drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9
  10 #include "dm-path-selector.h"
  11 #include "dm-uevent.h"
  12
  13 #include <linux/ctype.h>
  14 #include <linux/init.h>
  15 #include <linux/mempool.h>
  16 #include <linux/module.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/slab.h>
  19 #include <linux/time.h>
  20 #include <linux/workqueue.h>
  21 #include <scsi/scsi_dh.h>
  22 #include <linux/atomic.h>
  23
  24 #define DM_MSG_PREFIX "multipath"
  25 #define MESG_STR(x) x, sizeof(x)
  26 #define DM_PG_INIT_DELAY_MSECS 2000
  27 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  28
  29 /* Path properties */
  30 struct pgpath {
  31         struct list_head list;
  32
  33         struct priority_group *pg;      /* Owning PG */
  34         unsigned is_active;             /* Path status */
  35         unsigned fail_count;            /* Cumulative failure count */
  36
  37         struct dm_path path;
  38         struct delayed_work activate_path;
  39 };
  40
  41 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  42
  43 /*
  44  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  45  * Each has a path selector which controls which path gets used.
  46  */
  47 struct priority_group {
  48         struct list_head list;
  49
  50         struct multipath *m;            /* Owning multipath instance */
  51         struct path_selector ps;
  52
  53         unsigned pg_num;                /* Reference number */
  54         unsigned bypassed;              /* Temporarily bypass this PG? */
  55
  56         unsigned nr_pgpaths;            /* Number of paths in PG */
  57         struct list_head pgpaths;
  58 };
  59
  60 /* Multipath context */
  61 struct multipath {
  62         struct list_head list;
  63         struct dm_target *ti;
  64
  65         spinlock_t lock;
  66
  67         const char *hw_handler_name;
  68         char *hw_handler_params;
  69
  70         unsigned nr_priority_groups;
  71         struct list_head priority_groups;
  72
  73         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  74
  75         unsigned pg_init_required;      /* pg_init needs calling? */
  76         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  77         unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
  78
  79         unsigned nr_valid_paths;        /* Total number of usable paths */
  80         struct pgpath *current_pgpath;
  81         struct priority_group *current_pg;
  82         struct priority_group *next_pg; /* Switch to this PG if set */
  83         unsigned repeat_count;          /* I/Os left before calling PS again */
  84
  85         unsigned queue_io;              /* Must we queue all I/O? */
  86         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  87         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  88         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  89         unsigned pg_init_count;         /* Number of times pg_init called */
  90         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  91
  92         struct work_struct process_queued_ios;
  93         struct list_head queued_ios;
  94         unsigned queue_size;
  95
  96         struct work_struct trigger_event;
  97
  98         /*
  99          * We must use a mempool of dm_mpath_io structs so that we
 100          * can resubmit bios on error.
 101          */
 102         mempool_t *mpio_pool;
 103
 104         struct mutex work_mutex;
 105 };
 106
 107 /*
 108  * Context information attached to each bio we process.
 109  */
 110 struct dm_mpath_io {
 111         struct pgpath *pgpath;
 112         size_t nr_bytes;
 113 };
 114
 115 typedef int (*action_fn) (struct pgpath *pgpath);
 116
 117 #define MIN_IOS 256     /* Mempool size */
 118
 119 static struct kmem_cache *_mpio_cache;
 120
 121 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 122 static void process_queued_ios(struct work_struct *work);
 123 static void trigger_event(struct work_struct *work);
 124 static void activate_path(struct work_struct *work);
 125
 126
 127 /*-----------------------------------------------
 128  * Allocation routines
 129  *-----------------------------------------------*/
 130
 131 static struct pgpath *alloc_pgpath(void)
 132 {
 133         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 134
 135         if (pgpath) {
 136                 pgpath->is_active = 1;
 137                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 138         }
 139
 140         return pgpath;
 141 }
 142
 143 static void free_pgpath(struct pgpath *pgpath)
 144 {
 145         kfree(pgpath);
 146 }
 147
 148 static struct priority_group *alloc_priority_group(void)
 149 {
 150         struct priority_group *pg;
 151
 152         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 153
 154         if (pg)
 155                 INIT_LIST_HEAD(&pg->pgpaths);
 156
 157         return pg;
 158 }
 159
 160 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 161 {
 162         struct pgpath *pgpath, *tmp;
 163         struct multipath *m = ti->private;
 164
 165         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 166                 list_del(&pgpath->list);
 167                 if (m->hw_handler_name)
 168                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 169                 dm_put_device(ti, pgpath->path.dev);
 170                 free_pgpath(pgpath);
 171         }
 172 }
 173
 174 static void free_priority_group(struct priority_group *pg,
 175                                 struct dm_target *ti)
 176 {
 177         struct path_selector *ps = &pg->ps;
 178
 179         if (ps->type) {
 180                 ps->type->destroy(ps);
 181                 dm_put_path_selector(ps->type);
 182         }
 183
 184         free_pgpaths(&pg->pgpaths, ti);
 185         kfree(pg);
 186 }
 187
 188 static struct multipath *alloc_multipath(struct dm_target *ti)
 189 {
 190         struct multipath *m;
 191
 192         m = kzalloc(sizeof(*m), GFP_KERNEL);
 193         if (m) {
 194                 INIT_LIST_HEAD(&m->priority_groups);
 195                 INIT_LIST_HEAD(&m->queued_ios);
 196                 spin_lock_init(&m->lock);
 197                 m->queue_io = 1;
 198                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 199                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 200                 INIT_WORK(&m->trigger_event, trigger_event);
 201                 init_waitqueue_head(&m->pg_init_wait);
 202                 mutex_init(&m->work_mutex);
 203                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 204                 if (!m->mpio_pool) {
 205                         kfree(m);
 206                         return NULL;
 207                 }
 208                 m->ti = ti;
 209                 ti->private = m;
 210         }
 211
 212         return m;
 213 }
 214
 215 static void free_multipath(struct multipath *m)
 216 {
 217         struct priority_group *pg, *tmp;
 218
 219         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 220                 list_del(&pg->list);
 221                 free_priority_group(pg, m->ti);
 222         }
 223
 224         kfree(m->hw_handler_name);
 225         kfree(m->hw_handler_params);
 226         mempool_destroy(m->mpio_pool);
 227         kfree(m);
 228 }
 229
 230
 231 /*-----------------------------------------------
 232  * Path selection
 233  *-----------------------------------------------*/
 234
 235 static void __pg_init_all_paths(struct multipath *m)
 236 {
 237         struct pgpath *pgpath;
 238         unsigned long pg_init_delay = 0;
 239
 240         m->pg_init_count++;
 241         m->pg_init_required = 0;
 242         if (m->pg_init_delay_retry)
 243                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 244                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 245         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 246                 /* Skip failed paths */
 247                 if (!pgpath->is_active)
 248                         continue;
 249                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 250                                        pg_init_delay))
 251                         m->pg_init_in_progress++;
 252         }
 253 }
 254
 255 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 256 {
 257         m->current_pg = pgpath->pg;
 258
 259         /* Must we initialise the PG first, and queue I/O till it's ready? */
 260         if (m->hw_handler_name) {
 261                 m->pg_init_required = 1;
 262                 m->queue_io = 1;
 263         } else {
 264                 m->pg_init_required = 0;
 265                 m->queue_io = 0;
 266         }
 267
 268         m->pg_init_count = 0;
 269 }
 270
 271 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 272                                size_t nr_bytes)
 273 {
 274         struct dm_path *path;
 275
 276         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 277         if (!path)
 278                 return -ENXIO;
 279
 280         m->current_pgpath = path_to_pgpath(path);
 281
 282         if (m->current_pg != pg)
 283                 __switch_pg(m, m->current_pgpath);
 284
 285         return 0;
 286 }
 287
 288 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 289 {
 290         struct priority_group *pg;
 291         unsigned bypassed = 1;
 292
 293         if (!m->nr_valid_paths)
 294                 goto failed;
 295
 296         /* Were we instructed to switch PG? */
 297         if (m->next_pg) {
 298                 pg = m->next_pg;
 299                 m->next_pg = NULL;
 300                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 301                         return;
 302         }
 303
 304         /* Don't change PG until it has no remaining paths */
 305         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 306                 return;
 307
 308         /*
 309          * Loop through priority groups until we find a valid path.
 310          * First time we skip PGs marked 'bypassed'.
 311          * Second time we only try the ones we skipped.
 312          */
 313         do {
 314                 list_for_each_entry(pg, &m->priority_groups, list) {
 315                         if (pg->bypassed == bypassed)
 316                                 continue;
 317                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 318                                 return;
 319                 }
 320         } while (bypassed--);
 321
 322 failed:
 323         m->current_pgpath = NULL;
 324         m->current_pg = NULL;
 325 }
 326
 327 /*
 328  * Check whether bios must be queued in the device-mapper core rather
 329  * than here in the target.
 330  *
 331  * m->lock must be held on entry.
 332  *
 333  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 334  * same value then we are not between multipath_presuspend()
 335  * and multipath_resume() calls and we have no need to check
 336  * for the DMF_NOFLUSH_SUSPENDING flag.
 337  */
 338 static int __must_push_back(struct multipath *m)
 339 {
 340         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 341                 dm_noflush_suspending(m->ti));
 342 }
 343
 344 static int map_io(struct multipath *m, struct request *clone,
 345                   struct dm_mpath_io *mpio, unsigned was_queued)
 346 {
 347         int r = DM_MAPIO_REMAPPED;
 348         size_t nr_bytes = blk_rq_bytes(clone);
 349         unsigned long flags;
 350         struct pgpath *pgpath;
 351         struct block_device *bdev;
 352
 353         spin_lock_irqsave(&m->lock, flags);
 354
 355         /* Do we need to select a new pgpath? */
 356         if (!m->current_pgpath ||
 357             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 358                 __choose_pgpath(m, nr_bytes);
 359
 360         pgpath = m->current_pgpath;
 361
 362         if (was_queued)
 363                 m->queue_size--;
 364
 365         if ((pgpath && m->queue_io) ||
 366             (!pgpath && m->queue_if_no_path)) {
 367                 /* Queue for the daemon to resubmit */
 368                 list_add_tail(&clone->queuelist, &m->queued_ios);
 369                 m->queue_size++;
 370                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 371                     !m->queue_io)
 372                         queue_work(kmultipathd, &m->process_queued_ios);
 373                 pgpath = NULL;
 374                 r = DM_MAPIO_SUBMITTED;
 375         } else if (pgpath) {
 376                 bdev = pgpath->path.dev->bdev;
 377                 clone->q = bdev_get_queue(bdev);
 378                 clone->rq_disk = bdev->bd_disk;
 379         } else if (__must_push_back(m))
 380                 r = DM_MAPIO_REQUEUE;
 381         else
 382                 r = -EIO;       /* Failed */
 383
 384         mpio->pgpath = pgpath;
 385         mpio->nr_bytes = nr_bytes;
 386
 387         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 388                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 389                                               nr_bytes);
 390
 391         spin_unlock_irqrestore(&m->lock, flags);
 392
 393         return r;
 394 }
 395
 396 /*
 397  * If we run out of usable paths, should we queue I/O or error it?
 398  */
 399 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 400                             unsigned save_old_value)
 401 {
 402         unsigned long flags;
 403
 404         spin_lock_irqsave(&m->lock, flags);
 405
 406         if (save_old_value)
 407                 m->saved_queue_if_no_path = m->queue_if_no_path;
 408         else
 409                 m->saved_queue_if_no_path = queue_if_no_path;
 410         m->queue_if_no_path = queue_if_no_path;
 411         if (!m->queue_if_no_path && m->queue_size)
 412                 queue_work(kmultipathd, &m->process_queued_ios);
 413
 414         spin_unlock_irqrestore(&m->lock, flags);
 415
 416         return 0;
 417 }
 418
 419 /*-----------------------------------------------------------------
 420  * The multipath daemon is responsible for resubmitting queued ios.
 421  *---------------------------------------------------------------*/
 422
 423 static void dispatch_queued_ios(struct multipath *m)
 424 {
 425         int r;
 426         unsigned long flags;
 427         struct dm_mpath_io *mpio;
 428         union map_info *info;
 429         struct request *clone, *n;
 430         LIST_HEAD(cl);
 431
 432         spin_lock_irqsave(&m->lock, flags);
 433         list_splice_init(&m->queued_ios, &cl);
 434         spin_unlock_irqrestore(&m->lock, flags);
 435
 436         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 437                 list_del_init(&clone->queuelist);
 438
 439                 info = dm_get_rq_mapinfo(clone);
 440                 mpio = info->ptr;
 441
 442                 r = map_io(m, clone, mpio, 1);
 443                 if (r < 0) {
 444                         mempool_free(mpio, m->mpio_pool);
 445                         dm_kill_unmapped_request(clone, r);
 446                 } else if (r == DM_MAPIO_REMAPPED)
 447                         dm_dispatch_request(clone);
 448                 else if (r == DM_MAPIO_REQUEUE) {
 449                         mempool_free(mpio, m->mpio_pool);
 450                         dm_requeue_unmapped_request(clone);
 451                 }
 452         }
 453 }
 454
 455 static void process_queued_ios(struct work_struct *work)
 456 {
 457         struct multipath *m =
 458                 container_of(work, struct multipath, process_queued_ios);
 459         struct pgpath *pgpath = NULL;
 460         unsigned must_queue = 1;
 461         unsigned long flags;
 462
 463         spin_lock_irqsave(&m->lock, flags);
 464
 465         if (!m->queue_size)
 466                 goto out;
 467
 468         if (!m->current_pgpath)
 469                 __choose_pgpath(m, 0);
 470
 471         pgpath = m->current_pgpath;
 472
 473         if ((pgpath && !m->queue_io) ||
 474             (!pgpath && !m->queue_if_no_path))
 475                 must_queue = 0;
 476
 477         if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
 478                 __pg_init_all_paths(m);
 479
 480 out:
 481         spin_unlock_irqrestore(&m->lock, flags);
 482         if (!must_queue)
 483                 dispatch_queued_ios(m);
 484 }
 485
 486 /*
 487  * An event is triggered whenever a path is taken out of use.
 488  * Includes path failure and PG bypass.
 489  */
 490 static void trigger_event(struct work_struct *work)
 491 {
 492         struct multipath *m =
 493                 container_of(work, struct multipath, trigger_event);
 494
 495         dm_table_event(m->ti->table);
 496 }
 497
 498 /*-----------------------------------------------------------------
 499  * Constructor/argument parsing:
 500  * <#multipath feature args> [<arg>]*
 501  * <#hw_handler args> [hw_handler [<arg>]*]
 502  * <#priority groups>
 503  * <initial priority group>
 504  *     [<selector> <#selector args> [<arg>]*
 505  *      <#paths> <#per-path selector args>
 506  *         [<path> [<arg>]* ]+ ]+
 507  *---------------------------------------------------------------*/
 508 struct param {
 509         unsigned min;
 510         unsigned max;
 511         char *error;
 512 };
 513
 514 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 515 {
 516         if (!str ||
 517             (sscanf(str, "%u", v) != 1) ||
 518             (*v < param->min) ||
 519             (*v > param->max)) {
 520                 *error = param->error;
 521                 return -EINVAL;
 522         }
 523
 524         return 0;
 525 }
 526
 527 struct arg_set {
 528         unsigned argc;
 529         char **argv;
 530 };
 531
 532 static char *shift(struct arg_set *as)
 533 {
 534         char *r;
 535
 536         if (as->argc) {
 537                 as->argc--;
 538                 r = *as->argv;
 539                 as->argv++;
 540                 return r;
 541         }
 542
 543         return NULL;
 544 }
 545
 546 static void consume(struct arg_set *as, unsigned n)
 547 {
 548         BUG_ON (as->argc < n);
 549         as->argc -= n;
 550         as->argv += n;
 551 }
 552
 553 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 554                                struct dm_target *ti)
 555 {
 556         int r;
 557         struct path_selector_type *pst;
 558         unsigned ps_argc;
 559
 560         static struct param _params[] = {
 561                 {0, 1024, "invalid number of path selector args"},
 562         };
 563
 564         pst = dm_get_path_selector(shift(as));
 565         if (!pst) {
 566                 ti->error = "unknown path selector type";
 567                 return -EINVAL;
 568         }
 569
 570         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 571         if (r) {
 572                 dm_put_path_selector(pst);
 573                 return -EINVAL;
 574         }
 575
 576         if (ps_argc > as->argc) {
 577                 dm_put_path_selector(pst);
 578                 ti->error = "not enough arguments for path selector";
 579                 return -EINVAL;
 580         }
 581
 582         r = pst->create(&pg->ps, ps_argc, as->argv);
 583         if (r) {
 584                 dm_put_path_selector(pst);
 585                 ti->error = "path selector constructor failed";
 586                 return r;
 587         }
 588
 589         pg->ps.type = pst;
 590         consume(as, ps_argc);
 591
 592         return 0;
 593 }
 594
 595 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 596                                struct dm_target *ti)
 597 {
 598         int r;
 599         struct pgpath *p;
 600         struct multipath *m = ti->private;
 601
 602         /* we need at least a path arg */
 603         if (as->argc < 1) {
 604                 ti->error = "no device given";
 605                 return ERR_PTR(-EINVAL);
 606         }
 607
 608         p = alloc_pgpath();
 609         if (!p)
 610                 return ERR_PTR(-ENOMEM);
 611
 612         r = dm_get_device(ti, shift(as), dm_table_get_mode(ti->table),
 613                           &p->path.dev);
 614         if (r) {
 615                 ti->error = "error getting device";
 616                 goto bad;
 617         }
 618
 619         if (m->hw_handler_name) {
 620                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 621
 622                 r = scsi_dh_attach(q, m->hw_handler_name);
 623                 if (r == -EBUSY) {
 624                         /*
 625                          * Already attached to different hw_handler,
 626                          * try to reattach with correct one.
 627                          */
 628                         scsi_dh_detach(q);
 629                         r = scsi_dh_attach(q, m->hw_handler_name);
 630                 }
 631
 632                 if (r < 0) {
 633                         ti->error = "error attaching hardware handler";
 634                         dm_put_device(ti, p->path.dev);
 635                         goto bad;
 636                 }
 637
 638                 if (m->hw_handler_params) {
 639                         r = scsi_dh_set_params(q, m->hw_handler_params);
 640                         if (r < 0) {
 641                                 ti->error = "unable to set hardware "
 642                                                         "handler parameters";
 643                                 scsi_dh_detach(q);
 644                                 dm_put_device(ti, p->path.dev);
 645                                 goto bad;
 646                         }
 647                 }
 648         }
 649
 650         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 651         if (r) {
 652                 dm_put_device(ti, p->path.dev);
 653                 goto bad;
 654         }
 655
 656         return p;
 657
 658  bad:
 659         free_pgpath(p);
 660         return ERR_PTR(r);
 661 }
 662
 663 static struct priority_group *parse_priority_group(struct arg_set *as,
 664                                                    struct multipath *m)
 665 {
 666         static struct param _params[] = {
 667                 {1, 1024, "invalid number of paths"},
 668                 {0, 1024, "invalid number of selector args"}
 669         };
 670
 671         int r;
 672         unsigned i, nr_selector_args, nr_params;
 673         struct priority_group *pg;
 674         struct dm_target *ti = m->ti;
 675
 676         if (as->argc < 2) {
 677                 as->argc = 0;
 678                 ti->error = "not enough priority group arguments";
 679                 return ERR_PTR(-EINVAL);
 680         }
 681
 682         pg = alloc_priority_group();
 683         if (!pg) {
 684                 ti->error = "couldn't allocate priority group";
 685                 return ERR_PTR(-ENOMEM);
 686         }
 687         pg->m = m;
 688
 689         r = parse_path_selector(as, pg, ti);
 690         if (r)
 691                 goto bad;
 692
 693         /*
 694          * read the paths
 695          */
 696         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 697         if (r)
 698                 goto bad;
 699
 700         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 701         if (r)
 702                 goto bad;
 703
 704         nr_params = 1 + nr_selector_args;
 705         for (i = 0; i < pg->nr_pgpaths; i++) {
 706                 struct pgpath *pgpath;
 707                 struct arg_set path_args;
 708
 709                 if (as->argc < nr_params) {
 710                         ti->error = "not enough path parameters";
 711                         r = -EINVAL;
 712                         goto bad;
 713                 }
 714
 715                 path_args.argc = nr_params;
 716                 path_args.argv = as->argv;
 717
 718                 pgpath = parse_path(&path_args, &pg->ps, ti);
 719                 if (IS_ERR(pgpath)) {
 720                         r = PTR_ERR(pgpath);
 721                         goto bad;
 722                 }
 723
 724                 pgpath->pg = pg;
 725                 list_add_tail(&pgpath->list, &pg->pgpaths);
 726                 consume(as, nr_params);
 727         }
 728
 729         return pg;
 730
 731  bad:
 732         free_priority_group(pg, ti);
 733         return ERR_PTR(r);
 734 }
 735
 736 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 737 {
 738         unsigned hw_argc;
 739         int ret;
 740         struct dm_target *ti = m->ti;
 741
 742         static struct param _params[] = {
 743                 {0, 1024, "invalid number of hardware handler args"},
 744         };
 745
 746         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 747                 return -EINVAL;
 748
 749         if (!hw_argc)
 750                 return 0;
 751
 752         if (hw_argc > as->argc) {
 753                 ti->error = "not enough arguments for hardware handler";
 754                 return -EINVAL;
 755         }
 756
 757         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 758         request_module("scsi_dh_%s", m->hw_handler_name);
 759         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 760                 ti->error = "unknown hardware handler type";
 761                 ret = -EINVAL;
 762                 goto fail;
 763         }
 764
 765         if (hw_argc > 1) {
 766                 char *p;
 767                 int i, j, len = 4;
 768
 769                 for (i = 0; i <= hw_argc - 2; i++)
 770                         len += strlen(as->argv[i]) + 1;
 771                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 772                 if (!p) {
 773                         ti->error = "memory allocation failed";
 774                         ret = -ENOMEM;
 775                         goto fail;
 776                 }
 777                 j = sprintf(p, "%d", hw_argc - 1);
 778                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 779                         j = sprintf(p, "%s", as->argv[i]);
 780         }
 781         consume(as, hw_argc - 1);
 782
 783         return 0;
 784 fail:
 785         kfree(m->hw_handler_name);
 786         m->hw_handler_name = NULL;
 787         return ret;
 788 }
 789
 790 static int parse_features(struct arg_set *as, struct multipath *m)
 791 {
 792         int r;
 793         unsigned argc;
 794         struct dm_target *ti = m->ti;
 795         const char *param_name;
 796
 797         static struct param _params[] = {
 798                 {0, 5, "invalid number of feature args"},
 799                 {1, 50, "pg_init_retries must be between 1 and 50"},
 800                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 801         };
 802
 803         r = read_param(_params, shift(as), &argc, &ti->error);
 804         if (r)
 805                 return -EINVAL;
 806
 807         if (!argc)
 808                 return 0;
 809
 810         if (argc > as->argc) {
 811                 ti->error = "not enough arguments for features";
 812                 return -EINVAL;
 813         }
 814
 815         do {
 816                 param_name = shift(as);
 817                 argc--;
 818
 819                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 820                         r = queue_if_no_path(m, 1, 0);
 821                         continue;
 822                 }
 823
 824                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 825                     (argc >= 1)) {
 826                         r = read_param(_params + 1, shift(as),
 827                                        &m->pg_init_retries, &ti->error);
 828                         argc--;
 829                         continue;
 830                 }
 831
 832                 if (!strnicmp(param_name, MESG_STR("pg_init_delay_msecs")) &&
 833                     (argc >= 1)) {
 834                         r = read_param(_params + 2, shift(as),
 835                                        &m->pg_init_delay_msecs, &ti->error);
 836                         argc--;
 837                         continue;
 838                 }
 839
 840                 ti->error = "Unrecognised multipath feature request";
 841                 r = -EINVAL;
 842         } while (argc && !r);
 843
 844         return r;
 845 }
 846
 847 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 848                          char **argv)
 849 {
 850         /* target parameters */
 851         static struct param _params[] = {
 852                 {0, 1024, "invalid number of priority groups"},
 853                 {0, 1024, "invalid initial priority group number"},
 854         };
 855
 856         int r;
 857         struct multipath *m;
 858         struct arg_set as;
 859         unsigned pg_count = 0;
 860         unsigned next_pg_num;
 861
 862         as.argc = argc;
 863         as.argv = argv;
 864
 865         m = alloc_multipath(ti);
 866         if (!m) {
 867                 ti->error = "can't allocate multipath";
 868                 return -EINVAL;
 869         }
 870
 871         r = parse_features(&as, m);
 872         if (r)
 873                 goto bad;
 874
 875         r = parse_hw_handler(&as, m);
 876         if (r)
 877                 goto bad;
 878
 879         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 880         if (r)
 881                 goto bad;
 882
 883         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 884         if (r)
 885                 goto bad;
 886
 887         if ((!m->nr_priority_groups && next_pg_num) ||
 888             (m->nr_priority_groups && !next_pg_num)) {
 889                 ti->error = "invalid initial priority group";
 890                 r = -EINVAL;
 891                 goto bad;
 892         }
 893
 894         /* parse the priority groups */
 895         while (as.argc) {
 896                 struct priority_group *pg;
 897
 898                 pg = parse_priority_group(&as, m);
 899                 if (IS_ERR(pg)) {
 900                         r = PTR_ERR(pg);
 901                         goto bad;
 902                 }
 903
 904                 m->nr_valid_paths += pg->nr_pgpaths;
 905                 list_add_tail(&pg->list, &m->priority_groups);
 906                 pg_count++;
 907                 pg->pg_num = pg_count;
 908                 if (!--next_pg_num)
 909                         m->next_pg = pg;
 910         }
 911
 912         if (pg_count != m->nr_priority_groups) {
 913                 ti->error = "priority group count mismatch";
 914                 r = -EINVAL;
 915                 goto bad;
 916         }
 917
 918         ti->num_flush_requests = 1;
 919         ti->num_discard_requests = 1;
 920
 921         return 0;
 922
 923  bad:
 924         free_multipath(m);
 925         return r;
 926 }
 927
 928 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 929 {
 930         DECLARE_WAITQUEUE(wait, current);
 931         unsigned long flags;
 932
 933         add_wait_queue(&m->pg_init_wait, &wait);
 934
 935         while (1) {
 936                 set_current_state(TASK_UNINTERRUPTIBLE);
 937
 938                 spin_lock_irqsave(&m->lock, flags);
 939                 if (!m->pg_init_in_progress) {
 940                         spin_unlock_irqrestore(&m->lock, flags);
 941                         break;
 942                 }
 943                 spin_unlock_irqrestore(&m->lock, flags);
 944
 945                 io_schedule();
 946         }
 947         set_current_state(TASK_RUNNING);
 948
 949         remove_wait_queue(&m->pg_init_wait, &wait);
 950 }
 951
 952 static void flush_multipath_work(struct multipath *m)
 953 {
 954         flush_workqueue(kmpath_handlerd);
 955         multipath_wait_for_pg_init_completion(m);
 956         flush_workqueue(kmultipathd);
 957         flush_work_sync(&m->trigger_event);
 958 }
 959
 960 static void multipath_dtr(struct dm_target *ti)
 961 {
 962         struct multipath *m = ti->private;
 963
 964         flush_multipath_work(m);
 965         free_multipath(m);
 966 }
 967
 968 /*
 969  * Map cloned requests
 970  */
 971 static int multipath_map(struct dm_target *ti, struct request *clone,
 972                          union map_info *map_context)
 973 {
 974         int r;
 975         struct dm_mpath_io *mpio;
 976         struct multipath *m = (struct multipath *) ti->private;
 977
 978         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 979         if (!mpio)
 980                 /* ENOMEM, requeue */
 981                 return DM_MAPIO_REQUEUE;
 982         memset(mpio, 0, sizeof(*mpio));
 983
 984         map_context->ptr = mpio;
 985         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 986         r = map_io(m, clone, mpio, 0);
 987         if (r < 0 || r == DM_MAPIO_REQUEUE)
 988                 mempool_free(mpio, m->mpio_pool);
 989
 990         return r;
 991 }
 992
 993 /*
 994  * Take a path out of use.
 995  */
 996 static int fail_path(struct pgpath *pgpath)
 997 {
 998         unsigned long flags;
 999         struct multipath *m = pgpath->pg->m;
1000
1001         spin_lock_irqsave(&m->lock, flags);
1002
1003         if (!pgpath->is_active)
1004                 goto out;
1005
1006         DMWARN("Failing path %s.", pgpath->path.dev->name);
1007
1008         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1009         pgpath->is_active = 0;
1010         pgpath->fail_count++;
1011
1012         m->nr_valid_paths--;
1013
1014         if (pgpath == m->current_pgpath)
1015                 m->current_pgpath = NULL;
1016
1017         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1018                       pgpath->path.dev->name, m->nr_valid_paths);
1019
1020         schedule_work(&m->trigger_event);
1021
1022 out:
1023         spin_unlock_irqrestore(&m->lock, flags);
1024
1025         return 0;
1026 }
1027
1028 /*
1029  * Reinstate a previously-failed path
1030  */
1031 static int reinstate_path(struct pgpath *pgpath)
1032 {
1033         int r = 0;
1034         unsigned long flags;
1035         struct multipath *m = pgpath->pg->m;
1036
1037         spin_lock_irqsave(&m->lock, flags);
1038
1039         if (pgpath->is_active)
1040                 goto out;
1041
1042         if (!pgpath->pg->ps.type->reinstate_path) {
1043                 DMWARN("Reinstate path not supported by path selector %s",
1044                        pgpath->pg->ps.type->name);
1045                 r = -EINVAL;
1046                 goto out;
1047         }
1048
1049         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1050         if (r)
1051                 goto out;
1052
1053         pgpath->is_active = 1;
1054
1055         if (!m->nr_valid_paths++ && m->queue_size) {
1056                 m->current_pgpath = NULL;
1057                 queue_work(kmultipathd, &m->process_queued_ios);
1058         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1059                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1060                         m->pg_init_in_progress++;
1061         }
1062
1063         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1064                       pgpath->path.dev->name, m->nr_valid_paths);
1065
1066         schedule_work(&m->trigger_event);
1067
1068 out:
1069         spin_unlock_irqrestore(&m->lock, flags);
1070
1071         return r;
1072 }
1073
1074 /*
1075  * Fail or reinstate all paths that match the provided struct dm_dev.
1076  */
1077 static int action_dev(struct multipath *m, struct dm_dev *dev,
1078                       action_fn action)
1079 {
1080         int r = -EINVAL;
1081         struct pgpath *pgpath;
1082         struct priority_group *pg;
1083
1084         list_for_each_entry(pg, &m->priority_groups, list) {
1085                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1086                         if (pgpath->path.dev == dev)
1087                                 r = action(pgpath);
1088                 }
1089         }
1090
1091         return r;
1092 }
1093
1094 /*
1095  * Temporarily try to avoid having to use the specified PG
1096  */
1097 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1098                       int bypassed)
1099 {
1100         unsigned long flags;
1101
1102         spin_lock_irqsave(&m->lock, flags);
1103
1104         pg->bypassed = bypassed;
1105         m->current_pgpath = NULL;
1106         m->current_pg = NULL;
1107
1108         spin_unlock_irqrestore(&m->lock, flags);
1109
1110         schedule_work(&m->trigger_event);
1111 }
1112
1113 /*
1114  * Switch to using the specified PG from the next I/O that gets mapped
1115  */
1116 static int switch_pg_num(struct multipath *m, const char *pgstr)
1117 {
1118         struct priority_group *pg;
1119         unsigned pgnum;
1120         unsigned long flags;
1121
1122         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1123             (pgnum > m->nr_priority_groups)) {
1124                 DMWARN("invalid PG number supplied to switch_pg_num");
1125                 return -EINVAL;
1126         }
1127
1128         spin_lock_irqsave(&m->lock, flags);
1129         list_for_each_entry(pg, &m->priority_groups, list) {
1130                 pg->bypassed = 0;
1131                 if (--pgnum)
1132                         continue;
1133
1134                 m->current_pgpath = NULL;
1135                 m->current_pg = NULL;
1136                 m->next_pg = pg;
1137         }
1138         spin_unlock_irqrestore(&m->lock, flags);
1139
1140         schedule_work(&m->trigger_event);
1141         return 0;
1142 }
1143
1144 /*
1145  * Set/clear bypassed status of a PG.
1146  * PGs are numbered upwards from 1 in the order they were declared.
1147  */
1148 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1149 {
1150         struct priority_group *pg;
1151         unsigned pgnum;
1152
1153         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1154             (pgnum > m->nr_priority_groups)) {
1155                 DMWARN("invalid PG number supplied to bypass_pg");
1156                 return -EINVAL;
1157         }
1158
1159         list_for_each_entry(pg, &m->priority_groups, list) {
1160                 if (!--pgnum)
1161                         break;
1162         }
1163
1164         bypass_pg(m, pg, bypassed);
1165         return 0;
1166 }
1167
1168 /*
1169  * Should we retry pg_init immediately?
1170  */
1171 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1172 {
1173         unsigned long flags;
1174         int limit_reached = 0;
1175
1176         spin_lock_irqsave(&m->lock, flags);
1177
1178         if (m->pg_init_count <= m->pg_init_retries)
1179                 m->pg_init_required = 1;
1180         else
1181                 limit_reached = 1;
1182
1183         spin_unlock_irqrestore(&m->lock, flags);
1184
1185         return limit_reached;
1186 }
1187
1188 static void pg_init_done(void *data, int errors)
1189 {
1190         struct pgpath *pgpath = data;
1191         struct priority_group *pg = pgpath->pg;
1192         struct multipath *m = pg->m;
1193         unsigned long flags;
1194         unsigned delay_retry = 0;
1195
1196         /* device or driver problems */
1197         switch (errors) {
1198         case SCSI_DH_OK:
1199                 break;
1200         case SCSI_DH_NOSYS:
1201                 if (!m->hw_handler_name) {
1202                         errors = 0;
1203                         break;
1204                 }
1205                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1206                       "Error %d.", m->hw_handler_name, errors);
1207                 /*
1208                  * Fail path for now, so we do not ping pong
1209                  */
1210                 fail_path(pgpath);
1211                 break;
1212         case SCSI_DH_DEV_TEMP_BUSY:
1213                 /*
1214                  * Probably doing something like FW upgrade on the
1215                  * controller so try the other pg.
1216                  */
1217                 bypass_pg(m, pg, 1);
1218                 break;
1219         case SCSI_DH_RETRY:
1220                 /* Wait before retrying. */
1221                 delay_retry = 1;
1222         case SCSI_DH_IMM_RETRY:
1223         case SCSI_DH_RES_TEMP_UNAVAIL:
1224                 if (pg_init_limit_reached(m, pgpath))
1225                         fail_path(pgpath);
1226                 errors = 0;
1227                 break;
1228         default:
1229                 /*
1230                  * We probably do not want to fail the path for a device
1231                  * error, but this is what the old dm did. In future
1232                  * patches we can do more advanced handling.
1233                  */
1234                 fail_path(pgpath);
1235         }
1236
1237         spin_lock_irqsave(&m->lock, flags);
1238         if (errors) {
1239                 if (pgpath == m->current_pgpath) {
1240                         DMERR("Could not failover device. Error %d.", errors);
1241                         m->current_pgpath = NULL;
1242                         m->current_pg = NULL;
1243                 }
1244         } else if (!m->pg_init_required)
1245                 pg->bypassed = 0;
1246
1247         if (--m->pg_init_in_progress)
1248                 /* Activations of other paths are still on going */
1249                 goto out;
1250
1251         if (!m->pg_init_required)
1252                 m->queue_io = 0;
1253
1254         m->pg_init_delay_retry = delay_retry;
1255         queue_work(kmultipathd, &m->process_queued_ios);
1256
1257         /*
1258          * Wake up any thread waiting to suspend.
1259          */
1260         wake_up(&m->pg_init_wait);
1261
1262 out:
1263         spin_unlock_irqrestore(&m->lock, flags);
1264 }
1265
1266 static void activate_path(struct work_struct *work)
1267 {
1268         struct pgpath *pgpath =
1269                 container_of(work, struct pgpath, activate_path.work);
1270
1271         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1272                                 pg_init_done, pgpath);
1273 }
1274
1275 /*
1276  * end_io handling
1277  */
1278 static int do_end_io(struct multipath *m, struct request *clone,
1279                      int error, struct dm_mpath_io *mpio)
1280 {
1281         /*
1282          * We don't queue any clone request inside the multipath target
1283          * during end I/O handling, since those clone requests don't have
1284          * bio clones.  If we queue them inside the multipath target,
1285          * we need to make bio clones, that requires memory allocation.
1286          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1287          *  don't have bio clones.)
1288          * Instead of queueing the clone request here, we queue the original
1289          * request into dm core, which will remake a clone request and
1290          * clone bios for it and resubmit it later.
1291          */
1292         int r = DM_ENDIO_REQUEUE;
1293         unsigned long flags;
1294
1295         if (!error && !clone->errors)
1296                 return 0;       /* I/O complete */
1297
1298         if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1299                 return error;
1300
1301         if (mpio->pgpath)
1302                 fail_path(mpio->pgpath);
1303
1304         spin_lock_irqsave(&m->lock, flags);
1305         if (!m->nr_valid_paths) {
1306                 if (!m->queue_if_no_path) {
1307                         if (!__must_push_back(m))
1308                                 r = -EIO;
1309                 } else {
1310                         if (error == -EBADE)
1311                                 r = error;
1312                 }
1313         }
1314         spin_unlock_irqrestore(&m->lock, flags);
1315
1316         return r;
1317 }
1318
1319 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1320                             int error, union map_info *map_context)
1321 {
1322         struct multipath *m = ti->private;
1323         struct dm_mpath_io *mpio = map_context->ptr;
1324         struct pgpath *pgpath = mpio->pgpath;
1325         struct path_selector *ps;
1326         int r;
1327
1328         r  = do_end_io(m, clone, error, mpio);
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         mempool_free(mpio, m->mpio_pool);
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 = (struct multipath *) ti->private;
1348
1349         queue_if_no_path(m, 0, 1);
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 = (struct multipath *) 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 int multipath_status(struct dm_target *ti, status_type_t type,
1391                             char *result, unsigned int maxlen)
1392 {
1393         int sz = 0;
1394         unsigned long flags;
1395         struct multipath *m = (struct multipath *) 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_size, 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                 if (m->queue_if_no_path)
1411                         DMEMIT("queue_if_no_path ");
1412                 if (m->pg_init_retries)
1413                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1414                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1415                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1416         }
1417
1418         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1419                 DMEMIT("0 ");
1420         else
1421                 DMEMIT("1 %s ", m->hw_handler_name);
1422
1423         DMEMIT("%u ", m->nr_priority_groups);
1424
1425         if (m->next_pg)
1426                 pg_num = m->next_pg->pg_num;
1427         else if (m->current_pg)
1428                 pg_num = m->current_pg->pg_num;
1429         else
1430                 pg_num = (m->nr_priority_groups ? 1 : 0);
1431
1432         DMEMIT("%u ", pg_num);
1433
1434         switch (type) {
1435         case STATUSTYPE_INFO:
1436                 list_for_each_entry(pg, &m->priority_groups, list) {
1437                         if (pg->bypassed)
1438                                 state = 'D';    /* Disabled */
1439                         else if (pg == m->current_pg)
1440                                 state = 'A';    /* Currently Active */
1441                         else
1442                                 state = 'E';    /* Enabled */
1443
1444                         DMEMIT("%c ", state);
1445
1446                         if (pg->ps.type->status)
1447                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1448                                                           result + sz,
1449                                                           maxlen - sz);
1450                         else
1451                                 DMEMIT("0 ");
1452
1453                         DMEMIT("%u %u ", pg->nr_pgpaths,
1454                                pg->ps.type->info_args);
1455
1456                         list_for_each_entry(p, &pg->pgpaths, list) {
1457                                 DMEMIT("%s %s %u ", p->path.dev->name,
1458                                        p->is_active ? "A" : "F",
1459                                        p->fail_count);
1460                                 if (pg->ps.type->status)
1461                                         sz += pg->ps.type->status(&pg->ps,
1462                                               &p->path, type, result + sz,
1463                                               maxlen - sz);
1464                         }
1465                 }
1466                 break;
1467
1468         case STATUSTYPE_TABLE:
1469                 list_for_each_entry(pg, &m->priority_groups, list) {
1470                         DMEMIT("%s ", pg->ps.type->name);
1471
1472                         if (pg->ps.type->status)
1473                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1474                                                           result + sz,
1475                                                           maxlen - sz);
1476                         else
1477                                 DMEMIT("0 ");
1478
1479                         DMEMIT("%u %u ", pg->nr_pgpaths,
1480                                pg->ps.type->table_args);
1481
1482                         list_for_each_entry(p, &pg->pgpaths, list) {
1483                                 DMEMIT("%s ", p->path.dev->name);
1484                                 if (pg->ps.type->status)
1485                                         sz += pg->ps.type->status(&pg->ps,
1486                                               &p->path, type, result + sz,
1487                                               maxlen - sz);
1488                         }
1489                 }
1490                 break;
1491         }
1492
1493         spin_unlock_irqrestore(&m->lock, flags);
1494
1495         return 0;
1496 }
1497
1498 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1499 {
1500         int r = -EINVAL;
1501         struct dm_dev *dev;
1502         struct multipath *m = (struct multipath *) ti->private;
1503         action_fn action;
1504
1505         mutex_lock(&m->work_mutex);
1506
1507         if (dm_suspended(ti)) {
1508                 r = -EBUSY;
1509                 goto out;
1510         }
1511
1512         if (argc == 1) {
1513                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
1514                         r = queue_if_no_path(m, 1, 0);
1515                         goto out;
1516                 } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
1517                         r = queue_if_no_path(m, 0, 0);
1518                         goto out;
1519                 }
1520         }
1521
1522         if (argc != 2) {
1523                 DMWARN("Unrecognised multipath message received.");
1524                 goto out;
1525         }
1526
1527         if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
1528                 r = bypass_pg_num(m, argv[1], 1);
1529                 goto out;
1530         } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
1531                 r = bypass_pg_num(m, argv[1], 0);
1532                 goto out;
1533         } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
1534                 r = switch_pg_num(m, argv[1]);
1535                 goto out;
1536         } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1537                 action = reinstate_path;
1538         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1539                 action = fail_path;
1540         else {
1541                 DMWARN("Unrecognised multipath message received.");
1542                 goto out;
1543         }
1544
1545         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1546         if (r) {
1547                 DMWARN("message: error getting device %s",
1548                        argv[1]);
1549                 goto out;
1550         }
1551
1552         r = action_dev(m, dev, action);
1553
1554         dm_put_device(ti, dev);
1555
1556 out:
1557         mutex_unlock(&m->work_mutex);
1558         return r;
1559 }
1560
1561 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1562                            unsigned long arg)
1563 {
1564         struct multipath *m = (struct multipath *) ti->private;
1565         struct block_device *bdev = NULL;
1566         fmode_t mode = 0;
1567         unsigned long flags;
1568         int r = 0;
1569
1570         spin_lock_irqsave(&m->lock, flags);
1571
1572         if (!m->current_pgpath)
1573                 __choose_pgpath(m, 0);
1574
1575         if (m->current_pgpath) {
1576                 bdev = m->current_pgpath->path.dev->bdev;
1577                 mode = m->current_pgpath->path.dev->mode;
1578         }
1579
1580         if (m->queue_io)
1581                 r = -EAGAIN;
1582         else if (!bdev)
1583                 r = -EIO;
1584
1585         spin_unlock_irqrestore(&m->lock, flags);
1586
1587         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1588 }
1589
1590 static int multipath_iterate_devices(struct dm_target *ti,
1591                                      iterate_devices_callout_fn fn, void *data)
1592 {
1593         struct multipath *m = ti->private;
1594         struct priority_group *pg;
1595         struct pgpath *p;
1596         int ret = 0;
1597
1598         list_for_each_entry(pg, &m->priority_groups, list) {
1599                 list_for_each_entry(p, &pg->pgpaths, list) {
1600                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1601                         if (ret)
1602                                 goto out;
1603                 }
1604         }
1605
1606 out:
1607         return ret;
1608 }
1609
1610 static int __pgpath_busy(struct pgpath *pgpath)
1611 {
1612         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1613
1614         return dm_underlying_device_busy(q);
1615 }
1616
1617 /*
1618  * We return "busy", only when we can map I/Os but underlying devices
1619  * are busy (so even if we map I/Os now, the I/Os will wait on
1620  * the underlying queue).
1621  * In other words, if we want to kill I/Os or queue them inside us
1622  * due to map unavailability, we don't return "busy".  Otherwise,
1623  * dm core won't give us the I/Os and we can't do what we want.
1624  */
1625 static int multipath_busy(struct dm_target *ti)
1626 {
1627         int busy = 0, has_active = 0;
1628         struct multipath *m = ti->private;
1629         struct priority_group *pg;
1630         struct pgpath *pgpath;
1631         unsigned long flags;
1632
1633         spin_lock_irqsave(&m->lock, flags);
1634
1635         /* Guess which priority_group will be used at next mapping time */
1636         if (unlikely(!m->current_pgpath && m->next_pg))
1637                 pg = m->next_pg;
1638         else if (likely(m->current_pg))
1639                 pg = m->current_pg;
1640         else
1641                 /*
1642                  * We don't know which pg will be used at next mapping time.
1643                  * We don't call __choose_pgpath() here to avoid to trigger
1644                  * pg_init just by busy checking.
1645                  * So we don't know whether underlying devices we will be using
1646                  * at next mapping time are busy or not. Just try mapping.
1647                  */
1648                 goto out;
1649
1650         /*
1651          * If there is one non-busy active path at least, the path selector
1652          * will be able to select it. So we consider such a pg as not busy.
1653          */
1654         busy = 1;
1655         list_for_each_entry(pgpath, &pg->pgpaths, list)
1656                 if (pgpath->is_active) {
1657                         has_active = 1;
1658
1659                         if (!__pgpath_busy(pgpath)) {
1660                                 busy = 0;
1661                                 break;
1662                         }
1663                 }
1664
1665         if (!has_active)
1666                 /*
1667                  * No active path in this pg, so this pg won't be used and
1668                  * the current_pg will be changed at next mapping time.
1669                  * We need to try mapping to determine it.
1670                  */
1671                 busy = 0;
1672
1673 out:
1674         spin_unlock_irqrestore(&m->lock, flags);
1675
1676         return busy;
1677 }
1678
1679 /*-----------------------------------------------------------------
1680  * Module setup
1681  *---------------------------------------------------------------*/
1682 static struct target_type multipath_target = {
1683         .name = "multipath",
1684         .version = {1, 3, 0},
1685         .module = THIS_MODULE,
1686         .ctr = multipath_ctr,
1687         .dtr = multipath_dtr,
1688         .map_rq = multipath_map,
1689         .rq_end_io = multipath_end_io,
1690         .presuspend = multipath_presuspend,
1691         .postsuspend = multipath_postsuspend,
1692         .resume = multipath_resume,
1693         .status = multipath_status,
1694         .message = multipath_message,
1695         .ioctl  = multipath_ioctl,
1696         .iterate_devices = multipath_iterate_devices,
1697         .busy = multipath_busy,
1698 };
1699
1700 static int __init dm_multipath_init(void)
1701 {
1702         int r;
1703
1704         /* allocate a slab for the dm_ios */
1705         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1706         if (!_mpio_cache)
1707                 return -ENOMEM;
1708
1709         r = dm_register_target(&multipath_target);
1710         if (r < 0) {
1711                 DMERR("register failed %d", r);
1712                 kmem_cache_destroy(_mpio_cache);
1713                 return -EINVAL;
1714         }
1715
1716         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1717         if (!kmultipathd) {
1718                 DMERR("failed to create workqueue kmpathd");
1719                 dm_unregister_target(&multipath_target);
1720                 kmem_cache_destroy(_mpio_cache);
1721                 return -ENOMEM;
1722         }
1723
1724         /*
1725          * A separate workqueue is used to handle the device handlers
1726          * to avoid overloading existing workqueue. Overloading the
1727          * old workqueue would also create a bottleneck in the
1728          * path of the storage hardware device activation.
1729          */
1730         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1731                                                   WQ_MEM_RECLAIM);
1732         if (!kmpath_handlerd) {
1733                 DMERR("failed to create workqueue kmpath_handlerd");
1734                 destroy_workqueue(kmultipathd);
1735                 dm_unregister_target(&multipath_target);
1736                 kmem_cache_destroy(_mpio_cache);
1737                 return -ENOMEM;
1738         }
1739
1740         DMINFO("version %u.%u.%u loaded",
1741                multipath_target.version[0], multipath_target.version[1],
1742                multipath_target.version[2]);
1743
1744         return r;
1745 }
1746
1747 static void __exit dm_multipath_exit(void)
1748 {
1749         destroy_workqueue(kmpath_handlerd);
1750         destroy_workqueue(kmultipathd);
1751
1752         dm_unregister_target(&multipath_target);
1753         kmem_cache_destroy(_mpio_cache);
1754 }
1755
1756 module_init(dm_multipath_init);
1757 module_exit(dm_multipath_exit);
1758
1759 MODULE_DESCRIPTION(DM_NAME " multipath target");
1760 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1761 MODULE_LICENSE("GPL");