kernel/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.h"
  11 #include "dm-path-selector.h"
  12 #include "dm-uevent.h"
  13
  14 #include <linux/blkdev.h>
  15 #include <linux/ctype.h>
  16 #include <linux/init.h>
  17 #include <linux/mempool.h>
  18 #include <linux/module.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/slab.h>
  21 #include <linux/time.h>
  22 #include <linux/workqueue.h>
  23 #include <linux/delay.h>
  24 #include <scsi/scsi_dh.h>
  25 #include <linux/atomic.h>
  26
  27 #define DM_MSG_PREFIX "multipath"
  28 #define DM_PG_INIT_DELAY_MSECS 2000
  29 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  30
  31 /* Path properties */
  32 struct pgpath {
  33         struct list_head list;
  34
  35         struct priority_group *pg;      /* Owning PG */
  36         unsigned is_active;             /* Path status */
  37         unsigned fail_count;            /* Cumulative failure count */
  38
  39         struct dm_path path;
  40         struct delayed_work activate_path;
  41 };
  42
  43 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  44
  45 /*
  46  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  47  * Each has a path selector which controls which path gets used.
  48  */
  49 struct priority_group {
  50         struct list_head list;
  51
  52         struct multipath *m;            /* Owning multipath instance */
  53         struct path_selector ps;
  54
  55         unsigned pg_num;                /* Reference number */
  56         unsigned bypassed;              /* Temporarily bypass this PG? */
  57
  58         unsigned nr_pgpaths;            /* Number of paths in PG */
  59         struct list_head pgpaths;
  60 };
  61
  62 /* Multipath context */
  63 struct multipath {
  64         struct list_head list;
  65         struct dm_target *ti;
  66
  67         const char *hw_handler_name;
  68         char *hw_handler_params;
  69
  70         spinlock_t lock;
  71
  72         unsigned nr_priority_groups;
  73         struct list_head priority_groups;
  74
  75         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  76
  77         unsigned pg_init_required;      /* pg_init needs calling? */
  78         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  79         unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
  80
  81         unsigned nr_valid_paths;        /* Total number of usable paths */
  82         struct pgpath *current_pgpath;
  83         struct priority_group *current_pg;
  84         struct priority_group *next_pg; /* Switch to this PG if set */
  85         unsigned repeat_count;          /* I/Os left before calling PS again */
  86
  87         unsigned queue_io:1;            /* Must we queue all I/O? */
  88         unsigned queue_if_no_path:1;    /* Queue I/O if last path fails? */
  89         unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
  90         unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
  91         unsigned pg_init_disabled:1;    /* pg_init is not currently allowed */
  92
  93         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  94         unsigned pg_init_count;         /* Number of times pg_init called */
  95         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  96
  97         struct work_struct trigger_event;
  98
  99         /*
 100          * We must use a mempool of dm_mpath_io structs so that we
 101          * can resubmit bios on error.
 102          */
 103         mempool_t *mpio_pool;
 104
 105         struct mutex work_mutex;
 106 };
 107
 108 /*
 109  * Context information attached to each bio we process.
 110  */
 111 struct dm_mpath_io {
 112         struct pgpath *pgpath;
 113         size_t nr_bytes;
 114 };
 115
 116 typedef int (*action_fn) (struct pgpath *pgpath);
 117
 118 static struct kmem_cache *_mpio_cache;
 119
 120 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 121 static void trigger_event(struct work_struct *work);
 122 static void activate_path(struct work_struct *work);
 123 static int __pgpath_busy(struct pgpath *pgpath);
 124
 125
 126 /*-----------------------------------------------
 127  * Allocation routines
 128  *-----------------------------------------------*/
 129
 130 static struct pgpath *alloc_pgpath(void)
 131 {
 132         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 133
 134         if (pgpath) {
 135                 pgpath->is_active = 1;
 136                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 137         }
 138
 139         return pgpath;
 140 }
 141
 142 static void free_pgpath(struct pgpath *pgpath)
 143 {
 144         kfree(pgpath);
 145 }
 146
 147 static struct priority_group *alloc_priority_group(void)
 148 {
 149         struct priority_group *pg;
 150
 151         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 152
 153         if (pg)
 154                 INIT_LIST_HEAD(&pg->pgpaths);
 155
 156         return pg;
 157 }
 158
 159 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 160 {
 161         struct pgpath *pgpath, *tmp;
 162         struct multipath *m = ti->private;
 163
 164         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 165                 list_del(&pgpath->list);
 166                 if (m->hw_handler_name)
 167                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 168                 dm_put_device(ti, pgpath->path.dev);
 169                 free_pgpath(pgpath);
 170         }
 171 }
 172
 173 static void free_priority_group(struct priority_group *pg,
 174                                 struct dm_target *ti)
 175 {
 176         struct path_selector *ps = &pg->ps;
 177
 178         if (ps->type) {
 179                 ps->type->destroy(ps);
 180                 dm_put_path_selector(ps->type);
 181         }
 182
 183         free_pgpaths(&pg->pgpaths, ti);
 184         kfree(pg);
 185 }
 186
 187 static struct multipath *alloc_multipath(struct dm_target *ti)
 188 {
 189         struct multipath *m;
 190         unsigned min_ios = dm_get_reserved_rq_based_ios();
 191
 192         m = kzalloc(sizeof(*m), GFP_KERNEL);
 193         if (m) {
 194                 INIT_LIST_HEAD(&m->priority_groups);
 195                 spin_lock_init(&m->lock);
 196                 m->queue_io = 1;
 197                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 198                 INIT_WORK(&m->trigger_event, trigger_event);
 199                 init_waitqueue_head(&m->pg_init_wait);
 200                 mutex_init(&m->work_mutex);
 201                 m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
 202                 if (!m->mpio_pool) {
 203                         kfree(m);
 204                         return NULL;
 205                 }
 206                 m->ti = ti;
 207                 ti->private = m;
 208         }
 209
 210         return m;
 211 }
 212
 213 static void free_multipath(struct multipath *m)
 214 {
 215         struct priority_group *pg, *tmp;
 216
 217         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 218                 list_del(&pg->list);
 219                 free_priority_group(pg, m->ti);
 220         }
 221
 222         kfree(m->hw_handler_name);
 223         kfree(m->hw_handler_params);
 224         mempool_destroy(m->mpio_pool);
 225         kfree(m);
 226 }
 227
 228 static int set_mapinfo(struct multipath *m, union map_info *info)
 229 {
 230         struct dm_mpath_io *mpio;
 231
 232         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 233         if (!mpio)
 234                 return -ENOMEM;
 235
 236         memset(mpio, 0, sizeof(*mpio));
 237         info->ptr = mpio;
 238
 239         return 0;
 240 }
 241
 242 static void clear_mapinfo(struct multipath *m, union map_info *info)
 243 {
 244         struct dm_mpath_io *mpio = info->ptr;
 245
 246         info->ptr = NULL;
 247         mempool_free(mpio, m->mpio_pool);
 248 }
 249
 250 /*-----------------------------------------------
 251  * Path selection
 252  *-----------------------------------------------*/
 253
 254 static int __pg_init_all_paths(struct multipath *m)
 255 {
 256         struct pgpath *pgpath;
 257         unsigned long pg_init_delay = 0;
 258
 259         if (m->pg_init_in_progress || m->pg_init_disabled)
 260                 return 0;
 261
 262         m->pg_init_count++;
 263         m->pg_init_required = 0;
 264
 265         /* Check here to reset pg_init_required */
 266         if (!m->current_pg)
 267                 return 0;
 268
 269         if (m->pg_init_delay_retry)
 270                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 271                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 272         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 273                 /* Skip failed paths */
 274                 if (!pgpath->is_active)
 275                         continue;
 276                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 277                                        pg_init_delay))
 278                         m->pg_init_in_progress++;
 279         }
 280         return m->pg_init_in_progress;
 281 }
 282
 283 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 284 {
 285         m->current_pg = pgpath->pg;
 286
 287         /* Must we initialise the PG first, and queue I/O till it's ready? */
 288         if (m->hw_handler_name) {
 289                 m->pg_init_required = 1;
 290                 m->queue_io = 1;
 291         } else {
 292                 m->pg_init_required = 0;
 293                 m->queue_io = 0;
 294         }
 295
 296         m->pg_init_count = 0;
 297 }
 298
 299 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 300                                size_t nr_bytes)
 301 {
 302         struct dm_path *path;
 303
 304         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 305         if (!path)
 306                 return -ENXIO;
 307
 308         m->current_pgpath = path_to_pgpath(path);
 309
 310         if (m->current_pg != pg)
 311                 __switch_pg(m, m->current_pgpath);
 312
 313         return 0;
 314 }
 315
 316 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 317 {
 318         struct priority_group *pg;
 319         unsigned bypassed = 1;
 320
 321         if (!m->nr_valid_paths) {
 322                 m->queue_io = 0;
 323                 goto failed;
 324         }
 325
 326         /* Were we instructed to switch PG? */
 327         if (m->next_pg) {
 328                 pg = m->next_pg;
 329                 m->next_pg = NULL;
 330                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 331                         return;
 332         }
 333
 334         /* Don't change PG until it has no remaining paths */
 335         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 336                 return;
 337
 338         /*
 339          * Loop through priority groups until we find a valid path.
 340          * First time we skip PGs marked 'bypassed'.
 341          * Second time we only try the ones we skipped, but set
 342          * pg_init_delay_retry so we do not hammer controllers.
 343          */
 344         do {
 345                 list_for_each_entry(pg, &m->priority_groups, list) {
 346                         if (pg->bypassed == bypassed)
 347                                 continue;
 348                         if (!__choose_path_in_pg(m, pg, nr_bytes)) {
 349                                 if (!bypassed)
 350                                         m->pg_init_delay_retry = 1;
 351                                 return;
 352                         }
 353                 }
 354         } while (bypassed--);
 355
 356 failed:
 357         m->current_pgpath = NULL;
 358         m->current_pg = NULL;
 359 }
 360
 361 /*
 362  * Check whether bios must be queued in the device-mapper core rather
 363  * than here in the target.
 364  *
 365  * m->lock must be held on entry.
 366  *
 367  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 368  * same value then we are not between multipath_presuspend()
 369  * and multipath_resume() calls and we have no need to check
 370  * for the DMF_NOFLUSH_SUSPENDING flag.
 371  */
 372 static int __must_push_back(struct multipath *m)
 373 {
 374         return (m->queue_if_no_path ||
 375                 (m->queue_if_no_path != m->saved_queue_if_no_path &&
 376                  dm_noflush_suspending(m->ti)));
 377 }
 378
 379 /*
 380  * Map cloned requests
 381  */
 382 static int __multipath_map(struct dm_target *ti, struct request *clone,
 383                            union map_info *map_context,
 384                            struct request *rq, struct request **__clone)
 385 {
 386         struct multipath *m = (struct multipath *) ti->private;
 387         int r = DM_MAPIO_REQUEUE;
 388         size_t nr_bytes = clone ? blk_rq_bytes(clone) : blk_rq_bytes(rq);
 389         struct pgpath *pgpath;
 390         struct block_device *bdev;
 391         struct dm_mpath_io *mpio;
 392
 393         spin_lock_irq(&m->lock);
 394
 395         /* Do we need to select a new pgpath? */
 396         if (!m->current_pgpath ||
 397             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 398                 __choose_pgpath(m, nr_bytes);
 399
 400         pgpath = m->current_pgpath;
 401
 402         if (!pgpath) {
 403                 if (!__must_push_back(m))
 404                         r = -EIO;       /* Failed */
 405                 goto out_unlock;
 406         } else if (m->queue_io || m->pg_init_required) {
 407                 __pg_init_all_paths(m);
 408                 goto out_unlock;
 409         }
 410
 411         if (set_mapinfo(m, map_context) < 0)
 412                 /* ENOMEM, requeue */
 413                 goto out_unlock;
 414
 415         mpio = map_context->ptr;
 416         mpio->pgpath = pgpath;
 417         mpio->nr_bytes = nr_bytes;
 418
 419         bdev = pgpath->path.dev->bdev;
 420
 421         spin_unlock_irq(&m->lock);
 422
 423         if (clone) {
 424                 /* Old request-based interface: allocated clone is passed in */
 425                 clone->q = bdev_get_queue(bdev);
 426                 clone->rq_disk = bdev->bd_disk;
 427                 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 428         } else {
 429                 /* blk-mq request-based interface */
 430                 *__clone = blk_get_request(bdev_get_queue(bdev),
 431                                            rq_data_dir(rq), GFP_ATOMIC);
 432                 if (IS_ERR(*__clone)) {
 433                         /* ENOMEM, requeue */
 434                         clear_mapinfo(m, map_context);
 435                         return r;
 436                 }
 437                 (*__clone)->bio = (*__clone)->biotail = NULL;
 438                 (*__clone)->rq_disk = bdev->bd_disk;
 439                 (*__clone)->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 440         }
 441
 442         if (pgpath->pg->ps.type->start_io)
 443                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 444                                               &pgpath->path,
 445                                               nr_bytes);
 446         return DM_MAPIO_REMAPPED;
 447
 448 out_unlock:
 449         spin_unlock_irq(&m->lock);
 450
 451         return r;
 452 }
 453
 454 static int multipath_map(struct dm_target *ti, struct request *clone,
 455                          union map_info *map_context)
 456 {
 457         return __multipath_map(ti, clone, map_context, NULL, NULL);
 458 }
 459
 460 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 461                                    union map_info *map_context,
 462                                    struct request **clone)
 463 {
 464         return __multipath_map(ti, NULL, map_context, rq, clone);
 465 }
 466
 467 static void multipath_release_clone(struct request *clone)
 468 {
 469         blk_put_request(clone);
 470 }
 471
 472 /*
 473  * If we run out of usable paths, should we queue I/O or error it?
 474  */
 475 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 476                             unsigned save_old_value)
 477 {
 478         unsigned long flags;
 479
 480         spin_lock_irqsave(&m->lock, flags);
 481
 482         if (save_old_value)
 483                 m->saved_queue_if_no_path = m->queue_if_no_path;
 484         else
 485                 m->saved_queue_if_no_path = queue_if_no_path;
 486         m->queue_if_no_path = queue_if_no_path;
 487         spin_unlock_irqrestore(&m->lock, flags);
 488
 489         if (!queue_if_no_path)
 490                 dm_table_run_md_queue_async(m->ti->table);
 491
 492         return 0;
 493 }
 494
 495 /*
 496  * An event is triggered whenever a path is taken out of use.
 497  * Includes path failure and PG bypass.
 498  */
 499 static void trigger_event(struct work_struct *work)
 500 {
 501         struct multipath *m =
 502                 container_of(work, struct multipath, trigger_event);
 503
 504         dm_table_event(m->ti->table);
 505 }
 506
 507 /*-----------------------------------------------------------------
 508  * Constructor/argument parsing:
 509  * <#multipath feature args> [<arg>]*
 510  * <#hw_handler args> [hw_handler [<arg>]*]
 511  * <#priority groups>
 512  * <initial priority group>
 513  *     [<selector> <#selector args> [<arg>]*
 514  *      <#paths> <#per-path selector args>
 515  *         [<path> [<arg>]* ]+ ]+
 516  *---------------------------------------------------------------*/
 517 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 518                                struct dm_target *ti)
 519 {
 520         int r;
 521         struct path_selector_type *pst;
 522         unsigned ps_argc;
 523
 524         static struct dm_arg _args[] = {
 525                 {0, 1024, "invalid number of path selector args"},
 526         };
 527
 528         pst = dm_get_path_selector(dm_shift_arg(as));
 529         if (!pst) {
 530                 ti->error = "unknown path selector type";
 531                 return -EINVAL;
 532         }
 533
 534         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 535         if (r) {
 536                 dm_put_path_selector(pst);
 537                 return -EINVAL;
 538         }
 539
 540         r = pst->create(&pg->ps, ps_argc, as->argv);
 541         if (r) {
 542                 dm_put_path_selector(pst);
 543                 ti->error = "path selector constructor failed";
 544                 return r;
 545         }
 546
 547         pg->ps.type = pst;
 548         dm_consume_args(as, ps_argc);
 549
 550         return 0;
 551 }
 552
 553 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 554                                struct dm_target *ti)
 555 {
 556         int r;
 557         struct pgpath *p;
 558         struct multipath *m = ti->private;
 559         struct request_queue *q = NULL;
 560         const char *attached_handler_name;
 561
 562         /* we need at least a path arg */
 563         if (as->argc < 1) {
 564                 ti->error = "no device given";
 565                 return ERR_PTR(-EINVAL);
 566         }
 567
 568         p = alloc_pgpath();
 569         if (!p)
 570                 return ERR_PTR(-ENOMEM);
 571
 572         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 573                           &p->path.dev);
 574         if (r) {
 575                 ti->error = "error getting device";
 576                 goto bad;
 577         }
 578
 579         if (m->retain_attached_hw_handler || m->hw_handler_name)
 580                 q = bdev_get_queue(p->path.dev->bdev);
 581
 582         if (m->retain_attached_hw_handler) {
 583                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 584                 if (attached_handler_name) {
 585                         /*
 586                          * Reset hw_handler_name to match the attached handler
 587                          * and clear any hw_handler_params associated with the
 588                          * ignored handler.
 589                          *
 590                          * NB. This modifies the table line to show the actual
 591                          * handler instead of the original table passed in.
 592                          */
 593                         kfree(m->hw_handler_name);
 594                         m->hw_handler_name = attached_handler_name;
 595
 596                         kfree(m->hw_handler_params);
 597                         m->hw_handler_params = NULL;
 598                 }
 599         }
 600
 601         if (m->hw_handler_name) {
 602                 /*
 603                  * Increments scsi_dh reference, even when using an
 604                  * already-attached handler.
 605                  */
 606                 r = scsi_dh_attach(q, m->hw_handler_name);
 607                 if (r == -EBUSY) {
 608                         /*
 609                          * Already attached to different hw_handler:
 610                          * try to reattach with correct one.
 611                          */
 612                         scsi_dh_detach(q);
 613                         r = scsi_dh_attach(q, m->hw_handler_name);
 614                 }
 615
 616                 if (r < 0) {
 617                         ti->error = "error attaching hardware handler";
 618                         dm_put_device(ti, p->path.dev);
 619                         goto bad;
 620                 }
 621
 622                 if (m->hw_handler_params) {
 623                         r = scsi_dh_set_params(q, m->hw_handler_params);
 624                         if (r < 0) {
 625                                 ti->error = "unable to set hardware "
 626                                                         "handler parameters";
 627                                 scsi_dh_detach(q);
 628                                 dm_put_device(ti, p->path.dev);
 629                                 goto bad;
 630                         }
 631                 }
 632         }
 633
 634         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 635         if (r) {
 636                 dm_put_device(ti, p->path.dev);
 637                 goto bad;
 638         }
 639
 640         return p;
 641
 642  bad:
 643         free_pgpath(p);
 644         return ERR_PTR(r);
 645 }
 646
 647 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 648                                                    struct multipath *m)
 649 {
 650         static struct dm_arg _args[] = {
 651                 {1, 1024, "invalid number of paths"},
 652                 {0, 1024, "invalid number of selector args"}
 653         };
 654
 655         int r;
 656         unsigned i, nr_selector_args, nr_args;
 657         struct priority_group *pg;
 658         struct dm_target *ti = m->ti;
 659
 660         if (as->argc < 2) {
 661                 as->argc = 0;
 662                 ti->error = "not enough priority group arguments";
 663                 return ERR_PTR(-EINVAL);
 664         }
 665
 666         pg = alloc_priority_group();
 667         if (!pg) {
 668                 ti->error = "couldn't allocate priority group";
 669                 return ERR_PTR(-ENOMEM);
 670         }
 671         pg->m = m;
 672
 673         r = parse_path_selector(as, pg, ti);
 674         if (r)
 675                 goto bad;
 676
 677         /*
 678          * read the paths
 679          */
 680         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 681         if (r)
 682                 goto bad;
 683
 684         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 685         if (r)
 686                 goto bad;
 687
 688         nr_args = 1 + nr_selector_args;
 689         for (i = 0; i < pg->nr_pgpaths; i++) {
 690                 struct pgpath *pgpath;
 691                 struct dm_arg_set path_args;
 692
 693                 if (as->argc < nr_args) {
 694                         ti->error = "not enough path parameters";
 695                         r = -EINVAL;
 696                         goto bad;
 697                 }
 698
 699                 path_args.argc = nr_args;
 700                 path_args.argv = as->argv;
 701
 702                 pgpath = parse_path(&path_args, &pg->ps, ti);
 703                 if (IS_ERR(pgpath)) {
 704                         r = PTR_ERR(pgpath);
 705                         goto bad;
 706                 }
 707
 708                 pgpath->pg = pg;
 709                 list_add_tail(&pgpath->list, &pg->pgpaths);
 710                 dm_consume_args(as, nr_args);
 711         }
 712
 713         return pg;
 714
 715  bad:
 716         free_priority_group(pg, ti);
 717         return ERR_PTR(r);
 718 }
 719
 720 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 721 {
 722         unsigned hw_argc;
 723         int ret;
 724         struct dm_target *ti = m->ti;
 725
 726         static struct dm_arg _args[] = {
 727                 {0, 1024, "invalid number of hardware handler args"},
 728         };
 729
 730         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 731                 return -EINVAL;
 732
 733         if (!hw_argc)
 734                 return 0;
 735
 736         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 737         if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
 738                                      "scsi_dh_%s", m->hw_handler_name)) {
 739                 ti->error = "unknown hardware handler type";
 740                 ret = -EINVAL;
 741                 goto fail;
 742         }
 743
 744         if (hw_argc > 1) {
 745                 char *p;
 746                 int i, j, len = 4;
 747
 748                 for (i = 0; i <= hw_argc - 2; i++)
 749                         len += strlen(as->argv[i]) + 1;
 750                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 751                 if (!p) {
 752                         ti->error = "memory allocation failed";
 753                         ret = -ENOMEM;
 754                         goto fail;
 755                 }
 756                 j = sprintf(p, "%d", hw_argc - 1);
 757                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 758                         j = sprintf(p, "%s", as->argv[i]);
 759         }
 760         dm_consume_args(as, hw_argc - 1);
 761
 762         return 0;
 763 fail:
 764         kfree(m->hw_handler_name);
 765         m->hw_handler_name = NULL;
 766         return ret;
 767 }
 768
 769 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 770 {
 771         int r;
 772         unsigned argc;
 773         struct dm_target *ti = m->ti;
 774         const char *arg_name;
 775
 776         static struct dm_arg _args[] = {
 777                 {0, 6, "invalid number of feature args"},
 778                 {1, 50, "pg_init_retries must be between 1 and 50"},
 779                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 780         };
 781
 782         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 783         if (r)
 784                 return -EINVAL;
 785
 786         if (!argc)
 787                 return 0;
 788
 789         do {
 790                 arg_name = dm_shift_arg(as);
 791                 argc--;
 792
 793                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 794                         r = queue_if_no_path(m, 1, 0);
 795                         continue;
 796                 }
 797
 798                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
 799                         m->retain_attached_hw_handler = 1;
 800                         continue;
 801                 }
 802
 803                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 804                     (argc >= 1)) {
 805                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 806                         argc--;
 807                         continue;
 808                 }
 809
 810                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 811                     (argc >= 1)) {
 812                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 813                         argc--;
 814                         continue;
 815                 }
 816
 817                 ti->error = "Unrecognised multipath feature request";
 818                 r = -EINVAL;
 819         } while (argc && !r);
 820
 821         return r;
 822 }
 823
 824 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 825                          char **argv)
 826 {
 827         /* target arguments */
 828         static struct dm_arg _args[] = {
 829                 {0, 1024, "invalid number of priority groups"},
 830                 {0, 1024, "invalid initial priority group number"},
 831         };
 832
 833         int r;
 834         struct multipath *m;
 835         struct dm_arg_set as;
 836         unsigned pg_count = 0;
 837         unsigned next_pg_num;
 838
 839         as.argc = argc;
 840         as.argv = argv;
 841
 842         m = alloc_multipath(ti);
 843         if (!m) {
 844                 ti->error = "can't allocate multipath";
 845                 return -EINVAL;
 846         }
 847
 848         r = parse_features(&as, m);
 849         if (r)
 850                 goto bad;
 851
 852         r = parse_hw_handler(&as, m);
 853         if (r)
 854                 goto bad;
 855
 856         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 857         if (r)
 858                 goto bad;
 859
 860         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 861         if (r)
 862                 goto bad;
 863
 864         if ((!m->nr_priority_groups && next_pg_num) ||
 865             (m->nr_priority_groups && !next_pg_num)) {
 866                 ti->error = "invalid initial priority group";
 867                 r = -EINVAL;
 868                 goto bad;
 869         }
 870
 871         /* parse the priority groups */
 872         while (as.argc) {
 873                 struct priority_group *pg;
 874
 875                 pg = parse_priority_group(&as, m);
 876                 if (IS_ERR(pg)) {
 877                         r = PTR_ERR(pg);
 878                         goto bad;
 879                 }
 880
 881                 m->nr_valid_paths += pg->nr_pgpaths;
 882                 list_add_tail(&pg->list, &m->priority_groups);
 883                 pg_count++;
 884                 pg->pg_num = pg_count;
 885                 if (!--next_pg_num)
 886                         m->next_pg = pg;
 887         }
 888
 889         if (pg_count != m->nr_priority_groups) {
 890                 ti->error = "priority group count mismatch";
 891                 r = -EINVAL;
 892                 goto bad;
 893         }
 894
 895         ti->num_flush_bios = 1;
 896         ti->num_discard_bios = 1;
 897         ti->num_write_same_bios = 1;
 898
 899         return 0;
 900
 901  bad:
 902         free_multipath(m);
 903         return r;
 904 }
 905
 906 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 907 {
 908         DECLARE_WAITQUEUE(wait, current);
 909         unsigned long flags;
 910
 911         add_wait_queue(&m->pg_init_wait, &wait);
 912
 913         while (1) {
 914                 set_current_state(TASK_UNINTERRUPTIBLE);
 915
 916                 spin_lock_irqsave(&m->lock, flags);
 917                 if (!m->pg_init_in_progress) {
 918                         spin_unlock_irqrestore(&m->lock, flags);
 919                         break;
 920                 }
 921                 spin_unlock_irqrestore(&m->lock, flags);
 922
 923                 io_schedule();
 924         }
 925         set_current_state(TASK_RUNNING);
 926
 927         remove_wait_queue(&m->pg_init_wait, &wait);
 928 }
 929
 930 static void flush_multipath_work(struct multipath *m)
 931 {
 932         unsigned long flags;
 933
 934         spin_lock_irqsave(&m->lock, flags);
 935         m->pg_init_disabled = 1;
 936         spin_unlock_irqrestore(&m->lock, flags);
 937
 938         flush_workqueue(kmpath_handlerd);
 939         multipath_wait_for_pg_init_completion(m);
 940         flush_workqueue(kmultipathd);
 941         flush_work(&m->trigger_event);
 942
 943         spin_lock_irqsave(&m->lock, flags);
 944         m->pg_init_disabled = 0;
 945         spin_unlock_irqrestore(&m->lock, flags);
 946 }
 947
 948 static void multipath_dtr(struct dm_target *ti)
 949 {
 950         struct multipath *m = ti->private;
 951
 952         flush_multipath_work(m);
 953         free_multipath(m);
 954 }
 955
 956 /*
 957  * Take a path out of use.
 958  */
 959 static int fail_path(struct pgpath *pgpath)
 960 {
 961         unsigned long flags;
 962         struct multipath *m = pgpath->pg->m;
 963
 964         spin_lock_irqsave(&m->lock, flags);
 965
 966         if (!pgpath->is_active)
 967                 goto out;
 968
 969         DMWARN("Failing path %s.", pgpath->path.dev->name);
 970
 971         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 972         pgpath->is_active = 0;
 973         pgpath->fail_count++;
 974
 975         m->nr_valid_paths--;
 976
 977         if (pgpath == m->current_pgpath)
 978                 m->current_pgpath = NULL;
 979
 980         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 981                       pgpath->path.dev->name, m->nr_valid_paths);
 982
 983         schedule_work(&m->trigger_event);
 984
 985 out:
 986         spin_unlock_irqrestore(&m->lock, flags);
 987
 988         return 0;
 989 }
 990
 991 /*
 992  * Reinstate a previously-failed path
 993  */
 994 static int reinstate_path(struct pgpath *pgpath)
 995 {
 996         int r = 0, run_queue = 0;
 997         unsigned long flags;
 998         struct multipath *m = pgpath->pg->m;
 999
1000         spin_lock_irqsave(&m->lock, flags);
1001
1002         if (pgpath->is_active)
1003                 goto out;
1004
1005         if (!pgpath->pg->ps.type->reinstate_path) {
1006                 DMWARN("Reinstate path not supported by path selector %s",
1007                        pgpath->pg->ps.type->name);
1008                 r = -EINVAL;
1009                 goto out;
1010         }
1011
1012         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1013         if (r)
1014                 goto out;
1015
1016         pgpath->is_active = 1;
1017
1018         if (!m->nr_valid_paths++) {
1019                 m->current_pgpath = NULL;
1020                 run_queue = 1;
1021         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1022                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1023                         m->pg_init_in_progress++;
1024         }
1025
1026         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1027                       pgpath->path.dev->name, m->nr_valid_paths);
1028
1029         schedule_work(&m->trigger_event);
1030
1031 out:
1032         spin_unlock_irqrestore(&m->lock, flags);
1033         if (run_queue)
1034                 dm_table_run_md_queue_async(m->ti->table);
1035
1036         return r;
1037 }
1038
1039 /*
1040  * Fail or reinstate all paths that match the provided struct dm_dev.
1041  */
1042 static int action_dev(struct multipath *m, struct dm_dev *dev,
1043                       action_fn action)
1044 {
1045         int r = -EINVAL;
1046         struct pgpath *pgpath;
1047         struct priority_group *pg;
1048
1049         list_for_each_entry(pg, &m->priority_groups, list) {
1050                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1051                         if (pgpath->path.dev == dev)
1052                                 r = action(pgpath);
1053                 }
1054         }
1055
1056         return r;
1057 }
1058
1059 /*
1060  * Temporarily try to avoid having to use the specified PG
1061  */
1062 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1063                       int bypassed)
1064 {
1065         unsigned long flags;
1066
1067         spin_lock_irqsave(&m->lock, flags);
1068
1069         pg->bypassed = bypassed;
1070         m->current_pgpath = NULL;
1071         m->current_pg = NULL;
1072
1073         spin_unlock_irqrestore(&m->lock, flags);
1074
1075         schedule_work(&m->trigger_event);
1076 }
1077
1078 /*
1079  * Switch to using the specified PG from the next I/O that gets mapped
1080  */
1081 static int switch_pg_num(struct multipath *m, const char *pgstr)
1082 {
1083         struct priority_group *pg;
1084         unsigned pgnum;
1085         unsigned long flags;
1086         char dummy;
1087
1088         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1089             (pgnum > m->nr_priority_groups)) {
1090                 DMWARN("invalid PG number supplied to switch_pg_num");
1091                 return -EINVAL;
1092         }
1093
1094         spin_lock_irqsave(&m->lock, flags);
1095         list_for_each_entry(pg, &m->priority_groups, list) {
1096                 pg->bypassed = 0;
1097                 if (--pgnum)
1098                         continue;
1099
1100                 m->current_pgpath = NULL;
1101                 m->current_pg = NULL;
1102                 m->next_pg = pg;
1103         }
1104         spin_unlock_irqrestore(&m->lock, flags);
1105
1106         schedule_work(&m->trigger_event);
1107         return 0;
1108 }
1109
1110 /*
1111  * Set/clear bypassed status of a PG.
1112  * PGs are numbered upwards from 1 in the order they were declared.
1113  */
1114 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1115 {
1116         struct priority_group *pg;
1117         unsigned pgnum;
1118         char dummy;
1119
1120         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1121             (pgnum > m->nr_priority_groups)) {
1122                 DMWARN("invalid PG number supplied to bypass_pg");
1123                 return -EINVAL;
1124         }
1125
1126         list_for_each_entry(pg, &m->priority_groups, list) {
1127                 if (!--pgnum)
1128                         break;
1129         }
1130
1131         bypass_pg(m, pg, bypassed);
1132         return 0;
1133 }
1134
1135 /*
1136  * Should we retry pg_init immediately?
1137  */
1138 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1139 {
1140         unsigned long flags;
1141         int limit_reached = 0;
1142
1143         spin_lock_irqsave(&m->lock, flags);
1144
1145         if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1146                 m->pg_init_required = 1;
1147         else
1148                 limit_reached = 1;
1149
1150         spin_unlock_irqrestore(&m->lock, flags);
1151
1152         return limit_reached;
1153 }
1154
1155 static void pg_init_done(void *data, int errors)
1156 {
1157         struct pgpath *pgpath = data;
1158         struct priority_group *pg = pgpath->pg;
1159         struct multipath *m = pg->m;
1160         unsigned long flags;
1161         unsigned delay_retry = 0;
1162
1163         /* device or driver problems */
1164         switch (errors) {
1165         case SCSI_DH_OK:
1166                 break;
1167         case SCSI_DH_NOSYS:
1168                 if (!m->hw_handler_name) {
1169                         errors = 0;
1170                         break;
1171                 }
1172                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1173                       "Error %d.", m->hw_handler_name, errors);
1174                 /*
1175                  * Fail path for now, so we do not ping pong
1176                  */
1177                 fail_path(pgpath);
1178                 break;
1179         case SCSI_DH_DEV_TEMP_BUSY:
1180                 /*
1181                  * Probably doing something like FW upgrade on the
1182                  * controller so try the other pg.
1183                  */
1184                 bypass_pg(m, pg, 1);
1185                 break;
1186         case SCSI_DH_RETRY:
1187                 /* Wait before retrying. */
1188                 delay_retry = 1;
1189         case SCSI_DH_IMM_RETRY:
1190         case SCSI_DH_RES_TEMP_UNAVAIL:
1191                 if (pg_init_limit_reached(m, pgpath))
1192                         fail_path(pgpath);
1193                 errors = 0;
1194                 break;
1195         default:
1196                 /*
1197                  * We probably do not want to fail the path for a device
1198                  * error, but this is what the old dm did. In future
1199                  * patches we can do more advanced handling.
1200                  */
1201                 fail_path(pgpath);
1202         }
1203
1204         spin_lock_irqsave(&m->lock, flags);
1205         if (errors) {
1206                 if (pgpath == m->current_pgpath) {
1207                         DMERR("Could not failover device. Error %d.", errors);
1208                         m->current_pgpath = NULL;
1209                         m->current_pg = NULL;
1210                 }
1211         } else if (!m->pg_init_required)
1212                 pg->bypassed = 0;
1213
1214         if (--m->pg_init_in_progress)
1215                 /* Activations of other paths are still on going */
1216                 goto out;
1217
1218         if (m->pg_init_required) {
1219                 m->pg_init_delay_retry = delay_retry;
1220                 if (__pg_init_all_paths(m))
1221                         goto out;
1222         }
1223         m->queue_io = 0;
1224
1225         /*
1226          * Wake up any thread waiting to suspend.
1227          */
1228         wake_up(&m->pg_init_wait);
1229
1230 out:
1231         spin_unlock_irqrestore(&m->lock, flags);
1232 }
1233
1234 static void activate_path(struct work_struct *work)
1235 {
1236         struct pgpath *pgpath =
1237                 container_of(work, struct pgpath, activate_path.work);
1238
1239         if (pgpath->is_active)
1240                 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1241                                  pg_init_done, pgpath);
1242         else
1243                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1244 }
1245
1246 static int noretry_error(int error)
1247 {
1248         switch (error) {
1249         case -EOPNOTSUPP:
1250         case -EREMOTEIO:
1251         case -EILSEQ:
1252         case -ENODATA:
1253         case -ENOSPC:
1254                 return 1;
1255         }
1256
1257         /* Anything else could be a path failure, so should be retried */
1258         return 0;
1259 }
1260
1261 /*
1262  * end_io handling
1263  */
1264 static int do_end_io(struct multipath *m, struct request *clone,
1265                      int error, struct dm_mpath_io *mpio)
1266 {
1267         /*
1268          * We don't queue any clone request inside the multipath target
1269          * during end I/O handling, since those clone requests don't have
1270          * bio clones.  If we queue them inside the multipath target,
1271          * we need to make bio clones, that requires memory allocation.
1272          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1273          *  don't have bio clones.)
1274          * Instead of queueing the clone request here, we queue the original
1275          * request into dm core, which will remake a clone request and
1276          * clone bios for it and resubmit it later.
1277          */
1278         int r = DM_ENDIO_REQUEUE;
1279         unsigned long flags;
1280
1281         if (!error && !clone->errors)
1282                 return 0;       /* I/O complete */
1283
1284         if (noretry_error(error))
1285                 return error;
1286
1287         if (mpio->pgpath)
1288                 fail_path(mpio->pgpath);
1289
1290         spin_lock_irqsave(&m->lock, flags);
1291         if (!m->nr_valid_paths) {
1292                 if (!m->queue_if_no_path) {
1293                         if (!__must_push_back(m))
1294                                 r = -EIO;
1295                 } else {
1296                         if (error == -EBADE)
1297                                 r = error;
1298                 }
1299         }
1300         spin_unlock_irqrestore(&m->lock, flags);
1301
1302         return r;
1303 }
1304
1305 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1306                             int error, union map_info *map_context)
1307 {
1308         struct multipath *m = ti->private;
1309         struct dm_mpath_io *mpio = map_context->ptr;
1310         struct pgpath *pgpath;
1311         struct path_selector *ps;
1312         int r;
1313
1314         BUG_ON(!mpio);
1315
1316         r  = do_end_io(m, clone, error, mpio);
1317         pgpath = mpio->pgpath;
1318         if (pgpath) {
1319                 ps = &pgpath->pg->ps;
1320                 if (ps->type->end_io)
1321                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1322         }
1323         clear_mapinfo(m, map_context);
1324
1325         return r;
1326 }
1327
1328 /*
1329  * Suspend can't complete until all the I/O is processed so if
1330  * the last path fails we must error any remaining I/O.
1331  * Note that if the freeze_bdev fails while suspending, the
1332  * queue_if_no_path state is lost - userspace should reset it.
1333  */
1334 static void multipath_presuspend(struct dm_target *ti)
1335 {
1336         struct multipath *m = (struct multipath *) ti->private;
1337
1338         queue_if_no_path(m, 0, 1);
1339 }
1340
1341 static void multipath_postsuspend(struct dm_target *ti)
1342 {
1343         struct multipath *m = ti->private;
1344
1345         mutex_lock(&m->work_mutex);
1346         flush_multipath_work(m);
1347         mutex_unlock(&m->work_mutex);
1348 }
1349
1350 /*
1351  * Restore the queue_if_no_path setting.
1352  */
1353 static void multipath_resume(struct dm_target *ti)
1354 {
1355         struct multipath *m = (struct multipath *) ti->private;
1356         unsigned long flags;
1357
1358         spin_lock_irqsave(&m->lock, flags);
1359         m->queue_if_no_path = m->saved_queue_if_no_path;
1360         spin_unlock_irqrestore(&m->lock, flags);
1361 }
1362
1363 /*
1364  * Info output has the following format:
1365  * num_multipath_feature_args [multipath_feature_args]*
1366  * num_handler_status_args [handler_status_args]*
1367  * num_groups init_group_number
1368  *            [A|D|E num_ps_status_args [ps_status_args]*
1369  *             num_paths num_selector_args
1370  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1371  *
1372  * Table output has the following format (identical to the constructor string):
1373  * num_feature_args [features_args]*
1374  * num_handler_args hw_handler [hw_handler_args]*
1375  * num_groups init_group_number
1376  *     [priority selector-name num_ps_args [ps_args]*
1377  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1378  */
1379 static void multipath_status(struct dm_target *ti, status_type_t type,
1380                              unsigned status_flags, char *result, unsigned maxlen)
1381 {
1382         int sz = 0;
1383         unsigned long flags;
1384         struct multipath *m = (struct multipath *) ti->private;
1385         struct priority_group *pg;
1386         struct pgpath *p;
1387         unsigned pg_num;
1388         char state;
1389
1390         spin_lock_irqsave(&m->lock, flags);
1391
1392         /* Features */
1393         if (type == STATUSTYPE_INFO)
1394                 DMEMIT("2 %u %u ", m->queue_io, m->pg_init_count);
1395         else {
1396                 DMEMIT("%u ", m->queue_if_no_path +
1397                               (m->pg_init_retries > 0) * 2 +
1398                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1399                               m->retain_attached_hw_handler);
1400                 if (m->queue_if_no_path)
1401                         DMEMIT("queue_if_no_path ");
1402                 if (m->pg_init_retries)
1403                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1404                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1405                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1406                 if (m->retain_attached_hw_handler)
1407                         DMEMIT("retain_attached_hw_handler ");
1408         }
1409
1410         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1411                 DMEMIT("0 ");
1412         else
1413                 DMEMIT("1 %s ", m->hw_handler_name);
1414
1415         DMEMIT("%u ", m->nr_priority_groups);
1416
1417         if (m->next_pg)
1418                 pg_num = m->next_pg->pg_num;
1419         else if (m->current_pg)
1420                 pg_num = m->current_pg->pg_num;
1421         else
1422                 pg_num = (m->nr_priority_groups ? 1 : 0);
1423
1424         DMEMIT("%u ", pg_num);
1425
1426         switch (type) {
1427         case STATUSTYPE_INFO:
1428                 list_for_each_entry(pg, &m->priority_groups, list) {
1429                         if (pg->bypassed)
1430                                 state = 'D';    /* Disabled */
1431                         else if (pg == m->current_pg)
1432                                 state = 'A';    /* Currently Active */
1433                         else
1434                                 state = 'E';    /* Enabled */
1435
1436                         DMEMIT("%c ", state);
1437
1438                         if (pg->ps.type->status)
1439                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1440                                                           result + sz,
1441                                                           maxlen - sz);
1442                         else
1443                                 DMEMIT("0 ");
1444
1445                         DMEMIT("%u %u ", pg->nr_pgpaths,
1446                                pg->ps.type->info_args);
1447
1448                         list_for_each_entry(p, &pg->pgpaths, list) {
1449                                 DMEMIT("%s %s %u ", p->path.dev->name,
1450                                        p->is_active ? "A" : "F",
1451                                        p->fail_count);
1452                                 if (pg->ps.type->status)
1453                                         sz += pg->ps.type->status(&pg->ps,
1454                                               &p->path, type, result + sz,
1455                                               maxlen - sz);
1456                         }
1457                 }
1458                 break;
1459
1460         case STATUSTYPE_TABLE:
1461                 list_for_each_entry(pg, &m->priority_groups, list) {
1462                         DMEMIT("%s ", pg->ps.type->name);
1463
1464                         if (pg->ps.type->status)
1465                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1466                                                           result + sz,
1467                                                           maxlen - sz);
1468                         else
1469                                 DMEMIT("0 ");
1470
1471                         DMEMIT("%u %u ", pg->nr_pgpaths,
1472                                pg->ps.type->table_args);
1473
1474                         list_for_each_entry(p, &pg->pgpaths, list) {
1475                                 DMEMIT("%s ", p->path.dev->name);
1476                                 if (pg->ps.type->status)
1477                                         sz += pg->ps.type->status(&pg->ps,
1478                                               &p->path, type, result + sz,
1479                                               maxlen - sz);
1480                         }
1481                 }
1482                 break;
1483         }
1484
1485         spin_unlock_irqrestore(&m->lock, flags);
1486 }
1487
1488 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1489 {
1490         int r = -EINVAL;
1491         struct dm_dev *dev;
1492         struct multipath *m = (struct multipath *) ti->private;
1493         action_fn action;
1494
1495         mutex_lock(&m->work_mutex);
1496
1497         if (dm_suspended(ti)) {
1498                 r = -EBUSY;
1499                 goto out;
1500         }
1501
1502         if (argc == 1) {
1503                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1504                         r = queue_if_no_path(m, 1, 0);
1505                         goto out;
1506                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1507                         r = queue_if_no_path(m, 0, 0);
1508                         goto out;
1509                 }
1510         }
1511
1512         if (argc != 2) {
1513                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1514                 goto out;
1515         }
1516
1517         if (!strcasecmp(argv[0], "disable_group")) {
1518                 r = bypass_pg_num(m, argv[1], 1);
1519                 goto out;
1520         } else if (!strcasecmp(argv[0], "enable_group")) {
1521                 r = bypass_pg_num(m, argv[1], 0);
1522                 goto out;
1523         } else if (!strcasecmp(argv[0], "switch_group")) {
1524                 r = switch_pg_num(m, argv[1]);
1525                 goto out;
1526         } else if (!strcasecmp(argv[0], "reinstate_path"))
1527                 action = reinstate_path;
1528         else if (!strcasecmp(argv[0], "fail_path"))
1529                 action = fail_path;
1530         else {
1531                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1532                 goto out;
1533         }
1534
1535         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1536         if (r) {
1537                 DMWARN("message: error getting device %s",
1538                        argv[1]);
1539                 goto out;
1540         }
1541
1542         r = action_dev(m, dev, action);
1543
1544         dm_put_device(ti, dev);
1545
1546 out:
1547         mutex_unlock(&m->work_mutex);
1548         return r;
1549 }
1550
1551 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1552                            unsigned long arg)
1553 {
1554         struct multipath *m = ti->private;
1555         struct pgpath *pgpath;
1556         struct block_device *bdev;
1557         fmode_t mode;
1558         unsigned long flags;
1559         int r;
1560
1561         bdev = NULL;
1562         mode = 0;
1563         r = 0;
1564
1565         spin_lock_irqsave(&m->lock, flags);
1566
1567         if (!m->current_pgpath)
1568                 __choose_pgpath(m, 0);
1569
1570         pgpath = m->current_pgpath;
1571
1572         if (pgpath) {
1573                 bdev = pgpath->path.dev->bdev;
1574                 mode = pgpath->path.dev->mode;
1575         }
1576
1577         if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1578                 r = -ENOTCONN;
1579         else if (!bdev)
1580                 r = -EIO;
1581
1582         spin_unlock_irqrestore(&m->lock, flags);
1583
1584         /*
1585          * Only pass ioctls through if the device sizes match exactly.
1586          */
1587         if (!bdev || ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) {
1588                 int err = scsi_verify_blk_ioctl(NULL, cmd);
1589                 if (err)
1590                         r = err;
1591         }
1592
1593         if (r == -ENOTCONN && !fatal_signal_pending(current)) {
1594                 spin_lock_irqsave(&m->lock, flags);
1595                 if (!m->current_pg) {
1596                         /* Path status changed, redo selection */
1597                         __choose_pgpath(m, 0);
1598                 }
1599                 if (m->pg_init_required)
1600                         __pg_init_all_paths(m);
1601                 spin_unlock_irqrestore(&m->lock, flags);
1602                 dm_table_run_md_queue_async(m->ti->table);
1603         }
1604
1605         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1606 }
1607
1608 static int multipath_iterate_devices(struct dm_target *ti,
1609                                      iterate_devices_callout_fn fn, void *data)
1610 {
1611         struct multipath *m = ti->private;
1612         struct priority_group *pg;
1613         struct pgpath *p;
1614         int ret = 0;
1615
1616         list_for_each_entry(pg, &m->priority_groups, list) {
1617                 list_for_each_entry(p, &pg->pgpaths, list) {
1618                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1619                         if (ret)
1620                                 goto out;
1621                 }
1622         }
1623
1624 out:
1625         return ret;
1626 }
1627
1628 static int __pgpath_busy(struct pgpath *pgpath)
1629 {
1630         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1631
1632         return blk_lld_busy(q);
1633 }
1634
1635 /*
1636  * We return "busy", only when we can map I/Os but underlying devices
1637  * are busy (so even if we map I/Os now, the I/Os will wait on
1638  * the underlying queue).
1639  * In other words, if we want to kill I/Os or queue them inside us
1640  * due to map unavailability, we don't return "busy".  Otherwise,
1641  * dm core won't give us the I/Os and we can't do what we want.
1642  */
1643 static int multipath_busy(struct dm_target *ti)
1644 {
1645         int busy = 0, has_active = 0;
1646         struct multipath *m = ti->private;
1647         struct priority_group *pg;
1648         struct pgpath *pgpath;
1649         unsigned long flags;
1650
1651         spin_lock_irqsave(&m->lock, flags);
1652
1653         /* pg_init in progress or no paths available */
1654         if (m->pg_init_in_progress ||
1655             (!m->nr_valid_paths && m->queue_if_no_path)) {
1656                 busy = 1;
1657                 goto out;
1658         }
1659         /* Guess which priority_group will be used at next mapping time */
1660         if (unlikely(!m->current_pgpath && m->next_pg))
1661                 pg = m->next_pg;
1662         else if (likely(m->current_pg))
1663                 pg = m->current_pg;
1664         else
1665                 /*
1666                  * We don't know which pg will be used at next mapping time.
1667                  * We don't call __choose_pgpath() here to avoid to trigger
1668                  * pg_init just by busy checking.
1669                  * So we don't know whether underlying devices we will be using
1670                  * at next mapping time are busy or not. Just try mapping.
1671                  */
1672                 goto out;
1673
1674         /*
1675          * If there is one non-busy active path at least, the path selector
1676          * will be able to select it. So we consider such a pg as not busy.
1677          */
1678         busy = 1;
1679         list_for_each_entry(pgpath, &pg->pgpaths, list)
1680                 if (pgpath->is_active) {
1681                         has_active = 1;
1682
1683                         if (!__pgpath_busy(pgpath)) {
1684                                 busy = 0;
1685                                 break;
1686                         }
1687                 }
1688
1689         if (!has_active)
1690                 /*
1691                  * No active path in this pg, so this pg won't be used and
1692                  * the current_pg will be changed at next mapping time.
1693                  * We need to try mapping to determine it.
1694                  */
1695                 busy = 0;
1696
1697 out:
1698         spin_unlock_irqrestore(&m->lock, flags);
1699
1700         return busy;
1701 }
1702
1703 /*-----------------------------------------------------------------
1704  * Module setup
1705  *---------------------------------------------------------------*/
1706 static struct target_type multipath_target = {
1707         .name = "multipath",
1708         .version = {1, 9, 0},
1709         .module = THIS_MODULE,
1710         .ctr = multipath_ctr,
1711         .dtr = multipath_dtr,
1712         .map_rq = multipath_map,
1713         .clone_and_map_rq = multipath_clone_and_map,
1714         .release_clone_rq = multipath_release_clone,
1715         .rq_end_io = multipath_end_io,
1716         .presuspend = multipath_presuspend,
1717         .postsuspend = multipath_postsuspend,
1718         .resume = multipath_resume,
1719         .status = multipath_status,
1720         .message = multipath_message,
1721         .ioctl  = multipath_ioctl,
1722         .iterate_devices = multipath_iterate_devices,
1723         .busy = multipath_busy,
1724 };
1725
1726 static int __init dm_multipath_init(void)
1727 {
1728         int r;
1729
1730         /* allocate a slab for the dm_ios */
1731         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1732         if (!_mpio_cache)
1733                 return -ENOMEM;
1734
1735         r = dm_register_target(&multipath_target);
1736         if (r < 0) {
1737                 DMERR("register failed %d", r);
1738                 r = -EINVAL;
1739                 goto bad_register_target;
1740         }
1741
1742         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1743         if (!kmultipathd) {
1744                 DMERR("failed to create workqueue kmpathd");
1745                 r = -ENOMEM;
1746                 goto bad_alloc_kmultipathd;
1747         }
1748
1749         /*
1750          * A separate workqueue is used to handle the device handlers
1751          * to avoid overloading existing workqueue. Overloading the
1752          * old workqueue would also create a bottleneck in the
1753          * path of the storage hardware device activation.
1754          */
1755         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1756                                                   WQ_MEM_RECLAIM);
1757         if (!kmpath_handlerd) {
1758                 DMERR("failed to create workqueue kmpath_handlerd");
1759                 r = -ENOMEM;
1760                 goto bad_alloc_kmpath_handlerd;
1761         }
1762
1763         DMINFO("version %u.%u.%u loaded",
1764                multipath_target.version[0], multipath_target.version[1],
1765                multipath_target.version[2]);
1766
1767         return 0;
1768
1769 bad_alloc_kmpath_handlerd:
1770         destroy_workqueue(kmultipathd);
1771 bad_alloc_kmultipathd:
1772         dm_unregister_target(&multipath_target);
1773 bad_register_target:
1774         kmem_cache_destroy(_mpio_cache);
1775
1776         return r;
1777 }
1778
1779 static void __exit dm_multipath_exit(void)
1780 {
1781         destroy_workqueue(kmpath_handlerd);
1782         destroy_workqueue(kmultipathd);
1783
1784         dm_unregister_target(&multipath_target);
1785         kmem_cache_destroy(_mpio_cache);
1786 }
1787
1788 module_init(dm_multipath_init);
1789 module_exit(dm_multipath_exit);
1790
1791 MODULE_DESCRIPTION(DM_NAME " multipath target");
1792 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1793 MODULE_LICENSE("GPL");