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 #define DM_PG_INIT_DELAY_MSECS 2000
  27 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  28
  29 /* Path properties */
  30 struct pgpath {
  31         struct list_head list;
  32
  33         struct priority_group *pg;      /* Owning PG */
  34         unsigned is_active;             /* Path status */
  35         unsigned fail_count;            /* Cumulative failure count */
  36
  37         struct dm_path path;
  38         struct delayed_work activate_path;
  39 };
  40
  41 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  42
  43 /*
  44  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  45  * Each has a path selector which controls which path gets used.
  46  */
  47 struct priority_group {
  48         struct list_head list;
  49
  50         struct multipath *m;            /* Owning multipath instance */
  51         struct path_selector ps;
  52
  53         unsigned pg_num;                /* Reference number */
  54         unsigned bypassed;              /* Temporarily bypass this PG? */
  55
  56         unsigned nr_pgpaths;            /* Number of paths in PG */
  57         struct list_head pgpaths;
  58 };
  59
  60 #define FEATURE_NO_PARTITIONS 1
  61
  62 /* Multipath context */
  63 struct multipath {
  64         struct list_head list;
  65         struct dm_target *ti;
  66
  67         spinlock_t lock;
  68
  69         const char *hw_handler_name;
  70         char *hw_handler_params;
  71
  72         unsigned nr_priority_groups;
  73         struct list_head priority_groups;
  74
  75         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  76
  77         unsigned pg_init_required;      /* pg_init needs calling? */
  78         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  79         unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
  80
  81         unsigned nr_valid_paths;        /* Total number of usable paths */
  82         struct pgpath *current_pgpath;
  83         struct priority_group *current_pg;
  84         struct priority_group *next_pg; /* Switch to this PG if set */
  85         unsigned repeat_count;          /* I/Os left before calling PS again */
  86
  87         unsigned queue_io;              /* Must we queue all I/O? */
  88         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  89         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  90         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  91         unsigned pg_init_count;         /* Number of times pg_init called */
  92         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  93         unsigned features;              /* Additional selected features */
  94
  95         struct work_struct process_queued_ios;
  96         struct list_head queued_ios;
  97         unsigned queue_size;
  98
  99         struct work_struct trigger_event;
 100
 101         /*
 102          * We must use a mempool of dm_mpath_io structs so that we
 103          * can resubmit bios on error.
 104          */
 105         mempool_t *mpio_pool;
 106
 107         struct mutex work_mutex;
 108 };
 109
 110 /*
 111  * Context information attached to each bio we process.
 112  */
 113 struct dm_mpath_io {
 114         struct pgpath *pgpath;
 115         size_t nr_bytes;
 116 };
 117
 118 typedef int (*action_fn) (struct pgpath *pgpath);
 119
 120 #define MIN_IOS 256     /* Mempool size */
 121
 122 static struct kmem_cache *_mpio_cache;
 123
 124 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 125 static void process_queued_ios(struct work_struct *work);
 126 static void trigger_event(struct work_struct *work);
 127 static void activate_path(struct work_struct *work);
 128
 129
 130 /*-----------------------------------------------
 131  * Allocation routines
 132  *-----------------------------------------------*/
 133
 134 static struct pgpath *alloc_pgpath(void)
 135 {
 136         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 137
 138         if (pgpath) {
 139                 pgpath->is_active = 1;
 140                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 141         }
 142
 143         return pgpath;
 144 }
 145
 146 static void free_pgpath(struct pgpath *pgpath)
 147 {
 148         kfree(pgpath);
 149 }
 150
 151 static struct priority_group *alloc_priority_group(void)
 152 {
 153         struct priority_group *pg;
 154
 155         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 156
 157         if (pg)
 158                 INIT_LIST_HEAD(&pg->pgpaths);
 159
 160         return pg;
 161 }
 162
 163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 164 {
 165         struct pgpath *pgpath, *tmp;
 166
 167         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 168                 list_del(&pgpath->list);
 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                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 199                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 200                 INIT_WORK(&m->trigger_event, trigger_event);
 201                 init_waitqueue_head(&m->pg_init_wait);
 202                 mutex_init(&m->work_mutex);
 203                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 204                 if (!m->mpio_pool) {
 205                         kfree(m);
 206                         return NULL;
 207                 }
 208                 m->ti = ti;
 209                 ti->private = m;
 210         }
 211
 212         return m;
 213 }
 214
 215 static void free_multipath(struct multipath *m)
 216 {
 217         struct priority_group *pg, *tmp;
 218
 219         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 220                 list_del(&pg->list);
 221                 free_priority_group(pg, m->ti);
 222         }
 223
 224         kfree(m->hw_handler_name);
 225         kfree(m->hw_handler_params);
 226         mempool_destroy(m->mpio_pool);
 227         kfree(m);
 228 }
 229
 230
 231 /*-----------------------------------------------
 232  * Path selection
 233  *-----------------------------------------------*/
 234
 235 static void __pg_init_all_paths(struct multipath *m)
 236 {
 237         struct pgpath *pgpath;
 238         unsigned long pg_init_delay = 0;
 239
 240         m->pg_init_count++;
 241         m->pg_init_required = 0;
 242         if (m->pg_init_delay_retry)
 243                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 244                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 245         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 246                 /* Skip failed paths */
 247                 if (!pgpath->is_active)
 248                         continue;
 249                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 250                                        pg_init_delay))
 251                         m->pg_init_in_progress++;
 252         }
 253 }
 254
 255 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 256 {
 257         m->current_pg = pgpath->pg;
 258
 259         /* Must we initialise the PG first, and queue I/O till it's ready? */
 260         if (m->hw_handler_name) {
 261                 m->pg_init_required = 1;
 262                 m->queue_io = 1;
 263         } else {
 264                 m->pg_init_required = 0;
 265                 m->queue_io = 0;
 266         }
 267
 268         m->pg_init_count = 0;
 269 }
 270
 271 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 272                                size_t nr_bytes)
 273 {
 274         struct dm_path *path;
 275
 276         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 277         if (!path)
 278                 return -ENXIO;
 279
 280         m->current_pgpath = path_to_pgpath(path);
 281
 282         if (m->current_pg != pg)
 283                 __switch_pg(m, m->current_pgpath);
 284
 285         return 0;
 286 }
 287
 288 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 289 {
 290         struct priority_group *pg;
 291         unsigned bypassed = 1;
 292
 293         if (!m->nr_valid_paths)
 294                 goto failed;
 295
 296         /* Were we instructed to switch PG? */
 297         if (m->next_pg) {
 298                 pg = m->next_pg;
 299                 m->next_pg = NULL;
 300                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 301                         return;
 302         }
 303
 304         /* Don't change PG until it has no remaining paths */
 305         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 306                 return;
 307
 308         /*
 309          * Loop through priority groups until we find a valid path.
 310          * First time we skip PGs marked 'bypassed'.
 311          * Second time we only try the ones we skipped.
 312          */
 313         do {
 314                 list_for_each_entry(pg, &m->priority_groups, list) {
 315                         if (pg->bypassed == bypassed)
 316                                 continue;
 317                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 318                                 return;
 319                 }
 320         } while (bypassed--);
 321
 322 failed:
 323         m->current_pgpath = NULL;
 324         m->current_pg = NULL;
 325 }
 326
 327 /*
 328  * Check whether bios must be queued in the device-mapper core rather
 329  * than here in the target.
 330  *
 331  * m->lock must be held on entry.
 332  *
 333  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 334  * same value then we are not between multipath_presuspend()
 335  * and multipath_resume() calls and we have no need to check
 336  * for the DMF_NOFLUSH_SUSPENDING flag.
 337  */
 338 static int __must_push_back(struct multipath *m)
 339 {
 340         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 341                 dm_noflush_suspending(m->ti));
 342 }
 343
 344 static int map_io(struct multipath *m, struct request *clone,
 345                   struct dm_mpath_io *mpio, unsigned was_queued)
 346 {
 347         int r = DM_MAPIO_REMAPPED;
 348         size_t nr_bytes = blk_rq_bytes(clone);
 349         unsigned long flags;
 350         struct pgpath *pgpath;
 351         struct block_device *bdev;
 352
 353         spin_lock_irqsave(&m->lock, flags);
 354
 355         /* Do we need to select a new pgpath? */
 356         if (!m->current_pgpath ||
 357             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 358                 __choose_pgpath(m, nr_bytes);
 359
 360         pgpath = m->current_pgpath;
 361
 362         if (was_queued)
 363                 m->queue_size--;
 364
 365         if ((pgpath && m->queue_io) ||
 366             (!pgpath && m->queue_if_no_path)) {
 367                 /* Queue for the daemon to resubmit */
 368                 list_add_tail(&clone->queuelist, &m->queued_ios);
 369                 m->queue_size++;
 370                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 371                     !m->queue_io)
 372                         queue_work(kmultipathd, &m->process_queued_ios);
 373                 pgpath = NULL;
 374                 r = DM_MAPIO_SUBMITTED;
 375         } else if (pgpath) {
 376                 bdev = pgpath->path.dev->bdev;
 377                 clone->q = bdev_get_queue(bdev);
 378                 clone->rq_disk = bdev->bd_disk;
 379         } else if (__must_push_back(m))
 380                 r = DM_MAPIO_REQUEUE;
 381         else
 382                 r = -EIO;       /* Failed */
 383
 384         mpio->pgpath = pgpath;
 385         mpio->nr_bytes = nr_bytes;
 386
 387         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 388                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 389                                               nr_bytes);
 390
 391         spin_unlock_irqrestore(&m->lock, flags);
 392
 393         return r;
 394 }
 395
 396 /*
 397  * If we run out of usable paths, should we queue I/O or error it?
 398  */
 399 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 400                             unsigned save_old_value)
 401 {
 402         unsigned long flags;
 403
 404         spin_lock_irqsave(&m->lock, flags);
 405
 406         if (save_old_value)
 407                 m->saved_queue_if_no_path = m->queue_if_no_path;
 408         else
 409                 m->saved_queue_if_no_path = queue_if_no_path;
 410         m->queue_if_no_path = queue_if_no_path;
 411         if (!m->queue_if_no_path && m->queue_size)
 412                 queue_work(kmultipathd, &m->process_queued_ios);
 413
 414         spin_unlock_irqrestore(&m->lock, flags);
 415
 416         return 0;
 417 }
 418
 419 /*-----------------------------------------------------------------
 420  * The multipath daemon is responsible for resubmitting queued ios.
 421  *---------------------------------------------------------------*/
 422
 423 static void dispatch_queued_ios(struct multipath *m)
 424 {
 425         int r;
 426         unsigned long flags;
 427         struct dm_mpath_io *mpio;
 428         union map_info *info;
 429         struct request *clone, *n;
 430         LIST_HEAD(cl);
 431
 432         spin_lock_irqsave(&m->lock, flags);
 433         list_splice_init(&m->queued_ios, &cl);
 434         spin_unlock_irqrestore(&m->lock, flags);
 435
 436         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 437                 list_del_init(&clone->queuelist);
 438
 439                 info = dm_get_rq_mapinfo(clone);
 440                 mpio = info->ptr;
 441
 442                 r = map_io(m, clone, mpio, 1);
 443                 if (r < 0) {
 444                         mempool_free(mpio, m->mpio_pool);
 445                         dm_kill_unmapped_request(clone, r);
 446                 } else if (r == DM_MAPIO_REMAPPED)
 447                         dm_dispatch_request(clone);
 448                 else if (r == DM_MAPIO_REQUEUE) {
 449                         mempool_free(mpio, m->mpio_pool);
 450                         dm_requeue_unmapped_request(clone);
 451                 }
 452         }
 453 }
 454
 455 static void process_queued_ios(struct work_struct *work)
 456 {
 457         struct multipath *m =
 458                 container_of(work, struct multipath, process_queued_ios);
 459         struct pgpath *pgpath = NULL;
 460         unsigned must_queue = 1;
 461         unsigned long flags;
 462
 463         spin_lock_irqsave(&m->lock, flags);
 464
 465         if (!m->queue_size)
 466                 goto out;
 467
 468         if (!m->current_pgpath)
 469                 __choose_pgpath(m, 0);
 470
 471         pgpath = m->current_pgpath;
 472
 473         if ((pgpath && !m->queue_io) ||
 474             (!pgpath && !m->queue_if_no_path))
 475                 must_queue = 0;
 476
 477         if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
 478                 __pg_init_all_paths(m);
 479
 480 out:
 481         spin_unlock_irqrestore(&m->lock, flags);
 482         if (!must_queue)
 483                 dispatch_queued_ios(m);
 484 }
 485
 486 /*
 487  * An event is triggered whenever a path is taken out of use.
 488  * Includes path failure and PG bypass.
 489  */
 490 static void trigger_event(struct work_struct *work)
 491 {
 492         struct multipath *m =
 493                 container_of(work, struct multipath, trigger_event);
 494
 495         dm_table_event(m->ti->table);
 496 }
 497
 498 /*-----------------------------------------------------------------
 499  * Constructor/argument parsing:
 500  * <#multipath feature args> [<arg>]*
 501  * <#hw_handler args> [hw_handler [<arg>]*]
 502  * <#priority groups>
 503  * <initial priority group>
 504  *     [<selector> <#selector args> [<arg>]*
 505  *      <#paths> <#per-path selector args>
 506  *         [<path> [<arg>]* ]+ ]+
 507  *---------------------------------------------------------------*/
 508 struct param {
 509         unsigned min;
 510         unsigned max;
 511         char *error;
 512 };
 513
 514 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 515 {
 516         if (!str ||
 517             (sscanf(str, "%u", v) != 1) ||
 518             (*v < param->min) ||
 519             (*v > param->max)) {
 520                 *error = param->error;
 521                 return -EINVAL;
 522         }
 523
 524         return 0;
 525 }
 526
 527 struct arg_set {
 528         unsigned argc;
 529         char **argv;
 530 };
 531
 532 static char *shift(struct arg_set *as)
 533 {
 534         char *r;
 535
 536         if (as->argc) {
 537                 as->argc--;
 538                 r = *as->argv;
 539                 as->argv++;
 540                 return r;
 541         }
 542
 543         return NULL;
 544 }
 545
 546 static void consume(struct arg_set *as, unsigned n)
 547 {
 548         BUG_ON (as->argc < n);
 549         as->argc -= n;
 550         as->argv += n;
 551 }
 552
 553 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 554                                struct dm_target *ti)
 555 {
 556         int r;
 557         struct path_selector_type *pst;
 558         unsigned ps_argc;
 559
 560         static struct param _params[] = {
 561                 {0, 1024, "invalid number of path selector args"},
 562         };
 563
 564         pst = dm_get_path_selector(shift(as));
 565         if (!pst) {
 566                 ti->error = "unknown path selector type";
 567                 return -EINVAL;
 568         }
 569
 570         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 571         if (r) {
 572                 dm_put_path_selector(pst);
 573                 return -EINVAL;
 574         }
 575
 576         if (ps_argc > as->argc) {
 577                 dm_put_path_selector(pst);
 578                 ti->error = "not enough arguments for path selector";
 579                 return -EINVAL;
 580         }
 581
 582         r = pst->create(&pg->ps, ps_argc, as->argv);
 583         if (r) {
 584                 dm_put_path_selector(pst);
 585                 ti->error = "path selector constructor failed";
 586                 return r;
 587         }
 588
 589         pg->ps.type = pst;
 590         consume(as, ps_argc);
 591
 592         return 0;
 593 }
 594
 595 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 596                                struct dm_target *ti)
 597 {
 598         int r;
 599         struct pgpath *p;
 600         struct multipath *m = ti->private;
 601
 602         /* we need at least a path arg */
 603         if (as->argc < 1) {
 604                 ti->error = "no device given";
 605                 return ERR_PTR(-EINVAL);
 606         }
 607
 608         p = alloc_pgpath();
 609         if (!p)
 610                 return ERR_PTR(-ENOMEM);
 611
 612         r = dm_get_device(ti, shift(as), dm_table_get_mode(ti->table),
 613                           &p->path.dev);
 614         if (r) {
 615                 ti->error = "error getting device";
 616                 goto bad;
 617         }
 618
 619         if (m->hw_handler_name) {
 620                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 621
 622                 r = scsi_dh_attach(q, m->hw_handler_name);
 623                 if (r == -EBUSY) {
 624                         /*
 625                          * Already attached to different hw_handler,
 626                          * try to reattach with correct one.
 627                          */
 628                         scsi_dh_detach(q);
 629                         r = scsi_dh_attach(q, m->hw_handler_name);
 630                 }
 631
 632                 if (r < 0) {
 633                         ti->error = "error attaching hardware handler";
 634                         dm_put_device(ti, p->path.dev);
 635                         goto bad;
 636                 }
 637
 638                 if (m->hw_handler_params) {
 639                         r = scsi_dh_set_params(q, m->hw_handler_params);
 640                         if (r < 0) {
 641                                 ti->error = "unable to set hardware "
 642                                                         "handler parameters";
 643                                 scsi_dh_detach(q);
 644                                 dm_put_device(ti, p->path.dev);
 645                                 goto bad;
 646                         }
 647                 }
 648         }
 649
 650         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 651         if (r) {
 652                 dm_put_device(ti, p->path.dev);
 653                 goto bad;
 654         }
 655
 656         return p;
 657
 658  bad:
 659         free_pgpath(p);
 660         return ERR_PTR(r);
 661 }
 662
 663 static struct priority_group *parse_priority_group(struct arg_set *as,
 664                                                    struct multipath *m)
 665 {
 666         static struct param _params[] = {
 667                 {1, 1024, "invalid number of paths"},
 668                 {0, 1024, "invalid number of selector args"}
 669         };
 670
 671         int r;
 672         unsigned i, nr_selector_args, nr_params;
 673         struct priority_group *pg;
 674         struct dm_target *ti = m->ti;
 675
 676         if (as->argc < 2) {
 677                 as->argc = 0;
 678                 ti->error = "not enough priority group arguments";
 679                 return ERR_PTR(-EINVAL);
 680         }
 681
 682         pg = alloc_priority_group();
 683         if (!pg) {
 684                 ti->error = "couldn't allocate priority group";
 685                 return ERR_PTR(-ENOMEM);
 686         }
 687         pg->m = m;
 688
 689         r = parse_path_selector(as, pg, ti);
 690         if (r)
 691                 goto bad;
 692
 693         /*
 694          * read the paths
 695          */
 696         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 697         if (r)
 698                 goto bad;
 699
 700         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 701         if (r)
 702                 goto bad;
 703
 704         nr_params = 1 + nr_selector_args;
 705         for (i = 0; i < pg->nr_pgpaths; i++) {
 706                 struct pgpath *pgpath;
 707                 struct arg_set path_args;
 708
 709                 if (as->argc < nr_params) {
 710                         ti->error = "not enough path parameters";
 711                         r = -EINVAL;
 712                         goto bad;
 713                 }
 714
 715                 path_args.argc = nr_params;
 716                 path_args.argv = as->argv;
 717
 718                 pgpath = parse_path(&path_args, &pg->ps, ti);
 719                 if (IS_ERR(pgpath)) {
 720                         r = PTR_ERR(pgpath);
 721                         goto bad;
 722                 }
 723
 724                 pgpath->pg = pg;
 725                 list_add_tail(&pgpath->list, &pg->pgpaths);
 726                 consume(as, nr_params);
 727         }
 728
 729         return pg;
 730
 731  bad:
 732         free_priority_group(pg, ti);
 733         return ERR_PTR(r);
 734 }
 735
 736 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 737 {
 738         unsigned hw_argc;
 739         int ret;
 740         struct dm_target *ti = m->ti;
 741
 742         static struct param _params[] = {
 743                 {0, 1024, "invalid number of hardware handler args"},
 744         };
 745
 746         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 747                 return -EINVAL;
 748
 749         if (!hw_argc)
 750                 return 0;
 751
 752         if (hw_argc > as->argc) {
 753                 ti->error = "not enough arguments for hardware handler";
 754                 return -EINVAL;
 755         }
 756
 757         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 758         request_module("scsi_dh_%s", m->hw_handler_name);
 759         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 760                 ti->error = "unknown hardware handler type";
 761                 ret = -EINVAL;
 762                 goto fail;
 763         }
 764
 765         if (hw_argc > 1) {
 766                 char *p;
 767                 int i, j, len = 4;
 768
 769                 for (i = 0; i <= hw_argc - 2; i++)
 770                         len += strlen(as->argv[i]) + 1;
 771                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 772                 if (!p) {
 773                         ti->error = "memory allocation failed";
 774                         ret = -ENOMEM;
 775                         goto fail;
 776                 }
 777                 j = sprintf(p, "%d", hw_argc - 1);
 778                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 779                         j = sprintf(p, "%s", as->argv[i]);
 780         }
 781         consume(as, hw_argc - 1);
 782
 783         return 0;
 784 fail:
 785         kfree(m->hw_handler_name);
 786         m->hw_handler_name = NULL;
 787         return ret;
 788 }
 789
 790 static int parse_features(struct arg_set *as, struct multipath *m)
 791 {
 792         int r;
 793         unsigned argc;
 794         struct dm_target *ti = m->ti;
 795         const char *param_name;
 796
 797         static struct param _params[] = {
 798                 {0, 5, "invalid number of feature args"},
 799                 {1, 50, "pg_init_retries must be between 1 and 50"},
 800                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 801         };
 802
 803         r = read_param(_params, shift(as), &argc, &ti->error);
 804         if (r)
 805                 return -EINVAL;
 806
 807         if (!argc)
 808                 return 0;
 809
 810         do {
 811                 param_name = shift(as);
 812                 argc--;
 813
 814                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 815                         r = queue_if_no_path(m, 1, 0);
 816                         continue;
 817                 }
 818
 819                 if (!strnicmp(param_name, MESG_STR("no_partitions"))) {
 820                         m->features |= FEATURE_NO_PARTITIONS;
 821                         continue;
 822                 }
 823                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 824                     (argc >= 1)) {
 825                         r = read_param(_params + 1, shift(as),
 826                                        &m->pg_init_retries, &ti->error);
 827                         argc--;
 828                         continue;
 829                 }
 830
 831                 if (!strnicmp(param_name, MESG_STR("pg_init_delay_msecs")) &&
 832                     (argc >= 1)) {
 833                         r = read_param(_params + 2, shift(as),
 834                                        &m->pg_init_delay_msecs, &ti->error);
 835                         argc--;
 836                         continue;
 837                 }
 838
 839                 ti->error = "Unrecognised multipath feature request";
 840                 r = -EINVAL;
 841         } while (argc && !r);
 842
 843         return r;
 844 }
 845
 846 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 847                          char **argv)
 848 {
 849         /* target parameters */
 850         static struct param _params[] = {
 851                 {0, 1024, "invalid number of priority groups"},
 852                 {0, 1024, "invalid initial priority group number"},
 853         };
 854
 855         int r;
 856         struct multipath *m;
 857         struct arg_set as;
 858         unsigned pg_count = 0;
 859         unsigned next_pg_num;
 860
 861         as.argc = argc;
 862         as.argv = argv;
 863
 864         m = alloc_multipath(ti);
 865         if (!m) {
 866                 ti->error = "can't allocate multipath";
 867                 return -EINVAL;
 868         }
 869
 870         r = parse_features(&as, m);
 871         if (r)
 872                 goto bad;
 873
 874         r = parse_hw_handler(&as, m);
 875         if (r)
 876                 goto bad;
 877
 878         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 879         if (r)
 880                 goto bad;
 881
 882         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 883         if (r)
 884                 goto bad;
 885
 886         if ((!m->nr_priority_groups && next_pg_num) ||
 887             (m->nr_priority_groups && !next_pg_num)) {
 888                 ti->error = "invalid initial priority group";
 889                 r = -EINVAL;
 890                 goto bad;
 891         }
 892
 893         /* parse the priority groups */
 894         while (as.argc) {
 895                 struct priority_group *pg;
 896
 897                 pg = parse_priority_group(&as, m);
 898                 if (IS_ERR(pg)) {
 899                         r = PTR_ERR(pg);
 900                         goto bad;
 901                 }
 902
 903                 m->nr_valid_paths += pg->nr_pgpaths;
 904                 list_add_tail(&pg->list, &m->priority_groups);
 905                 pg_count++;
 906                 pg->pg_num = pg_count;
 907                 if (!--next_pg_num)
 908                         m->next_pg = pg;
 909         }
 910
 911         if (pg_count != m->nr_priority_groups) {
 912                 ti->error = "priority group count mismatch";
 913                 r = -EINVAL;
 914                 goto bad;
 915         }
 916
 917         ti->num_flush_requests = 1;
 918         ti->num_discard_requests = 1;
 919
 920         return 0;
 921
 922  bad:
 923         free_multipath(m);
 924         return r;
 925 }
 926
 927 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 928 {
 929         DECLARE_WAITQUEUE(wait, current);
 930         unsigned long flags;
 931
 932         add_wait_queue(&m->pg_init_wait, &wait);
 933
 934         while (1) {
 935                 set_current_state(TASK_UNINTERRUPTIBLE);
 936
 937                 spin_lock_irqsave(&m->lock, flags);
 938                 if (!m->pg_init_in_progress) {
 939                         spin_unlock_irqrestore(&m->lock, flags);
 940                         break;
 941                 }
 942                 spin_unlock_irqrestore(&m->lock, flags);
 943
 944                 io_schedule();
 945         }
 946         set_current_state(TASK_RUNNING);
 947
 948         remove_wait_queue(&m->pg_init_wait, &wait);
 949 }
 950
 951 static void flush_multipath_work(struct multipath *m)
 952 {
 953         flush_workqueue(kmpath_handlerd);
 954         multipath_wait_for_pg_init_completion(m);
 955         flush_workqueue(kmultipathd);
 956         flush_work_sync(&m->trigger_event);
 957 }
 958
 959 static void multipath_dtr(struct dm_target *ti)
 960 {
 961         struct multipath *m = ti->private;
 962
 963         flush_multipath_work(m);
 964         free_multipath(m);
 965 }
 966
 967 /*
 968  * Map cloned requests
 969  */
 970 static int multipath_map(struct dm_target *ti, struct request *clone,
 971                          union map_info *map_context)
 972 {
 973         int r;
 974         struct dm_mpath_io *mpio;
 975         struct multipath *m = (struct multipath *) ti->private;
 976
 977         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 978         if (!mpio)
 979                 /* ENOMEM, requeue */
 980                 return DM_MAPIO_REQUEUE;
 981         memset(mpio, 0, sizeof(*mpio));
 982
 983         map_context->ptr = mpio;
 984         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 985         r = map_io(m, clone, mpio, 0);
 986         if (r < 0 || r == DM_MAPIO_REQUEUE)
 987                 mempool_free(mpio, m->mpio_pool);
 988
 989         return r;
 990 }
 991
 992 /*
 993  * Take a path out of use.
 994  */
 995 static int fail_path(struct pgpath *pgpath)
 996 {
 997         unsigned long flags;
 998         struct multipath *m = pgpath->pg->m;
 999
1000         spin_lock_irqsave(&m->lock, flags);
1001
1002         if (!pgpath->is_active)
1003                 goto out;
1004
1005         DMWARN("Failing path %s.", pgpath->path.dev->name);
1006
1007         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1008         pgpath->is_active = 0;
1009         pgpath->fail_count++;
1010
1011         m->nr_valid_paths--;
1012
1013         if (pgpath == m->current_pgpath)
1014                 m->current_pgpath = NULL;
1015
1016         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1017                       pgpath->path.dev->name, m->nr_valid_paths);
1018
1019         schedule_work(&m->trigger_event);
1020
1021 out:
1022         spin_unlock_irqrestore(&m->lock, flags);
1023
1024         return 0;
1025 }
1026
1027 /*
1028  * Reinstate a previously-failed path
1029  */
1030 static int reinstate_path(struct pgpath *pgpath)
1031 {
1032         int r = 0;
1033         unsigned long flags;
1034         struct multipath *m = pgpath->pg->m;
1035
1036         spin_lock_irqsave(&m->lock, flags);
1037
1038         if (pgpath->is_active)
1039                 goto out;
1040
1041         if (!pgpath->pg->ps.type->reinstate_path) {
1042                 DMWARN("Reinstate path not supported by path selector %s",
1043                        pgpath->pg->ps.type->name);
1044                 r = -EINVAL;
1045                 goto out;
1046         }
1047
1048         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1049         if (r)
1050                 goto out;
1051
1052         pgpath->is_active = 1;
1053
1054         if (!m->nr_valid_paths++ && m->queue_size) {
1055                 m->current_pgpath = NULL;
1056                 queue_work(kmultipathd, &m->process_queued_ios);
1057         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1058                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1059                         m->pg_init_in_progress++;
1060         }
1061
1062         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1063                       pgpath->path.dev->name, m->nr_valid_paths);
1064
1065         schedule_work(&m->trigger_event);
1066
1067 out:
1068         spin_unlock_irqrestore(&m->lock, flags);
1069
1070         return r;
1071 }
1072
1073 /*
1074  * Fail or reinstate all paths that match the provided struct dm_dev.
1075  */
1076 static int action_dev(struct multipath *m, struct dm_dev *dev,
1077                       action_fn action)
1078 {
1079         int r = -EINVAL;
1080         struct pgpath *pgpath;
1081         struct priority_group *pg;
1082
1083         list_for_each_entry(pg, &m->priority_groups, list) {
1084                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1085                         if (pgpath->path.dev == dev)
1086                                 r = action(pgpath);
1087                 }
1088         }
1089
1090         return r;
1091 }
1092
1093 /*
1094  * Temporarily try to avoid having to use the specified PG
1095  */
1096 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1097                       int bypassed)
1098 {
1099         unsigned long flags;
1100
1101         spin_lock_irqsave(&m->lock, flags);
1102
1103         pg->bypassed = bypassed;
1104         m->current_pgpath = NULL;
1105         m->current_pg = NULL;
1106
1107         spin_unlock_irqrestore(&m->lock, flags);
1108
1109         schedule_work(&m->trigger_event);
1110 }
1111
1112 /*
1113  * Switch to using the specified PG from the next I/O that gets mapped
1114  */
1115 static int switch_pg_num(struct multipath *m, const char *pgstr)
1116 {
1117         struct priority_group *pg;
1118         unsigned pgnum;
1119         unsigned long flags;
1120
1121         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1122             (pgnum > m->nr_priority_groups)) {
1123                 DMWARN("invalid PG number supplied to switch_pg_num");
1124                 return -EINVAL;
1125         }
1126
1127         spin_lock_irqsave(&m->lock, flags);
1128         list_for_each_entry(pg, &m->priority_groups, list) {
1129                 pg->bypassed = 0;
1130                 if (--pgnum)
1131                         continue;
1132
1133                 m->current_pgpath = NULL;
1134                 m->current_pg = NULL;
1135                 m->next_pg = pg;
1136         }
1137         spin_unlock_irqrestore(&m->lock, flags);
1138
1139         schedule_work(&m->trigger_event);
1140         return 0;
1141 }
1142
1143 /*
1144  * Set/clear bypassed status of a PG.
1145  * PGs are numbered upwards from 1 in the order they were declared.
1146  */
1147 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1148 {
1149         struct priority_group *pg;
1150         unsigned pgnum;
1151
1152         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1153             (pgnum > m->nr_priority_groups)) {
1154                 DMWARN("invalid PG number supplied to bypass_pg");
1155                 return -EINVAL;
1156         }
1157
1158         list_for_each_entry(pg, &m->priority_groups, list) {
1159                 if (!--pgnum)
1160                         break;
1161         }
1162
1163         bypass_pg(m, pg, bypassed);
1164         return 0;
1165 }
1166
1167 /*
1168  * Should we retry pg_init immediately?
1169  */
1170 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1171 {
1172         unsigned long flags;
1173         int limit_reached = 0;
1174
1175         spin_lock_irqsave(&m->lock, flags);
1176
1177         if (m->pg_init_count <= m->pg_init_retries)
1178                 m->pg_init_required = 1;
1179         else
1180                 limit_reached = 1;
1181
1182         spin_unlock_irqrestore(&m->lock, flags);
1183
1184         return limit_reached;
1185 }
1186
1187 static void pg_init_done(void *data, int errors)
1188 {
1189         struct pgpath *pgpath = data;
1190         struct priority_group *pg = pgpath->pg;
1191         struct multipath *m = pg->m;
1192         unsigned long flags;
1193         unsigned delay_retry = 0;
1194
1195         /* device or driver problems */
1196         switch (errors) {
1197         case SCSI_DH_OK:
1198                 break;
1199         case SCSI_DH_NOSYS:
1200                 if (!m->hw_handler_name) {
1201                         errors = 0;
1202                         break;
1203                 }
1204                 DMERR("Count not failover device %s: Handler scsi_dh_%s "
1205                       "was not loaded.", pgpath->path.dev->name,
1206                       m->hw_handler_name);
1207                 /*
1208                  * Fail path for now, so we do not ping pong
1209                  */
1210                 fail_path(pgpath);
1211                 break;
1212         case SCSI_DH_DEV_TEMP_BUSY:
1213                 /*
1214                  * Probably doing something like FW upgrade on the
1215                  * controller so try the other pg.
1216                  */
1217                 bypass_pg(m, pg, 1);
1218                 break;
1219         case SCSI_DH_DEV_OFFLINED:
1220                 DMWARN("Device %s offlined.", pgpath->path.dev->name);
1221                 errors = 0;
1222                 break;
1223         case SCSI_DH_RETRY:
1224                 /* Wait before retrying. */
1225                 delay_retry = 1;
1226         case SCSI_DH_IMM_RETRY:
1227         case SCSI_DH_RES_TEMP_UNAVAIL:
1228                 if (pg_init_limit_reached(m, pgpath))
1229                         fail_path(pgpath);
1230                 errors = 0;
1231                 break;
1232         default:
1233                 /*
1234                  * We probably do not want to fail the path for a device
1235                  * error, but this is what the old dm did. In future
1236                  * patches we can do more advanced handling.
1237                  */
1238                 fail_path(pgpath);
1239         }
1240
1241         spin_lock_irqsave(&m->lock, flags);
1242         if (errors) {
1243                 if (pgpath == m->current_pgpath) {
1244                         DMERR("Could not failover device %s, error %d.",
1245                               pgpath->path.dev->name, errors);
1246                         m->current_pgpath = NULL;
1247                         m->current_pg = NULL;
1248                 }
1249         } else if (!m->pg_init_required)
1250                 pg->bypassed = 0;
1251
1252         if (--m->pg_init_in_progress)
1253                 /* Activations of other paths are still on going */
1254                 goto out;
1255
1256         if (!m->pg_init_required)
1257                 m->queue_io = 0;
1258
1259         m->pg_init_delay_retry = delay_retry;
1260         queue_work(kmultipathd, &m->process_queued_ios);
1261
1262         /*
1263          * Wake up any thread waiting to suspend.
1264          */
1265         wake_up(&m->pg_init_wait);
1266
1267 out:
1268         spin_unlock_irqrestore(&m->lock, flags);
1269 }
1270
1271 static void activate_path(struct work_struct *work)
1272 {
1273         struct pgpath *pgpath =
1274                 container_of(work, struct pgpath, activate_path.work);
1275
1276         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1277                                 pg_init_done, pgpath);
1278 }
1279
1280 /*
1281  * end_io handling
1282  */
1283 static int do_end_io(struct multipath *m, struct request *clone,
1284                      int error, struct dm_mpath_io *mpio)
1285 {
1286         /*
1287          * We don't queue any clone request inside the multipath target
1288          * during end I/O handling, since those clone requests don't have
1289          * bio clones.  If we queue them inside the multipath target,
1290          * we need to make bio clones, that requires memory allocation.
1291          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1292          *  don't have bio clones.)
1293          * Instead of queueing the clone request here, we queue the original
1294          * request into dm core, which will remake a clone request and
1295          * clone bios for it and resubmit it later.
1296          */
1297         int r = DM_ENDIO_REQUEUE;
1298         unsigned long flags;
1299
1300         if (!error && !clone->errors)
1301                 return 0;       /* I/O complete */
1302
1303         if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1304                 return error;
1305
1306         if (mpio->pgpath)
1307                 fail_path(mpio->pgpath);
1308
1309         spin_lock_irqsave(&m->lock, flags);
1310         if (!m->nr_valid_paths) {
1311                 if (!m->queue_if_no_path) {
1312                         if (!__must_push_back(m))
1313                                 r = -EIO;
1314                 } else {
1315                         if (error == -EBADE)
1316                                 r = error;
1317                 }
1318         }
1319         spin_unlock_irqrestore(&m->lock, flags);
1320
1321         return r;
1322 }
1323
1324 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1325                             int error, union map_info *map_context)
1326 {
1327         struct multipath *m = ti->private;
1328         struct dm_mpath_io *mpio = map_context->ptr;
1329         struct pgpath *pgpath = mpio->pgpath;
1330         struct path_selector *ps;
1331         int r;
1332
1333         r  = do_end_io(m, clone, error, mpio);
1334         if (pgpath) {
1335                 ps = &pgpath->pg->ps;
1336                 if (ps->type->end_io)
1337                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1338         }
1339         mempool_free(mpio, m->mpio_pool);
1340
1341         return r;
1342 }
1343
1344 /*
1345  * Suspend can't complete until all the I/O is processed so if
1346  * the last path fails we must error any remaining I/O.
1347  * Note that if the freeze_bdev fails while suspending, the
1348  * queue_if_no_path state is lost - userspace should reset it.
1349  */
1350 static void multipath_presuspend(struct dm_target *ti)
1351 {
1352         struct multipath *m = (struct multipath *) ti->private;
1353
1354         queue_if_no_path(m, 0, 1);
1355 }
1356
1357 static void multipath_postsuspend(struct dm_target *ti)
1358 {
1359         struct multipath *m = ti->private;
1360
1361         mutex_lock(&m->work_mutex);
1362         flush_multipath_work(m);
1363         mutex_unlock(&m->work_mutex);
1364 }
1365
1366 /*
1367  * Restore the queue_if_no_path setting.
1368  */
1369 static void multipath_resume(struct dm_target *ti)
1370 {
1371         struct multipath *m = (struct multipath *) ti->private;
1372         unsigned long flags;
1373
1374         spin_lock_irqsave(&m->lock, flags);
1375         m->queue_if_no_path = m->saved_queue_if_no_path;
1376         spin_unlock_irqrestore(&m->lock, flags);
1377 }
1378
1379 /*
1380  * Info output has the following format:
1381  * num_multipath_feature_args [multipath_feature_args]*
1382  * num_handler_status_args [handler_status_args]*
1383  * num_groups init_group_number
1384  *            [A|D|E num_ps_status_args [ps_status_args]*
1385  *             num_paths num_selector_args
1386  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1387  *
1388  * Table output has the following format (identical to the constructor string):
1389  * num_feature_args [features_args]*
1390  * num_handler_args hw_handler [hw_handler_args]*
1391  * num_groups init_group_number
1392  *     [priority selector-name num_ps_args [ps_args]*
1393  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1394  */
1395 static int multipath_status(struct dm_target *ti, status_type_t type,
1396                             char *result, unsigned int maxlen)
1397 {
1398         int sz = 0;
1399         unsigned long flags;
1400         struct multipath *m = (struct multipath *) ti->private;
1401         struct priority_group *pg;
1402         struct pgpath *p;
1403         unsigned pg_num;
1404         char state;
1405
1406         spin_lock_irqsave(&m->lock, flags);
1407
1408         /* Features */
1409         if (type == STATUSTYPE_INFO)
1410                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1411         else {
1412                 DMEMIT("%u ", m->queue_if_no_path +
1413                               (m->pg_init_retries > 0) * 2 +
1414                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1415                               (m->features & FEATURE_NO_PARTITIONS));
1416                 if (m->queue_if_no_path)
1417                         DMEMIT("queue_if_no_path ");
1418                 if (m->pg_init_retries)
1419                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1420                 if (m->features & FEATURE_NO_PARTITIONS)
1421                         DMEMIT("no_partitions ");
1422                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1423                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1424         }
1425
1426         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1427                 DMEMIT("0 ");
1428         else
1429                 DMEMIT("1 %s ", m->hw_handler_name);
1430
1431         DMEMIT("%u ", m->nr_priority_groups);
1432
1433         if (m->next_pg)
1434                 pg_num = m->next_pg->pg_num;
1435         else if (m->current_pg)
1436                 pg_num = m->current_pg->pg_num;
1437         else
1438                 pg_num = (m->nr_priority_groups ? 1 : 0);
1439
1440         DMEMIT("%u ", pg_num);
1441
1442         switch (type) {
1443         case STATUSTYPE_INFO:
1444                 list_for_each_entry(pg, &m->priority_groups, list) {
1445                         if (pg->bypassed)
1446                                 state = 'D';    /* Disabled */
1447                         else if (pg == m->current_pg)
1448                                 state = 'A';    /* Currently Active */
1449                         else
1450                                 state = 'E';    /* Enabled */
1451
1452                         DMEMIT("%c ", state);
1453
1454                         if (pg->ps.type->status)
1455                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1456                                                           result + sz,
1457                                                           maxlen - sz);
1458                         else
1459                                 DMEMIT("0 ");
1460
1461                         DMEMIT("%u %u ", pg->nr_pgpaths,
1462                                pg->ps.type->info_args);
1463
1464                         list_for_each_entry(p, &pg->pgpaths, list) {
1465                                 DMEMIT("%s %s %u ", p->path.dev->name,
1466                                        p->is_active ? "A" : "F",
1467                                        p->fail_count);
1468                                 if (pg->ps.type->status)
1469                                         sz += pg->ps.type->status(&pg->ps,
1470                                               &p->path, type, result + sz,
1471                                               maxlen - sz);
1472                         }
1473                 }
1474                 break;
1475
1476         case STATUSTYPE_TABLE:
1477                 list_for_each_entry(pg, &m->priority_groups, list) {
1478                         DMEMIT("%s ", pg->ps.type->name);
1479
1480                         if (pg->ps.type->status)
1481                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1482                                                           result + sz,
1483                                                           maxlen - sz);
1484                         else
1485                                 DMEMIT("0 ");
1486
1487                         DMEMIT("%u %u ", pg->nr_pgpaths,
1488                                pg->ps.type->table_args);
1489
1490                         list_for_each_entry(p, &pg->pgpaths, list) {
1491                                 DMEMIT("%s ", p->path.dev->name);
1492                                 if (pg->ps.type->status)
1493                                         sz += pg->ps.type->status(&pg->ps,
1494                                               &p->path, type, result + sz,
1495                                               maxlen - sz);
1496                         }
1497                 }
1498                 break;
1499         }
1500
1501         spin_unlock_irqrestore(&m->lock, flags);
1502
1503         return 0;
1504 }
1505
1506 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1507 {
1508         int r = -EINVAL;
1509         struct dm_dev *dev;
1510         struct multipath *m = (struct multipath *) ti->private;
1511         action_fn action;
1512
1513         mutex_lock(&m->work_mutex);
1514
1515         if (dm_suspended(ti)) {
1516                 r = -EBUSY;
1517                 goto out;
1518         }
1519
1520         if (argc == 1) {
1521                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
1522                         r = queue_if_no_path(m, 1, 0);
1523                         goto out;
1524                 } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
1525                         r = queue_if_no_path(m, 0, 0);
1526                         goto out;
1527                 }
1528         }
1529
1530         if (argc != 2) {
1531                 DMWARN("Unrecognised multipath message received.");
1532                 goto out;
1533         }
1534
1535         if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
1536                 r = bypass_pg_num(m, argv[1], 1);
1537                 goto out;
1538         } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
1539                 r = bypass_pg_num(m, argv[1], 0);
1540                 goto out;
1541         } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
1542                 r = switch_pg_num(m, argv[1]);
1543                 goto out;
1544         } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1545                 action = reinstate_path;
1546         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1547                 action = fail_path;
1548         else {
1549                 DMWARN("Unrecognised multipath message received.");
1550                 goto out;
1551         }
1552
1553         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1554         if (r) {
1555                 DMWARN("message: error getting device %s",
1556                        argv[1]);
1557                 goto out;
1558         }
1559
1560         r = action_dev(m, dev, action);
1561
1562         dm_put_device(ti, dev);
1563
1564 out:
1565         mutex_unlock(&m->work_mutex);
1566         return r;
1567 }
1568
1569 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1570                            unsigned long arg)
1571 {
1572         struct multipath *m = (struct multipath *) ti->private;
1573         struct block_device *bdev = NULL;
1574         fmode_t mode = 0;
1575         unsigned long flags;
1576         int r = 0;
1577
1578         spin_lock_irqsave(&m->lock, flags);
1579
1580         if (!m->current_pgpath)
1581                 __choose_pgpath(m, 0);
1582
1583         if (m->current_pgpath) {
1584                 bdev = m->current_pgpath->path.dev->bdev;
1585                 mode = m->current_pgpath->path.dev->mode;
1586         }
1587
1588         if (m->queue_io)
1589                 r = -EAGAIN;
1590         else if (!bdev)
1591                 r = -EIO;
1592
1593         spin_unlock_irqrestore(&m->lock, flags);
1594
1595         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1596 }
1597
1598 static int multipath_iterate_devices(struct dm_target *ti,
1599                                      iterate_devices_callout_fn fn, void *data)
1600 {
1601         struct multipath *m = ti->private;
1602         struct priority_group *pg;
1603         struct pgpath *p;
1604         int ret = 0;
1605
1606         list_for_each_entry(pg, &m->priority_groups, list) {
1607                 list_for_each_entry(p, &pg->pgpaths, list) {
1608                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1609                         if (ret)
1610                                 goto out;
1611                 }
1612         }
1613
1614 out:
1615         return ret;
1616 }
1617
1618 static int __pgpath_busy(struct pgpath *pgpath)
1619 {
1620         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1621
1622         return dm_underlying_device_busy(q);
1623 }
1624
1625 /*
1626  * We return "busy", only when we can map I/Os but underlying devices
1627  * are busy (so even if we map I/Os now, the I/Os will wait on
1628  * the underlying queue).
1629  * In other words, if we want to kill I/Os or queue them inside us
1630  * due to map unavailability, we don't return "busy".  Otherwise,
1631  * dm core won't give us the I/Os and we can't do what we want.
1632  */
1633 static int multipath_busy(struct dm_target *ti)
1634 {
1635         int busy = 0, has_active = 0;
1636         struct multipath *m = ti->private;
1637         struct priority_group *pg;
1638         struct pgpath *pgpath;
1639         unsigned long flags;
1640
1641         spin_lock_irqsave(&m->lock, flags);
1642
1643         /* Guess which priority_group will be used at next mapping time */
1644         if (unlikely(!m->current_pgpath && m->next_pg))
1645                 pg = m->next_pg;
1646         else if (likely(m->current_pg))
1647                 pg = m->current_pg;
1648         else
1649                 /*
1650                  * We don't know which pg will be used at next mapping time.
1651                  * We don't call __choose_pgpath() here to avoid to trigger
1652                  * pg_init just by busy checking.
1653                  * So we don't know whether underlying devices we will be using
1654                  * at next mapping time are busy or not. Just try mapping.
1655                  */
1656                 goto out;
1657
1658         /*
1659          * If there is one non-busy active path at least, the path selector
1660          * will be able to select it. So we consider such a pg as not busy.
1661          */
1662         busy = 1;
1663         list_for_each_entry(pgpath, &pg->pgpaths, list)
1664                 if (pgpath->is_active) {
1665                         has_active = 1;
1666
1667                         if (!__pgpath_busy(pgpath)) {
1668                                 busy = 0;
1669                                 break;
1670                         }
1671                 }
1672
1673         if (!has_active)
1674                 /*
1675                  * No active path in this pg, so this pg won't be used and
1676                  * the current_pg will be changed at next mapping time.
1677                  * We need to try mapping to determine it.
1678                  */
1679                 busy = 0;
1680
1681 out:
1682         spin_unlock_irqrestore(&m->lock, flags);
1683
1684         return busy;
1685 }
1686
1687 /*-----------------------------------------------------------------
1688  * Module setup
1689  *---------------------------------------------------------------*/
1690 static struct target_type multipath_target = {
1691         .name = "multipath",
1692         .version = {1, 3, 0},
1693         .module = THIS_MODULE,
1694         .ctr = multipath_ctr,
1695         .dtr = multipath_dtr,
1696         .map_rq = multipath_map,
1697         .rq_end_io = multipath_end_io,
1698         .presuspend = multipath_presuspend,
1699         .postsuspend = multipath_postsuspend,
1700         .resume = multipath_resume,
1701         .status = multipath_status,
1702         .message = multipath_message,
1703         .ioctl  = multipath_ioctl,
1704         .iterate_devices = multipath_iterate_devices,
1705         .busy = multipath_busy,
1706 };
1707
1708 static int __init dm_multipath_init(void)
1709 {
1710         int r;
1711
1712         /* allocate a slab for the dm_ios */
1713         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1714         if (!_mpio_cache)
1715                 return -ENOMEM;
1716
1717         r = dm_register_target(&multipath_target);
1718         if (r < 0) {
1719                 DMERR("register failed %d", r);
1720                 kmem_cache_destroy(_mpio_cache);
1721                 return -EINVAL;
1722         }
1723
1724         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1725         if (!kmultipathd) {
1726                 DMERR("failed to create workqueue kmpathd");
1727                 dm_unregister_target(&multipath_target);
1728                 kmem_cache_destroy(_mpio_cache);
1729                 return -ENOMEM;
1730         }
1731
1732         /*
1733          * A separate workqueue is used to handle the device handlers
1734          * to avoid overloading existing workqueue. Overloading the
1735          * old workqueue would also create a bottleneck in the
1736          * path of the storage hardware device activation.
1737          */
1738         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1739                                                   WQ_MEM_RECLAIM);
1740         if (!kmpath_handlerd) {
1741                 DMERR("failed to create workqueue kmpath_handlerd");
1742                 destroy_workqueue(kmultipathd);
1743                 dm_unregister_target(&multipath_target);
1744                 kmem_cache_destroy(_mpio_cache);
1745                 return -ENOMEM;
1746         }
1747
1748         DMINFO("version %u.%u.%u loaded",
1749                multipath_target.version[0], multipath_target.version[1],
1750                multipath_target.version[2]);
1751
1752         return r;
1753 }
1754
1755 static void __exit dm_multipath_exit(void)
1756 {
1757         destroy_workqueue(kmpath_handlerd);
1758         destroy_workqueue(kmultipathd);
1759
1760         dm_unregister_target(&multipath_target);
1761         kmem_cache_destroy(_mpio_cache);
1762 }
1763
1764 module_init(dm_multipath_init);
1765 module_exit(dm_multipath_exit);
1766
1767 MODULE_DESCRIPTION(DM_NAME " multipath target");
1768 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1769 MODULE_LICENSE("GPL");