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 <linux/atomic.h>
  23
  24 #define DM_MSG_PREFIX "multipath"
  25 #define DM_PG_INIT_DELAY_MSECS 2000
  26 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  27
  28 /* Path properties */
  29 struct pgpath {
  30         struct list_head list;
  31
  32         struct priority_group *pg;      /* Owning PG */
  33         unsigned is_active;             /* Path status */
  34         unsigned fail_count;            /* Cumulative failure count */
  35
  36         struct dm_path path;
  37         struct delayed_work 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         const char *hw_handler_name;
  65         char *hw_handler_params;
  66
  67         spinlock_t lock;
  68
  69         unsigned nr_priority_groups;
  70         struct list_head priority_groups;
  71
  72         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  73
  74         unsigned pg_init_required;      /* pg_init needs calling? */
  75         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  76         unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
  77
  78         unsigned nr_valid_paths;        /* Total number of usable paths */
  79         struct pgpath *current_pgpath;
  80         struct priority_group *current_pg;
  81         struct priority_group *next_pg; /* Switch to this PG if set */
  82         unsigned repeat_count;          /* I/Os left before calling PS again */
  83
  84         unsigned queue_io:1;            /* Must we queue all I/O? */
  85         unsigned queue_if_no_path:1;    /* Queue I/O if last path fails? */
  86         unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
  87
  88         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  89         unsigned pg_init_count;         /* Number of times pg_init called */
  90         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  91
  92         unsigned queue_size;
  93         struct work_struct process_queued_ios;
  94         struct list_head queued_ios;
  95
  96         struct work_struct trigger_event;
  97
  98         /*
  99          * We must use a mempool of dm_mpath_io structs so that we
 100          * can resubmit bios on error.
 101          */
 102         mempool_t *mpio_pool;
 103
 104         struct mutex work_mutex;
 105 };
 106
 107 /*
 108  * Context information attached to each bio we process.
 109  */
 110 struct dm_mpath_io {
 111         struct pgpath *pgpath;
 112         size_t nr_bytes;
 113 };
 114
 115 typedef int (*action_fn) (struct pgpath *pgpath);
 116
 117 #define MIN_IOS 256     /* Mempool size */
 118
 119 static struct kmem_cache *_mpio_cache;
 120
 121 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 122 static void process_queued_ios(struct work_struct *work);
 123 static void trigger_event(struct work_struct *work);
 124 static void activate_path(struct work_struct *work);
 125
 126
 127 /*-----------------------------------------------
 128  * Allocation routines
 129  *-----------------------------------------------*/
 130
 131 static struct pgpath *alloc_pgpath(void)
 132 {
 133         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 134
 135         if (pgpath) {
 136                 pgpath->is_active = 1;
 137                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 138         }
 139
 140         return pgpath;
 141 }
 142
 143 static void free_pgpath(struct pgpath *pgpath)
 144 {
 145         kfree(pgpath);
 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                 if (m->hw_handler_name)
 168                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 169                 dm_put_device(ti, pgpath->path.dev);
 170                 free_pgpath(pgpath);
 171         }
 172 }
 173
 174 static void free_priority_group(struct priority_group *pg,
 175                                 struct dm_target *ti)
 176 {
 177         struct path_selector *ps = &pg->ps;
 178
 179         if (ps->type) {
 180                 ps->type->destroy(ps);
 181                 dm_put_path_selector(ps->type);
 182         }
 183
 184         free_pgpaths(&pg->pgpaths, ti);
 185         kfree(pg);
 186 }
 187
 188 static struct multipath *alloc_multipath(struct dm_target *ti)
 189 {
 190         struct multipath *m;
 191
 192         m = kzalloc(sizeof(*m), GFP_KERNEL);
 193         if (m) {
 194                 INIT_LIST_HEAD(&m->priority_groups);
 195                 INIT_LIST_HEAD(&m->queued_ios);
 196                 spin_lock_init(&m->lock);
 197                 m->queue_io = 1;
 198                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 199                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 200                 INIT_WORK(&m->trigger_event, trigger_event);
 201                 init_waitqueue_head(&m->pg_init_wait);
 202                 mutex_init(&m->work_mutex);
 203                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 204                 if (!m->mpio_pool) {
 205                         kfree(m);
 206                         return NULL;
 207                 }
 208                 m->ti = ti;
 209                 ti->private = m;
 210         }
 211
 212         return m;
 213 }
 214
 215 static void free_multipath(struct multipath *m)
 216 {
 217         struct priority_group *pg, *tmp;
 218
 219         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 220                 list_del(&pg->list);
 221                 free_priority_group(pg, m->ti);
 222         }
 223
 224         kfree(m->hw_handler_name);
 225         kfree(m->hw_handler_params);
 226         mempool_destroy(m->mpio_pool);
 227         kfree(m);
 228 }
 229
 230 static int set_mapinfo(struct multipath *m, union map_info *info)
 231 {
 232         struct dm_mpath_io *mpio;
 233
 234         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 235         if (!mpio)
 236                 return -ENOMEM;
 237
 238         memset(mpio, 0, sizeof(*mpio));
 239         info->ptr = mpio;
 240
 241         return 0;
 242 }
 243
 244 static void clear_mapinfo(struct multipath *m, union map_info *info)
 245 {
 246         struct dm_mpath_io *mpio = info->ptr;
 247
 248         info->ptr = NULL;
 249         mempool_free(mpio, m->mpio_pool);
 250 }
 251
 252 /*-----------------------------------------------
 253  * Path selection
 254  *-----------------------------------------------*/
 255
 256 static void __pg_init_all_paths(struct multipath *m)
 257 {
 258         struct pgpath *pgpath;
 259         unsigned long pg_init_delay = 0;
 260
 261         m->pg_init_count++;
 262         m->pg_init_required = 0;
 263         if (m->pg_init_delay_retry)
 264                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 265                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 266         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 267                 /* Skip failed paths */
 268                 if (!pgpath->is_active)
 269                         continue;
 270                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 271                                        pg_init_delay))
 272                         m->pg_init_in_progress++;
 273         }
 274 }
 275
 276 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 277 {
 278         m->current_pg = pgpath->pg;
 279
 280         /* Must we initialise the PG first, and queue I/O till it's ready? */
 281         if (m->hw_handler_name) {
 282                 m->pg_init_required = 1;
 283                 m->queue_io = 1;
 284         } else {
 285                 m->pg_init_required = 0;
 286                 m->queue_io = 0;
 287         }
 288
 289         m->pg_init_count = 0;
 290 }
 291
 292 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 293                                size_t nr_bytes)
 294 {
 295         struct dm_path *path;
 296
 297         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 298         if (!path)
 299                 return -ENXIO;
 300
 301         m->current_pgpath = path_to_pgpath(path);
 302
 303         if (m->current_pg != pg)
 304                 __switch_pg(m, m->current_pgpath);
 305
 306         return 0;
 307 }
 308
 309 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 310 {
 311         struct priority_group *pg;
 312         unsigned bypassed = 1;
 313
 314         if (!m->nr_valid_paths)
 315                 goto failed;
 316
 317         /* Were we instructed to switch PG? */
 318         if (m->next_pg) {
 319                 pg = m->next_pg;
 320                 m->next_pg = NULL;
 321                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 322                         return;
 323         }
 324
 325         /* Don't change PG until it has no remaining paths */
 326         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 327                 return;
 328
 329         /*
 330          * Loop through priority groups until we find a valid path.
 331          * First time we skip PGs marked 'bypassed'.
 332          * Second time we only try the ones we skipped, but set
 333          * pg_init_delay_retry so we do not hammer controllers.
 334          */
 335         do {
 336                 list_for_each_entry(pg, &m->priority_groups, list) {
 337                         if (pg->bypassed == bypassed)
 338                                 continue;
 339                         if (!__choose_path_in_pg(m, pg, nr_bytes)) {
 340                                 if (!bypassed)
 341                                         m->pg_init_delay_retry = 1;
 342                                 return;
 343                         }
 344                 }
 345         } while (bypassed--);
 346
 347 failed:
 348         m->current_pgpath = NULL;
 349         m->current_pg = NULL;
 350 }
 351
 352 /*
 353  * Check whether bios must be queued in the device-mapper core rather
 354  * than here in the target.
 355  *
 356  * m->lock must be held on entry.
 357  *
 358  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 359  * same value then we are not between multipath_presuspend()
 360  * and multipath_resume() calls and we have no need to check
 361  * for the DMF_NOFLUSH_SUSPENDING flag.
 362  */
 363 static int __must_push_back(struct multipath *m)
 364 {
 365         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 366                 dm_noflush_suspending(m->ti));
 367 }
 368
 369 static int map_io(struct multipath *m, struct request *clone,
 370                   union map_info *map_context, unsigned was_queued)
 371 {
 372         int r = DM_MAPIO_REMAPPED;
 373         size_t nr_bytes = blk_rq_bytes(clone);
 374         unsigned long flags;
 375         struct pgpath *pgpath;
 376         struct block_device *bdev;
 377         struct dm_mpath_io *mpio = map_context->ptr;
 378
 379         spin_lock_irqsave(&m->lock, flags);
 380
 381         /* Do we need to select a new pgpath? */
 382         if (!m->current_pgpath ||
 383             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 384                 __choose_pgpath(m, nr_bytes);
 385
 386         pgpath = m->current_pgpath;
 387
 388         if (was_queued)
 389                 m->queue_size--;
 390
 391         if ((pgpath && m->queue_io) ||
 392             (!pgpath && m->queue_if_no_path)) {
 393                 /* Queue for the daemon to resubmit */
 394                 list_add_tail(&clone->queuelist, &m->queued_ios);
 395                 m->queue_size++;
 396                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 397                     !m->queue_io)
 398                         queue_work(kmultipathd, &m->process_queued_ios);
 399                 pgpath = NULL;
 400                 r = DM_MAPIO_SUBMITTED;
 401         } else if (pgpath) {
 402                 bdev = pgpath->path.dev->bdev;
 403                 clone->q = bdev_get_queue(bdev);
 404                 clone->rq_disk = bdev->bd_disk;
 405         } else if (__must_push_back(m))
 406                 r = DM_MAPIO_REQUEUE;
 407         else
 408                 r = -EIO;       /* Failed */
 409
 410         mpio->pgpath = pgpath;
 411         mpio->nr_bytes = nr_bytes;
 412
 413         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 414                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 415                                               nr_bytes);
 416
 417         spin_unlock_irqrestore(&m->lock, flags);
 418
 419         return r;
 420 }
 421
 422 /*
 423  * If we run out of usable paths, should we queue I/O or error it?
 424  */
 425 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 426                             unsigned save_old_value)
 427 {
 428         unsigned long flags;
 429
 430         spin_lock_irqsave(&m->lock, flags);
 431
 432         if (save_old_value)
 433                 m->saved_queue_if_no_path = m->queue_if_no_path;
 434         else
 435                 m->saved_queue_if_no_path = queue_if_no_path;
 436         m->queue_if_no_path = queue_if_no_path;
 437         if (!m->queue_if_no_path && m->queue_size)
 438                 queue_work(kmultipathd, &m->process_queued_ios);
 439
 440         spin_unlock_irqrestore(&m->lock, flags);
 441
 442         return 0;
 443 }
 444
 445 /*-----------------------------------------------------------------
 446  * The multipath daemon is responsible for resubmitting queued ios.
 447  *---------------------------------------------------------------*/
 448
 449 static void dispatch_queued_ios(struct multipath *m)
 450 {
 451         int r;
 452         unsigned long flags;
 453         union map_info *info;
 454         struct request *clone, *n;
 455         LIST_HEAD(cl);
 456
 457         spin_lock_irqsave(&m->lock, flags);
 458         list_splice_init(&m->queued_ios, &cl);
 459         spin_unlock_irqrestore(&m->lock, flags);
 460
 461         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 462                 list_del_init(&clone->queuelist);
 463
 464                 info = dm_get_rq_mapinfo(clone);
 465
 466                 r = map_io(m, clone, info, 1);
 467                 if (r < 0) {
 468                         clear_mapinfo(m, info);
 469                         dm_kill_unmapped_request(clone, r);
 470                 } else if (r == DM_MAPIO_REMAPPED)
 471                         dm_dispatch_request(clone);
 472                 else if (r == DM_MAPIO_REQUEUE) {
 473                         clear_mapinfo(m, info);
 474                         dm_requeue_unmapped_request(clone);
 475                 }
 476         }
 477 }
 478
 479 static void process_queued_ios(struct work_struct *work)
 480 {
 481         struct multipath *m =
 482                 container_of(work, struct multipath, process_queued_ios);
 483         struct pgpath *pgpath = NULL;
 484         unsigned must_queue = 1;
 485         unsigned long flags;
 486
 487         spin_lock_irqsave(&m->lock, flags);
 488
 489         if (!m->queue_size)
 490                 goto out;
 491
 492         if (!m->current_pgpath)
 493                 __choose_pgpath(m, 0);
 494
 495         pgpath = m->current_pgpath;
 496
 497         if ((pgpath && !m->queue_io) ||
 498             (!pgpath && !m->queue_if_no_path))
 499                 must_queue = 0;
 500
 501         if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
 502                 __pg_init_all_paths(m);
 503
 504 out:
 505         spin_unlock_irqrestore(&m->lock, flags);
 506         if (!must_queue)
 507                 dispatch_queued_ios(m);
 508 }
 509
 510 /*
 511  * An event is triggered whenever a path is taken out of use.
 512  * Includes path failure and PG bypass.
 513  */
 514 static void trigger_event(struct work_struct *work)
 515 {
 516         struct multipath *m =
 517                 container_of(work, struct multipath, trigger_event);
 518
 519         dm_table_event(m->ti->table);
 520 }
 521
 522 /*-----------------------------------------------------------------
 523  * Constructor/argument parsing:
 524  * <#multipath feature args> [<arg>]*
 525  * <#hw_handler args> [hw_handler [<arg>]*]
 526  * <#priority groups>
 527  * <initial priority group>
 528  *     [<selector> <#selector args> [<arg>]*
 529  *      <#paths> <#per-path selector args>
 530  *         [<path> [<arg>]* ]+ ]+
 531  *---------------------------------------------------------------*/
 532 static int parse_path_selector(struct dm_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 dm_arg _args[] = {
 540                 {0, 1024, "invalid number of path selector args"},
 541         };
 542
 543         pst = dm_get_path_selector(dm_shift_arg(as));
 544         if (!pst) {
 545                 ti->error = "unknown path selector type";
 546                 return -EINVAL;
 547         }
 548
 549         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 550         if (r) {
 551                 dm_put_path_selector(pst);
 552                 return -EINVAL;
 553         }
 554
 555         r = pst->create(&pg->ps, ps_argc, as->argv);
 556         if (r) {
 557                 dm_put_path_selector(pst);
 558                 ti->error = "path selector constructor failed";
 559                 return r;
 560         }
 561
 562         pg->ps.type = pst;
 563         dm_consume_args(as, ps_argc);
 564
 565         return 0;
 566 }
 567
 568 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 569                                struct dm_target *ti)
 570 {
 571         int r;
 572         struct pgpath *p;
 573         struct multipath *m = ti->private;
 574
 575         /* we need at least a path arg */
 576         if (as->argc < 1) {
 577                 ti->error = "no device given";
 578                 return ERR_PTR(-EINVAL);
 579         }
 580
 581         p = alloc_pgpath();
 582         if (!p)
 583                 return ERR_PTR(-ENOMEM);
 584
 585         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 586                           &p->path.dev);
 587         if (r) {
 588                 ti->error = "error getting device";
 589                 goto bad;
 590         }
 591
 592         if (m->hw_handler_name) {
 593                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 594
 595                 r = scsi_dh_attach(q, m->hw_handler_name);
 596                 if (r == -EBUSY) {
 597                         /*
 598                          * Already attached to different hw_handler,
 599                          * try to reattach with correct one.
 600                          */
 601                         scsi_dh_detach(q);
 602                         r = scsi_dh_attach(q, m->hw_handler_name);
 603                 }
 604
 605                 if (r < 0) {
 606                         ti->error = "error attaching hardware handler";
 607                         dm_put_device(ti, p->path.dev);
 608                         goto bad;
 609                 }
 610
 611                 if (m->hw_handler_params) {
 612                         r = scsi_dh_set_params(q, m->hw_handler_params);
 613                         if (r < 0) {
 614                                 ti->error = "unable to set hardware "
 615                                                         "handler parameters";
 616                                 scsi_dh_detach(q);
 617                                 dm_put_device(ti, p->path.dev);
 618                                 goto bad;
 619                         }
 620                 }
 621         }
 622
 623         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 624         if (r) {
 625                 dm_put_device(ti, p->path.dev);
 626                 goto bad;
 627         }
 628
 629         return p;
 630
 631  bad:
 632         free_pgpath(p);
 633         return ERR_PTR(r);
 634 }
 635
 636 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 637                                                    struct multipath *m)
 638 {
 639         static struct dm_arg _args[] = {
 640                 {1, 1024, "invalid number of paths"},
 641                 {0, 1024, "invalid number of selector args"}
 642         };
 643
 644         int r;
 645         unsigned i, nr_selector_args, nr_args;
 646         struct priority_group *pg;
 647         struct dm_target *ti = m->ti;
 648
 649         if (as->argc < 2) {
 650                 as->argc = 0;
 651                 ti->error = "not enough priority group arguments";
 652                 return ERR_PTR(-EINVAL);
 653         }
 654
 655         pg = alloc_priority_group();
 656         if (!pg) {
 657                 ti->error = "couldn't allocate priority group";
 658                 return ERR_PTR(-ENOMEM);
 659         }
 660         pg->m = m;
 661
 662         r = parse_path_selector(as, pg, ti);
 663         if (r)
 664                 goto bad;
 665
 666         /*
 667          * read the paths
 668          */
 669         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 670         if (r)
 671                 goto bad;
 672
 673         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 674         if (r)
 675                 goto bad;
 676
 677         nr_args = 1 + nr_selector_args;
 678         for (i = 0; i < pg->nr_pgpaths; i++) {
 679                 struct pgpath *pgpath;
 680                 struct dm_arg_set path_args;
 681
 682                 if (as->argc < nr_args) {
 683                         ti->error = "not enough path parameters";
 684                         r = -EINVAL;
 685                         goto bad;
 686                 }
 687
 688                 path_args.argc = nr_args;
 689                 path_args.argv = as->argv;
 690
 691                 pgpath = parse_path(&path_args, &pg->ps, ti);
 692                 if (IS_ERR(pgpath)) {
 693                         r = PTR_ERR(pgpath);
 694                         goto bad;
 695                 }
 696
 697                 pgpath->pg = pg;
 698                 list_add_tail(&pgpath->list, &pg->pgpaths);
 699                 dm_consume_args(as, nr_args);
 700         }
 701
 702         return pg;
 703
 704  bad:
 705         free_priority_group(pg, ti);
 706         return ERR_PTR(r);
 707 }
 708
 709 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 710 {
 711         unsigned hw_argc;
 712         int ret;
 713         struct dm_target *ti = m->ti;
 714
 715         static struct dm_arg _args[] = {
 716                 {0, 1024, "invalid number of hardware handler args"},
 717         };
 718
 719         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 720                 return -EINVAL;
 721
 722         if (!hw_argc)
 723                 return 0;
 724
 725         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 726         if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
 727                                      "scsi_dh_%s", m->hw_handler_name)) {
 728                 ti->error = "unknown hardware handler type";
 729                 ret = -EINVAL;
 730                 goto fail;
 731         }
 732
 733         if (hw_argc > 1) {
 734                 char *p;
 735                 int i, j, len = 4;
 736
 737                 for (i = 0; i <= hw_argc - 2; i++)
 738                         len += strlen(as->argv[i]) + 1;
 739                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 740                 if (!p) {
 741                         ti->error = "memory allocation failed";
 742                         ret = -ENOMEM;
 743                         goto fail;
 744                 }
 745                 j = sprintf(p, "%d", hw_argc - 1);
 746                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 747                         j = sprintf(p, "%s", as->argv[i]);
 748         }
 749         dm_consume_args(as, hw_argc - 1);
 750
 751         return 0;
 752 fail:
 753         kfree(m->hw_handler_name);
 754         m->hw_handler_name = NULL;
 755         return ret;
 756 }
 757
 758 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 759 {
 760         int r;
 761         unsigned argc;
 762         struct dm_target *ti = m->ti;
 763         const char *arg_name;
 764
 765         static struct dm_arg _args[] = {
 766                 {0, 5, "invalid number of feature args"},
 767                 {1, 50, "pg_init_retries must be between 1 and 50"},
 768                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 769         };
 770
 771         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 772         if (r)
 773                 return -EINVAL;
 774
 775         if (!argc)
 776                 return 0;
 777
 778         do {
 779                 arg_name = dm_shift_arg(as);
 780                 argc--;
 781
 782                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 783                         r = queue_if_no_path(m, 1, 0);
 784                         continue;
 785                 }
 786
 787                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 788                     (argc >= 1)) {
 789                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 790                         argc--;
 791                         continue;
 792                 }
 793
 794                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 795                     (argc >= 1)) {
 796                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 797                         argc--;
 798                         continue;
 799                 }
 800
 801                 ti->error = "Unrecognised multipath feature request";
 802                 r = -EINVAL;
 803         } while (argc && !r);
 804
 805         return r;
 806 }
 807
 808 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 809                          char **argv)
 810 {
 811         /* target arguments */
 812         static struct dm_arg _args[] = {
 813                 {0, 1024, "invalid number of priority groups"},
 814                 {0, 1024, "invalid initial priority group number"},
 815         };
 816
 817         int r;
 818         struct multipath *m;
 819         struct dm_arg_set as;
 820         unsigned pg_count = 0;
 821         unsigned next_pg_num;
 822
 823         as.argc = argc;
 824         as.argv = argv;
 825
 826         m = alloc_multipath(ti);
 827         if (!m) {
 828                 ti->error = "can't allocate multipath";
 829                 return -EINVAL;
 830         }
 831
 832         r = parse_features(&as, m);
 833         if (r)
 834                 goto bad;
 835
 836         r = parse_hw_handler(&as, m);
 837         if (r)
 838                 goto bad;
 839
 840         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 841         if (r)
 842                 goto bad;
 843
 844         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 845         if (r)
 846                 goto bad;
 847
 848         if ((!m->nr_priority_groups && next_pg_num) ||
 849             (m->nr_priority_groups && !next_pg_num)) {
 850                 ti->error = "invalid initial priority group";
 851                 r = -EINVAL;
 852                 goto bad;
 853         }
 854
 855         /* parse the priority groups */
 856         while (as.argc) {
 857                 struct priority_group *pg;
 858
 859                 pg = parse_priority_group(&as, m);
 860                 if (IS_ERR(pg)) {
 861                         r = PTR_ERR(pg);
 862                         goto bad;
 863                 }
 864
 865                 m->nr_valid_paths += pg->nr_pgpaths;
 866                 list_add_tail(&pg->list, &m->priority_groups);
 867                 pg_count++;
 868                 pg->pg_num = pg_count;
 869                 if (!--next_pg_num)
 870                         m->next_pg = pg;
 871         }
 872
 873         if (pg_count != m->nr_priority_groups) {
 874                 ti->error = "priority group count mismatch";
 875                 r = -EINVAL;
 876                 goto bad;
 877         }
 878
 879         ti->num_flush_requests = 1;
 880         ti->num_discard_requests = 1;
 881
 882         return 0;
 883
 884  bad:
 885         free_multipath(m);
 886         return r;
 887 }
 888
 889 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 890 {
 891         DECLARE_WAITQUEUE(wait, current);
 892         unsigned long flags;
 893
 894         add_wait_queue(&m->pg_init_wait, &wait);
 895
 896         while (1) {
 897                 set_current_state(TASK_UNINTERRUPTIBLE);
 898
 899                 spin_lock_irqsave(&m->lock, flags);
 900                 if (!m->pg_init_in_progress) {
 901                         spin_unlock_irqrestore(&m->lock, flags);
 902                         break;
 903                 }
 904                 spin_unlock_irqrestore(&m->lock, flags);
 905
 906                 io_schedule();
 907         }
 908         set_current_state(TASK_RUNNING);
 909
 910         remove_wait_queue(&m->pg_init_wait, &wait);
 911 }
 912
 913 static void flush_multipath_work(struct multipath *m)
 914 {
 915         flush_workqueue(kmpath_handlerd);
 916         multipath_wait_for_pg_init_completion(m);
 917         flush_workqueue(kmultipathd);
 918         flush_work_sync(&m->trigger_event);
 919 }
 920
 921 static void multipath_dtr(struct dm_target *ti)
 922 {
 923         struct multipath *m = ti->private;
 924
 925         flush_multipath_work(m);
 926         free_multipath(m);
 927 }
 928
 929 /*
 930  * Map cloned requests
 931  */
 932 static int multipath_map(struct dm_target *ti, struct request *clone,
 933                          union map_info *map_context)
 934 {
 935         int r;
 936         struct multipath *m = (struct multipath *) ti->private;
 937
 938         if (set_mapinfo(m, map_context) < 0)
 939                 /* ENOMEM, requeue */
 940                 return DM_MAPIO_REQUEUE;
 941
 942         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 943         r = map_io(m, clone, map_context, 0);
 944         if (r < 0 || r == DM_MAPIO_REQUEUE)
 945                 clear_mapinfo(m, map_context);
 946
 947         return r;
 948 }
 949
 950 /*
 951  * Take a path out of use.
 952  */
 953 static int fail_path(struct pgpath *pgpath)
 954 {
 955         unsigned long flags;
 956         struct multipath *m = pgpath->pg->m;
 957
 958         spin_lock_irqsave(&m->lock, flags);
 959
 960         if (!pgpath->is_active)
 961                 goto out;
 962
 963         DMWARN("Failing path %s.", pgpath->path.dev->name);
 964
 965         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 966         pgpath->is_active = 0;
 967         pgpath->fail_count++;
 968
 969         m->nr_valid_paths--;
 970
 971         if (pgpath == m->current_pgpath)
 972                 m->current_pgpath = NULL;
 973
 974         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 975                       pgpath->path.dev->name, m->nr_valid_paths);
 976
 977         schedule_work(&m->trigger_event);
 978
 979 out:
 980         spin_unlock_irqrestore(&m->lock, flags);
 981
 982         return 0;
 983 }
 984
 985 /*
 986  * Reinstate a previously-failed path
 987  */
 988 static int reinstate_path(struct pgpath *pgpath)
 989 {
 990         int r = 0;
 991         unsigned long flags;
 992         struct multipath *m = pgpath->pg->m;
 993
 994         spin_lock_irqsave(&m->lock, flags);
 995
 996         if (pgpath->is_active)
 997                 goto out;
 998
 999         if (!pgpath->pg->ps.type->reinstate_path) {
1000                 DMWARN("Reinstate path not supported by path selector %s",
1001                        pgpath->pg->ps.type->name);
1002                 r = -EINVAL;
1003                 goto out;
1004         }
1005
1006         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1007         if (r)
1008                 goto out;
1009
1010         pgpath->is_active = 1;
1011
1012         if (!m->nr_valid_paths++ && m->queue_size) {
1013                 m->current_pgpath = NULL;
1014                 queue_work(kmultipathd, &m->process_queued_ios);
1015         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1016                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1017                         m->pg_init_in_progress++;
1018         }
1019
1020         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1021                       pgpath->path.dev->name, m->nr_valid_paths);
1022
1023         schedule_work(&m->trigger_event);
1024
1025 out:
1026         spin_unlock_irqrestore(&m->lock, flags);
1027
1028         return r;
1029 }
1030
1031 /*
1032  * Fail or reinstate all paths that match the provided struct dm_dev.
1033  */
1034 static int action_dev(struct multipath *m, struct dm_dev *dev,
1035                       action_fn action)
1036 {
1037         int r = -EINVAL;
1038         struct pgpath *pgpath;
1039         struct priority_group *pg;
1040
1041         list_for_each_entry(pg, &m->priority_groups, list) {
1042                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1043                         if (pgpath->path.dev == dev)
1044                                 r = action(pgpath);
1045                 }
1046         }
1047
1048         return r;
1049 }
1050
1051 /*
1052  * Temporarily try to avoid having to use the specified PG
1053  */
1054 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1055                       int bypassed)
1056 {
1057         unsigned long flags;
1058
1059         spin_lock_irqsave(&m->lock, flags);
1060
1061         pg->bypassed = bypassed;
1062         m->current_pgpath = NULL;
1063         m->current_pg = NULL;
1064
1065         spin_unlock_irqrestore(&m->lock, flags);
1066
1067         schedule_work(&m->trigger_event);
1068 }
1069
1070 /*
1071  * Switch to using the specified PG from the next I/O that gets mapped
1072  */
1073 static int switch_pg_num(struct multipath *m, const char *pgstr)
1074 {
1075         struct priority_group *pg;
1076         unsigned pgnum;
1077         unsigned long flags;
1078         char dummy;
1079
1080         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1081             (pgnum > m->nr_priority_groups)) {
1082                 DMWARN("invalid PG number supplied to switch_pg_num");
1083                 return -EINVAL;
1084         }
1085
1086         spin_lock_irqsave(&m->lock, flags);
1087         list_for_each_entry(pg, &m->priority_groups, list) {
1088                 pg->bypassed = 0;
1089                 if (--pgnum)
1090                         continue;
1091
1092                 m->current_pgpath = NULL;
1093                 m->current_pg = NULL;
1094                 m->next_pg = pg;
1095         }
1096         spin_unlock_irqrestore(&m->lock, flags);
1097
1098         schedule_work(&m->trigger_event);
1099         return 0;
1100 }
1101
1102 /*
1103  * Set/clear bypassed status of a PG.
1104  * PGs are numbered upwards from 1 in the order they were declared.
1105  */
1106 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1107 {
1108         struct priority_group *pg;
1109         unsigned pgnum;
1110         char dummy;
1111
1112         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1113             (pgnum > m->nr_priority_groups)) {
1114                 DMWARN("invalid PG number supplied to bypass_pg");
1115                 return -EINVAL;
1116         }
1117
1118         list_for_each_entry(pg, &m->priority_groups, list) {
1119                 if (!--pgnum)
1120                         break;
1121         }
1122
1123         bypass_pg(m, pg, bypassed);
1124         return 0;
1125 }
1126
1127 /*
1128  * Should we retry pg_init immediately?
1129  */
1130 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1131 {
1132         unsigned long flags;
1133         int limit_reached = 0;
1134
1135         spin_lock_irqsave(&m->lock, flags);
1136
1137         if (m->pg_init_count <= m->pg_init_retries)
1138                 m->pg_init_required = 1;
1139         else
1140                 limit_reached = 1;
1141
1142         spin_unlock_irqrestore(&m->lock, flags);
1143
1144         return limit_reached;
1145 }
1146
1147 static void pg_init_done(void *data, int errors)
1148 {
1149         struct pgpath *pgpath = data;
1150         struct priority_group *pg = pgpath->pg;
1151         struct multipath *m = pg->m;
1152         unsigned long flags;
1153         unsigned delay_retry = 0;
1154
1155         /* device or driver problems */
1156         switch (errors) {
1157         case SCSI_DH_OK:
1158                 break;
1159         case SCSI_DH_NOSYS:
1160                 if (!m->hw_handler_name) {
1161                         errors = 0;
1162                         break;
1163                 }
1164                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1165                       "Error %d.", m->hw_handler_name, errors);
1166                 /*
1167                  * Fail path for now, so we do not ping pong
1168                  */
1169                 fail_path(pgpath);
1170                 break;
1171         case SCSI_DH_DEV_TEMP_BUSY:
1172                 /*
1173                  * Probably doing something like FW upgrade on the
1174                  * controller so try the other pg.
1175                  */
1176                 bypass_pg(m, pg, 1);
1177                 break;
1178         case SCSI_DH_RETRY:
1179                 /* Wait before retrying. */
1180                 delay_retry = 1;
1181         case SCSI_DH_IMM_RETRY:
1182         case SCSI_DH_RES_TEMP_UNAVAIL:
1183                 if (pg_init_limit_reached(m, pgpath))
1184                         fail_path(pgpath);
1185                 errors = 0;
1186                 break;
1187         default:
1188                 /*
1189                  * We probably do not want to fail the path for a device
1190                  * error, but this is what the old dm did. In future
1191                  * patches we can do more advanced handling.
1192                  */
1193                 fail_path(pgpath);
1194         }
1195
1196         spin_lock_irqsave(&m->lock, flags);
1197         if (errors) {
1198                 if (pgpath == m->current_pgpath) {
1199                         DMERR("Could not failover device. Error %d.", errors);
1200                         m->current_pgpath = NULL;
1201                         m->current_pg = NULL;
1202                 }
1203         } else if (!m->pg_init_required)
1204                 pg->bypassed = 0;
1205
1206         if (--m->pg_init_in_progress)
1207                 /* Activations of other paths are still on going */
1208                 goto out;
1209
1210         if (!m->pg_init_required)
1211                 m->queue_io = 0;
1212
1213         m->pg_init_delay_retry = delay_retry;
1214         queue_work(kmultipathd, &m->process_queued_ios);
1215
1216         /*
1217          * Wake up any thread waiting to suspend.
1218          */
1219         wake_up(&m->pg_init_wait);
1220
1221 out:
1222         spin_unlock_irqrestore(&m->lock, flags);
1223 }
1224
1225 static void activate_path(struct work_struct *work)
1226 {
1227         struct pgpath *pgpath =
1228                 container_of(work, struct pgpath, activate_path.work);
1229
1230         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1231                                 pg_init_done, pgpath);
1232 }
1233
1234 /*
1235  * end_io handling
1236  */
1237 static int do_end_io(struct multipath *m, struct request *clone,
1238                      int error, struct dm_mpath_io *mpio)
1239 {
1240         /*
1241          * We don't queue any clone request inside the multipath target
1242          * during end I/O handling, since those clone requests don't have
1243          * bio clones.  If we queue them inside the multipath target,
1244          * we need to make bio clones, that requires memory allocation.
1245          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1246          *  don't have bio clones.)
1247          * Instead of queueing the clone request here, we queue the original
1248          * request into dm core, which will remake a clone request and
1249          * clone bios for it and resubmit it later.
1250          */
1251         int r = DM_ENDIO_REQUEUE;
1252         unsigned long flags;
1253
1254         if (!error && !clone->errors)
1255                 return 0;       /* I/O complete */
1256
1257         if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1258                 return error;
1259
1260         if (mpio->pgpath)
1261                 fail_path(mpio->pgpath);
1262
1263         spin_lock_irqsave(&m->lock, flags);
1264         if (!m->nr_valid_paths) {
1265                 if (!m->queue_if_no_path) {
1266                         if (!__must_push_back(m))
1267                                 r = -EIO;
1268                 } else {
1269                         if (error == -EBADE)
1270                                 r = error;
1271                 }
1272         }
1273         spin_unlock_irqrestore(&m->lock, flags);
1274
1275         return r;
1276 }
1277
1278 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1279                             int error, union map_info *map_context)
1280 {
1281         struct multipath *m = ti->private;
1282         struct dm_mpath_io *mpio = map_context->ptr;
1283         struct pgpath *pgpath = mpio->pgpath;
1284         struct path_selector *ps;
1285         int r;
1286
1287         BUG_ON(!mpio);
1288
1289         r  = do_end_io(m, clone, error, mpio);
1290         if (pgpath) {
1291                 ps = &pgpath->pg->ps;
1292                 if (ps->type->end_io)
1293                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1294         }
1295         clear_mapinfo(m, map_context);
1296
1297         return r;
1298 }
1299
1300 /*
1301  * Suspend can't complete until all the I/O is processed so if
1302  * the last path fails we must error any remaining I/O.
1303  * Note that if the freeze_bdev fails while suspending, the
1304  * queue_if_no_path state is lost - userspace should reset it.
1305  */
1306 static void multipath_presuspend(struct dm_target *ti)
1307 {
1308         struct multipath *m = (struct multipath *) ti->private;
1309
1310         queue_if_no_path(m, 0, 1);
1311 }
1312
1313 static void multipath_postsuspend(struct dm_target *ti)
1314 {
1315         struct multipath *m = ti->private;
1316
1317         mutex_lock(&m->work_mutex);
1318         flush_multipath_work(m);
1319         mutex_unlock(&m->work_mutex);
1320 }
1321
1322 /*
1323  * Restore the queue_if_no_path setting.
1324  */
1325 static void multipath_resume(struct dm_target *ti)
1326 {
1327         struct multipath *m = (struct multipath *) ti->private;
1328         unsigned long flags;
1329
1330         spin_lock_irqsave(&m->lock, flags);
1331         m->queue_if_no_path = m->saved_queue_if_no_path;
1332         spin_unlock_irqrestore(&m->lock, flags);
1333 }
1334
1335 /*
1336  * Info output has the following format:
1337  * num_multipath_feature_args [multipath_feature_args]*
1338  * num_handler_status_args [handler_status_args]*
1339  * num_groups init_group_number
1340  *            [A|D|E num_ps_status_args [ps_status_args]*
1341  *             num_paths num_selector_args
1342  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1343  *
1344  * Table output has the following format (identical to the constructor string):
1345  * num_feature_args [features_args]*
1346  * num_handler_args hw_handler [hw_handler_args]*
1347  * num_groups init_group_number
1348  *     [priority selector-name num_ps_args [ps_args]*
1349  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1350  */
1351 static int multipath_status(struct dm_target *ti, status_type_t type,
1352                             char *result, unsigned int maxlen)
1353 {
1354         int sz = 0;
1355         unsigned long flags;
1356         struct multipath *m = (struct multipath *) ti->private;
1357         struct priority_group *pg;
1358         struct pgpath *p;
1359         unsigned pg_num;
1360         char state;
1361
1362         spin_lock_irqsave(&m->lock, flags);
1363
1364         /* Features */
1365         if (type == STATUSTYPE_INFO)
1366                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1367         else {
1368                 DMEMIT("%u ", m->queue_if_no_path +
1369                               (m->pg_init_retries > 0) * 2 +
1370                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2);
1371                 if (m->queue_if_no_path)
1372                         DMEMIT("queue_if_no_path ");
1373                 if (m->pg_init_retries)
1374                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1375                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1376                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1377         }
1378
1379         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1380                 DMEMIT("0 ");
1381         else
1382                 DMEMIT("1 %s ", m->hw_handler_name);
1383
1384         DMEMIT("%u ", m->nr_priority_groups);
1385
1386         if (m->next_pg)
1387                 pg_num = m->next_pg->pg_num;
1388         else if (m->current_pg)
1389                 pg_num = m->current_pg->pg_num;
1390         else
1391                 pg_num = (m->nr_priority_groups ? 1 : 0);
1392
1393         DMEMIT("%u ", pg_num);
1394
1395         switch (type) {
1396         case STATUSTYPE_INFO:
1397                 list_for_each_entry(pg, &m->priority_groups, list) {
1398                         if (pg->bypassed)
1399                                 state = 'D';    /* Disabled */
1400                         else if (pg == m->current_pg)
1401                                 state = 'A';    /* Currently Active */
1402                         else
1403                                 state = 'E';    /* Enabled */
1404
1405                         DMEMIT("%c ", state);
1406
1407                         if (pg->ps.type->status)
1408                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1409                                                           result + sz,
1410                                                           maxlen - sz);
1411                         else
1412                                 DMEMIT("0 ");
1413
1414                         DMEMIT("%u %u ", pg->nr_pgpaths,
1415                                pg->ps.type->info_args);
1416
1417                         list_for_each_entry(p, &pg->pgpaths, list) {
1418                                 DMEMIT("%s %s %u ", p->path.dev->name,
1419                                        p->is_active ? "A" : "F",
1420                                        p->fail_count);
1421                                 if (pg->ps.type->status)
1422                                         sz += pg->ps.type->status(&pg->ps,
1423                                               &p->path, type, result + sz,
1424                                               maxlen - sz);
1425                         }
1426                 }
1427                 break;
1428
1429         case STATUSTYPE_TABLE:
1430                 list_for_each_entry(pg, &m->priority_groups, list) {
1431                         DMEMIT("%s ", pg->ps.type->name);
1432
1433                         if (pg->ps.type->status)
1434                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1435                                                           result + sz,
1436                                                           maxlen - sz);
1437                         else
1438                                 DMEMIT("0 ");
1439
1440                         DMEMIT("%u %u ", pg->nr_pgpaths,
1441                                pg->ps.type->table_args);
1442
1443                         list_for_each_entry(p, &pg->pgpaths, list) {
1444                                 DMEMIT("%s ", p->path.dev->name);
1445                                 if (pg->ps.type->status)
1446                                         sz += pg->ps.type->status(&pg->ps,
1447                                               &p->path, type, result + sz,
1448                                               maxlen - sz);
1449                         }
1450                 }
1451                 break;
1452         }
1453
1454         spin_unlock_irqrestore(&m->lock, flags);
1455
1456         return 0;
1457 }
1458
1459 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1460 {
1461         int r = -EINVAL;
1462         struct dm_dev *dev;
1463         struct multipath *m = (struct multipath *) ti->private;
1464         action_fn action;
1465
1466         mutex_lock(&m->work_mutex);
1467
1468         if (dm_suspended(ti)) {
1469                 r = -EBUSY;
1470                 goto out;
1471         }
1472
1473         if (argc == 1) {
1474                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1475                         r = queue_if_no_path(m, 1, 0);
1476                         goto out;
1477                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1478                         r = queue_if_no_path(m, 0, 0);
1479                         goto out;
1480                 }
1481         }
1482
1483         if (argc != 2) {
1484                 DMWARN("Unrecognised multipath message received.");
1485                 goto out;
1486         }
1487
1488         if (!strcasecmp(argv[0], "disable_group")) {
1489                 r = bypass_pg_num(m, argv[1], 1);
1490                 goto out;
1491         } else if (!strcasecmp(argv[0], "enable_group")) {
1492                 r = bypass_pg_num(m, argv[1], 0);
1493                 goto out;
1494         } else if (!strcasecmp(argv[0], "switch_group")) {
1495                 r = switch_pg_num(m, argv[1]);
1496                 goto out;
1497         } else if (!strcasecmp(argv[0], "reinstate_path"))
1498                 action = reinstate_path;
1499         else if (!strcasecmp(argv[0], "fail_path"))
1500                 action = fail_path;
1501         else {
1502                 DMWARN("Unrecognised multipath message received.");
1503                 goto out;
1504         }
1505
1506         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1507         if (r) {
1508                 DMWARN("message: error getting device %s",
1509                        argv[1]);
1510                 goto out;
1511         }
1512
1513         r = action_dev(m, dev, action);
1514
1515         dm_put_device(ti, dev);
1516
1517 out:
1518         mutex_unlock(&m->work_mutex);
1519         return r;
1520 }
1521
1522 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1523                            unsigned long arg)
1524 {
1525         struct multipath *m = (struct multipath *) ti->private;
1526         struct block_device *bdev = NULL;
1527         fmode_t mode = 0;
1528         unsigned long flags;
1529         int r = 0;
1530
1531         spin_lock_irqsave(&m->lock, flags);
1532
1533         if (!m->current_pgpath)
1534                 __choose_pgpath(m, 0);
1535
1536         if (m->current_pgpath) {
1537                 bdev = m->current_pgpath->path.dev->bdev;
1538                 mode = m->current_pgpath->path.dev->mode;
1539         }
1540
1541         if (m->queue_io)
1542                 r = -EAGAIN;
1543         else if (!bdev)
1544                 r = -EIO;
1545
1546         spin_unlock_irqrestore(&m->lock, flags);
1547
1548         /*
1549          * Only pass ioctls through if the device sizes match exactly.
1550          */
1551         if (!r && ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT)
1552                 r = scsi_verify_blk_ioctl(NULL, cmd);
1553
1554         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1555 }
1556
1557 static int multipath_iterate_devices(struct dm_target *ti,
1558                                      iterate_devices_callout_fn fn, void *data)
1559 {
1560         struct multipath *m = ti->private;
1561         struct priority_group *pg;
1562         struct pgpath *p;
1563         int ret = 0;
1564
1565         list_for_each_entry(pg, &m->priority_groups, list) {
1566                 list_for_each_entry(p, &pg->pgpaths, list) {
1567                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1568                         if (ret)
1569                                 goto out;
1570                 }
1571         }
1572
1573 out:
1574         return ret;
1575 }
1576
1577 static int __pgpath_busy(struct pgpath *pgpath)
1578 {
1579         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1580
1581         return dm_underlying_device_busy(q);
1582 }
1583
1584 /*
1585  * We return "busy", only when we can map I/Os but underlying devices
1586  * are busy (so even if we map I/Os now, the I/Os will wait on
1587  * the underlying queue).
1588  * In other words, if we want to kill I/Os or queue them inside us
1589  * due to map unavailability, we don't return "busy".  Otherwise,
1590  * dm core won't give us the I/Os and we can't do what we want.
1591  */
1592 static int multipath_busy(struct dm_target *ti)
1593 {
1594         int busy = 0, has_active = 0;
1595         struct multipath *m = ti->private;
1596         struct priority_group *pg;
1597         struct pgpath *pgpath;
1598         unsigned long flags;
1599
1600         spin_lock_irqsave(&m->lock, flags);
1601
1602         /* Guess which priority_group will be used at next mapping time */
1603         if (unlikely(!m->current_pgpath && m->next_pg))
1604                 pg = m->next_pg;
1605         else if (likely(m->current_pg))
1606                 pg = m->current_pg;
1607         else
1608                 /*
1609                  * We don't know which pg will be used at next mapping time.
1610                  * We don't call __choose_pgpath() here to avoid to trigger
1611                  * pg_init just by busy checking.
1612                  * So we don't know whether underlying devices we will be using
1613                  * at next mapping time are busy or not. Just try mapping.
1614                  */
1615                 goto out;
1616
1617         /*
1618          * If there is one non-busy active path at least, the path selector
1619          * will be able to select it. So we consider such a pg as not busy.
1620          */
1621         busy = 1;
1622         list_for_each_entry(pgpath, &pg->pgpaths, list)
1623                 if (pgpath->is_active) {
1624                         has_active = 1;
1625
1626                         if (!__pgpath_busy(pgpath)) {
1627                                 busy = 0;
1628                                 break;
1629                         }
1630                 }
1631
1632         if (!has_active)
1633                 /*
1634                  * No active path in this pg, so this pg won't be used and
1635                  * the current_pg will be changed at next mapping time.
1636                  * We need to try mapping to determine it.
1637                  */
1638                 busy = 0;
1639
1640 out:
1641         spin_unlock_irqrestore(&m->lock, flags);
1642
1643         return busy;
1644 }
1645
1646 /*-----------------------------------------------------------------
1647  * Module setup
1648  *---------------------------------------------------------------*/
1649 static struct target_type multipath_target = {
1650         .name = "multipath",
1651         .version = {1, 3, 0},
1652         .module = THIS_MODULE,
1653         .ctr = multipath_ctr,
1654         .dtr = multipath_dtr,
1655         .map_rq = multipath_map,
1656         .rq_end_io = multipath_end_io,
1657         .presuspend = multipath_presuspend,
1658         .postsuspend = multipath_postsuspend,
1659         .resume = multipath_resume,
1660         .status = multipath_status,
1661         .message = multipath_message,
1662         .ioctl  = multipath_ioctl,
1663         .iterate_devices = multipath_iterate_devices,
1664         .busy = multipath_busy,
1665 };
1666
1667 static int __init dm_multipath_init(void)
1668 {
1669         int r;
1670
1671         /* allocate a slab for the dm_ios */
1672         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1673         if (!_mpio_cache)
1674                 return -ENOMEM;
1675
1676         r = dm_register_target(&multipath_target);
1677         if (r < 0) {
1678                 DMERR("register failed %d", r);
1679                 kmem_cache_destroy(_mpio_cache);
1680                 return -EINVAL;
1681         }
1682
1683         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1684         if (!kmultipathd) {
1685                 DMERR("failed to create workqueue kmpathd");
1686                 dm_unregister_target(&multipath_target);
1687                 kmem_cache_destroy(_mpio_cache);
1688                 return -ENOMEM;
1689         }
1690
1691         /*
1692          * A separate workqueue is used to handle the device handlers
1693          * to avoid overloading existing workqueue. Overloading the
1694          * old workqueue would also create a bottleneck in the
1695          * path of the storage hardware device activation.
1696          */
1697         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1698                                                   WQ_MEM_RECLAIM);
1699         if (!kmpath_handlerd) {
1700                 DMERR("failed to create workqueue kmpath_handlerd");
1701                 destroy_workqueue(kmultipathd);
1702                 dm_unregister_target(&multipath_target);
1703                 kmem_cache_destroy(_mpio_cache);
1704                 return -ENOMEM;
1705         }
1706
1707         DMINFO("version %u.%u.%u loaded",
1708                multipath_target.version[0], multipath_target.version[1],
1709                multipath_target.version[2]);
1710
1711         return r;
1712 }
1713
1714 static void __exit dm_multipath_exit(void)
1715 {
1716         destroy_workqueue(kmpath_handlerd);
1717         destroy_workqueue(kmultipathd);
1718
1719         dm_unregister_target(&multipath_target);
1720         kmem_cache_destroy(_mpio_cache);
1721 }
1722
1723 module_init(dm_multipath_init);
1724 module_exit(dm_multipath_exit);
1725
1726 MODULE_DESCRIPTION(DM_NAME " multipath target");
1727 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1728 MODULE_LICENSE("GPL");