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_KERNEL);
 432                 if (IS_ERR(*__clone))
 433                         /* ENOMEM, requeue */
 434                         return r;
 435                 (*__clone)->bio = (*__clone)->biotail = NULL;
 436                 (*__clone)->rq_disk = bdev->bd_disk;
 437                 (*__clone)->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 438         }
 439
 440         if (pgpath->pg->ps.type->start_io)
 441                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 442                                               &pgpath->path,
 443                                               nr_bytes);
 444         return DM_MAPIO_REMAPPED;
 445
 446 out_unlock:
 447         spin_unlock_irq(&m->lock);
 448
 449         return r;
 450 }
 451
 452 static int multipath_map(struct dm_target *ti, struct request *clone,
 453                          union map_info *map_context)
 454 {
 455         return __multipath_map(ti, clone, map_context, NULL, NULL);
 456 }
 457
 458 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 459                                    union map_info *map_context,
 460                                    struct request **clone)
 461 {
 462         return __multipath_map(ti, NULL, map_context, rq, clone);
 463 }
 464
 465 static void multipath_release_clone(struct request *clone)
 466 {
 467         blk_put_request(clone);
 468 }
 469
 470 /*
 471  * If we run out of usable paths, should we queue I/O or error it?
 472  */
 473 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 474                             unsigned save_old_value)
 475 {
 476         unsigned long flags;
 477
 478         spin_lock_irqsave(&m->lock, flags);
 479
 480         if (save_old_value)
 481                 m->saved_queue_if_no_path = m->queue_if_no_path;
 482         else
 483                 m->saved_queue_if_no_path = queue_if_no_path;
 484         m->queue_if_no_path = queue_if_no_path;
 485         spin_unlock_irqrestore(&m->lock, flags);
 486
 487         if (!queue_if_no_path)
 488                 dm_table_run_md_queue_async(m->ti->table);
 489
 490         return 0;
 491 }
 492
 493 /*
 494  * An event is triggered whenever a path is taken out of use.
 495  * Includes path failure and PG bypass.
 496  */
 497 static void trigger_event(struct work_struct *work)
 498 {
 499         struct multipath *m =
 500                 container_of(work, struct multipath, trigger_event);
 501
 502         dm_table_event(m->ti->table);
 503 }
 504
 505 /*-----------------------------------------------------------------
 506  * Constructor/argument parsing:
 507  * <#multipath feature args> [<arg>]*
 508  * <#hw_handler args> [hw_handler [<arg>]*]
 509  * <#priority groups>
 510  * <initial priority group>
 511  *     [<selector> <#selector args> [<arg>]*
 512  *      <#paths> <#per-path selector args>
 513  *         [<path> [<arg>]* ]+ ]+
 514  *---------------------------------------------------------------*/
 515 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 516                                struct dm_target *ti)
 517 {
 518         int r;
 519         struct path_selector_type *pst;
 520         unsigned ps_argc;
 521
 522         static struct dm_arg _args[] = {
 523                 {0, 1024, "invalid number of path selector args"},
 524         };
 525
 526         pst = dm_get_path_selector(dm_shift_arg(as));
 527         if (!pst) {
 528                 ti->error = "unknown path selector type";
 529                 return -EINVAL;
 530         }
 531
 532         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 533         if (r) {
 534                 dm_put_path_selector(pst);
 535                 return -EINVAL;
 536         }
 537
 538         r = pst->create(&pg->ps, ps_argc, as->argv);
 539         if (r) {
 540                 dm_put_path_selector(pst);
 541                 ti->error = "path selector constructor failed";
 542                 return r;
 543         }
 544
 545         pg->ps.type = pst;
 546         dm_consume_args(as, ps_argc);
 547
 548         return 0;
 549 }
 550
 551 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 552                                struct dm_target *ti)
 553 {
 554         int r;
 555         struct pgpath *p;
 556         struct multipath *m = ti->private;
 557         struct request_queue *q = NULL;
 558         const char *attached_handler_name;
 559
 560         /* we need at least a path arg */
 561         if (as->argc < 1) {
 562                 ti->error = "no device given";
 563                 return ERR_PTR(-EINVAL);
 564         }
 565
 566         p = alloc_pgpath();
 567         if (!p)
 568                 return ERR_PTR(-ENOMEM);
 569
 570         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 571                           &p->path.dev);
 572         if (r) {
 573                 ti->error = "error getting device";
 574                 goto bad;
 575         }
 576
 577         if (m->retain_attached_hw_handler || m->hw_handler_name)
 578                 q = bdev_get_queue(p->path.dev->bdev);
 579
 580         if (m->retain_attached_hw_handler) {
 581                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 582                 if (attached_handler_name) {
 583                         /*
 584                          * Reset hw_handler_name to match the attached handler
 585                          * and clear any hw_handler_params associated with the
 586                          * ignored handler.
 587                          *
 588                          * NB. This modifies the table line to show the actual
 589                          * handler instead of the original table passed in.
 590                          */
 591                         kfree(m->hw_handler_name);
 592                         m->hw_handler_name = attached_handler_name;
 593
 594                         kfree(m->hw_handler_params);
 595                         m->hw_handler_params = NULL;
 596                 }
 597         }
 598
 599         if (m->hw_handler_name) {
 600                 /*
 601                  * Increments scsi_dh reference, even when using an
 602                  * already-attached handler.
 603                  */
 604                 r = scsi_dh_attach(q, m->hw_handler_name);
 605                 if (r == -EBUSY) {
 606                         /*
 607                          * Already attached to different hw_handler:
 608                          * try to reattach with correct one.
 609                          */
 610                         scsi_dh_detach(q);
 611                         r = scsi_dh_attach(q, m->hw_handler_name);
 612                 }
 613
 614                 if (r < 0) {
 615                         ti->error = "error attaching hardware handler";
 616                         dm_put_device(ti, p->path.dev);
 617                         goto bad;
 618                 }
 619
 620                 if (m->hw_handler_params) {
 621                         r = scsi_dh_set_params(q, m->hw_handler_params);
 622                         if (r < 0) {
 623                                 ti->error = "unable to set hardware "
 624                                                         "handler parameters";
 625                                 scsi_dh_detach(q);
 626                                 dm_put_device(ti, p->path.dev);
 627                                 goto bad;
 628                         }
 629                 }
 630         }
 631
 632         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 633         if (r) {
 634                 dm_put_device(ti, p->path.dev);
 635                 goto bad;
 636         }
 637
 638         return p;
 639
 640  bad:
 641         free_pgpath(p);
 642         return ERR_PTR(r);
 643 }
 644
 645 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 646                                                    struct multipath *m)
 647 {
 648         static struct dm_arg _args[] = {
 649                 {1, 1024, "invalid number of paths"},
 650                 {0, 1024, "invalid number of selector args"}
 651         };
 652
 653         int r;
 654         unsigned i, nr_selector_args, nr_args;
 655         struct priority_group *pg;
 656         struct dm_target *ti = m->ti;
 657
 658         if (as->argc < 2) {
 659                 as->argc = 0;
 660                 ti->error = "not enough priority group arguments";
 661                 return ERR_PTR(-EINVAL);
 662         }
 663
 664         pg = alloc_priority_group();
 665         if (!pg) {
 666                 ti->error = "couldn't allocate priority group";
 667                 return ERR_PTR(-ENOMEM);
 668         }
 669         pg->m = m;
 670
 671         r = parse_path_selector(as, pg, ti);
 672         if (r)
 673                 goto bad;
 674
 675         /*
 676          * read the paths
 677          */
 678         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 679         if (r)
 680                 goto bad;
 681
 682         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 683         if (r)
 684                 goto bad;
 685
 686         nr_args = 1 + nr_selector_args;
 687         for (i = 0; i < pg->nr_pgpaths; i++) {
 688                 struct pgpath *pgpath;
 689                 struct dm_arg_set path_args;
 690
 691                 if (as->argc < nr_args) {
 692                         ti->error = "not enough path parameters";
 693                         r = -EINVAL;
 694                         goto bad;
 695                 }
 696
 697                 path_args.argc = nr_args;
 698                 path_args.argv = as->argv;
 699
 700                 pgpath = parse_path(&path_args, &pg->ps, ti);
 701                 if (IS_ERR(pgpath)) {
 702                         r = PTR_ERR(pgpath);
 703                         goto bad;
 704                 }
 705
 706                 pgpath->pg = pg;
 707                 list_add_tail(&pgpath->list, &pg->pgpaths);
 708                 dm_consume_args(as, nr_args);
 709         }
 710
 711         return pg;
 712
 713  bad:
 714         free_priority_group(pg, ti);
 715         return ERR_PTR(r);
 716 }
 717
 718 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 719 {
 720         unsigned hw_argc;
 721         int ret;
 722         struct dm_target *ti = m->ti;
 723
 724         static struct dm_arg _args[] = {
 725                 {0, 1024, "invalid number of hardware handler args"},
 726         };
 727
 728         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 729                 return -EINVAL;
 730
 731         if (!hw_argc)
 732                 return 0;
 733
 734         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 735         if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
 736                                      "scsi_dh_%s", m->hw_handler_name)) {
 737                 ti->error = "unknown hardware handler type";
 738                 ret = -EINVAL;
 739                 goto fail;
 740         }
 741
 742         if (hw_argc > 1) {
 743                 char *p;
 744                 int i, j, len = 4;
 745
 746                 for (i = 0; i <= hw_argc - 2; i++)
 747                         len += strlen(as->argv[i]) + 1;
 748                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 749                 if (!p) {
 750                         ti->error = "memory allocation failed";
 751                         ret = -ENOMEM;
 752                         goto fail;
 753                 }
 754                 j = sprintf(p, "%d", hw_argc - 1);
 755                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 756                         j = sprintf(p, "%s", as->argv[i]);
 757         }
 758         dm_consume_args(as, hw_argc - 1);
 759
 760         return 0;
 761 fail:
 762         kfree(m->hw_handler_name);
 763         m->hw_handler_name = NULL;
 764         return ret;
 765 }
 766
 767 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 768 {
 769         int r;
 770         unsigned argc;
 771         struct dm_target *ti = m->ti;
 772         const char *arg_name;
 773
 774         static struct dm_arg _args[] = {
 775                 {0, 6, "invalid number of feature args"},
 776                 {1, 50, "pg_init_retries must be between 1 and 50"},
 777                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 778         };
 779
 780         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 781         if (r)
 782                 return -EINVAL;
 783
 784         if (!argc)
 785                 return 0;
 786
 787         do {
 788                 arg_name = dm_shift_arg(as);
 789                 argc--;
 790
 791                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 792                         r = queue_if_no_path(m, 1, 0);
 793                         continue;
 794                 }
 795
 796                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
 797                         m->retain_attached_hw_handler = 1;
 798                         continue;
 799                 }
 800
 801                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 802                     (argc >= 1)) {
 803                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 804                         argc--;
 805                         continue;
 806                 }
 807
 808                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 809                     (argc >= 1)) {
 810                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 811                         argc--;
 812                         continue;
 813                 }
 814
 815                 ti->error = "Unrecognised multipath feature request";
 816                 r = -EINVAL;
 817         } while (argc && !r);
 818
 819         return r;
 820 }
 821
 822 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 823                          char **argv)
 824 {
 825         /* target arguments */
 826         static struct dm_arg _args[] = {
 827                 {0, 1024, "invalid number of priority groups"},
 828                 {0, 1024, "invalid initial priority group number"},
 829         };
 830
 831         int r;
 832         struct multipath *m;
 833         struct dm_arg_set as;
 834         unsigned pg_count = 0;
 835         unsigned next_pg_num;
 836
 837         as.argc = argc;
 838         as.argv = argv;
 839
 840         m = alloc_multipath(ti);
 841         if (!m) {
 842                 ti->error = "can't allocate multipath";
 843                 return -EINVAL;
 844         }
 845
 846         r = parse_features(&as, m);
 847         if (r)
 848                 goto bad;
 849
 850         r = parse_hw_handler(&as, m);
 851         if (r)
 852                 goto bad;
 853
 854         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 855         if (r)
 856                 goto bad;
 857
 858         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 859         if (r)
 860                 goto bad;
 861
 862         if ((!m->nr_priority_groups && next_pg_num) ||
 863             (m->nr_priority_groups && !next_pg_num)) {
 864                 ti->error = "invalid initial priority group";
 865                 r = -EINVAL;
 866                 goto bad;
 867         }
 868
 869         /* parse the priority groups */
 870         while (as.argc) {
 871                 struct priority_group *pg;
 872
 873                 pg = parse_priority_group(&as, m);
 874                 if (IS_ERR(pg)) {
 875                         r = PTR_ERR(pg);
 876                         goto bad;
 877                 }
 878
 879                 m->nr_valid_paths += pg->nr_pgpaths;
 880                 list_add_tail(&pg->list, &m->priority_groups);
 881                 pg_count++;
 882                 pg->pg_num = pg_count;
 883                 if (!--next_pg_num)
 884                         m->next_pg = pg;
 885         }
 886
 887         if (pg_count != m->nr_priority_groups) {
 888                 ti->error = "priority group count mismatch";
 889                 r = -EINVAL;
 890                 goto bad;
 891         }
 892
 893         ti->num_flush_bios = 1;
 894         ti->num_discard_bios = 1;
 895         ti->num_write_same_bios = 1;
 896
 897         return 0;
 898
 899  bad:
 900         free_multipath(m);
 901         return r;
 902 }
 903
 904 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 905 {
 906         DECLARE_WAITQUEUE(wait, current);
 907         unsigned long flags;
 908
 909         add_wait_queue(&m->pg_init_wait, &wait);
 910
 911         while (1) {
 912                 set_current_state(TASK_UNINTERRUPTIBLE);
 913
 914                 spin_lock_irqsave(&m->lock, flags);
 915                 if (!m->pg_init_in_progress) {
 916                         spin_unlock_irqrestore(&m->lock, flags);
 917                         break;
 918                 }
 919                 spin_unlock_irqrestore(&m->lock, flags);
 920
 921                 io_schedule();
 922         }
 923         set_current_state(TASK_RUNNING);
 924
 925         remove_wait_queue(&m->pg_init_wait, &wait);
 926 }
 927
 928 static void flush_multipath_work(struct multipath *m)
 929 {
 930         unsigned long flags;
 931
 932         spin_lock_irqsave(&m->lock, flags);
 933         m->pg_init_disabled = 1;
 934         spin_unlock_irqrestore(&m->lock, flags);
 935
 936         flush_workqueue(kmpath_handlerd);
 937         multipath_wait_for_pg_init_completion(m);
 938         flush_workqueue(kmultipathd);
 939         flush_work(&m->trigger_event);
 940
 941         spin_lock_irqsave(&m->lock, flags);
 942         m->pg_init_disabled = 0;
 943         spin_unlock_irqrestore(&m->lock, flags);
 944 }
 945
 946 static void multipath_dtr(struct dm_target *ti)
 947 {
 948         struct multipath *m = ti->private;
 949
 950         flush_multipath_work(m);
 951         free_multipath(m);
 952 }
 953
 954 /*
 955  * Take a path out of use.
 956  */
 957 static int fail_path(struct pgpath *pgpath)
 958 {
 959         unsigned long flags;
 960         struct multipath *m = pgpath->pg->m;
 961
 962         spin_lock_irqsave(&m->lock, flags);
 963
 964         if (!pgpath->is_active)
 965                 goto out;
 966
 967         DMWARN("Failing path %s.", pgpath->path.dev->name);
 968
 969         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 970         pgpath->is_active = 0;
 971         pgpath->fail_count++;
 972
 973         m->nr_valid_paths--;
 974
 975         if (pgpath == m->current_pgpath)
 976                 m->current_pgpath = NULL;
 977
 978         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 979                       pgpath->path.dev->name, m->nr_valid_paths);
 980
 981         schedule_work(&m->trigger_event);
 982
 983 out:
 984         spin_unlock_irqrestore(&m->lock, flags);
 985
 986         return 0;
 987 }
 988
 989 /*
 990  * Reinstate a previously-failed path
 991  */
 992 static int reinstate_path(struct pgpath *pgpath)
 993 {
 994         int r = 0, run_queue = 0;
 995         unsigned long flags;
 996         struct multipath *m = pgpath->pg->m;
 997
 998         spin_lock_irqsave(&m->lock, flags);
 999
1000         if (pgpath->is_active)
1001                 goto out;
1002
1003         if (!pgpath->pg->ps.type->reinstate_path) {
1004                 DMWARN("Reinstate path not supported by path selector %s",
1005                        pgpath->pg->ps.type->name);
1006                 r = -EINVAL;
1007                 goto out;
1008         }
1009
1010         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1011         if (r)
1012                 goto out;
1013
1014         pgpath->is_active = 1;
1015
1016         if (!m->nr_valid_paths++) {
1017                 m->current_pgpath = NULL;
1018                 run_queue = 1;
1019         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1020                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1021                         m->pg_init_in_progress++;
1022         }
1023
1024         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1025                       pgpath->path.dev->name, m->nr_valid_paths);
1026
1027         schedule_work(&m->trigger_event);
1028
1029 out:
1030         spin_unlock_irqrestore(&m->lock, flags);
1031         if (run_queue)
1032                 dm_table_run_md_queue_async(m->ti->table);
1033
1034         return r;
1035 }
1036
1037 /*
1038  * Fail or reinstate all paths that match the provided struct dm_dev.
1039  */
1040 static int action_dev(struct multipath *m, struct dm_dev *dev,
1041                       action_fn action)
1042 {
1043         int r = -EINVAL;
1044         struct pgpath *pgpath;
1045         struct priority_group *pg;
1046
1047         list_for_each_entry(pg, &m->priority_groups, list) {
1048                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1049                         if (pgpath->path.dev == dev)
1050                                 r = action(pgpath);
1051                 }
1052         }
1053
1054         return r;
1055 }
1056
1057 /*
1058  * Temporarily try to avoid having to use the specified PG
1059  */
1060 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1061                       int bypassed)
1062 {
1063         unsigned long flags;
1064
1065         spin_lock_irqsave(&m->lock, flags);
1066
1067         pg->bypassed = bypassed;
1068         m->current_pgpath = NULL;
1069         m->current_pg = NULL;
1070
1071         spin_unlock_irqrestore(&m->lock, flags);
1072
1073         schedule_work(&m->trigger_event);
1074 }
1075
1076 /*
1077  * Switch to using the specified PG from the next I/O that gets mapped
1078  */
1079 static int switch_pg_num(struct multipath *m, const char *pgstr)
1080 {
1081         struct priority_group *pg;
1082         unsigned pgnum;
1083         unsigned long flags;
1084         char dummy;
1085
1086         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1087             (pgnum > m->nr_priority_groups)) {
1088                 DMWARN("invalid PG number supplied to switch_pg_num");
1089                 return -EINVAL;
1090         }
1091
1092         spin_lock_irqsave(&m->lock, flags);
1093         list_for_each_entry(pg, &m->priority_groups, list) {
1094                 pg->bypassed = 0;
1095                 if (--pgnum)
1096                         continue;
1097
1098                 m->current_pgpath = NULL;
1099                 m->current_pg = NULL;
1100                 m->next_pg = pg;
1101         }
1102         spin_unlock_irqrestore(&m->lock, flags);
1103
1104         schedule_work(&m->trigger_event);
1105         return 0;
1106 }
1107
1108 /*
1109  * Set/clear bypassed status of a PG.
1110  * PGs are numbered upwards from 1 in the order they were declared.
1111  */
1112 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1113 {
1114         struct priority_group *pg;
1115         unsigned pgnum;
1116         char dummy;
1117
1118         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1119             (pgnum > m->nr_priority_groups)) {
1120                 DMWARN("invalid PG number supplied to bypass_pg");
1121                 return -EINVAL;
1122         }
1123
1124         list_for_each_entry(pg, &m->priority_groups, list) {
1125                 if (!--pgnum)
1126                         break;
1127         }
1128
1129         bypass_pg(m, pg, bypassed);
1130         return 0;
1131 }
1132
1133 /*
1134  * Should we retry pg_init immediately?
1135  */
1136 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1137 {
1138         unsigned long flags;
1139         int limit_reached = 0;
1140
1141         spin_lock_irqsave(&m->lock, flags);
1142
1143         if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1144                 m->pg_init_required = 1;
1145         else
1146                 limit_reached = 1;
1147
1148         spin_unlock_irqrestore(&m->lock, flags);
1149
1150         return limit_reached;
1151 }
1152
1153 static void pg_init_done(void *data, int errors)
1154 {
1155         struct pgpath *pgpath = data;
1156         struct priority_group *pg = pgpath->pg;
1157         struct multipath *m = pg->m;
1158         unsigned long flags;
1159         unsigned delay_retry = 0;
1160
1161         /* device or driver problems */
1162         switch (errors) {
1163         case SCSI_DH_OK:
1164                 break;
1165         case SCSI_DH_NOSYS:
1166                 if (!m->hw_handler_name) {
1167                         errors = 0;
1168                         break;
1169                 }
1170                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1171                       "Error %d.", m->hw_handler_name, errors);
1172                 /*
1173                  * Fail path for now, so we do not ping pong
1174                  */
1175                 fail_path(pgpath);
1176                 break;
1177         case SCSI_DH_DEV_TEMP_BUSY:
1178                 /*
1179                  * Probably doing something like FW upgrade on the
1180                  * controller so try the other pg.
1181                  */
1182                 bypass_pg(m, pg, 1);
1183                 break;
1184         case SCSI_DH_RETRY:
1185                 /* Wait before retrying. */
1186                 delay_retry = 1;
1187         case SCSI_DH_IMM_RETRY:
1188         case SCSI_DH_RES_TEMP_UNAVAIL:
1189                 if (pg_init_limit_reached(m, pgpath))
1190                         fail_path(pgpath);
1191                 errors = 0;
1192                 break;
1193         default:
1194                 /*
1195                  * We probably do not want to fail the path for a device
1196                  * error, but this is what the old dm did. In future
1197                  * patches we can do more advanced handling.
1198                  */
1199                 fail_path(pgpath);
1200         }
1201
1202         spin_lock_irqsave(&m->lock, flags);
1203         if (errors) {
1204                 if (pgpath == m->current_pgpath) {
1205                         DMERR("Could not failover device. Error %d.", errors);
1206                         m->current_pgpath = NULL;
1207                         m->current_pg = NULL;
1208                 }
1209         } else if (!m->pg_init_required)
1210                 pg->bypassed = 0;
1211
1212         if (--m->pg_init_in_progress)
1213                 /* Activations of other paths are still on going */
1214                 goto out;
1215
1216         if (m->pg_init_required) {
1217                 m->pg_init_delay_retry = delay_retry;
1218                 if (__pg_init_all_paths(m))
1219                         goto out;
1220         }
1221         m->queue_io = 0;
1222
1223         /*
1224          * Wake up any thread waiting to suspend.
1225          */
1226         wake_up(&m->pg_init_wait);
1227
1228 out:
1229         spin_unlock_irqrestore(&m->lock, flags);
1230 }
1231
1232 static void activate_path(struct work_struct *work)
1233 {
1234         struct pgpath *pgpath =
1235                 container_of(work, struct pgpath, activate_path.work);
1236
1237         if (pgpath->is_active)
1238                 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1239                                  pg_init_done, pgpath);
1240         else
1241                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1242 }
1243
1244 static int noretry_error(int error)
1245 {
1246         switch (error) {
1247         case -EOPNOTSUPP:
1248         case -EREMOTEIO:
1249         case -EILSEQ:
1250         case -ENODATA:
1251         case -ENOSPC:
1252                 return 1;
1253         }
1254
1255         /* Anything else could be a path failure, so should be retried */
1256         return 0;
1257 }
1258
1259 /*
1260  * end_io handling
1261  */
1262 static int do_end_io(struct multipath *m, struct request *clone,
1263                      int error, struct dm_mpath_io *mpio)
1264 {
1265         /*
1266          * We don't queue any clone request inside the multipath target
1267          * during end I/O handling, since those clone requests don't have
1268          * bio clones.  If we queue them inside the multipath target,
1269          * we need to make bio clones, that requires memory allocation.
1270          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1271          *  don't have bio clones.)
1272          * Instead of queueing the clone request here, we queue the original
1273          * request into dm core, which will remake a clone request and
1274          * clone bios for it and resubmit it later.
1275          */
1276         int r = DM_ENDIO_REQUEUE;
1277         unsigned long flags;
1278
1279         if (!error && !clone->errors)
1280                 return 0;       /* I/O complete */
1281
1282         if (noretry_error(error))
1283                 return error;
1284
1285         if (mpio->pgpath)
1286                 fail_path(mpio->pgpath);
1287
1288         spin_lock_irqsave(&m->lock, flags);
1289         if (!m->nr_valid_paths) {
1290                 if (!m->queue_if_no_path) {
1291                         if (!__must_push_back(m))
1292                                 r = -EIO;
1293                 } else {
1294                         if (error == -EBADE)
1295                                 r = error;
1296                 }
1297         }
1298         spin_unlock_irqrestore(&m->lock, flags);
1299
1300         return r;
1301 }
1302
1303 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1304                             int error, union map_info *map_context)
1305 {
1306         struct multipath *m = ti->private;
1307         struct dm_mpath_io *mpio = map_context->ptr;
1308         struct pgpath *pgpath;
1309         struct path_selector *ps;
1310         int r;
1311
1312         BUG_ON(!mpio);
1313
1314         r  = do_end_io(m, clone, error, mpio);
1315         pgpath = mpio->pgpath;
1316         if (pgpath) {
1317                 ps = &pgpath->pg->ps;
1318                 if (ps->type->end_io)
1319                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1320         }
1321         clear_mapinfo(m, map_context);
1322
1323         return r;
1324 }
1325
1326 /*
1327  * Suspend can't complete until all the I/O is processed so if
1328  * the last path fails we must error any remaining I/O.
1329  * Note that if the freeze_bdev fails while suspending, the
1330  * queue_if_no_path state is lost - userspace should reset it.
1331  */
1332 static void multipath_presuspend(struct dm_target *ti)
1333 {
1334         struct multipath *m = (struct multipath *) ti->private;
1335
1336         queue_if_no_path(m, 0, 1);
1337 }
1338
1339 static void multipath_postsuspend(struct dm_target *ti)
1340 {
1341         struct multipath *m = ti->private;
1342
1343         mutex_lock(&m->work_mutex);
1344         flush_multipath_work(m);
1345         mutex_unlock(&m->work_mutex);
1346 }
1347
1348 /*
1349  * Restore the queue_if_no_path setting.
1350  */
1351 static void multipath_resume(struct dm_target *ti)
1352 {
1353         struct multipath *m = (struct multipath *) ti->private;
1354         unsigned long flags;
1355
1356         spin_lock_irqsave(&m->lock, flags);
1357         m->queue_if_no_path = m->saved_queue_if_no_path;
1358         spin_unlock_irqrestore(&m->lock, flags);
1359 }
1360
1361 /*
1362  * Info output has the following format:
1363  * num_multipath_feature_args [multipath_feature_args]*
1364  * num_handler_status_args [handler_status_args]*
1365  * num_groups init_group_number
1366  *            [A|D|E num_ps_status_args [ps_status_args]*
1367  *             num_paths num_selector_args
1368  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1369  *
1370  * Table output has the following format (identical to the constructor string):
1371  * num_feature_args [features_args]*
1372  * num_handler_args hw_handler [hw_handler_args]*
1373  * num_groups init_group_number
1374  *     [priority selector-name num_ps_args [ps_args]*
1375  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1376  */
1377 static void multipath_status(struct dm_target *ti, status_type_t type,
1378                              unsigned status_flags, char *result, unsigned maxlen)
1379 {
1380         int sz = 0;
1381         unsigned long flags;
1382         struct multipath *m = (struct multipath *) ti->private;
1383         struct priority_group *pg;
1384         struct pgpath *p;
1385         unsigned pg_num;
1386         char state;
1387
1388         spin_lock_irqsave(&m->lock, flags);
1389
1390         /* Features */
1391         if (type == STATUSTYPE_INFO)
1392                 DMEMIT("2 %u %u ", m->queue_io, m->pg_init_count);
1393         else {
1394                 DMEMIT("%u ", m->queue_if_no_path +
1395                               (m->pg_init_retries > 0) * 2 +
1396                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1397                               m->retain_attached_hw_handler);
1398                 if (m->queue_if_no_path)
1399                         DMEMIT("queue_if_no_path ");
1400                 if (m->pg_init_retries)
1401                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1402                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1403                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1404                 if (m->retain_attached_hw_handler)
1405                         DMEMIT("retain_attached_hw_handler ");
1406         }
1407
1408         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1409                 DMEMIT("0 ");
1410         else
1411                 DMEMIT("1 %s ", m->hw_handler_name);
1412
1413         DMEMIT("%u ", m->nr_priority_groups);
1414
1415         if (m->next_pg)
1416                 pg_num = m->next_pg->pg_num;
1417         else if (m->current_pg)
1418                 pg_num = m->current_pg->pg_num;
1419         else
1420                 pg_num = (m->nr_priority_groups ? 1 : 0);
1421
1422         DMEMIT("%u ", pg_num);
1423
1424         switch (type) {
1425         case STATUSTYPE_INFO:
1426                 list_for_each_entry(pg, &m->priority_groups, list) {
1427                         if (pg->bypassed)
1428                                 state = 'D';    /* Disabled */
1429                         else if (pg == m->current_pg)
1430                                 state = 'A';    /* Currently Active */
1431                         else
1432                                 state = 'E';    /* Enabled */
1433
1434                         DMEMIT("%c ", state);
1435
1436                         if (pg->ps.type->status)
1437                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1438                                                           result + sz,
1439                                                           maxlen - sz);
1440                         else
1441                                 DMEMIT("0 ");
1442
1443                         DMEMIT("%u %u ", pg->nr_pgpaths,
1444                                pg->ps.type->info_args);
1445
1446                         list_for_each_entry(p, &pg->pgpaths, list) {
1447                                 DMEMIT("%s %s %u ", p->path.dev->name,
1448                                        p->is_active ? "A" : "F",
1449                                        p->fail_count);
1450                                 if (pg->ps.type->status)
1451                                         sz += pg->ps.type->status(&pg->ps,
1452                                               &p->path, type, result + sz,
1453                                               maxlen - sz);
1454                         }
1455                 }
1456                 break;
1457
1458         case STATUSTYPE_TABLE:
1459                 list_for_each_entry(pg, &m->priority_groups, list) {
1460                         DMEMIT("%s ", pg->ps.type->name);
1461
1462                         if (pg->ps.type->status)
1463                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1464                                                           result + sz,
1465                                                           maxlen - sz);
1466                         else
1467                                 DMEMIT("0 ");
1468
1469                         DMEMIT("%u %u ", pg->nr_pgpaths,
1470                                pg->ps.type->table_args);
1471
1472                         list_for_each_entry(p, &pg->pgpaths, list) {
1473                                 DMEMIT("%s ", p->path.dev->name);
1474                                 if (pg->ps.type->status)
1475                                         sz += pg->ps.type->status(&pg->ps,
1476                                               &p->path, type, result + sz,
1477                                               maxlen - sz);
1478                         }
1479                 }
1480                 break;
1481         }
1482
1483         spin_unlock_irqrestore(&m->lock, flags);
1484 }
1485
1486 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1487 {
1488         int r = -EINVAL;
1489         struct dm_dev *dev;
1490         struct multipath *m = (struct multipath *) ti->private;
1491         action_fn action;
1492
1493         mutex_lock(&m->work_mutex);
1494
1495         if (dm_suspended(ti)) {
1496                 r = -EBUSY;
1497                 goto out;
1498         }
1499
1500         if (argc == 1) {
1501                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1502                         r = queue_if_no_path(m, 1, 0);
1503                         goto out;
1504                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1505                         r = queue_if_no_path(m, 0, 0);
1506                         goto out;
1507                 }
1508         }
1509
1510         if (argc != 2) {
1511                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1512                 goto out;
1513         }
1514
1515         if (!strcasecmp(argv[0], "disable_group")) {
1516                 r = bypass_pg_num(m, argv[1], 1);
1517                 goto out;
1518         } else if (!strcasecmp(argv[0], "enable_group")) {
1519                 r = bypass_pg_num(m, argv[1], 0);
1520                 goto out;
1521         } else if (!strcasecmp(argv[0], "switch_group")) {
1522                 r = switch_pg_num(m, argv[1]);
1523                 goto out;
1524         } else if (!strcasecmp(argv[0], "reinstate_path"))
1525                 action = reinstate_path;
1526         else if (!strcasecmp(argv[0], "fail_path"))
1527                 action = fail_path;
1528         else {
1529                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1530                 goto out;
1531         }
1532
1533         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1534         if (r) {
1535                 DMWARN("message: error getting device %s",
1536                        argv[1]);
1537                 goto out;
1538         }
1539
1540         r = action_dev(m, dev, action);
1541
1542         dm_put_device(ti, dev);
1543
1544 out:
1545         mutex_unlock(&m->work_mutex);
1546         return r;
1547 }
1548
1549 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1550                            unsigned long arg)
1551 {
1552         struct multipath *m = ti->private;
1553         struct pgpath *pgpath;
1554         struct block_device *bdev;
1555         fmode_t mode;
1556         unsigned long flags;
1557         int r;
1558
1559         bdev = NULL;
1560         mode = 0;
1561         r = 0;
1562
1563         spin_lock_irqsave(&m->lock, flags);
1564
1565         if (!m->current_pgpath)
1566                 __choose_pgpath(m, 0);
1567
1568         pgpath = m->current_pgpath;
1569
1570         if (pgpath) {
1571                 bdev = pgpath->path.dev->bdev;
1572                 mode = pgpath->path.dev->mode;
1573         }
1574
1575         if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1576                 r = -ENOTCONN;
1577         else if (!bdev)
1578                 r = -EIO;
1579
1580         spin_unlock_irqrestore(&m->lock, flags);
1581
1582         /*
1583          * Only pass ioctls through if the device sizes match exactly.
1584          */
1585         if (!bdev || ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) {
1586                 int err = scsi_verify_blk_ioctl(NULL, cmd);
1587                 if (err)
1588                         r = err;
1589         }
1590
1591         if (r == -ENOTCONN && !fatal_signal_pending(current)) {
1592                 spin_lock_irqsave(&m->lock, flags);
1593                 if (!m->current_pg) {
1594                         /* Path status changed, redo selection */
1595                         __choose_pgpath(m, 0);
1596                 }
1597                 if (m->pg_init_required)
1598                         __pg_init_all_paths(m);
1599                 spin_unlock_irqrestore(&m->lock, flags);
1600                 dm_table_run_md_queue_async(m->ti->table);
1601         }
1602
1603         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1604 }
1605
1606 static int multipath_iterate_devices(struct dm_target *ti,
1607                                      iterate_devices_callout_fn fn, void *data)
1608 {
1609         struct multipath *m = ti->private;
1610         struct priority_group *pg;
1611         struct pgpath *p;
1612         int ret = 0;
1613
1614         list_for_each_entry(pg, &m->priority_groups, list) {
1615                 list_for_each_entry(p, &pg->pgpaths, list) {
1616                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1617                         if (ret)
1618                                 goto out;
1619                 }
1620         }
1621
1622 out:
1623         return ret;
1624 }
1625
1626 static int __pgpath_busy(struct pgpath *pgpath)
1627 {
1628         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1629
1630         return dm_underlying_device_busy(q);
1631 }
1632
1633 /*
1634  * We return "busy", only when we can map I/Os but underlying devices
1635  * are busy (so even if we map I/Os now, the I/Os will wait on
1636  * the underlying queue).
1637  * In other words, if we want to kill I/Os or queue them inside us
1638  * due to map unavailability, we don't return "busy".  Otherwise,
1639  * dm core won't give us the I/Os and we can't do what we want.
1640  */
1641 static int multipath_busy(struct dm_target *ti)
1642 {
1643         int busy = 0, has_active = 0;
1644         struct multipath *m = ti->private;
1645         struct priority_group *pg;
1646         struct pgpath *pgpath;
1647         unsigned long flags;
1648
1649         spin_lock_irqsave(&m->lock, flags);
1650
1651         /* pg_init in progress or no paths available */
1652         if (m->pg_init_in_progress ||
1653             (!m->nr_valid_paths && m->queue_if_no_path)) {
1654                 busy = 1;
1655                 goto out;
1656         }
1657         /* Guess which priority_group will be used at next mapping time */
1658         if (unlikely(!m->current_pgpath && m->next_pg))
1659                 pg = m->next_pg;
1660         else if (likely(m->current_pg))
1661                 pg = m->current_pg;
1662         else
1663                 /*
1664                  * We don't know which pg will be used at next mapping time.
1665                  * We don't call __choose_pgpath() here to avoid to trigger
1666                  * pg_init just by busy checking.
1667                  * So we don't know whether underlying devices we will be using
1668                  * at next mapping time are busy or not. Just try mapping.
1669                  */
1670                 goto out;
1671
1672         /*
1673          * If there is one non-busy active path at least, the path selector
1674          * will be able to select it. So we consider such a pg as not busy.
1675          */
1676         busy = 1;
1677         list_for_each_entry(pgpath, &pg->pgpaths, list)
1678                 if (pgpath->is_active) {
1679                         has_active = 1;
1680
1681                         if (!__pgpath_busy(pgpath)) {
1682                                 busy = 0;
1683                                 break;
1684                         }
1685                 }
1686
1687         if (!has_active)
1688                 /*
1689                  * No active path in this pg, so this pg won't be used and
1690                  * the current_pg will be changed at next mapping time.
1691                  * We need to try mapping to determine it.
1692                  */
1693                 busy = 0;
1694
1695 out:
1696         spin_unlock_irqrestore(&m->lock, flags);
1697
1698         return busy;
1699 }
1700
1701 /*-----------------------------------------------------------------
1702  * Module setup
1703  *---------------------------------------------------------------*/
1704 static struct target_type multipath_target = {
1705         .name = "multipath",
1706         .version = {1, 8, 0},
1707         .module = THIS_MODULE,
1708         .ctr = multipath_ctr,
1709         .dtr = multipath_dtr,
1710         .map_rq = multipath_map,
1711         .clone_and_map_rq = multipath_clone_and_map,
1712         .release_clone_rq = multipath_release_clone,
1713         .rq_end_io = multipath_end_io,
1714         .presuspend = multipath_presuspend,
1715         .postsuspend = multipath_postsuspend,
1716         .resume = multipath_resume,
1717         .status = multipath_status,
1718         .message = multipath_message,
1719         .ioctl  = multipath_ioctl,
1720         .iterate_devices = multipath_iterate_devices,
1721         .busy = multipath_busy,
1722 };
1723
1724 static int __init dm_multipath_init(void)
1725 {
1726         int r;
1727
1728         /* allocate a slab for the dm_ios */
1729         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1730         if (!_mpio_cache)
1731                 return -ENOMEM;
1732
1733         r = dm_register_target(&multipath_target);
1734         if (r < 0) {
1735                 DMERR("register failed %d", r);
1736                 r = -EINVAL;
1737                 goto bad_register_target;
1738         }
1739
1740         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1741         if (!kmultipathd) {
1742                 DMERR("failed to create workqueue kmpathd");
1743                 r = -ENOMEM;
1744                 goto bad_alloc_kmultipathd;
1745         }
1746
1747         /*
1748          * A separate workqueue is used to handle the device handlers
1749          * to avoid overloading existing workqueue. Overloading the
1750          * old workqueue would also create a bottleneck in the
1751          * path of the storage hardware device activation.
1752          */
1753         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1754                                                   WQ_MEM_RECLAIM);
1755         if (!kmpath_handlerd) {
1756                 DMERR("failed to create workqueue kmpath_handlerd");
1757                 r = -ENOMEM;
1758                 goto bad_alloc_kmpath_handlerd;
1759         }
1760
1761         DMINFO("version %u.%u.%u loaded",
1762                multipath_target.version[0], multipath_target.version[1],
1763                multipath_target.version[2]);
1764
1765         return 0;
1766
1767 bad_alloc_kmpath_handlerd:
1768         destroy_workqueue(kmultipathd);
1769 bad_alloc_kmultipathd:
1770         dm_unregister_target(&multipath_target);
1771 bad_register_target:
1772         kmem_cache_destroy(_mpio_cache);
1773
1774         return r;
1775 }
1776
1777 static void __exit dm_multipath_exit(void)
1778 {
1779         destroy_workqueue(kmpath_handlerd);
1780         destroy_workqueue(kmultipathd);
1781
1782         dm_unregister_target(&multipath_target);
1783         kmem_cache_destroy(_mpio_cache);
1784 }
1785
1786 module_init(dm_multipath_init);
1787 module_exit(dm_multipath_exit);
1788
1789 MODULE_DESCRIPTION(DM_NAME " multipath target");
1790 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1791 MODULE_LICENSE("GPL");