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 "dm.h"
   9 #include "dm-path-selector.h"
  10 #include "dm-uevent.h"
  11
  12 #include <linux/ctype.h>
  13 #include <linux/init.h>
  14 #include <linux/mempool.h>
  15 #include <linux/module.h>
  16 #include <linux/pagemap.h>
  17 #include <linux/slab.h>
  18 #include <linux/time.h>
  19 #include <linux/workqueue.h>
  20 #include <scsi/scsi_dh.h>
  21 #include <asm/atomic.h>
  22
  23 #define DM_MSG_PREFIX "multipath"
  24 #define MESG_STR(x) x, sizeof(x)
  25
  26 /* Path properties */
  27 struct pgpath {
  28         struct list_head list;
  29
  30         struct priority_group *pg;      /* Owning PG */
  31         unsigned is_active;             /* Path status */
  32         unsigned fail_count;            /* Cumulative failure count */
  33
  34         struct dm_path path;
  35         struct work_struct deactivate_path;
  36        struct work_struct activate_path;
  37 };
  38
  39 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  40
  41 /*
  42  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  43  * Each has a path selector which controls which path gets used.
  44  */
  45 struct priority_group {
  46         struct list_head list;
  47
  48         struct multipath *m;            /* Owning multipath instance */
  49         struct path_selector ps;
  50
  51         unsigned pg_num;                /* Reference number */
  52         unsigned bypassed;              /* Temporarily bypass this PG? */
  53
  54         unsigned nr_pgpaths;            /* Number of paths in PG */
  55         struct list_head pgpaths;
  56 };
  57
  58 /* Multipath context */
  59 struct multipath {
  60         struct list_head list;
  61         struct dm_target *ti;
  62
  63         spinlock_t lock;
  64
  65         const char *hw_handler_name;
  66         unsigned nr_priority_groups;
  67         struct list_head priority_groups;
  68         unsigned pg_init_required;      /* pg_init needs calling? */
  69         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  70
  71         unsigned nr_valid_paths;        /* Total number of usable paths */
  72         struct pgpath *current_pgpath;
  73         struct priority_group *current_pg;
  74         struct priority_group *next_pg; /* Switch to this PG if set */
  75         unsigned repeat_count;          /* I/Os left before calling PS again */
  76
  77         unsigned queue_io;              /* Must we queue all I/O? */
  78         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  79         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  80         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  81         unsigned pg_init_count;         /* Number of times pg_init called */
  82
  83         struct work_struct process_queued_ios;
  84         struct list_head queued_ios;
  85         unsigned queue_size;
  86
  87         struct work_struct trigger_event;
  88
  89         /*
  90          * We must use a mempool of dm_mpath_io structs so that we
  91          * can resubmit bios on error.
  92          */
  93         mempool_t *mpio_pool;
  94 };
  95
  96 /*
  97  * Context information attached to each bio we process.
  98  */
  99 struct dm_mpath_io {
 100         struct pgpath *pgpath;
 101         size_t nr_bytes;
 102 };
 103
 104 typedef int (*action_fn) (struct pgpath *pgpath);
 105
 106 #define MIN_IOS 256     /* Mempool size */
 107
 108 static struct kmem_cache *_mpio_cache;
 109
 110 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 111 static void process_queued_ios(struct work_struct *work);
 112 static void trigger_event(struct work_struct *work);
 113 static void activate_path(struct work_struct *work);
 114 static void deactivate_path(struct work_struct *work);
 115
 116
 117 /*-----------------------------------------------
 118  * Allocation routines
 119  *-----------------------------------------------*/
 120
 121 static struct pgpath *alloc_pgpath(void)
 122 {
 123         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 124
 125         if (pgpath) {
 126                 pgpath->is_active = 1;
 127                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
 128                 INIT_WORK(&pgpath->activate_path, activate_path);
 129         }
 130
 131         return pgpath;
 132 }
 133
 134 static void free_pgpath(struct pgpath *pgpath)
 135 {
 136         kfree(pgpath);
 137 }
 138
 139 static void deactivate_path(struct work_struct *work)
 140 {
 141         struct pgpath *pgpath =
 142                 container_of(work, struct pgpath, deactivate_path);
 143
 144         if (pgpath->path.dev)
 145                 blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
 146 }
 147
 148 static struct priority_group *alloc_priority_group(void)
 149 {
 150         struct priority_group *pg;
 151
 152         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 153
 154         if (pg)
 155                 INIT_LIST_HEAD(&pg->pgpaths);
 156
 157         return pg;
 158 }
 159
 160 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 161 {
 162         unsigned long flags;
 163         struct pgpath *pgpath, *tmp;
 164         struct multipath *m = ti->private;
 165
 166         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 167                 list_del(&pgpath->list);
 168                 dm_put_device(ti, pgpath->path.dev);
 169                 free_pgpath(pgpath);
 170         }
 171 }
 172
 173 static void free_priority_group(struct priority_group *pg,
 174                                 struct dm_target *ti)
 175 {
 176         struct path_selector *ps = &pg->ps;
 177
 178         if (ps->type) {
 179                 ps->type->destroy(ps);
 180                 dm_put_path_selector(ps->type);
 181         }
 182
 183         free_pgpaths(&pg->pgpaths, ti);
 184         kfree(pg);
 185 }
 186
 187 static struct multipath *alloc_multipath(struct dm_target *ti)
 188 {
 189         struct multipath *m;
 190
 191         m = kzalloc(sizeof(*m), GFP_KERNEL);
 192         if (m) {
 193                 INIT_LIST_HEAD(&m->priority_groups);
 194                 INIT_LIST_HEAD(&m->queued_ios);
 195                 spin_lock_init(&m->lock);
 196                 m->queue_io = 1;
 197                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 198                 INIT_WORK(&m->trigger_event, trigger_event);
 199                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 200                 if (!m->mpio_pool) {
 201                         kfree(m);
 202                         return NULL;
 203                 }
 204                 m->ti = ti;
 205                 ti->private = m;
 206         }
 207
 208         return m;
 209 }
 210
 211 static void free_multipath(struct multipath *m)
 212 {
 213         struct priority_group *pg, *tmp;
 214
 215         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 216                 list_del(&pg->list);
 217                 free_priority_group(pg, m->ti);
 218         }
 219
 220         kfree(m->hw_handler_name);
 221         mempool_destroy(m->mpio_pool);
 222         kfree(m);
 223 }
 224
 225
 226 /*-----------------------------------------------
 227  * Path selection
 228  *-----------------------------------------------*/
 229
 230 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 231 {
 232         m->current_pg = pgpath->pg;
 233
 234         /* Must we initialise the PG first, and queue I/O till it's ready? */
 235         if (m->hw_handler_name) {
 236                 m->pg_init_required = 1;
 237                 m->queue_io = 1;
 238         } else {
 239                 m->pg_init_required = 0;
 240                 m->queue_io = 0;
 241         }
 242
 243         m->pg_init_count = 0;
 244 }
 245
 246 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 247                                size_t nr_bytes)
 248 {
 249         struct dm_path *path;
 250
 251         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 252         if (!path)
 253                 return -ENXIO;
 254
 255         m->current_pgpath = path_to_pgpath(path);
 256
 257         if (!m->current_pgpath->path.dev) {
 258                 m->current_pgpath = NULL;
 259                 return -ENODEV;
 260         }
 261
 262         if (m->current_pg != pg)
 263                 __switch_pg(m, m->current_pgpath);
 264
 265         return 0;
 266 }
 267
 268 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 269 {
 270         struct priority_group *pg;
 271         unsigned bypassed = 1;
 272
 273         if (!m->nr_valid_paths)
 274                 goto failed;
 275
 276         /* Were we instructed to switch PG? */
 277         if (m->next_pg) {
 278                 pg = m->next_pg;
 279                 m->next_pg = NULL;
 280                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 281                         return;
 282         }
 283
 284         /* Don't change PG until it has no remaining paths */
 285         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 286                 return;
 287
 288         /*
 289          * Loop through priority groups until we find a valid path.
 290          * First time we skip PGs marked 'bypassed'.
 291          * Second time we only try the ones we skipped.
 292          */
 293         do {
 294                 list_for_each_entry(pg, &m->priority_groups, list) {
 295                         if (pg->bypassed == bypassed)
 296                                 continue;
 297                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 298                                 return;
 299                 }
 300         } while (bypassed--);
 301
 302 failed:
 303         m->current_pgpath = NULL;
 304         m->current_pg = NULL;
 305 }
 306
 307 /*
 308  * Check whether bios must be queued in the device-mapper core rather
 309  * than here in the target.
 310  *
 311  * m->lock must be held on entry.
 312  *
 313  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 314  * same value then we are not between multipath_presuspend()
 315  * and multipath_resume() calls and we have no need to check
 316  * for the DMF_NOFLUSH_SUSPENDING flag.
 317  */
 318 static int __must_push_back(struct multipath *m)
 319 {
 320         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 321                 dm_noflush_suspending(m->ti));
 322 }
 323
 324 static int map_io(struct multipath *m, struct request *clone,
 325                   struct dm_mpath_io *mpio, unsigned was_queued)
 326 {
 327         int r = DM_MAPIO_REMAPPED;
 328         size_t nr_bytes = blk_rq_bytes(clone);
 329         unsigned long flags;
 330         struct pgpath *pgpath;
 331         struct block_device *bdev;
 332
 333         spin_lock_irqsave(&m->lock, flags);
 334
 335         /* Do we need to select a new pgpath? */
 336         if (!m->current_pgpath ||
 337             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 338                 __choose_pgpath(m, nr_bytes);
 339
 340         pgpath = m->current_pgpath;
 341
 342         if (was_queued)
 343                 m->queue_size--;
 344
 345         if ((pgpath && m->queue_io) ||
 346             (!pgpath && m->queue_if_no_path)) {
 347                 /* Queue for the daemon to resubmit */
 348                 list_add_tail(&clone->queuelist, &m->queued_ios);
 349                 m->queue_size++;
 350                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 351                     !m->queue_io)
 352                         queue_work(kmultipathd, &m->process_queued_ios);
 353                 pgpath = NULL;
 354                 r = DM_MAPIO_SUBMITTED;
 355         } else if (pgpath) {
 356                 bdev = pgpath->path.dev->bdev;
 357                 clone->q = bdev_get_queue(bdev);
 358                 clone->rq_disk = bdev->bd_disk;
 359         } else if (__must_push_back(m))
 360                 r = DM_MAPIO_REQUEUE;
 361         else
 362                 r = -EIO;       /* Failed */
 363
 364         mpio->pgpath = pgpath;
 365         mpio->nr_bytes = nr_bytes;
 366
 367         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 368                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 369                                               nr_bytes);
 370
 371         spin_unlock_irqrestore(&m->lock, flags);
 372
 373         return r;
 374 }
 375
 376 /*
 377  * If we run out of usable paths, should we queue I/O or error it?
 378  */
 379 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 380                             unsigned save_old_value)
 381 {
 382         unsigned long flags;
 383
 384         spin_lock_irqsave(&m->lock, flags);
 385
 386         if (save_old_value)
 387                 m->saved_queue_if_no_path = m->queue_if_no_path;
 388         else
 389                 m->saved_queue_if_no_path = queue_if_no_path;
 390         m->queue_if_no_path = queue_if_no_path;
 391         if (!m->queue_if_no_path && m->queue_size)
 392                 queue_work(kmultipathd, &m->process_queued_ios);
 393
 394         spin_unlock_irqrestore(&m->lock, flags);
 395
 396         return 0;
 397 }
 398
 399 /*-----------------------------------------------------------------
 400  * The multipath daemon is responsible for resubmitting queued ios.
 401  *---------------------------------------------------------------*/
 402
 403 static void dispatch_queued_ios(struct multipath *m)
 404 {
 405         int r;
 406         unsigned long flags;
 407         struct dm_mpath_io *mpio;
 408         union map_info *info;
 409         struct request *clone, *n;
 410         LIST_HEAD(cl);
 411
 412         spin_lock_irqsave(&m->lock, flags);
 413         list_splice_init(&m->queued_ios, &cl);
 414         spin_unlock_irqrestore(&m->lock, flags);
 415
 416         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 417                 list_del_init(&clone->queuelist);
 418
 419                 info = dm_get_rq_mapinfo(clone);
 420                 mpio = info->ptr;
 421
 422                 r = map_io(m, clone, mpio, 1);
 423                 if (r < 0) {
 424                         mempool_free(mpio, m->mpio_pool);
 425                         dm_kill_request(clone, r);
 426                 } else if (r == DM_MAPIO_REMAPPED)
 427                         dm_dispatch_request(clone);
 428                 else if (r == DM_MAPIO_REQUEUE) {
 429                         mempool_free(mpio, m->mpio_pool);
 430                         dm_requeue_request(clone);
 431                 }
 432         }
 433 }
 434
 435 static void process_queued_ios(struct work_struct *work)
 436 {
 437         struct multipath *m =
 438                 container_of(work, struct multipath, process_queued_ios);
 439        struct pgpath *pgpath = NULL, *tmp;
 440        unsigned must_queue = 1;
 441         unsigned long flags;
 442
 443         spin_lock_irqsave(&m->lock, flags);
 444
 445         if (!m->queue_size)
 446                 goto out;
 447
 448         if (!m->current_pgpath)
 449                 __choose_pgpath(m, 1 << 19); /* Assume 512 KB */
 450
 451         pgpath = m->current_pgpath;
 452
 453         if ((pgpath && !m->queue_io) ||
 454             (!pgpath && !m->queue_if_no_path))
 455                 must_queue = 0;
 456
 457         if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
 458                 m->pg_init_count++;
 459                 m->pg_init_required = 0;
 460                 list_for_each_entry(tmp, &pgpath->pg->pgpaths, list) {
 461                         queue_work(kmpath_handlerd, &tmp->activate_path);
 462                         m->pg_init_in_progress++;
 463                 }
 464         }
 465 out:
 466         spin_unlock_irqrestore(&m->lock, flags);
 467         if (!must_queue)
 468                 dispatch_queued_ios(m);
 469 }
 470
 471 /*
 472  * An event is triggered whenever a path is taken out of use.
 473  * Includes path failure and PG bypass.
 474  */
 475 static void trigger_event(struct work_struct *work)
 476 {
 477         struct multipath *m =
 478                 container_of(work, struct multipath, trigger_event);
 479
 480         dm_table_event(m->ti->table);
 481 }
 482
 483 /*-----------------------------------------------------------------
 484  * Constructor/argument parsing:
 485  * <#multipath feature args> [<arg>]*
 486  * <#hw_handler args> [hw_handler [<arg>]*]
 487  * <#priority groups>
 488  * <initial priority group>
 489  *     [<selector> <#selector args> [<arg>]*
 490  *      <#paths> <#per-path selector args>
 491  *         [<path> [<arg>]* ]+ ]+
 492  *---------------------------------------------------------------*/
 493 struct param {
 494         unsigned min;
 495         unsigned max;
 496         char *error;
 497 };
 498
 499 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 500 {
 501         if (!str ||
 502             (sscanf(str, "%u", v) != 1) ||
 503             (*v < param->min) ||
 504             (*v > param->max)) {
 505                 *error = param->error;
 506                 return -EINVAL;
 507         }
 508
 509         return 0;
 510 }
 511
 512 struct arg_set {
 513         unsigned argc;
 514         char **argv;
 515 };
 516
 517 static char *shift(struct arg_set *as)
 518 {
 519         char *r;
 520
 521         if (as->argc) {
 522                 as->argc--;
 523                 r = *as->argv;
 524                 as->argv++;
 525                 return r;
 526         }
 527
 528         return NULL;
 529 }
 530
 531 static void consume(struct arg_set *as, unsigned n)
 532 {
 533         BUG_ON (as->argc < n);
 534         as->argc -= n;
 535         as->argv += n;
 536 }
 537
 538 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 539                                struct dm_target *ti)
 540 {
 541         int r;
 542         struct path_selector_type *pst;
 543         unsigned ps_argc;
 544
 545         static struct param _params[] = {
 546                 {0, 1024, "invalid number of path selector args"},
 547         };
 548
 549         pst = dm_get_path_selector(shift(as));
 550         if (!pst) {
 551                 ti->error = "unknown path selector type";
 552                 return -EINVAL;
 553         }
 554
 555         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 556         if (r) {
 557                 dm_put_path_selector(pst);
 558                 return -EINVAL;
 559         }
 560
 561         r = pst->create(&pg->ps, ps_argc, as->argv);
 562         if (r) {
 563                 dm_put_path_selector(pst);
 564                 ti->error = "path selector constructor failed";
 565                 return r;
 566         }
 567
 568         pg->ps.type = pst;
 569         consume(as, ps_argc);
 570
 571         return 0;
 572 }
 573
 574 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 575                                struct dm_target *ti)
 576 {
 577         int r;
 578         struct pgpath *p;
 579         char *path;
 580         struct multipath *m = ti->private;
 581
 582         /* we need at least a path arg */
 583         if (as->argc < 1) {
 584                 ti->error = "no device given";
 585                 return NULL;
 586         }
 587
 588         p = alloc_pgpath();
 589         if (!p)
 590                 return NULL;
 591
 592         path = shift(as);
 593         r = dm_get_device(ti, path, ti->begin, ti->len,
 594                           dm_table_get_mode(ti->table), &p->path.dev);
 595         if (r) {
 596                 unsigned major, minor;
 597
 598                 /* Try to add a failed device */
 599                 if (r == -ENXIO && sscanf(path, "%u:%u", &major, &minor) == 2) {
 600                         dev_t dev;
 601
 602                         /* Extract the major/minor numbers */
 603                         dev = MKDEV(major, minor);
 604                         if (MAJOR(dev) != major || MINOR(dev) != minor) {
 605                                 /* Nice try, didn't work */
 606                                 DMWARN("Invalid device path %s", path);
 607                                 ti->error = "error converting devnum";
 608                                 goto bad;
 609                         }
 610                         DMWARN("adding disabled device %d:%d", major, minor);
 611                         p->path.dev = NULL;
 612                         format_dev_t(p->path.pdev, dev);
 613                         p->is_active = 0;
 614                 } else {
 615                         ti->error = "error getting device";
 616                         goto bad;
 617                 }
 618         } else {
 619                 memcpy(p->path.pdev, p->path.dev->name, 16);
 620         }
 621
 622         if (m->hw_handler_name && p->path.dev) {
 623                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 624
 625                 r = scsi_dh_attach(q, m->hw_handler_name);
 626                 if (r == -EBUSY) {
 627                         /*
 628                          * Already attached to different hw_handler,
 629                          * try to reattach with correct one.
 630                          */
 631                         scsi_dh_detach(q);
 632                         r = scsi_dh_attach(q, m->hw_handler_name);
 633                 }
 634                 if (r < 0) {
 635                         ti->error = "error attaching hardware handler";
 636                         dm_put_device(ti, p->path.dev);
 637                         goto bad;
 638                 }
 639         }
 640
 641         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 642         if (r) {
 643                 dm_put_device(ti, p->path.dev);
 644                 goto bad;
 645         }
 646
 647         return p;
 648
 649  bad:
 650         free_pgpath(p);
 651         return NULL;
 652 }
 653
 654 static struct priority_group *parse_priority_group(struct arg_set *as,
 655                                                    struct multipath *m)
 656 {
 657         static struct param _params[] = {
 658                 {1, 1024, "invalid number of paths"},
 659                 {0, 1024, "invalid number of selector args"}
 660         };
 661
 662         int r;
 663         unsigned i, nr_selector_args, nr_params;
 664         struct priority_group *pg;
 665         struct dm_target *ti = m->ti;
 666
 667         if (as->argc < 2) {
 668                 as->argc = 0;
 669                 ti->error = "not enough priority group aruments";
 670                 return NULL;
 671         }
 672
 673         pg = alloc_priority_group();
 674         if (!pg) {
 675                 ti->error = "couldn't allocate priority group";
 676                 return NULL;
 677         }
 678         pg->m = m;
 679
 680         r = parse_path_selector(as, pg, ti);
 681         if (r)
 682                 goto bad;
 683
 684         /*
 685          * read the paths
 686          */
 687         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 688         if (r)
 689                 goto bad;
 690
 691         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 692         if (r)
 693                 goto bad;
 694
 695         nr_params = 1 + nr_selector_args;
 696         for (i = 0; i < pg->nr_pgpaths; i++) {
 697                 struct pgpath *pgpath;
 698                 struct arg_set path_args;
 699
 700                 if (as->argc < nr_params) {
 701                         ti->error = "not enough path parameters";
 702                         goto bad;
 703                 }
 704
 705                 path_args.argc = nr_params;
 706                 path_args.argv = as->argv;
 707
 708                 pgpath = parse_path(&path_args, &pg->ps, ti);
 709                 if (!pgpath)
 710                         goto bad;
 711
 712                 pgpath->pg = pg;
 713                 list_add_tail(&pgpath->list, &pg->pgpaths);
 714                 consume(as, nr_params);
 715         }
 716
 717         return pg;
 718
 719  bad:
 720         free_priority_group(pg, ti);
 721         return NULL;
 722 }
 723
 724 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 725 {
 726         unsigned hw_argc;
 727         struct dm_target *ti = m->ti;
 728
 729         static struct param _params[] = {
 730                 {0, 1024, "invalid number of hardware handler args"},
 731         };
 732
 733         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 734                 return -EINVAL;
 735
 736         if (!hw_argc)
 737                 return 0;
 738
 739         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 740         request_module("scsi_dh_%s", m->hw_handler_name);
 741         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 742                 ti->error = "unknown hardware handler type";
 743                 kfree(m->hw_handler_name);
 744                 m->hw_handler_name = NULL;
 745                 return -EINVAL;
 746         }
 747         consume(as, hw_argc - 1);
 748
 749         return 0;
 750 }
 751
 752 static int parse_features(struct arg_set *as, struct multipath *m)
 753 {
 754         int r;
 755         unsigned argc;
 756         struct dm_target *ti = m->ti;
 757         const char *param_name;
 758
 759         static struct param _params[] = {
 760                 {0, 3, "invalid number of feature args"},
 761                 {1, 50, "pg_init_retries must be between 1 and 50"},
 762         };
 763
 764         r = read_param(_params, shift(as), &argc, &ti->error);
 765         if (r)
 766                 return -EINVAL;
 767
 768         if (!argc)
 769                 return 0;
 770
 771         do {
 772                 param_name = shift(as);
 773                 argc--;
 774
 775                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 776                         r = queue_if_no_path(m, 1, 0);
 777                         continue;
 778                 }
 779
 780                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 781                     (argc >= 1)) {
 782                         r = read_param(_params + 1, shift(as),
 783                                        &m->pg_init_retries, &ti->error);
 784                         argc--;
 785                         continue;
 786                 }
 787
 788                 ti->error = "Unrecognised multipath feature request";
 789                 r = -EINVAL;
 790         } while (argc && !r);
 791
 792         return r;
 793 }
 794
 795 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 796                          char **argv)
 797 {
 798         /* target parameters */
 799         static struct param _params[] = {
 800                 {0, 1024, "invalid number of priority groups"},
 801                 {0, 1024, "invalid initial priority group number"},
 802         };
 803
 804         int r;
 805         struct multipath *m;
 806         struct arg_set as;
 807         unsigned pg_count = 0;
 808         unsigned next_pg_num;
 809
 810         as.argc = argc;
 811         as.argv = argv;
 812
 813         m = alloc_multipath(ti);
 814         if (!m) {
 815                 ti->error = "can't allocate multipath";
 816                 return -EINVAL;
 817         }
 818
 819         r = parse_features(&as, m);
 820         if (r)
 821                 goto bad;
 822
 823         r = parse_hw_handler(&as, m);
 824         if (r)
 825                 goto bad;
 826
 827         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 828         if (r)
 829                 goto bad;
 830
 831         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 832         if (r)
 833                 goto bad;
 834
 835         /* parse the priority groups */
 836         while (as.argc) {
 837                 struct priority_group *pg;
 838
 839                 pg = parse_priority_group(&as, m);
 840                 if (!pg) {
 841                         r = -EINVAL;
 842                         goto bad;
 843                 }
 844
 845                 m->nr_valid_paths += pg->nr_pgpaths;
 846                 list_add_tail(&pg->list, &m->priority_groups);
 847                 pg_count++;
 848                 pg->pg_num = pg_count;
 849                 if (!--next_pg_num)
 850                         m->next_pg = pg;
 851         }
 852
 853         if (pg_count != m->nr_priority_groups) {
 854                 ti->error = "priority group count mismatch";
 855                 r = -EINVAL;
 856                 goto bad;
 857         }
 858
 859         return 0;
 860
 861  bad:
 862         free_multipath(m);
 863         return r;
 864 }
 865
 866 static void multipath_dtr(struct dm_target *ti)
 867 {
 868         struct multipath *m = (struct multipath *) ti->private;
 869
 870         flush_workqueue(kmpath_handlerd);
 871         flush_workqueue(kmultipathd);
 872         free_multipath(m);
 873 }
 874
 875 /*
 876  * Map cloned requests
 877  */
 878 static int multipath_map(struct dm_target *ti, struct request *clone,
 879                          union map_info *map_context)
 880 {
 881         int r;
 882         struct dm_mpath_io *mpio;
 883         struct multipath *m = (struct multipath *) ti->private;
 884
 885         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 886         if (!mpio)
 887                 /* ENOMEM, requeue */
 888                 return DM_MAPIO_REQUEUE;
 889         memset(mpio, 0, sizeof(*mpio));
 890
 891         map_context->ptr = mpio;
 892         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 893         r = map_io(m, clone, mpio, 0);
 894         if (r < 0 || r == DM_MAPIO_REQUEUE)
 895                 mempool_free(mpio, m->mpio_pool);
 896
 897         return r;
 898 }
 899
 900 /*
 901  * Take a path out of use.
 902  */
 903 static int fail_path(struct pgpath *pgpath)
 904 {
 905         unsigned long flags;
 906         struct multipath *m = pgpath->pg->m;
 907
 908         spin_lock_irqsave(&m->lock, flags);
 909
 910         if (!pgpath->is_active)
 911                 goto out;
 912
 913         DMWARN("Failing path %s.", pgpath->path.pdev);
 914
 915         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 916         pgpath->is_active = 0;
 917         pgpath->fail_count++;
 918
 919         m->nr_valid_paths--;
 920
 921         if (pgpath == m->current_pgpath)
 922                 m->current_pgpath = NULL;
 923
 924         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 925                        pgpath->path.pdev, m->nr_valid_paths);
 926
 927         queue_work(kmultipathd, &m->trigger_event);
 928         queue_work(kmultipathd, &pgpath->deactivate_path);
 929
 930 out:
 931         spin_unlock_irqrestore(&m->lock, flags);
 932
 933         return 0;
 934 }
 935
 936 /*
 937  * Reinstate a previously-failed path
 938  */
 939 static int reinstate_path(struct pgpath *pgpath)
 940 {
 941         int r = 0;
 942         unsigned long flags;
 943         struct multipath *m = pgpath->pg->m;
 944
 945         spin_lock_irqsave(&m->lock, flags);
 946
 947         if (pgpath->is_active)
 948                 goto out;
 949
 950         if (!pgpath->path.dev) {
 951                 DMWARN("Cannot reinstate disabled path %s", pgpath->path.pdev);
 952                 r = -ENODEV;
 953                 goto out;
 954         }
 955
 956         if (!pgpath->pg->ps.type->reinstate_path) {
 957                 DMWARN("Reinstate path not supported by path selector %s",
 958                        pgpath->pg->ps.type->name);
 959                 r = -EINVAL;
 960                 goto out;
 961         }
 962
 963         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 964         if (r)
 965                 goto out;
 966
 967         pgpath->is_active = 1;
 968
 969         m->current_pgpath = NULL;
 970         if (!m->nr_valid_paths++ && m->queue_size)
 971                 queue_work(kmultipathd, &m->process_queued_ios);
 972
 973         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 974                        pgpath->path.pdev, m->nr_valid_paths);
 975
 976         queue_work(kmultipathd, &m->trigger_event);
 977
 978 out:
 979         spin_unlock_irqrestore(&m->lock, flags);
 980
 981         return r;
 982 }
 983
 984 /*
 985  * Fail or reinstate all paths that match the provided struct dm_dev.
 986  */
 987 static int action_dev(struct multipath *m, struct dm_dev *dev,
 988                       action_fn action)
 989 {
 990         int r = 0;
 991         struct pgpath *pgpath;
 992         struct priority_group *pg;
 993
 994         if (!dev)
 995                 return 0;
 996
 997         list_for_each_entry(pg, &m->priority_groups, list) {
 998                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 999                         if (pgpath->path.dev == dev)
1000                                 r = action(pgpath);
1001                 }
1002         }
1003
1004         return r;
1005 }
1006
1007 /*
1008  * Temporarily try to avoid having to use the specified PG
1009  */
1010 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1011                       int bypassed)
1012 {
1013         unsigned long flags;
1014
1015         spin_lock_irqsave(&m->lock, flags);
1016
1017         pg->bypassed = bypassed;
1018         m->current_pgpath = NULL;
1019         m->current_pg = NULL;
1020
1021         spin_unlock_irqrestore(&m->lock, flags);
1022
1023         queue_work(kmultipathd, &m->trigger_event);
1024 }
1025
1026 /*
1027  * Switch to using the specified PG from the next I/O that gets mapped
1028  */
1029 static int switch_pg_num(struct multipath *m, const char *pgstr)
1030 {
1031         struct priority_group *pg;
1032         unsigned pgnum;
1033         unsigned long flags;
1034
1035         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1036             (pgnum > m->nr_priority_groups)) {
1037                 DMWARN("invalid PG number supplied to switch_pg_num");
1038                 return -EINVAL;
1039         }
1040
1041         spin_lock_irqsave(&m->lock, flags);
1042         list_for_each_entry(pg, &m->priority_groups, list) {
1043                 pg->bypassed = 0;
1044                 if (--pgnum)
1045                         continue;
1046
1047                 m->current_pgpath = NULL;
1048                 m->current_pg = NULL;
1049                 m->next_pg = pg;
1050         }
1051         spin_unlock_irqrestore(&m->lock, flags);
1052
1053         queue_work(kmultipathd, &m->trigger_event);
1054         return 0;
1055 }
1056
1057 /*
1058  * Set/clear bypassed status of a PG.
1059  * PGs are numbered upwards from 1 in the order they were declared.
1060  */
1061 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1062 {
1063         struct priority_group *pg;
1064         unsigned pgnum;
1065
1066         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1067             (pgnum > m->nr_priority_groups)) {
1068                 DMWARN("invalid PG number supplied to bypass_pg");
1069                 return -EINVAL;
1070         }
1071
1072         list_for_each_entry(pg, &m->priority_groups, list) {
1073                 if (!--pgnum)
1074                         break;
1075         }
1076
1077         bypass_pg(m, pg, bypassed);
1078         return 0;
1079 }
1080
1081 /*
1082  * Should we retry pg_init immediately?
1083  */
1084 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1085 {
1086         unsigned long flags;
1087         int limit_reached = 0;
1088
1089         spin_lock_irqsave(&m->lock, flags);
1090
1091         if (m->pg_init_count <= m->pg_init_retries)
1092                 m->pg_init_required = 1;
1093         else
1094                 limit_reached = 1;
1095
1096         spin_unlock_irqrestore(&m->lock, flags);
1097
1098         return limit_reached;
1099 }
1100
1101 static void pg_init_done(struct dm_path *path, int errors)
1102 {
1103         struct pgpath *pgpath = path_to_pgpath(path);
1104         struct priority_group *pg = pgpath->pg;
1105         struct multipath *m = pg->m;
1106         unsigned long flags;
1107
1108         /* device or driver problems */
1109         switch (errors) {
1110         case SCSI_DH_OK:
1111                 break;
1112         case SCSI_DH_NOSYS:
1113                 if (!m->hw_handler_name) {
1114                         errors = 0;
1115                         break;
1116                 }
1117                 DMERR("Cannot failover device because scsi_dh_%s was not "
1118                       "loaded.", m->hw_handler_name);
1119                 /*
1120                  * Fail path for now, so we do not ping pong
1121                  */
1122                 fail_path(pgpath);
1123                 break;
1124         case SCSI_DH_DEV_TEMP_BUSY:
1125                 /*
1126                  * Probably doing something like FW upgrade on the
1127                  * controller so try the other pg.
1128                  */
1129                 bypass_pg(m, pg, 1);
1130                 break;
1131         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1132         case SCSI_DH_RETRY:
1133         case SCSI_DH_IMM_RETRY:
1134         case SCSI_DH_RES_TEMP_UNAVAIL:
1135                 if (pg_init_limit_reached(m, pgpath))
1136                         fail_path(pgpath);
1137                 errors = 0;
1138                 break;
1139         default:
1140                 /*
1141                  * We probably do not want to fail the path for a device
1142                  * error, but this is what the old dm did. In future
1143                  * patches we can do more advanced handling.
1144                  */
1145                 fail_path(pgpath);
1146         }
1147
1148         spin_lock_irqsave(&m->lock, flags);
1149         if (errors) {
1150                 DMERR("Could not failover device. Error %d.", errors);
1151                 m->current_pgpath = NULL;
1152                 m->current_pg = NULL;
1153         } else if (!m->pg_init_required) {
1154                 m->queue_io = 0;
1155                 pg->bypassed = 0;
1156         }
1157
1158        m->pg_init_in_progress--;
1159        if (!m->pg_init_in_progress)
1160                queue_work(kmultipathd, &m->process_queued_ios);
1161        spin_unlock_irqrestore(&m->lock, flags);
1162 }
1163
1164 static void activate_path(struct work_struct *work)
1165 {
1166         int ret = SCSI_DH_DEV_OFFLINED;
1167         struct pgpath *pgpath =
1168                 container_of(work, struct pgpath, activate_path);
1169
1170         if (pgpath->path.dev)
1171                 ret = scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev));
1172         pg_init_done(&pgpath->path, ret);
1173 }
1174
1175 /*
1176  * end_io handling
1177  */
1178 static int do_end_io(struct multipath *m, struct request *clone,
1179                      int error, struct dm_mpath_io *mpio)
1180 {
1181         /*
1182          * We don't queue any clone request inside the multipath target
1183          * during end I/O handling, since those clone requests don't have
1184          * bio clones.  If we queue them inside the multipath target,
1185          * we need to make bio clones, that requires memory allocation.
1186          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1187          *  don't have bio clones.)
1188          * Instead of queueing the clone request here, we queue the original
1189          * request into dm core, which will remake a clone request and
1190          * clone bios for it and resubmit it later.
1191          */
1192         int r = DM_ENDIO_REQUEUE;
1193         unsigned long flags;
1194
1195         if (!error && !clone->errors)
1196                 return 0;       /* I/O complete */
1197
1198         if (error == -EOPNOTSUPP)
1199                 return error;
1200
1201         if (mpio->pgpath)
1202                 fail_path(mpio->pgpath);
1203
1204         spin_lock_irqsave(&m->lock, flags);
1205         if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1206                 r = -EIO;
1207         spin_unlock_irqrestore(&m->lock, flags);
1208
1209         return r;
1210 }
1211
1212 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1213                             int error, union map_info *map_context)
1214 {
1215         struct multipath *m = ti->private;
1216         struct dm_mpath_io *mpio = map_context->ptr;
1217         struct pgpath *pgpath = mpio->pgpath;
1218         struct path_selector *ps;
1219         int r;
1220
1221         r  = do_end_io(m, clone, error, mpio);
1222         if (pgpath) {
1223                 ps = &pgpath->pg->ps;
1224                 if (ps->type->end_io)
1225                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1226         }
1227         mempool_free(mpio, m->mpio_pool);
1228
1229         return r;
1230 }
1231
1232 /*
1233  * Suspend can't complete until all the I/O is processed so if
1234  * the last path fails we must error any remaining I/O.
1235  * Note that if the freeze_bdev fails while suspending, the
1236  * queue_if_no_path state is lost - userspace should reset it.
1237  */
1238 static void multipath_presuspend(struct dm_target *ti)
1239 {
1240         struct multipath *m = (struct multipath *) ti->private;
1241
1242         queue_if_no_path(m, 0, 1);
1243 }
1244
1245 /*
1246  * Restore the queue_if_no_path setting.
1247  */
1248 static void multipath_resume(struct dm_target *ti)
1249 {
1250         struct multipath *m = (struct multipath *) ti->private;
1251         unsigned long flags;
1252
1253         spin_lock_irqsave(&m->lock, flags);
1254         m->queue_if_no_path = m->saved_queue_if_no_path;
1255         spin_unlock_irqrestore(&m->lock, flags);
1256 }
1257
1258 /*
1259  * Info output has the following format:
1260  * num_multipath_feature_args [multipath_feature_args]*
1261  * num_handler_status_args [handler_status_args]*
1262  * num_groups init_group_number
1263  *            [A|D|E num_ps_status_args [ps_status_args]*
1264  *             num_paths num_selector_args
1265  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1266  *
1267  * Table output has the following format (identical to the constructor string):
1268  * num_feature_args [features_args]*
1269  * num_handler_args hw_handler [hw_handler_args]*
1270  * num_groups init_group_number
1271  *     [priority selector-name num_ps_args [ps_args]*
1272  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1273  */
1274 static int multipath_status(struct dm_target *ti, status_type_t type,
1275                             char *result, unsigned int maxlen)
1276 {
1277         int sz = 0;
1278         unsigned long flags;
1279         struct multipath *m = (struct multipath *) ti->private;
1280         struct priority_group *pg;
1281         struct pgpath *p;
1282         unsigned pg_num;
1283         char state;
1284
1285         spin_lock_irqsave(&m->lock, flags);
1286
1287         /* Features */
1288         if (type == STATUSTYPE_INFO)
1289                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1290         else {
1291                 DMEMIT("%u ", m->queue_if_no_path +
1292                               (m->pg_init_retries > 0) * 2);
1293                 if (m->queue_if_no_path)
1294                         DMEMIT("queue_if_no_path ");
1295                 if (m->pg_init_retries)
1296                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1297         }
1298
1299         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1300                 DMEMIT("0 ");
1301         else
1302                 DMEMIT("1 %s ", m->hw_handler_name);
1303
1304         DMEMIT("%u ", m->nr_priority_groups);
1305
1306         if (m->next_pg)
1307                 pg_num = m->next_pg->pg_num;
1308         else if (m->current_pg)
1309                 pg_num = m->current_pg->pg_num;
1310         else
1311                         pg_num = 1;
1312
1313         DMEMIT("%u ", pg_num);
1314
1315         switch (type) {
1316         case STATUSTYPE_INFO:
1317                 list_for_each_entry(pg, &m->priority_groups, list) {
1318                         if (pg->bypassed)
1319                                 state = 'D';    /* Disabled */
1320                         else if (pg == m->current_pg)
1321                                 state = 'A';    /* Currently Active */
1322                         else
1323                                 state = 'E';    /* Enabled */
1324
1325                         DMEMIT("%c ", state);
1326
1327                         if (pg->ps.type->status)
1328                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1329                                                           result + sz,
1330                                                           maxlen - sz);
1331                         else
1332                                 DMEMIT("0 ");
1333
1334                         DMEMIT("%u %u ", pg->nr_pgpaths,
1335                                pg->ps.type->info_args);
1336
1337                         list_for_each_entry(p, &pg->pgpaths, list) {
1338                                 DMEMIT("%s %s %u ", p->path.pdev,
1339                                        p->is_active ? "A" : "F",
1340                                        p->fail_count);
1341                                 if (pg->ps.type->status)
1342                                         sz += pg->ps.type->status(&pg->ps,
1343                                               &p->path, type, result + sz,
1344                                               maxlen - sz);
1345                         }
1346                 }
1347                 break;
1348
1349         case STATUSTYPE_TABLE:
1350                 list_for_each_entry(pg, &m->priority_groups, list) {
1351                         DMEMIT("%s ", pg->ps.type->name);
1352
1353                         if (pg->ps.type->status)
1354                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1355                                                           result + sz,
1356                                                           maxlen - sz);
1357                         else
1358                                 DMEMIT("0 ");
1359
1360                         DMEMIT("%u %u ", pg->nr_pgpaths,
1361                                pg->ps.type->table_args);
1362
1363                         list_for_each_entry(p, &pg->pgpaths, list) {
1364                                 DMEMIT("%s ", p->path.pdev);
1365                                 if (pg->ps.type->status)
1366                                         sz += pg->ps.type->status(&pg->ps,
1367                                               &p->path, type, result + sz,
1368                                               maxlen - sz);
1369                         }
1370                 }
1371                 break;
1372         }
1373
1374         spin_unlock_irqrestore(&m->lock, flags);
1375
1376         return 0;
1377 }
1378
1379 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1380 {
1381         int r;
1382         struct dm_dev *dev;
1383         struct multipath *m = (struct multipath *) ti->private;
1384         action_fn action;
1385
1386         if (argc == 1) {
1387                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1388                         return queue_if_no_path(m, 1, 0);
1389                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1390                         return queue_if_no_path(m, 0, 0);
1391         }
1392
1393         if (argc != 2)
1394                 goto error;
1395
1396         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1397                 return bypass_pg_num(m, argv[1], 1);
1398         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1399                 return bypass_pg_num(m, argv[1], 0);
1400         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1401                 return switch_pg_num(m, argv[1]);
1402         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1403                 action = reinstate_path;
1404         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1405                 action = fail_path;
1406         else
1407                 goto error;
1408
1409         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1410                           dm_table_get_mode(ti->table), &dev);
1411         if (r) {
1412                 DMWARN("message: error getting device %s",
1413                        argv[1]);
1414                 return -EINVAL;
1415         }
1416
1417         r = action_dev(m, dev, action);
1418
1419         dm_put_device(ti, dev);
1420
1421         return r;
1422
1423 error:
1424         DMWARN("Unrecognised multipath message received.");
1425         return -EINVAL;
1426 }
1427
1428 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1429                            struct file *filp, unsigned int cmd,
1430                            unsigned long arg)
1431 {
1432         struct multipath *m = (struct multipath *) ti->private;
1433         struct block_device *bdev = NULL;
1434         unsigned long flags;
1435         struct file fake_file = {};
1436         struct dentry fake_dentry = {};
1437         int r = 0;
1438
1439         fake_file.f_path.dentry = &fake_dentry;
1440
1441         spin_lock_irqsave(&m->lock, flags);
1442
1443         if (!m->current_pgpath)
1444                 __choose_pgpath(m, 1 << 19); /* Assume 512KB */
1445
1446         if (m->current_pgpath && m->current_pgpath->path.dev) {
1447                 bdev = m->current_pgpath->path.dev->bdev;
1448                 fake_dentry.d_inode = bdev->bd_inode;
1449                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1450         }
1451
1452         if (m->queue_io)
1453                 r = -EAGAIN;
1454         else if (!bdev)
1455                 r = -EIO;
1456
1457         spin_unlock_irqrestore(&m->lock, flags);
1458
1459         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1460                                          bdev->bd_disk, cmd, arg);
1461 }
1462
1463 static int __pgpath_busy(struct pgpath *pgpath)
1464 {
1465         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1466
1467         return dm_underlying_device_busy(q);
1468 }
1469
1470 /*
1471  * We return "busy", only when we can map I/Os but underlying devices
1472  * are busy (so even if we map I/Os now, the I/Os will wait on
1473  * the underlying queue).
1474  * In other words, if we want to kill I/Os or queue them inside us
1475  * due to map unavailability, we don't return "busy".  Otherwise,
1476  * dm core won't give us the I/Os and we can't do what we want.
1477  */
1478 static int multipath_busy(struct dm_target *ti)
1479 {
1480         int busy = 0, has_active = 0;
1481         struct multipath *m = (struct multipath *) ti->private;
1482         struct priority_group *pg;
1483         struct pgpath *pgpath;
1484         unsigned long flags;
1485
1486         spin_lock_irqsave(&m->lock, flags);
1487
1488         /* Guess which priority_group will be used at next mapping time */
1489         if (unlikely(!m->current_pgpath && m->next_pg))
1490                 pg = m->next_pg;
1491         else if (likely(m->current_pg))
1492                 pg = m->current_pg;
1493         else
1494                 /*
1495                  * We don't know which pg will be used at next mapping time.
1496                  * We don't call __choose_pgpath() here to avoid to trigger
1497                  * pg_init just by busy checking.
1498                  * So we don't know whether underlying devices we will be using
1499                  * at next mapping time are busy or not. Just try mapping.
1500                  */
1501                 goto out;
1502
1503         /*
1504          * If there is one non-busy active path at least, the path selector
1505          * will be able to select it. So we consider such a pg as not busy.
1506          */
1507         busy = 1;
1508         list_for_each_entry(pgpath, &pg->pgpaths, list)
1509                 if (pgpath->is_active) {
1510                         has_active = 1;
1511
1512                         if (!__pgpath_busy(pgpath)) {
1513                                 busy = 0;
1514                                 break;
1515                         }
1516                 }
1517
1518         if (!has_active)
1519                 /*
1520                  * No active path in this pg, so this pg won't be used and
1521                  * the current_pg will be changed at next mapping time.
1522                  * We need to try mapping to determine it.
1523                  */
1524                 busy = 0;
1525
1526 out:
1527         spin_unlock_irqrestore(&m->lock, flags);
1528
1529         return busy;
1530 }
1531
1532 /*-----------------------------------------------------------------
1533  * Module setup
1534  *---------------------------------------------------------------*/
1535 static struct target_type multipath_target = {
1536         .name = "multipath",
1537         .version = {1, 0, 5},
1538         .module = THIS_MODULE,
1539         .ctr = multipath_ctr,
1540         .dtr = multipath_dtr,
1541         .map_rq = multipath_map,
1542         .rq_end_io = multipath_end_io,
1543         .presuspend = multipath_presuspend,
1544         .resume = multipath_resume,
1545         .status = multipath_status,
1546         .message = multipath_message,
1547         .ioctl  = multipath_ioctl,
1548         .busy = multipath_busy,
1549 };
1550
1551 static int __init dm_multipath_init(void)
1552 {
1553         int r;
1554
1555         /* allocate a slab for the dm_ios */
1556         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1557         if (!_mpio_cache)
1558                 return -ENOMEM;
1559
1560         r = dm_register_target(&multipath_target);
1561         if (r < 0) {
1562                 DMERR("register failed %d", r);
1563                 kmem_cache_destroy(_mpio_cache);
1564                 return -EINVAL;
1565         }
1566
1567         kmultipathd = create_workqueue("kmpathd");
1568         if (!kmultipathd) {
1569                 DMERR("failed to create workqueue kmpathd");
1570                 dm_unregister_target(&multipath_target);
1571                 kmem_cache_destroy(_mpio_cache);
1572                 return -ENOMEM;
1573         }
1574
1575         /*
1576          * A separate workqueue is used to handle the device handlers
1577          * to avoid overloading existing workqueue. Overloading the
1578          * old workqueue would also create a bottleneck in the
1579          * path of the storage hardware device activation.
1580          */
1581         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1582         if (!kmpath_handlerd) {
1583                 DMERR("failed to create workqueue kmpath_handlerd");
1584                 destroy_workqueue(kmultipathd);
1585                 dm_unregister_target(&multipath_target);
1586                 kmem_cache_destroy(_mpio_cache);
1587                 return -ENOMEM;
1588         }
1589
1590         DMINFO("version %u.%u.%u loaded",
1591                multipath_target.version[0], multipath_target.version[1],
1592                multipath_target.version[2]);
1593
1594         return r;
1595 }
1596
1597 static void __exit dm_multipath_exit(void)
1598 {
1599         int r;
1600
1601         destroy_workqueue(kmpath_handlerd);
1602         destroy_workqueue(kmultipathd);
1603
1604         r = dm_unregister_target(&multipath_target);
1605         if (r < 0)
1606                 DMERR("target unregister failed %d", r);
1607         kmem_cache_destroy(_mpio_cache);
1608 }
1609
1610 module_init(dm_multipath_init);
1611 module_exit(dm_multipath_exit);
1612
1613 MODULE_DESCRIPTION(DM_NAME " multipath target");
1614 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1615 MODULE_LICENSE("GPL");