drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9
  10 #include "dm-path-selector.h"
  11 #include "dm-uevent.h"
  12
  13 #include <linux/ctype.h>
  14 #include <linux/init.h>
  15 #include <linux/mempool.h>
  16 #include <linux/module.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/slab.h>
  19 #include <linux/time.h>
  20 #include <linux/workqueue.h>
  21 #include <scsi/scsi_dh.h>
  22 #include <asm/atomic.h>
  23
  24 #define DM_MSG_PREFIX "multipath"
  25 #define MESG_STR(x) x, sizeof(x)
  26
  27 /* Path properties */
  28 struct pgpath {
  29         struct list_head list;
  30
  31         struct priority_group *pg;      /* Owning PG */
  32         unsigned is_active;             /* Path status */
  33         unsigned fail_count;            /* Cumulative failure count */
  34
  35         struct dm_path path;
  36         struct work_struct deactivate_path;
  37         struct work_struct activate_path;
  38 };
  39
  40 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  41
  42 /*
  43  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  44  * Each has a path selector which controls which path gets used.
  45  */
  46 struct priority_group {
  47         struct list_head list;
  48
  49         struct multipath *m;            /* Owning multipath instance */
  50         struct path_selector ps;
  51
  52         unsigned pg_num;                /* Reference number */
  53         unsigned bypassed;              /* Temporarily bypass this PG? */
  54
  55         unsigned nr_pgpaths;            /* Number of paths in PG */
  56         struct list_head pgpaths;
  57 };
  58
  59 /* Multipath context */
  60 struct multipath {
  61         struct list_head list;
  62         struct dm_target *ti;
  63
  64         spinlock_t lock;
  65
  66         const char *hw_handler_name;
  67         char *hw_handler_params;
  68         unsigned nr_priority_groups;
  69         struct list_head priority_groups;
  70         unsigned pg_init_required;      /* pg_init needs calling? */
  71         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  72
  73         unsigned nr_valid_paths;        /* Total number of usable paths */
  74         struct pgpath *current_pgpath;
  75         struct priority_group *current_pg;
  76         struct priority_group *next_pg; /* Switch to this PG if set */
  77         unsigned repeat_count;          /* I/Os left before calling PS again */
  78
  79         unsigned queue_io;              /* Must we queue all I/O? */
  80         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  81         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  82         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  83         unsigned pg_init_count;         /* Number of times pg_init called */
  84
  85         struct work_struct process_queued_ios;
  86         struct list_head queued_ios;
  87         unsigned queue_size;
  88
  89         struct work_struct trigger_event;
  90
  91         /*
  92          * We must use a mempool of dm_mpath_io structs so that we
  93          * can resubmit bios on error.
  94          */
  95         mempool_t *mpio_pool;
  96
  97         struct mutex work_mutex;
  98 };
  99
 100 /*
 101  * Context information attached to each bio we process.
 102  */
 103 struct dm_mpath_io {
 104         struct pgpath *pgpath;
 105         size_t nr_bytes;
 106 };
 107
 108 typedef int (*action_fn) (struct pgpath *pgpath);
 109
 110 #define MIN_IOS 256     /* Mempool size */
 111
 112 static struct kmem_cache *_mpio_cache;
 113
 114 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 115 static void process_queued_ios(struct work_struct *work);
 116 static void trigger_event(struct work_struct *work);
 117 static void activate_path(struct work_struct *work);
 118 static void deactivate_path(struct work_struct *work);
 119
 120
 121 /*-----------------------------------------------
 122  * Allocation routines
 123  *-----------------------------------------------*/
 124
 125 static struct pgpath *alloc_pgpath(void)
 126 {
 127         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 128
 129         if (pgpath) {
 130                 pgpath->is_active = 1;
 131                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
 132                 INIT_WORK(&pgpath->activate_path, activate_path);
 133         }
 134
 135         return pgpath;
 136 }
 137
 138 static void free_pgpath(struct pgpath *pgpath)
 139 {
 140         kfree(pgpath);
 141 }
 142
 143 static void deactivate_path(struct work_struct *work)
 144 {
 145         struct pgpath *pgpath =
 146                 container_of(work, struct pgpath, deactivate_path);
 147
 148         blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
 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         struct multipath *m = ti->private;
 167
 168         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 169                 list_del(&pgpath->list);
 170                 if (m->hw_handler_name)
 171                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 172                 dm_put_device(ti, pgpath->path.dev);
 173                 free_pgpath(pgpath);
 174         }
 175 }
 176
 177 static void free_priority_group(struct priority_group *pg,
 178                                 struct dm_target *ti)
 179 {
 180         struct path_selector *ps = &pg->ps;
 181
 182         if (ps->type) {
 183                 ps->type->destroy(ps);
 184                 dm_put_path_selector(ps->type);
 185         }
 186
 187         free_pgpaths(&pg->pgpaths, ti);
 188         kfree(pg);
 189 }
 190
 191 static struct multipath *alloc_multipath(struct dm_target *ti)
 192 {
 193         struct multipath *m;
 194
 195         m = kzalloc(sizeof(*m), GFP_KERNEL);
 196         if (m) {
 197                 INIT_LIST_HEAD(&m->priority_groups);
 198                 INIT_LIST_HEAD(&m->queued_ios);
 199                 spin_lock_init(&m->lock);
 200                 m->queue_io = 1;
 201                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 202                 INIT_WORK(&m->trigger_event, trigger_event);
 203                 mutex_init(&m->work_mutex);
 204                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 205                 if (!m->mpio_pool) {
 206                         kfree(m);
 207                         return NULL;
 208                 }
 209                 m->ti = ti;
 210                 ti->private = m;
 211         }
 212
 213         return m;
 214 }
 215
 216 static void free_multipath(struct multipath *m)
 217 {
 218         struct priority_group *pg, *tmp;
 219
 220         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 221                 list_del(&pg->list);
 222                 free_priority_group(pg, m->ti);
 223         }
 224
 225         kfree(m->hw_handler_name);
 226         kfree(m->hw_handler_params);
 227         mempool_destroy(m->mpio_pool);
 228         kfree(m);
 229 }
 230
 231
 232 /*-----------------------------------------------
 233  * Path selection
 234  *-----------------------------------------------*/
 235
 236 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 237 {
 238         m->current_pg = pgpath->pg;
 239
 240         /* Must we initialise the PG first, and queue I/O till it's ready? */
 241         if (m->hw_handler_name) {
 242                 m->pg_init_required = 1;
 243                 m->queue_io = 1;
 244         } else {
 245                 m->pg_init_required = 0;
 246                 m->queue_io = 0;
 247         }
 248
 249         m->pg_init_count = 0;
 250 }
 251
 252 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 253                                size_t nr_bytes)
 254 {
 255         struct dm_path *path;
 256
 257         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 258         if (!path)
 259                 return -ENXIO;
 260
 261         m->current_pgpath = path_to_pgpath(path);
 262
 263         if (m->current_pg != pg)
 264                 __switch_pg(m, m->current_pgpath);
 265
 266         return 0;
 267 }
 268
 269 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 270 {
 271         struct priority_group *pg;
 272         unsigned bypassed = 1;
 273
 274         if (!m->nr_valid_paths)
 275                 goto failed;
 276
 277         /* Were we instructed to switch PG? */
 278         if (m->next_pg) {
 279                 pg = m->next_pg;
 280                 m->next_pg = NULL;
 281                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 282                         return;
 283         }
 284
 285         /* Don't change PG until it has no remaining paths */
 286         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 287                 return;
 288
 289         /*
 290          * Loop through priority groups until we find a valid path.
 291          * First time we skip PGs marked 'bypassed'.
 292          * Second time we only try the ones we skipped.
 293          */
 294         do {
 295                 list_for_each_entry(pg, &m->priority_groups, list) {
 296                         if (pg->bypassed == bypassed)
 297                                 continue;
 298                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 299                                 return;
 300                 }
 301         } while (bypassed--);
 302
 303 failed:
 304         m->current_pgpath = NULL;
 305         m->current_pg = NULL;
 306 }
 307
 308 /*
 309  * Check whether bios must be queued in the device-mapper core rather
 310  * than here in the target.
 311  *
 312  * m->lock must be held on entry.
 313  *
 314  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 315  * same value then we are not between multipath_presuspend()
 316  * and multipath_resume() calls and we have no need to check
 317  * for the DMF_NOFLUSH_SUSPENDING flag.
 318  */
 319 static int __must_push_back(struct multipath *m)
 320 {
 321         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 322                 dm_noflush_suspending(m->ti));
 323 }
 324
 325 static int map_io(struct multipath *m, struct request *clone,
 326                   struct dm_mpath_io *mpio, unsigned was_queued)
 327 {
 328         int r = DM_MAPIO_REMAPPED;
 329         size_t nr_bytes = blk_rq_bytes(clone);
 330         unsigned long flags;
 331         struct pgpath *pgpath;
 332         struct block_device *bdev;
 333
 334         spin_lock_irqsave(&m->lock, flags);
 335
 336         /* Do we need to select a new pgpath? */
 337         if (!m->current_pgpath ||
 338             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 339                 __choose_pgpath(m, nr_bytes);
 340
 341         pgpath = m->current_pgpath;
 342
 343         if (was_queued)
 344                 m->queue_size--;
 345
 346         if ((pgpath && m->queue_io) ||
 347             (!pgpath && m->queue_if_no_path)) {
 348                 /* Queue for the daemon to resubmit */
 349                 list_add_tail(&clone->queuelist, &m->queued_ios);
 350                 m->queue_size++;
 351                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 352                     !m->queue_io)
 353                         queue_work(kmultipathd, &m->process_queued_ios);
 354                 pgpath = NULL;
 355                 r = DM_MAPIO_SUBMITTED;
 356         } else if (pgpath) {
 357                 bdev = pgpath->path.dev->bdev;
 358                 clone->q = bdev_get_queue(bdev);
 359                 clone->rq_disk = bdev->bd_disk;
 360         } else if (__must_push_back(m))
 361                 r = DM_MAPIO_REQUEUE;
 362         else
 363                 r = -EIO;       /* Failed */
 364
 365         mpio->pgpath = pgpath;
 366         mpio->nr_bytes = nr_bytes;
 367
 368         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 369                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 370                                               nr_bytes);
 371
 372         spin_unlock_irqrestore(&m->lock, flags);
 373
 374         return r;
 375 }
 376
 377 /*
 378  * If we run out of usable paths, should we queue I/O or error it?
 379  */
 380 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 381                             unsigned save_old_value)
 382 {
 383         unsigned long flags;
 384
 385         spin_lock_irqsave(&m->lock, flags);
 386
 387         if (save_old_value)
 388                 m->saved_queue_if_no_path = m->queue_if_no_path;
 389         else
 390                 m->saved_queue_if_no_path = queue_if_no_path;
 391         m->queue_if_no_path = queue_if_no_path;
 392         if (!m->queue_if_no_path && m->queue_size)
 393                 queue_work(kmultipathd, &m->process_queued_ios);
 394
 395         spin_unlock_irqrestore(&m->lock, flags);
 396
 397         return 0;
 398 }
 399
 400 /*-----------------------------------------------------------------
 401  * The multipath daemon is responsible for resubmitting queued ios.
 402  *---------------------------------------------------------------*/
 403
 404 static void dispatch_queued_ios(struct multipath *m)
 405 {
 406         int r;
 407         unsigned long flags;
 408         struct dm_mpath_io *mpio;
 409         union map_info *info;
 410         struct request *clone, *n;
 411         LIST_HEAD(cl);
 412
 413         spin_lock_irqsave(&m->lock, flags);
 414         list_splice_init(&m->queued_ios, &cl);
 415         spin_unlock_irqrestore(&m->lock, flags);
 416
 417         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 418                 list_del_init(&clone->queuelist);
 419
 420                 info = dm_get_rq_mapinfo(clone);
 421                 mpio = info->ptr;
 422
 423                 r = map_io(m, clone, mpio, 1);
 424                 if (r < 0) {
 425                         mempool_free(mpio, m->mpio_pool);
 426                         dm_kill_unmapped_request(clone, r);
 427                 } else if (r == DM_MAPIO_REMAPPED)
 428                         dm_dispatch_request(clone);
 429                 else if (r == DM_MAPIO_REQUEUE) {
 430                         mempool_free(mpio, m->mpio_pool);
 431                         dm_requeue_unmapped_request(clone);
 432                 }
 433         }
 434 }
 435
 436 static void process_queued_ios(struct work_struct *work)
 437 {
 438         struct multipath *m =
 439                 container_of(work, struct multipath, process_queued_ios);
 440         struct pgpath *pgpath = NULL, *tmp;
 441         unsigned must_queue = 1;
 442         unsigned long flags;
 443
 444         spin_lock_irqsave(&m->lock, flags);
 445
 446         if (!m->queue_size)
 447                 goto out;
 448
 449         if (!m->current_pgpath)
 450                 __choose_pgpath(m, 0);
 451
 452         pgpath = m->current_pgpath;
 453
 454         if ((pgpath && !m->queue_io) ||
 455             (!pgpath && !m->queue_if_no_path))
 456                 must_queue = 0;
 457
 458         if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
 459                 m->pg_init_count++;
 460                 m->pg_init_required = 0;
 461                 list_for_each_entry(tmp, &pgpath->pg->pgpaths, list) {
 462                         if (queue_work(kmpath_handlerd, &tmp->activate_path))
 463                                 m->pg_init_in_progress++;
 464                 }
 465         }
 466 out:
 467         spin_unlock_irqrestore(&m->lock, flags);
 468         if (!must_queue)
 469                 dispatch_queued_ios(m);
 470 }
 471
 472 /*
 473  * An event is triggered whenever a path is taken out of use.
 474  * Includes path failure and PG bypass.
 475  */
 476 static void trigger_event(struct work_struct *work)
 477 {
 478         struct multipath *m =
 479                 container_of(work, struct multipath, trigger_event);
 480
 481         dm_table_event(m->ti->table);
 482 }
 483
 484 /*-----------------------------------------------------------------
 485  * Constructor/argument parsing:
 486  * <#multipath feature args> [<arg>]*
 487  * <#hw_handler args> [hw_handler [<arg>]*]
 488  * <#priority groups>
 489  * <initial priority group>
 490  *     [<selector> <#selector args> [<arg>]*
 491  *      <#paths> <#per-path selector args>
 492  *         [<path> [<arg>]* ]+ ]+
 493  *---------------------------------------------------------------*/
 494 struct param {
 495         unsigned min;
 496         unsigned max;
 497         char *error;
 498 };
 499
 500 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 501 {
 502         if (!str ||
 503             (sscanf(str, "%u", v) != 1) ||
 504             (*v < param->min) ||
 505             (*v > param->max)) {
 506                 *error = param->error;
 507                 return -EINVAL;
 508         }
 509
 510         return 0;
 511 }
 512
 513 struct arg_set {
 514         unsigned argc;
 515         char **argv;
 516 };
 517
 518 static char *shift(struct arg_set *as)
 519 {
 520         char *r;
 521
 522         if (as->argc) {
 523                 as->argc--;
 524                 r = *as->argv;
 525                 as->argv++;
 526                 return r;
 527         }
 528
 529         return NULL;
 530 }
 531
 532 static void consume(struct arg_set *as, unsigned n)
 533 {
 534         BUG_ON (as->argc < n);
 535         as->argc -= n;
 536         as->argv += n;
 537 }
 538
 539 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 540                                struct dm_target *ti)
 541 {
 542         int r;
 543         struct path_selector_type *pst;
 544         unsigned ps_argc;
 545
 546         static struct param _params[] = {
 547                 {0, 1024, "invalid number of path selector args"},
 548         };
 549
 550         pst = dm_get_path_selector(shift(as));
 551         if (!pst) {
 552                 ti->error = "unknown path selector type";
 553                 return -EINVAL;
 554         }
 555
 556         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 557         if (r) {
 558                 dm_put_path_selector(pst);
 559                 return -EINVAL;
 560         }
 561
 562         if (ps_argc > as->argc) {
 563                 dm_put_path_selector(pst);
 564                 ti->error = "not enough arguments for path selector";
 565                 return -EINVAL;
 566         }
 567
 568         r = pst->create(&pg->ps, ps_argc, as->argv);
 569         if (r) {
 570                 dm_put_path_selector(pst);
 571                 ti->error = "path selector constructor failed";
 572                 return r;
 573         }
 574
 575         pg->ps.type = pst;
 576         consume(as, ps_argc);
 577
 578         return 0;
 579 }
 580
 581 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 582                                struct dm_target *ti)
 583 {
 584         int r;
 585         struct pgpath *p;
 586         struct multipath *m = ti->private;
 587
 588         /* we need at least a path arg */
 589         if (as->argc < 1) {
 590                 ti->error = "no device given";
 591                 return ERR_PTR(-EINVAL);
 592         }
 593
 594         p = alloc_pgpath();
 595         if (!p)
 596                 return ERR_PTR(-ENOMEM);
 597
 598         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 599                           dm_table_get_mode(ti->table), &p->path.dev);
 600         if (r) {
 601                 ti->error = "error getting device";
 602                 goto bad;
 603         }
 604
 605         if (m->hw_handler_name) {
 606                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 607
 608                 r = scsi_dh_attach(q, m->hw_handler_name);
 609                 if (r == -EBUSY) {
 610                         /*
 611                          * Already attached to different hw_handler,
 612                          * try to reattach with correct one.
 613                          */
 614                         scsi_dh_detach(q);
 615                         r = scsi_dh_attach(q, m->hw_handler_name);
 616                 }
 617
 618                 if (r < 0) {
 619                         ti->error = "error attaching hardware handler";
 620                         dm_put_device(ti, p->path.dev);
 621                         goto bad;
 622                 }
 623
 624                 if (m->hw_handler_params) {
 625                         r = scsi_dh_set_params(q, m->hw_handler_params);
 626                         if (r < 0) {
 627                                 ti->error = "unable to set hardware "
 628                                                         "handler parameters";
 629                                 scsi_dh_detach(q);
 630                                 dm_put_device(ti, p->path.dev);
 631                                 goto bad;
 632                         }
 633                 }
 634         }
 635
 636         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 637         if (r) {
 638                 dm_put_device(ti, p->path.dev);
 639                 goto bad;
 640         }
 641
 642         return p;
 643
 644  bad:
 645         free_pgpath(p);
 646         return ERR_PTR(r);
 647 }
 648
 649 static struct priority_group *parse_priority_group(struct arg_set *as,
 650                                                    struct multipath *m)
 651 {
 652         static struct param _params[] = {
 653                 {1, 1024, "invalid number of paths"},
 654                 {0, 1024, "invalid number of selector args"}
 655         };
 656
 657         int r;
 658         unsigned i, nr_selector_args, nr_params;
 659         struct priority_group *pg;
 660         struct dm_target *ti = m->ti;
 661
 662         if (as->argc < 2) {
 663                 as->argc = 0;
 664                 ti->error = "not enough priority group arguments";
 665                 return ERR_PTR(-EINVAL);
 666         }
 667
 668         pg = alloc_priority_group();
 669         if (!pg) {
 670                 ti->error = "couldn't allocate priority group";
 671                 return ERR_PTR(-ENOMEM);
 672         }
 673         pg->m = m;
 674
 675         r = parse_path_selector(as, pg, ti);
 676         if (r)
 677                 goto bad;
 678
 679         /*
 680          * read the paths
 681          */
 682         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 683         if (r)
 684                 goto bad;
 685
 686         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 687         if (r)
 688                 goto bad;
 689
 690         nr_params = 1 + nr_selector_args;
 691         for (i = 0; i < pg->nr_pgpaths; i++) {
 692                 struct pgpath *pgpath;
 693                 struct arg_set path_args;
 694
 695                 if (as->argc < nr_params) {
 696                         ti->error = "not enough path parameters";
 697                         goto bad;
 698                 }
 699
 700                 path_args.argc = nr_params;
 701                 path_args.argv = as->argv;
 702
 703                 pgpath = parse_path(&path_args, &pg->ps, ti);
 704                 if (IS_ERR(pgpath)) {
 705                         r = PTR_ERR(pgpath);
 706                         goto bad;
 707                 }
 708
 709                 pgpath->pg = pg;
 710                 list_add_tail(&pgpath->list, &pg->pgpaths);
 711                 consume(as, nr_params);
 712         }
 713
 714         return pg;
 715
 716  bad:
 717         free_priority_group(pg, ti);
 718         return ERR_PTR(r);
 719 }
 720
 721 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 722 {
 723         unsigned hw_argc;
 724         int ret;
 725         struct dm_target *ti = m->ti;
 726
 727         static struct param _params[] = {
 728                 {0, 1024, "invalid number of hardware handler args"},
 729         };
 730
 731         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 732                 return -EINVAL;
 733
 734         if (!hw_argc)
 735                 return 0;
 736
 737         if (hw_argc > as->argc) {
 738                 ti->error = "not enough arguments for hardware handler";
 739                 return -EINVAL;
 740         }
 741
 742         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 743         request_module("scsi_dh_%s", m->hw_handler_name);
 744         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 745                 ti->error = "unknown hardware handler type";
 746                 ret = -EINVAL;
 747                 goto fail;
 748         }
 749
 750         if (hw_argc > 1) {
 751                 char *p;
 752                 int i, j, len = 4;
 753
 754                 for (i = 0; i <= hw_argc - 2; i++)
 755                         len += strlen(as->argv[i]) + 1;
 756                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 757                 if (!p) {
 758                         ti->error = "memory allocation failed";
 759                         ret = -ENOMEM;
 760                         goto fail;
 761                 }
 762                 j = sprintf(p, "%d", hw_argc - 1);
 763                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 764                         j = sprintf(p, "%s", as->argv[i]);
 765         }
 766         consume(as, hw_argc - 1);
 767
 768         return 0;
 769 fail:
 770         kfree(m->hw_handler_name);
 771         m->hw_handler_name = NULL;
 772         return ret;
 773 }
 774
 775 static int parse_features(struct arg_set *as, struct multipath *m)
 776 {
 777         int r;
 778         unsigned argc;
 779         struct dm_target *ti = m->ti;
 780         const char *param_name;
 781
 782         static struct param _params[] = {
 783                 {0, 3, "invalid number of feature args"},
 784                 {1, 50, "pg_init_retries must be between 1 and 50"},
 785         };
 786
 787         r = read_param(_params, shift(as), &argc, &ti->error);
 788         if (r)
 789                 return -EINVAL;
 790
 791         if (!argc)
 792                 return 0;
 793
 794         do {
 795                 param_name = shift(as);
 796                 argc--;
 797
 798                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 799                         r = queue_if_no_path(m, 1, 0);
 800                         continue;
 801                 }
 802
 803                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 804                     (argc >= 1)) {
 805                         r = read_param(_params + 1, shift(as),
 806                                        &m->pg_init_retries, &ti->error);
 807                         argc--;
 808                         continue;
 809                 }
 810
 811                 ti->error = "Unrecognised multipath feature request";
 812                 r = -EINVAL;
 813         } while (argc && !r);
 814
 815         return r;
 816 }
 817
 818 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 819                          char **argv)
 820 {
 821         /* target parameters */
 822         static struct param _params[] = {
 823                 {1, 1024, "invalid number of priority groups"},
 824                 {1, 1024, "invalid initial priority group number"},
 825         };
 826
 827         int r;
 828         struct multipath *m;
 829         struct arg_set as;
 830         unsigned pg_count = 0;
 831         unsigned next_pg_num;
 832
 833         as.argc = argc;
 834         as.argv = argv;
 835
 836         m = alloc_multipath(ti);
 837         if (!m) {
 838                 ti->error = "can't allocate multipath";
 839                 return -EINVAL;
 840         }
 841
 842         r = parse_features(&as, m);
 843         if (r)
 844                 goto bad;
 845
 846         r = parse_hw_handler(&as, m);
 847         if (r)
 848                 goto bad;
 849
 850         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 851         if (r)
 852                 goto bad;
 853
 854         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 855         if (r)
 856                 goto bad;
 857
 858         /* parse the priority groups */
 859         while (as.argc) {
 860                 struct priority_group *pg;
 861
 862                 pg = parse_priority_group(&as, m);
 863                 if (IS_ERR(pg)) {
 864                         r = PTR_ERR(pg);
 865                         goto bad;
 866                 }
 867
 868                 m->nr_valid_paths += pg->nr_pgpaths;
 869                 list_add_tail(&pg->list, &m->priority_groups);
 870                 pg_count++;
 871                 pg->pg_num = pg_count;
 872                 if (!--next_pg_num)
 873                         m->next_pg = pg;
 874         }
 875
 876         if (pg_count != m->nr_priority_groups) {
 877                 ti->error = "priority group count mismatch";
 878                 r = -EINVAL;
 879                 goto bad;
 880         }
 881
 882         ti->num_flush_requests = 1;
 883
 884         return 0;
 885
 886  bad:
 887         free_multipath(m);
 888         return r;
 889 }
 890
 891 static void flush_multipath_work(void)
 892 {
 893         flush_workqueue(kmpath_handlerd);
 894         flush_workqueue(kmultipathd);
 895         flush_scheduled_work();
 896 }
 897
 898 static void multipath_dtr(struct dm_target *ti)
 899 {
 900         struct multipath *m = ti->private;
 901
 902         flush_multipath_work();
 903         free_multipath(m);
 904 }
 905
 906 /*
 907  * Map cloned requests
 908  */
 909 static int multipath_map(struct dm_target *ti, struct request *clone,
 910                          union map_info *map_context)
 911 {
 912         int r;
 913         struct dm_mpath_io *mpio;
 914         struct multipath *m = (struct multipath *) ti->private;
 915
 916         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 917         if (!mpio)
 918                 /* ENOMEM, requeue */
 919                 return DM_MAPIO_REQUEUE;
 920         memset(mpio, 0, sizeof(*mpio));
 921
 922         map_context->ptr = mpio;
 923         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 924         r = map_io(m, clone, mpio, 0);
 925         if (r < 0 || r == DM_MAPIO_REQUEUE)
 926                 mempool_free(mpio, m->mpio_pool);
 927
 928         return r;
 929 }
 930
 931 /*
 932  * Take a path out of use.
 933  */
 934 static int fail_path(struct pgpath *pgpath)
 935 {
 936         unsigned long flags;
 937         struct multipath *m = pgpath->pg->m;
 938
 939         spin_lock_irqsave(&m->lock, flags);
 940
 941         if (!pgpath->is_active)
 942                 goto out;
 943
 944         DMWARN("Failing path %s.", pgpath->path.dev->name);
 945
 946         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 947         pgpath->is_active = 0;
 948         pgpath->fail_count++;
 949
 950         m->nr_valid_paths--;
 951
 952         if (pgpath == m->current_pgpath)
 953                 m->current_pgpath = NULL;
 954
 955         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 956                       pgpath->path.dev->name, m->nr_valid_paths);
 957
 958         schedule_work(&m->trigger_event);
 959         queue_work(kmultipathd, &pgpath->deactivate_path);
 960
 961 out:
 962         spin_unlock_irqrestore(&m->lock, flags);
 963
 964         return 0;
 965 }
 966
 967 /*
 968  * Reinstate a previously-failed path
 969  */
 970 static int reinstate_path(struct pgpath *pgpath)
 971 {
 972         int r = 0;
 973         unsigned long flags;
 974         struct multipath *m = pgpath->pg->m;
 975
 976         spin_lock_irqsave(&m->lock, flags);
 977
 978         if (pgpath->is_active)
 979                 goto out;
 980
 981         if (!pgpath->pg->ps.type->reinstate_path) {
 982                 DMWARN("Reinstate path not supported by path selector %s",
 983                        pgpath->pg->ps.type->name);
 984                 r = -EINVAL;
 985                 goto out;
 986         }
 987
 988         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 989         if (r)
 990                 goto out;
 991
 992         pgpath->is_active = 1;
 993
 994         if (!m->nr_valid_paths++ && m->queue_size) {
 995                 m->current_pgpath = NULL;
 996                 queue_work(kmultipathd, &m->process_queued_ios);
 997         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
 998                 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
 999                         m->pg_init_in_progress++;
1000         }
1001
1002         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1003                       pgpath->path.dev->name, m->nr_valid_paths);
1004
1005         schedule_work(&m->trigger_event);
1006
1007 out:
1008         spin_unlock_irqrestore(&m->lock, flags);
1009
1010         return r;
1011 }
1012
1013 /*
1014  * Fail or reinstate all paths that match the provided struct dm_dev.
1015  */
1016 static int action_dev(struct multipath *m, struct dm_dev *dev,
1017                       action_fn action)
1018 {
1019         int r = 0;
1020         struct pgpath *pgpath;
1021         struct priority_group *pg;
1022
1023         list_for_each_entry(pg, &m->priority_groups, list) {
1024                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1025                         if (pgpath->path.dev == dev)
1026                                 r = action(pgpath);
1027                 }
1028         }
1029
1030         return r;
1031 }
1032
1033 /*
1034  * Temporarily try to avoid having to use the specified PG
1035  */
1036 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1037                       int bypassed)
1038 {
1039         unsigned long flags;
1040
1041         spin_lock_irqsave(&m->lock, flags);
1042
1043         pg->bypassed = bypassed;
1044         m->current_pgpath = NULL;
1045         m->current_pg = NULL;
1046
1047         spin_unlock_irqrestore(&m->lock, flags);
1048
1049         schedule_work(&m->trigger_event);
1050 }
1051
1052 /*
1053  * Switch to using the specified PG from the next I/O that gets mapped
1054  */
1055 static int switch_pg_num(struct multipath *m, const char *pgstr)
1056 {
1057         struct priority_group *pg;
1058         unsigned pgnum;
1059         unsigned long flags;
1060
1061         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1062             (pgnum > m->nr_priority_groups)) {
1063                 DMWARN("invalid PG number supplied to switch_pg_num");
1064                 return -EINVAL;
1065         }
1066
1067         spin_lock_irqsave(&m->lock, flags);
1068         list_for_each_entry(pg, &m->priority_groups, list) {
1069                 pg->bypassed = 0;
1070                 if (--pgnum)
1071                         continue;
1072
1073                 m->current_pgpath = NULL;
1074                 m->current_pg = NULL;
1075                 m->next_pg = pg;
1076         }
1077         spin_unlock_irqrestore(&m->lock, flags);
1078
1079         schedule_work(&m->trigger_event);
1080         return 0;
1081 }
1082
1083 /*
1084  * Set/clear bypassed status of a PG.
1085  * PGs are numbered upwards from 1 in the order they were declared.
1086  */
1087 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1088 {
1089         struct priority_group *pg;
1090         unsigned pgnum;
1091
1092         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1093             (pgnum > m->nr_priority_groups)) {
1094                 DMWARN("invalid PG number supplied to bypass_pg");
1095                 return -EINVAL;
1096         }
1097
1098         list_for_each_entry(pg, &m->priority_groups, list) {
1099                 if (!--pgnum)
1100                         break;
1101         }
1102
1103         bypass_pg(m, pg, bypassed);
1104         return 0;
1105 }
1106
1107 /*
1108  * Should we retry pg_init immediately?
1109  */
1110 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1111 {
1112         unsigned long flags;
1113         int limit_reached = 0;
1114
1115         spin_lock_irqsave(&m->lock, flags);
1116
1117         if (m->pg_init_count <= m->pg_init_retries)
1118                 m->pg_init_required = 1;
1119         else
1120                 limit_reached = 1;
1121
1122         spin_unlock_irqrestore(&m->lock, flags);
1123
1124         return limit_reached;
1125 }
1126
1127 static void pg_init_done(void *data, int errors)
1128 {
1129         struct dm_path *path = data;
1130         struct pgpath *pgpath = path_to_pgpath(path);
1131         struct priority_group *pg = pgpath->pg;
1132         struct multipath *m = pg->m;
1133         unsigned long flags;
1134
1135         /* device or driver problems */
1136         switch (errors) {
1137         case SCSI_DH_OK:
1138                 break;
1139         case SCSI_DH_NOSYS:
1140                 if (!m->hw_handler_name) {
1141                         errors = 0;
1142                         break;
1143                 }
1144                 DMERR("Cannot failover device because scsi_dh_%s was not "
1145                       "loaded.", m->hw_handler_name);
1146                 /*
1147                  * Fail path for now, so we do not ping pong
1148                  */
1149                 fail_path(pgpath);
1150                 break;
1151         case SCSI_DH_DEV_TEMP_BUSY:
1152                 /*
1153                  * Probably doing something like FW upgrade on the
1154                  * controller so try the other pg.
1155                  */
1156                 bypass_pg(m, pg, 1);
1157                 break;
1158         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1159         case SCSI_DH_RETRY:
1160         case SCSI_DH_IMM_RETRY:
1161         case SCSI_DH_RES_TEMP_UNAVAIL:
1162                 if (pg_init_limit_reached(m, pgpath))
1163                         fail_path(pgpath);
1164                 errors = 0;
1165                 break;
1166         default:
1167                 /*
1168                  * We probably do not want to fail the path for a device
1169                  * error, but this is what the old dm did. In future
1170                  * patches we can do more advanced handling.
1171                  */
1172                 fail_path(pgpath);
1173         }
1174
1175         spin_lock_irqsave(&m->lock, flags);
1176         if (errors) {
1177                 if (pgpath == m->current_pgpath) {
1178                         DMERR("Could not failover device. Error %d.", errors);
1179                         m->current_pgpath = NULL;
1180                         m->current_pg = NULL;
1181                 }
1182         } else if (!m->pg_init_required) {
1183                 m->queue_io = 0;
1184                 pg->bypassed = 0;
1185         }
1186
1187         m->pg_init_in_progress--;
1188         if (!m->pg_init_in_progress)
1189                 queue_work(kmultipathd, &m->process_queued_ios);
1190         spin_unlock_irqrestore(&m->lock, flags);
1191 }
1192
1193 static void activate_path(struct work_struct *work)
1194 {
1195         struct pgpath *pgpath =
1196                 container_of(work, struct pgpath, activate_path);
1197
1198         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1199                                 pg_init_done, &pgpath->path);
1200 }
1201
1202 /*
1203  * end_io handling
1204  */
1205 static int do_end_io(struct multipath *m, struct request *clone,
1206                      int error, struct dm_mpath_io *mpio)
1207 {
1208         /*
1209          * We don't queue any clone request inside the multipath target
1210          * during end I/O handling, since those clone requests don't have
1211          * bio clones.  If we queue them inside the multipath target,
1212          * we need to make bio clones, that requires memory allocation.
1213          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1214          *  don't have bio clones.)
1215          * Instead of queueing the clone request here, we queue the original
1216          * request into dm core, which will remake a clone request and
1217          * clone bios for it and resubmit it later.
1218          */
1219         int r = DM_ENDIO_REQUEUE;
1220         unsigned long flags;
1221
1222         if (!error && !clone->errors)
1223                 return 0;       /* I/O complete */
1224
1225         if (error == -EOPNOTSUPP)
1226                 return error;
1227
1228         if (mpio->pgpath)
1229                 fail_path(mpio->pgpath);
1230
1231         spin_lock_irqsave(&m->lock, flags);
1232         if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1233                 r = -EIO;
1234         spin_unlock_irqrestore(&m->lock, flags);
1235
1236         return r;
1237 }
1238
1239 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1240                             int error, union map_info *map_context)
1241 {
1242         struct multipath *m = ti->private;
1243         struct dm_mpath_io *mpio = map_context->ptr;
1244         struct pgpath *pgpath = mpio->pgpath;
1245         struct path_selector *ps;
1246         int r;
1247
1248         r  = do_end_io(m, clone, error, mpio);
1249         if (pgpath) {
1250                 ps = &pgpath->pg->ps;
1251                 if (ps->type->end_io)
1252                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1253         }
1254         mempool_free(mpio, m->mpio_pool);
1255
1256         return r;
1257 }
1258
1259 /*
1260  * Suspend can't complete until all the I/O is processed so if
1261  * the last path fails we must error any remaining I/O.
1262  * Note that if the freeze_bdev fails while suspending, the
1263  * queue_if_no_path state is lost - userspace should reset it.
1264  */
1265 static void multipath_presuspend(struct dm_target *ti)
1266 {
1267         struct multipath *m = (struct multipath *) ti->private;
1268
1269         queue_if_no_path(m, 0, 1);
1270 }
1271
1272 static void multipath_postsuspend(struct dm_target *ti)
1273 {
1274         struct multipath *m = ti->private;
1275
1276         mutex_lock(&m->work_mutex);
1277         flush_multipath_work();
1278         mutex_unlock(&m->work_mutex);
1279 }
1280
1281 /*
1282  * Restore the queue_if_no_path setting.
1283  */
1284 static void multipath_resume(struct dm_target *ti)
1285 {
1286         struct multipath *m = (struct multipath *) ti->private;
1287         unsigned long flags;
1288
1289         spin_lock_irqsave(&m->lock, flags);
1290         m->queue_if_no_path = m->saved_queue_if_no_path;
1291         spin_unlock_irqrestore(&m->lock, flags);
1292 }
1293
1294 /*
1295  * Info output has the following format:
1296  * num_multipath_feature_args [multipath_feature_args]*
1297  * num_handler_status_args [handler_status_args]*
1298  * num_groups init_group_number
1299  *            [A|D|E num_ps_status_args [ps_status_args]*
1300  *             num_paths num_selector_args
1301  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1302  *
1303  * Table output has the following format (identical to the constructor string):
1304  * num_feature_args [features_args]*
1305  * num_handler_args hw_handler [hw_handler_args]*
1306  * num_groups init_group_number
1307  *     [priority selector-name num_ps_args [ps_args]*
1308  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1309  */
1310 static int multipath_status(struct dm_target *ti, status_type_t type,
1311                             char *result, unsigned int maxlen)
1312 {
1313         int sz = 0;
1314         unsigned long flags;
1315         struct multipath *m = (struct multipath *) ti->private;
1316         struct priority_group *pg;
1317         struct pgpath *p;
1318         unsigned pg_num;
1319         char state;
1320
1321         spin_lock_irqsave(&m->lock, flags);
1322
1323         /* Features */
1324         if (type == STATUSTYPE_INFO)
1325                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1326         else {
1327                 DMEMIT("%u ", m->queue_if_no_path +
1328                               (m->pg_init_retries > 0) * 2);
1329                 if (m->queue_if_no_path)
1330                         DMEMIT("queue_if_no_path ");
1331                 if (m->pg_init_retries)
1332                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1333         }
1334
1335         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1336                 DMEMIT("0 ");
1337         else
1338                 DMEMIT("1 %s ", m->hw_handler_name);
1339
1340         DMEMIT("%u ", m->nr_priority_groups);
1341
1342         if (m->next_pg)
1343                 pg_num = m->next_pg->pg_num;
1344         else if (m->current_pg)
1345                 pg_num = m->current_pg->pg_num;
1346         else
1347                         pg_num = 1;
1348
1349         DMEMIT("%u ", pg_num);
1350
1351         switch (type) {
1352         case STATUSTYPE_INFO:
1353                 list_for_each_entry(pg, &m->priority_groups, list) {
1354                         if (pg->bypassed)
1355                                 state = 'D';    /* Disabled */
1356                         else if (pg == m->current_pg)
1357                                 state = 'A';    /* Currently Active */
1358                         else
1359                                 state = 'E';    /* Enabled */
1360
1361                         DMEMIT("%c ", state);
1362
1363                         if (pg->ps.type->status)
1364                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1365                                                           result + sz,
1366                                                           maxlen - sz);
1367                         else
1368                                 DMEMIT("0 ");
1369
1370                         DMEMIT("%u %u ", pg->nr_pgpaths,
1371                                pg->ps.type->info_args);
1372
1373                         list_for_each_entry(p, &pg->pgpaths, list) {
1374                                 DMEMIT("%s %s %u ", p->path.dev->name,
1375                                        p->is_active ? "A" : "F",
1376                                        p->fail_count);
1377                                 if (pg->ps.type->status)
1378                                         sz += pg->ps.type->status(&pg->ps,
1379                                               &p->path, type, result + sz,
1380                                               maxlen - sz);
1381                         }
1382                 }
1383                 break;
1384
1385         case STATUSTYPE_TABLE:
1386                 list_for_each_entry(pg, &m->priority_groups, list) {
1387                         DMEMIT("%s ", pg->ps.type->name);
1388
1389                         if (pg->ps.type->status)
1390                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1391                                                           result + sz,
1392                                                           maxlen - sz);
1393                         else
1394                                 DMEMIT("0 ");
1395
1396                         DMEMIT("%u %u ", pg->nr_pgpaths,
1397                                pg->ps.type->table_args);
1398
1399                         list_for_each_entry(p, &pg->pgpaths, list) {
1400                                 DMEMIT("%s ", p->path.dev->name);
1401                                 if (pg->ps.type->status)
1402                                         sz += pg->ps.type->status(&pg->ps,
1403                                               &p->path, type, result + sz,
1404                                               maxlen - sz);
1405                         }
1406                 }
1407                 break;
1408         }
1409
1410         spin_unlock_irqrestore(&m->lock, flags);
1411
1412         return 0;
1413 }
1414
1415 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1416 {
1417         int r = -EINVAL;
1418         struct dm_dev *dev;
1419         struct multipath *m = (struct multipath *) ti->private;
1420         action_fn action;
1421
1422         mutex_lock(&m->work_mutex);
1423
1424         if (argc == 1) {
1425                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
1426                         r = queue_if_no_path(m, 1, 0);
1427                         goto out;
1428                 } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
1429                         r = queue_if_no_path(m, 0, 0);
1430                         goto out;
1431                 }
1432         }
1433
1434         if (argc != 2) {
1435                 DMWARN("Unrecognised multipath message received.");
1436                 goto out;
1437         }
1438
1439         if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
1440                 r = bypass_pg_num(m, argv[1], 1);
1441                 goto out;
1442         } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
1443                 r = bypass_pg_num(m, argv[1], 0);
1444                 goto out;
1445         } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
1446                 r = switch_pg_num(m, argv[1]);
1447                 goto out;
1448         } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1449                 action = reinstate_path;
1450         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1451                 action = fail_path;
1452         else {
1453                 DMWARN("Unrecognised multipath message received.");
1454                 goto out;
1455         }
1456
1457         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1458                           dm_table_get_mode(ti->table), &dev);
1459         if (r) {
1460                 DMWARN("message: error getting device %s",
1461                        argv[1]);
1462                 goto out;
1463         }
1464
1465         r = action_dev(m, dev, action);
1466
1467         dm_put_device(ti, dev);
1468
1469 out:
1470         mutex_unlock(&m->work_mutex);
1471         return r;
1472 }
1473
1474 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1475                            unsigned long arg)
1476 {
1477         struct multipath *m = (struct multipath *) ti->private;
1478         struct block_device *bdev = NULL;
1479         fmode_t mode = 0;
1480         unsigned long flags;
1481         int r = 0;
1482
1483         spin_lock_irqsave(&m->lock, flags);
1484
1485         if (!m->current_pgpath)
1486                 __choose_pgpath(m, 0);
1487
1488         if (m->current_pgpath) {
1489                 bdev = m->current_pgpath->path.dev->bdev;
1490                 mode = m->current_pgpath->path.dev->mode;
1491         }
1492
1493         if (m->queue_io)
1494                 r = -EAGAIN;
1495         else if (!bdev)
1496                 r = -EIO;
1497
1498         spin_unlock_irqrestore(&m->lock, flags);
1499
1500         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1501 }
1502
1503 static int multipath_iterate_devices(struct dm_target *ti,
1504                                      iterate_devices_callout_fn fn, void *data)
1505 {
1506         struct multipath *m = ti->private;
1507         struct priority_group *pg;
1508         struct pgpath *p;
1509         int ret = 0;
1510
1511         list_for_each_entry(pg, &m->priority_groups, list) {
1512                 list_for_each_entry(p, &pg->pgpaths, list) {
1513                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1514                         if (ret)
1515                                 goto out;
1516                 }
1517         }
1518
1519 out:
1520         return ret;
1521 }
1522
1523 static int __pgpath_busy(struct pgpath *pgpath)
1524 {
1525         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1526
1527         return dm_underlying_device_busy(q);
1528 }
1529
1530 /*
1531  * We return "busy", only when we can map I/Os but underlying devices
1532  * are busy (so even if we map I/Os now, the I/Os will wait on
1533  * the underlying queue).
1534  * In other words, if we want to kill I/Os or queue them inside us
1535  * due to map unavailability, we don't return "busy".  Otherwise,
1536  * dm core won't give us the I/Os and we can't do what we want.
1537  */
1538 static int multipath_busy(struct dm_target *ti)
1539 {
1540         int busy = 0, has_active = 0;
1541         struct multipath *m = ti->private;
1542         struct priority_group *pg;
1543         struct pgpath *pgpath;
1544         unsigned long flags;
1545
1546         spin_lock_irqsave(&m->lock, flags);
1547
1548         /* Guess which priority_group will be used at next mapping time */
1549         if (unlikely(!m->current_pgpath && m->next_pg))
1550                 pg = m->next_pg;
1551         else if (likely(m->current_pg))
1552                 pg = m->current_pg;
1553         else
1554                 /*
1555                  * We don't know which pg will be used at next mapping time.
1556                  * We don't call __choose_pgpath() here to avoid to trigger
1557                  * pg_init just by busy checking.
1558                  * So we don't know whether underlying devices we will be using
1559                  * at next mapping time are busy or not. Just try mapping.
1560                  */
1561                 goto out;
1562
1563         /*
1564          * If there is one non-busy active path at least, the path selector
1565          * will be able to select it. So we consider such a pg as not busy.
1566          */
1567         busy = 1;
1568         list_for_each_entry(pgpath, &pg->pgpaths, list)
1569                 if (pgpath->is_active) {
1570                         has_active = 1;
1571
1572                         if (!__pgpath_busy(pgpath)) {
1573                                 busy = 0;
1574                                 break;
1575                         }
1576                 }
1577
1578         if (!has_active)
1579                 /*
1580                  * No active path in this pg, so this pg won't be used and
1581                  * the current_pg will be changed at next mapping time.
1582                  * We need to try mapping to determine it.
1583                  */
1584                 busy = 0;
1585
1586 out:
1587         spin_unlock_irqrestore(&m->lock, flags);
1588
1589         return busy;
1590 }
1591
1592 /*-----------------------------------------------------------------
1593  * Module setup
1594  *---------------------------------------------------------------*/
1595 static struct target_type multipath_target = {
1596         .name = "multipath",
1597         .version = {1, 1, 1},
1598         .module = THIS_MODULE,
1599         .ctr = multipath_ctr,
1600         .dtr = multipath_dtr,
1601         .map_rq = multipath_map,
1602         .rq_end_io = multipath_end_io,
1603         .presuspend = multipath_presuspend,
1604         .postsuspend = multipath_postsuspend,
1605         .resume = multipath_resume,
1606         .status = multipath_status,
1607         .message = multipath_message,
1608         .ioctl  = multipath_ioctl,
1609         .iterate_devices = multipath_iterate_devices,
1610         .busy = multipath_busy,
1611 };
1612
1613 static int __init dm_multipath_init(void)
1614 {
1615         int r;
1616
1617         /* allocate a slab for the dm_ios */
1618         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1619         if (!_mpio_cache)
1620                 return -ENOMEM;
1621
1622         r = dm_register_target(&multipath_target);
1623         if (r < 0) {
1624                 DMERR("register failed %d", r);
1625                 kmem_cache_destroy(_mpio_cache);
1626                 return -EINVAL;
1627         }
1628
1629         kmultipathd = create_workqueue("kmpathd");
1630         if (!kmultipathd) {
1631                 DMERR("failed to create workqueue kmpathd");
1632                 dm_unregister_target(&multipath_target);
1633                 kmem_cache_destroy(_mpio_cache);
1634                 return -ENOMEM;
1635         }
1636
1637         /*
1638          * A separate workqueue is used to handle the device handlers
1639          * to avoid overloading existing workqueue. Overloading the
1640          * old workqueue would also create a bottleneck in the
1641          * path of the storage hardware device activation.
1642          */
1643         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1644         if (!kmpath_handlerd) {
1645                 DMERR("failed to create workqueue kmpath_handlerd");
1646                 destroy_workqueue(kmultipathd);
1647                 dm_unregister_target(&multipath_target);
1648                 kmem_cache_destroy(_mpio_cache);
1649                 return -ENOMEM;
1650         }
1651
1652         DMINFO("version %u.%u.%u loaded",
1653                multipath_target.version[0], multipath_target.version[1],
1654                multipath_target.version[2]);
1655
1656         return r;
1657 }
1658
1659 static void __exit dm_multipath_exit(void)
1660 {
1661         destroy_workqueue(kmpath_handlerd);
1662         destroy_workqueue(kmultipathd);
1663
1664         dm_unregister_target(&multipath_target);
1665         kmem_cache_destroy(_mpio_cache);
1666 }
1667
1668 module_init(dm_multipath_init);
1669 module_exit(dm_multipath_exit);
1670
1671 MODULE_DESCRIPTION(DM_NAME " multipath target");
1672 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1673 MODULE_LICENSE("GPL");