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         struct pgpath *pgpath, *tmp;
 163
 164         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 165                 list_del(&pgpath->list);
 166                 dm_put_device(ti, pgpath->path.dev);
 167                 free_pgpath(pgpath);
 168         }
 169 }
 170
 171 static void free_priority_group(struct priority_group *pg,
 172                                 struct dm_target *ti)
 173 {
 174         struct path_selector *ps = &pg->ps;
 175
 176         if (ps->type) {
 177                 ps->type->destroy(ps);
 178                 dm_put_path_selector(ps->type);
 179         }
 180
 181         free_pgpaths(&pg->pgpaths, ti);
 182         kfree(pg);
 183 }
 184
 185 static struct multipath *alloc_multipath(struct dm_target *ti)
 186 {
 187         struct multipath *m;
 188
 189         m = kzalloc(sizeof(*m), GFP_KERNEL);
 190         if (m) {
 191                 INIT_LIST_HEAD(&m->priority_groups);
 192                 INIT_LIST_HEAD(&m->queued_ios);
 193                 spin_lock_init(&m->lock);
 194                 m->queue_io = 1;
 195                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 196                 INIT_WORK(&m->trigger_event, trigger_event);
 197                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 198                 if (!m->mpio_pool) {
 199                         kfree(m);
 200                         return NULL;
 201                 }
 202                 m->ti = ti;
 203                 ti->private = m;
 204         }
 205
 206         return m;
 207 }
 208
 209 static void free_multipath(struct multipath *m)
 210 {
 211         struct priority_group *pg, *tmp;
 212
 213         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 214                 list_del(&pg->list);
 215                 free_priority_group(pg, m->ti);
 216         }
 217
 218         kfree(m->hw_handler_name);
 219         mempool_destroy(m->mpio_pool);
 220         kfree(m);
 221 }
 222
 223
 224 /*-----------------------------------------------
 225  * Path selection
 226  *-----------------------------------------------*/
 227
 228 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 229 {
 230         m->current_pg = pgpath->pg;
 231
 232         /* Must we initialise the PG first, and queue I/O till it's ready? */
 233         if (m->hw_handler_name) {
 234                 m->pg_init_required = 1;
 235                 m->queue_io = 1;
 236         } else {
 237                 m->pg_init_required = 0;
 238                 m->queue_io = 0;
 239         }
 240
 241         m->pg_init_count = 0;
 242 }
 243
 244 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 245                                size_t nr_bytes)
 246 {
 247         struct dm_path *path;
 248
 249         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 250         if (!path)
 251                 return -ENXIO;
 252
 253         m->current_pgpath = path_to_pgpath(path);
 254
 255         if (!m->current_pgpath->path.dev) {
 256                 m->current_pgpath = NULL;
 257                 return -ENODEV;
 258         }
 259
 260         if (m->current_pg != pg)
 261                 __switch_pg(m, m->current_pgpath);
 262
 263         return 0;
 264 }
 265
 266 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 267 {
 268         struct priority_group *pg;
 269         unsigned bypassed = 1;
 270
 271         if (!m->nr_valid_paths)
 272                 goto failed;
 273
 274         /* Were we instructed to switch PG? */
 275         if (m->next_pg) {
 276                 pg = m->next_pg;
 277                 m->next_pg = NULL;
 278                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 279                         return;
 280         }
 281
 282         /* Don't change PG until it has no remaining paths */
 283         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 284                 return;
 285
 286         /*
 287          * Loop through priority groups until we find a valid path.
 288          * First time we skip PGs marked 'bypassed'.
 289          * Second time we only try the ones we skipped.
 290          */
 291         do {
 292                 list_for_each_entry(pg, &m->priority_groups, list) {
 293                         if (pg->bypassed == bypassed)
 294                                 continue;
 295                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 296                                 return;
 297                 }
 298         } while (bypassed--);
 299
 300 failed:
 301         m->current_pgpath = NULL;
 302         m->current_pg = NULL;
 303 }
 304
 305 /*
 306  * Check whether bios must be queued in the device-mapper core rather
 307  * than here in the target.
 308  *
 309  * m->lock must be held on entry.
 310  *
 311  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 312  * same value then we are not between multipath_presuspend()
 313  * and multipath_resume() calls and we have no need to check
 314  * for the DMF_NOFLUSH_SUSPENDING flag.
 315  */
 316 static int __must_push_back(struct multipath *m)
 317 {
 318         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 319                 dm_noflush_suspending(m->ti));
 320 }
 321
 322 static int map_io(struct multipath *m, struct request *clone,
 323                   struct dm_mpath_io *mpio, unsigned was_queued)
 324 {
 325         int r = DM_MAPIO_REMAPPED;
 326         size_t nr_bytes = blk_rq_bytes(clone);
 327         unsigned long flags;
 328         struct pgpath *pgpath;
 329         struct block_device *bdev;
 330
 331         spin_lock_irqsave(&m->lock, flags);
 332
 333         /* Do we need to select a new pgpath? */
 334         if (!m->current_pgpath ||
 335             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 336                 __choose_pgpath(m, nr_bytes);
 337
 338         pgpath = m->current_pgpath;
 339
 340         if (was_queued)
 341                 m->queue_size--;
 342
 343         if ((pgpath && m->queue_io) ||
 344             (!pgpath && m->queue_if_no_path)) {
 345                 /* Queue for the daemon to resubmit */
 346                 list_add_tail(&clone->queuelist, &m->queued_ios);
 347                 m->queue_size++;
 348                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 349                     !m->queue_io)
 350                         queue_work(kmultipathd, &m->process_queued_ios);
 351                 pgpath = NULL;
 352                 r = DM_MAPIO_SUBMITTED;
 353         } else if (pgpath) {
 354                 bdev = pgpath->path.dev->bdev;
 355                 clone->q = bdev_get_queue(bdev);
 356                 clone->rq_disk = bdev->bd_disk;
 357         } else if (__must_push_back(m))
 358                 r = DM_MAPIO_REQUEUE;
 359         else
 360                 r = -EIO;       /* Failed */
 361
 362         mpio->pgpath = pgpath;
 363         mpio->nr_bytes = nr_bytes;
 364
 365         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 366                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 367                                               nr_bytes);
 368
 369         spin_unlock_irqrestore(&m->lock, flags);
 370
 371         return r;
 372 }
 373
 374 /*
 375  * If we run out of usable paths, should we queue I/O or error it?
 376  */
 377 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 378                             unsigned save_old_value)
 379 {
 380         unsigned long flags;
 381
 382         spin_lock_irqsave(&m->lock, flags);
 383
 384         if (save_old_value)
 385                 m->saved_queue_if_no_path = m->queue_if_no_path;
 386         else
 387                 m->saved_queue_if_no_path = queue_if_no_path;
 388         m->queue_if_no_path = queue_if_no_path;
 389         if (!m->queue_if_no_path && m->queue_size)
 390                 queue_work(kmultipathd, &m->process_queued_ios);
 391
 392         spin_unlock_irqrestore(&m->lock, flags);
 393
 394         return 0;
 395 }
 396
 397 /*-----------------------------------------------------------------
 398  * The multipath daemon is responsible for resubmitting queued ios.
 399  *---------------------------------------------------------------*/
 400
 401 static void dispatch_queued_ios(struct multipath *m)
 402 {
 403         int r;
 404         unsigned long flags;
 405         struct dm_mpath_io *mpio;
 406         union map_info *info;
 407         struct request *clone, *n;
 408         LIST_HEAD(cl);
 409
 410         spin_lock_irqsave(&m->lock, flags);
 411         list_splice_init(&m->queued_ios, &cl);
 412         spin_unlock_irqrestore(&m->lock, flags);
 413
 414         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 415                 list_del_init(&clone->queuelist);
 416
 417                 info = dm_get_rq_mapinfo(clone);
 418                 mpio = info->ptr;
 419
 420                 r = map_io(m, clone, mpio, 1);
 421                 if (r < 0) {
 422                         mempool_free(mpio, m->mpio_pool);
 423                         dm_kill_request(clone, r);
 424                 } else if (r == DM_MAPIO_REMAPPED)
 425                         dm_dispatch_request(clone);
 426                 else if (r == DM_MAPIO_REQUEUE) {
 427                         mempool_free(mpio, m->mpio_pool);
 428                         dm_requeue_request(clone);
 429                 }
 430         }
 431 }
 432
 433 static void process_queued_ios(struct work_struct *work)
 434 {
 435         struct multipath *m =
 436                 container_of(work, struct multipath, process_queued_ios);
 437         struct pgpath *pgpath = NULL, *tmp;
 438         unsigned must_queue = 1;
 439         unsigned long flags;
 440
 441         spin_lock_irqsave(&m->lock, flags);
 442
 443         if (!m->queue_size)
 444                 goto out;
 445
 446         if (!m->current_pgpath)
 447                 __choose_pgpath(m, 1 << 19); /* Assume 512 KB */
 448
 449         pgpath = m->current_pgpath;
 450
 451         if (pgpath) {
 452                 struct block_device *bdev = pgpath->path.dev->bdev;
 453                 if (unlikely(blk_queue_stopped(bdev_get_queue(bdev))))
 454                         goto out;
 455         }
 456
 457         if ((pgpath && !m->queue_io) ||
 458             (!pgpath && !m->queue_if_no_path))
 459                 must_queue = 0;
 460
 461         if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
 462                 m->pg_init_count++;
 463                 m->pg_init_required = 0;
 464                 list_for_each_entry(tmp, &pgpath->pg->pgpaths, list) {
 465                         queue_work(kmpath_handlerd, &tmp->activate_path);
 466                         m->pg_init_in_progress++;
 467                 }
 468         }
 469 out:
 470         spin_unlock_irqrestore(&m->lock, flags);
 471         if (!must_queue)
 472                 dispatch_queued_ios(m);
 473 }
 474
 475 /*
 476  * An event is triggered whenever a path is taken out of use.
 477  * Includes path failure and PG bypass.
 478  */
 479 static void trigger_event(struct work_struct *work)
 480 {
 481         struct multipath *m =
 482                 container_of(work, struct multipath, trigger_event);
 483
 484         dm_table_event(m->ti->table);
 485 }
 486
 487 /*-----------------------------------------------------------------
 488  * Constructor/argument parsing:
 489  * <#multipath feature args> [<arg>]*
 490  * <#hw_handler args> [hw_handler [<arg>]*]
 491  * <#priority groups>
 492  * <initial priority group>
 493  *     [<selector> <#selector args> [<arg>]*
 494  *      <#paths> <#per-path selector args>
 495  *         [<path> [<arg>]* ]+ ]+
 496  *---------------------------------------------------------------*/
 497 struct param {
 498         unsigned min;
 499         unsigned max;
 500         char *error;
 501 };
 502
 503 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 504 {
 505         if (!str ||
 506             (sscanf(str, "%u", v) != 1) ||
 507             (*v < param->min) ||
 508             (*v > param->max)) {
 509                 *error = param->error;
 510                 return -EINVAL;
 511         }
 512
 513         return 0;
 514 }
 515
 516 struct arg_set {
 517         unsigned argc;
 518         char **argv;
 519 };
 520
 521 static char *shift(struct arg_set *as)
 522 {
 523         char *r;
 524
 525         if (as->argc) {
 526                 as->argc--;
 527                 r = *as->argv;
 528                 as->argv++;
 529                 return r;
 530         }
 531
 532         return NULL;
 533 }
 534
 535 static void consume(struct arg_set *as, unsigned n)
 536 {
 537         BUG_ON (as->argc < n);
 538         as->argc -= n;
 539         as->argv += n;
 540 }
 541
 542 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 543                                struct dm_target *ti)
 544 {
 545         int r;
 546         struct path_selector_type *pst;
 547         unsigned ps_argc;
 548
 549         static struct param _params[] = {
 550                 {0, 1024, "invalid number of path selector args"},
 551         };
 552
 553         pst = dm_get_path_selector(shift(as));
 554         if (!pst) {
 555                 ti->error = "unknown path selector type";
 556                 return -EINVAL;
 557         }
 558
 559         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 560         if (r) {
 561                 dm_put_path_selector(pst);
 562                 return -EINVAL;
 563         }
 564
 565         r = pst->create(&pg->ps, ps_argc, as->argv);
 566         if (r) {
 567                 dm_put_path_selector(pst);
 568                 ti->error = "path selector constructor failed";
 569                 return r;
 570         }
 571
 572         pg->ps.type = pst;
 573         consume(as, ps_argc);
 574
 575         return 0;
 576 }
 577
 578 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 579                                struct dm_target *ti)
 580 {
 581         int r;
 582         struct pgpath *p;
 583         char *path;
 584         struct multipath *m = ti->private;
 585
 586         /* we need at least a path arg */
 587         if (as->argc < 1) {
 588                 ti->error = "no device given";
 589                 return NULL;
 590         }
 591
 592         p = alloc_pgpath();
 593         if (!p)
 594                 return NULL;
 595
 596         path = shift(as);
 597         r = dm_get_device(ti, path, ti->begin, ti->len,
 598                           dm_table_get_mode(ti->table), &p->path.dev);
 599         if (r) {
 600                 unsigned major, minor;
 601
 602                 /* Try to add a failed device */
 603                 if (r == -ENXIO && sscanf(path, "%u:%u", &major, &minor) == 2) {
 604                         dev_t dev;
 605
 606                         /* Extract the major/minor numbers */
 607                         dev = MKDEV(major, minor);
 608                         if (MAJOR(dev) != major || MINOR(dev) != minor) {
 609                                 /* Nice try, didn't work */
 610                                 DMWARN("Invalid device path %s", path);
 611                                 ti->error = "error converting devnum";
 612                                 goto bad;
 613                         }
 614                         DMWARN("adding disabled device %d:%d", major, minor);
 615                         p->path.dev = NULL;
 616                         format_dev_t(p->path.pdev, dev);
 617                         p->is_active = 0;
 618                 } else {
 619                         ti->error = "error getting device";
 620                         goto bad;
 621                 }
 622         } else {
 623                 memcpy(p->path.pdev, p->path.dev->name, 16);
 624         }
 625
 626         if (p->path.dev) {
 627                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 628
 629                 if (m->hw_handler_name) {
 630                         r = scsi_dh_attach(q, m->hw_handler_name);
 631                         if (r == -EBUSY) {
 632                                 /*
 633                                  * Already attached to different hw_handler,
 634                                  * try to reattach with correct one.
 635                                  */
 636                                 scsi_dh_detach(q);
 637                                 r = scsi_dh_attach(q, m->hw_handler_name);
 638                         }
 639                         if (r < 0) {
 640                                 ti->error = "error attaching hardware handler";
 641                                 dm_put_device(ti, p->path.dev);
 642                                 goto bad;
 643                         }
 644                 } else {
 645                         /* Play safe and detach hardware handler */
 646                         scsi_dh_detach(q);
 647                 }
 648         }
 649
 650         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 651         if (r) {
 652                 dm_put_device(ti, p->path.dev);
 653                 goto bad;
 654         }
 655
 656         if (!p->is_active) {
 657                 ps->type->fail_path(ps, &p->path);
 658                 p->fail_count++;
 659                 m->nr_valid_paths--;
 660         }
 661         return p;
 662
 663  bad:
 664         free_pgpath(p);
 665         return NULL;
 666 }
 667
 668 static struct priority_group *parse_priority_group(struct arg_set *as,
 669                                                    struct multipath *m)
 670 {
 671         static struct param _params[] = {
 672                 {1, 1024, "invalid number of paths"},
 673                 {0, 1024, "invalid number of selector args"}
 674         };
 675
 676         int r;
 677         unsigned i, nr_selector_args, nr_params;
 678         struct priority_group *pg;
 679         struct dm_target *ti = m->ti;
 680
 681         if (as->argc < 2) {
 682                 as->argc = 0;
 683                 ti->error = "not enough priority group aruments";
 684                 return NULL;
 685         }
 686
 687         pg = alloc_priority_group();
 688         if (!pg) {
 689                 ti->error = "couldn't allocate priority group";
 690                 return NULL;
 691         }
 692         pg->m = m;
 693
 694         r = parse_path_selector(as, pg, ti);
 695         if (r)
 696                 goto bad;
 697
 698         /*
 699          * read the paths
 700          */
 701         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 702         if (r)
 703                 goto bad;
 704
 705         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 706         if (r)
 707                 goto bad;
 708
 709         nr_params = 1 + nr_selector_args;
 710         for (i = 0; i < pg->nr_pgpaths; i++) {
 711                 struct pgpath *pgpath;
 712                 struct arg_set path_args;
 713
 714                 if (as->argc < nr_params) {
 715                         ti->error = "not enough path parameters";
 716                         goto bad;
 717                 }
 718
 719                 path_args.argc = nr_params;
 720                 path_args.argv = as->argv;
 721
 722                 pgpath = parse_path(&path_args, &pg->ps, ti);
 723                 if (!pgpath)
 724                         goto bad;
 725
 726                 pgpath->pg = pg;
 727                 list_add_tail(&pgpath->list, &pg->pgpaths);
 728                 consume(as, nr_params);
 729         }
 730
 731         return pg;
 732
 733  bad:
 734         free_priority_group(pg, ti);
 735         return NULL;
 736 }
 737
 738 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 739 {
 740         unsigned hw_argc;
 741         struct dm_target *ti = m->ti;
 742
 743         static struct param _params[] = {
 744                 {0, 1024, "invalid number of hardware handler args"},
 745         };
 746
 747         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 748                 return -EINVAL;
 749
 750         if (!hw_argc)
 751                 return 0;
 752
 753         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 754         request_module("scsi_dh_%s", m->hw_handler_name);
 755         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 756                 ti->error = "unknown hardware handler type";
 757                 kfree(m->hw_handler_name);
 758                 m->hw_handler_name = NULL;
 759                 return -EINVAL;
 760         }
 761         consume(as, hw_argc - 1);
 762
 763         return 0;
 764 }
 765
 766 static int parse_features(struct arg_set *as, struct multipath *m)
 767 {
 768         int r;
 769         unsigned argc;
 770         struct dm_target *ti = m->ti;
 771         const char *param_name;
 772
 773         static struct param _params[] = {
 774                 {0, 3, "invalid number of feature args"},
 775                 {1, 50, "pg_init_retries must be between 1 and 50"},
 776         };
 777
 778         r = read_param(_params, shift(as), &argc, &ti->error);
 779         if (r)
 780                 return -EINVAL;
 781
 782         if (!argc)
 783                 return 0;
 784
 785         do {
 786                 param_name = shift(as);
 787                 argc--;
 788
 789                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 790                         r = queue_if_no_path(m, 1, 0);
 791                         continue;
 792                 }
 793
 794                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 795                     (argc >= 1)) {
 796                         r = read_param(_params + 1, shift(as),
 797                                        &m->pg_init_retries, &ti->error);
 798                         argc--;
 799                         continue;
 800                 }
 801
 802                 ti->error = "Unrecognised multipath feature request";
 803                 r = -EINVAL;
 804         } while (argc && !r);
 805
 806         return r;
 807 }
 808
 809 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 810                          char **argv)
 811 {
 812         /* target parameters */
 813         static struct param _params[] = {
 814                 {0, 1024, "invalid number of priority groups"},
 815                 {0, 1024, "invalid initial priority group number"},
 816         };
 817
 818         int r;
 819         struct multipath *m;
 820         struct arg_set as;
 821         unsigned pg_count = 0;
 822         unsigned next_pg_num;
 823
 824         as.argc = argc;
 825         as.argv = argv;
 826
 827         m = alloc_multipath(ti);
 828         if (!m) {
 829                 ti->error = "can't allocate multipath";
 830                 return -EINVAL;
 831         }
 832
 833         r = parse_features(&as, m);
 834         if (r)
 835                 goto bad;
 836
 837         r = parse_hw_handler(&as, m);
 838         if (r)
 839                 goto bad;
 840
 841         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 842         if (r)
 843                 goto bad;
 844
 845         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 846         if (r)
 847                 goto bad;
 848
 849         /* parse the priority groups */
 850         while (as.argc) {
 851                 struct priority_group *pg;
 852
 853                 pg = parse_priority_group(&as, m);
 854                 if (!pg) {
 855                         r = -EINVAL;
 856                         goto bad;
 857                 }
 858
 859                 m->nr_valid_paths += pg->nr_pgpaths;
 860                 list_add_tail(&pg->list, &m->priority_groups);
 861                 pg_count++;
 862                 pg->pg_num = pg_count;
 863                 if (!--next_pg_num)
 864                         m->next_pg = pg;
 865         }
 866
 867         if (pg_count != m->nr_priority_groups) {
 868                 ti->error = "priority group count mismatch";
 869                 r = -EINVAL;
 870                 goto bad;
 871         }
 872
 873         return 0;
 874
 875  bad:
 876         free_multipath(m);
 877         return r;
 878 }
 879
 880 static void multipath_dtr(struct dm_target *ti)
 881 {
 882         struct multipath *m = (struct multipath *) ti->private;
 883
 884         flush_workqueue(kmpath_handlerd);
 885         flush_workqueue(kmultipathd);
 886         free_multipath(m);
 887 }
 888
 889 /*
 890  * Map cloned requests
 891  */
 892 static int multipath_map(struct dm_target *ti, struct request *clone,
 893                          union map_info *map_context)
 894 {
 895         int r;
 896         struct dm_mpath_io *mpio;
 897         struct multipath *m = (struct multipath *) ti->private;
 898
 899         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 900         if (!mpio)
 901                 /* ENOMEM, requeue */
 902                 return DM_MAPIO_REQUEUE;
 903         memset(mpio, 0, sizeof(*mpio));
 904
 905         map_context->ptr = mpio;
 906         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 907         r = map_io(m, clone, mpio, 0);
 908         if (r < 0 || r == DM_MAPIO_REQUEUE)
 909                 mempool_free(mpio, m->mpio_pool);
 910
 911         return r;
 912 }
 913
 914 /*
 915  * Take a path out of use.
 916  */
 917 static int fail_path(struct pgpath *pgpath)
 918 {
 919         unsigned long flags;
 920         struct multipath *m = pgpath->pg->m;
 921
 922         spin_lock_irqsave(&m->lock, flags);
 923
 924         if (!pgpath->is_active)
 925                 goto out;
 926
 927         DMWARN("Failing path %s.", pgpath->path.pdev);
 928
 929         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 930         pgpath->is_active = 0;
 931         pgpath->fail_count++;
 932
 933         m->nr_valid_paths--;
 934
 935         if (pgpath == m->current_pgpath)
 936                 m->current_pgpath = NULL;
 937
 938         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 939                        pgpath->path.pdev, m->nr_valid_paths);
 940
 941         queue_work(kmultipathd, &m->trigger_event);
 942         queue_work(kmultipathd, &pgpath->deactivate_path);
 943
 944 out:
 945         spin_unlock_irqrestore(&m->lock, flags);
 946
 947         return 0;
 948 }
 949
 950 /*
 951  * Reinstate a previously-failed path
 952  */
 953 static int reinstate_path(struct pgpath *pgpath)
 954 {
 955         int r = 0;
 956         unsigned long flags;
 957         struct multipath *m = pgpath->pg->m;
 958
 959         spin_lock_irqsave(&m->lock, flags);
 960
 961         if (pgpath->is_active)
 962                 goto out;
 963
 964         if (!pgpath->path.dev) {
 965                 DMWARN("Cannot reinstate disabled path %s", pgpath->path.pdev);
 966                 r = -ENODEV;
 967                 goto out;
 968         }
 969
 970         if (!pgpath->pg->ps.type->reinstate_path) {
 971                 DMWARN("Reinstate path not supported by path selector %s",
 972                        pgpath->pg->ps.type->name);
 973                 r = -EINVAL;
 974                 goto out;
 975         }
 976
 977         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 978         if (r)
 979                 goto out;
 980
 981         pgpath->is_active = 1;
 982
 983         m->current_pgpath = NULL;
 984         if (!m->nr_valid_paths++ && m->queue_size)
 985                 queue_work(kmultipathd, &m->process_queued_ios);
 986
 987         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 988                        pgpath->path.pdev, m->nr_valid_paths);
 989
 990         queue_work(kmultipathd, &m->trigger_event);
 991
 992 out:
 993         spin_unlock_irqrestore(&m->lock, flags);
 994
 995         return r;
 996 }
 997
 998 /*
 999  * Fail or reinstate all paths that match the provided struct dm_dev.
1000  */
1001 static int action_dev(struct multipath *m, struct dm_dev *dev,
1002                       action_fn action)
1003 {
1004         int r = 0;
1005         struct pgpath *pgpath;
1006         struct priority_group *pg;
1007
1008         if (!dev)
1009                 return 0;
1010
1011         list_for_each_entry(pg, &m->priority_groups, list) {
1012                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1013                         if (pgpath->path.dev == dev)
1014                                 r = action(pgpath);
1015                 }
1016         }
1017
1018         return r;
1019 }
1020
1021 /*
1022  * Temporarily try to avoid having to use the specified PG
1023  */
1024 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1025                       int bypassed)
1026 {
1027         unsigned long flags;
1028
1029         spin_lock_irqsave(&m->lock, flags);
1030
1031         pg->bypassed = bypassed;
1032         m->current_pgpath = NULL;
1033         m->current_pg = NULL;
1034
1035         spin_unlock_irqrestore(&m->lock, flags);
1036
1037         queue_work(kmultipathd, &m->trigger_event);
1038 }
1039
1040 /*
1041  * Switch to using the specified PG from the next I/O that gets mapped
1042  */
1043 static int switch_pg_num(struct multipath *m, const char *pgstr)
1044 {
1045         struct priority_group *pg;
1046         unsigned pgnum;
1047         unsigned long flags;
1048
1049         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1050             (pgnum > m->nr_priority_groups)) {
1051                 DMWARN("invalid PG number supplied to switch_pg_num");
1052                 return -EINVAL;
1053         }
1054
1055         spin_lock_irqsave(&m->lock, flags);
1056         list_for_each_entry(pg, &m->priority_groups, list) {
1057                 pg->bypassed = 0;
1058                 if (--pgnum)
1059                         continue;
1060
1061                 m->current_pgpath = NULL;
1062                 m->current_pg = NULL;
1063                 m->next_pg = pg;
1064         }
1065         spin_unlock_irqrestore(&m->lock, flags);
1066
1067         queue_work(kmultipathd, &m->trigger_event);
1068         return 0;
1069 }
1070
1071 /*
1072  * Set/clear bypassed status of a PG.
1073  * PGs are numbered upwards from 1 in the order they were declared.
1074  */
1075 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1076 {
1077         struct priority_group *pg;
1078         unsigned pgnum;
1079
1080         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1081             (pgnum > m->nr_priority_groups)) {
1082                 DMWARN("invalid PG number supplied to bypass_pg");
1083                 return -EINVAL;
1084         }
1085
1086         list_for_each_entry(pg, &m->priority_groups, list) {
1087                 if (!--pgnum)
1088                         break;
1089         }
1090
1091         bypass_pg(m, pg, bypassed);
1092         return 0;
1093 }
1094
1095 /*
1096  * Should we retry pg_init immediately?
1097  */
1098 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1099 {
1100         unsigned long flags;
1101         int limit_reached = 0;
1102
1103         spin_lock_irqsave(&m->lock, flags);
1104
1105         if (m->pg_init_count <= m->pg_init_retries)
1106                 m->pg_init_required = 1;
1107         else
1108                 limit_reached = 1;
1109
1110         spin_unlock_irqrestore(&m->lock, flags);
1111
1112         return limit_reached;
1113 }
1114
1115 static void pg_init_done(struct dm_path *path, int errors)
1116 {
1117         struct pgpath *pgpath = path_to_pgpath(path);
1118         struct priority_group *pg = pgpath->pg;
1119         struct multipath *m = pg->m;
1120         unsigned long flags;
1121
1122         /* device or driver problems */
1123         switch (errors) {
1124         case SCSI_DH_OK:
1125                 break;
1126         case SCSI_DH_NOSYS:
1127                 if (!m->hw_handler_name) {
1128                         errors = 0;
1129                         break;
1130                 }
1131                 DMERR("Cannot failover device because scsi_dh_%s was not "
1132                       "loaded.", m->hw_handler_name);
1133                 /*
1134                  * Fail path for now, so we do not ping pong
1135                  */
1136                 fail_path(pgpath);
1137                 break;
1138         case SCSI_DH_DEV_TEMP_BUSY:
1139                 /*
1140                  * Probably doing something like FW upgrade on the
1141                  * controller so try the other pg.
1142                  */
1143                 bypass_pg(m, pg, 1);
1144                 break;
1145         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1146         case SCSI_DH_RETRY:
1147         case SCSI_DH_IMM_RETRY:
1148         case SCSI_DH_RES_TEMP_UNAVAIL:
1149                 if (pg_init_limit_reached(m, pgpath))
1150                         fail_path(pgpath);
1151                 errors = 0;
1152                 break;
1153         default:
1154                 /*
1155                  * We probably do not want to fail the path for a device
1156                  * error, but this is what the old dm did. In future
1157                  * patches we can do more advanced handling.
1158                  */
1159                 fail_path(pgpath);
1160         }
1161
1162         spin_lock_irqsave(&m->lock, flags);
1163         if (errors) {
1164                 DMERR("Could not failover device. Error %d.", errors);
1165                 m->current_pgpath = NULL;
1166                 m->current_pg = NULL;
1167         } else if (!m->pg_init_required) {
1168                 m->queue_io = 0;
1169                 pg->bypassed = 0;
1170         }
1171
1172        m->pg_init_in_progress--;
1173        if (!m->pg_init_in_progress)
1174                queue_work(kmultipathd, &m->process_queued_ios);
1175        spin_unlock_irqrestore(&m->lock, flags);
1176 }
1177
1178 static void activate_path(struct work_struct *work)
1179 {
1180         int ret = SCSI_DH_DEV_OFFLINED;
1181         struct pgpath *pgpath =
1182                 container_of(work, struct pgpath, activate_path);
1183
1184         if (pgpath->path.dev)
1185                 ret = scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev));
1186         pg_init_done(&pgpath->path, ret);
1187 }
1188
1189 /*
1190  * end_io handling
1191  */
1192 static int do_end_io(struct multipath *m, struct request *clone,
1193                      int error, struct dm_mpath_io *mpio)
1194 {
1195         /*
1196          * We don't queue any clone request inside the multipath target
1197          * during end I/O handling, since those clone requests don't have
1198          * bio clones.  If we queue them inside the multipath target,
1199          * we need to make bio clones, that requires memory allocation.
1200          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1201          *  don't have bio clones.)
1202          * Instead of queueing the clone request here, we queue the original
1203          * request into dm core, which will remake a clone request and
1204          * clone bios for it and resubmit it later.
1205          */
1206         int r = DM_ENDIO_REQUEUE;
1207         unsigned long flags;
1208
1209         if (!error && !clone->errors)
1210                 return 0;       /* I/O complete */
1211
1212         if (error == -EOPNOTSUPP)
1213                 return error;
1214
1215         if (mpio->pgpath)
1216                 fail_path(mpio->pgpath);
1217
1218         spin_lock_irqsave(&m->lock, flags);
1219         if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1220                 r = -EIO;
1221         spin_unlock_irqrestore(&m->lock, flags);
1222
1223         return r;
1224 }
1225
1226 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1227                             int error, union map_info *map_context)
1228 {
1229         struct multipath *m = ti->private;
1230         struct dm_mpath_io *mpio = map_context->ptr;
1231         struct pgpath *pgpath = mpio->pgpath;
1232         struct path_selector *ps;
1233         int r;
1234
1235         r  = do_end_io(m, clone, error, mpio);
1236         if (pgpath) {
1237                 ps = &pgpath->pg->ps;
1238                 if (ps->type->end_io)
1239                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1240         }
1241         mempool_free(mpio, m->mpio_pool);
1242
1243         return r;
1244 }
1245
1246 /*
1247  * Suspend can't complete until all the I/O is processed so if
1248  * the last path fails we must error any remaining I/O.
1249  * Note that if the freeze_bdev fails while suspending, the
1250  * queue_if_no_path state is lost - userspace should reset it.
1251  */
1252 static void multipath_presuspend(struct dm_target *ti)
1253 {
1254         struct multipath *m = (struct multipath *) ti->private;
1255
1256         queue_if_no_path(m, 0, 1);
1257 }
1258
1259 /*
1260  * Restore the queue_if_no_path setting.
1261  */
1262 static void multipath_resume(struct dm_target *ti)
1263 {
1264         struct multipath *m = (struct multipath *) ti->private;
1265         unsigned long flags;
1266
1267         spin_lock_irqsave(&m->lock, flags);
1268         m->queue_if_no_path = m->saved_queue_if_no_path;
1269         spin_unlock_irqrestore(&m->lock, flags);
1270 }
1271
1272 /*
1273  * Info output has the following format:
1274  * num_multipath_feature_args [multipath_feature_args]*
1275  * num_handler_status_args [handler_status_args]*
1276  * num_groups init_group_number
1277  *            [A|D|E num_ps_status_args [ps_status_args]*
1278  *             num_paths num_selector_args
1279  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1280  *
1281  * Table output has the following format (identical to the constructor string):
1282  * num_feature_args [features_args]*
1283  * num_handler_args hw_handler [hw_handler_args]*
1284  * num_groups init_group_number
1285  *     [priority selector-name num_ps_args [ps_args]*
1286  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1287  */
1288 static int multipath_status(struct dm_target *ti, status_type_t type,
1289                             char *result, unsigned int maxlen)
1290 {
1291         int sz = 0;
1292         unsigned long flags;
1293         struct multipath *m = (struct multipath *) ti->private;
1294         struct priority_group *pg;
1295         struct pgpath *p;
1296         unsigned pg_num;
1297         char state;
1298
1299         spin_lock_irqsave(&m->lock, flags);
1300
1301         /* Features */
1302         if (type == STATUSTYPE_INFO)
1303                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1304         else {
1305                 DMEMIT("%u ", m->queue_if_no_path +
1306                               (m->pg_init_retries > 0) * 2);
1307                 if (m->queue_if_no_path)
1308                         DMEMIT("queue_if_no_path ");
1309                 if (m->pg_init_retries)
1310                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1311         }
1312
1313         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1314                 DMEMIT("0 ");
1315         else
1316                 DMEMIT("1 %s ", m->hw_handler_name);
1317
1318         DMEMIT("%u ", m->nr_priority_groups);
1319
1320         if (m->next_pg)
1321                 pg_num = m->next_pg->pg_num;
1322         else if (m->current_pg)
1323                 pg_num = m->current_pg->pg_num;
1324         else
1325                         pg_num = 1;
1326
1327         DMEMIT("%u ", pg_num);
1328
1329         switch (type) {
1330         case STATUSTYPE_INFO:
1331                 list_for_each_entry(pg, &m->priority_groups, list) {
1332                         if (pg->bypassed)
1333                                 state = 'D';    /* Disabled */
1334                         else if (pg == m->current_pg)
1335                                 state = 'A';    /* Currently Active */
1336                         else
1337                                 state = 'E';    /* Enabled */
1338
1339                         DMEMIT("%c ", state);
1340
1341                         if (pg->ps.type->status)
1342                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1343                                                           result + sz,
1344                                                           maxlen - sz);
1345                         else
1346                                 DMEMIT("0 ");
1347
1348                         DMEMIT("%u %u ", pg->nr_pgpaths,
1349                                pg->ps.type->info_args);
1350
1351                         list_for_each_entry(p, &pg->pgpaths, list) {
1352                                 DMEMIT("%s %s %u ", p->path.pdev,
1353                                        p->is_active ? "A" : "F",
1354                                        p->fail_count);
1355                                 if (pg->ps.type->status)
1356                                         sz += pg->ps.type->status(&pg->ps,
1357                                               &p->path, type, result + sz,
1358                                               maxlen - sz);
1359                         }
1360                 }
1361                 break;
1362
1363         case STATUSTYPE_TABLE:
1364                 list_for_each_entry(pg, &m->priority_groups, list) {
1365                         DMEMIT("%s ", pg->ps.type->name);
1366
1367                         if (pg->ps.type->status)
1368                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1369                                                           result + sz,
1370                                                           maxlen - sz);
1371                         else
1372                                 DMEMIT("0 ");
1373
1374                         DMEMIT("%u %u ", pg->nr_pgpaths,
1375                                pg->ps.type->table_args);
1376
1377                         list_for_each_entry(p, &pg->pgpaths, list) {
1378                                 DMEMIT("%s ", p->path.pdev);
1379                                 if (pg->ps.type->status)
1380                                         sz += pg->ps.type->status(&pg->ps,
1381                                               &p->path, type, result + sz,
1382                                               maxlen - sz);
1383                         }
1384                 }
1385                 break;
1386         }
1387
1388         spin_unlock_irqrestore(&m->lock, flags);
1389
1390         return 0;
1391 }
1392
1393 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1394 {
1395         int r;
1396         struct dm_dev *dev;
1397         struct multipath *m = (struct multipath *) ti->private;
1398         action_fn action;
1399
1400         if (argc == 1) {
1401                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1402                         return queue_if_no_path(m, 1, 0);
1403                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1404                         return queue_if_no_path(m, 0, 0);
1405         }
1406
1407         if (argc != 2)
1408                 goto error;
1409
1410         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1411                 return bypass_pg_num(m, argv[1], 1);
1412         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1413                 return bypass_pg_num(m, argv[1], 0);
1414         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1415                 return switch_pg_num(m, argv[1]);
1416         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1417                 action = reinstate_path;
1418         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1419                 action = fail_path;
1420         else
1421                 goto error;
1422
1423         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1424                           dm_table_get_mode(ti->table), &dev);
1425         if (r) {
1426                 DMWARN("message: error getting device %s",
1427                        argv[1]);
1428                 return -EINVAL;
1429         }
1430
1431         r = action_dev(m, dev, action);
1432
1433         dm_put_device(ti, dev);
1434
1435         return r;
1436
1437 error:
1438         DMWARN("Unrecognised multipath message received.");
1439         return -EINVAL;
1440 }
1441
1442 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1443                            struct file *filp, unsigned int cmd,
1444                            unsigned long arg)
1445 {
1446         struct multipath *m = (struct multipath *) ti->private;
1447         struct block_device *bdev = NULL;
1448         unsigned long flags;
1449         struct file fake_file = {};
1450         struct dentry fake_dentry = {};
1451         int r = 0;
1452
1453         fake_file.f_path.dentry = &fake_dentry;
1454
1455         spin_lock_irqsave(&m->lock, flags);
1456
1457         if (!m->current_pgpath)
1458                 __choose_pgpath(m, 1 << 19); /* Assume 512KB */
1459
1460         if (m->current_pgpath && m->current_pgpath->path.dev) {
1461                 bdev = m->current_pgpath->path.dev->bdev;
1462                 fake_dentry.d_inode = bdev->bd_inode;
1463                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1464         }
1465
1466         if (m->queue_io)
1467                 r = -EAGAIN;
1468         else if (!bdev)
1469                 r = -EIO;
1470
1471         spin_unlock_irqrestore(&m->lock, flags);
1472
1473         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1474                                          bdev->bd_disk, cmd, arg);
1475 }
1476
1477 static int __pgpath_busy(struct pgpath *pgpath)
1478 {
1479         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1480
1481         return dm_underlying_device_busy(q);
1482 }
1483
1484 /*
1485  * We return "busy", only when we can map I/Os but underlying devices
1486  * are busy (so even if we map I/Os now, the I/Os will wait on
1487  * the underlying queue).
1488  * In other words, if we want to kill I/Os or queue them inside us
1489  * due to map unavailability, we don't return "busy".  Otherwise,
1490  * dm core won't give us the I/Os and we can't do what we want.
1491  */
1492 static int multipath_busy(struct dm_target *ti)
1493 {
1494         int busy = 0, has_active = 0;
1495         struct multipath *m = (struct multipath *) ti->private;
1496         struct priority_group *pg;
1497         struct pgpath *pgpath;
1498         unsigned long flags;
1499
1500         spin_lock_irqsave(&m->lock, flags);
1501
1502         /* Guess which priority_group will be used at next mapping time */
1503         if (unlikely(!m->current_pgpath && m->next_pg))
1504                 pg = m->next_pg;
1505         else if (likely(m->current_pg))
1506                 pg = m->current_pg;
1507         else
1508                 /*
1509                  * We don't know which pg will be used at next mapping time.
1510                  * We don't call __choose_pgpath() here to avoid to trigger
1511                  * pg_init just by busy checking.
1512                  * So we don't know whether underlying devices we will be using
1513                  * at next mapping time are busy or not. Just try mapping.
1514                  */
1515                 goto out;
1516
1517         /*
1518          * If there is one non-busy active path at least, the path selector
1519          * will be able to select it. So we consider such a pg as not busy.
1520          */
1521         busy = 1;
1522         list_for_each_entry(pgpath, &pg->pgpaths, list)
1523                 if (pgpath->is_active) {
1524                         has_active = 1;
1525
1526                         if (!__pgpath_busy(pgpath)) {
1527                                 busy = 0;
1528                                 break;
1529                         }
1530                 }
1531
1532         if (!has_active)
1533                 /*
1534                  * No active path in this pg, so this pg won't be used and
1535                  * the current_pg will be changed at next mapping time.
1536                  * We need to try mapping to determine it.
1537                  */
1538                 busy = 0;
1539
1540 out:
1541         spin_unlock_irqrestore(&m->lock, flags);
1542
1543         return busy;
1544 }
1545
1546 /*-----------------------------------------------------------------
1547  * Module setup
1548  *---------------------------------------------------------------*/
1549 static struct target_type multipath_target = {
1550         .name = "multipath",
1551         .version = {1, 0, 5},
1552         .module = THIS_MODULE,
1553         .ctr = multipath_ctr,
1554         .dtr = multipath_dtr,
1555         .map_rq = multipath_map,
1556         .rq_end_io = multipath_end_io,
1557         .presuspend = multipath_presuspend,
1558         .resume = multipath_resume,
1559         .status = multipath_status,
1560         .message = multipath_message,
1561         .ioctl  = multipath_ioctl,
1562         .busy = multipath_busy,
1563 };
1564
1565 static int __init dm_multipath_init(void)
1566 {
1567         int r;
1568
1569         /* allocate a slab for the dm_ios */
1570         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1571         if (!_mpio_cache)
1572                 return -ENOMEM;
1573
1574         r = dm_register_target(&multipath_target);
1575         if (r < 0) {
1576                 DMERR("register failed %d", r);
1577                 kmem_cache_destroy(_mpio_cache);
1578                 return -EINVAL;
1579         }
1580
1581         kmultipathd = create_workqueue("kmpathd");
1582         if (!kmultipathd) {
1583                 DMERR("failed to create workqueue kmpathd");
1584                 dm_unregister_target(&multipath_target);
1585                 kmem_cache_destroy(_mpio_cache);
1586                 return -ENOMEM;
1587         }
1588
1589         /*
1590          * A separate workqueue is used to handle the device handlers
1591          * to avoid overloading existing workqueue. Overloading the
1592          * old workqueue would also create a bottleneck in the
1593          * path of the storage hardware device activation.
1594          */
1595         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1596         if (!kmpath_handlerd) {
1597                 DMERR("failed to create workqueue kmpath_handlerd");
1598                 destroy_workqueue(kmultipathd);
1599                 dm_unregister_target(&multipath_target);
1600                 kmem_cache_destroy(_mpio_cache);
1601                 return -ENOMEM;
1602         }
1603
1604         DMINFO("version %u.%u.%u loaded",
1605                multipath_target.version[0], multipath_target.version[1],
1606                multipath_target.version[2]);
1607
1608         return r;
1609 }
1610
1611 static void __exit dm_multipath_exit(void)
1612 {
1613         int r;
1614
1615         destroy_workqueue(kmpath_handlerd);
1616         destroy_workqueue(kmultipathd);
1617
1618         r = dm_unregister_target(&multipath_target);
1619         if (r < 0)
1620                 DMERR("target unregister failed %d", r);
1621         kmem_cache_destroy(_mpio_cache);
1622 }
1623
1624 module_init(dm_multipath_init);
1625 module_exit(dm_multipath_exit);
1626
1627 MODULE_DESCRIPTION(DM_NAME " multipath target");
1628 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1629 MODULE_LICENSE("GPL");