git.oblomov.eu Git - linux-2.6/blob - 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         unsigned nr_priority_groups;
  68         struct list_head priority_groups;
  69         unsigned pg_init_required;      /* pg_init needs calling? */
  70         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  71
  72         unsigned nr_valid_paths;        /* Total number of usable paths */
  73         struct pgpath *current_pgpath;
  74         struct priority_group *current_pg;
  75         struct priority_group *next_pg; /* Switch to this PG if set */
  76         unsigned repeat_count;          /* I/Os left before calling PS again */
  77
  78         unsigned queue_io;              /* Must we queue all I/O? */
  79         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  80         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  81         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  82         unsigned pg_init_count;         /* Number of times pg_init called */
  83
  84         struct work_struct process_queued_ios;
  85         struct list_head queued_ios;
  86         unsigned queue_size;
  87
  88         struct work_struct trigger_event;
  89
  90         /*
  91          * We must use a mempool of dm_mpath_io structs so that we
  92          * can resubmit bios on error.
  93          */
  94         mempool_t *mpio_pool;
  95 };
  96
  97 /*
  98  * Context information attached to each bio we process.
  99  */
 100 struct dm_mpath_io {
 101         struct pgpath *pgpath;
 102         size_t nr_bytes;
 103 };
 104
 105 typedef int (*action_fn) (struct pgpath *pgpath);
 106
 107 #define MIN_IOS 256     /* Mempool size */
 108
 109 static struct kmem_cache *_mpio_cache;
 110
 111 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 112 static void process_queued_ios(struct work_struct *work);
 113 static void trigger_event(struct work_struct *work);
 114 static void activate_path(struct work_struct *work);
 115 static void deactivate_path(struct work_struct *work);
 116
 117
 118 /*-----------------------------------------------
 119  * Allocation routines
 120  *-----------------------------------------------*/
 121
 122 static struct pgpath *alloc_pgpath(void)
 123 {
 124         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 125
 126         if (pgpath) {
 127                 pgpath->is_active = 1;
 128                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
 129                 INIT_WORK(&pgpath->activate_path, activate_path);
 130         }
 131
 132         return pgpath;
 133 }
 134
 135 static void free_pgpath(struct pgpath *pgpath)
 136 {
 137         kfree(pgpath);
 138 }
 139
 140 static void deactivate_path(struct work_struct *work)
 141 {
 142         struct pgpath *pgpath =
 143                 container_of(work, struct pgpath, deactivate_path);
 144
 145         blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
 146 }
 147
 148 static struct priority_group *alloc_priority_group(void)
 149 {
 150         struct priority_group *pg;
 151
 152         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 153
 154         if (pg)
 155                 INIT_LIST_HEAD(&pg->pgpaths);
 156
 157         return pg;
 158 }
 159
 160 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 161 {
 162         struct pgpath *pgpath, *tmp;
 163         struct multipath *m = ti->private;
 164
 165         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 166                 list_del(&pgpath->list);
 167                 if (m->hw_handler_name)
 168                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 169                 dm_put_device(ti, pgpath->path.dev);
 170                 free_pgpath(pgpath);
 171         }
 172 }
 173
 174 static void free_priority_group(struct priority_group *pg,
 175                                 struct dm_target *ti)
 176 {
 177         struct path_selector *ps = &pg->ps;
 178
 179         if (ps->type) {
 180                 ps->type->destroy(ps);
 181                 dm_put_path_selector(ps->type);
 182         }
 183
 184         free_pgpaths(&pg->pgpaths, ti);
 185         kfree(pg);
 186 }
 187
 188 static struct multipath *alloc_multipath(struct dm_target *ti)
 189 {
 190         struct multipath *m;
 191
 192         m = kzalloc(sizeof(*m), GFP_KERNEL);
 193         if (m) {
 194                 INIT_LIST_HEAD(&m->priority_groups);
 195                 INIT_LIST_HEAD(&m->queued_ios);
 196                 spin_lock_init(&m->lock);
 197                 m->queue_io = 1;
 198                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 199                 INIT_WORK(&m->trigger_event, trigger_event);
 200                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 201                 if (!m->mpio_pool) {
 202                         kfree(m);
 203                         return NULL;
 204                 }
 205                 m->ti = ti;
 206                 ti->private = m;
 207         }
 208
 209         return m;
 210 }
 211
 212 static void free_multipath(struct multipath *m)
 213 {
 214         struct priority_group *pg, *tmp;
 215
 216         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 217                 list_del(&pg->list);
 218                 free_priority_group(pg, m->ti);
 219         }
 220
 221         kfree(m->hw_handler_name);
 222         mempool_destroy(m->mpio_pool);
 223         kfree(m);
 224 }
 225
 226
 227 /*-----------------------------------------------
 228  * Path selection
 229  *-----------------------------------------------*/
 230
 231 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 232 {
 233         m->current_pg = pgpath->pg;
 234
 235         /* Must we initialise the PG first, and queue I/O till it's ready? */
 236         if (m->hw_handler_name) {
 237                 m->pg_init_required = 1;
 238                 m->queue_io = 1;
 239         } else {
 240                 m->pg_init_required = 0;
 241                 m->queue_io = 0;
 242         }
 243
 244         m->pg_init_count = 0;
 245 }
 246
 247 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 248                                size_t nr_bytes)
 249 {
 250         struct dm_path *path;
 251
 252         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 253         if (!path)
 254                 return -ENXIO;
 255
 256         m->current_pgpath = path_to_pgpath(path);
 257
 258         if (m->current_pg != pg)
 259                 __switch_pg(m, m->current_pgpath);
 260
 261         return 0;
 262 }
 263
 264 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 265 {
 266         struct priority_group *pg;
 267         unsigned bypassed = 1;
 268
 269         if (!m->nr_valid_paths)
 270                 goto failed;
 271
 272         /* Were we instructed to switch PG? */
 273         if (m->next_pg) {
 274                 pg = m->next_pg;
 275                 m->next_pg = NULL;
 276                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 277                         return;
 278         }
 279
 280         /* Don't change PG until it has no remaining paths */
 281         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 282                 return;
 283
 284         /*
 285          * Loop through priority groups until we find a valid path.
 286          * First time we skip PGs marked 'bypassed'.
 287          * Second time we only try the ones we skipped.
 288          */
 289         do {
 290                 list_for_each_entry(pg, &m->priority_groups, list) {
 291                         if (pg->bypassed == bypassed)
 292                                 continue;
 293                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 294                                 return;
 295                 }
 296         } while (bypassed--);
 297
 298 failed:
 299         m->current_pgpath = NULL;
 300         m->current_pg = NULL;
 301 }
 302
 303 /*
 304  * Check whether bios must be queued in the device-mapper core rather
 305  * than here in the target.
 306  *
 307  * m->lock must be held on entry.
 308  *
 309  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 310  * same value then we are not between multipath_presuspend()
 311  * and multipath_resume() calls and we have no need to check
 312  * for the DMF_NOFLUSH_SUSPENDING flag.
 313  */
 314 static int __must_push_back(struct multipath *m)
 315 {
 316         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 317                 dm_noflush_suspending(m->ti));
 318 }
 319
 320 static int map_io(struct multipath *m, struct request *clone,
 321                   struct dm_mpath_io *mpio, unsigned was_queued)
 322 {
 323         int r = DM_MAPIO_REMAPPED;
 324         size_t nr_bytes = blk_rq_bytes(clone);
 325         unsigned long flags;
 326         struct pgpath *pgpath;
 327         struct block_device *bdev;
 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                 list_add_tail(&clone->queuelist, &m->queued_ios);
 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                 bdev = pgpath->path.dev->bdev;
 353                 clone->q = bdev_get_queue(bdev);
 354                 clone->rq_disk = bdev->bd_disk;
 355         } else if (__must_push_back(m))
 356                 r = DM_MAPIO_REQUEUE;
 357         else
 358                 r = -EIO;       /* Failed */
 359
 360         mpio->pgpath = pgpath;
 361         mpio->nr_bytes = nr_bytes;
 362
 363         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 364                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 365                                               nr_bytes);
 366
 367         spin_unlock_irqrestore(&m->lock, flags);
 368
 369         return r;
 370 }
 371
 372 /*
 373  * If we run out of usable paths, should we queue I/O or error it?
 374  */
 375 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 376                             unsigned save_old_value)
 377 {
 378         unsigned long flags;
 379
 380         spin_lock_irqsave(&m->lock, flags);
 381
 382         if (save_old_value)
 383                 m->saved_queue_if_no_path = m->queue_if_no_path;
 384         else
 385                 m->saved_queue_if_no_path = queue_if_no_path;
 386         m->queue_if_no_path = queue_if_no_path;
 387         if (!m->queue_if_no_path && m->queue_size)
 388                 queue_work(kmultipathd, &m->process_queued_ios);
 389
 390         spin_unlock_irqrestore(&m->lock, flags);
 391
 392         return 0;
 393 }
 394
 395 /*-----------------------------------------------------------------
 396  * The multipath daemon is responsible for resubmitting queued ios.
 397  *---------------------------------------------------------------*/
 398
 399 static void dispatch_queued_ios(struct multipath *m)
 400 {
 401         int r;
 402         unsigned long flags;
 403         struct dm_mpath_io *mpio;
 404         union map_info *info;
 405         struct request *clone, *n;
 406         LIST_HEAD(cl);
 407
 408         spin_lock_irqsave(&m->lock, flags);
 409         list_splice_init(&m->queued_ios, &cl);
 410         spin_unlock_irqrestore(&m->lock, flags);
 411
 412         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 413                 list_del_init(&clone->queuelist);
 414
 415                 info = dm_get_rq_mapinfo(clone);
 416                 mpio = info->ptr;
 417
 418                 r = map_io(m, clone, mpio, 1);
 419                 if (r < 0) {
 420                         mempool_free(mpio, m->mpio_pool);
 421                         dm_kill_unmapped_request(clone, r);
 422                 } else if (r == DM_MAPIO_REMAPPED)
 423                         dm_dispatch_request(clone);
 424                 else if (r == DM_MAPIO_REQUEUE) {
 425                         mempool_free(mpio, m->mpio_pool);
 426                         dm_requeue_unmapped_request(clone);
 427                 }
 428         }
 429 }
 430
 431 static void process_queued_ios(struct work_struct *work)
 432 {
 433         struct multipath *m =
 434                 container_of(work, struct multipath, process_queued_ios);
 435         struct pgpath *pgpath = NULL, *tmp;
 436         unsigned must_queue = 1;
 437         unsigned long flags;
 438
 439         spin_lock_irqsave(&m->lock, flags);
 440
 441         if (!m->queue_size)
 442                 goto out;
 443
 444         if (!m->current_pgpath)
 445                 __choose_pgpath(m, 0);
 446
 447         pgpath = m->current_pgpath;
 448
 449         if ((pgpath && !m->queue_io) ||
 450             (!pgpath && !m->queue_if_no_path))
 451                 must_queue = 0;
 452
 453         if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
 454                 m->pg_init_count++;
 455                 m->pg_init_required = 0;
 456                 list_for_each_entry(tmp, &pgpath->pg->pgpaths, list) {
 457                         if (queue_work(kmpath_handlerd, &tmp->activate_path))
 458                                 m->pg_init_in_progress++;
 459                 }
 460         }
 461 out:
 462         spin_unlock_irqrestore(&m->lock, flags);
 463         if (!must_queue)
 464                 dispatch_queued_ios(m);
 465 }
 466
 467 /*
 468  * An event is triggered whenever a path is taken out of use.
 469  * Includes path failure and PG bypass.
 470  */
 471 static void trigger_event(struct work_struct *work)
 472 {
 473         struct multipath *m =
 474                 container_of(work, struct multipath, trigger_event);
 475
 476         dm_table_event(m->ti->table);
 477 }
 478
 479 /*-----------------------------------------------------------------
 480  * Constructor/argument parsing:
 481  * <#multipath feature args> [<arg>]*
 482  * <#hw_handler args> [hw_handler [<arg>]*]
 483  * <#priority groups>
 484  * <initial priority group>
 485  *     [<selector> <#selector args> [<arg>]*
 486  *      <#paths> <#per-path selector args>
 487  *         [<path> [<arg>]* ]+ ]+
 488  *---------------------------------------------------------------*/
 489 struct param {
 490         unsigned min;
 491         unsigned max;
 492         char *error;
 493 };
 494
 495 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 496 {
 497         if (!str ||
 498             (sscanf(str, "%u", v) != 1) ||
 499             (*v < param->min) ||
 500             (*v > param->max)) {
 501                 *error = param->error;
 502                 return -EINVAL;
 503         }
 504
 505         return 0;
 506 }
 507
 508 struct arg_set {
 509         unsigned argc;
 510         char **argv;
 511 };
 512
 513 static char *shift(struct arg_set *as)
 514 {
 515         char *r;
 516
 517         if (as->argc) {
 518                 as->argc--;
 519                 r = *as->argv;
 520                 as->argv++;
 521                 return r;
 522         }
 523
 524         return NULL;
 525 }
 526
 527 static void consume(struct arg_set *as, unsigned n)
 528 {
 529         BUG_ON (as->argc < n);
 530         as->argc -= n;
 531         as->argv += n;
 532 }
 533
 534 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 535                                struct dm_target *ti)
 536 {
 537         int r;
 538         struct path_selector_type *pst;
 539         unsigned ps_argc;
 540
 541         static struct param _params[] = {
 542                 {0, 1024, "invalid number of path selector args"},
 543         };
 544
 545         pst = dm_get_path_selector(shift(as));
 546         if (!pst) {
 547                 ti->error = "unknown path selector type";
 548                 return -EINVAL;
 549         }
 550
 551         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 552         if (r) {
 553                 dm_put_path_selector(pst);
 554                 return -EINVAL;
 555         }
 556
 557         if (ps_argc > as->argc) {
 558                 dm_put_path_selector(pst);
 559                 ti->error = "not enough arguments for path selector";
 560                 return -EINVAL;
 561         }
 562
 563         r = pst->create(&pg->ps, ps_argc, as->argv);
 564         if (r) {
 565                 dm_put_path_selector(pst);
 566                 ti->error = "path selector constructor failed";
 567                 return r;
 568         }
 569
 570         pg->ps.type = pst;
 571         consume(as, ps_argc);
 572
 573         return 0;
 574 }
 575
 576 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 577                                struct dm_target *ti)
 578 {
 579         int r;
 580         struct pgpath *p;
 581         struct multipath *m = ti->private;
 582
 583         /* we need at least a path arg */
 584         if (as->argc < 1) {
 585                 ti->error = "no device given";
 586                 return ERR_PTR(-EINVAL);
 587         }
 588
 589         p = alloc_pgpath();
 590         if (!p)
 591                 return ERR_PTR(-ENOMEM);
 592
 593         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 594                           dm_table_get_mode(ti->table), &p->path.dev);
 595         if (r) {
 596                 ti->error = "error getting device";
 597                 goto bad;
 598         }
 599
 600         if (m->hw_handler_name) {
 601                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 602
 603                 r = scsi_dh_attach(q, m->hw_handler_name);
 604                 if (r == -EBUSY) {
 605                         /*
 606                          * Already attached to different hw_handler,
 607                          * try to reattach with correct one.
 608                          */
 609                         scsi_dh_detach(q);
 610                         r = scsi_dh_attach(q, m->hw_handler_name);
 611                 }
 612
 613                 if (r < 0) {
 614                         ti->error = "error attaching hardware handler";
 615                         dm_put_device(ti, p->path.dev);
 616                         goto bad;
 617                 }
 618         }
 619
 620         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 621         if (r) {
 622                 dm_put_device(ti, p->path.dev);
 623                 goto bad;
 624         }
 625
 626         return p;
 627
 628  bad:
 629         free_pgpath(p);
 630         return ERR_PTR(r);
 631 }
 632
 633 static struct priority_group *parse_priority_group(struct arg_set *as,
 634                                                    struct multipath *m)
 635 {
 636         static struct param _params[] = {
 637                 {1, 1024, "invalid number of paths"},
 638                 {0, 1024, "invalid number of selector args"}
 639         };
 640
 641         int r;
 642         unsigned i, nr_selector_args, nr_params;
 643         struct priority_group *pg;
 644         struct dm_target *ti = m->ti;
 645
 646         if (as->argc < 2) {
 647                 as->argc = 0;
 648                 ti->error = "not enough priority group arguments";
 649                 return ERR_PTR(-EINVAL);
 650         }
 651
 652         pg = alloc_priority_group();
 653         if (!pg) {
 654                 ti->error = "couldn't allocate priority group";
 655                 return ERR_PTR(-ENOMEM);
 656         }
 657         pg->m = m;
 658
 659         r = parse_path_selector(as, pg, ti);
 660         if (r)
 661                 goto bad;
 662
 663         /*
 664          * read the paths
 665          */
 666         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 667         if (r)
 668                 goto bad;
 669
 670         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 671         if (r)
 672                 goto bad;
 673
 674         nr_params = 1 + nr_selector_args;
 675         for (i = 0; i < pg->nr_pgpaths; i++) {
 676                 struct pgpath *pgpath;
 677                 struct arg_set path_args;
 678
 679                 if (as->argc < nr_params) {
 680                         ti->error = "not enough path parameters";
 681                         goto bad;
 682                 }
 683
 684                 path_args.argc = nr_params;
 685                 path_args.argv = as->argv;
 686
 687                 pgpath = parse_path(&path_args, &pg->ps, ti);
 688                 if (IS_ERR(pgpath)) {
 689                         r = PTR_ERR(pgpath);
 690                         goto bad;
 691                 }
 692
 693                 pgpath->pg = pg;
 694                 list_add_tail(&pgpath->list, &pg->pgpaths);
 695                 consume(as, nr_params);
 696         }
 697
 698         return pg;
 699
 700  bad:
 701         free_priority_group(pg, ti);
 702         return ERR_PTR(r);
 703 }
 704
 705 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 706 {
 707         unsigned hw_argc;
 708         struct dm_target *ti = m->ti;
 709
 710         static struct param _params[] = {
 711                 {0, 1024, "invalid number of hardware handler args"},
 712         };
 713
 714         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 715                 return -EINVAL;
 716
 717         if (!hw_argc)
 718                 return 0;
 719
 720         if (hw_argc > as->argc) {
 721                 ti->error = "not enough arguments for hardware handler";
 722                 return -EINVAL;
 723         }
 724
 725         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 726         request_module("scsi_dh_%s", m->hw_handler_name);
 727         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 728                 ti->error = "unknown hardware handler type";
 729                 kfree(m->hw_handler_name);
 730                 m->hw_handler_name = NULL;
 731                 return -EINVAL;
 732         }
 733
 734         if (hw_argc > 1)
 735                 DMWARN("Ignoring user-specified arguments for "
 736                        "hardware handler \"%s\"", m->hw_handler_name);
 737         consume(as, hw_argc - 1);
 738
 739         return 0;
 740 }
 741
 742 static int parse_features(struct arg_set *as, struct multipath *m)
 743 {
 744         int r;
 745         unsigned argc;
 746         struct dm_target *ti = m->ti;
 747         const char *param_name;
 748
 749         static struct param _params[] = {
 750                 {0, 3, "invalid number of feature args"},
 751                 {1, 50, "pg_init_retries must be between 1 and 50"},
 752         };
 753
 754         r = read_param(_params, shift(as), &argc, &ti->error);
 755         if (r)
 756                 return -EINVAL;
 757
 758         if (!argc)
 759                 return 0;
 760
 761         do {
 762                 param_name = shift(as);
 763                 argc--;
 764
 765                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 766                         r = queue_if_no_path(m, 1, 0);
 767                         continue;
 768                 }
 769
 770                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 771                     (argc >= 1)) {
 772                         r = read_param(_params + 1, shift(as),
 773                                        &m->pg_init_retries, &ti->error);
 774                         argc--;
 775                         continue;
 776                 }
 777
 778                 ti->error = "Unrecognised multipath feature request";
 779                 r = -EINVAL;
 780         } while (argc && !r);
 781
 782         return r;
 783 }
 784
 785 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 786                          char **argv)
 787 {
 788         /* target parameters */
 789         static struct param _params[] = {
 790                 {1, 1024, "invalid number of priority groups"},
 791                 {1, 1024, "invalid initial priority group number"},
 792         };
 793
 794         int r;
 795         struct multipath *m;
 796         struct arg_set as;
 797         unsigned pg_count = 0;
 798         unsigned next_pg_num;
 799
 800         as.argc = argc;
 801         as.argv = argv;
 802
 803         m = alloc_multipath(ti);
 804         if (!m) {
 805                 ti->error = "can't allocate multipath";
 806                 return -EINVAL;
 807         }
 808
 809         r = parse_features(&as, m);
 810         if (r)
 811                 goto bad;
 812
 813         r = parse_hw_handler(&as, m);
 814         if (r)
 815                 goto bad;
 816
 817         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 818         if (r)
 819                 goto bad;
 820
 821         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 822         if (r)
 823                 goto bad;
 824
 825         /* parse the priority groups */
 826         while (as.argc) {
 827                 struct priority_group *pg;
 828
 829                 pg = parse_priority_group(&as, m);
 830                 if (IS_ERR(pg)) {
 831                         r = PTR_ERR(pg);
 832                         goto bad;
 833                 }
 834
 835                 m->nr_valid_paths += pg->nr_pgpaths;
 836                 list_add_tail(&pg->list, &m->priority_groups);
 837                 pg_count++;
 838                 pg->pg_num = pg_count;
 839                 if (!--next_pg_num)
 840                         m->next_pg = pg;
 841         }
 842
 843         if (pg_count != m->nr_priority_groups) {
 844                 ti->error = "priority group count mismatch";
 845                 r = -EINVAL;
 846                 goto bad;
 847         }
 848
 849         ti->num_flush_requests = 1;
 850
 851         return 0;
 852
 853  bad:
 854         free_multipath(m);
 855         return r;
 856 }
 857
 858 static void multipath_dtr(struct dm_target *ti)
 859 {
 860         struct multipath *m = (struct multipath *) ti->private;
 861
 862         flush_workqueue(kmpath_handlerd);
 863         flush_workqueue(kmultipathd);
 864         flush_scheduled_work();
 865         free_multipath(m);
 866 }
 867
 868 /*
 869  * Map cloned requests
 870  */
 871 static int multipath_map(struct dm_target *ti, struct request *clone,
 872                          union map_info *map_context)
 873 {
 874         int r;
 875         struct dm_mpath_io *mpio;
 876         struct multipath *m = (struct multipath *) ti->private;
 877
 878         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 879         if (!mpio)
 880                 /* ENOMEM, requeue */
 881                 return DM_MAPIO_REQUEUE;
 882         memset(mpio, 0, sizeof(*mpio));
 883
 884         map_context->ptr = mpio;
 885         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 886         r = map_io(m, clone, mpio, 0);
 887         if (r < 0 || r == DM_MAPIO_REQUEUE)
 888                 mempool_free(mpio, m->mpio_pool);
 889
 890         return r;
 891 }
 892
 893 /*
 894  * Take a path out of use.
 895  */
 896 static int fail_path(struct pgpath *pgpath)
 897 {
 898         unsigned long flags;
 899         struct multipath *m = pgpath->pg->m;
 900
 901         spin_lock_irqsave(&m->lock, flags);
 902
 903         if (!pgpath->is_active)
 904                 goto out;
 905
 906         DMWARN("Failing path %s.", pgpath->path.dev->name);
 907
 908         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 909         pgpath->is_active = 0;
 910         pgpath->fail_count++;
 911
 912         m->nr_valid_paths--;
 913
 914         if (pgpath == m->current_pgpath)
 915                 m->current_pgpath = NULL;
 916
 917         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 918                       pgpath->path.dev->name, m->nr_valid_paths);
 919
 920         schedule_work(&m->trigger_event);
 921         queue_work(kmultipathd, &pgpath->deactivate_path);
 922
 923 out:
 924         spin_unlock_irqrestore(&m->lock, flags);
 925
 926         return 0;
 927 }
 928
 929 /*
 930  * Reinstate a previously-failed path
 931  */
 932 static int reinstate_path(struct pgpath *pgpath)
 933 {
 934         int r = 0;
 935         unsigned long flags;
 936         struct multipath *m = pgpath->pg->m;
 937
 938         spin_lock_irqsave(&m->lock, flags);
 939
 940         if (pgpath->is_active)
 941                 goto out;
 942
 943         if (!pgpath->pg->ps.type->reinstate_path) {
 944                 DMWARN("Reinstate path not supported by path selector %s",
 945                        pgpath->pg->ps.type->name);
 946                 r = -EINVAL;
 947                 goto out;
 948         }
 949
 950         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 951         if (r)
 952                 goto out;
 953
 954         pgpath->is_active = 1;
 955
 956         if (!m->nr_valid_paths++ && m->queue_size) {
 957                 m->current_pgpath = NULL;
 958                 queue_work(kmultipathd, &m->process_queued_ios);
 959         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
 960                 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
 961                         m->pg_init_in_progress++;
 962         }
 963
 964         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 965                       pgpath->path.dev->name, m->nr_valid_paths);
 966
 967         schedule_work(&m->trigger_event);
 968
 969 out:
 970         spin_unlock_irqrestore(&m->lock, flags);
 971
 972         return r;
 973 }
 974
 975 /*
 976  * Fail or reinstate all paths that match the provided struct dm_dev.
 977  */
 978 static int action_dev(struct multipath *m, struct dm_dev *dev,
 979                       action_fn action)
 980 {
 981         int r = 0;
 982         struct pgpath *pgpath;
 983         struct priority_group *pg;
 984
 985         list_for_each_entry(pg, &m->priority_groups, list) {
 986                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 987                         if (pgpath->path.dev == dev)
 988                                 r = action(pgpath);
 989                 }
 990         }
 991
 992         return r;
 993 }
 994
 995 /*
 996  * Temporarily try to avoid having to use the specified PG
 997  */
 998 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 999                       int bypassed)
1000 {
1001         unsigned long flags;
1002
1003         spin_lock_irqsave(&m->lock, flags);
1004
1005         pg->bypassed = bypassed;
1006         m->current_pgpath = NULL;
1007         m->current_pg = NULL;
1008
1009         spin_unlock_irqrestore(&m->lock, flags);
1010
1011         schedule_work(&m->trigger_event);
1012 }
1013
1014 /*
1015  * Switch to using the specified PG from the next I/O that gets mapped
1016  */
1017 static int switch_pg_num(struct multipath *m, const char *pgstr)
1018 {
1019         struct priority_group *pg;
1020         unsigned pgnum;
1021         unsigned long flags;
1022
1023         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1024             (pgnum > m->nr_priority_groups)) {
1025                 DMWARN("invalid PG number supplied to switch_pg_num");
1026                 return -EINVAL;
1027         }
1028
1029         spin_lock_irqsave(&m->lock, flags);
1030         list_for_each_entry(pg, &m->priority_groups, list) {
1031                 pg->bypassed = 0;
1032                 if (--pgnum)
1033                         continue;
1034
1035                 m->current_pgpath = NULL;
1036                 m->current_pg = NULL;
1037                 m->next_pg = pg;
1038         }
1039         spin_unlock_irqrestore(&m->lock, flags);
1040
1041         schedule_work(&m->trigger_event);
1042         return 0;
1043 }
1044
1045 /*
1046  * Set/clear bypassed status of a PG.
1047  * PGs are numbered upwards from 1 in the order they were declared.
1048  */
1049 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1050 {
1051         struct priority_group *pg;
1052         unsigned pgnum;
1053
1054         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1055             (pgnum > m->nr_priority_groups)) {
1056                 DMWARN("invalid PG number supplied to bypass_pg");
1057                 return -EINVAL;
1058         }
1059
1060         list_for_each_entry(pg, &m->priority_groups, list) {
1061                 if (!--pgnum)
1062                         break;
1063         }
1064
1065         bypass_pg(m, pg, bypassed);
1066         return 0;
1067 }
1068
1069 /*
1070  * Should we retry pg_init immediately?
1071  */
1072 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1073 {
1074         unsigned long flags;
1075         int limit_reached = 0;
1076
1077         spin_lock_irqsave(&m->lock, flags);
1078
1079         if (m->pg_init_count <= m->pg_init_retries)
1080                 m->pg_init_required = 1;
1081         else
1082                 limit_reached = 1;
1083
1084         spin_unlock_irqrestore(&m->lock, flags);
1085
1086         return limit_reached;
1087 }
1088
1089 static void pg_init_done(struct dm_path *path, int errors)
1090 {
1091         struct pgpath *pgpath = path_to_pgpath(path);
1092         struct priority_group *pg = pgpath->pg;
1093         struct multipath *m = pg->m;
1094         unsigned long flags;
1095
1096         /* device or driver problems */
1097         switch (errors) {
1098         case SCSI_DH_OK:
1099                 break;
1100         case SCSI_DH_NOSYS:
1101                 if (!m->hw_handler_name) {
1102                         errors = 0;
1103                         break;
1104                 }
1105                 DMERR("Cannot failover device because scsi_dh_%s was not "
1106                       "loaded.", m->hw_handler_name);
1107                 /*
1108                  * Fail path for now, so we do not ping pong
1109                  */
1110                 fail_path(pgpath);
1111                 break;
1112         case SCSI_DH_DEV_TEMP_BUSY:
1113                 /*
1114                  * Probably doing something like FW upgrade on the
1115                  * controller so try the other pg.
1116                  */
1117                 bypass_pg(m, pg, 1);
1118                 break;
1119         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1120         case SCSI_DH_RETRY:
1121         case SCSI_DH_IMM_RETRY:
1122         case SCSI_DH_RES_TEMP_UNAVAIL:
1123                 if (pg_init_limit_reached(m, pgpath))
1124                         fail_path(pgpath);
1125                 errors = 0;
1126                 break;
1127         default:
1128                 /*
1129                  * We probably do not want to fail the path for a device
1130                  * error, but this is what the old dm did. In future
1131                  * patches we can do more advanced handling.
1132                  */
1133                 fail_path(pgpath);
1134         }
1135
1136         spin_lock_irqsave(&m->lock, flags);
1137         if (errors) {
1138                 if (pgpath == m->current_pgpath) {
1139                         DMERR("Could not failover device. Error %d.", errors);
1140                         m->current_pgpath = NULL;
1141                         m->current_pg = NULL;
1142                 }
1143         } else if (!m->pg_init_required) {
1144                 m->queue_io = 0;
1145                 pg->bypassed = 0;
1146         }
1147
1148         m->pg_init_in_progress--;
1149         if (!m->pg_init_in_progress)
1150                 queue_work(kmultipathd, &m->process_queued_ios);
1151         spin_unlock_irqrestore(&m->lock, flags);
1152 }
1153
1154 static void activate_path(struct work_struct *work)
1155 {
1156         int ret;
1157         struct pgpath *pgpath =
1158                 container_of(work, struct pgpath, activate_path);
1159
1160         ret = scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev));
1161         pg_init_done(&pgpath->path, ret);
1162 }
1163
1164 /*
1165  * end_io handling
1166  */
1167 static int do_end_io(struct multipath *m, struct request *clone,
1168                      int error, struct dm_mpath_io *mpio)
1169 {
1170         /*
1171          * We don't queue any clone request inside the multipath target
1172          * during end I/O handling, since those clone requests don't have
1173          * bio clones.  If we queue them inside the multipath target,
1174          * we need to make bio clones, that requires memory allocation.
1175          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1176          *  don't have bio clones.)
1177          * Instead of queueing the clone request here, we queue the original
1178          * request into dm core, which will remake a clone request and
1179          * clone bios for it and resubmit it later.
1180          */
1181         int r = DM_ENDIO_REQUEUE;
1182         unsigned long flags;
1183
1184         if (!error && !clone->errors)
1185                 return 0;       /* I/O complete */
1186
1187         if (error == -EOPNOTSUPP)
1188                 return error;
1189
1190         if (mpio->pgpath)
1191                 fail_path(mpio->pgpath);
1192
1193         spin_lock_irqsave(&m->lock, flags);
1194         if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1195                 r = -EIO;
1196         spin_unlock_irqrestore(&m->lock, flags);
1197
1198         return r;
1199 }
1200
1201 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1202                             int error, union map_info *map_context)
1203 {
1204         struct multipath *m = ti->private;
1205         struct dm_mpath_io *mpio = map_context->ptr;
1206         struct pgpath *pgpath = mpio->pgpath;
1207         struct path_selector *ps;
1208         int r;
1209
1210         r  = do_end_io(m, clone, error, mpio);
1211         if (pgpath) {
1212                 ps = &pgpath->pg->ps;
1213                 if (ps->type->end_io)
1214                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1215         }
1216         mempool_free(mpio, m->mpio_pool);
1217
1218         return r;
1219 }
1220
1221 /*
1222  * Suspend can't complete until all the I/O is processed so if
1223  * the last path fails we must error any remaining I/O.
1224  * Note that if the freeze_bdev fails while suspending, the
1225  * queue_if_no_path state is lost - userspace should reset it.
1226  */
1227 static void multipath_presuspend(struct dm_target *ti)
1228 {
1229         struct multipath *m = (struct multipath *) ti->private;
1230
1231         queue_if_no_path(m, 0, 1);
1232 }
1233
1234 /*
1235  * Restore the queue_if_no_path setting.
1236  */
1237 static void multipath_resume(struct dm_target *ti)
1238 {
1239         struct multipath *m = (struct multipath *) ti->private;
1240         unsigned long flags;
1241
1242         spin_lock_irqsave(&m->lock, flags);
1243         m->queue_if_no_path = m->saved_queue_if_no_path;
1244         spin_unlock_irqrestore(&m->lock, flags);
1245 }
1246
1247 /*
1248  * Info output has the following format:
1249  * num_multipath_feature_args [multipath_feature_args]*
1250  * num_handler_status_args [handler_status_args]*
1251  * num_groups init_group_number
1252  *            [A|D|E num_ps_status_args [ps_status_args]*
1253  *             num_paths num_selector_args
1254  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1255  *
1256  * Table output has the following format (identical to the constructor string):
1257  * num_feature_args [features_args]*
1258  * num_handler_args hw_handler [hw_handler_args]*
1259  * num_groups init_group_number
1260  *     [priority selector-name num_ps_args [ps_args]*
1261  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1262  */
1263 static int multipath_status(struct dm_target *ti, status_type_t type,
1264                             char *result, unsigned int maxlen)
1265 {
1266         int sz = 0;
1267         unsigned long flags;
1268         struct multipath *m = (struct multipath *) ti->private;
1269         struct priority_group *pg;
1270         struct pgpath *p;
1271         unsigned pg_num;
1272         char state;
1273
1274         spin_lock_irqsave(&m->lock, flags);
1275
1276         /* Features */
1277         if (type == STATUSTYPE_INFO)
1278                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1279         else {
1280                 DMEMIT("%u ", m->queue_if_no_path +
1281                               (m->pg_init_retries > 0) * 2);
1282                 if (m->queue_if_no_path)
1283                         DMEMIT("queue_if_no_path ");
1284                 if (m->pg_init_retries)
1285                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1286         }
1287
1288         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1289                 DMEMIT("0 ");
1290         else
1291                 DMEMIT("1 %s ", m->hw_handler_name);
1292
1293         DMEMIT("%u ", m->nr_priority_groups);
1294
1295         if (m->next_pg)
1296                 pg_num = m->next_pg->pg_num;
1297         else if (m->current_pg)
1298                 pg_num = m->current_pg->pg_num;
1299         else
1300                         pg_num = 1;
1301
1302         DMEMIT("%u ", pg_num);
1303
1304         switch (type) {
1305         case STATUSTYPE_INFO:
1306                 list_for_each_entry(pg, &m->priority_groups, list) {
1307                         if (pg->bypassed)
1308                                 state = 'D';    /* Disabled */
1309                         else if (pg == m->current_pg)
1310                                 state = 'A';    /* Currently Active */
1311                         else
1312                                 state = 'E';    /* Enabled */
1313
1314                         DMEMIT("%c ", state);
1315
1316                         if (pg->ps.type->status)
1317                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1318                                                           result + sz,
1319                                                           maxlen - sz);
1320                         else
1321                                 DMEMIT("0 ");
1322
1323                         DMEMIT("%u %u ", pg->nr_pgpaths,
1324                                pg->ps.type->info_args);
1325
1326                         list_for_each_entry(p, &pg->pgpaths, list) {
1327                                 DMEMIT("%s %s %u ", p->path.dev->name,
1328                                        p->is_active ? "A" : "F",
1329                                        p->fail_count);
1330                                 if (pg->ps.type->status)
1331                                         sz += pg->ps.type->status(&pg->ps,
1332                                               &p->path, type, result + sz,
1333                                               maxlen - sz);
1334                         }
1335                 }
1336                 break;
1337
1338         case STATUSTYPE_TABLE:
1339                 list_for_each_entry(pg, &m->priority_groups, list) {
1340                         DMEMIT("%s ", pg->ps.type->name);
1341
1342                         if (pg->ps.type->status)
1343                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1344                                                           result + sz,
1345                                                           maxlen - sz);
1346                         else
1347                                 DMEMIT("0 ");
1348
1349                         DMEMIT("%u %u ", pg->nr_pgpaths,
1350                                pg->ps.type->table_args);
1351
1352                         list_for_each_entry(p, &pg->pgpaths, list) {
1353                                 DMEMIT("%s ", p->path.dev->name);
1354                                 if (pg->ps.type->status)
1355                                         sz += pg->ps.type->status(&pg->ps,
1356                                               &p->path, type, result + sz,
1357                                               maxlen - sz);
1358                         }
1359                 }
1360                 break;
1361         }
1362
1363         spin_unlock_irqrestore(&m->lock, flags);
1364
1365         return 0;
1366 }
1367
1368 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1369 {
1370         int r;
1371         struct dm_dev *dev;
1372         struct multipath *m = (struct multipath *) ti->private;
1373         action_fn action;
1374
1375         if (argc == 1) {
1376                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1377                         return queue_if_no_path(m, 1, 0);
1378                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1379                         return queue_if_no_path(m, 0, 0);
1380         }
1381
1382         if (argc != 2)
1383                 goto error;
1384
1385         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1386                 return bypass_pg_num(m, argv[1], 1);
1387         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1388                 return bypass_pg_num(m, argv[1], 0);
1389         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1390                 return switch_pg_num(m, argv[1]);
1391         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1392                 action = reinstate_path;
1393         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1394                 action = fail_path;
1395         else
1396                 goto error;
1397
1398         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1399                           dm_table_get_mode(ti->table), &dev);
1400         if (r) {
1401                 DMWARN("message: error getting device %s",
1402                        argv[1]);
1403                 return -EINVAL;
1404         }
1405
1406         r = action_dev(m, dev, action);
1407
1408         dm_put_device(ti, dev);
1409
1410         return r;
1411
1412 error:
1413         DMWARN("Unrecognised multipath message received.");
1414         return -EINVAL;
1415 }
1416
1417 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1418                            unsigned long arg)
1419 {
1420         struct multipath *m = (struct multipath *) ti->private;
1421         struct block_device *bdev = NULL;
1422         fmode_t mode = 0;
1423         unsigned long flags;
1424         int r = 0;
1425
1426         spin_lock_irqsave(&m->lock, flags);
1427
1428         if (!m->current_pgpath)
1429                 __choose_pgpath(m, 0);
1430
1431         if (m->current_pgpath) {
1432                 bdev = m->current_pgpath->path.dev->bdev;
1433                 mode = m->current_pgpath->path.dev->mode;
1434         }
1435
1436         if (m->queue_io)
1437                 r = -EAGAIN;
1438         else if (!bdev)
1439                 r = -EIO;
1440
1441         spin_unlock_irqrestore(&m->lock, flags);
1442
1443         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1444 }
1445
1446 static int multipath_iterate_devices(struct dm_target *ti,
1447                                      iterate_devices_callout_fn fn, void *data)
1448 {
1449         struct multipath *m = ti->private;
1450         struct priority_group *pg;
1451         struct pgpath *p;
1452         int ret = 0;
1453
1454         list_for_each_entry(pg, &m->priority_groups, list) {
1455                 list_for_each_entry(p, &pg->pgpaths, list) {
1456                         ret = fn(ti, p->path.dev, ti->begin, data);
1457                         if (ret)
1458                                 goto out;
1459                 }
1460         }
1461
1462 out:
1463         return ret;
1464 }
1465
1466 static int __pgpath_busy(struct pgpath *pgpath)
1467 {
1468         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1469
1470         return dm_underlying_device_busy(q);
1471 }
1472
1473 /*
1474  * We return "busy", only when we can map I/Os but underlying devices
1475  * are busy (so even if we map I/Os now, the I/Os will wait on
1476  * the underlying queue).
1477  * In other words, if we want to kill I/Os or queue them inside us
1478  * due to map unavailability, we don't return "busy".  Otherwise,
1479  * dm core won't give us the I/Os and we can't do what we want.
1480  */
1481 static int multipath_busy(struct dm_target *ti)
1482 {
1483         int busy = 0, has_active = 0;
1484         struct multipath *m = ti->private;
1485         struct priority_group *pg;
1486         struct pgpath *pgpath;
1487         unsigned long flags;
1488
1489         spin_lock_irqsave(&m->lock, flags);
1490
1491         /* Guess which priority_group will be used at next mapping time */
1492         if (unlikely(!m->current_pgpath && m->next_pg))
1493                 pg = m->next_pg;
1494         else if (likely(m->current_pg))
1495                 pg = m->current_pg;
1496         else
1497                 /*
1498                  * We don't know which pg will be used at next mapping time.
1499                  * We don't call __choose_pgpath() here to avoid to trigger
1500                  * pg_init just by busy checking.
1501                  * So we don't know whether underlying devices we will be using
1502                  * at next mapping time are busy or not. Just try mapping.
1503                  */
1504                 goto out;
1505
1506         /*
1507          * If there is one non-busy active path at least, the path selector
1508          * will be able to select it. So we consider such a pg as not busy.
1509          */
1510         busy = 1;
1511         list_for_each_entry(pgpath, &pg->pgpaths, list)
1512                 if (pgpath->is_active) {
1513                         has_active = 1;
1514
1515                         if (!__pgpath_busy(pgpath)) {
1516                                 busy = 0;
1517                                 break;
1518                         }
1519                 }
1520
1521         if (!has_active)
1522                 /*
1523                  * No active path in this pg, so this pg won't be used and
1524                  * the current_pg will be changed at next mapping time.
1525                  * We need to try mapping to determine it.
1526                  */
1527                 busy = 0;
1528
1529 out:
1530         spin_unlock_irqrestore(&m->lock, flags);
1531
1532         return busy;
1533 }
1534
1535 /*-----------------------------------------------------------------
1536  * Module setup
1537  *---------------------------------------------------------------*/
1538 static struct target_type multipath_target = {
1539         .name = "multipath",
1540         .version = {1, 1, 0},
1541         .module = THIS_MODULE,
1542         .ctr = multipath_ctr,
1543         .dtr = multipath_dtr,
1544         .map_rq = multipath_map,
1545         .rq_end_io = multipath_end_io,
1546         .presuspend = multipath_presuspend,
1547         .resume = multipath_resume,
1548         .status = multipath_status,
1549         .message = multipath_message,
1550         .ioctl  = multipath_ioctl,
1551         .iterate_devices = multipath_iterate_devices,
1552         .busy = multipath_busy,
1553 };
1554
1555 static int __init dm_multipath_init(void)
1556 {
1557         int r;
1558
1559         /* allocate a slab for the dm_ios */
1560         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1561         if (!_mpio_cache)
1562                 return -ENOMEM;
1563
1564         r = dm_register_target(&multipath_target);
1565         if (r < 0) {
1566                 DMERR("register failed %d", r);
1567                 kmem_cache_destroy(_mpio_cache);
1568                 return -EINVAL;
1569         }
1570
1571         kmultipathd = create_workqueue("kmpathd");
1572         if (!kmultipathd) {
1573                 DMERR("failed to create workqueue kmpathd");
1574                 dm_unregister_target(&multipath_target);
1575                 kmem_cache_destroy(_mpio_cache);
1576                 return -ENOMEM;
1577         }
1578
1579         /*
1580          * A separate workqueue is used to handle the device handlers
1581          * to avoid overloading existing workqueue. Overloading the
1582          * old workqueue would also create a bottleneck in the
1583          * path of the storage hardware device activation.
1584          */
1585         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1586         if (!kmpath_handlerd) {
1587                 DMERR("failed to create workqueue kmpath_handlerd");
1588                 destroy_workqueue(kmultipathd);
1589                 dm_unregister_target(&multipath_target);
1590                 kmem_cache_destroy(_mpio_cache);
1591                 return -ENOMEM;
1592         }
1593
1594         DMINFO("version %u.%u.%u loaded",
1595                multipath_target.version[0], multipath_target.version[1],
1596                multipath_target.version[2]);
1597
1598         return r;
1599 }
1600
1601 static void __exit dm_multipath_exit(void)
1602 {
1603         destroy_workqueue(kmpath_handlerd);
1604         destroy_workqueue(kmultipathd);
1605
1606         dm_unregister_target(&multipath_target);
1607         kmem_cache_destroy(_mpio_cache);
1608 }
1609
1610 module_init(dm_multipath_init);
1611 module_exit(dm_multipath_exit);
1612
1613 MODULE_DESCRIPTION(DM_NAME " multipath target");
1614 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1615 MODULE_LICENSE("GPL");