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 #include <linux/blk-mq.h>
  27
  28 #define DM_MSG_PREFIX "multipath"
  29 #define DM_PG_INIT_DELAY_MSECS 2000
  30 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  31
  32 /* Path properties */
  33 struct pgpath {
  34         struct list_head list;
  35
  36         struct priority_group *pg;      /* Owning PG */
  37         unsigned fail_count;            /* Cumulative failure count */
  38
  39         struct dm_path path;
  40         struct delayed_work activate_path;
  41
  42         bool is_active:1;               /* Path status */
  43 };
  44
  45 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  46
  47 /*
  48  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  49  * Each has a path selector which controls which path gets used.
  50  */
  51 struct priority_group {
  52         struct list_head list;
  53
  54         struct multipath *m;            /* Owning multipath instance */
  55         struct path_selector ps;
  56
  57         unsigned pg_num;                /* Reference number */
  58         unsigned nr_pgpaths;            /* Number of paths in PG */
  59         struct list_head pgpaths;
  60
  61         bool bypassed:1;                /* Temporarily bypass this PG? */
  62 };
  63
  64 /* Multipath context */
  65 struct multipath {
  66         struct list_head list;
  67         struct dm_target *ti;
  68
  69         const char *hw_handler_name;
  70         char *hw_handler_params;
  71
  72         spinlock_t lock;
  73
  74         unsigned nr_priority_groups;
  75         struct list_head priority_groups;
  76
  77         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  78
  79         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  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
  86         bool queue_io:1;                /* Must we queue all I/O? */
  87         bool queue_if_no_path:1;        /* Queue I/O if last path fails? */
  88         bool saved_queue_if_no_path:1;  /* Saved state during suspension */
  89         bool retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
  90         bool pg_init_disabled:1;        /* pg_init is not currently allowed */
  91         bool pg_init_required:1;        /* pg_init needs calling? */
  92         bool pg_init_delay_retry:1;     /* Delay pg_init retry? */
  93
  94         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  95         unsigned pg_init_count;         /* Number of times pg_init called */
  96         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  97
  98         struct work_struct trigger_event;
  99
 100         /*
 101          * We must use a mempool of dm_mpath_io structs so that we
 102          * can resubmit bios on error.
 103          */
 104         mempool_t *mpio_pool;
 105
 106         struct mutex work_mutex;
 107 };
 108
 109 /*
 110  * Context information attached to each bio we process.
 111  */
 112 struct dm_mpath_io {
 113         struct pgpath *pgpath;
 114         size_t nr_bytes;
 115 };
 116
 117 typedef int (*action_fn) (struct pgpath *pgpath);
 118
 119 static struct kmem_cache *_mpio_cache;
 120
 121 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 122 static void trigger_event(struct work_struct *work);
 123 static void activate_path(struct work_struct *work);
 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 = true;
 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
 163         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 164                 list_del(&pgpath->list);
 165                 dm_put_device(ti, pgpath->path.dev);
 166                 free_pgpath(pgpath);
 167         }
 168 }
 169
 170 static void free_priority_group(struct priority_group *pg,
 171                                 struct dm_target *ti)
 172 {
 173         struct path_selector *ps = &pg->ps;
 174
 175         if (ps->type) {
 176                 ps->type->destroy(ps);
 177                 dm_put_path_selector(ps->type);
 178         }
 179
 180         free_pgpaths(&pg->pgpaths, ti);
 181         kfree(pg);
 182 }
 183
 184 static struct multipath *alloc_multipath(struct dm_target *ti, bool use_blk_mq)
 185 {
 186         struct multipath *m;
 187
 188         m = kzalloc(sizeof(*m), GFP_KERNEL);
 189         if (m) {
 190                 INIT_LIST_HEAD(&m->priority_groups);
 191                 spin_lock_init(&m->lock);
 192                 m->queue_io = true;
 193                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 194                 INIT_WORK(&m->trigger_event, trigger_event);
 195                 init_waitqueue_head(&m->pg_init_wait);
 196                 mutex_init(&m->work_mutex);
 197
 198                 m->mpio_pool = NULL;
 199                 if (!use_blk_mq) {
 200                         unsigned min_ios = dm_get_reserved_rq_based_ios();
 201
 202                         m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
 203                         if (!m->mpio_pool) {
 204                                 kfree(m);
 205                                 return NULL;
 206                         }
 207                 }
 208
 209                 m->ti = ti;
 210                 ti->private = m;
 211         }
 212
 213         return m;
 214 }
 215
 216 static void free_multipath(struct multipath *m)
 217 {
 218         struct priority_group *pg, *tmp;
 219
 220         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 221                 list_del(&pg->list);
 222                 free_priority_group(pg, m->ti);
 223         }
 224
 225         kfree(m->hw_handler_name);
 226         kfree(m->hw_handler_params);
 227         mempool_destroy(m->mpio_pool);
 228         kfree(m);
 229 }
 230
 231 static struct dm_mpath_io *get_mpio(union map_info *info)
 232 {
 233         return info->ptr;
 234 }
 235
 236 static struct dm_mpath_io *set_mpio(struct multipath *m, union map_info *info)
 237 {
 238         struct dm_mpath_io *mpio;
 239
 240         if (!m->mpio_pool) {
 241                 /* Use blk-mq pdu memory requested via per_io_data_size */
 242                 mpio = get_mpio(info);
 243                 memset(mpio, 0, sizeof(*mpio));
 244                 return mpio;
 245         }
 246
 247         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 248         if (!mpio)
 249                 return NULL;
 250
 251         memset(mpio, 0, sizeof(*mpio));
 252         info->ptr = mpio;
 253
 254         return mpio;
 255 }
 256
 257 static void clear_request_fn_mpio(struct multipath *m, union map_info *info)
 258 {
 259         /* Only needed for non blk-mq (.request_fn) multipath */
 260         if (m->mpio_pool) {
 261                 struct dm_mpath_io *mpio = info->ptr;
 262
 263                 info->ptr = NULL;
 264                 mempool_free(mpio, m->mpio_pool);
 265         }
 266 }
 267
 268 /*-----------------------------------------------
 269  * Path selection
 270  *-----------------------------------------------*/
 271
 272 static int __pg_init_all_paths(struct multipath *m)
 273 {
 274         struct pgpath *pgpath;
 275         unsigned long pg_init_delay = 0;
 276
 277         if (m->pg_init_in_progress || m->pg_init_disabled)
 278                 return 0;
 279
 280         m->pg_init_count++;
 281         m->pg_init_required = false;
 282
 283         /* Check here to reset pg_init_required */
 284         if (!m->current_pg)
 285                 return 0;
 286
 287         if (m->pg_init_delay_retry)
 288                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 289                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 290         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 291                 /* Skip failed paths */
 292                 if (!pgpath->is_active)
 293                         continue;
 294                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 295                                        pg_init_delay))
 296                         m->pg_init_in_progress++;
 297         }
 298         return m->pg_init_in_progress;
 299 }
 300
 301 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 302 {
 303         m->current_pg = pgpath->pg;
 304
 305         /* Must we initialise the PG first, and queue I/O till it's ready? */
 306         if (m->hw_handler_name) {
 307                 m->pg_init_required = true;
 308                 m->queue_io = true;
 309         } else {
 310                 m->pg_init_required = false;
 311                 m->queue_io = false;
 312         }
 313
 314         m->pg_init_count = 0;
 315 }
 316
 317 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 318                                size_t nr_bytes)
 319 {
 320         struct dm_path *path;
 321
 322         path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 323         if (!path)
 324                 return -ENXIO;
 325
 326         m->current_pgpath = path_to_pgpath(path);
 327
 328         if (m->current_pg != pg)
 329                 __switch_pg(m, m->current_pgpath);
 330
 331         return 0;
 332 }
 333
 334 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 335 {
 336         struct priority_group *pg;
 337         bool bypassed = true;
 338
 339         if (!m->nr_valid_paths) {
 340                 m->queue_io = false;
 341                 goto failed;
 342         }
 343
 344         /* Were we instructed to switch PG? */
 345         if (m->next_pg) {
 346                 pg = m->next_pg;
 347                 m->next_pg = NULL;
 348                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 349                         return;
 350         }
 351
 352         /* Don't change PG until it has no remaining paths */
 353         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 354                 return;
 355
 356         /*
 357          * Loop through priority groups until we find a valid path.
 358          * First time we skip PGs marked 'bypassed'.
 359          * Second time we only try the ones we skipped, but set
 360          * pg_init_delay_retry so we do not hammer controllers.
 361          */
 362         do {
 363                 list_for_each_entry(pg, &m->priority_groups, list) {
 364                         if (pg->bypassed == bypassed)
 365                                 continue;
 366                         if (!__choose_path_in_pg(m, pg, nr_bytes)) {
 367                                 if (!bypassed)
 368                                         m->pg_init_delay_retry = true;
 369                                 return;
 370                         }
 371                 }
 372         } while (bypassed--);
 373
 374 failed:
 375         m->current_pgpath = NULL;
 376         m->current_pg = NULL;
 377 }
 378
 379 /*
 380  * Check whether bios must be queued in the device-mapper core rather
 381  * than here in the target.
 382  *
 383  * m->lock must be held on entry.
 384  *
 385  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 386  * same value then we are not between multipath_presuspend()
 387  * and multipath_resume() calls and we have no need to check
 388  * for the DMF_NOFLUSH_SUSPENDING flag.
 389  */
 390 static int __must_push_back(struct multipath *m)
 391 {
 392         return (m->queue_if_no_path ||
 393                 (m->queue_if_no_path != m->saved_queue_if_no_path &&
 394                  dm_noflush_suspending(m->ti)));
 395 }
 396
 397 /*
 398  * Map cloned requests
 399  */
 400 static int __multipath_map(struct dm_target *ti, struct request *clone,
 401                            union map_info *map_context,
 402                            struct request *rq, struct request **__clone)
 403 {
 404         struct multipath *m = ti->private;
 405         int r = DM_MAPIO_REQUEUE;
 406         size_t nr_bytes = clone ? blk_rq_bytes(clone) : blk_rq_bytes(rq);
 407         struct pgpath *pgpath;
 408         struct block_device *bdev;
 409         struct dm_mpath_io *mpio;
 410
 411         spin_lock_irq(&m->lock);
 412
 413         /* Do we need to select a new pgpath? */
 414         if (!m->current_pgpath || !m->queue_io)
 415                 __choose_pgpath(m, nr_bytes);
 416
 417         pgpath = m->current_pgpath;
 418
 419         if (!pgpath) {
 420                 if (!__must_push_back(m))
 421                         r = -EIO;       /* Failed */
 422                 goto out_unlock;
 423         } else if (m->queue_io || m->pg_init_required) {
 424                 __pg_init_all_paths(m);
 425                 goto out_unlock;
 426         }
 427
 428         mpio = set_mpio(m, map_context);
 429         if (!mpio)
 430                 /* ENOMEM, requeue */
 431                 goto out_unlock;
 432
 433         mpio->pgpath = pgpath;
 434         mpio->nr_bytes = nr_bytes;
 435
 436         bdev = pgpath->path.dev->bdev;
 437
 438         spin_unlock_irq(&m->lock);
 439
 440         if (clone) {
 441                 /*
 442                  * Old request-based interface: allocated clone is passed in.
 443                  * Used by: .request_fn stacked on .request_fn path(s).
 444                  */
 445                 clone->q = bdev_get_queue(bdev);
 446                 clone->rq_disk = bdev->bd_disk;
 447                 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 448         } else {
 449                 /*
 450                  * blk-mq request-based interface; used by both:
 451                  * .request_fn stacked on blk-mq path(s) and
 452                  * blk-mq stacked on blk-mq path(s).
 453                  */
 454                 *__clone = blk_mq_alloc_request(bdev_get_queue(bdev),
 455                                                 rq_data_dir(rq), BLK_MQ_REQ_NOWAIT);
 456                 if (IS_ERR(*__clone)) {
 457                         /* ENOMEM, requeue */
 458                         clear_request_fn_mpio(m, map_context);
 459                         return r;
 460                 }
 461                 (*__clone)->bio = (*__clone)->biotail = NULL;
 462                 (*__clone)->rq_disk = bdev->bd_disk;
 463                 (*__clone)->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 464         }
 465
 466         if (pgpath->pg->ps.type->start_io)
 467                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 468                                               &pgpath->path,
 469                                               nr_bytes);
 470         return DM_MAPIO_REMAPPED;
 471
 472 out_unlock:
 473         spin_unlock_irq(&m->lock);
 474
 475         return r;
 476 }
 477
 478 static int multipath_map(struct dm_target *ti, struct request *clone,
 479                          union map_info *map_context)
 480 {
 481         return __multipath_map(ti, clone, map_context, NULL, NULL);
 482 }
 483
 484 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 485                                    union map_info *map_context,
 486                                    struct request **clone)
 487 {
 488         return __multipath_map(ti, NULL, map_context, rq, clone);
 489 }
 490
 491 static void multipath_release_clone(struct request *clone)
 492 {
 493         blk_mq_free_request(clone);
 494 }
 495
 496 /*
 497  * If we run out of usable paths, should we queue I/O or error it?
 498  */
 499 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
 500                             bool save_old_value)
 501 {
 502         unsigned long flags;
 503
 504         spin_lock_irqsave(&m->lock, flags);
 505
 506         if (save_old_value)
 507                 m->saved_queue_if_no_path = m->queue_if_no_path;
 508         else
 509                 m->saved_queue_if_no_path = queue_if_no_path;
 510         m->queue_if_no_path = queue_if_no_path;
 511         spin_unlock_irqrestore(&m->lock, flags);
 512
 513         if (!queue_if_no_path)
 514                 dm_table_run_md_queue_async(m->ti->table);
 515
 516         return 0;
 517 }
 518
 519 /*
 520  * An event is triggered whenever a path is taken out of use.
 521  * Includes path failure and PG bypass.
 522  */
 523 static void trigger_event(struct work_struct *work)
 524 {
 525         struct multipath *m =
 526                 container_of(work, struct multipath, trigger_event);
 527
 528         dm_table_event(m->ti->table);
 529 }
 530
 531 /*-----------------------------------------------------------------
 532  * Constructor/argument parsing:
 533  * <#multipath feature args> [<arg>]*
 534  * <#hw_handler args> [hw_handler [<arg>]*]
 535  * <#priority groups>
 536  * <initial priority group>
 537  *     [<selector> <#selector args> [<arg>]*
 538  *      <#paths> <#per-path selector args>
 539  *         [<path> [<arg>]* ]+ ]+
 540  *---------------------------------------------------------------*/
 541 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 542                                struct dm_target *ti)
 543 {
 544         int r;
 545         struct path_selector_type *pst;
 546         unsigned ps_argc;
 547
 548         static struct dm_arg _args[] = {
 549                 {0, 1024, "invalid number of path selector args"},
 550         };
 551
 552         pst = dm_get_path_selector(dm_shift_arg(as));
 553         if (!pst) {
 554                 ti->error = "unknown path selector type";
 555                 return -EINVAL;
 556         }
 557
 558         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 559         if (r) {
 560                 dm_put_path_selector(pst);
 561                 return -EINVAL;
 562         }
 563
 564         r = pst->create(&pg->ps, ps_argc, as->argv);
 565         if (r) {
 566                 dm_put_path_selector(pst);
 567                 ti->error = "path selector constructor failed";
 568                 return r;
 569         }
 570
 571         pg->ps.type = pst;
 572         dm_consume_args(as, ps_argc);
 573
 574         return 0;
 575 }
 576
 577 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 578                                struct dm_target *ti)
 579 {
 580         int r;
 581         struct pgpath *p;
 582         struct multipath *m = ti->private;
 583         struct request_queue *q = NULL;
 584         const char *attached_handler_name;
 585
 586         /* we need at least a path arg */
 587         if (as->argc < 1) {
 588                 ti->error = "no device given";
 589                 return ERR_PTR(-EINVAL);
 590         }
 591
 592         p = alloc_pgpath();
 593         if (!p)
 594                 return ERR_PTR(-ENOMEM);
 595
 596         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 597                           &p->path.dev);
 598         if (r) {
 599                 ti->error = "error getting device";
 600                 goto bad;
 601         }
 602
 603         if (m->retain_attached_hw_handler || m->hw_handler_name)
 604                 q = bdev_get_queue(p->path.dev->bdev);
 605
 606         if (m->retain_attached_hw_handler) {
 607 retain:
 608                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 609                 if (attached_handler_name) {
 610                         /*
 611                          * Reset hw_handler_name to match the attached handler
 612                          * and clear any hw_handler_params associated with the
 613                          * ignored handler.
 614                          *
 615                          * NB. This modifies the table line to show the actual
 616                          * handler instead of the original table passed in.
 617                          */
 618                         kfree(m->hw_handler_name);
 619                         m->hw_handler_name = attached_handler_name;
 620
 621                         kfree(m->hw_handler_params);
 622                         m->hw_handler_params = NULL;
 623                 }
 624         }
 625
 626         if (m->hw_handler_name) {
 627                 r = scsi_dh_attach(q, m->hw_handler_name);
 628                 if (r == -EBUSY) {
 629                         char b[BDEVNAME_SIZE];
 630
 631                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
 632                                 bdevname(p->path.dev->bdev, b));
 633                         goto retain;
 634                 }
 635                 if (r < 0) {
 636                         ti->error = "error attaching hardware handler";
 637                         dm_put_device(ti, p->path.dev);
 638                         goto bad;
 639                 }
 640
 641                 if (m->hw_handler_params) {
 642                         r = scsi_dh_set_params(q, m->hw_handler_params);
 643                         if (r < 0) {
 644                                 ti->error = "unable to set hardware "
 645                                                         "handler parameters";
 646                                 dm_put_device(ti, p->path.dev);
 647                                 goto bad;
 648                         }
 649                 }
 650         }
 651
 652         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 653         if (r) {
 654                 dm_put_device(ti, p->path.dev);
 655                 goto bad;
 656         }
 657
 658         return p;
 659
 660  bad:
 661         free_pgpath(p);
 662         return ERR_PTR(r);
 663 }
 664
 665 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 666                                                    struct multipath *m)
 667 {
 668         static struct dm_arg _args[] = {
 669                 {1, 1024, "invalid number of paths"},
 670                 {0, 1024, "invalid number of selector args"}
 671         };
 672
 673         int r;
 674         unsigned i, nr_selector_args, nr_args;
 675         struct priority_group *pg;
 676         struct dm_target *ti = m->ti;
 677
 678         if (as->argc < 2) {
 679                 as->argc = 0;
 680                 ti->error = "not enough priority group arguments";
 681                 return ERR_PTR(-EINVAL);
 682         }
 683
 684         pg = alloc_priority_group();
 685         if (!pg) {
 686                 ti->error = "couldn't allocate priority group";
 687                 return ERR_PTR(-ENOMEM);
 688         }
 689         pg->m = m;
 690
 691         r = parse_path_selector(as, pg, ti);
 692         if (r)
 693                 goto bad;
 694
 695         /*
 696          * read the paths
 697          */
 698         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 699         if (r)
 700                 goto bad;
 701
 702         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 703         if (r)
 704                 goto bad;
 705
 706         nr_args = 1 + nr_selector_args;
 707         for (i = 0; i < pg->nr_pgpaths; i++) {
 708                 struct pgpath *pgpath;
 709                 struct dm_arg_set path_args;
 710
 711                 if (as->argc < nr_args) {
 712                         ti->error = "not enough path parameters";
 713                         r = -EINVAL;
 714                         goto bad;
 715                 }
 716
 717                 path_args.argc = nr_args;
 718                 path_args.argv = as->argv;
 719
 720                 pgpath = parse_path(&path_args, &pg->ps, ti);
 721                 if (IS_ERR(pgpath)) {
 722                         r = PTR_ERR(pgpath);
 723                         goto bad;
 724                 }
 725
 726                 pgpath->pg = pg;
 727                 list_add_tail(&pgpath->list, &pg->pgpaths);
 728                 dm_consume_args(as, nr_args);
 729         }
 730
 731         return pg;
 732
 733  bad:
 734         free_priority_group(pg, ti);
 735         return ERR_PTR(r);
 736 }
 737
 738 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 739 {
 740         unsigned hw_argc;
 741         int ret;
 742         struct dm_target *ti = m->ti;
 743
 744         static struct dm_arg _args[] = {
 745                 {0, 1024, "invalid number of hardware handler args"},
 746         };
 747
 748         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 749                 return -EINVAL;
 750
 751         if (!hw_argc)
 752                 return 0;
 753
 754         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 755
 756         if (hw_argc > 1) {
 757                 char *p;
 758                 int i, j, len = 4;
 759
 760                 for (i = 0; i <= hw_argc - 2; i++)
 761                         len += strlen(as->argv[i]) + 1;
 762                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 763                 if (!p) {
 764                         ti->error = "memory allocation failed";
 765                         ret = -ENOMEM;
 766                         goto fail;
 767                 }
 768                 j = sprintf(p, "%d", hw_argc - 1);
 769                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 770                         j = sprintf(p, "%s", as->argv[i]);
 771         }
 772         dm_consume_args(as, hw_argc - 1);
 773
 774         return 0;
 775 fail:
 776         kfree(m->hw_handler_name);
 777         m->hw_handler_name = NULL;
 778         return ret;
 779 }
 780
 781 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 782 {
 783         int r;
 784         unsigned argc;
 785         struct dm_target *ti = m->ti;
 786         const char *arg_name;
 787
 788         static struct dm_arg _args[] = {
 789                 {0, 6, "invalid number of feature args"},
 790                 {1, 50, "pg_init_retries must be between 1 and 50"},
 791                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 792         };
 793
 794         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 795         if (r)
 796                 return -EINVAL;
 797
 798         if (!argc)
 799                 return 0;
 800
 801         do {
 802                 arg_name = dm_shift_arg(as);
 803                 argc--;
 804
 805                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 806                         r = queue_if_no_path(m, true, false);
 807                         continue;
 808                 }
 809
 810                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
 811                         m->retain_attached_hw_handler = true;
 812                         continue;
 813                 }
 814
 815                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 816                     (argc >= 1)) {
 817                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 818                         argc--;
 819                         continue;
 820                 }
 821
 822                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 823                     (argc >= 1)) {
 824                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 825                         argc--;
 826                         continue;
 827                 }
 828
 829                 ti->error = "Unrecognised multipath feature request";
 830                 r = -EINVAL;
 831         } while (argc && !r);
 832
 833         return r;
 834 }
 835
 836 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 837                          char **argv)
 838 {
 839         /* target arguments */
 840         static struct dm_arg _args[] = {
 841                 {0, 1024, "invalid number of priority groups"},
 842                 {0, 1024, "invalid initial priority group number"},
 843         };
 844
 845         int r;
 846         struct multipath *m;
 847         struct dm_arg_set as;
 848         unsigned pg_count = 0;
 849         unsigned next_pg_num;
 850         bool use_blk_mq = dm_use_blk_mq(dm_table_get_md(ti->table));
 851
 852         as.argc = argc;
 853         as.argv = argv;
 854
 855         m = alloc_multipath(ti, use_blk_mq);
 856         if (!m) {
 857                 ti->error = "can't allocate multipath";
 858                 return -EINVAL;
 859         }
 860
 861         r = parse_features(&as, m);
 862         if (r)
 863                 goto bad;
 864
 865         r = parse_hw_handler(&as, m);
 866         if (r)
 867                 goto bad;
 868
 869         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 870         if (r)
 871                 goto bad;
 872
 873         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 874         if (r)
 875                 goto bad;
 876
 877         if ((!m->nr_priority_groups && next_pg_num) ||
 878             (m->nr_priority_groups && !next_pg_num)) {
 879                 ti->error = "invalid initial priority group";
 880                 r = -EINVAL;
 881                 goto bad;
 882         }
 883
 884         /* parse the priority groups */
 885         while (as.argc) {
 886                 struct priority_group *pg;
 887
 888                 pg = parse_priority_group(&as, m);
 889                 if (IS_ERR(pg)) {
 890                         r = PTR_ERR(pg);
 891                         goto bad;
 892                 }
 893
 894                 m->nr_valid_paths += pg->nr_pgpaths;
 895                 list_add_tail(&pg->list, &m->priority_groups);
 896                 pg_count++;
 897                 pg->pg_num = pg_count;
 898                 if (!--next_pg_num)
 899                         m->next_pg = pg;
 900         }
 901
 902         if (pg_count != m->nr_priority_groups) {
 903                 ti->error = "priority group count mismatch";
 904                 r = -EINVAL;
 905                 goto bad;
 906         }
 907
 908         ti->num_flush_bios = 1;
 909         ti->num_discard_bios = 1;
 910         ti->num_write_same_bios = 1;
 911         if (use_blk_mq)
 912                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
 913
 914         return 0;
 915
 916  bad:
 917         free_multipath(m);
 918         return r;
 919 }
 920
 921 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 922 {
 923         DECLARE_WAITQUEUE(wait, current);
 924         unsigned long flags;
 925
 926         add_wait_queue(&m->pg_init_wait, &wait);
 927
 928         while (1) {
 929                 set_current_state(TASK_UNINTERRUPTIBLE);
 930
 931                 spin_lock_irqsave(&m->lock, flags);
 932                 if (!m->pg_init_in_progress) {
 933                         spin_unlock_irqrestore(&m->lock, flags);
 934                         break;
 935                 }
 936                 spin_unlock_irqrestore(&m->lock, flags);
 937
 938                 io_schedule();
 939         }
 940         set_current_state(TASK_RUNNING);
 941
 942         remove_wait_queue(&m->pg_init_wait, &wait);
 943 }
 944
 945 static void flush_multipath_work(struct multipath *m)
 946 {
 947         unsigned long flags;
 948
 949         spin_lock_irqsave(&m->lock, flags);
 950         m->pg_init_disabled = true;
 951         spin_unlock_irqrestore(&m->lock, flags);
 952
 953         flush_workqueue(kmpath_handlerd);
 954         multipath_wait_for_pg_init_completion(m);
 955         flush_workqueue(kmultipathd);
 956         flush_work(&m->trigger_event);
 957
 958         spin_lock_irqsave(&m->lock, flags);
 959         m->pg_init_disabled = false;
 960         spin_unlock_irqrestore(&m->lock, flags);
 961 }
 962
 963 static void multipath_dtr(struct dm_target *ti)
 964 {
 965         struct multipath *m = ti->private;
 966
 967         flush_multipath_work(m);
 968         free_multipath(m);
 969 }
 970
 971 /*
 972  * Take a path out of use.
 973  */
 974 static int fail_path(struct pgpath *pgpath)
 975 {
 976         unsigned long flags;
 977         struct multipath *m = pgpath->pg->m;
 978
 979         spin_lock_irqsave(&m->lock, flags);
 980
 981         if (!pgpath->is_active)
 982                 goto out;
 983
 984         DMWARN("Failing path %s.", pgpath->path.dev->name);
 985
 986         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 987         pgpath->is_active = false;
 988         pgpath->fail_count++;
 989
 990         m->nr_valid_paths--;
 991
 992         if (pgpath == m->current_pgpath)
 993                 m->current_pgpath = NULL;
 994
 995         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 996                       pgpath->path.dev->name, m->nr_valid_paths);
 997
 998         schedule_work(&m->trigger_event);
 999
1000 out:
1001         spin_unlock_irqrestore(&m->lock, flags);
1002
1003         return 0;
1004 }
1005
1006 /*
1007  * Reinstate a previously-failed path
1008  */
1009 static int reinstate_path(struct pgpath *pgpath)
1010 {
1011         int r = 0, run_queue = 0;
1012         unsigned long flags;
1013         struct multipath *m = pgpath->pg->m;
1014
1015         spin_lock_irqsave(&m->lock, flags);
1016
1017         if (pgpath->is_active)
1018                 goto out;
1019
1020         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1021
1022         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1023         if (r)
1024                 goto out;
1025
1026         pgpath->is_active = true;
1027
1028         if (!m->nr_valid_paths++) {
1029                 m->current_pgpath = NULL;
1030                 run_queue = 1;
1031         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1032                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1033                         m->pg_init_in_progress++;
1034         }
1035
1036         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1037                       pgpath->path.dev->name, m->nr_valid_paths);
1038
1039         schedule_work(&m->trigger_event);
1040
1041 out:
1042         spin_unlock_irqrestore(&m->lock, flags);
1043         if (run_queue)
1044                 dm_table_run_md_queue_async(m->ti->table);
1045
1046         return r;
1047 }
1048
1049 /*
1050  * Fail or reinstate all paths that match the provided struct dm_dev.
1051  */
1052 static int action_dev(struct multipath *m, struct dm_dev *dev,
1053                       action_fn action)
1054 {
1055         int r = -EINVAL;
1056         struct pgpath *pgpath;
1057         struct priority_group *pg;
1058
1059         list_for_each_entry(pg, &m->priority_groups, list) {
1060                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1061                         if (pgpath->path.dev == dev)
1062                                 r = action(pgpath);
1063                 }
1064         }
1065
1066         return r;
1067 }
1068
1069 /*
1070  * Temporarily try to avoid having to use the specified PG
1071  */
1072 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1073                       bool bypassed)
1074 {
1075         unsigned long flags;
1076
1077         spin_lock_irqsave(&m->lock, flags);
1078
1079         pg->bypassed = bypassed;
1080         m->current_pgpath = NULL;
1081         m->current_pg = NULL;
1082
1083         spin_unlock_irqrestore(&m->lock, flags);
1084
1085         schedule_work(&m->trigger_event);
1086 }
1087
1088 /*
1089  * Switch to using the specified PG from the next I/O that gets mapped
1090  */
1091 static int switch_pg_num(struct multipath *m, const char *pgstr)
1092 {
1093         struct priority_group *pg;
1094         unsigned pgnum;
1095         unsigned long flags;
1096         char dummy;
1097
1098         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1099             (pgnum > m->nr_priority_groups)) {
1100                 DMWARN("invalid PG number supplied to switch_pg_num");
1101                 return -EINVAL;
1102         }
1103
1104         spin_lock_irqsave(&m->lock, flags);
1105         list_for_each_entry(pg, &m->priority_groups, list) {
1106                 pg->bypassed = false;
1107                 if (--pgnum)
1108                         continue;
1109
1110                 m->current_pgpath = NULL;
1111                 m->current_pg = NULL;
1112                 m->next_pg = pg;
1113         }
1114         spin_unlock_irqrestore(&m->lock, flags);
1115
1116         schedule_work(&m->trigger_event);
1117         return 0;
1118 }
1119
1120 /*
1121  * Set/clear bypassed status of a PG.
1122  * PGs are numbered upwards from 1 in the order they were declared.
1123  */
1124 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1125 {
1126         struct priority_group *pg;
1127         unsigned pgnum;
1128         char dummy;
1129
1130         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1131             (pgnum > m->nr_priority_groups)) {
1132                 DMWARN("invalid PG number supplied to bypass_pg");
1133                 return -EINVAL;
1134         }
1135
1136         list_for_each_entry(pg, &m->priority_groups, list) {
1137                 if (!--pgnum)
1138                         break;
1139         }
1140
1141         bypass_pg(m, pg, bypassed);
1142         return 0;
1143 }
1144
1145 /*
1146  * Should we retry pg_init immediately?
1147  */
1148 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1149 {
1150         unsigned long flags;
1151         bool limit_reached = false;
1152
1153         spin_lock_irqsave(&m->lock, flags);
1154
1155         if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1156                 m->pg_init_required = true;
1157         else
1158                 limit_reached = true;
1159
1160         spin_unlock_irqrestore(&m->lock, flags);
1161
1162         return limit_reached;
1163 }
1164
1165 static void pg_init_done(void *data, int errors)
1166 {
1167         struct pgpath *pgpath = data;
1168         struct priority_group *pg = pgpath->pg;
1169         struct multipath *m = pg->m;
1170         unsigned long flags;
1171         bool delay_retry = false;
1172
1173         /* device or driver problems */
1174         switch (errors) {
1175         case SCSI_DH_OK:
1176                 break;
1177         case SCSI_DH_NOSYS:
1178                 if (!m->hw_handler_name) {
1179                         errors = 0;
1180                         break;
1181                 }
1182                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1183                       "Error %d.", m->hw_handler_name, errors);
1184                 /*
1185                  * Fail path for now, so we do not ping pong
1186                  */
1187                 fail_path(pgpath);
1188                 break;
1189         case SCSI_DH_DEV_TEMP_BUSY:
1190                 /*
1191                  * Probably doing something like FW upgrade on the
1192                  * controller so try the other pg.
1193                  */
1194                 bypass_pg(m, pg, true);
1195                 break;
1196         case SCSI_DH_RETRY:
1197                 /* Wait before retrying. */
1198                 delay_retry = 1;
1199         case SCSI_DH_IMM_RETRY:
1200         case SCSI_DH_RES_TEMP_UNAVAIL:
1201                 if (pg_init_limit_reached(m, pgpath))
1202                         fail_path(pgpath);
1203                 errors = 0;
1204                 break;
1205         case SCSI_DH_DEV_OFFLINED:
1206         default:
1207                 /*
1208                  * We probably do not want to fail the path for a device
1209                  * error, but this is what the old dm did. In future
1210                  * patches we can do more advanced handling.
1211                  */
1212                 fail_path(pgpath);
1213         }
1214
1215         spin_lock_irqsave(&m->lock, flags);
1216         if (errors) {
1217                 if (pgpath == m->current_pgpath) {
1218                         DMERR("Could not failover device. Error %d.", errors);
1219                         m->current_pgpath = NULL;
1220                         m->current_pg = NULL;
1221                 }
1222         } else if (!m->pg_init_required)
1223                 pg->bypassed = false;
1224
1225         if (--m->pg_init_in_progress)
1226                 /* Activations of other paths are still on going */
1227                 goto out;
1228
1229         if (m->pg_init_required) {
1230                 m->pg_init_delay_retry = delay_retry;
1231                 if (__pg_init_all_paths(m))
1232                         goto out;
1233         }
1234         m->queue_io = false;
1235
1236         /*
1237          * Wake up any thread waiting to suspend.
1238          */
1239         wake_up(&m->pg_init_wait);
1240
1241 out:
1242         spin_unlock_irqrestore(&m->lock, flags);
1243 }
1244
1245 static void activate_path(struct work_struct *work)
1246 {
1247         struct pgpath *pgpath =
1248                 container_of(work, struct pgpath, activate_path.work);
1249
1250         if (pgpath->is_active)
1251                 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1252                                  pg_init_done, pgpath);
1253         else
1254                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1255 }
1256
1257 static int noretry_error(int error)
1258 {
1259         switch (error) {
1260         case -EOPNOTSUPP:
1261         case -EREMOTEIO:
1262         case -EILSEQ:
1263         case -ENODATA:
1264         case -ENOSPC:
1265                 return 1;
1266         }
1267
1268         /* Anything else could be a path failure, so should be retried */
1269         return 0;
1270 }
1271
1272 /*
1273  * end_io handling
1274  */
1275 static int do_end_io(struct multipath *m, struct request *clone,
1276                      int error, struct dm_mpath_io *mpio)
1277 {
1278         /*
1279          * We don't queue any clone request inside the multipath target
1280          * during end I/O handling, since those clone requests don't have
1281          * bio clones.  If we queue them inside the multipath target,
1282          * we need to make bio clones, that requires memory allocation.
1283          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1284          *  don't have bio clones.)
1285          * Instead of queueing the clone request here, we queue the original
1286          * request into dm core, which will remake a clone request and
1287          * clone bios for it and resubmit it later.
1288          */
1289         int r = DM_ENDIO_REQUEUE;
1290         unsigned long flags;
1291
1292         if (!error && !clone->errors)
1293                 return 0;       /* I/O complete */
1294
1295         if (noretry_error(error))
1296                 return error;
1297
1298         if (mpio->pgpath)
1299                 fail_path(mpio->pgpath);
1300
1301         spin_lock_irqsave(&m->lock, flags);
1302         if (!m->nr_valid_paths) {
1303                 if (!m->queue_if_no_path) {
1304                         if (!__must_push_back(m))
1305                                 r = -EIO;
1306                 } else {
1307                         if (error == -EBADE)
1308                                 r = error;
1309                 }
1310         }
1311         spin_unlock_irqrestore(&m->lock, flags);
1312
1313         return r;
1314 }
1315
1316 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1317                             int error, union map_info *map_context)
1318 {
1319         struct multipath *m = ti->private;
1320         struct dm_mpath_io *mpio = get_mpio(map_context);
1321         struct pgpath *pgpath;
1322         struct path_selector *ps;
1323         int r;
1324
1325         BUG_ON(!mpio);
1326
1327         r = do_end_io(m, clone, error, mpio);
1328         pgpath = mpio->pgpath;
1329         if (pgpath) {
1330                 ps = &pgpath->pg->ps;
1331                 if (ps->type->end_io)
1332                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1333         }
1334         clear_request_fn_mpio(m, map_context);
1335
1336         return r;
1337 }
1338
1339 /*
1340  * Suspend can't complete until all the I/O is processed so if
1341  * the last path fails we must error any remaining I/O.
1342  * Note that if the freeze_bdev fails while suspending, the
1343  * queue_if_no_path state is lost - userspace should reset it.
1344  */
1345 static void multipath_presuspend(struct dm_target *ti)
1346 {
1347         struct multipath *m = ti->private;
1348
1349         queue_if_no_path(m, false, true);
1350 }
1351
1352 static void multipath_postsuspend(struct dm_target *ti)
1353 {
1354         struct multipath *m = ti->private;
1355
1356         mutex_lock(&m->work_mutex);
1357         flush_multipath_work(m);
1358         mutex_unlock(&m->work_mutex);
1359 }
1360
1361 /*
1362  * Restore the queue_if_no_path setting.
1363  */
1364 static void multipath_resume(struct dm_target *ti)
1365 {
1366         struct multipath *m = ti->private;
1367         unsigned long flags;
1368
1369         spin_lock_irqsave(&m->lock, flags);
1370         m->queue_if_no_path = m->saved_queue_if_no_path;
1371         spin_unlock_irqrestore(&m->lock, flags);
1372 }
1373
1374 /*
1375  * Info output has the following format:
1376  * num_multipath_feature_args [multipath_feature_args]*
1377  * num_handler_status_args [handler_status_args]*
1378  * num_groups init_group_number
1379  *            [A|D|E num_ps_status_args [ps_status_args]*
1380  *             num_paths num_selector_args
1381  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1382  *
1383  * Table output has the following format (identical to the constructor string):
1384  * num_feature_args [features_args]*
1385  * num_handler_args hw_handler [hw_handler_args]*
1386  * num_groups init_group_number
1387  *     [priority selector-name num_ps_args [ps_args]*
1388  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1389  */
1390 static void multipath_status(struct dm_target *ti, status_type_t type,
1391                              unsigned status_flags, char *result, unsigned maxlen)
1392 {
1393         int sz = 0;
1394         unsigned long flags;
1395         struct multipath *m = ti->private;
1396         struct priority_group *pg;
1397         struct pgpath *p;
1398         unsigned pg_num;
1399         char state;
1400
1401         spin_lock_irqsave(&m->lock, flags);
1402
1403         /* Features */
1404         if (type == STATUSTYPE_INFO)
1405                 DMEMIT("2 %u %u ", m->queue_io, m->pg_init_count);
1406         else {
1407                 DMEMIT("%u ", m->queue_if_no_path +
1408                               (m->pg_init_retries > 0) * 2 +
1409                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1410                               m->retain_attached_hw_handler);
1411                 if (m->queue_if_no_path)
1412                         DMEMIT("queue_if_no_path ");
1413                 if (m->pg_init_retries)
1414                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1415                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1416                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1417                 if (m->retain_attached_hw_handler)
1418                         DMEMIT("retain_attached_hw_handler ");
1419         }
1420
1421         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1422                 DMEMIT("0 ");
1423         else
1424                 DMEMIT("1 %s ", m->hw_handler_name);
1425
1426         DMEMIT("%u ", m->nr_priority_groups);
1427
1428         if (m->next_pg)
1429                 pg_num = m->next_pg->pg_num;
1430         else if (m->current_pg)
1431                 pg_num = m->current_pg->pg_num;
1432         else
1433                 pg_num = (m->nr_priority_groups ? 1 : 0);
1434
1435         DMEMIT("%u ", pg_num);
1436
1437         switch (type) {
1438         case STATUSTYPE_INFO:
1439                 list_for_each_entry(pg, &m->priority_groups, list) {
1440                         if (pg->bypassed)
1441                                 state = 'D';    /* Disabled */
1442                         else if (pg == m->current_pg)
1443                                 state = 'A';    /* Currently Active */
1444                         else
1445                                 state = 'E';    /* Enabled */
1446
1447                         DMEMIT("%c ", state);
1448
1449                         if (pg->ps.type->status)
1450                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1451                                                           result + sz,
1452                                                           maxlen - sz);
1453                         else
1454                                 DMEMIT("0 ");
1455
1456                         DMEMIT("%u %u ", pg->nr_pgpaths,
1457                                pg->ps.type->info_args);
1458
1459                         list_for_each_entry(p, &pg->pgpaths, list) {
1460                                 DMEMIT("%s %s %u ", p->path.dev->name,
1461                                        p->is_active ? "A" : "F",
1462                                        p->fail_count);
1463                                 if (pg->ps.type->status)
1464                                         sz += pg->ps.type->status(&pg->ps,
1465                                               &p->path, type, result + sz,
1466                                               maxlen - sz);
1467                         }
1468                 }
1469                 break;
1470
1471         case STATUSTYPE_TABLE:
1472                 list_for_each_entry(pg, &m->priority_groups, list) {
1473                         DMEMIT("%s ", pg->ps.type->name);
1474
1475                         if (pg->ps.type->status)
1476                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1477                                                           result + sz,
1478                                                           maxlen - sz);
1479                         else
1480                                 DMEMIT("0 ");
1481
1482                         DMEMIT("%u %u ", pg->nr_pgpaths,
1483                                pg->ps.type->table_args);
1484
1485                         list_for_each_entry(p, &pg->pgpaths, list) {
1486                                 DMEMIT("%s ", p->path.dev->name);
1487                                 if (pg->ps.type->status)
1488                                         sz += pg->ps.type->status(&pg->ps,
1489                                               &p->path, type, result + sz,
1490                                               maxlen - sz);
1491                         }
1492                 }
1493                 break;
1494         }
1495
1496         spin_unlock_irqrestore(&m->lock, flags);
1497 }
1498
1499 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1500 {
1501         int r = -EINVAL;
1502         struct dm_dev *dev;
1503         struct multipath *m = ti->private;
1504         action_fn action;
1505
1506         mutex_lock(&m->work_mutex);
1507
1508         if (dm_suspended(ti)) {
1509                 r = -EBUSY;
1510                 goto out;
1511         }
1512
1513         if (argc == 1) {
1514                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1515                         r = queue_if_no_path(m, true, false);
1516                         goto out;
1517                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1518                         r = queue_if_no_path(m, false, false);
1519                         goto out;
1520                 }
1521         }
1522
1523         if (argc != 2) {
1524                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1525                 goto out;
1526         }
1527
1528         if (!strcasecmp(argv[0], "disable_group")) {
1529                 r = bypass_pg_num(m, argv[1], true);
1530                 goto out;
1531         } else if (!strcasecmp(argv[0], "enable_group")) {
1532                 r = bypass_pg_num(m, argv[1], false);
1533                 goto out;
1534         } else if (!strcasecmp(argv[0], "switch_group")) {
1535                 r = switch_pg_num(m, argv[1]);
1536                 goto out;
1537         } else if (!strcasecmp(argv[0], "reinstate_path"))
1538                 action = reinstate_path;
1539         else if (!strcasecmp(argv[0], "fail_path"))
1540                 action = fail_path;
1541         else {
1542                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1543                 goto out;
1544         }
1545
1546         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1547         if (r) {
1548                 DMWARN("message: error getting device %s",
1549                        argv[1]);
1550                 goto out;
1551         }
1552
1553         r = action_dev(m, dev, action);
1554
1555         dm_put_device(ti, dev);
1556
1557 out:
1558         mutex_unlock(&m->work_mutex);
1559         return r;
1560 }
1561
1562 static int multipath_prepare_ioctl(struct dm_target *ti,
1563                 struct block_device **bdev, fmode_t *mode)
1564 {
1565         struct multipath *m = ti->private;
1566         unsigned long flags;
1567         int r;
1568
1569         spin_lock_irqsave(&m->lock, flags);
1570
1571         if (!m->current_pgpath)
1572                 __choose_pgpath(m, 0);
1573
1574         if (m->current_pgpath) {
1575                 if (!m->queue_io) {
1576                         *bdev = m->current_pgpath->path.dev->bdev;
1577                         *mode = m->current_pgpath->path.dev->mode;
1578                         r = 0;
1579                 } else {
1580                         /* pg_init has not started or completed */
1581                         r = -ENOTCONN;
1582                 }
1583         } else {
1584                 /* No path is available */
1585                 if (m->queue_if_no_path)
1586                         r = -ENOTCONN;
1587                 else
1588                         r = -EIO;
1589         }
1590
1591         spin_unlock_irqrestore(&m->lock, flags);
1592
1593         if (r == -ENOTCONN) {
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         /*
1606          * Only pass ioctls through if the device sizes match exactly.
1607          */
1608         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1609                 return 1;
1610         return r;
1611 }
1612
1613 static int multipath_iterate_devices(struct dm_target *ti,
1614                                      iterate_devices_callout_fn fn, void *data)
1615 {
1616         struct multipath *m = ti->private;
1617         struct priority_group *pg;
1618         struct pgpath *p;
1619         int ret = 0;
1620
1621         list_for_each_entry(pg, &m->priority_groups, list) {
1622                 list_for_each_entry(p, &pg->pgpaths, list) {
1623                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1624                         if (ret)
1625                                 goto out;
1626                 }
1627         }
1628
1629 out:
1630         return ret;
1631 }
1632
1633 static int pgpath_busy(struct pgpath *pgpath)
1634 {
1635         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1636
1637         return blk_lld_busy(q);
1638 }
1639
1640 /*
1641  * We return "busy", only when we can map I/Os but underlying devices
1642  * are busy (so even if we map I/Os now, the I/Os will wait on
1643  * the underlying queue).
1644  * In other words, if we want to kill I/Os or queue them inside us
1645  * due to map unavailability, we don't return "busy".  Otherwise,
1646  * dm core won't give us the I/Os and we can't do what we want.
1647  */
1648 static int multipath_busy(struct dm_target *ti)
1649 {
1650         bool busy = false, has_active = false;
1651         struct multipath *m = ti->private;
1652         struct priority_group *pg;
1653         struct pgpath *pgpath;
1654         unsigned long flags;
1655
1656         spin_lock_irqsave(&m->lock, flags);
1657
1658         /* pg_init in progress or no paths available */
1659         if (m->pg_init_in_progress ||
1660             (!m->nr_valid_paths && m->queue_if_no_path)) {
1661                 busy = true;
1662                 goto out;
1663         }
1664         /* Guess which priority_group will be used at next mapping time */
1665         if (unlikely(!m->current_pgpath && m->next_pg))
1666                 pg = m->next_pg;
1667         else if (likely(m->current_pg))
1668                 pg = m->current_pg;
1669         else
1670                 /*
1671                  * We don't know which pg will be used at next mapping time.
1672                  * We don't call __choose_pgpath() here to avoid to trigger
1673                  * pg_init just by busy checking.
1674                  * So we don't know whether underlying devices we will be using
1675                  * at next mapping time are busy or not. Just try mapping.
1676                  */
1677                 goto out;
1678
1679         /*
1680          * If there is one non-busy active path at least, the path selector
1681          * will be able to select it. So we consider such a pg as not busy.
1682          */
1683         busy = true;
1684         list_for_each_entry(pgpath, &pg->pgpaths, list)
1685                 if (pgpath->is_active) {
1686                         has_active = true;
1687                         if (!pgpath_busy(pgpath)) {
1688                                 busy = false;
1689                                 break;
1690                         }
1691                 }
1692
1693         if (!has_active)
1694                 /*
1695                  * No active path in this pg, so this pg won't be used and
1696                  * the current_pg will be changed at next mapping time.
1697                  * We need to try mapping to determine it.
1698                  */
1699                 busy = false;
1700
1701 out:
1702         spin_unlock_irqrestore(&m->lock, flags);
1703
1704         return busy;
1705 }
1706
1707 /*-----------------------------------------------------------------
1708  * Module setup
1709  *---------------------------------------------------------------*/
1710 static struct target_type multipath_target = {
1711         .name = "multipath",
1712         .version = {1, 11, 0},
1713         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1714         .module = THIS_MODULE,
1715         .ctr = multipath_ctr,
1716         .dtr = multipath_dtr,
1717         .map_rq = multipath_map,
1718         .clone_and_map_rq = multipath_clone_and_map,
1719         .release_clone_rq = multipath_release_clone,
1720         .rq_end_io = multipath_end_io,
1721         .presuspend = multipath_presuspend,
1722         .postsuspend = multipath_postsuspend,
1723         .resume = multipath_resume,
1724         .status = multipath_status,
1725         .message = multipath_message,
1726         .prepare_ioctl = multipath_prepare_ioctl,
1727         .iterate_devices = multipath_iterate_devices,
1728         .busy = multipath_busy,
1729 };
1730
1731 static int __init dm_multipath_init(void)
1732 {
1733         int r;
1734
1735         /* allocate a slab for the dm_ios */
1736         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1737         if (!_mpio_cache)
1738                 return -ENOMEM;
1739
1740         r = dm_register_target(&multipath_target);
1741         if (r < 0) {
1742                 DMERR("register failed %d", r);
1743                 r = -EINVAL;
1744                 goto bad_register_target;
1745         }
1746
1747         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1748         if (!kmultipathd) {
1749                 DMERR("failed to create workqueue kmpathd");
1750                 r = -ENOMEM;
1751                 goto bad_alloc_kmultipathd;
1752         }
1753
1754         /*
1755          * A separate workqueue is used to handle the device handlers
1756          * to avoid overloading existing workqueue. Overloading the
1757          * old workqueue would also create a bottleneck in the
1758          * path of the storage hardware device activation.
1759          */
1760         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1761                                                   WQ_MEM_RECLAIM);
1762         if (!kmpath_handlerd) {
1763                 DMERR("failed to create workqueue kmpath_handlerd");
1764                 r = -ENOMEM;
1765                 goto bad_alloc_kmpath_handlerd;
1766         }
1767
1768         DMINFO("version %u.%u.%u loaded",
1769                multipath_target.version[0], multipath_target.version[1],
1770                multipath_target.version[2]);
1771
1772         return 0;
1773
1774 bad_alloc_kmpath_handlerd:
1775         destroy_workqueue(kmultipathd);
1776 bad_alloc_kmultipathd:
1777         dm_unregister_target(&multipath_target);
1778 bad_register_target:
1779         kmem_cache_destroy(_mpio_cache);
1780
1781         return r;
1782 }
1783
1784 static void __exit dm_multipath_exit(void)
1785 {
1786         destroy_workqueue(kmpath_handlerd);
1787         destroy_workqueue(kmultipathd);
1788
1789         dm_unregister_target(&multipath_target);
1790         kmem_cache_destroy(_mpio_cache);
1791 }
1792
1793 module_init(dm_multipath_init);
1794 module_exit(dm_multipath_exit);
1795
1796 MODULE_DESCRIPTION(DM_NAME " multipath target");
1797 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1798 MODULE_LICENSE("GPL");