invalid number of path selector args"},
 553         };
 554
 555         pst = dm_get_path_selector(shift(as));
 556         if (!pst) {
 557                 ti->error = "unknown path selector type";
 558                 return -EINVAL;
 559         }
 560
 561         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 562         if (r) {
 563                 dm_put_path_selector(pst);
 564                 return -EINVAL;
 565         }
 566
 567         if (ps_argc > as->argc) {
 568                 dm_put_path_selector(pst);
 569                 ti->error = "not enough arguments for path selector";
 570                 return -EINVAL;
 571         }
 572
 573         r = pst->create(&pg->ps, ps_argc, as->argv);
 574         if (r) {
 575                 dm_put_path_selector(pst);
 576                 ti->error = "path selector constructor failed";
 577                 return r;
 578         }
 579
 580         pg->ps.type = pst;
 581         consume(as, ps_argc);
 582
 583         return 0;
 584 }
 585
 586 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 587                                struct dm_target *ti)
 588 {
 589         int r;
 590         struct pgpath *p;
 591         struct multipath *m = ti->private;
 592
 593         /* we need at least a path arg */
 594         if (as->argc < 1) {
 595                 ti->error = "no device given";
 596                 return ERR_PTR(-EINVAL);
 597         }
 598
 599         p = alloc_pgpath();
 600         if (!p)
 601                 return ERR_PTR(-ENOMEM);
 602
 603         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 604                           dm_table_get_mode(ti->table), &p->path.dev);
 605         if (r) {
 606                 ti->error = "error getting device";
 607                 goto bad;
 608         }
 609
 610         if (m->hw_handler_name) {
 611                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 612
 613                 r = scsi_dh_attach(q, m->hw_handler_name);
 614                 if (r == -EBUSY) {
 615                         /*
 616                          * Already attached to different hw_handler,
 617                          * try to reattach with correct one.
 618                          */
 619                         scsi_dh_detach(q);
 620                         r = scsi_dh_attach(q, m->hw_handler_name);
 621                 }
 622
 623                 if (r < 0) {
 624                         ti->error = "error attaching hardware handler";
 625                         dm_put_device(ti, p->path.dev);
 626                         goto bad;
 627                 }
 628
 629                 if (m->hw_handler_params) {
 630                         r = scsi_dh_set_params(q, m->hw_handler_params);
 631                         if (r < 0) {
 632                                 ti->error = "unable to set hardware "
 633                                                         "handler parameters";
 634                                 scsi_dh_detach(q);
 635                                 dm_put_device(ti, p->path.dev);
 636                                 goto bad;
 637                         }
 638                 }
 639         }
 640
 641         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 642         if (r) {
 643                 dm_put_device(ti, p->path.dev);
 644                 goto bad;
 645         }
 646
 647         return p;
 648
 649  bad:
 650         free_pgpath(p);
 651         return ERR_PTR(r);
 652 }
 653
 654 static struct priority_group *parse_priority_group(struct arg_set *as,
 655                                                    struct multipath *m)
 656 {
 657         static struct param _params[] = {
 658                 {1, 1024, "invalid number of paths"},
 659                 {0, 1024, "invalid number of selector args"}
 660         };
 661
 662         int r;
 663         unsigned i, nr_selector_args, nr_params;
 664         struct priority_group *pg;
 665         struct dm_target *ti = m->ti;
 666
 667         if (as->argc < 2) {
 668                 as->argc = 0;
 669                 ti->error = "not enough priority group arguments";
 670                 return ERR_PTR(-EINVAL);
 671         }
 672
 673         pg = alloc_priority_group();
 674         if (!pg) {
 675                 ti->error = "couldn't allocate priority group";
 676                 return ERR_PTR(-ENOMEM);
 677         }
 678         pg->m = m;
 679
 680         r = parse_path_selector(as, pg, ti);
 681         if (r)
 682                 goto bad;
 683
 684         /*
 685          * read the paths
 686          */
 687         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 688         if (r)
 689                 goto bad;
 690
 691         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 692         if (r)
 693                 goto bad;
 694
 695         nr_params = 1 + nr_selector_args;
 696         for (i = 0; i < pg->nr_pgpaths; i++) {
 697                 struct pgpath *pgpath;
 698                 struct arg_set path_args;
 699
 700                 if (as->argc < nr_params) {
 701                         ti->error = "not enough path parameters";
 702                         goto bad;
 703                 }
 704
 705                 path_args.argc = nr_params;
 706                 path_args.argv = as->argv;
 707
 708                 pgpath = parse_path(&path_args, &pg->ps, ti);
 709                 if (IS_ERR(pgpath)) {
 710                         r = PTR_ERR(pgpath);
 711                         goto bad;
 712                 }
 713
 714                 pgpath->pg = pg;
 715                 list_add_tail(&pgpath->list, &pg->pgpaths);
 716                 consume(as, nr_params);
 717         }
 718
 719         return pg;
 720
 721  bad:
 722         free_priority_group(pg, ti);
 723         return ERR_PTR(r);
 724 }
 725
 726 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 727 {
 728         unsigned hw_argc;
 729         int ret;
 730         struct dm_target *ti = m->ti;
 731
 732         static struct param _params[] = {
 733                 {0, 1024, "invalid number of hardware handler args"},
 734         };
 735
 736         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 737                 return -EINVAL;
 738
 739         if (!hw_argc)
 740                 return 0;
 741
 742         if (hw_argc > as->argc) {
 743                 ti->error = "not enough arguments for hardware handler";
 744                 return -EINVAL;
 745         }
 746
 747         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 748         request_module("scsi_dh_%s", m->hw_handler_name);
 749         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 750                 ti->error = "unknown hardware handler type";
 751                 ret = -EINVAL;
 752                 goto fail;
 753         }
 754
 755         if (hw_argc > 1) {
 756                 char *p;
 757                 int i, j, len = 4;
 758
 759                 for (i = 0; i <= hw_argc - 2; i++)
 760                         len += strlen(as->argv[i]) + 1;
 761                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 762                 if (!p) {
 763                         ti->error = "memory allocation failed";
 764                         ret = -ENOMEM;
 765                         goto fail;
 766                 }
 767                 j = sprintf(p, "%d", hw_argc - 1);
 768                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 769                         j = sprintf(p, "%s", as->argv[i]);
 770         }
 771         consume(as, hw_argc - 1);
 772
 773         return 0;
 774 fail:
 775         kfree(m->hw_handler_name);
 776         m->hw_handler_name = NULL;
 777         return ret;
 778 }
 779
 780 static int parse_features(struct arg_set *as, struct multipath *m)
 781 {
 782         int r;
 783         unsigned argc;
 784         struct dm_target *ti = m->ti;
 785         const char *param_name;
 786
 787         static struct param _params[] = {
 788                 {0, 3, "invalid number of feature args"},
 789                 {1, 50, "pg_init_retries must be between 1 and 50"},
 790         };
 791
 792         r = read_param(_params, shift(as), &argc, &ti->error);
 793         if (r)
 794                 return -EINVAL;
 795
 796         if (!argc)
 797                 return 0;
 798
 799         do {
 800                 param_name = shift(as);
 801                 argc--;
 802
 803                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 804                         r = queue_if_no_path(m, 1, 0);
 805                         continue;
 806                 }
 807
 808                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 809                     (argc >= 1)) {
 810                         r = read_param(_params + 1, shift(as),
 811                                        &m->pg_init_retries, &ti->error);
 812                         argc--;
 813                         continue;
 814                 }
 815
 816                 ti->error = "Unrecognised multipath feature request";
 817                 r = -EINVAL;
 818         } while (argc && !r);
 819
 820         return r;
 821 }
 822
 823 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 824                          char **argv)
 825 {
 826         /* target parameters */
 827         static struct param _params[] = {
 828                 {1, 1024, "invalid number of priority groups"},
 829                 {1, 1024, "invalid initial priority group number"},
 830         };
 831
 832         int r;
 833         struct multipath *m;
 834         struct arg_set as;
 835         unsigned pg_count = 0;
 836         unsigned next_pg_num;
 837
 838         as.argc = argc;
 839         as.argv = argv;
 840
 841         m = alloc_multipath(ti);
 842         if (!m) {
 843                 ti->error = "can't allocate multipath";
 844                 return -EINVAL;
 845         }
 846
 847         r = parse_features(&as, m);
 848         if (r)
 849                 goto bad;
 850
 851         r = parse_hw_handler(&as, m);
 852         if (r)
 853                 goto bad;
 854
 855         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 856         if (r)
 857                 goto bad;
 858
 859         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 860         if (r)
 861                 goto bad;
 862
 863         /* parse the priority groups */
 864         while (as.argc) {
 865                 struct priority_group *pg;
 866
 867                 pg = parse_priority_group(&as, m);
 868                 if (IS_ERR(pg)) {
 869                         r = PTR_ERR(pg);
 870                         goto bad;
 871                 }
 872
 873                 m->nr_valid_paths += pg->nr_pgpaths;
 874                 list_add_tail(&pg->list, &m->priority_groups);
 875                 pg_count++;
 876                 pg->pg_num = pg_count;
 877                 if (!--next_pg_num)
 878                         m->next_pg = pg;
 879         }
 880
 881         if (pg_count != m->nr_priority_groups) {
 882                 ti->error = "priority group count mismatch";
 883                 r = -EINVAL;
 884                 goto bad;
 885         }
 886
 887         ti->num_flush_requests = 1;
 888
 889         return 0;
 890
 891  bad:
 892         free_multipath(m);
 893         return r;
 894 }
 895
 896 static void flush_multipath_work(void)
 897 {
 898         flush_workqueue(kmpath_handlerd);
 899         flush_workqueue(kmultipathd);
 900         flush_scheduled_work();
 901 }
 902
 903 static void multipath_dtr(struct dm_target *ti)
 904 {
 905         struct multipath *m = ti->private;
 906
 907         flush_multipath_work();
 908         free_multipath(m);
 909 }
 910
 911 /*
 912  * Map cloned requests
 913  */
 914 static int multipath_map(struct dm_target *ti, struct request *clone,
 915                          union map_info *map_context)
 916 {
 917         int r;
 918         struct dm_mpath_io *mpio;
 919         struct multipath *m = (struct multipath *) ti->private;
 920
 921         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 922         if (!mpio)
 923                 /* ENOMEM, requeue */
 924                 return DM_MAPIO_REQUEUE;
 925         memset(mpio, 0, sizeof(*mpio));
 926
 927         map_context->ptr = mpio;
 928         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 929         r = map_io(m, clone, mpio, 0);
 930         if (r < 0 || r == DM_MAPIO_REQUEUE)
 931                 mempool_free(mpio, m->mpio_pool);
 932
 933         return r;
 934 }
 935
 936 /*
 937  * Take a path out of use.
 938  */
 939 static int fail_path(struct pgpath *pgpath)
 940 {
 941         unsigned long flags;
 942         struct multipath *m = pgpath->pg->m;
 943
 944         spin_lock_irqsave(&m->lock, flags);
 945
 946         if (!pgpath->is_active)
 947                 goto out;
 948
 949         DMWARN("Failing path %s.", pgpath->path.dev->name);
 950
 951         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 952         pgpath->is_active = 0;
 953         pgpath->fail_count++;
 954
 955         m->nr_valid_paths--;
 956
 957         if (pgpath == m->current_pgpath)
 958                 m->current_pgpath = NULL;
 959
 960         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 961                       pgpath->path.dev->name, m->nr_valid_paths);
 962
 963         schedule_work(&m->trigger_event);
 964         queue_work(kmultipathd, &pgpath->deactivate_path);
 965
 966 out:
 967         spin_unlock_irqrestore(&m->lock, flags);
 968
 969         return 0;
 970 }
 971
 972 /*
 973  * Reinstate a previously-failed path
 974  */
 975 static int reinstate_path(struct pgpath *pgpath)
 976 {
 977         int r = 0;
 978         unsigned long flags;
 979         struct multipath *m = pgpath->pg->m;
 980
 981         spin_lock_irqsave(&m->lock, flags);
 982
 983         if (pgpath->is_active)
 984                 goto out;
 985
 986         if (!pgpath->pg->ps.type->reinstate_path) {
 987                 DMWARN("Reinstate path not supported by path selector %s",
 988                        pgpath->pg->ps.type->name);
 989                 r = -EINVAL;
 990                 goto out;
 991         }
 992
 993         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 994         if (r)
 995                 goto out;
 996
 997         pgpath->is_active = 1;
 998
 999         if (!m->nr_valid_paths++ && m->queue_size) {
1000                 m->current_pgpath = NULL;
1001                 queue_work(kmultipathd, &m->process_queued_ios);
1002         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1003                 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
1004                         m->pg_init_in_progress++;
1005         }
1006
1007         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1008                       pgpath->path.dev->name, m->nr_valid_paths);
1009
1010         schedule_work(&m->trigger_event);
1011
1012 out:
1013         spin_unlock_irqrestore(&m->lock, flags);
1014
1015         return r;
1016 }
1017
1018 /*
1019  * Fail or reinstate all paths that match the provided struct dm_dev.
1020  */
1021 static int action_dev(struct multipath *m, struct dm_dev *dev,
1022                       action_fn action)
1023 {
1024         int r = 0;
1025         struct pgpath *pgpath;
1026         struct priority_group *pg;
1027
1028         list_for_each_entry(pg, &m->priority_groups, list) {
1029                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1030                         if (pgpath->path.dev == dev)
1031                                 r = action(pgpath);
1032                 }
1033         }
1034
1035         return r;
1036 }
1037
1038 /*
1039  * Temporarily try to avoid having to use the specified PG
1040  */
1041 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1042                       int bypassed)
1043 {
1044         unsigned long flags;
1045
1046         spin_lock_irqsave(&m->lock, flags);
1047
1048         pg->bypassed = bypassed;
1049         m->current_pgpath = NULL;
1050         m->current_pg = NULL;
1051
1052         spin_unlock_irqrestore(&m->lock, flags);
1053
1054         schedule_work(&m->trigger_event);
1055 }
1056
1057 /*
1058  * Switch to using the specified PG from the next I/O that gets mapped
1059  */
1060 static int switch_pg_num(struct multipath *m, const char *pgstr)
1061 {
1062         struct priority_group *pg;
1063         unsigned pgnum;
1064         unsigned long flags;
1065
1066         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1067             (pgnum > m->nr_priority_groups)) {
1068                 DMWARN("invalid PG number supplied to switch_pg_num");
1069                 return -EINVAL;
1070         }
1071
1072         spin_lock_irqsave(&m->lock, flags);
1073         list_for_each_entry(pg, &m->priority_groups, list) {
1074                 pg->bypassed = 0;
1075                 if (--pgnum)
1076                         continue;
1077
1078                 m->current_pgpath = NULL;
1079                 m->current_pg = NULL;
1080                 m->next_pg = pg;
1081         }
1082         spin_unlock_irqrestore(&m->lock, flags);
1083
1084         schedule_work(&m->trigger_event);
1085         return 0;
1086 }
1087
1088 /*
1089  * Set/clear bypassed status of a PG.
1090  * PGs are numbered upwards from 1 in the order they were declared.
1091  */
1092 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1093 {
1094         struct priority_group *pg;
1095         unsigned pgnum;
1096
1097         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1098             (pgnum > m->nr_priority_groups)) {
1099                 DMWARN("invalid PG number supplied to bypass_pg");
1100                 return -EINVAL;
1101         }
1102
1103         list_for_each_entry(pg, &m->priority_groups, list) {
1104                 if (!--pgnum)
1105                         break;
1106         }
1107
1108         bypass_pg(m, pg, bypassed);
1109         return 0;
1110 }
1111
1112 /*
1113  * Should we retry pg_init immediately?
1114  */
1115 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1116 {
1117         unsigned long flags;
1118         int limit_reached = 0;
1119
1120         spin_lock_irqsave(&m->lock, flags);
1121
1122         if (m->pg_init_count <= m->pg_init_retries)
1123                 m->pg_init_required = 1;
1124         else
1125                 limit_reached = 1;
1126
1127         spin_unlock_irqrestore(&m->lock, flags);
1128
1129         return limit_reached;
1130 }
1131
1132 static void pg_init_done(void *data, int errors)
1133 {
1134         struct pgpath *pgpath = data;
1135         struct priority_group *pg = pgpath->pg;
1136         struct multipath *m = pg->m;
1137         unsigned long flags;
1138
1139         /* device or driver problems */
1140         switch (errors) {
1141         case SCSI_DH_OK:
1142                 break;
1143         case SCSI_DH_NOSYS:
1144                 if (!m->hw_handler_name) {
1145                         errors = 0;
1146                         break;
1147                 }
1148                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1149                       "Error %d.", m->hw_handler_name, errors);
1150                 /*
1151                  * Fail path for now, so we do not ping pong
1152                  */
1153                 fail_path(pgpath);
1154                 break;
1155         case SCSI_DH_DEV_TEMP_BUSY:
1156                 /*
1157                  * Probably doing something like FW upgrade on the
1158                  * controller so try the other pg.
1159                  */
1160                 bypass_pg(m, pg, 1);
1161                 break;
1162         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1163         case SCSI_DH_RETRY:
1164         case SCSI_DH_IMM_RETRY:
1165         case SCSI_DH_RES_TEMP_UNAVAIL:
1166                 if (pg_init_limit_reached(m, pgpath))
1167                         fail_path(pgpath);
1168                 errors = 0;
1169                 break;
1170         default:
1171                 /*
1172                  * We probably do not want to fail the path for a device
1173                  * error, but this is what the old dm did. In future
1174                  * patches we can do more advanced handling.
1175                  */
1176                 fail_path(pgpath);
1177         }
1178
1179         spin_lock_irqsave(&m->lock, flags);
1180         if (errors) {
1181                 if (pgpath == m->current_pgpath) {
1182                         DMERR("Could not failover device. Error %d.", errors);
1183                         m->current_pgpath = NULL;
1184                         m->current_pg = NULL;
1185                 }
1186         } else if (!m->pg_init_required) {
1187                 m->queue_io = 0;
1188                 pg->bypassed = 0;
1189         }
1190
1191         m->pg_init_in_progress--;
1192         if (!m->pg_init_in_progress)
1193                 queue_work(kmultipathd, &m->process_queued_ios);
1194         spin_unlock_irqrestore(&m->lock, flags);
1195 }
1196
1197 static void activate_path(struct work_struct *work)
1198 {
1199         struct pgpath *pgpath =
1200                 container_of(work, struct pgpath, activate_path);
1201
1202         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1203                                 pg_init_done, pgpath);
1204 }
1205
1206 /*
1207  * end_io handling
1208  */
1209 static int do_end_io(struct multipath *m, struct request *clone,
1210                      int error, struct dm_mpath_io *mpio)
1211 {
1212         /*
1213          * We don't queue any clone request inside the multipath target
1214          * during end I/O handling, since those clone requests don't have
1215          * bio clones.  If we queue them inside the multipath target,
1216          * we need to make bio clones, that requires memory allocation.
1217          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1218          *  don't have bio clones.)
1219          * Instead of queueing the clone request here, we queue the original
1220          * request into dm core, which will remake a clone request and
1221          * clone bios for it and resubmit it later.
1222          */
1223         int r = DM_ENDIO_REQUEUE;
1224         unsigned long flags;
1225
1226         if (!error && !clone->errors)
1227                 return 0;       /* I/O complete */
1228
1229         if (error == -EOPNOTSUPP)
1230                 return error;
1231
1232         if (mpio->pgpath)
1233                 fail_path(mpio->pgpath);
1234
1235         spin_lock_irqsave(&m->lock, flags);
1236         if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1237                 r = -EIO;
1238         spin_unlock_irqrestore(&m->lock, flags);
1239
1240         return r;
1241 }
1242
1243 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1244                             int error, union map_info *map_context)
1245 {
1246         struct multipath *m = ti->private;
1247         struct dm_mpath_io *mpio = map_context->ptr;
1248         struct pgpath *pgpath = mpio->pgpath;
1249         struct path_selector *ps;
1250         int r;
1251
1252         r  = do_end_io(m, clone, error, mpio);
1253         if (pgpath) {
1254                 ps = &pgpath->pg->ps;
1255                 if (ps->type->end_io)
1256                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1257         }
1258         mempool_free(mpio, m->mpio_pool);
1259
1260         return r;
1261 }
1262
1263 /*
1264  * Suspend can't complete until all the I/O is processed so if
1265  * the last path fails we must error any remaining I/O.
1266  * Note that if the freeze_bdev fails while suspending, the
1267  * queue_if_no_path state is lost - userspace should reset it.
1268  */
1269 static void multipath_presuspend(struct dm_target *ti)
1270 {
1271         struct multipath *m = (struct multipath *) ti->private;
1272
1273         queue_if_no_path(m, 0, 1);
1274 }
1275
1276 static void multipath_postsuspend(struct dm_target *ti)
1277 {
1278         struct multipath *m = ti->private;
1279
1280         mutex_lock(&m->work_mutex);
1281         m->suspended = 1;
1282         flush_multipath_work();
1283         mutex_unlock(&m->work_mutex);
1284 }
1285
1286 /*
1287  * Restore the queue_if_no_path setting.
1288  */
1289 static void multipath_resume(struct dm_target *ti)
1290 {
1291         struct multipath *m = (struct multipath *) ti->private;
1292         unsigned long flags;
1293
1294         mutex_lock(&m->work_mutex);
1295         m->suspended = 0;
1296         mutex_unlock(&m->work_mutex);
1297
1298         spin_lock_irqsave(&m->lock, flags);
1299         m->queue_if_no_path = m->saved_queue_if_no_path;
1300         spin_unlock_irqrestore(&m->lock, flags);
1301 }
1302
1303 /*
1304  * Info output has the following format:
1305  * num_multipath_feature_args [multipath_feature_args]*
1306  * num_handler_status_args [handler_status_args]*
1307  * num_groups init_group_number
1308  *            [A|D|E num_ps_status_args [ps_status_args]*
1309  *             num_paths num_selector_args
1310  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1311  *
1312  * Table output has the following format (identical to the constructor string):
1313  * num_feature_args [features_args]*
1314  * num_handler_args hw_handler [hw_handler_args]*
1315  * num_groups init_group_number
1316  *     [priority selector-name num_ps_args [ps_args]*
1317  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1318  */
1319 static int multipath_status(struct dm_target *ti, status_type_t type,
1320                             char *result, unsigned int maxlen)
1321 {
1322         int sz = 0;
1323         unsigned long flags;
1324         struct multipath *m = (struct multipath *) ti->private;
1325         struct priority_group *pg;
1326         struct pgpath *p;
1327         unsigned pg_num;
1328         char state;
1329
1330         spin_lock_irqsave(&m->lock, flags);
1331
1332         /* Features */
1333         if (type == STATUSTYPE_INFO)
1334                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1335         else {
1336                 DMEMIT("%u ", m->queue_if_no_path +
1337                               (m->pg_init_retries > 0) * 2);
1338                 if (m->queue_if_no_path)
1339                         DMEMIT("queue_if_no_path ");
1340                 if (m->pg_init_retries)
1341                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1342         }
1343
1344         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1345                 DMEMIT("0 ");
1346         else
1347                 DMEMIT("1 %s ", m->hw_handler_name);
1348
1349         DMEMIT("%u ", m->nr_priority_groups);
1350
1351         if (m->next_pg)
1352                 pg_num = m->next_pg->pg_num;
1353         else if (m->current_pg)
1354                 pg_num = m->current_pg->pg_num;
1355         else
1356                         pg_num = 1;
1357
1358         DMEMIT("%u ", pg_num);
1359
1360         switch (type) {
1361         case STATUSTYPE_INFO:
1362                 list_for_each_entry(pg, &m->priority_groups, list) {
1363                         if (pg->bypassed)
1364                                 state = 'D';    /* Disabled */
1365                         else if (pg == m->current_pg)
1366                                 state = 'A';    /* Currently Active */
1367                         else
1368                                 state = 'E';    /* Enabled */
1369
1370                         DMEMIT("%c ", state);
1371
1372                         if (pg->ps.type->status)
1373                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1374                                                           result + sz,
1375                                                           maxlen - sz);
1376                         else
1377                                 DMEMIT("0 ");
1378
1379                         DMEMIT("%u %u ", pg->nr_pgpaths,
1380                                pg->ps.type->info_args);
1381
1382                         list_for_each_entry(p, &pg->pgpaths, list) {
1383                                 DMEMIT("%s %s %u ", p->path.dev->name,
1384                                        p->is_active ? "A" : "F",
1385                                        p->fail_count);
1386                                 if (pg->ps.type->status)
1387                                         sz += pg->ps.type->status(&pg->ps,
1388                                               &p->path, type, result + sz,
1389                                               maxlen - sz);
1390                         }
1391                 }
1392                 break;
1393
1394         case STATUSTYPE_TABLE:
1395                 list_for_each_entry(pg, &m->priority_groups, list) {
1396                         DMEMIT("%s ", pg->ps.type->name);
1397
1398                         if (pg->ps.type->status)
1399                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1400                                                           result + sz,
1401                                                           maxlen - sz);
1402                         else
1403                                 DMEMIT("0 ");
1404
1405                         DMEMIT("%u %u ", pg->nr_pgpaths,
1406                                pg->ps.type->table_args);
1407
1408                         list_for_each_entry(p, &pg->pgpaths, list) {
1409                                 DMEMIT("%s ", p->path.dev->name);
1410                                 if (pg->ps.type->status)
1411                                         sz += pg->ps.type->status(&pg->ps,
1412                                               &p->path, type, result + sz,
1413                                               maxlen - sz);
1414                         }
1415                 }
1416                 break;
1417         }
1418
1419         spin_unlock_irqrestore(&m->lock, flags);
1420
1421         return 0;
1422 }
1423
1424 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1425 {
1426         int r = -EINVAL;
1427         struct dm_dev *dev;
1428         struct multipath *m = (struct multipath *) ti->private;
1429         action_fn action;
1430
1431         mutex_lock(&m->work_mutex);
1432
1433         if (m->suspended) {
1434                 r = -EBUSY;
1435                 goto out;
1436         }
1437
1438         if (dm_suspended(ti)) {
1439                 r = -EBUSY;
1440                 goto out;
1441         }
1442
1443         if (argc == 1) {
1444                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
1445                         r = queue_if_no_path(m, 1, 0);
1446                         goto out;
1447                 } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
1448                         r = queue_if_no_path(m, 0, 0);
1449                         goto out;
1450                 }
1451         }
1452
1453         if (argc != 2) {
1454                 DMWARN("Unrecognised multipath message received.");
1455                 goto out;
1456         }
1457
1458         if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
1459                 r = bypass_pg_num(m, argv[1], 1);
1460                 goto out;
1461         } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
1462                 r = bypass_pg_num(m, argv[1], 0);
1463                 goto out;
1464         } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
1465                 r = switch_pg_num(m, argv[1]);
1466                 goto out;
1467         } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1468                 action = reinstate_path;
1469         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1470                 action = fail_path;
1471         else {
1472                 DMWARN("Unrecognised multipath message received.");
1473                 goto out;
1474         }
1475
1476         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1477                           dm_table_get_mode(ti->table), &dev);
1478         if (r) {
1479                 DMWARN("message: error getting device %s",
1480                        argv[1]);
1481                 goto out;
1482         }
1483
1484         r = action_dev(m, dev, action);
1485
1486         dm_put_device(ti, dev);
1487
1488 out:
1489         mutex_unlock(&m->work_mutex);
1490         return r;
1491 }
1492
1493 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1494                            unsigned long arg)
1495 {
1496         struct multipath *m = (struct multipath *) ti->private;
1497         struct block_device *bdev = NULL;
1498         fmode_t mode = 0;
1499         unsigned long flags;
1500         int r = 0;
1501
1502         spin_lock_irqsave(&m->lock, flags);
1503
1504         if (!m->current_pgpath)
1505                 __choose_pgpath(m, 0);
1506
1507         if (m->current_pgpath) {
1508                 bdev = m->current_pgpath->path.dev->bdev;
1509                 mode = m->current_pgpath->path.dev->mode;
1510         }
1511
1512         if (m->queue_io)
1513                 r = -EAGAIN;
1514         else if (!bdev)
1515                 r = -EIO;
1516
1517         spin_unlock_irqrestore(&m->lock, flags);
1518
1519         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1520 }
1521
1522 static int multipath_iterate_devices(struct dm_target *ti,
1523                                      iterate_devices_callout_fn fn, void *data)
1524 {
1525         struct multipath *m = ti->private;
1526         struct priority_group *pg;
1527         struct pgpath *p;
1528         int ret = 0;
1529
1530         list_for_each_entry(pg, &m->priority_groups, list) {
1531                 list_for_each_entry(p, &pg->pgpaths, list) {
1532                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1533                         if (ret)
1534                                 goto out;
1535                 }
1536         }
1537
1538 out:
1539         return ret;
1540 }
1541
1542 static int __pgpath_busy(struct pgpath *pgpath)
1543 {
1544         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1545
1546         return dm_underlying_device_busy(q);
1547 }
1548
1549 /*
1550  * We return "busy", only when we can map I/Os but underlying devices
1551  * are busy (so even if we map I/Os now, the I/Os will wait on
1552  * the underlying queue).
1553  * In other words, if we want to kill I/Os or queue them inside us
1554  * due to map unavailability, we don't return "busy".  Otherwise,
1555  * dm core won't give us the I/Os and we can't do what we want.
1556  */
1557 static int multipath_busy(struct dm_target *ti)
1558 {
1559         int busy = 0, has_active = 0;
1560         struct multipath *m = ti->private;
1561         struct priority_group *pg;
1562         struct pgpath *pgpath;
1563         unsigned long flags;
1564
1565         spin_lock_irqsave(&m->lock, flags);
1566
1567         /* Guess which priority_group will be used at next mapping time */
1568         if (unlikely(!m->current_pgpath && m->next_pg))
1569                 pg = m->next_pg;
1570         else if (likely(m->current_pg))
1571                 pg = m->current_pg;
1572         else
1573                 /*
1574                  * We don't know which pg will be used at next mapping time.
1575                  * We don't call __choose_pgpath() here to avoid to trigger
1576                  * pg_init just by busy checking.
1577                  * So we don't know whether underlying devices we will be using
1578                  * at next mapping time are busy or not. Just try mapping.
1579                  */
1580                 goto out;
1581
1582         /*
1583          * If there is one non-busy active path at least, the path selector
1584          * will be able to select it. So we consider such a pg as not busy.
1585          */
1586         busy = 1;
1587         list_for_each_entry(pgpath, &pg->pgpaths, list)
1588                 if (pgpath->is_active) {
1589                         has_active = 1;
1590
1591                         if (!__pgpath_busy(pgpath)) {
1592                                 busy = 0;
1593                                 break;
1594                         }
1595                 }
1596
1597         if (!has_active)
1598                 /*
1599                  * No active path in this pg, so this pg won't be used and
1600                  * the current_pg will be changed at next mapping time.
1601                  * We need to try mapping to determine it.
1602                  */
1603                 busy = 0;
1604
1605 out:
1606         spin_unlock_irqrestore(&m->lock, flags);
1607
1608         return busy;
1609 }
1610
1611 /*-----------------------------------------------------------------
1612  * Module setup
1613  *---------------------------------------------------------------*/
1614 static struct target_type multipath_target = {
1615         .name = "multipath",
1616         .version = {1, 1, 1},
1617         .module = THIS_MODULE,
1618         .ctr = multipath_ctr,
1619         .dtr = multipath_dtr,
1620         .map_rq = multipath_map,
1621         .rq_end_io = multipath_end_io,
1622         .presuspend = multipath_presuspend,
1623         .postsuspend = multipath_postsuspend,
1624         .resume = multipath_resume,
1625         .status = multipath_status,
1626         .message = multipath_message,
1627         .ioctl  = multipath_ioctl,
1628         .iterate_devices = multipath_iterate_devices,
1629         .busy = multipath_busy,
1630 };
1631
1632 static int __init dm_multipath_init(void)
1633 {
1634         int r;
1635
1636         /* allocate a slab for the dm_ios */
1637         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1638         if (!_mpio_cache)
1639                 return -ENOMEM;
1640
1641         r = dm_register_target(&multipath_target);
1642         if (r < 0) {
1643                 DMERR("register failed %d", r);
1644                 kmem_cache_destroy(_mpio_cache);
1645                 return -EINVAL;
1646         }
1647
1648         kmultipathd = create_workqueue("kmpathd");
1649         if (!kmultipathd) {
1650                 DMERR("failed to create workqueue kmpathd");
1651                 dm_unregister_target(&multipath_target);
1652                 kmem_cache_destroy(_mpio_cache);
1653                 return -ENOMEM;
1654         }
1655
1656         /*
1657          * A separate workqueue is used to handle the device handlers
1658          * to avoid overloading existing workqueue. Overloading the
1659          * old workqueue would also create a bottleneck in the
1660          * path of the storage hardware device activation.
1661          */
1662         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1663         if (!kmpath_handlerd) {
1664                 DMERR("failed to create workqueue kmpath_handlerd");
1665                 destroy_workqueue(kmultipathd);
1666                 dm_unregister_target(&multipath_target);
1667                 kmem_cache_destroy(_mpio_cache);
1668                 return -ENOMEM;
1669         }
1670
1671         DMINFO("version %u.%u.%u loaded",
1672                multipath_target.version[0], multipath_target.version[1],
1673                multipath_target.version[2]);
1674
1675         return r;
1676 }
1677
1678 static void __exit dm_multipath_exit(void)
1679 {
1680         destroy_workqueue(kmpath_handlerd);
1681         destroy_workqueue(kmultipathd);
1682
1683         dm_unregister_target(&multipath_target);
1684         kmem_cache_destroy(_mpio_cache);
1685 }
1686
1687 module_init(dm_multipath_init);
1688 module_exit(dm_multipath_exit);
1689
1690 MODULE_DESCRIPTION(DM_NAME " multipath target");
1691 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1692 MODULE_LICENSE("GPL");