drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9
  10 #include "dm-path-selector.h"
  11 #include "dm-uevent.h"
  12
  13 #include <linux/ctype.h>
  14 #include <linux/init.h>
  15 #include <linux/mempool.h>
  16 #include <linux/module.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/slab.h>
  19 #include <linux/time.h>
  20 #include <linux/workqueue.h>
  21 #include <scsi/scsi_dh.h>
  22 #include <asm/atomic.h>
  23
  24 #define DM_MSG_PREFIX "multipath"
  25 #define MESG_STR(x) x, sizeof(x)
  26
  27 /* Path properties */
  28 struct pgpath {
  29         struct list_head list;
  30
  31         struct priority_group *pg;      /* Owning PG */
  32         unsigned is_active;             /* Path status */
  33         unsigned fail_count;            /* Cumulative failure count */
  34
  35         struct dm_path path;
  36         struct work_struct deactivate_path;
  37         struct work_struct activate_path;
  38 };
  39
  40 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  41
  42 /*
  43  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  44  * Each has a path selector which controls which path gets used.
  45  */
  46 struct priority_group {
  47         struct list_head list;
  48
  49         struct multipath *m;            /* Owning multipath instance */
  50         struct path_selector ps;
  51
  52         unsigned pg_num;                /* Reference number */
  53         unsigned bypassed;              /* Temporarily bypass this PG? */
  54
  55         unsigned nr_pgpaths;            /* Number of paths in PG */
  56         struct list_head pgpaths;
  57 };
  58
  59 /* Multipath context */
  60 struct multipath {
  61         struct list_head list;
  62         struct dm_target *ti;
  63
  64         spinlock_t lock;
  65
  66         const char *hw_handler_name;
  67         unsigned nr_priority_groups;
  68         struct list_head priority_groups;
  69         unsigned pg_init_required;      /* pg_init needs calling? */
  70         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  71
  72         unsigned nr_valid_paths;        /* Total number of usable paths */
  73         struct pgpath *current_pgpath;
  74         struct priority_group *current_pg;
  75         struct priority_group *next_pg; /* Switch to this PG if set */
  76         unsigned repeat_count;          /* I/Os left before calling PS again */
  77
  78         unsigned queue_io;              /* Must we queue all I/O? */
  79         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  80         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  81         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  82         unsigned pg_init_count;         /* Number of times pg_init called */
  83
  84         struct work_struct process_queued_ios;
  85         struct list_head queued_ios;
  86         unsigned queue_size;
  87
  88         struct work_struct trigger_event;
  89
  90         /*
  91          * We must use a mempool of dm_mpath_io structs so that we
  92          * can resubmit bios on error.
  93          */
  94         mempool_t *mpio_pool;
  95 };
  96
  97 /*
  98  * Context information attached to each bio we process.
  99  */
 100 struct dm_mpath_io {
 101         struct pgpath *pgpath;
 102         size_t nr_bytes;
 103 };
 104
 105 typedef int (*action_fn) (struct pgpath *pgpath);
 106
 107 #define MIN_IOS 256     /* Mempool size */
 108
 109 static struct kmem_cache *_mpio_cache;
 110
 111 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 112 static void process_queued_ios(struct work_struct *work);
 113 static void trigger_event(struct work_struct *work);
 114 static void activate_path(struct work_struct *work);
 115 static void deactivate_path(struct work_struct *work);
 116
 117
 118 /*-----------------------------------------------
 119  * Allocation routines
 120  *-----------------------------------------------*/
 121
 122 static struct pgpath *alloc_pgpath(void)
 123 {
 124         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 125
 126         if (pgpath) {
 127                 pgpath->is_active = 1;
 128                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
 129                 INIT_WORK(&pgpath->activate_path, activate_path);
 130         }
 131
 132         return pgpath;
 133 }
 134
 135 static void free_pgpath(struct pgpath *pgpath)
 136 {
 137         kfree(pgpath);
 138 }
 139
 140 static void deactivate_path(struct work_struct *work)
 141 {
 142         struct pgpath *pgpath =
 143                 container_of(work, struct pgpath, deactivate_path);
 144
 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         struct multipath *m = ti->private;
 164
 165         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 166                 list_del(&pgpath->list);
 167                 dm_put_device(ti, pgpath->path.dev);
 168                 free_pgpath(pgpath);
 169         }
 170 }
 171
 172 static void free_priority_group(struct priority_group *pg,
 173                                 struct dm_target *ti)
 174 {
 175         struct path_selector *ps = &pg->ps;
 176
 177         if (ps->type) {
 178                 ps->type->destroy(ps);
 179                 dm_put_path_selector(ps->type);
 180         }
 181
 182         free_pgpaths(&pg->pgpaths, ti);
 183         kfree(pg);
 184 }
 185
 186 static struct multipath *alloc_multipath(struct dm_target *ti)
 187 {
 188         struct multipath *m;
 189
 190         m = kzalloc(sizeof(*m), GFP_KERNEL);
 191         if (m) {
 192                 INIT_LIST_HEAD(&m->priority_groups);
 193                 INIT_LIST_HEAD(&m->queued_ios);
 194                 spin_lock_init(&m->lock);
 195                 m->queue_io = 1;
 196                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 197                 INIT_WORK(&m->trigger_event, trigger_event);
 198                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 199                 if (!m->mpio_pool) {
 200                         kfree(m);
 201                         return NULL;
 202                 }
 203                 m->ti = ti;
 204                 ti->private = m;
 205         }
 206
 207         return m;
 208 }
 209
 210 static void free_multipath(struct multipath *m)
 211 {
 212         struct priority_group *pg, *tmp;
 213
 214         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 215                 list_del(&pg->list);
 216                 free_priority_group(pg, m->ti);
 217         }
 218
 219         kfree(m->hw_handler_name);
 220         mempool_destroy(m->mpio_pool);
 221         kfree(m);
 222 }
 223
 224
 225 /*-----------------------------------------------
 226  * Path selection
 227  *-----------------------------------------------*/
 228
 229 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 230 {
 231         m->current_pg = pgpath->pg;
 232
 233         /* Must we initialise the PG first, and queue I/O till it's ready? */
 234         if (m->hw_handler_name) {
 235                 m->pg_init_required = 1;
 236                 m->queue_io = 1;
 237         } else {
 238                 m->pg_init_required = 0;
 239                 m->queue_io = 0;
 240         }
 241
 242         m->pg_init_count = 0;
 243 }
 244
 245 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 246                                size_t nr_bytes)
 247 {
 248         struct dm_path *path;
 249
 250         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 251         if (!path)
 252                 return -ENXIO;
 253
 254         m->current_pgpath = path_to_pgpath(path);
 255
 256         if (m->current_pg != pg)
 257                 __switch_pg(m, m->current_pgpath);
 258
 259         return 0;
 260 }
 261
 262 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 263 {
 264         struct priority_group *pg;
 265         unsigned bypassed = 1;
 266
 267         if (!m->nr_valid_paths)
 268                 goto failed;
 269
 270         /* Were we instructed to switch PG? */
 271         if (m->next_pg) {
 272                 pg = m->next_pg;
 273                 m->next_pg = NULL;
 274                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 275                         return;
 276         }
 277
 278         /* Don't change PG until it has no remaining paths */
 279         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 280                 return;
 281
 282         /*
 283          * Loop through priority groups until we find a valid path.
 284          * First time we skip PGs marked 'bypassed'.
 285          * Second time we only try the ones we skipped.
 286          */
 287         do {
 288                 list_for_each_entry(pg, &m->priority_groups, list) {
 289                         if (pg->bypassed == bypassed)
 290                                 continue;
 291                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 292                                 return;
 293                 }
 294         } while (bypassed--);
 295
 296 failed:
 297         m->current_pgpath = NULL;
 298         m->current_pg = NULL;
 299 }
 300
 301 /*
 302  * Check whether bios must be queued in the device-mapper core rather
 303  * than here in the target.
 304  *
 305  * m->lock must be held on entry.
 306  *
 307  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 308  * same value then we are not between multipath_presuspend()
 309  * and multipath_resume() calls and we have no need to check
 310  * for the DMF_NOFLUSH_SUSPENDING flag.
 311  */
 312 static int __must_push_back(struct multipath *m)
 313 {
 314         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 315                 dm_noflush_suspending(m->ti));
 316 }
 317
 318 static int map_io(struct multipath *m, struct request *clone,
 319                   struct dm_mpath_io *mpio, unsigned was_queued)
 320 {
 321         int r = DM_MAPIO_REMAPPED;
 322         size_t nr_bytes = blk_rq_bytes(clone);
 323         unsigned long flags;
 324         struct pgpath *pgpath;
 325         struct block_device *bdev;
 326
 327         spin_lock_irqsave(&m->lock, flags);
 328
 329         /* Do we need to select a new pgpath? */
 330         if (!m->current_pgpath ||
 331             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 332                 __choose_pgpath(m, nr_bytes);
 333
 334         pgpath = m->current_pgpath;
 335
 336         if (was_queued)
 337                 m->queue_size--;
 338
 339         if ((pgpath && m->queue_io) ||
 340             (!pgpath && m->queue_if_no_path)) {
 341                 /* Queue for the daemon to resubmit */
 342                 list_add_tail(&clone->queuelist, &m->queued_ios);
 343                 m->queue_size++;
 344                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 345                     !m->queue_io)
 346                         queue_work(kmultipathd, &m->process_queued_ios);
 347                 pgpath = NULL;
 348                 r = DM_MAPIO_SUBMITTED;
 349         } else if (pgpath) {
 350                 bdev = pgpath->path.dev->bdev;
 351                 clone->q = bdev_get_queue(bdev);
 352                 clone->rq_disk = bdev->bd_disk;
 353         } else if (__must_push_back(m))
 354                 r = DM_MAPIO_REQUEUE;
 355         else
 356                 r = -EIO;       /* Failed */
 357
 358         mpio->pgpath = pgpath;
 359         mpio->nr_bytes = nr_bytes;
 360
 361         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 362                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 363                                               nr_bytes);
 364
 365         spin_unlock_irqrestore(&m->lock, flags);
 366
 367         return r;
 368 }
 369
 370 /*
 371  * If we run out of usable paths, should we queue I/O or error it?
 372  */
 373 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 374                             unsigned save_old_value)
 375 {
 376         unsigned long flags;
 377
 378         spin_lock_irqsave(&m->lock, flags);
 379
 380         if (save_old_value)
 381                 m->saved_queue_if_no_path = m->queue_if_no_path;
 382         else
 383                 m->saved_queue_if_no_path = queue_if_no_path;
 384         m->queue_if_no_path = queue_if_no_path;
 385         if (!m->queue_if_no_path && m->queue_size)
 386                 queue_work(kmultipathd, &m->process_queued_ios);
 387
 388         spin_unlock_irqrestore(&m->lock, flags);
 389
 390         return 0;
 391 }
 392
 393 /*-----------------------------------------------------------------
 394  * The multipath daemon is responsible for resubmitting queued ios.
 395  *---------------------------------------------------------------*/
 396
 397 static void dispatch_queued_ios(struct multipath *m)
 398 {
 399         int r;
 400         unsigned long flags;
 401         struct dm_mpath_io *mpio;
 402         union map_info *info;
 403         struct request *clone, *n;
 404         LIST_HEAD(cl);
 405
 406         spin_lock_irqsave(&m->lock, flags);
 407         list_splice_init(&m->queued_ios, &cl);
 408         spin_unlock_irqrestore(&m->lock, flags);
 409
 410         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 411                 list_del_init(&clone->queuelist);
 412
 413                 info = dm_get_rq_mapinfo(clone);
 414                 mpio = info->ptr;
 415
 416                 r = map_io(m, clone, mpio, 1);
 417                 if (r < 0) {
 418                         mempool_free(mpio, m->mpio_pool);
 419                         dm_kill_unmapped_request(clone, r);
 420                 } else if (r == DM_MAPIO_REMAPPED)
 421                         dm_dispatch_request(clone);
 422                 else if (r == DM_MAPIO_REQUEUE) {
 423                         mempool_free(mpio, m->mpio_pool);
 424                         dm_requeue_unmapped_request(clone);
 425                 }
 426         }
 427 }
 428
 429 static void process_queued_ios(struct work_struct *work)
 430 {
 431         struct multipath *m =
 432                 container_of(work, struct multipath, process_queued_ios);
 433         struct pgpath *pgpath = NULL, *tmp;
 434         unsigned must_queue = 1;
 435         unsigned long flags;
 436
 437         spin_lock_irqsave(&m->lock, flags);
 438
 439         if (!m->queue_size)
 440                 goto out;
 441
 442         if (!m->current_pgpath)
 443                 __choose_pgpath(m, 0);
 444
 445         pgpath = m->current_pgpath;
 446
 447         if ((pgpath && !m->queue_io) ||
 448             (!pgpath && !m->queue_if_no_path))
 449                 must_queue = 0;
 450
 451         if (m->pg_init_required && !m->pg_init_in_progress && pgpath) {
 452                 m->pg_init_count++;
 453                 m->pg_init_required = 0;
 454                 list_for_each_entry(tmp, &pgpath->pg->pgpaths, list) {
 455                         if (queue_work(kmpath_handlerd, &tmp->activate_path))
 456                                 m->pg_init_in_progress++;
 457                 }
 458         }
 459 out:
 460         spin_unlock_irqrestore(&m->lock, flags);
 461         if (!must_queue)
 462                 dispatch_queued_ios(m);
 463 }
 464
 465 /*
 466  * An event is triggered whenever a path is taken out of use.
 467  * Includes path failure and PG bypass.
 468  */
 469 static void trigger_event(struct work_struct *work)
 470 {
 471         struct multipath *m =
 472                 container_of(work, struct multipath, trigger_event);
 473
 474         dm_table_event(m->ti->table);
 475 }
 476
 477 /*-----------------------------------------------------------------
 478  * Constructor/argument parsing:
 479  * <#multipath feature args> [<arg>]*
 480  * <#hw_handler args> [hw_handler [<arg>]*]
 481  * <#priority groups>
 482  * <initial priority group>
 483  *     [<selector> <#selector args> [<arg>]*
 484  *      <#paths> <#per-path selector args>
 485  *         [<path> [<arg>]* ]+ ]+
 486  *---------------------------------------------------------------*/
 487 struct param {
 488         unsigned min;
 489         unsigned max;
 490         char *error;
 491 };
 492
 493 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 494 {
 495         if (!str ||
 496             (sscanf(str, "%u", v) != 1) ||
 497             (*v < param->min) ||
 498             (*v > param->max)) {
 499                 *error = param->error;
 500                 return -EINVAL;
 501         }
 502
 503         return 0;
 504 }
 505
 506 struct arg_set {
 507         unsigned argc;
 508         char **argv;
 509 };
 510
 511 static char *shift(struct arg_set *as)
 512 {
 513         char *r;
 514
 515         if (as->argc) {
 516                 as->argc--;
 517                 r = *as->argv;
 518                 as->argv++;
 519                 return r;
 520         }
 521
 522         return NULL;
 523 }
 524
 525 static void consume(struct arg_set *as, unsigned n)
 526 {
 527         BUG_ON (as->argc < n);
 528         as->argc -= n;
 529         as->argv += n;
 530 }
 531
 532 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 533                                struct dm_target *ti)
 534 {
 535         int r;
 536         struct path_selector_type *pst;
 537         unsigned ps_argc;
 538
 539         static struct param _params[] = {
 540                 {0, 1024, "invalid number of path selector args"},
 541         };
 542
 543         pst = dm_get_path_selector(shift(as));
 544         if (!pst) {
 545                 ti->error = "unknown path selector type";
 546                 return -EINVAL;
 547         }
 548
 549         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 550         if (r) {
 551                 dm_put_path_selector(pst);
 552                 return -EINVAL;
 553         }
 554
 555         if (ps_argc > as->argc) {
 556                 dm_put_path_selector(pst);
 557                 ti->error = "not enough arguments for path selector";
 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         struct multipath *m = ti->private;
 580
 581         /* we need at least a path arg */
 582         if (as->argc < 1) {
 583                 ti->error = "no device given";
 584                 return ERR_PTR(-EINVAL);
 585         }
 586
 587         p = alloc_pgpath();
 588         if (!p)
 589                 return ERR_PTR(-ENOMEM);
 590
 591         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 592                           dm_table_get_mode(ti->table), &p->path.dev);
 593         if (r) {
 594                 ti->error = "error getting device";
 595                 goto bad;
 596         }
 597
 598         if (m->hw_handler_name) {
 599                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 600
 601                 r = scsi_dh_attach(q, m->hw_handler_name);
 602                 if (r == -EBUSY) {
 603                         /*
 604                          * Already attached to different hw_handler,
 605                          * try to reattach with correct one.
 606                          */
 607                         scsi_dh_detach(q);
 608                         r = scsi_dh_attach(q, m->hw_handler_name);
 609                 }
 610
 611                 if (r < 0) {
 612                         ti->error = "error attaching hardware handler";
 613                         dm_put_device(ti, p->path.dev);
 614                         goto bad;
 615                 }
 616         }
 617
 618         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 619         if (r) {
 620                 dm_put_device(ti, p->path.dev);
 621                 goto bad;
 622         }
 623
 624         return p;
 625
 626  bad:
 627         free_pgpath(p);
 628         return ERR_PTR(r);
 629 }
 630
 631 static struct priority_group *parse_priority_group(struct arg_set *as,
 632                                                    struct multipath *m)
 633 {
 634         static struct param _params[] = {
 635                 {1, 1024, "invalid number of paths"},
 636                 {0, 1024, "invalid number of selector args"}
 637         };
 638
 639         int r;
 640         unsigned i, nr_selector_args, nr_params;
 641         struct priority_group *pg;
 642         struct dm_target *ti = m->ti;
 643
 644         if (as->argc < 2) {
 645                 as->argc = 0;
 646                 ti->error = "not enough priority group arguments";
 647                 return ERR_PTR(-EINVAL);
 648         }
 649
 650         pg = alloc_priority_group();
 651         if (!pg) {
 652                 ti->error = "couldn't allocate priority group";
 653                 return ERR_PTR(-ENOMEM);
 654         }
 655         pg->m = m;
 656
 657         r = parse_path_selector(as, pg, ti);
 658         if (r)
 659                 goto bad;
 660
 661         /*
 662          * read the paths
 663          */
 664         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 665         if (r)
 666                 goto bad;
 667
 668         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 669         if (r)
 670                 goto bad;
 671
 672         nr_params = 1 + nr_selector_args;
 673         for (i = 0; i < pg->nr_pgpaths; i++) {
 674                 struct pgpath *pgpath;
 675                 struct arg_set path_args;
 676
 677                 if (as->argc < nr_params) {
 678                         ti->error = "not enough path parameters";
 679                         goto bad;
 680                 }
 681
 682                 path_args.argc = nr_params;
 683                 path_args.argv = as->argv;
 684
 685                 pgpath = parse_path(&path_args, &pg->ps, ti);
 686                 if (IS_ERR(pgpath)) {
 687                         r = PTR_ERR(pgpath);
 688                         goto bad;
 689                 }
 690
 691                 pgpath->pg = pg;
 692                 list_add_tail(&pgpath->list, &pg->pgpaths);
 693                 consume(as, nr_params);
 694         }
 695
 696         return pg;
 697
 698  bad:
 699         free_priority_group(pg, ti);
 700         return ERR_PTR(r);
 701 }
 702
 703 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 704 {
 705         unsigned hw_argc;
 706         struct dm_target *ti = m->ti;
 707
 708         static struct param _params[] = {
 709                 {0, 1024, "invalid number of hardware handler args"},
 710         };
 711
 712         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 713                 return -EINVAL;
 714
 715         if (!hw_argc)
 716                 return 0;
 717
 718         if (hw_argc > as->argc) {
 719                 ti->error = "not enough arguments for hardware handler";
 720                 return -EINVAL;
 721         }
 722
 723         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 724         request_module("scsi_dh_%s", m->hw_handler_name);
 725         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 726                 ti->error = "unknown hardware handler type";
 727                 kfree(m->hw_handler_name);
 728                 m->hw_handler_name = NULL;
 729                 return -EINVAL;
 730         }
 731
 732         if (hw_argc > 1)
 733                 DMWARN("Ignoring user-specified arguments for "
 734                        "hardware handler \"%s\"", m->hw_handler_name);
 735         consume(as, hw_argc - 1);
 736
 737         return 0;
 738 }
 739
 740 static int parse_features(struct arg_set *as, struct multipath *m)
 741 {
 742         int r;
 743         unsigned argc;
 744         struct dm_target *ti = m->ti;
 745         const char *param_name;
 746
 747         static struct param _params[] = {
 748                 {0, 3, "invalid number of feature args"},
 749                 {1, 50, "pg_init_retries must be between 1 and 50"},
 750         };
 751
 752         r = read_param(_params, shift(as), &argc, &ti->error);
 753         if (r)
 754                 return -EINVAL;
 755
 756         if (!argc)
 757                 return 0;
 758
 759         do {
 760                 param_name = shift(as);
 761                 argc--;
 762
 763                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 764                         r = queue_if_no_path(m, 1, 0);
 765                         continue;
 766                 }
 767
 768                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 769                     (argc >= 1)) {
 770                         r = read_param(_params + 1, shift(as),
 771                                        &m->pg_init_retries, &ti->error);
 772                         argc--;
 773                         continue;
 774                 }
 775
 776                 ti->error = "Unrecognised multipath feature request";
 777                 r = -EINVAL;
 778         } while (argc && !r);
 779
 780         return r;
 781 }
 782
 783 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 784                          char **argv)
 785 {
 786         /* target parameters */
 787         static struct param _params[] = {
 788                 {0, 1024, "invalid number of priority groups"},
 789                 {0, 1024, "invalid initial priority group number"},
 790         };
 791
 792         int r;
 793         struct multipath *m;
 794         struct arg_set as;
 795         unsigned pg_count = 0;
 796         unsigned next_pg_num;
 797
 798         as.argc = argc;
 799         as.argv = argv;
 800
 801         m = alloc_multipath(ti);
 802         if (!m) {
 803                 ti->error = "can't allocate multipath";
 804                 return -EINVAL;
 805         }
 806
 807         r = parse_features(&as, m);
 808         if (r)
 809                 goto bad;
 810
 811         r = parse_hw_handler(&as, m);
 812         if (r)
 813                 goto bad;
 814
 815         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 816         if (r)
 817                 goto bad;
 818
 819         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 820         if (r)
 821                 goto bad;
 822
 823         /* parse the priority groups */
 824         while (as.argc) {
 825                 struct priority_group *pg;
 826
 827                 pg = parse_priority_group(&as, m);
 828                 if (IS_ERR(pg)) {
 829                         r = PTR_ERR(pg);
 830                         goto bad;
 831                 }
 832
 833                 m->nr_valid_paths += pg->nr_pgpaths;
 834                 list_add_tail(&pg->list, &m->priority_groups);
 835                 pg_count++;
 836                 pg->pg_num = pg_count;
 837                 if (!--next_pg_num)
 838                         m->next_pg = pg;
 839         }
 840
 841         if (pg_count != m->nr_priority_groups) {
 842                 ti->error = "priority group count mismatch";
 843                 r = -EINVAL;
 844                 goto bad;
 845         }
 846
 847         ti->num_flush_requests = 1;
 848
 849         return 0;
 850
 851  bad:
 852         free_multipath(m);
 853         return r;
 854 }
 855
 856 static void multipath_dtr(struct dm_target *ti)
 857 {
 858         struct multipath *m = (struct multipath *) ti->private;
 859
 860         flush_workqueue(kmpath_handlerd);
 861         flush_workqueue(kmultipathd);
 862         flush_scheduled_work();
 863         free_multipath(m);
 864 }
 865
 866 /*
 867  * Map cloned requests
 868  */
 869 static int multipath_map(struct dm_target *ti, struct request *clone,
 870                          union map_info *map_context)
 871 {
 872         int r;
 873         struct dm_mpath_io *mpio;
 874         struct multipath *m = (struct multipath *) ti->private;
 875
 876         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 877         if (!mpio)
 878                 /* ENOMEM, requeue */
 879                 return DM_MAPIO_REQUEUE;
 880         memset(mpio, 0, sizeof(*mpio));
 881
 882         map_context->ptr = mpio;
 883         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 884         r = map_io(m, clone, mpio, 0);
 885         if (r < 0 || r == DM_MAPIO_REQUEUE)
 886                 mempool_free(mpio, m->mpio_pool);
 887
 888         return r;
 889 }
 890
 891 /*
 892  * Take a path out of use.
 893  */
 894 static int fail_path(struct pgpath *pgpath)
 895 {
 896         unsigned long flags;
 897         struct multipath *m = pgpath->pg->m;
 898
 899         spin_lock_irqsave(&m->lock, flags);
 900
 901         if (!pgpath->is_active)
 902                 goto out;
 903
 904         DMWARN("Failing path %s.", pgpath->path.dev->name);
 905
 906         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 907         pgpath->is_active = 0;
 908         pgpath->fail_count++;
 909
 910         m->nr_valid_paths--;
 911
 912         if (pgpath == m->current_pgpath)
 913                 m->current_pgpath = NULL;
 914
 915         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 916                       pgpath->path.dev->name, m->nr_valid_paths);
 917
 918         schedule_work(&m->trigger_event);
 919         queue_work(kmultipathd, &pgpath->deactivate_path);
 920
 921 out:
 922         spin_unlock_irqrestore(&m->lock, flags);
 923
 924         return 0;
 925 }
 926
 927 /*
 928  * Reinstate a previously-failed path
 929  */
 930 static int reinstate_path(struct pgpath *pgpath)
 931 {
 932         int r = 0;
 933         unsigned long flags;
 934         struct multipath *m = pgpath->pg->m;
 935
 936         spin_lock_irqsave(&m->lock, flags);
 937
 938         if (pgpath->is_active)
 939                 goto out;
 940
 941         if (!pgpath->pg->ps.type->reinstate_path) {
 942                 DMWARN("Reinstate path not supported by path selector %s",
 943                        pgpath->pg->ps.type->name);
 944                 r = -EINVAL;
 945                 goto out;
 946         }
 947
 948         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 949         if (r)
 950                 goto out;
 951
 952         pgpath->is_active = 1;
 953
 954         if (!m->nr_valid_paths++ && m->queue_size) {
 955                 m->current_pgpath = NULL;
 956                 queue_work(kmultipathd, &m->process_queued_ios);
 957         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
 958                 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
 959                         m->pg_init_in_progress++;
 960         }
 961
 962         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 963                       pgpath->path.dev->name, m->nr_valid_paths);
 964
 965         schedule_work(&m->trigger_event);
 966
 967 out:
 968         spin_unlock_irqrestore(&m->lock, flags);
 969
 970         return r;
 971 }
 972
 973 /*
 974  * Fail or reinstate all paths that match the provided struct dm_dev.
 975  */
 976 static int action_dev(struct multipath *m, struct dm_dev *dev,
 977                       action_fn action)
 978 {
 979         int r = 0;
 980         struct pgpath *pgpath;
 981         struct priority_group *pg;
 982
 983         list_for_each_entry(pg, &m->priority_groups, list) {
 984                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 985                         if (pgpath->path.dev == dev)
 986                                 r = action(pgpath);
 987                 }
 988         }
 989
 990         return r;
 991 }
 992
 993 /*
 994  * Temporarily try to avoid having to use the specified PG
 995  */
 996 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 997                       int bypassed)
 998 {
 999         unsigned long flags;
1000
1001         spin_lock_irqsave(&m->lock, flags);
1002
1003         pg->bypassed = bypassed;
1004         m->current_pgpath = NULL;
1005         m->current_pg = NULL;
1006
1007         spin_unlock_irqrestore(&m->lock, flags);
1008
1009         schedule_work(&m->trigger_event);
1010 }
1011
1012 /*
1013  * Switch to using the specified PG from the next I/O that gets mapped
1014  */
1015 static int switch_pg_num(struct multipath *m, const char *pgstr)
1016 {
1017         struct priority_group *pg;
1018         unsigned pgnum;
1019         unsigned long flags;
1020
1021         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1022             (pgnum > m->nr_priority_groups)) {
1023                 DMWARN("invalid PG number supplied to switch_pg_num");
1024                 return -EINVAL;
1025         }
1026
1027         spin_lock_irqsave(&m->lock, flags);
1028         list_for_each_entry(pg, &m->priority_groups, list) {
1029                 pg->bypassed = 0;
1030                 if (--pgnum)
1031                         continue;
1032
1033                 m->current_pgpath = NULL;
1034                 m->current_pg = NULL;
1035                 m->next_pg = pg;
1036         }
1037         spin_unlock_irqrestore(&m->lock, flags);
1038
1039         schedule_work(&m->trigger_event);
1040         return 0;
1041 }
1042
1043 /*
1044  * Set/clear bypassed status of a PG.
1045  * PGs are numbered upwards from 1 in the order they were declared.
1046  */
1047 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1048 {
1049         struct priority_group *pg;
1050         unsigned pgnum;
1051
1052         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1053             (pgnum > m->nr_priority_groups)) {
1054                 DMWARN("invalid PG number supplied to bypass_pg");
1055                 return -EINVAL;
1056         }
1057
1058         list_for_each_entry(pg, &m->priority_groups, list) {
1059                 if (!--pgnum)
1060                         break;
1061         }
1062
1063         bypass_pg(m, pg, bypassed);
1064         return 0;
1065 }
1066
1067 /*
1068  * Should we retry pg_init immediately?
1069  */
1070 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1071 {
1072         unsigned long flags;
1073         int limit_reached = 0;
1074
1075         spin_lock_irqsave(&m->lock, flags);
1076
1077         if (m->pg_init_count <= m->pg_init_retries)
1078                 m->pg_init_required = 1;
1079         else
1080                 limit_reached = 1;
1081
1082         spin_unlock_irqrestore(&m->lock, flags);
1083
1084         return limit_reached;
1085 }
1086
1087 static void pg_init_done(struct dm_path *path, int errors)
1088 {
1089         struct pgpath *pgpath = path_to_pgpath(path);
1090         struct priority_group *pg = pgpath->pg;
1091         struct multipath *m = pg->m;
1092         unsigned long flags;
1093
1094         /* device or driver problems */
1095         switch (errors) {
1096         case SCSI_DH_OK:
1097                 break;
1098         case SCSI_DH_NOSYS:
1099                 if (!m->hw_handler_name) {
1100                         errors = 0;
1101                         break;
1102                 }
1103                 DMERR("Cannot failover device because scsi_dh_%s was not "
1104                       "loaded.", m->hw_handler_name);
1105                 /*
1106                  * Fail path for now, so we do not ping pong
1107                  */
1108                 fail_path(pgpath);
1109                 break;
1110         case SCSI_DH_DEV_TEMP_BUSY:
1111                 /*
1112                  * Probably doing something like FW upgrade on the
1113                  * controller so try the other pg.
1114                  */
1115                 bypass_pg(m, pg, 1);
1116                 break;
1117         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1118         case SCSI_DH_RETRY:
1119         case SCSI_DH_IMM_RETRY:
1120         case SCSI_DH_RES_TEMP_UNAVAIL:
1121                 if (pg_init_limit_reached(m, pgpath))
1122                         fail_path(pgpath);
1123                 errors = 0;
1124                 break;
1125         default:
1126                 /*
1127                  * We probably do not want to fail the path for a device
1128                  * error, but this is what the old dm did. In future
1129                  * patches we can do more advanced handling.
1130                  */
1131                 fail_path(pgpath);
1132         }
1133
1134         spin_lock_irqsave(&m->lock, flags);
1135         if (errors) {
1136                 if (pgpath == m->current_pgpath) {
1137                         DMERR("Could not failover device. Error %d.", errors);
1138                         m->current_pgpath = NULL;
1139                         m->current_pg = NULL;
1140                 }
1141         } else if (!m->pg_init_required) {
1142                 m->queue_io = 0;
1143                 pg->bypassed = 0;
1144         }
1145
1146         m->pg_init_in_progress--;
1147         if (!m->pg_init_in_progress)
1148                 queue_work(kmultipathd, &m->process_queued_ios);
1149         spin_unlock_irqrestore(&m->lock, flags);
1150 }
1151
1152 static void activate_path(struct work_struct *work)
1153 {
1154         int ret;
1155         struct pgpath *pgpath =
1156                 container_of(work, struct pgpath, activate_path);
1157
1158         ret = scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev));
1159         pg_init_done(&pgpath->path, ret);
1160 }
1161
1162 /*
1163  * end_io handling
1164  */
1165 static int do_end_io(struct multipath *m, struct request *clone,
1166                      int error, struct dm_mpath_io *mpio)
1167 {
1168         /*
1169          * We don't queue any clone request inside the multipath target
1170          * during end I/O handling, since those clone requests don't have
1171          * bio clones.  If we queue them inside the multipath target,
1172          * we need to make bio clones, that requires memory allocation.
1173          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1174          *  don't have bio clones.)
1175          * Instead of queueing the clone request here, we queue the original
1176          * request into dm core, which will remake a clone request and
1177          * clone bios for it and resubmit it later.
1178          */
1179         int r = DM_ENDIO_REQUEUE;
1180         unsigned long flags;
1181
1182         if (!error && !clone->errors)
1183                 return 0;       /* I/O complete */
1184
1185         if (error == -EOPNOTSUPP)
1186                 return error;
1187
1188         if (mpio->pgpath)
1189                 fail_path(mpio->pgpath);
1190
1191         spin_lock_irqsave(&m->lock, flags);
1192         if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1193                 r = -EIO;
1194         spin_unlock_irqrestore(&m->lock, flags);
1195
1196         return r;
1197 }
1198
1199 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1200                             int error, union map_info *map_context)
1201 {
1202         struct multipath *m = ti->private;
1203         struct dm_mpath_io *mpio = map_context->ptr;
1204         struct pgpath *pgpath = mpio->pgpath;
1205         struct path_selector *ps;
1206         int r;
1207
1208         r  = do_end_io(m, clone, error, mpio);
1209         if (pgpath) {
1210                 ps = &pgpath->pg->ps;
1211                 if (ps->type->end_io)
1212                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1213         }
1214         mempool_free(mpio, m->mpio_pool);
1215
1216         return r;
1217 }
1218
1219 /*
1220  * Suspend can't complete until all the I/O is processed so if
1221  * the last path fails we must error any remaining I/O.
1222  * Note that if the freeze_bdev fails while suspending, the
1223  * queue_if_no_path state is lost - userspace should reset it.
1224  */
1225 static void multipath_presuspend(struct dm_target *ti)
1226 {
1227         struct multipath *m = (struct multipath *) ti->private;
1228
1229         queue_if_no_path(m, 0, 1);
1230 }
1231
1232 /*
1233  * Restore the queue_if_no_path setting.
1234  */
1235 static void multipath_resume(struct dm_target *ti)
1236 {
1237         struct multipath *m = (struct multipath *) ti->private;
1238         unsigned long flags;
1239
1240         spin_lock_irqsave(&m->lock, flags);
1241         m->queue_if_no_path = m->saved_queue_if_no_path;
1242         spin_unlock_irqrestore(&m->lock, flags);
1243 }
1244
1245 /*
1246  * Info output has the following format:
1247  * num_multipath_feature_args [multipath_feature_args]*
1248  * num_handler_status_args [handler_status_args]*
1249  * num_groups init_group_number
1250  *            [A|D|E num_ps_status_args [ps_status_args]*
1251  *             num_paths num_selector_args
1252  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1253  *
1254  * Table output has the following format (identical to the constructor string):
1255  * num_feature_args [features_args]*
1256  * num_handler_args hw_handler [hw_handler_args]*
1257  * num_groups init_group_number
1258  *     [priority selector-name num_ps_args [ps_args]*
1259  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1260  */
1261 static int multipath_status(struct dm_target *ti, status_type_t type,
1262                             char *result, unsigned int maxlen)
1263 {
1264         int sz = 0;
1265         unsigned long flags;
1266         struct multipath *m = (struct multipath *) ti->private;
1267         struct priority_group *pg;
1268         struct pgpath *p;
1269         unsigned pg_num;
1270         char state;
1271
1272         spin_lock_irqsave(&m->lock, flags);
1273
1274         /* Features */
1275         if (type == STATUSTYPE_INFO)
1276                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1277         else {
1278                 DMEMIT("%u ", m->queue_if_no_path +
1279                               (m->pg_init_retries > 0) * 2);
1280                 if (m->queue_if_no_path)
1281                         DMEMIT("queue_if_no_path ");
1282                 if (m->pg_init_retries)
1283                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1284         }
1285
1286         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1287                 DMEMIT("0 ");
1288         else
1289                 DMEMIT("1 %s ", m->hw_handler_name);
1290
1291         DMEMIT("%u ", m->nr_priority_groups);
1292
1293         if (m->next_pg)
1294                 pg_num = m->next_pg->pg_num;
1295         else if (m->current_pg)
1296                 pg_num = m->current_pg->pg_num;
1297         else
1298                         pg_num = 1;
1299
1300         DMEMIT("%u ", pg_num);
1301
1302         switch (type) {
1303         case STATUSTYPE_INFO:
1304                 list_for_each_entry(pg, &m->priority_groups, list) {
1305                         if (pg->bypassed)
1306                                 state = 'D';    /* Disabled */
1307                         else if (pg == m->current_pg)
1308                                 state = 'A';    /* Currently Active */
1309                         else
1310                                 state = 'E';    /* Enabled */
1311
1312                         DMEMIT("%c ", state);
1313
1314                         if (pg->ps.type->status)
1315                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1316                                                           result + sz,
1317                                                           maxlen - sz);
1318                         else
1319                                 DMEMIT("0 ");
1320
1321                         DMEMIT("%u %u ", pg->nr_pgpaths,
1322                                pg->ps.type->info_args);
1323
1324                         list_for_each_entry(p, &pg->pgpaths, list) {
1325                                 DMEMIT("%s %s %u ", p->path.dev->name,
1326                                        p->is_active ? "A" : "F",
1327                                        p->fail_count);
1328                                 if (pg->ps.type->status)
1329                                         sz += pg->ps.type->status(&pg->ps,
1330                                               &p->path, type, result + sz,
1331                                               maxlen - sz);
1332                         }
1333                 }
1334                 break;
1335
1336         case STATUSTYPE_TABLE:
1337                 list_for_each_entry(pg, &m->priority_groups, list) {
1338                         DMEMIT("%s ", pg->ps.type->name);
1339
1340                         if (pg->ps.type->status)
1341                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1342                                                           result + sz,
1343                                                           maxlen - sz);
1344                         else
1345                                 DMEMIT("0 ");
1346
1347                         DMEMIT("%u %u ", pg->nr_pgpaths,
1348                                pg->ps.type->table_args);
1349
1350                         list_for_each_entry(p, &pg->pgpaths, list) {
1351                                 DMEMIT("%s ", p->path.dev->name);
1352                                 if (pg->ps.type->status)
1353                                         sz += pg->ps.type->status(&pg->ps,
1354                                               &p->path, type, result + sz,
1355                                               maxlen - sz);
1356                         }
1357                 }
1358                 break;
1359         }
1360
1361         spin_unlock_irqrestore(&m->lock, flags);
1362
1363         return 0;
1364 }
1365
1366 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1367 {
1368         int r;
1369         struct dm_dev *dev;
1370         struct multipath *m = (struct multipath *) ti->private;
1371         action_fn action;
1372
1373         if (argc == 1) {
1374                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1375                         return queue_if_no_path(m, 1, 0);
1376                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1377                         return queue_if_no_path(m, 0, 0);
1378         }
1379
1380         if (argc != 2)
1381                 goto error;
1382
1383         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1384                 return bypass_pg_num(m, argv[1], 1);
1385         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1386                 return bypass_pg_num(m, argv[1], 0);
1387         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1388                 return switch_pg_num(m, argv[1]);
1389         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1390                 action = reinstate_path;
1391         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1392                 action = fail_path;
1393         else
1394                 goto error;
1395
1396         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1397                           dm_table_get_mode(ti->table), &dev);
1398         if (r) {
1399                 DMWARN("message: error getting device %s",
1400                        argv[1]);
1401                 return -EINVAL;
1402         }
1403
1404         r = action_dev(m, dev, action);
1405
1406         dm_put_device(ti, dev);
1407
1408         return r;
1409
1410 error:
1411         DMWARN("Unrecognised multipath message received.");
1412         return -EINVAL;
1413 }
1414
1415 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1416                            unsigned long arg)
1417 {
1418         struct multipath *m = (struct multipath *) ti->private;
1419         struct block_device *bdev = NULL;
1420         fmode_t mode = 0;
1421         unsigned long flags;
1422         int r = 0;
1423
1424         spin_lock_irqsave(&m->lock, flags);
1425
1426         if (!m->current_pgpath)
1427                 __choose_pgpath(m, 0);
1428
1429         if (m->current_pgpath) {
1430                 bdev = m->current_pgpath->path.dev->bdev;
1431                 mode = m->current_pgpath->path.dev->mode;
1432         }
1433
1434         if (m->queue_io)
1435                 r = -EAGAIN;
1436         else if (!bdev)
1437                 r = -EIO;
1438
1439         spin_unlock_irqrestore(&m->lock, flags);
1440
1441         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1442 }
1443
1444 static int multipath_iterate_devices(struct dm_target *ti,
1445                                      iterate_devices_callout_fn fn, void *data)
1446 {
1447         struct multipath *m = ti->private;
1448         struct priority_group *pg;
1449         struct pgpath *p;
1450         int ret = 0;
1451
1452         list_for_each_entry(pg, &m->priority_groups, list) {
1453                 list_for_each_entry(p, &pg->pgpaths, list) {
1454                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1455                         if (ret)
1456                                 goto out;
1457                 }
1458         }
1459
1460 out:
1461         return ret;
1462 }
1463
1464 static int __pgpath_busy(struct pgpath *pgpath)
1465 {
1466         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1467
1468         return dm_underlying_device_busy(q);
1469 }
1470
1471 /*
1472  * We return "busy", only when we can map I/Os but underlying devices
1473  * are busy (so even if we map I/Os now, the I/Os will wait on
1474  * the underlying queue).
1475  * In other words, if we want to kill I/Os or queue them inside us
1476  * due to map unavailability, we don't return "busy".  Otherwise,
1477  * dm core won't give us the I/Os and we can't do what we want.
1478  */
1479 static int multipath_busy(struct dm_target *ti)
1480 {
1481         int busy = 0, has_active = 0;
1482         struct multipath *m = ti->private;
1483         struct priority_group *pg;
1484         struct pgpath *pgpath;
1485         unsigned long flags;
1486
1487         spin_lock_irqsave(&m->lock, flags);
1488
1489         /* Guess which priority_group will be used at next mapping time */
1490         if (unlikely(!m->current_pgpath && m->next_pg))
1491                 pg = m->next_pg;
1492         else if (likely(m->current_pg))
1493                 pg = m->current_pg;
1494         else
1495                 /*
1496                  * We don't know which pg will be used at next mapping time.
1497                  * We don't call __choose_pgpath() here to avoid to trigger
1498                  * pg_init just by busy checking.
1499                  * So we don't know whether underlying devices we will be using
1500                  * at next mapping time are busy or not. Just try mapping.
1501                  */
1502                 goto out;
1503
1504         /*
1505          * If there is one non-busy active path at least, the path selector
1506          * will be able to select it. So we consider such a pg as not busy.
1507          */
1508         busy = 1;
1509         list_for_each_entry(pgpath, &pg->pgpaths, list)
1510                 if (pgpath->is_active) {
1511                         has_active = 1;
1512
1513                         if (!__pgpath_busy(pgpath)) {
1514                                 busy = 0;
1515                                 break;
1516                         }
1517                 }
1518
1519         if (!has_active)
1520                 /*
1521                  * No active path in this pg, so this pg won't be used and
1522                  * the current_pg will be changed at next mapping time.
1523                  * We need to try mapping to determine it.
1524                  */
1525                 busy = 0;
1526
1527 out:
1528         spin_unlock_irqrestore(&m->lock, flags);
1529
1530         return busy;
1531 }
1532
1533 /*-----------------------------------------------------------------
1534  * Module setup
1535  *---------------------------------------------------------------*/
1536 static struct target_type multipath_target = {
1537         .name = "multipath",
1538         .version = {1, 1, 0},
1539         .module = THIS_MODULE,
1540         .ctr = multipath_ctr,
1541         .dtr = multipath_dtr,
1542         .map_rq = multipath_map,
1543         .rq_end_io = multipath_end_io,
1544         .presuspend = multipath_presuspend,
1545         .resume = multipath_resume,
1546         .status = multipath_status,
1547         .message = multipath_message,
1548         .ioctl  = multipath_ioctl,
1549         .iterate_devices = multipath_iterate_devices,
1550         .busy = multipath_busy,
1551 };
1552
1553 static int __init dm_multipath_init(void)
1554 {
1555         int r;
1556
1557         /* allocate a slab for the dm_ios */
1558         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1559         if (!_mpio_cache)
1560                 return -ENOMEM;
1561
1562         r = dm_register_target(&multipath_target);
1563         if (r < 0) {
1564                 DMERR("register failed %d", r);
1565                 kmem_cache_destroy(_mpio_cache);
1566                 return -EINVAL;
1567         }
1568
1569         kmultipathd = create_workqueue("kmpathd");
1570         if (!kmultipathd) {
1571                 DMERR("failed to create workqueue kmpathd");
1572                 dm_unregister_target(&multipath_target);
1573                 kmem_cache_destroy(_mpio_cache);
1574                 return -ENOMEM;
1575         }
1576
1577         /*
1578          * A separate workqueue is used to handle the device handlers
1579          * to avoid overloading existing workqueue. Overloading the
1580          * old workqueue would also create a bottleneck in the
1581          * path of the storage hardware device activation.
1582          */
1583         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1584         if (!kmpath_handlerd) {
1585                 DMERR("failed to create workqueue kmpath_handlerd");
1586                 destroy_workqueue(kmultipathd);
1587                 dm_unregister_target(&multipath_target);
1588                 kmem_cache_destroy(_mpio_cache);
1589                 return -ENOMEM;
1590         }
1591
1592         DMINFO("version %u.%u.%u loaded",
1593                multipath_target.version[0], multipath_target.version[1],
1594                multipath_target.version[2]);
1595
1596         return r;
1597 }
1598
1599 static void __exit dm_multipath_exit(void)
1600 {
1601         destroy_workqueue(kmpath_handlerd);
1602         destroy_workqueue(kmultipathd);
1603
1604         dm_unregister_target(&multipath_target);
1605         kmem_cache_destroy(_mpio_cache);
1606 }
1607
1608 module_init(dm_multipath_init);
1609 module_exit(dm_multipath_exit);
1610
1611 MODULE_DESCRIPTION(DM_NAME " multipath target");
1612 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1613 MODULE_LICENSE("GPL");