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
  98 /*
  99  * Context information attached to each bio we process.
 100  */
 101 struct dm_mpath_io {
 102         struct pgpath *pgpath;
 103         size_t nr_bytes;
 104 };
 105
 106 typedef int (*action_fn) (struct pgpath *pgpath);
 107
 108 #define MIN_IOS 256     /* Mempool size */
 109
 110 static struct kmem_cache *_mpio_cache;
 111
 112 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 113 static void process_queued_ios(struct work_struct *work);
 114 static void trigger_event(struct work_struct *work);
 115 static void activate_path(struct work_struct *work);
 116 static void deactivate_path(struct work_struct *work);
 117
 118
 119 /*-----------------------------------------------
 120  * Allocation routines
 121  *-----------------------------------------------*/
 122
 123 static struct pgpath *alloc_pgpath(void)
 124 {
 125         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 126
 127         if (pgpath) {
 128                 pgpath->is_active = 1;
 129                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
 130                 INIT_WORK(&pgpath->activate_path, activate_path);
 131         }
 132
 133         return pgpath;
 134 }
 135
 136 static void free_pgpath(struct pgpath *pgpath)
 137 {
 138         kfree(pgpath);
 139 }
 140
 141 static void deactivate_path(struct work_struct *work)
 142 {
 143         struct pgpath *pgpath =
 144                 container_of(work, struct pgpath, deactivate_path);
 145
 146         blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
 147 }
 148
 149 static struct priority_group *alloc_priority_group(void)
 150 {
 151         struct priority_group *pg;
 152
 153         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 154
 155         if (pg)
 156                 INIT_LIST_HEAD(&pg->pgpaths);
 157
 158         return pg;
 159 }
 160
 161 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 162 {
 163         struct pgpath *pgpath, *tmp;
 164         struct multipath *m = ti->private;
 165
 166         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 167                 list_del(&pgpath->list);
 168                 dm_put_device(ti, pgpath->path.dev);
 169                 free_pgpath(pgpath);
 170         }
 171 }
 172
 173 static void free_priority_group(struct priority_group *pg,
 174                                 struct dm_target *ti)
 175 {
 176         struct path_selector *ps = &pg->ps;
 177
 178         if (ps->type) {
 179                 ps->type->destroy(ps);
 180                 dm_put_path_selector(ps->type);
 181         }
 182
 183         free_pgpaths(&pg->pgpaths, ti);
 184         kfree(pg);
 185 }
 186
 187 static struct multipath *alloc_multipath(struct dm_target *ti)
 188 {
 189         struct multipath *m;
 190
 191         m = kzalloc(sizeof(*m), GFP_KERNEL);
 192         if (m) {
 193                 INIT_LIST_HEAD(&m->priority_groups);
 194                 INIT_LIST_HEAD(&m->queued_ios);
 195                 spin_lock_init(&m->lock);
 196                 m->queue_io = 1;
 197                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 198                 INIT_WORK(&m->trigger_event, trigger_event);
 199                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 200                 if (!m->mpio_pool) {
 201                         kfree(m);
 202                         return NULL;
 203                 }
 204                 m->ti = ti;
 205                 ti->private = m;
 206         }
 207
 208         return m;
 209 }
 210
 211 static void free_multipath(struct multipath *m)
 212 {
 213         struct priority_group *pg, *tmp;
 214
 215         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 216                 list_del(&pg->list);
 217                 free_priority_group(pg, m->ti);
 218         }
 219
 220         kfree(m->hw_handler_name);
 221         kfree(m->hw_handler_params);
 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                 if (m->hw_handler_params) {
 620                         r = scsi_dh_set_params(q, m->hw_handler_params);
 621                         if (r < 0) {
 622                                 ti->error = "unable to set hardware "
 623                                                         "handler parameters";
 624                                 scsi_dh_detach(q);
 625                                 dm_put_device(ti, p->path.dev);
 626                                 goto bad;
 627                         }
 628                 }
 629         }
 630
 631         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 632         if (r) {
 633                 dm_put_device(ti, p->path.dev);
 634                 goto bad;
 635         }
 636
 637         return p;
 638
 639  bad:
 640         free_pgpath(p);
 641         return ERR_PTR(r);
 642 }
 643
 644 static struct priority_group *parse_priority_group(struct arg_set *as,
 645                                                    struct multipath *m)
 646 {
 647         static struct param _params[] = {
 648                 {1, 1024, "invalid number of paths"},
 649                 {0, 1024, "invalid number of selector args"}
 650         };
 651
 652         int r;
 653         unsigned i, nr_selector_args, nr_params;
 654         struct priority_group *pg;
 655         struct dm_target *ti = m->ti;
 656
 657         if (as->argc < 2) {
 658                 as->argc = 0;
 659                 ti->error = "not enough priority group arguments";
 660                 return ERR_PTR(-EINVAL);
 661         }
 662
 663         pg = alloc_priority_group();
 664         if (!pg) {
 665                 ti->error = "couldn't allocate priority group";
 666                 return ERR_PTR(-ENOMEM);
 667         }
 668         pg->m = m;
 669
 670         r = parse_path_selector(as, pg, ti);
 671         if (r)
 672                 goto bad;
 673
 674         /*
 675          * read the paths
 676          */
 677         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 678         if (r)
 679                 goto bad;
 680
 681         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 682         if (r)
 683                 goto bad;
 684
 685         nr_params = 1 + nr_selector_args;
 686         for (i = 0; i < pg->nr_pgpaths; i++) {
 687                 struct pgpath *pgpath;
 688                 struct arg_set path_args;
 689
 690                 if (as->argc < nr_params) {
 691                         ti->error = "not enough path parameters";
 692                         goto bad;
 693                 }
 694
 695                 path_args.argc = nr_params;
 696                 path_args.argv = as->argv;
 697
 698                 pgpath = parse_path(&path_args, &pg->ps, ti);
 699                 if (IS_ERR(pgpath)) {
 700                         r = PTR_ERR(pgpath);
 701                         goto bad;
 702                 }
 703
 704                 pgpath->pg = pg;
 705                 list_add_tail(&pgpath->list, &pg->pgpaths);
 706                 consume(as, nr_params);
 707         }
 708
 709         return pg;
 710
 711  bad:
 712         free_priority_group(pg, ti);
 713         return ERR_PTR(r);
 714 }
 715
 716 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 717 {
 718         unsigned hw_argc;
 719         int ret;
 720         struct dm_target *ti = m->ti;
 721
 722         static struct param _params[] = {
 723                 {0, 1024, "invalid number of hardware handler args"},
 724         };
 725
 726         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 727                 return -EINVAL;
 728
 729         if (!hw_argc)
 730                 return 0;
 731
 732         if (hw_argc > as->argc) {
 733                 ti->error = "not enough arguments for hardware handler";
 734                 return -EINVAL;
 735         }
 736
 737         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 738         request_module("scsi_dh_%s", m->hw_handler_name);
 739         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 740                 ti->error = "unknown hardware handler type";
 741                 ret = -EINVAL;
 742                 goto fail;
 743         }
 744
 745         if (hw_argc > 1) {
 746                 char *p;
 747                 int i, j, len = 4;
 748
 749                 for (i = 0; i <= hw_argc - 2; i++)
 750                         len += strlen(as->argv[i]) + 1;
 751                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 752                 if (!p) {
 753                         ti->error = "memory allocation failed";
 754                         ret = -ENOMEM;
 755                         goto fail;
 756                 }
 757                 j = sprintf(p, "%d", hw_argc - 1);
 758                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 759                         j = sprintf(p, "%s", as->argv[i]);
 760         }
 761         consume(as, hw_argc - 1);
 762
 763         return 0;
 764 fail:
 765         kfree(m->hw_handler_name);
 766         m->hw_handler_name = NULL;
 767         return ret;
 768 }
 769
 770 static int parse_features(struct arg_set *as, struct multipath *m)
 771 {
 772         int r;
 773         unsigned argc;
 774         struct dm_target *ti = m->ti;
 775         const char *param_name;
 776
 777         static struct param _params[] = {
 778                 {0, 3, "invalid number of feature args"},
 779                 {1, 50, "pg_init_retries must be between 1 and 50"},
 780         };
 781
 782         r = read_param(_params, shift(as), &argc, &ti->error);
 783         if (r)
 784                 return -EINVAL;
 785
 786         if (!argc)
 787                 return 0;
 788
 789         do {
 790                 param_name = shift(as);
 791                 argc--;
 792
 793                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 794                         r = queue_if_no_path(m, 1, 0);
 795                         continue;
 796                 }
 797
 798                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 799                     (argc >= 1)) {
 800                         r = read_param(_params + 1, shift(as),
 801                                        &m->pg_init_retries, &ti->error);
 802                         argc--;
 803                         continue;
 804                 }
 805
 806                 ti->error = "Unrecognised multipath feature request";
 807                 r = -EINVAL;
 808         } while (argc && !r);
 809
 810         return r;
 811 }
 812
 813 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 814                          char **argv)
 815 {
 816         /* target parameters */
 817         static struct param _params[] = {
 818                 {0, 1024, "invalid number of priority groups"},
 819                 {0, 1024, "invalid initial priority group number"},
 820         };
 821
 822         int r;
 823         struct multipath *m;
 824         struct arg_set as;
 825         unsigned pg_count = 0;
 826         unsigned next_pg_num;
 827
 828         as.argc = argc;
 829         as.argv = argv;
 830
 831         m = alloc_multipath(ti);
 832         if (!m) {
 833                 ti->error = "can't allocate multipath";
 834                 return -EINVAL;
 835         }
 836
 837         r = parse_features(&as, m);
 838         if (r)
 839                 goto bad;
 840
 841         r = parse_hw_handler(&as, m);
 842         if (r)
 843                 goto bad;
 844
 845         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 846         if (r)
 847                 goto bad;
 848
 849         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 850         if (r)
 851                 goto bad;
 852
 853         /* parse the priority groups */
 854         while (as.argc) {
 855                 struct priority_group *pg;
 856
 857                 pg = parse_priority_group(&as, m);
 858                 if (IS_ERR(pg)) {
 859                         r = PTR_ERR(pg);
 860                         goto bad;
 861                 }
 862
 863                 m->nr_valid_paths += pg->nr_pgpaths;
 864                 list_add_tail(&pg->list, &m->priority_groups);
 865                 pg_count++;
 866                 pg->pg_num = pg_count;
 867                 if (!--next_pg_num)
 868                         m->next_pg = pg;
 869         }
 870
 871         if (pg_count != m->nr_priority_groups) {
 872                 ti->error = "priority group count mismatch";
 873                 r = -EINVAL;
 874                 goto bad;
 875         }
 876
 877         ti->num_flush_requests = 1;
 878
 879         return 0;
 880
 881  bad:
 882         free_multipath(m);
 883         return r;
 884 }
 885
 886 static void multipath_dtr(struct dm_target *ti)
 887 {
 888         struct multipath *m = (struct multipath *) ti->private;
 889
 890         flush_workqueue(kmpath_handlerd);
 891         flush_workqueue(kmultipathd);
 892         flush_scheduled_work();
 893         free_multipath(m);
 894 }
 895
 896 /*
 897  * Map cloned requests
 898  */
 899 static int multipath_map(struct dm_target *ti, struct request *clone,
 900                          union map_info *map_context)
 901 {
 902         int r;
 903         struct dm_mpath_io *mpio;
 904         struct multipath *m = (struct multipath *) ti->private;
 905
 906         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 907         if (!mpio)
 908                 /* ENOMEM, requeue */
 909                 return DM_MAPIO_REQUEUE;
 910         memset(mpio, 0, sizeof(*mpio));
 911
 912         map_context->ptr = mpio;
 913         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 914         r = map_io(m, clone, mpio, 0);
 915         if (r < 0 || r == DM_MAPIO_REQUEUE)
 916                 mempool_free(mpio, m->mpio_pool);
 917
 918         return r;
 919 }
 920
 921 /*
 922  * Take a path out of use.
 923  */
 924 static int fail_path(struct pgpath *pgpath)
 925 {
 926         unsigned long flags;
 927         struct multipath *m = pgpath->pg->m;
 928
 929         spin_lock_irqsave(&m->lock, flags);
 930
 931         if (!pgpath->is_active)
 932                 goto out;
 933
 934         DMWARN("Failing path %s.", pgpath->path.dev->name);
 935
 936         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 937         pgpath->is_active = 0;
 938         pgpath->fail_count++;
 939
 940         m->nr_valid_paths--;
 941
 942         if (pgpath == m->current_pgpath)
 943                 m->current_pgpath = NULL;
 944
 945         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 946                       pgpath->path.dev->name, m->nr_valid_paths);
 947
 948         schedule_work(&m->trigger_event);
 949         queue_work(kmultipathd, &pgpath->deactivate_path);
 950
 951 out:
 952         spin_unlock_irqrestore(&m->lock, flags);
 953
 954         return 0;
 955 }
 956
 957 /*
 958  * Reinstate a previously-failed path
 959  */
 960 static int reinstate_path(struct pgpath *pgpath)
 961 {
 962         int r = 0;
 963         unsigned long flags;
 964         struct multipath *m = pgpath->pg->m;
 965
 966         spin_lock_irqsave(&m->lock, flags);
 967
 968         if (pgpath->is_active)
 969                 goto out;
 970
 971         if (!pgpath->pg->ps.type->reinstate_path) {
 972                 DMWARN("Reinstate path not supported by path selector %s",
 973                        pgpath->pg->ps.type->name);
 974                 r = -EINVAL;
 975                 goto out;
 976         }
 977
 978         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 979         if (r)
 980                 goto out;
 981
 982         pgpath->is_active = 1;
 983
 984         if (!m->nr_valid_paths++ && m->queue_size) {
 985                 m->current_pgpath = NULL;
 986                 queue_work(kmultipathd, &m->process_queued_ios);
 987         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
 988                 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
 989                         m->pg_init_in_progress++;
 990         }
 991
 992         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 993                       pgpath->path.dev->name, m->nr_valid_paths);
 994
 995         schedule_work(&m->trigger_event);
 996
 997 out:
 998         spin_unlock_irqrestore(&m->lock, flags);
 999
1000         return r;
1001 }
1002
1003 /*
1004  * Fail or reinstate all paths that match the provided struct dm_dev.
1005  */
1006 static int action_dev(struct multipath *m, struct dm_dev *dev,
1007                       action_fn action)
1008 {
1009         int r = 0;
1010         struct pgpath *pgpath;
1011         struct priority_group *pg;
1012
1013         list_for_each_entry(pg, &m->priority_groups, list) {
1014                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1015                         if (pgpath->path.dev == dev)
1016                                 r = action(pgpath);
1017                 }
1018         }
1019
1020         return r;
1021 }
1022
1023 /*
1024  * Temporarily try to avoid having to use the specified PG
1025  */
1026 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1027                       int bypassed)
1028 {
1029         unsigned long flags;
1030
1031         spin_lock_irqsave(&m->lock, flags);
1032
1033         pg->bypassed = bypassed;
1034         m->current_pgpath = NULL;
1035         m->current_pg = NULL;
1036
1037         spin_unlock_irqrestore(&m->lock, flags);
1038
1039         schedule_work(&m->trigger_event);
1040 }
1041
1042 /*
1043  * Switch to using the specified PG from the next I/O that gets mapped
1044  */
1045 static int switch_pg_num(struct multipath *m, const char *pgstr)
1046 {
1047         struct priority_group *pg;
1048         unsigned pgnum;
1049         unsigned long flags;
1050
1051         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1052             (pgnum > m->nr_priority_groups)) {
1053                 DMWARN("invalid PG number supplied to switch_pg_num");
1054                 return -EINVAL;
1055         }
1056
1057         spin_lock_irqsave(&m->lock, flags);
1058         list_for_each_entry(pg, &m->priority_groups, list) {
1059                 pg->bypassed = 0;
1060                 if (--pgnum)
1061                         continue;
1062
1063                 m->current_pgpath = NULL;
1064                 m->current_pg = NULL;
1065                 m->next_pg = pg;
1066         }
1067         spin_unlock_irqrestore(&m->lock, flags);
1068
1069         schedule_work(&m->trigger_event);
1070         return 0;
1071 }
1072
1073 /*
1074  * Set/clear bypassed status of a PG.
1075  * PGs are numbered upwards from 1 in the order they were declared.
1076  */
1077 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1078 {
1079         struct priority_group *pg;
1080         unsigned pgnum;
1081
1082         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1083             (pgnum > m->nr_priority_groups)) {
1084                 DMWARN("invalid PG number supplied to bypass_pg");
1085                 return -EINVAL;
1086         }
1087
1088         list_for_each_entry(pg, &m->priority_groups, list) {
1089                 if (!--pgnum)
1090                         break;
1091         }
1092
1093         bypass_pg(m, pg, bypassed);
1094         return 0;
1095 }
1096
1097 /*
1098  * Should we retry pg_init immediately?
1099  */
1100 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1101 {
1102         unsigned long flags;
1103         int limit_reached = 0;
1104
1105         spin_lock_irqsave(&m->lock, flags);
1106
1107         if (m->pg_init_count <= m->pg_init_retries)
1108                 m->pg_init_required = 1;
1109         else
1110                 limit_reached = 1;
1111
1112         spin_unlock_irqrestore(&m->lock, flags);
1113
1114         return limit_reached;
1115 }
1116
1117 static void pg_init_done(struct dm_path *path, int errors)
1118 {
1119         struct pgpath *pgpath = path_to_pgpath(path);
1120         struct priority_group *pg = pgpath->pg;
1121         struct multipath *m = pg->m;
1122         unsigned long flags;
1123
1124         /* device or driver problems */
1125         switch (errors) {
1126         case SCSI_DH_OK:
1127                 break;
1128         case SCSI_DH_NOSYS:
1129                 if (!m->hw_handler_name) {
1130                         errors = 0;
1131                         break;
1132                 }
1133                 DMERR("Cannot failover device because scsi_dh_%s was not "
1134                       "loaded.", m->hw_handler_name);
1135                 /*
1136                  * Fail path for now, so we do not ping pong
1137                  */
1138                 fail_path(pgpath);
1139                 break;
1140         case SCSI_DH_DEV_TEMP_BUSY:
1141                 /*
1142                  * Probably doing something like FW upgrade on the
1143                  * controller so try the other pg.
1144                  */
1145                 bypass_pg(m, pg, 1);
1146                 break;
1147         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1148         case SCSI_DH_RETRY:
1149         case SCSI_DH_IMM_RETRY:
1150         case SCSI_DH_RES_TEMP_UNAVAIL:
1151                 if (pg_init_limit_reached(m, pgpath))
1152                         fail_path(pgpath);
1153                 errors = 0;
1154                 break;
1155         default:
1156                 /*
1157                  * We probably do not want to fail the path for a device
1158                  * error, but this is what the old dm did. In future
1159                  * patches we can do more advanced handling.
1160                  */
1161                 fail_path(pgpath);
1162         }
1163
1164         spin_lock_irqsave(&m->lock, flags);
1165         if (errors) {
1166                 if (pgpath == m->current_pgpath) {
1167                         DMERR("Could not failover device. Error %d.", errors);
1168                         m->current_pgpath = NULL;
1169                         m->current_pg = NULL;
1170                 }
1171         } else if (!m->pg_init_required) {
1172                 m->queue_io = 0;
1173                 pg->bypassed = 0;
1174         }
1175
1176         m->pg_init_in_progress--;
1177         if (!m->pg_init_in_progress)
1178                 queue_work(kmultipathd, &m->process_queued_ios);
1179         spin_unlock_irqrestore(&m->lock, flags);
1180 }
1181
1182 static void activate_path(struct work_struct *work)
1183 {
1184         int ret;
1185         struct pgpath *pgpath =
1186                 container_of(work, struct pgpath, activate_path);
1187
1188         ret = scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev));
1189         pg_init_done(&pgpath->path, ret);
1190 }
1191
1192 /*
1193  * end_io handling
1194  */
1195 static int do_end_io(struct multipath *m, struct request *clone,
1196                      int error, struct dm_mpath_io *mpio)
1197 {
1198         /*
1199          * We don't queue any clone request inside the multipath target
1200          * during end I/O handling, since those clone requests don't have
1201          * bio clones.  If we queue them inside the multipath target,
1202          * we need to make bio clones, that requires memory allocation.
1203          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1204          *  don't have bio clones.)
1205          * Instead of queueing the clone request here, we queue the original
1206          * request into dm core, which will remake a clone request and
1207          * clone bios for it and resubmit it later.
1208          */
1209         int r = DM_ENDIO_REQUEUE;
1210         unsigned long flags;
1211
1212         if (!error && !clone->errors)
1213                 return 0;       /* I/O complete */
1214
1215         if (error == -EOPNOTSUPP)
1216                 return error;
1217
1218         if (mpio->pgpath)
1219                 fail_path(mpio->pgpath);
1220
1221         spin_lock_irqsave(&m->lock, flags);
1222         if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1223                 r = -EIO;
1224         spin_unlock_irqrestore(&m->lock, flags);
1225
1226         return r;
1227 }
1228
1229 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1230                             int error, union map_info *map_context)
1231 {
1232         struct multipath *m = ti->private;
1233         struct dm_mpath_io *mpio = map_context->ptr;
1234         struct pgpath *pgpath = mpio->pgpath;
1235         struct path_selector *ps;
1236         int r;
1237
1238         r  = do_end_io(m, clone, error, mpio);
1239         if (pgpath) {
1240                 ps = &pgpath->pg->ps;
1241                 if (ps->type->end_io)
1242                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1243         }
1244         mempool_free(mpio, m->mpio_pool);
1245
1246         return r;
1247 }
1248
1249 /*
1250  * Suspend can't complete until all the I/O is processed so if
1251  * the last path fails we must error any remaining I/O.
1252  * Note that if the freeze_bdev fails while suspending, the
1253  * queue_if_no_path state is lost - userspace should reset it.
1254  */
1255 static void multipath_presuspend(struct dm_target *ti)
1256 {
1257         struct multipath *m = (struct multipath *) ti->private;
1258
1259         queue_if_no_path(m, 0, 1);
1260 }
1261
1262 /*
1263  * Restore the queue_if_no_path setting.
1264  */
1265 static void multipath_resume(struct dm_target *ti)
1266 {
1267         struct multipath *m = (struct multipath *) ti->private;
1268         unsigned long flags;
1269
1270         spin_lock_irqsave(&m->lock, flags);
1271         m->queue_if_no_path = m->saved_queue_if_no_path;
1272         spin_unlock_irqrestore(&m->lock, flags);
1273 }
1274
1275 /*
1276  * Info output has the following format:
1277  * num_multipath_feature_args [multipath_feature_args]*
1278  * num_handler_status_args [handler_status_args]*
1279  * num_groups init_group_number
1280  *            [A|D|E num_ps_status_args [ps_status_args]*
1281  *             num_paths num_selector_args
1282  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1283  *
1284  * Table output has the following format (identical to the constructor string):
1285  * num_feature_args [features_args]*
1286  * num_handler_args hw_handler [hw_handler_args]*
1287  * num_groups init_group_number
1288  *     [priority selector-name num_ps_args [ps_args]*
1289  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1290  */
1291 static int multipath_status(struct dm_target *ti, status_type_t type,
1292                             char *result, unsigned int maxlen)
1293 {
1294         int sz = 0;
1295         unsigned long flags;
1296         struct multipath *m = (struct multipath *) ti->private;
1297         struct priority_group *pg;
1298         struct pgpath *p;
1299         unsigned pg_num;
1300         char state;
1301
1302         spin_lock_irqsave(&m->lock, flags);
1303
1304         /* Features */
1305         if (type == STATUSTYPE_INFO)
1306                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1307         else {
1308                 DMEMIT("%u ", m->queue_if_no_path +
1309                               (m->pg_init_retries > 0) * 2);
1310                 if (m->queue_if_no_path)
1311                         DMEMIT("queue_if_no_path ");
1312                 if (m->pg_init_retries)
1313                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1314         }
1315
1316         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1317                 DMEMIT("0 ");
1318         else
1319                 DMEMIT("1 %s ", m->hw_handler_name);
1320
1321         DMEMIT("%u ", m->nr_priority_groups);
1322
1323         if (m->next_pg)
1324                 pg_num = m->next_pg->pg_num;
1325         else if (m->current_pg)
1326                 pg_num = m->current_pg->pg_num;
1327         else
1328                         pg_num = 1;
1329
1330         DMEMIT("%u ", pg_num);
1331
1332         switch (type) {
1333         case STATUSTYPE_INFO:
1334                 list_for_each_entry(pg, &m->priority_groups, list) {
1335                         if (pg->bypassed)
1336                                 state = 'D';    /* Disabled */
1337                         else if (pg == m->current_pg)
1338                                 state = 'A';    /* Currently Active */
1339                         else
1340                                 state = 'E';    /* Enabled */
1341
1342                         DMEMIT("%c ", state);
1343
1344                         if (pg->ps.type->status)
1345                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1346                                                           result + sz,
1347                                                           maxlen - sz);
1348                         else
1349                                 DMEMIT("0 ");
1350
1351                         DMEMIT("%u %u ", pg->nr_pgpaths,
1352                                pg->ps.type->info_args);
1353
1354                         list_for_each_entry(p, &pg->pgpaths, list) {
1355                                 DMEMIT("%s %s %u ", p->path.dev->name,
1356                                        p->is_active ? "A" : "F",
1357                                        p->fail_count);
1358                                 if (pg->ps.type->status)
1359                                         sz += pg->ps.type->status(&pg->ps,
1360                                               &p->path, type, result + sz,
1361                                               maxlen - sz);
1362                         }
1363                 }
1364                 break;
1365
1366         case STATUSTYPE_TABLE:
1367                 list_for_each_entry(pg, &m->priority_groups, list) {
1368                         DMEMIT("%s ", pg->ps.type->name);
1369
1370                         if (pg->ps.type->status)
1371                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1372                                                           result + sz,
1373                                                           maxlen - sz);
1374                         else
1375                                 DMEMIT("0 ");
1376
1377                         DMEMIT("%u %u ", pg->nr_pgpaths,
1378                                pg->ps.type->table_args);
1379
1380                         list_for_each_entry(p, &pg->pgpaths, list) {
1381                                 DMEMIT("%s ", p->path.dev->name);
1382                                 if (pg->ps.type->status)
1383                                         sz += pg->ps.type->status(&pg->ps,
1384                                               &p->path, type, result + sz,
1385                                               maxlen - sz);
1386                         }
1387                 }
1388                 break;
1389         }
1390
1391         spin_unlock_irqrestore(&m->lock, flags);
1392
1393         return 0;
1394 }
1395
1396 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1397 {
1398         int r;
1399         struct dm_dev *dev;
1400         struct multipath *m = (struct multipath *) ti->private;
1401         action_fn action;
1402
1403         if (argc == 1) {
1404                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1405                         return queue_if_no_path(m, 1, 0);
1406                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1407                         return queue_if_no_path(m, 0, 0);
1408         }
1409
1410         if (argc != 2)
1411                 goto error;
1412
1413         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1414                 return bypass_pg_num(m, argv[1], 1);
1415         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1416                 return bypass_pg_num(m, argv[1], 0);
1417         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1418                 return switch_pg_num(m, argv[1]);
1419         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1420                 action = reinstate_path;
1421         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1422                 action = fail_path;
1423         else
1424                 goto error;
1425
1426         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1427                           dm_table_get_mode(ti->table), &dev);
1428         if (r) {
1429                 DMWARN("message: error getting device %s",
1430                        argv[1]);
1431                 return -EINVAL;
1432         }
1433
1434         r = action_dev(m, dev, action);
1435
1436         dm_put_device(ti, dev);
1437
1438         return r;
1439
1440 error:
1441         DMWARN("Unrecognised multipath message received.");
1442         return -EINVAL;
1443 }
1444
1445 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1446                            unsigned long arg)
1447 {
1448         struct multipath *m = (struct multipath *) ti->private;
1449         struct block_device *bdev = NULL;
1450         fmode_t mode = 0;
1451         unsigned long flags;
1452         int r = 0;
1453
1454         spin_lock_irqsave(&m->lock, flags);
1455
1456         if (!m->current_pgpath)
1457                 __choose_pgpath(m, 0);
1458
1459         if (m->current_pgpath) {
1460                 bdev = m->current_pgpath->path.dev->bdev;
1461                 mode = m->current_pgpath->path.dev->mode;
1462         }
1463
1464         if (m->queue_io)
1465                 r = -EAGAIN;
1466         else if (!bdev)
1467                 r = -EIO;
1468
1469         spin_unlock_irqrestore(&m->lock, flags);
1470
1471         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1472 }
1473
1474 static int multipath_iterate_devices(struct dm_target *ti,
1475                                      iterate_devices_callout_fn fn, void *data)
1476 {
1477         struct multipath *m = ti->private;
1478         struct priority_group *pg;
1479         struct pgpath *p;
1480         int ret = 0;
1481
1482         list_for_each_entry(pg, &m->priority_groups, list) {
1483                 list_for_each_entry(p, &pg->pgpaths, list) {
1484                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1485                         if (ret)
1486                                 goto out;
1487                 }
1488         }
1489
1490 out:
1491         return ret;
1492 }
1493
1494 static int __pgpath_busy(struct pgpath *pgpath)
1495 {
1496         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1497
1498         return dm_underlying_device_busy(q);
1499 }
1500
1501 /*
1502  * We return "busy", only when we can map I/Os but underlying devices
1503  * are busy (so even if we map I/Os now, the I/Os will wait on
1504  * the underlying queue).
1505  * In other words, if we want to kill I/Os or queue them inside us
1506  * due to map unavailability, we don't return "busy".  Otherwise,
1507  * dm core won't give us the I/Os and we can't do what we want.
1508  */
1509 static int multipath_busy(struct dm_target *ti)
1510 {
1511         int busy = 0, has_active = 0;
1512         struct multipath *m = ti->private;
1513         struct priority_group *pg;
1514         struct pgpath *pgpath;
1515         unsigned long flags;
1516
1517         spin_lock_irqsave(&m->lock, flags);
1518
1519         /* Guess which priority_group will be used at next mapping time */
1520         if (unlikely(!m->current_pgpath && m->next_pg))
1521                 pg = m->next_pg;
1522         else if (likely(m->current_pg))
1523                 pg = m->current_pg;
1524         else
1525                 /*
1526                  * We don't know which pg will be used at next mapping time.
1527                  * We don't call __choose_pgpath() here to avoid to trigger
1528                  * pg_init just by busy checking.
1529                  * So we don't know whether underlying devices we will be using
1530                  * at next mapping time are busy or not. Just try mapping.
1531                  */
1532                 goto out;
1533
1534         /*
1535          * If there is one non-busy active path at least, the path selector
1536          * will be able to select it. So we consider such a pg as not busy.
1537          */
1538         busy = 1;
1539         list_for_each_entry(pgpath, &pg->pgpaths, list)
1540                 if (pgpath->is_active) {
1541                         has_active = 1;
1542
1543                         if (!__pgpath_busy(pgpath)) {
1544                                 busy = 0;
1545                                 break;
1546                         }
1547                 }
1548
1549         if (!has_active)
1550                 /*
1551                  * No active path in this pg, so this pg won't be used and
1552                  * the current_pg will be changed at next mapping time.
1553                  * We need to try mapping to determine it.
1554                  */
1555                 busy = 0;
1556
1557 out:
1558         spin_unlock_irqrestore(&m->lock, flags);
1559
1560         return busy;
1561 }
1562
1563 /*-----------------------------------------------------------------
1564  * Module setup
1565  *---------------------------------------------------------------*/
1566 static struct target_type multipath_target = {
1567         .name = "multipath",
1568         .version = {1, 1, 0},
1569         .module = THIS_MODULE,
1570         .ctr = multipath_ctr,
1571         .dtr = multipath_dtr,
1572         .map_rq = multipath_map,
1573         .rq_end_io = multipath_end_io,
1574         .presuspend = multipath_presuspend,
1575         .resume = multipath_resume,
1576         .status = multipath_status,
1577         .message = multipath_message,
1578         .ioctl  = multipath_ioctl,
1579         .iterate_devices = multipath_iterate_devices,
1580         .busy = multipath_busy,
1581 };
1582
1583 static int __init dm_multipath_init(void)
1584 {
1585         int r;
1586
1587         /* allocate a slab for the dm_ios */
1588         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1589         if (!_mpio_cache)
1590                 return -ENOMEM;
1591
1592         r = dm_register_target(&multipath_target);
1593         if (r < 0) {
1594                 DMERR("register failed %d", r);
1595                 kmem_cache_destroy(_mpio_cache);
1596                 return -EINVAL;
1597         }
1598
1599         kmultipathd = create_workqueue("kmpathd");
1600         if (!kmultipathd) {
1601                 DMERR("failed to create workqueue kmpathd");
1602                 dm_unregister_target(&multipath_target);
1603                 kmem_cache_destroy(_mpio_cache);
1604                 return -ENOMEM;
1605         }
1606
1607         /*
1608          * A separate workqueue is used to handle the device handlers
1609          * to avoid overloading existing workqueue. Overloading the
1610          * old workqueue would also create a bottleneck in the
1611          * path of the storage hardware device activation.
1612          */
1613         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1614         if (!kmpath_handlerd) {
1615                 DMERR("failed to create workqueue kmpath_handlerd");
1616                 destroy_workqueue(kmultipathd);
1617                 dm_unregister_target(&multipath_target);
1618                 kmem_cache_destroy(_mpio_cache);
1619                 return -ENOMEM;
1620         }
1621
1622         DMINFO("version %u.%u.%u loaded",
1623                multipath_target.version[0], multipath_target.version[1],
1624                multipath_target.version[2]);
1625
1626         return r;
1627 }
1628
1629 static void __exit dm_multipath_exit(void)
1630 {
1631         destroy_workqueue(kmpath_handlerd);
1632         destroy_workqueue(kmultipathd);
1633
1634         dm_unregister_target(&multipath_target);
1635         kmem_cache_destroy(_mpio_cache);
1636 }
1637
1638 module_init(dm_multipath_init);
1639 module_exit(dm_multipath_exit);
1640
1641 MODULE_DESCRIPTION(DM_NAME " multipath target");
1642 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1643 MODULE_LICENSE("GPL");