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         struct pgpath *current_pgpath;
  80         struct priority_group *current_pg;
  81         struct priority_group *next_pg; /* Switch to this PG if set */
  82
  83         unsigned long flags;            /* Multipath state flags */
  84
  85         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  86         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  87
  88         atomic_t nr_valid_paths;        /* Total number of usable paths */
  89         atomic_t pg_init_in_progress;   /* Only one pg_init allowed at once */
  90         atomic_t pg_init_count;         /* Number of times pg_init called */
  91
  92         /*
  93          * We must use a mempool of dm_mpath_io structs so that we
  94          * can resubmit bios on error.
  95          */
  96         mempool_t *mpio_pool;
  97
  98         struct mutex work_mutex;
  99         struct work_struct trigger_event;
 100 };
 101
 102 /*
 103  * Context information attached to each bio we process.
 104  */
 105 struct dm_mpath_io {
 106         struct pgpath *pgpath;
 107         size_t nr_bytes;
 108 };
 109
 110 typedef int (*action_fn) (struct pgpath *pgpath);
 111
 112 static struct kmem_cache *_mpio_cache;
 113
 114 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 115 static void trigger_event(struct work_struct *work);
 116 static void activate_path(struct work_struct *work);
 117
 118 /*-----------------------------------------------
 119  * Multipath state flags.
 120  *-----------------------------------------------*/
 121
 122 #define MPATHF_QUEUE_IO 0                       /* Must we queue all I/O? */
 123 #define MPATHF_QUEUE_IF_NO_PATH 1               /* Queue I/O if last path fails? */
 124 #define MPATHF_SAVED_QUEUE_IF_NO_PATH 2         /* Saved state during suspension */
 125 #define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3     /* If there's already a hw_handler present, don't change it. */
 126 #define MPATHF_PG_INIT_DISABLED 4               /* pg_init is not currently allowed */
 127 #define MPATHF_PG_INIT_REQUIRED 5               /* pg_init needs calling? */
 128 #define MPATHF_PG_INIT_DELAY_RETRY 6            /* Delay pg_init retry? */
 129
 130 /*-----------------------------------------------
 131  * Allocation routines
 132  *-----------------------------------------------*/
 133
 134 static struct pgpath *alloc_pgpath(void)
 135 {
 136         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 137
 138         if (pgpath) {
 139                 pgpath->is_active = true;
 140                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 141         }
 142
 143         return pgpath;
 144 }
 145
 146 static void free_pgpath(struct pgpath *pgpath)
 147 {
 148         kfree(pgpath);
 149 }
 150
 151 static struct priority_group *alloc_priority_group(void)
 152 {
 153         struct priority_group *pg;
 154
 155         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 156
 157         if (pg)
 158                 INIT_LIST_HEAD(&pg->pgpaths);
 159
 160         return pg;
 161 }
 162
 163 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 164 {
 165         struct pgpath *pgpath, *tmp;
 166
 167         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 168                 list_del(&pgpath->list);
 169                 dm_put_device(ti, pgpath->path.dev);
 170                 free_pgpath(pgpath);
 171         }
 172 }
 173
 174 static void free_priority_group(struct priority_group *pg,
 175                                 struct dm_target *ti)
 176 {
 177         struct path_selector *ps = &pg->ps;
 178
 179         if (ps->type) {
 180                 ps->type->destroy(ps);
 181                 dm_put_path_selector(ps->type);
 182         }
 183
 184         free_pgpaths(&pg->pgpaths, ti);
 185         kfree(pg);
 186 }
 187
 188 static struct multipath *alloc_multipath(struct dm_target *ti, bool use_blk_mq)
 189 {
 190         struct multipath *m;
 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                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 197                 atomic_set(&m->nr_valid_paths, 0);
 198                 atomic_set(&m->pg_init_in_progress, 0);
 199                 atomic_set(&m->pg_init_count, 0);
 200                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 201                 INIT_WORK(&m->trigger_event, trigger_event);
 202                 init_waitqueue_head(&m->pg_init_wait);
 203                 mutex_init(&m->work_mutex);
 204
 205                 m->mpio_pool = NULL;
 206                 if (!use_blk_mq) {
 207                         unsigned min_ios = dm_get_reserved_rq_based_ios();
 208
 209                         m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
 210                         if (!m->mpio_pool) {
 211                                 kfree(m);
 212                                 return NULL;
 213                         }
 214                 }
 215
 216                 m->ti = ti;
 217                 ti->private = m;
 218         }
 219
 220         return m;
 221 }
 222
 223 static void free_multipath(struct multipath *m)
 224 {
 225         struct priority_group *pg, *tmp;
 226
 227         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 228                 list_del(&pg->list);
 229                 free_priority_group(pg, m->ti);
 230         }
 231
 232         kfree(m->hw_handler_name);
 233         kfree(m->hw_handler_params);
 234         mempool_destroy(m->mpio_pool);
 235         kfree(m);
 236 }
 237
 238 static struct dm_mpath_io *get_mpio(union map_info *info)
 239 {
 240         return info->ptr;
 241 }
 242
 243 static struct dm_mpath_io *set_mpio(struct multipath *m, union map_info *info)
 244 {
 245         struct dm_mpath_io *mpio;
 246
 247         if (!m->mpio_pool) {
 248                 /* Use blk-mq pdu memory requested via per_io_data_size */
 249                 mpio = get_mpio(info);
 250                 memset(mpio, 0, sizeof(*mpio));
 251                 return mpio;
 252         }
 253
 254         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 255         if (!mpio)
 256                 return NULL;
 257
 258         memset(mpio, 0, sizeof(*mpio));
 259         info->ptr = mpio;
 260
 261         return mpio;
 262 }
 263
 264 static void clear_request_fn_mpio(struct multipath *m, union map_info *info)
 265 {
 266         /* Only needed for non blk-mq (.request_fn) multipath */
 267         if (m->mpio_pool) {
 268                 struct dm_mpath_io *mpio = info->ptr;
 269
 270                 info->ptr = NULL;
 271                 mempool_free(mpio, m->mpio_pool);
 272         }
 273 }
 274
 275 /*-----------------------------------------------
 276  * Path selection
 277  *-----------------------------------------------*/
 278
 279 static int __pg_init_all_paths(struct multipath *m)
 280 {
 281         struct pgpath *pgpath;
 282         unsigned long pg_init_delay = 0;
 283
 284         if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
 285                 return 0;
 286
 287         atomic_inc(&m->pg_init_count);
 288         clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 289
 290         /* Check here to reset pg_init_required */
 291         if (!m->current_pg)
 292                 return 0;
 293
 294         if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
 295                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 296                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 297         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 298                 /* Skip failed paths */
 299                 if (!pgpath->is_active)
 300                         continue;
 301                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 302                                        pg_init_delay))
 303                         atomic_inc(&m->pg_init_in_progress);
 304         }
 305         return atomic_read(&m->pg_init_in_progress);
 306 }
 307
 308 static int pg_init_all_paths(struct multipath *m)
 309 {
 310         int r;
 311         unsigned long flags;
 312
 313         spin_lock_irqsave(&m->lock, flags);
 314         r = __pg_init_all_paths(m);
 315         spin_unlock_irqrestore(&m->lock, flags);
 316
 317         return r;
 318 }
 319
 320 static void __switch_pg(struct multipath *m, struct priority_group *pg)
 321 {
 322         m->current_pg = pg;
 323
 324         /* Must we initialise the PG first, and queue I/O till it's ready? */
 325         if (m->hw_handler_name) {
 326                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 327                 set_bit(MPATHF_QUEUE_IO, &m->flags);
 328         } else {
 329                 clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 330                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 331         }
 332
 333         atomic_set(&m->pg_init_count, 0);
 334 }
 335
 336 static struct pgpath *choose_path_in_pg(struct multipath *m,
 337                                         struct priority_group *pg,
 338                                         size_t nr_bytes)
 339 {
 340         unsigned long flags;
 341         struct dm_path *path;
 342         struct pgpath *pgpath;
 343
 344         path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 345         if (!path)
 346                 return ERR_PTR(-ENXIO);
 347
 348         pgpath = path_to_pgpath(path);
 349
 350         if (unlikely(lockless_dereference(m->current_pg) != pg)) {
 351                 /* Only update current_pgpath if pg changed */
 352                 spin_lock_irqsave(&m->lock, flags);
 353                 m->current_pgpath = pgpath;
 354                 __switch_pg(m, pg);
 355                 spin_unlock_irqrestore(&m->lock, flags);
 356         }
 357
 358         return pgpath;
 359 }
 360
 361 static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
 362 {
 363         unsigned long flags;
 364         struct priority_group *pg;
 365         struct pgpath *pgpath;
 366         bool bypassed = true;
 367
 368         if (!atomic_read(&m->nr_valid_paths)) {
 369                 clear_bit(MPATHF_QUEUE_IO, &m->flags);
 370                 goto failed;
 371         }
 372
 373         /* Were we instructed to switch PG? */
 374         if (lockless_dereference(m->next_pg)) {
 375                 spin_lock_irqsave(&m->lock, flags);
 376                 pg = m->next_pg;
 377                 if (!pg) {
 378                         spin_unlock_irqrestore(&m->lock, flags);
 379                         goto check_current_pg;
 380                 }
 381                 m->next_pg = NULL;
 382                 spin_unlock_irqrestore(&m->lock, flags);
 383                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 384                 if (!IS_ERR_OR_NULL(pgpath))
 385                         return pgpath;
 386         }
 387
 388         /* Don't change PG until it has no remaining paths */
 389 check_current_pg:
 390         pg = lockless_dereference(m->current_pg);
 391         if (pg) {
 392                 pgpath = choose_path_in_pg(m, pg, nr_bytes);
 393                 if (!IS_ERR_OR_NULL(pgpath))
 394                         return pgpath;
 395         }
 396
 397         /*
 398          * Loop through priority groups until we find a valid path.
 399          * First time we skip PGs marked 'bypassed'.
 400          * Second time we only try the ones we skipped, but set
 401          * pg_init_delay_retry so we do not hammer controllers.
 402          */
 403         do {
 404                 list_for_each_entry(pg, &m->priority_groups, list) {
 405                         if (pg->bypassed == bypassed)
 406                                 continue;
 407                         pgpath = choose_path_in_pg(m, pg, nr_bytes);
 408                         if (!IS_ERR_OR_NULL(pgpath)) {
 409                                 if (!bypassed)
 410                                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
 411                                 return pgpath;
 412                         }
 413                 }
 414         } while (bypassed--);
 415
 416 failed:
 417         spin_lock_irqsave(&m->lock, flags);
 418         m->current_pgpath = NULL;
 419         m->current_pg = NULL;
 420         spin_unlock_irqrestore(&m->lock, flags);
 421
 422         return NULL;
 423 }
 424
 425 /*
 426  * Check whether bios must be queued in the device-mapper core rather
 427  * than here in the target.
 428  *
 429  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 430  * same value then we are not between multipath_presuspend()
 431  * and multipath_resume() calls and we have no need to check
 432  * for the DMF_NOFLUSH_SUSPENDING flag.
 433  */
 434 static int must_push_back(struct multipath *m)
 435 {
 436         return (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) ||
 437                 ((test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) !=
 438                   test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) &&
 439                  dm_noflush_suspending(m->ti)));
 440 }
 441
 442 /*
 443  * Map cloned requests
 444  */
 445 static int __multipath_map(struct dm_target *ti, struct request *clone,
 446                            union map_info *map_context,
 447                            struct request *rq, struct request **__clone)
 448 {
 449         struct multipath *m = ti->private;
 450         int r = DM_MAPIO_REQUEUE;
 451         size_t nr_bytes = clone ? blk_rq_bytes(clone) : blk_rq_bytes(rq);
 452         struct pgpath *pgpath;
 453         struct block_device *bdev;
 454         struct dm_mpath_io *mpio;
 455
 456         /* Do we need to select a new pgpath? */
 457         pgpath = lockless_dereference(m->current_pgpath);
 458         if (!pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
 459                 pgpath = choose_pgpath(m, nr_bytes);
 460
 461         if (!pgpath) {
 462                 if (!must_push_back(m))
 463                         r = -EIO;       /* Failed */
 464                 return r;
 465         } else if (test_bit(MPATHF_QUEUE_IO, &m->flags) ||
 466                    test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
 467                 pg_init_all_paths(m);
 468                 return r;
 469         }
 470
 471         mpio = set_mpio(m, map_context);
 472         if (!mpio)
 473                 /* ENOMEM, requeue */
 474                 return r;
 475
 476         mpio->pgpath = pgpath;
 477         mpio->nr_bytes = nr_bytes;
 478
 479         bdev = pgpath->path.dev->bdev;
 480
 481         if (clone) {
 482                 /*
 483                  * Old request-based interface: allocated clone is passed in.
 484                  * Used by: .request_fn stacked on .request_fn path(s).
 485                  */
 486                 clone->q = bdev_get_queue(bdev);
 487                 clone->rq_disk = bdev->bd_disk;
 488                 clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 489         } else {
 490                 /*
 491                  * blk-mq request-based interface; used by both:
 492                  * .request_fn stacked on blk-mq path(s) and
 493                  * blk-mq stacked on blk-mq path(s).
 494                  */
 495                 *__clone = blk_mq_alloc_request(bdev_get_queue(bdev),
 496                                                 rq_data_dir(rq), BLK_MQ_REQ_NOWAIT);
 497                 if (IS_ERR(*__clone)) {
 498                         /* ENOMEM, requeue */
 499                         clear_request_fn_mpio(m, map_context);
 500                         return r;
 501                 }
 502                 (*__clone)->bio = (*__clone)->biotail = NULL;
 503                 (*__clone)->rq_disk = bdev->bd_disk;
 504                 (*__clone)->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 505         }
 506
 507         if (pgpath->pg->ps.type->start_io)
 508                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 509                                               &pgpath->path,
 510                                               nr_bytes);
 511         return DM_MAPIO_REMAPPED;
 512 }
 513
 514 static int multipath_map(struct dm_target *ti, struct request *clone,
 515                          union map_info *map_context)
 516 {
 517         return __multipath_map(ti, clone, map_context, NULL, NULL);
 518 }
 519
 520 static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 521                                    union map_info *map_context,
 522                                    struct request **clone)
 523 {
 524         return __multipath_map(ti, NULL, map_context, rq, clone);
 525 }
 526
 527 static void multipath_release_clone(struct request *clone)
 528 {
 529         blk_mq_free_request(clone);
 530 }
 531
 532 /*
 533  * If we run out of usable paths, should we queue I/O or error it?
 534  */
 535 static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
 536                             bool save_old_value)
 537 {
 538         unsigned long flags;
 539
 540         spin_lock_irqsave(&m->lock, flags);
 541
 542         if (save_old_value) {
 543                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
 544                         set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 545                 else
 546                         clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 547         } else {
 548                 if (queue_if_no_path)
 549                         set_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 550                 else
 551                         clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 552         }
 553         if (queue_if_no_path)
 554                 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 555         else
 556                 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 557
 558         spin_unlock_irqrestore(&m->lock, flags);
 559
 560         if (!queue_if_no_path)
 561                 dm_table_run_md_queue_async(m->ti->table);
 562
 563         return 0;
 564 }
 565
 566 /*
 567  * An event is triggered whenever a path is taken out of use.
 568  * Includes path failure and PG bypass.
 569  */
 570 static void trigger_event(struct work_struct *work)
 571 {
 572         struct multipath *m =
 573                 container_of(work, struct multipath, trigger_event);
 574
 575         dm_table_event(m->ti->table);
 576 }
 577
 578 /*-----------------------------------------------------------------
 579  * Constructor/argument parsing:
 580  * <#multipath feature args> [<arg>]*
 581  * <#hw_handler args> [hw_handler [<arg>]*]
 582  * <#priority groups>
 583  * <initial priority group>
 584  *     [<selector> <#selector args> [<arg>]*
 585  *      <#paths> <#per-path selector args>
 586  *         [<path> [<arg>]* ]+ ]+
 587  *---------------------------------------------------------------*/
 588 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 589                                struct dm_target *ti)
 590 {
 591         int r;
 592         struct path_selector_type *pst;
 593         unsigned ps_argc;
 594
 595         static struct dm_arg _args[] = {
 596                 {0, 1024, "invalid number of path selector args"},
 597         };
 598
 599         pst = dm_get_path_selector(dm_shift_arg(as));
 600         if (!pst) {
 601                 ti->error = "unknown path selector type";
 602                 return -EINVAL;
 603         }
 604
 605         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 606         if (r) {
 607                 dm_put_path_selector(pst);
 608                 return -EINVAL;
 609         }
 610
 611         r = pst->create(&pg->ps, ps_argc, as->argv);
 612         if (r) {
 613                 dm_put_path_selector(pst);
 614                 ti->error = "path selector constructor failed";
 615                 return r;
 616         }
 617
 618         pg->ps.type = pst;
 619         dm_consume_args(as, ps_argc);
 620
 621         return 0;
 622 }
 623
 624 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 625                                struct dm_target *ti)
 626 {
 627         int r;
 628         struct pgpath *p;
 629         struct multipath *m = ti->private;
 630         struct request_queue *q = NULL;
 631         const char *attached_handler_name;
 632
 633         /* we need at least a path arg */
 634         if (as->argc < 1) {
 635                 ti->error = "no device given";
 636                 return ERR_PTR(-EINVAL);
 637         }
 638
 639         p = alloc_pgpath();
 640         if (!p)
 641                 return ERR_PTR(-ENOMEM);
 642
 643         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 644                           &p->path.dev);
 645         if (r) {
 646                 ti->error = "error getting device";
 647                 goto bad;
 648         }
 649
 650         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) || m->hw_handler_name)
 651                 q = bdev_get_queue(p->path.dev->bdev);
 652
 653         if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags)) {
 654 retain:
 655                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 656                 if (attached_handler_name) {
 657                         /*
 658                          * Reset hw_handler_name to match the attached handler
 659                          * and clear any hw_handler_params associated with the
 660                          * ignored handler.
 661                          *
 662                          * NB. This modifies the table line to show the actual
 663                          * handler instead of the original table passed in.
 664                          */
 665                         kfree(m->hw_handler_name);
 666                         m->hw_handler_name = attached_handler_name;
 667
 668                         kfree(m->hw_handler_params);
 669                         m->hw_handler_params = NULL;
 670                 }
 671         }
 672
 673         if (m->hw_handler_name) {
 674                 r = scsi_dh_attach(q, m->hw_handler_name);
 675                 if (r == -EBUSY) {
 676                         char b[BDEVNAME_SIZE];
 677
 678                         printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
 679                                 bdevname(p->path.dev->bdev, b));
 680                         goto retain;
 681                 }
 682                 if (r < 0) {
 683                         ti->error = "error attaching hardware handler";
 684                         dm_put_device(ti, p->path.dev);
 685                         goto bad;
 686                 }
 687
 688                 if (m->hw_handler_params) {
 689                         r = scsi_dh_set_params(q, m->hw_handler_params);
 690                         if (r < 0) {
 691                                 ti->error = "unable to set hardware "
 692                                                         "handler parameters";
 693                                 dm_put_device(ti, p->path.dev);
 694                                 goto bad;
 695                         }
 696                 }
 697         }
 698
 699         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 700         if (r) {
 701                 dm_put_device(ti, p->path.dev);
 702                 goto bad;
 703         }
 704
 705         return p;
 706
 707  bad:
 708         free_pgpath(p);
 709         return ERR_PTR(r);
 710 }
 711
 712 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 713                                                    struct multipath *m)
 714 {
 715         static struct dm_arg _args[] = {
 716                 {1, 1024, "invalid number of paths"},
 717                 {0, 1024, "invalid number of selector args"}
 718         };
 719
 720         int r;
 721         unsigned i, nr_selector_args, nr_args;
 722         struct priority_group *pg;
 723         struct dm_target *ti = m->ti;
 724
 725         if (as->argc < 2) {
 726                 as->argc = 0;
 727                 ti->error = "not enough priority group arguments";
 728                 return ERR_PTR(-EINVAL);
 729         }
 730
 731         pg = alloc_priority_group();
 732         if (!pg) {
 733                 ti->error = "couldn't allocate priority group";
 734                 return ERR_PTR(-ENOMEM);
 735         }
 736         pg->m = m;
 737
 738         r = parse_path_selector(as, pg, ti);
 739         if (r)
 740                 goto bad;
 741
 742         /*
 743          * read the paths
 744          */
 745         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 746         if (r)
 747                 goto bad;
 748
 749         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 750         if (r)
 751                 goto bad;
 752
 753         nr_args = 1 + nr_selector_args;
 754         for (i = 0; i < pg->nr_pgpaths; i++) {
 755                 struct pgpath *pgpath;
 756                 struct dm_arg_set path_args;
 757
 758                 if (as->argc < nr_args) {
 759                         ti->error = "not enough path parameters";
 760                         r = -EINVAL;
 761                         goto bad;
 762                 }
 763
 764                 path_args.argc = nr_args;
 765                 path_args.argv = as->argv;
 766
 767                 pgpath = parse_path(&path_args, &pg->ps, ti);
 768                 if (IS_ERR(pgpath)) {
 769                         r = PTR_ERR(pgpath);
 770                         goto bad;
 771                 }
 772
 773                 pgpath->pg = pg;
 774                 list_add_tail(&pgpath->list, &pg->pgpaths);
 775                 dm_consume_args(as, nr_args);
 776         }
 777
 778         return pg;
 779
 780  bad:
 781         free_priority_group(pg, ti);
 782         return ERR_PTR(r);
 783 }
 784
 785 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 786 {
 787         unsigned hw_argc;
 788         int ret;
 789         struct dm_target *ti = m->ti;
 790
 791         static struct dm_arg _args[] = {
 792                 {0, 1024, "invalid number of hardware handler args"},
 793         };
 794
 795         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 796                 return -EINVAL;
 797
 798         if (!hw_argc)
 799                 return 0;
 800
 801         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 802
 803         if (hw_argc > 1) {
 804                 char *p;
 805                 int i, j, len = 4;
 806
 807                 for (i = 0; i <= hw_argc - 2; i++)
 808                         len += strlen(as->argv[i]) + 1;
 809                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 810                 if (!p) {
 811                         ti->error = "memory allocation failed";
 812                         ret = -ENOMEM;
 813                         goto fail;
 814                 }
 815                 j = sprintf(p, "%d", hw_argc - 1);
 816                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 817                         j = sprintf(p, "%s", as->argv[i]);
 818         }
 819         dm_consume_args(as, hw_argc - 1);
 820
 821         return 0;
 822 fail:
 823         kfree(m->hw_handler_name);
 824         m->hw_handler_name = NULL;
 825         return ret;
 826 }
 827
 828 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 829 {
 830         int r;
 831         unsigned argc;
 832         struct dm_target *ti = m->ti;
 833         const char *arg_name;
 834
 835         static struct dm_arg _args[] = {
 836                 {0, 6, "invalid number of feature args"},
 837                 {1, 50, "pg_init_retries must be between 1 and 50"},
 838                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 839         };
 840
 841         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 842         if (r)
 843                 return -EINVAL;
 844
 845         if (!argc)
 846                 return 0;
 847
 848         do {
 849                 arg_name = dm_shift_arg(as);
 850                 argc--;
 851
 852                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 853                         r = queue_if_no_path(m, true, false);
 854                         continue;
 855                 }
 856
 857                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
 858                         set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 859                         continue;
 860                 }
 861
 862                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 863                     (argc >= 1)) {
 864                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 865                         argc--;
 866                         continue;
 867                 }
 868
 869                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 870                     (argc >= 1)) {
 871                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 872                         argc--;
 873                         continue;
 874                 }
 875
 876                 ti->error = "Unrecognised multipath feature request";
 877                 r = -EINVAL;
 878         } while (argc && !r);
 879
 880         return r;
 881 }
 882
 883 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 884                          char **argv)
 885 {
 886         /* target arguments */
 887         static struct dm_arg _args[] = {
 888                 {0, 1024, "invalid number of priority groups"},
 889                 {0, 1024, "invalid initial priority group number"},
 890         };
 891
 892         int r;
 893         struct multipath *m;
 894         struct dm_arg_set as;
 895         unsigned pg_count = 0;
 896         unsigned next_pg_num;
 897         bool use_blk_mq = dm_use_blk_mq(dm_table_get_md(ti->table));
 898
 899         as.argc = argc;
 900         as.argv = argv;
 901
 902         m = alloc_multipath(ti, use_blk_mq);
 903         if (!m) {
 904                 ti->error = "can't allocate multipath";
 905                 return -EINVAL;
 906         }
 907
 908         r = parse_features(&as, m);
 909         if (r)
 910                 goto bad;
 911
 912         r = parse_hw_handler(&as, m);
 913         if (r)
 914                 goto bad;
 915
 916         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 917         if (r)
 918                 goto bad;
 919
 920         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 921         if (r)
 922                 goto bad;
 923
 924         if ((!m->nr_priority_groups && next_pg_num) ||
 925             (m->nr_priority_groups && !next_pg_num)) {
 926                 ti->error = "invalid initial priority group";
 927                 r = -EINVAL;
 928                 goto bad;
 929         }
 930
 931         /* parse the priority groups */
 932         while (as.argc) {
 933                 struct priority_group *pg;
 934                 unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
 935
 936                 pg = parse_priority_group(&as, m);
 937                 if (IS_ERR(pg)) {
 938                         r = PTR_ERR(pg);
 939                         goto bad;
 940                 }
 941
 942                 nr_valid_paths += pg->nr_pgpaths;
 943                 atomic_set(&m->nr_valid_paths, nr_valid_paths);
 944
 945                 list_add_tail(&pg->list, &m->priority_groups);
 946                 pg_count++;
 947                 pg->pg_num = pg_count;
 948                 if (!--next_pg_num)
 949                         m->next_pg = pg;
 950         }
 951
 952         if (pg_count != m->nr_priority_groups) {
 953                 ti->error = "priority group count mismatch";
 954                 r = -EINVAL;
 955                 goto bad;
 956         }
 957
 958         ti->num_flush_bios = 1;
 959         ti->num_discard_bios = 1;
 960         ti->num_write_same_bios = 1;
 961         if (use_blk_mq)
 962                 ti->per_io_data_size = sizeof(struct dm_mpath_io);
 963
 964         return 0;
 965
 966  bad:
 967         free_multipath(m);
 968         return r;
 969 }
 970
 971 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 972 {
 973         DECLARE_WAITQUEUE(wait, current);
 974
 975         add_wait_queue(&m->pg_init_wait, &wait);
 976
 977         while (1) {
 978                 set_current_state(TASK_UNINTERRUPTIBLE);
 979
 980                 if (!atomic_read(&m->pg_init_in_progress))
 981                         break;
 982
 983                 io_schedule();
 984         }
 985         set_current_state(TASK_RUNNING);
 986
 987         remove_wait_queue(&m->pg_init_wait, &wait);
 988 }
 989
 990 static void flush_multipath_work(struct multipath *m)
 991 {
 992         set_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
 993         smp_mb__after_atomic();
 994
 995         flush_workqueue(kmpath_handlerd);
 996         multipath_wait_for_pg_init_completion(m);
 997         flush_workqueue(kmultipathd);
 998         flush_work(&m->trigger_event);
 999
1000         clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1001         smp_mb__after_atomic();
1002 }
1003
1004 static void multipath_dtr(struct dm_target *ti)
1005 {
1006         struct multipath *m = ti->private;
1007
1008         flush_multipath_work(m);
1009         free_multipath(m);
1010 }
1011
1012 /*
1013  * Take a path out of use.
1014  */
1015 static int fail_path(struct pgpath *pgpath)
1016 {
1017         unsigned long flags;
1018         struct multipath *m = pgpath->pg->m;
1019
1020         spin_lock_irqsave(&m->lock, flags);
1021
1022         if (!pgpath->is_active)
1023                 goto out;
1024
1025         DMWARN("Failing path %s.", pgpath->path.dev->name);
1026
1027         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1028         pgpath->is_active = false;
1029         pgpath->fail_count++;
1030
1031         atomic_dec(&m->nr_valid_paths);
1032
1033         if (pgpath == m->current_pgpath)
1034                 m->current_pgpath = NULL;
1035
1036         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1037                        pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1038
1039         schedule_work(&m->trigger_event);
1040
1041 out:
1042         spin_unlock_irqrestore(&m->lock, flags);
1043
1044         return 0;
1045 }
1046
1047 /*
1048  * Reinstate a previously-failed path
1049  */
1050 static int reinstate_path(struct pgpath *pgpath)
1051 {
1052         int r = 0, run_queue = 0;
1053         unsigned long flags;
1054         struct multipath *m = pgpath->pg->m;
1055         unsigned nr_valid_paths;
1056
1057         spin_lock_irqsave(&m->lock, flags);
1058
1059         if (pgpath->is_active)
1060                 goto out;
1061
1062         DMWARN("Reinstating path %s.", pgpath->path.dev->name);
1063
1064         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1065         if (r)
1066                 goto out;
1067
1068         pgpath->is_active = true;
1069
1070         nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1071         if (nr_valid_paths == 1) {
1072                 m->current_pgpath = NULL;
1073                 run_queue = 1;
1074         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1075                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1076                         atomic_inc(&m->pg_init_in_progress);
1077         }
1078
1079         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1080                        pgpath->path.dev->name, nr_valid_paths);
1081
1082         schedule_work(&m->trigger_event);
1083
1084 out:
1085         spin_unlock_irqrestore(&m->lock, flags);
1086         if (run_queue)
1087                 dm_table_run_md_queue_async(m->ti->table);
1088
1089         return r;
1090 }
1091
1092 /*
1093  * Fail or reinstate all paths that match the provided struct dm_dev.
1094  */
1095 static int action_dev(struct multipath *m, struct dm_dev *dev,
1096                       action_fn action)
1097 {
1098         int r = -EINVAL;
1099         struct pgpath *pgpath;
1100         struct priority_group *pg;
1101
1102         list_for_each_entry(pg, &m->priority_groups, list) {
1103                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1104                         if (pgpath->path.dev == dev)
1105                                 r = action(pgpath);
1106                 }
1107         }
1108
1109         return r;
1110 }
1111
1112 /*
1113  * Temporarily try to avoid having to use the specified PG
1114  */
1115 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1116                       bool bypassed)
1117 {
1118         unsigned long flags;
1119
1120         spin_lock_irqsave(&m->lock, flags);
1121
1122         pg->bypassed = bypassed;
1123         m->current_pgpath = NULL;
1124         m->current_pg = NULL;
1125
1126         spin_unlock_irqrestore(&m->lock, flags);
1127
1128         schedule_work(&m->trigger_event);
1129 }
1130
1131 /*
1132  * Switch to using the specified PG from the next I/O that gets mapped
1133  */
1134 static int switch_pg_num(struct multipath *m, const char *pgstr)
1135 {
1136         struct priority_group *pg;
1137         unsigned pgnum;
1138         unsigned long flags;
1139         char dummy;
1140
1141         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1142             (pgnum > m->nr_priority_groups)) {
1143                 DMWARN("invalid PG number supplied to switch_pg_num");
1144                 return -EINVAL;
1145         }
1146
1147         spin_lock_irqsave(&m->lock, flags);
1148         list_for_each_entry(pg, &m->priority_groups, list) {
1149                 pg->bypassed = false;
1150                 if (--pgnum)
1151                         continue;
1152
1153                 m->current_pgpath = NULL;
1154                 m->current_pg = NULL;
1155                 m->next_pg = pg;
1156         }
1157         spin_unlock_irqrestore(&m->lock, flags);
1158
1159         schedule_work(&m->trigger_event);
1160         return 0;
1161 }
1162
1163 /*
1164  * Set/clear bypassed status of a PG.
1165  * PGs are numbered upwards from 1 in the order they were declared.
1166  */
1167 static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1168 {
1169         struct priority_group *pg;
1170         unsigned pgnum;
1171         char dummy;
1172
1173         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1174             (pgnum > m->nr_priority_groups)) {
1175                 DMWARN("invalid PG number supplied to bypass_pg");
1176                 return -EINVAL;
1177         }
1178
1179         list_for_each_entry(pg, &m->priority_groups, list) {
1180                 if (!--pgnum)
1181                         break;
1182         }
1183
1184         bypass_pg(m, pg, bypassed);
1185         return 0;
1186 }
1187
1188 /*
1189  * Should we retry pg_init immediately?
1190  */
1191 static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1192 {
1193         unsigned long flags;
1194         bool limit_reached = false;
1195
1196         spin_lock_irqsave(&m->lock, flags);
1197
1198         if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1199             !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1200                 set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1201         else
1202                 limit_reached = true;
1203
1204         spin_unlock_irqrestore(&m->lock, flags);
1205
1206         return limit_reached;
1207 }
1208
1209 static void pg_init_done(void *data, int errors)
1210 {
1211         struct pgpath *pgpath = data;
1212         struct priority_group *pg = pgpath->pg;
1213         struct multipath *m = pg->m;
1214         unsigned long flags;
1215         bool delay_retry = false;
1216
1217         /* device or driver problems */
1218         switch (errors) {
1219         case SCSI_DH_OK:
1220                 break;
1221         case SCSI_DH_NOSYS:
1222                 if (!m->hw_handler_name) {
1223                         errors = 0;
1224                         break;
1225                 }
1226                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1227                       "Error %d.", m->hw_handler_name, errors);
1228                 /*
1229                  * Fail path for now, so we do not ping pong
1230                  */
1231                 fail_path(pgpath);
1232                 break;
1233         case SCSI_DH_DEV_TEMP_BUSY:
1234                 /*
1235                  * Probably doing something like FW upgrade on the
1236                  * controller so try the other pg.
1237                  */
1238                 bypass_pg(m, pg, true);
1239                 break;
1240         case SCSI_DH_RETRY:
1241                 /* Wait before retrying. */
1242                 delay_retry = 1;
1243         case SCSI_DH_IMM_RETRY:
1244         case SCSI_DH_RES_TEMP_UNAVAIL:
1245                 if (pg_init_limit_reached(m, pgpath))
1246                         fail_path(pgpath);
1247                 errors = 0;
1248                 break;
1249         case SCSI_DH_DEV_OFFLINED:
1250         default:
1251                 /*
1252                  * We probably do not want to fail the path for a device
1253                  * error, but this is what the old dm did. In future
1254                  * patches we can do more advanced handling.
1255                  */
1256                 fail_path(pgpath);
1257         }
1258
1259         spin_lock_irqsave(&m->lock, flags);
1260         if (errors) {
1261                 if (pgpath == m->current_pgpath) {
1262                         DMERR("Could not failover device. Error %d.", errors);
1263                         m->current_pgpath = NULL;
1264                         m->current_pg = NULL;
1265                 }
1266         } else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1267                 pg->bypassed = false;
1268
1269         if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1270                 /* Activations of other paths are still on going */
1271                 goto out;
1272
1273         if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1274                 if (delay_retry)
1275                         set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1276                 else
1277                         clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1278
1279                 if (__pg_init_all_paths(m))
1280                         goto out;
1281         }
1282         clear_bit(MPATHF_QUEUE_IO, &m->flags);
1283
1284         /*
1285          * Wake up any thread waiting to suspend.
1286          */
1287         wake_up(&m->pg_init_wait);
1288
1289 out:
1290         spin_unlock_irqrestore(&m->lock, flags);
1291 }
1292
1293 static void activate_path(struct work_struct *work)
1294 {
1295         struct pgpath *pgpath =
1296                 container_of(work, struct pgpath, activate_path.work);
1297
1298         if (pgpath->is_active)
1299                 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1300                                  pg_init_done, pgpath);
1301         else
1302                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1303 }
1304
1305 static int noretry_error(int error)
1306 {
1307         switch (error) {
1308         case -EOPNOTSUPP:
1309         case -EREMOTEIO:
1310         case -EILSEQ:
1311         case -ENODATA:
1312         case -ENOSPC:
1313                 return 1;
1314         }
1315
1316         /* Anything else could be a path failure, so should be retried */
1317         return 0;
1318 }
1319
1320 /*
1321  * end_io handling
1322  */
1323 static int do_end_io(struct multipath *m, struct request *clone,
1324                      int error, struct dm_mpath_io *mpio)
1325 {
1326         /*
1327          * We don't queue any clone request inside the multipath target
1328          * during end I/O handling, since those clone requests don't have
1329          * bio clones.  If we queue them inside the multipath target,
1330          * we need to make bio clones, that requires memory allocation.
1331          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1332          *  don't have bio clones.)
1333          * Instead of queueing the clone request here, we queue the original
1334          * request into dm core, which will remake a clone request and
1335          * clone bios for it and resubmit it later.
1336          */
1337         int r = DM_ENDIO_REQUEUE;
1338
1339         if (!error && !clone->errors)
1340                 return 0;       /* I/O complete */
1341
1342         if (noretry_error(error))
1343                 return error;
1344
1345         if (mpio->pgpath)
1346                 fail_path(mpio->pgpath);
1347
1348         if (!atomic_read(&m->nr_valid_paths)) {
1349                 if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1350                         if (!must_push_back(m))
1351                                 r = -EIO;
1352                 } else {
1353                         if (error == -EBADE)
1354                                 r = error;
1355                 }
1356         }
1357
1358         return r;
1359 }
1360
1361 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1362                             int error, union map_info *map_context)
1363 {
1364         struct multipath *m = ti->private;
1365         struct dm_mpath_io *mpio = get_mpio(map_context);
1366         struct pgpath *pgpath;
1367         struct path_selector *ps;
1368         int r;
1369
1370         BUG_ON(!mpio);
1371
1372         r = do_end_io(m, clone, error, mpio);
1373         pgpath = mpio->pgpath;
1374         if (pgpath) {
1375                 ps = &pgpath->pg->ps;
1376                 if (ps->type->end_io)
1377                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1378         }
1379         clear_request_fn_mpio(m, map_context);
1380
1381         return r;
1382 }
1383
1384 /*
1385  * Suspend can't complete until all the I/O is processed so if
1386  * the last path fails we must error any remaining I/O.
1387  * Note that if the freeze_bdev fails while suspending, the
1388  * queue_if_no_path state is lost - userspace should reset it.
1389  */
1390 static void multipath_presuspend(struct dm_target *ti)
1391 {
1392         struct multipath *m = ti->private;
1393
1394         queue_if_no_path(m, false, true);
1395 }
1396
1397 static void multipath_postsuspend(struct dm_target *ti)
1398 {
1399         struct multipath *m = ti->private;
1400
1401         mutex_lock(&m->work_mutex);
1402         flush_multipath_work(m);
1403         mutex_unlock(&m->work_mutex);
1404 }
1405
1406 /*
1407  * Restore the queue_if_no_path setting.
1408  */
1409 static void multipath_resume(struct dm_target *ti)
1410 {
1411         struct multipath *m = ti->private;
1412
1413         if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags))
1414                 set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1415         else
1416                 clear_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1417         smp_mb__after_atomic();
1418 }
1419
1420 /*
1421  * Info output has the following format:
1422  * num_multipath_feature_args [multipath_feature_args]*
1423  * num_handler_status_args [handler_status_args]*
1424  * num_groups init_group_number
1425  *            [A|D|E num_ps_status_args [ps_status_args]*
1426  *             num_paths num_selector_args
1427  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1428  *
1429  * Table output has the following format (identical to the constructor string):
1430  * num_feature_args [features_args]*
1431  * num_handler_args hw_handler [hw_handler_args]*
1432  * num_groups init_group_number
1433  *     [priority selector-name num_ps_args [ps_args]*
1434  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1435  */
1436 static void multipath_status(struct dm_target *ti, status_type_t type,
1437                              unsigned status_flags, char *result, unsigned maxlen)
1438 {
1439         int sz = 0;
1440         unsigned long flags;
1441         struct multipath *m = ti->private;
1442         struct priority_group *pg;
1443         struct pgpath *p;
1444         unsigned pg_num;
1445         char state;
1446
1447         spin_lock_irqsave(&m->lock, flags);
1448
1449         /* Features */
1450         if (type == STATUSTYPE_INFO)
1451                 DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1452                        atomic_read(&m->pg_init_count));
1453         else {
1454                 DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1455                               (m->pg_init_retries > 0) * 2 +
1456                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1457                               test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags));
1458                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1459                         DMEMIT("queue_if_no_path ");
1460                 if (m->pg_init_retries)
1461                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1462                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1463                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1464                 if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1465                         DMEMIT("retain_attached_hw_handler ");
1466         }
1467
1468         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1469                 DMEMIT("0 ");
1470         else
1471                 DMEMIT("1 %s ", m->hw_handler_name);
1472
1473         DMEMIT("%u ", m->nr_priority_groups);
1474
1475         if (m->next_pg)
1476                 pg_num = m->next_pg->pg_num;
1477         else if (m->current_pg)
1478                 pg_num = m->current_pg->pg_num;
1479         else
1480                 pg_num = (m->nr_priority_groups ? 1 : 0);
1481
1482         DMEMIT("%u ", pg_num);
1483
1484         switch (type) {
1485         case STATUSTYPE_INFO:
1486                 list_for_each_entry(pg, &m->priority_groups, list) {
1487                         if (pg->bypassed)
1488                                 state = 'D';    /* Disabled */
1489                         else if (pg == m->current_pg)
1490                                 state = 'A';    /* Currently Active */
1491                         else
1492                                 state = 'E';    /* Enabled */
1493
1494                         DMEMIT("%c ", state);
1495
1496                         if (pg->ps.type->status)
1497                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1498                                                           result + sz,
1499                                                           maxlen - sz);
1500                         else
1501                                 DMEMIT("0 ");
1502
1503                         DMEMIT("%u %u ", pg->nr_pgpaths,
1504                                pg->ps.type->info_args);
1505
1506                         list_for_each_entry(p, &pg->pgpaths, list) {
1507                                 DMEMIT("%s %s %u ", p->path.dev->name,
1508                                        p->is_active ? "A" : "F",
1509                                        p->fail_count);
1510                                 if (pg->ps.type->status)
1511                                         sz += pg->ps.type->status(&pg->ps,
1512                                               &p->path, type, result + sz,
1513                                               maxlen - sz);
1514                         }
1515                 }
1516                 break;
1517
1518         case STATUSTYPE_TABLE:
1519                 list_for_each_entry(pg, &m->priority_groups, list) {
1520                         DMEMIT("%s ", pg->ps.type->name);
1521
1522                         if (pg->ps.type->status)
1523                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1524                                                           result + sz,
1525                                                           maxlen - sz);
1526                         else
1527                                 DMEMIT("0 ");
1528
1529                         DMEMIT("%u %u ", pg->nr_pgpaths,
1530                                pg->ps.type->table_args);
1531
1532                         list_for_each_entry(p, &pg->pgpaths, list) {
1533                                 DMEMIT("%s ", p->path.dev->name);
1534                                 if (pg->ps.type->status)
1535                                         sz += pg->ps.type->status(&pg->ps,
1536                                               &p->path, type, result + sz,
1537                                               maxlen - sz);
1538                         }
1539                 }
1540                 break;
1541         }
1542
1543         spin_unlock_irqrestore(&m->lock, flags);
1544 }
1545
1546 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1547 {
1548         int r = -EINVAL;
1549         struct dm_dev *dev;
1550         struct multipath *m = ti->private;
1551         action_fn action;
1552
1553         mutex_lock(&m->work_mutex);
1554
1555         if (dm_suspended(ti)) {
1556                 r = -EBUSY;
1557                 goto out;
1558         }
1559
1560         if (argc == 1) {
1561                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1562                         r = queue_if_no_path(m, true, false);
1563                         goto out;
1564                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1565                         r = queue_if_no_path(m, false, false);
1566                         goto out;
1567                 }
1568         }
1569
1570         if (argc != 2) {
1571                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1572                 goto out;
1573         }
1574
1575         if (!strcasecmp(argv[0], "disable_group")) {
1576                 r = bypass_pg_num(m, argv[1], true);
1577                 goto out;
1578         } else if (!strcasecmp(argv[0], "enable_group")) {
1579                 r = bypass_pg_num(m, argv[1], false);
1580                 goto out;
1581         } else if (!strcasecmp(argv[0], "switch_group")) {
1582                 r = switch_pg_num(m, argv[1]);
1583                 goto out;
1584         } else if (!strcasecmp(argv[0], "reinstate_path"))
1585                 action = reinstate_path;
1586         else if (!strcasecmp(argv[0], "fail_path"))
1587                 action = fail_path;
1588         else {
1589                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1590                 goto out;
1591         }
1592
1593         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1594         if (r) {
1595                 DMWARN("message: error getting device %s",
1596                        argv[1]);
1597                 goto out;
1598         }
1599
1600         r = action_dev(m, dev, action);
1601
1602         dm_put_device(ti, dev);
1603
1604 out:
1605         mutex_unlock(&m->work_mutex);
1606         return r;
1607 }
1608
1609 static int multipath_prepare_ioctl(struct dm_target *ti,
1610                 struct block_device **bdev, fmode_t *mode)
1611 {
1612         struct multipath *m = ti->private;
1613         struct pgpath *current_pgpath;
1614         int r;
1615
1616         current_pgpath = lockless_dereference(m->current_pgpath);
1617         if (!current_pgpath)
1618                 current_pgpath = choose_pgpath(m, 0);
1619
1620         if (current_pgpath) {
1621                 if (!test_bit(MPATHF_QUEUE_IO, &m->flags)) {
1622                         *bdev = current_pgpath->path.dev->bdev;
1623                         *mode = current_pgpath->path.dev->mode;
1624                         r = 0;
1625                 } else {
1626                         /* pg_init has not started or completed */
1627                         r = -ENOTCONN;
1628                 }
1629         } else {
1630                 /* No path is available */
1631                 if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1632                         r = -ENOTCONN;
1633                 else
1634                         r = -EIO;
1635         }
1636
1637         if (r == -ENOTCONN) {
1638                 if (!lockless_dereference(m->current_pg)) {
1639                         /* Path status changed, redo selection */
1640                         (void) choose_pgpath(m, 0);
1641                 }
1642                 if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1643                         pg_init_all_paths(m);
1644                 dm_table_run_md_queue_async(m->ti->table);
1645         }
1646
1647         /*
1648          * Only pass ioctls through if the device sizes match exactly.
1649          */
1650         if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
1651                 return 1;
1652         return r;
1653 }
1654
1655 static int multipath_iterate_devices(struct dm_target *ti,
1656                                      iterate_devices_callout_fn fn, void *data)
1657 {
1658         struct multipath *m = ti->private;
1659         struct priority_group *pg;
1660         struct pgpath *p;
1661         int ret = 0;
1662
1663         list_for_each_entry(pg, &m->priority_groups, list) {
1664                 list_for_each_entry(p, &pg->pgpaths, list) {
1665                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1666                         if (ret)
1667                                 goto out;
1668                 }
1669         }
1670
1671 out:
1672         return ret;
1673 }
1674
1675 static int pgpath_busy(struct pgpath *pgpath)
1676 {
1677         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1678
1679         return blk_lld_busy(q);
1680 }
1681
1682 /*
1683  * We return "busy", only when we can map I/Os but underlying devices
1684  * are busy (so even if we map I/Os now, the I/Os will wait on
1685  * the underlying queue).
1686  * In other words, if we want to kill I/Os or queue them inside us
1687  * due to map unavailability, we don't return "busy".  Otherwise,
1688  * dm core won't give us the I/Os and we can't do what we want.
1689  */
1690 static int multipath_busy(struct dm_target *ti)
1691 {
1692         bool busy = false, has_active = false;
1693         struct multipath *m = ti->private;
1694         struct priority_group *pg, *next_pg;
1695         struct pgpath *pgpath;
1696
1697         /* pg_init in progress or no paths available */
1698         if (atomic_read(&m->pg_init_in_progress) ||
1699             (!atomic_read(&m->nr_valid_paths) && test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)))
1700                 return true;
1701
1702         /* Guess which priority_group will be used at next mapping time */
1703         pg = lockless_dereference(m->current_pg);
1704         next_pg = lockless_dereference(m->next_pg);
1705         if (unlikely(!lockless_dereference(m->current_pgpath) && next_pg))
1706                 pg = next_pg;
1707
1708         if (!pg) {
1709                 /*
1710                  * We don't know which pg will be used at next mapping time.
1711                  * We don't call choose_pgpath() here to avoid to trigger
1712                  * pg_init just by busy checking.
1713                  * So we don't know whether underlying devices we will be using
1714                  * at next mapping time are busy or not. Just try mapping.
1715                  */
1716                 return busy;
1717         }
1718
1719         /*
1720          * If there is one non-busy active path at least, the path selector
1721          * will be able to select it. So we consider such a pg as not busy.
1722          */
1723         busy = true;
1724         list_for_each_entry(pgpath, &pg->pgpaths, list) {
1725                 if (pgpath->is_active) {
1726                         has_active = true;
1727                         if (!pgpath_busy(pgpath)) {
1728                                 busy = false;
1729                                 break;
1730                         }
1731                 }
1732         }
1733
1734         if (!has_active) {
1735                 /*
1736                  * No active path in this pg, so this pg won't be used and
1737                  * the current_pg will be changed at next mapping time.
1738                  * We need to try mapping to determine it.
1739                  */
1740                 busy = false;
1741         }
1742
1743         return busy;
1744 }
1745
1746 /*-----------------------------------------------------------------
1747  * Module setup
1748  *---------------------------------------------------------------*/
1749 static struct target_type multipath_target = {
1750         .name = "multipath",
1751         .version = {1, 11, 0},
1752         .features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE,
1753         .module = THIS_MODULE,
1754         .ctr = multipath_ctr,
1755         .dtr = multipath_dtr,
1756         .map_rq = multipath_map,
1757         .clone_and_map_rq = multipath_clone_and_map,
1758         .release_clone_rq = multipath_release_clone,
1759         .rq_end_io = multipath_end_io,
1760         .presuspend = multipath_presuspend,
1761         .postsuspend = multipath_postsuspend,
1762         .resume = multipath_resume,
1763         .status = multipath_status,
1764         .message = multipath_message,
1765         .prepare_ioctl = multipath_prepare_ioctl,
1766         .iterate_devices = multipath_iterate_devices,
1767         .busy = multipath_busy,
1768 };
1769
1770 static int __init dm_multipath_init(void)
1771 {
1772         int r;
1773
1774         /* allocate a slab for the dm_ios */
1775         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1776         if (!_mpio_cache)
1777                 return -ENOMEM;
1778
1779         r = dm_register_target(&multipath_target);
1780         if (r < 0) {
1781                 DMERR("register failed %d", r);
1782                 r = -EINVAL;
1783                 goto bad_register_target;
1784         }
1785
1786         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1787         if (!kmultipathd) {
1788                 DMERR("failed to create workqueue kmpathd");
1789                 r = -ENOMEM;
1790                 goto bad_alloc_kmultipathd;
1791         }
1792
1793         /*
1794          * A separate workqueue is used to handle the device handlers
1795          * to avoid overloading existing workqueue. Overloading the
1796          * old workqueue would also create a bottleneck in the
1797          * path of the storage hardware device activation.
1798          */
1799         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1800                                                   WQ_MEM_RECLAIM);
1801         if (!kmpath_handlerd) {
1802                 DMERR("failed to create workqueue kmpath_handlerd");
1803                 r = -ENOMEM;
1804                 goto bad_alloc_kmpath_handlerd;
1805         }
1806
1807         DMINFO("version %u.%u.%u loaded",
1808                multipath_target.version[0], multipath_target.version[1],
1809                multipath_target.version[2]);
1810
1811         return 0;
1812
1813 bad_alloc_kmpath_handlerd:
1814         destroy_workqueue(kmultipathd);
1815 bad_alloc_kmultipathd:
1816         dm_unregister_target(&multipath_target);
1817 bad_register_target:
1818         kmem_cache_destroy(_mpio_cache);
1819
1820         return r;
1821 }
1822
1823 static void __exit dm_multipath_exit(void)
1824 {
1825         destroy_workqueue(kmpath_handlerd);
1826         destroy_workqueue(kmultipathd);
1827
1828         dm_unregister_target(&multipath_target);
1829         kmem_cache_destroy(_mpio_cache);
1830 }
1831
1832 module_init(dm_multipath_init);
1833 module_exit(dm_multipath_exit);
1834
1835 MODULE_DESCRIPTION(DM_NAME " multipath target");
1836 MODULE_AUTHOR("Sistina Software <[email protected]>");
1837 MODULE_LICENSE("GPL");