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 <linux/delay.h>
  22 #include <scsi/scsi_dh.h>
  23 #include <linux/atomic.h>
  24
  25 #define DM_MSG_PREFIX "multipath"
  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         const char *hw_handler_name;
  66         char *hw_handler_params;
  67
  68         spinlock_t lock;
  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:1;            /* Must we queue all I/O? */
  86         unsigned queue_if_no_path:1;    /* Queue I/O if last path fails? */
  87         unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
  88         unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
  89
  90         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  91         unsigned pg_init_count;         /* Number of times pg_init called */
  92         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  93
  94         unsigned queue_size;
  95         struct work_struct process_queued_ios;
  96         struct list_head queued_ios;
  97
  98         struct work_struct trigger_event;
  99
 100         /*
 101          * We must use a mempool of dm_mpath_io structs so that we
 102          * can resubmit bios on error.
 103          */
 104         mempool_t *mpio_pool;
 105
 106         struct mutex work_mutex;
 107 };
 108
 109 /*
 110  * Context information attached to each bio we process.
 111  */
 112 struct dm_mpath_io {
 113         struct pgpath *pgpath;
 114         size_t nr_bytes;
 115 };
 116
 117 typedef int (*action_fn) (struct pgpath *pgpath);
 118
 119 #define MIN_IOS 256     /* Mempool size */
 120
 121 static struct kmem_cache *_mpio_cache;
 122
 123 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 124 static void process_queued_ios(struct work_struct *work);
 125 static void trigger_event(struct work_struct *work);
 126 static void activate_path(struct work_struct *work);
 127
 128
 129 /*-----------------------------------------------
 130  * Allocation routines
 131  *-----------------------------------------------*/
 132
 133 static struct pgpath *alloc_pgpath(void)
 134 {
 135         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 136
 137         if (pgpath) {
 138                 pgpath->is_active = 1;
 139                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 140         }
 141
 142         return pgpath;
 143 }
 144
 145 static void free_pgpath(struct pgpath *pgpath)
 146 {
 147         kfree(pgpath);
 148 }
 149
 150 static struct priority_group *alloc_priority_group(void)
 151 {
 152         struct priority_group *pg;
 153
 154         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 155
 156         if (pg)
 157                 INIT_LIST_HEAD(&pg->pgpaths);
 158
 159         return pg;
 160 }
 161
 162 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 163 {
 164         struct pgpath *pgpath, *tmp;
 165         struct multipath *m = ti->private;
 166
 167         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 168                 list_del(&pgpath->list);
 169                 if (m->hw_handler_name)
 170                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 171                 dm_put_device(ti, pgpath->path.dev);
 172                 free_pgpath(pgpath);
 173         }
 174 }
 175
 176 static void free_priority_group(struct priority_group *pg,
 177                                 struct dm_target *ti)
 178 {
 179         struct path_selector *ps = &pg->ps;
 180
 181         if (ps->type) {
 182                 ps->type->destroy(ps);
 183                 dm_put_path_selector(ps->type);
 184         }
 185
 186         free_pgpaths(&pg->pgpaths, ti);
 187         kfree(pg);
 188 }
 189
 190 static struct multipath *alloc_multipath(struct dm_target *ti)
 191 {
 192         struct multipath *m;
 193
 194         m = kzalloc(sizeof(*m), GFP_KERNEL);
 195         if (m) {
 196                 INIT_LIST_HEAD(&m->priority_groups);
 197                 INIT_LIST_HEAD(&m->queued_ios);
 198                 spin_lock_init(&m->lock);
 199                 m->queue_io = 1;
 200                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 201                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 202                 INIT_WORK(&m->trigger_event, trigger_event);
 203                 init_waitqueue_head(&m->pg_init_wait);
 204                 mutex_init(&m->work_mutex);
 205                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 206                 if (!m->mpio_pool) {
 207                         kfree(m);
 208                         return NULL;
 209                 }
 210                 m->ti = ti;
 211                 ti->private = m;
 212         }
 213
 214         return m;
 215 }
 216
 217 static void free_multipath(struct multipath *m)
 218 {
 219         struct priority_group *pg, *tmp;
 220
 221         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 222                 list_del(&pg->list);
 223                 free_priority_group(pg, m->ti);
 224         }
 225
 226         kfree(m->hw_handler_name);
 227         kfree(m->hw_handler_params);
 228         mempool_destroy(m->mpio_pool);
 229         kfree(m);
 230 }
 231
 232 static int set_mapinfo(struct multipath *m, union map_info *info)
 233 {
 234         struct dm_mpath_io *mpio;
 235
 236         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 237         if (!mpio)
 238                 return -ENOMEM;
 239
 240         memset(mpio, 0, sizeof(*mpio));
 241         info->ptr = mpio;
 242
 243         return 0;
 244 }
 245
 246 static void clear_mapinfo(struct multipath *m, union map_info *info)
 247 {
 248         struct dm_mpath_io *mpio = info->ptr;
 249
 250         info->ptr = NULL;
 251         mempool_free(mpio, m->mpio_pool);
 252 }
 253
 254 /*-----------------------------------------------
 255  * Path selection
 256  *-----------------------------------------------*/
 257
 258 static void __pg_init_all_paths(struct multipath *m)
 259 {
 260         struct pgpath *pgpath;
 261         unsigned long pg_init_delay = 0;
 262
 263         m->pg_init_count++;
 264         m->pg_init_required = 0;
 265         if (m->pg_init_delay_retry)
 266                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 267                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 268         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 269                 /* Skip failed paths */
 270                 if (!pgpath->is_active)
 271                         continue;
 272                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 273                                        pg_init_delay))
 274                         m->pg_init_in_progress++;
 275         }
 276 }
 277
 278 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 279 {
 280         m->current_pg = pgpath->pg;
 281
 282         /* Must we initialise the PG first, and queue I/O till it's ready? */
 283         if (m->hw_handler_name) {
 284                 m->pg_init_required = 1;
 285                 m->queue_io = 1;
 286         } else {
 287                 m->pg_init_required = 0;
 288                 m->queue_io = 0;
 289         }
 290
 291         m->pg_init_count = 0;
 292 }
 293
 294 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 295                                size_t nr_bytes)
 296 {
 297         struct dm_path *path;
 298
 299         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 300         if (!path)
 301                 return -ENXIO;
 302
 303         m->current_pgpath = path_to_pgpath(path);
 304
 305         if (m->current_pg != pg)
 306                 __switch_pg(m, m->current_pgpath);
 307
 308         return 0;
 309 }
 310
 311 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 312 {
 313         struct priority_group *pg;
 314         unsigned bypassed = 1;
 315
 316         if (!m->nr_valid_paths)
 317                 goto failed;
 318
 319         /* Were we instructed to switch PG? */
 320         if (m->next_pg) {
 321                 pg = m->next_pg;
 322                 m->next_pg = NULL;
 323                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 324                         return;
 325         }
 326
 327         /* Don't change PG until it has no remaining paths */
 328         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 329                 return;
 330
 331         /*
 332          * Loop through priority groups until we find a valid path.
 333          * First time we skip PGs marked 'bypassed'.
 334          * Second time we only try the ones we skipped, but set
 335          * pg_init_delay_retry so we do not hammer controllers.
 336          */
 337         do {
 338                 list_for_each_entry(pg, &m->priority_groups, list) {
 339                         if (pg->bypassed == bypassed)
 340                                 continue;
 341                         if (!__choose_path_in_pg(m, pg, nr_bytes)) {
 342                                 if (!bypassed)
 343                                         m->pg_init_delay_retry = 1;
 344                                 return;
 345                         }
 346                 }
 347         } while (bypassed--);
 348
 349 failed:
 350         m->current_pgpath = NULL;
 351         m->current_pg = NULL;
 352 }
 353
 354 /*
 355  * Check whether bios must be queued in the device-mapper core rather
 356  * than here in the target.
 357  *
 358  * m->lock must be held on entry.
 359  *
 360  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 361  * same value then we are not between multipath_presuspend()
 362  * and multipath_resume() calls and we have no need to check
 363  * for the DMF_NOFLUSH_SUSPENDING flag.
 364  */
 365 static int __must_push_back(struct multipath *m)
 366 {
 367         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 368                 dm_noflush_suspending(m->ti));
 369 }
 370
 371 static int map_io(struct multipath *m, struct request *clone,
 372                   union map_info *map_context, unsigned was_queued)
 373 {
 374         int r = DM_MAPIO_REMAPPED;
 375         size_t nr_bytes = blk_rq_bytes(clone);
 376         unsigned long flags;
 377         struct pgpath *pgpath;
 378         struct block_device *bdev;
 379         struct dm_mpath_io *mpio = map_context->ptr;
 380
 381         spin_lock_irqsave(&m->lock, flags);
 382
 383         /* Do we need to select a new pgpath? */
 384         if (!m->current_pgpath ||
 385             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 386                 __choose_pgpath(m, nr_bytes);
 387
 388         pgpath = m->current_pgpath;
 389
 390         if (was_queued)
 391                 m->queue_size--;
 392
 393         if ((pgpath && m->queue_io) ||
 394             (!pgpath && m->queue_if_no_path)) {
 395                 /* Queue for the daemon to resubmit */
 396                 list_add_tail(&clone->queuelist, &m->queued_ios);
 397                 m->queue_size++;
 398                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 399                     !m->queue_io)
 400                         queue_work(kmultipathd, &m->process_queued_ios);
 401                 pgpath = NULL;
 402                 r = DM_MAPIO_SUBMITTED;
 403         } else if (pgpath) {
 404                 bdev = pgpath->path.dev->bdev;
 405                 clone->q = bdev_get_queue(bdev);
 406                 clone->rq_disk = bdev->bd_disk;
 407         } else if (__must_push_back(m))
 408                 r = DM_MAPIO_REQUEUE;
 409         else
 410                 r = -EIO;       /* Failed */
 411
 412         mpio->pgpath = pgpath;
 413         mpio->nr_bytes = nr_bytes;
 414
 415         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 416                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 417                                               nr_bytes);
 418
 419         spin_unlock_irqrestore(&m->lock, flags);
 420
 421         return r;
 422 }
 423
 424 /*
 425  * If we run out of usable paths, should we queue I/O or error it?
 426  */
 427 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 428                             unsigned save_old_value)
 429 {
 430         unsigned long flags;
 431
 432         spin_lock_irqsave(&m->lock, flags);
 433
 434         if (save_old_value)
 435                 m->saved_queue_if_no_path = m->queue_if_no_path;
 436         else
 437                 m->saved_queue_if_no_path = queue_if_no_path;
 438         m->queue_if_no_path = queue_if_no_path;
 439         if (!m->queue_if_no_path && m->queue_size)
 440                 queue_work(kmultipathd, &m->process_queued_ios);
 441
 442         spin_unlock_irqrestore(&m->lock, flags);
 443
 444         return 0;
 445 }
 446
 447 /*-----------------------------------------------------------------
 448  * The multipath daemon is responsible for resubmitting queued ios.
 449  *---------------------------------------------------------------*/
 450
 451 static void dispatch_queued_ios(struct multipath *m)
 452 {
 453         int r;
 454         unsigned long flags;
 455         union map_info *info;
 456         struct request *clone, *n;
 457         LIST_HEAD(cl);
 458
 459         spin_lock_irqsave(&m->lock, flags);
 460         list_splice_init(&m->queued_ios, &cl);
 461         spin_unlock_irqrestore(&m->lock, flags);
 462
 463         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 464                 list_del_init(&clone->queuelist);
 465
 466                 info = dm_get_rq_mapinfo(clone);
 467
 468                 r = map_io(m, clone, info, 1);
 469                 if (r < 0) {
 470                         clear_mapinfo(m, info);
 471                         dm_kill_unmapped_request(clone, r);
 472                 } else if (r == DM_MAPIO_REMAPPED)
 473                         dm_dispatch_request(clone);
 474                 else if (r == DM_MAPIO_REQUEUE) {
 475                         clear_mapinfo(m, info);
 476                         dm_requeue_unmapped_request(clone);
 477                 }
 478         }
 479 }
 480
 481 static void process_queued_ios(struct work_struct *work)
 482 {
 483         struct multipath *m =
 484                 container_of(work, struct multipath, process_queued_ios);
 485         struct pgpath *pgpath = NULL;
 486         unsigned must_queue = 1;
 487         unsigned long flags;
 488
 489         spin_lock_irqsave(&m->lock, flags);
 490
 491         if (!m->current_pgpath)
 492                 __choose_pgpath(m, 0);
 493
 494         pgpath = m->current_pgpath;
 495
 496         if ((pgpath && !m->queue_io) ||
 497             (!pgpath && !m->queue_if_no_path))
 498                 must_queue = 0;
 499
 500         if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
 501                 __pg_init_all_paths(m);
 502
 503         spin_unlock_irqrestore(&m->lock, flags);
 504         if (!must_queue)
 505                 dispatch_queued_ios(m);
 506 }
 507
 508 /*
 509  * An event is triggered whenever a path is taken out of use.
 510  * Includes path failure and PG bypass.
 511  */
 512 static void trigger_event(struct work_struct *work)
 513 {
 514         struct multipath *m =
 515                 container_of(work, struct multipath, trigger_event);
 516
 517         dm_table_event(m->ti->table);
 518 }
 519
 520 /*-----------------------------------------------------------------
 521  * Constructor/argument parsing:
 522  * <#multipath feature args> [<arg>]*
 523  * <#hw_handler args> [hw_handler [<arg>]*]
 524  * <#priority groups>
 525  * <initial priority group>
 526  *     [<selector> <#selector args> [<arg>]*
 527  *      <#paths> <#per-path selector args>
 528  *         [<path> [<arg>]* ]+ ]+
 529  *---------------------------------------------------------------*/
 530 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 531                                struct dm_target *ti)
 532 {
 533         int r;
 534         struct path_selector_type *pst;
 535         unsigned ps_argc;
 536
 537         static struct dm_arg _args[] = {
 538                 {0, 1024, "invalid number of path selector args"},
 539         };
 540
 541         pst = dm_get_path_selector(dm_shift_arg(as));
 542         if (!pst) {
 543                 ti->error = "unknown path selector type";
 544                 return -EINVAL;
 545         }
 546
 547         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 548         if (r) {
 549                 dm_put_path_selector(pst);
 550                 return -EINVAL;
 551         }
 552
 553         r = pst->create(&pg->ps, ps_argc, as->argv);
 554         if (r) {
 555                 dm_put_path_selector(pst);
 556                 ti->error = "path selector constructor failed";
 557                 return r;
 558         }
 559
 560         pg->ps.type = pst;
 561         dm_consume_args(as, ps_argc);
 562
 563         return 0;
 564 }
 565
 566 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 567                                struct dm_target *ti)
 568 {
 569         int r;
 570         struct pgpath *p;
 571         struct multipath *m = ti->private;
 572         struct request_queue *q = NULL;
 573         const char *attached_handler_name;
 574
 575         /* we need at least a path arg */
 576         if (as->argc < 1) {
 577                 ti->error = "no device given";
 578                 return ERR_PTR(-EINVAL);
 579         }
 580
 581         p = alloc_pgpath();
 582         if (!p)
 583                 return ERR_PTR(-ENOMEM);
 584
 585         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 586                           &p->path.dev);
 587         if (r) {
 588                 ti->error = "error getting device";
 589                 goto bad;
 590         }
 591
 592         if (m->retain_attached_hw_handler || m->hw_handler_name)
 593                 q = bdev_get_queue(p->path.dev->bdev);
 594
 595         if (m->retain_attached_hw_handler) {
 596                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 597                 if (attached_handler_name) {
 598                         /*
 599                          * Reset hw_handler_name to match the attached handler
 600                          * and clear any hw_handler_params associated with the
 601                          * ignored handler.
 602                          *
 603                          * NB. This modifies the table line to show the actual
 604                          * handler instead of the original table passed in.
 605                          */
 606                         kfree(m->hw_handler_name);
 607                         m->hw_handler_name = attached_handler_name;
 608
 609                         kfree(m->hw_handler_params);
 610                         m->hw_handler_params = NULL;
 611                 }
 612         }
 613
 614         if (m->hw_handler_name) {
 615                 /*
 616                  * Increments scsi_dh reference, even when using an
 617                  * already-attached handler.
 618                  */
 619                 r = scsi_dh_attach(q, m->hw_handler_name);
 620                 if (r == -EBUSY) {
 621                         /*
 622                          * Already attached to different hw_handler:
 623                          * try to reattach with correct one.
 624                          */
 625                         scsi_dh_detach(q);
 626                         r = scsi_dh_attach(q, m->hw_handler_name);
 627                 }
 628
 629                 if (r < 0) {
 630                         ti->error = "error attaching hardware handler";
 631                         dm_put_device(ti, p->path.dev);
 632                         goto bad;
 633                 }
 634
 635                 if (m->hw_handler_params) {
 636                         r = scsi_dh_set_params(q, m->hw_handler_params);
 637                         if (r < 0) {
 638                                 ti->error = "unable to set hardware "
 639                                                         "handler parameters";
 640                                 scsi_dh_detach(q);
 641                                 dm_put_device(ti, p->path.dev);
 642                                 goto bad;
 643                         }
 644                 }
 645         }
 646
 647         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 648         if (r) {
 649                 dm_put_device(ti, p->path.dev);
 650                 goto bad;
 651         }
 652
 653         return p;
 654
 655  bad:
 656         free_pgpath(p);
 657         return ERR_PTR(r);
 658 }
 659
 660 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 661                                                    struct multipath *m)
 662 {
 663         static struct dm_arg _args[] = {
 664                 {1, 1024, "invalid number of paths"},
 665                 {0, 1024, "invalid number of selector args"}
 666         };
 667
 668         int r;
 669         unsigned i, nr_selector_args, nr_args;
 670         struct priority_group *pg;
 671         struct dm_target *ti = m->ti;
 672
 673         if (as->argc < 2) {
 674                 as->argc = 0;
 675                 ti->error = "not enough priority group arguments";
 676                 return ERR_PTR(-EINVAL);
 677         }
 678
 679         pg = alloc_priority_group();
 680         if (!pg) {
 681                 ti->error = "couldn't allocate priority group";
 682                 return ERR_PTR(-ENOMEM);
 683         }
 684         pg->m = m;
 685
 686         r = parse_path_selector(as, pg, ti);
 687         if (r)
 688                 goto bad;
 689
 690         /*
 691          * read the paths
 692          */
 693         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 694         if (r)
 695                 goto bad;
 696
 697         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 698         if (r)
 699                 goto bad;
 700
 701         nr_args = 1 + nr_selector_args;
 702         for (i = 0; i < pg->nr_pgpaths; i++) {
 703                 struct pgpath *pgpath;
 704                 struct dm_arg_set path_args;
 705
 706                 if (as->argc < nr_args) {
 707                         ti->error = "not enough path parameters";
 708                         r = -EINVAL;
 709                         goto bad;
 710                 }
 711
 712                 path_args.argc = nr_args;
 713                 path_args.argv = as->argv;
 714
 715                 pgpath = parse_path(&path_args, &pg->ps, ti);
 716                 if (IS_ERR(pgpath)) {
 717                         r = PTR_ERR(pgpath);
 718                         goto bad;
 719                 }
 720
 721                 pgpath->pg = pg;
 722                 list_add_tail(&pgpath->list, &pg->pgpaths);
 723                 dm_consume_args(as, nr_args);
 724         }
 725
 726         return pg;
 727
 728  bad:
 729         free_priority_group(pg, ti);
 730         return ERR_PTR(r);
 731 }
 732
 733 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 734 {
 735         unsigned hw_argc;
 736         int ret;
 737         struct dm_target *ti = m->ti;
 738
 739         static struct dm_arg _args[] = {
 740                 {0, 1024, "invalid number of hardware handler args"},
 741         };
 742
 743         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 744                 return -EINVAL;
 745
 746         if (!hw_argc)
 747                 return 0;
 748
 749         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 750         if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
 751                                      "scsi_dh_%s", m->hw_handler_name)) {
 752                 ti->error = "unknown hardware handler type";
 753                 ret = -EINVAL;
 754                 goto fail;
 755         }
 756
 757         if (hw_argc > 1) {
 758                 char *p;
 759                 int i, j, len = 4;
 760
 761                 for (i = 0; i <= hw_argc - 2; i++)
 762                         len += strlen(as->argv[i]) + 1;
 763                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 764                 if (!p) {
 765                         ti->error = "memory allocation failed";
 766                         ret = -ENOMEM;
 767                         goto fail;
 768                 }
 769                 j = sprintf(p, "%d", hw_argc - 1);
 770                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 771                         j = sprintf(p, "%s", as->argv[i]);
 772         }
 773         dm_consume_args(as, hw_argc - 1);
 774
 775         return 0;
 776 fail:
 777         kfree(m->hw_handler_name);
 778         m->hw_handler_name = NULL;
 779         return ret;
 780 }
 781
 782 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 783 {
 784         int r;
 785         unsigned argc;
 786         struct dm_target *ti = m->ti;
 787         const char *arg_name;
 788
 789         static struct dm_arg _args[] = {
 790                 {0, 6, "invalid number of feature args"},
 791                 {1, 50, "pg_init_retries must be between 1 and 50"},
 792                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 793         };
 794
 795         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 796         if (r)
 797                 return -EINVAL;
 798
 799         if (!argc)
 800                 return 0;
 801
 802         do {
 803                 arg_name = dm_shift_arg(as);
 804                 argc--;
 805
 806                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 807                         r = queue_if_no_path(m, 1, 0);
 808                         continue;
 809                 }
 810
 811                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
 812                         m->retain_attached_hw_handler = 1;
 813                         continue;
 814                 }
 815
 816                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 817                     (argc >= 1)) {
 818                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 819                         argc--;
 820                         continue;
 821                 }
 822
 823                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 824                     (argc >= 1)) {
 825                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 826                         argc--;
 827                         continue;
 828                 }
 829
 830                 ti->error = "Unrecognised multipath feature request";
 831                 r = -EINVAL;
 832         } while (argc && !r);
 833
 834         return r;
 835 }
 836
 837 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 838                          char **argv)
 839 {
 840         /* target arguments */
 841         static struct dm_arg _args[] = {
 842                 {0, 1024, "invalid number of priority groups"},
 843                 {0, 1024, "invalid initial priority group number"},
 844         };
 845
 846         int r;
 847         struct multipath *m;
 848         struct dm_arg_set as;
 849         unsigned pg_count = 0;
 850         unsigned next_pg_num;
 851
 852         as.argc = argc;
 853         as.argv = argv;
 854
 855         m = alloc_multipath(ti);
 856         if (!m) {
 857                 ti->error = "can't allocate multipath";
 858                 return -EINVAL;
 859         }
 860
 861         r = parse_features(&as, m);
 862         if (r)
 863                 goto bad;
 864
 865         r = parse_hw_handler(&as, m);
 866         if (r)
 867                 goto bad;
 868
 869         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 870         if (r)
 871                 goto bad;
 872
 873         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 874         if (r)
 875                 goto bad;
 876
 877         if ((!m->nr_priority_groups && next_pg_num) ||
 878             (m->nr_priority_groups && !next_pg_num)) {
 879                 ti->error = "invalid initial priority group";
 880                 r = -EINVAL;
 881                 goto bad;
 882         }
 883
 884         /* parse the priority groups */
 885         while (as.argc) {
 886                 struct priority_group *pg;
 887
 888                 pg = parse_priority_group(&as, m);
 889                 if (IS_ERR(pg)) {
 890                         r = PTR_ERR(pg);
 891                         goto bad;
 892                 }
 893
 894                 m->nr_valid_paths += pg->nr_pgpaths;
 895                 list_add_tail(&pg->list, &m->priority_groups);
 896                 pg_count++;
 897                 pg->pg_num = pg_count;
 898                 if (!--next_pg_num)
 899                         m->next_pg = pg;
 900         }
 901
 902         if (pg_count != m->nr_priority_groups) {
 903                 ti->error = "priority group count mismatch";
 904                 r = -EINVAL;
 905                 goto bad;
 906         }
 907
 908         ti->num_flush_bios = 1;
 909         ti->num_discard_bios = 1;
 910         ti->num_write_same_bios = 1;
 911
 912         return 0;
 913
 914  bad:
 915         free_multipath(m);
 916         return r;
 917 }
 918
 919 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 920 {
 921         DECLARE_WAITQUEUE(wait, current);
 922         unsigned long flags;
 923
 924         add_wait_queue(&m->pg_init_wait, &wait);
 925
 926         while (1) {
 927                 set_current_state(TASK_UNINTERRUPTIBLE);
 928
 929                 spin_lock_irqsave(&m->lock, flags);
 930                 if (!m->pg_init_in_progress) {
 931                         spin_unlock_irqrestore(&m->lock, flags);
 932                         break;
 933                 }
 934                 spin_unlock_irqrestore(&m->lock, flags);
 935
 936                 io_schedule();
 937         }
 938         set_current_state(TASK_RUNNING);
 939
 940         remove_wait_queue(&m->pg_init_wait, &wait);
 941 }
 942
 943 static void flush_multipath_work(struct multipath *m)
 944 {
 945         flush_workqueue(kmpath_handlerd);
 946         multipath_wait_for_pg_init_completion(m);
 947         flush_workqueue(kmultipathd);
 948         flush_work(&m->trigger_event);
 949 }
 950
 951 static void multipath_dtr(struct dm_target *ti)
 952 {
 953         struct multipath *m = ti->private;
 954
 955         flush_multipath_work(m);
 956         free_multipath(m);
 957 }
 958
 959 /*
 960  * Map cloned requests
 961  */
 962 static int multipath_map(struct dm_target *ti, struct request *clone,
 963                          union map_info *map_context)
 964 {
 965         int r;
 966         struct multipath *m = (struct multipath *) ti->private;
 967
 968         if (set_mapinfo(m, map_context) < 0)
 969                 /* ENOMEM, requeue */
 970                 return DM_MAPIO_REQUEUE;
 971
 972         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 973         r = map_io(m, clone, map_context, 0);
 974         if (r < 0 || r == DM_MAPIO_REQUEUE)
 975                 clear_mapinfo(m, map_context);
 976
 977         return r;
 978 }
 979
 980 /*
 981  * Take a path out of use.
 982  */
 983 static int fail_path(struct pgpath *pgpath)
 984 {
 985         unsigned long flags;
 986         struct multipath *m = pgpath->pg->m;
 987
 988         spin_lock_irqsave(&m->lock, flags);
 989
 990         if (!pgpath->is_active)
 991                 goto out;
 992
 993         DMWARN("Failing path %s.", pgpath->path.dev->name);
 994
 995         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 996         pgpath->is_active = 0;
 997         pgpath->fail_count++;
 998
 999         m->nr_valid_paths--;
1000
1001         if (pgpath == m->current_pgpath)
1002                 m->current_pgpath = NULL;
1003
1004         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1005                       pgpath->path.dev->name, m->nr_valid_paths);
1006
1007         schedule_work(&m->trigger_event);
1008
1009 out:
1010         spin_unlock_irqrestore(&m->lock, flags);
1011
1012         return 0;
1013 }
1014
1015 /*
1016  * Reinstate a previously-failed path
1017  */
1018 static int reinstate_path(struct pgpath *pgpath)
1019 {
1020         int r = 0;
1021         unsigned long flags;
1022         struct multipath *m = pgpath->pg->m;
1023
1024         spin_lock_irqsave(&m->lock, flags);
1025
1026         if (pgpath->is_active)
1027                 goto out;
1028
1029         if (!pgpath->pg->ps.type->reinstate_path) {
1030                 DMWARN("Reinstate path not supported by path selector %s",
1031                        pgpath->pg->ps.type->name);
1032                 r = -EINVAL;
1033                 goto out;
1034         }
1035
1036         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1037         if (r)
1038                 goto out;
1039
1040         pgpath->is_active = 1;
1041
1042         if (!m->nr_valid_paths++ && m->queue_size) {
1043                 m->current_pgpath = NULL;
1044                 queue_work(kmultipathd, &m->process_queued_ios);
1045         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1046                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1047                         m->pg_init_in_progress++;
1048         }
1049
1050         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1051                       pgpath->path.dev->name, m->nr_valid_paths);
1052
1053         schedule_work(&m->trigger_event);
1054
1055 out:
1056         spin_unlock_irqrestore(&m->lock, flags);
1057
1058         return r;
1059 }
1060
1061 /*
1062  * Fail or reinstate all paths that match the provided struct dm_dev.
1063  */
1064 static int action_dev(struct multipath *m, struct dm_dev *dev,
1065                       action_fn action)
1066 {
1067         int r = -EINVAL;
1068         struct pgpath *pgpath;
1069         struct priority_group *pg;
1070
1071         list_for_each_entry(pg, &m->priority_groups, list) {
1072                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1073                         if (pgpath->path.dev == dev)
1074                                 r = action(pgpath);
1075                 }
1076         }
1077
1078         return r;
1079 }
1080
1081 /*
1082  * Temporarily try to avoid having to use the specified PG
1083  */
1084 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1085                       int bypassed)
1086 {
1087         unsigned long flags;
1088
1089         spin_lock_irqsave(&m->lock, flags);
1090
1091         pg->bypassed = bypassed;
1092         m->current_pgpath = NULL;
1093         m->current_pg = NULL;
1094
1095         spin_unlock_irqrestore(&m->lock, flags);
1096
1097         schedule_work(&m->trigger_event);
1098 }
1099
1100 /*
1101  * Switch to using the specified PG from the next I/O that gets mapped
1102  */
1103 static int switch_pg_num(struct multipath *m, const char *pgstr)
1104 {
1105         struct priority_group *pg;
1106         unsigned pgnum;
1107         unsigned long flags;
1108         char dummy;
1109
1110         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1111             (pgnum > m->nr_priority_groups)) {
1112                 DMWARN("invalid PG number supplied to switch_pg_num");
1113                 return -EINVAL;
1114         }
1115
1116         spin_lock_irqsave(&m->lock, flags);
1117         list_for_each_entry(pg, &m->priority_groups, list) {
1118                 pg->bypassed = 0;
1119                 if (--pgnum)
1120                         continue;
1121
1122                 m->current_pgpath = NULL;
1123                 m->current_pg = NULL;
1124                 m->next_pg = pg;
1125         }
1126         spin_unlock_irqrestore(&m->lock, flags);
1127
1128         schedule_work(&m->trigger_event);
1129         return 0;
1130 }
1131
1132 /*
1133  * Set/clear bypassed status of a PG.
1134  * PGs are numbered upwards from 1 in the order they were declared.
1135  */
1136 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1137 {
1138         struct priority_group *pg;
1139         unsigned pgnum;
1140         char dummy;
1141
1142         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1143             (pgnum > m->nr_priority_groups)) {
1144                 DMWARN("invalid PG number supplied to bypass_pg");
1145                 return -EINVAL;
1146         }
1147
1148         list_for_each_entry(pg, &m->priority_groups, list) {
1149                 if (!--pgnum)
1150                         break;
1151         }
1152
1153         bypass_pg(m, pg, bypassed);
1154         return 0;
1155 }
1156
1157 /*
1158  * Should we retry pg_init immediately?
1159  */
1160 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1161 {
1162         unsigned long flags;
1163         int limit_reached = 0;
1164
1165         spin_lock_irqsave(&m->lock, flags);
1166
1167         if (m->pg_init_count <= m->pg_init_retries)
1168                 m->pg_init_required = 1;
1169         else
1170                 limit_reached = 1;
1171
1172         spin_unlock_irqrestore(&m->lock, flags);
1173
1174         return limit_reached;
1175 }
1176
1177 static void pg_init_done(void *data, int errors)
1178 {
1179         struct pgpath *pgpath = data;
1180         struct priority_group *pg = pgpath->pg;
1181         struct multipath *m = pg->m;
1182         unsigned long flags;
1183         unsigned delay_retry = 0;
1184
1185         /* device or driver problems */
1186         switch (errors) {
1187         case SCSI_DH_OK:
1188                 break;
1189         case SCSI_DH_NOSYS:
1190                 if (!m->hw_handler_name) {
1191                         errors = 0;
1192                         break;
1193                 }
1194                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1195                       "Error %d.", m->hw_handler_name, errors);
1196                 /*
1197                  * Fail path for now, so we do not ping pong
1198                  */
1199                 fail_path(pgpath);
1200                 break;
1201         case SCSI_DH_DEV_TEMP_BUSY:
1202                 /*
1203                  * Probably doing something like FW upgrade on the
1204                  * controller so try the other pg.
1205                  */
1206                 bypass_pg(m, pg, 1);
1207                 break;
1208         case SCSI_DH_RETRY:
1209                 /* Wait before retrying. */
1210                 delay_retry = 1;
1211         case SCSI_DH_IMM_RETRY:
1212         case SCSI_DH_RES_TEMP_UNAVAIL:
1213                 if (pg_init_limit_reached(m, pgpath))
1214                         fail_path(pgpath);
1215                 errors = 0;
1216                 break;
1217         default:
1218                 /*
1219                  * We probably do not want to fail the path for a device
1220                  * error, but this is what the old dm did. In future
1221                  * patches we can do more advanced handling.
1222                  */
1223                 fail_path(pgpath);
1224         }
1225
1226         spin_lock_irqsave(&m->lock, flags);
1227         if (errors) {
1228                 if (pgpath == m->current_pgpath) {
1229                         DMERR("Could not failover device. Error %d.", errors);
1230                         m->current_pgpath = NULL;
1231                         m->current_pg = NULL;
1232                 }
1233         } else if (!m->pg_init_required)
1234                 pg->bypassed = 0;
1235
1236         if (--m->pg_init_in_progress)
1237                 /* Activations of other paths are still on going */
1238                 goto out;
1239
1240         if (!m->pg_init_required)
1241                 m->queue_io = 0;
1242
1243         m->pg_init_delay_retry = delay_retry;
1244         queue_work(kmultipathd, &m->process_queued_ios);
1245
1246         /*
1247          * Wake up any thread waiting to suspend.
1248          */
1249         wake_up(&m->pg_init_wait);
1250
1251 out:
1252         spin_unlock_irqrestore(&m->lock, flags);
1253 }
1254
1255 static void activate_path(struct work_struct *work)
1256 {
1257         struct pgpath *pgpath =
1258                 container_of(work, struct pgpath, activate_path.work);
1259
1260         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1261                                 pg_init_done, pgpath);
1262 }
1263
1264 /*
1265  * end_io handling
1266  */
1267 static int do_end_io(struct multipath *m, struct request *clone,
1268                      int error, struct dm_mpath_io *mpio)
1269 {
1270         /*
1271          * We don't queue any clone request inside the multipath target
1272          * during end I/O handling, since those clone requests don't have
1273          * bio clones.  If we queue them inside the multipath target,
1274          * we need to make bio clones, that requires memory allocation.
1275          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1276          *  don't have bio clones.)
1277          * Instead of queueing the clone request here, we queue the original
1278          * request into dm core, which will remake a clone request and
1279          * clone bios for it and resubmit it later.
1280          */
1281         int r = DM_ENDIO_REQUEUE;
1282         unsigned long flags;
1283
1284         if (!error && !clone->errors)
1285                 return 0;       /* I/O complete */
1286
1287         if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1288                 return error;
1289
1290         if (mpio->pgpath)
1291                 fail_path(mpio->pgpath);
1292
1293         spin_lock_irqsave(&m->lock, flags);
1294         if (!m->nr_valid_paths) {
1295                 if (!m->queue_if_no_path) {
1296                         if (!__must_push_back(m))
1297                                 r = -EIO;
1298                 } else {
1299                         if (error == -EBADE)
1300                                 r = error;
1301                 }
1302         }
1303         spin_unlock_irqrestore(&m->lock, flags);
1304
1305         return r;
1306 }
1307
1308 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1309                             int error, union map_info *map_context)
1310 {
1311         struct multipath *m = ti->private;
1312         struct dm_mpath_io *mpio = map_context->ptr;
1313         struct pgpath *pgpath;
1314         struct path_selector *ps;
1315         int r;
1316
1317         BUG_ON(!mpio);
1318
1319         r  = do_end_io(m, clone, error, mpio);
1320         pgpath = mpio->pgpath;
1321         if (pgpath) {
1322                 ps = &pgpath->pg->ps;
1323                 if (ps->type->end_io)
1324                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1325         }
1326         clear_mapinfo(m, map_context);
1327
1328         return r;
1329 }
1330
1331 /*
1332  * Suspend can't complete until all the I/O is processed so if
1333  * the last path fails we must error any remaining I/O.
1334  * Note that if the freeze_bdev fails while suspending, the
1335  * queue_if_no_path state is lost - userspace should reset it.
1336  */
1337 static void multipath_presuspend(struct dm_target *ti)
1338 {
1339         struct multipath *m = (struct multipath *) ti->private;
1340
1341         queue_if_no_path(m, 0, 1);
1342 }
1343
1344 static void multipath_postsuspend(struct dm_target *ti)
1345 {
1346         struct multipath *m = ti->private;
1347
1348         mutex_lock(&m->work_mutex);
1349         flush_multipath_work(m);
1350         mutex_unlock(&m->work_mutex);
1351 }
1352
1353 /*
1354  * Restore the queue_if_no_path setting.
1355  */
1356 static void multipath_resume(struct dm_target *ti)
1357 {
1358         struct multipath *m = (struct multipath *) ti->private;
1359         unsigned long flags;
1360
1361         spin_lock_irqsave(&m->lock, flags);
1362         m->queue_if_no_path = m->saved_queue_if_no_path;
1363         spin_unlock_irqrestore(&m->lock, flags);
1364 }
1365
1366 /*
1367  * Info output has the following format:
1368  * num_multipath_feature_args [multipath_feature_args]*
1369  * num_handler_status_args [handler_status_args]*
1370  * num_groups init_group_number
1371  *            [A|D|E num_ps_status_args [ps_status_args]*
1372  *             num_paths num_selector_args
1373  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1374  *
1375  * Table output has the following format (identical to the constructor string):
1376  * num_feature_args [features_args]*
1377  * num_handler_args hw_handler [hw_handler_args]*
1378  * num_groups init_group_number
1379  *     [priority selector-name num_ps_args [ps_args]*
1380  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1381  */
1382 static void multipath_status(struct dm_target *ti, status_type_t type,
1383                              unsigned status_flags, char *result, unsigned maxlen)
1384 {
1385         int sz = 0;
1386         unsigned long flags;
1387         struct multipath *m = (struct multipath *) ti->private;
1388         struct priority_group *pg;
1389         struct pgpath *p;
1390         unsigned pg_num;
1391         char state;
1392
1393         spin_lock_irqsave(&m->lock, flags);
1394
1395         /* Features */
1396         if (type == STATUSTYPE_INFO)
1397                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1398         else {
1399                 DMEMIT("%u ", m->queue_if_no_path +
1400                               (m->pg_init_retries > 0) * 2 +
1401                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1402                               m->retain_attached_hw_handler);
1403                 if (m->queue_if_no_path)
1404                         DMEMIT("queue_if_no_path ");
1405                 if (m->pg_init_retries)
1406                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1407                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1408                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1409                 if (m->retain_attached_hw_handler)
1410                         DMEMIT("retain_attached_hw_handler ");
1411         }
1412
1413         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1414                 DMEMIT("0 ");
1415         else
1416                 DMEMIT("1 %s ", m->hw_handler_name);
1417
1418         DMEMIT("%u ", m->nr_priority_groups);
1419
1420         if (m->next_pg)
1421                 pg_num = m->next_pg->pg_num;
1422         else if (m->current_pg)
1423                 pg_num = m->current_pg->pg_num;
1424         else
1425                 pg_num = (m->nr_priority_groups ? 1 : 0);
1426
1427         DMEMIT("%u ", pg_num);
1428
1429         switch (type) {
1430         case STATUSTYPE_INFO:
1431                 list_for_each_entry(pg, &m->priority_groups, list) {
1432                         if (pg->bypassed)
1433                                 state = 'D';    /* Disabled */
1434                         else if (pg == m->current_pg)
1435                                 state = 'A';    /* Currently Active */
1436                         else
1437                                 state = 'E';    /* Enabled */
1438
1439                         DMEMIT("%c ", state);
1440
1441                         if (pg->ps.type->status)
1442                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1443                                                           result + sz,
1444                                                           maxlen - sz);
1445                         else
1446                                 DMEMIT("0 ");
1447
1448                         DMEMIT("%u %u ", pg->nr_pgpaths,
1449                                pg->ps.type->info_args);
1450
1451                         list_for_each_entry(p, &pg->pgpaths, list) {
1452                                 DMEMIT("%s %s %u ", p->path.dev->name,
1453                                        p->is_active ? "A" : "F",
1454                                        p->fail_count);
1455                                 if (pg->ps.type->status)
1456                                         sz += pg->ps.type->status(&pg->ps,
1457                                               &p->path, type, result + sz,
1458                                               maxlen - sz);
1459                         }
1460                 }
1461                 break;
1462
1463         case STATUSTYPE_TABLE:
1464                 list_for_each_entry(pg, &m->priority_groups, list) {
1465                         DMEMIT("%s ", pg->ps.type->name);
1466
1467                         if (pg->ps.type->status)
1468                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1469                                                           result + sz,
1470                                                           maxlen - sz);
1471                         else
1472                                 DMEMIT("0 ");
1473
1474                         DMEMIT("%u %u ", pg->nr_pgpaths,
1475                                pg->ps.type->table_args);
1476
1477                         list_for_each_entry(p, &pg->pgpaths, list) {
1478                                 DMEMIT("%s ", p->path.dev->name);
1479                                 if (pg->ps.type->status)
1480                                         sz += pg->ps.type->status(&pg->ps,
1481                                               &p->path, type, result + sz,
1482                                               maxlen - sz);
1483                         }
1484                 }
1485                 break;
1486         }
1487
1488         spin_unlock_irqrestore(&m->lock, flags);
1489 }
1490
1491 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1492 {
1493         int r = -EINVAL;
1494         struct dm_dev *dev;
1495         struct multipath *m = (struct multipath *) ti->private;
1496         action_fn action;
1497
1498         mutex_lock(&m->work_mutex);
1499
1500         if (dm_suspended(ti)) {
1501                 r = -EBUSY;
1502                 goto out;
1503         }
1504
1505         if (argc == 1) {
1506                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1507                         r = queue_if_no_path(m, 1, 0);
1508                         goto out;
1509                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1510                         r = queue_if_no_path(m, 0, 0);
1511                         goto out;
1512                 }
1513         }
1514
1515         if (argc != 2) {
1516                 DMWARN("Unrecognised multipath message received.");
1517                 goto out;
1518         }
1519
1520         if (!strcasecmp(argv[0], "disable_group")) {
1521                 r = bypass_pg_num(m, argv[1], 1);
1522                 goto out;
1523         } else if (!strcasecmp(argv[0], "enable_group")) {
1524                 r = bypass_pg_num(m, argv[1], 0);
1525                 goto out;
1526         } else if (!strcasecmp(argv[0], "switch_group")) {
1527                 r = switch_pg_num(m, argv[1]);
1528                 goto out;
1529         } else if (!strcasecmp(argv[0], "reinstate_path"))
1530                 action = reinstate_path;
1531         else if (!strcasecmp(argv[0], "fail_path"))
1532                 action = fail_path;
1533         else {
1534                 DMWARN("Unrecognised multipath message received.");
1535                 goto out;
1536         }
1537
1538         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1539         if (r) {
1540                 DMWARN("message: error getting device %s",
1541                        argv[1]);
1542                 goto out;
1543         }
1544
1545         r = action_dev(m, dev, action);
1546
1547         dm_put_device(ti, dev);
1548
1549 out:
1550         mutex_unlock(&m->work_mutex);
1551         return r;
1552 }
1553
1554 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1555                            unsigned long arg)
1556 {
1557         struct multipath *m = ti->private;
1558         struct pgpath *pgpath;
1559         struct block_device *bdev;
1560         fmode_t mode;
1561         unsigned long flags;
1562         int r;
1563
1564 again:
1565         bdev = NULL;
1566         mode = 0;
1567         r = 0;
1568
1569         spin_lock_irqsave(&m->lock, flags);
1570
1571         if (!m->current_pgpath)
1572                 __choose_pgpath(m, 0);
1573
1574         pgpath = m->current_pgpath;
1575
1576         if (pgpath) {
1577                 bdev = pgpath->path.dev->bdev;
1578                 mode = pgpath->path.dev->mode;
1579         }
1580
1581         if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1582                 r = -EAGAIN;
1583         else if (!bdev)
1584                 r = -EIO;
1585
1586         spin_unlock_irqrestore(&m->lock, flags);
1587
1588         /*
1589          * Only pass ioctls through if the device sizes match exactly.
1590          */
1591         if (!r && ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT)
1592                 r = scsi_verify_blk_ioctl(NULL, cmd);
1593
1594         if (r == -EAGAIN && !fatal_signal_pending(current)) {
1595                 queue_work(kmultipathd, &m->process_queued_ios);
1596                 msleep(10);
1597                 goto again;
1598         }
1599
1600         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1601 }
1602
1603 static int multipath_iterate_devices(struct dm_target *ti,
1604                                      iterate_devices_callout_fn fn, void *data)
1605 {
1606         struct multipath *m = ti->private;
1607         struct priority_group *pg;
1608         struct pgpath *p;
1609         int ret = 0;
1610
1611         list_for_each_entry(pg, &m->priority_groups, list) {
1612                 list_for_each_entry(p, &pg->pgpaths, list) {
1613                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1614                         if (ret)
1615                                 goto out;
1616                 }
1617         }
1618
1619 out:
1620         return ret;
1621 }
1622
1623 static int __pgpath_busy(struct pgpath *pgpath)
1624 {
1625         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1626
1627         return dm_underlying_device_busy(q);
1628 }
1629
1630 /*
1631  * We return "busy", only when we can map I/Os but underlying devices
1632  * are busy (so even if we map I/Os now, the I/Os will wait on
1633  * the underlying queue).
1634  * In other words, if we want to kill I/Os or queue them inside us
1635  * due to map unavailability, we don't return "busy".  Otherwise,
1636  * dm core won't give us the I/Os and we can't do what we want.
1637  */
1638 static int multipath_busy(struct dm_target *ti)
1639 {
1640         int busy = 0, has_active = 0;
1641         struct multipath *m = ti->private;
1642         struct priority_group *pg;
1643         struct pgpath *pgpath;
1644         unsigned long flags;
1645
1646         spin_lock_irqsave(&m->lock, flags);
1647
1648         /* Guess which priority_group will be used at next mapping time */
1649         if (unlikely(!m->current_pgpath && m->next_pg))
1650                 pg = m->next_pg;
1651         else if (likely(m->current_pg))
1652                 pg = m->current_pg;
1653         else
1654                 /*
1655                  * We don't know which pg will be used at next mapping time.
1656                  * We don't call __choose_pgpath() here to avoid to trigger
1657                  * pg_init just by busy checking.
1658                  * So we don't know whether underlying devices we will be using
1659                  * at next mapping time are busy or not. Just try mapping.
1660                  */
1661                 goto out;
1662
1663         /*
1664          * If there is one non-busy active path at least, the path selector
1665          * will be able to select it. So we consider such a pg as not busy.
1666          */
1667         busy = 1;
1668         list_for_each_entry(pgpath, &pg->pgpaths, list)
1669                 if (pgpath->is_active) {
1670                         has_active = 1;
1671
1672                         if (!__pgpath_busy(pgpath)) {
1673                                 busy = 0;
1674                                 break;
1675                         }
1676                 }
1677
1678         if (!has_active)
1679                 /*
1680                  * No active path in this pg, so this pg won't be used and
1681                  * the current_pg will be changed at next mapping time.
1682                  * We need to try mapping to determine it.
1683                  */
1684                 busy = 0;
1685
1686 out:
1687         spin_unlock_irqrestore(&m->lock, flags);
1688
1689         return busy;
1690 }
1691
1692 /*-----------------------------------------------------------------
1693  * Module setup
1694  *---------------------------------------------------------------*/
1695 static struct target_type multipath_target = {
1696         .name = "multipath",
1697         .version = {1, 5, 1},
1698         .module = THIS_MODULE,
1699         .ctr = multipath_ctr,
1700         .dtr = multipath_dtr,
1701         .map_rq = multipath_map,
1702         .rq_end_io = multipath_end_io,
1703         .presuspend = multipath_presuspend,
1704         .postsuspend = multipath_postsuspend,
1705         .resume = multipath_resume,
1706         .status = multipath_status,
1707         .message = multipath_message,
1708         .ioctl  = multipath_ioctl,
1709         .iterate_devices = multipath_iterate_devices,
1710         .busy = multipath_busy,
1711 };
1712
1713 static int __init dm_multipath_init(void)
1714 {
1715         int r;
1716
1717         /* allocate a slab for the dm_ios */
1718         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1719         if (!_mpio_cache)
1720                 return -ENOMEM;
1721
1722         r = dm_register_target(&multipath_target);
1723         if (r < 0) {
1724                 DMERR("register failed %d", r);
1725                 kmem_cache_destroy(_mpio_cache);
1726                 return -EINVAL;
1727         }
1728
1729         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1730         if (!kmultipathd) {
1731                 DMERR("failed to create workqueue kmpathd");
1732                 dm_unregister_target(&multipath_target);
1733                 kmem_cache_destroy(_mpio_cache);
1734                 return -ENOMEM;
1735         }
1736
1737         /*
1738          * A separate workqueue is used to handle the device handlers
1739          * to avoid overloading existing workqueue. Overloading the
1740          * old workqueue would also create a bottleneck in the
1741          * path of the storage hardware device activation.
1742          */
1743         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1744                                                   WQ_MEM_RECLAIM);
1745         if (!kmpath_handlerd) {
1746                 DMERR("failed to create workqueue kmpath_handlerd");
1747                 destroy_workqueue(kmultipathd);
1748                 dm_unregister_target(&multipath_target);
1749                 kmem_cache_destroy(_mpio_cache);
1750                 return -ENOMEM;
1751         }
1752
1753         DMINFO("version %u.%u.%u loaded",
1754                multipath_target.version[0], multipath_target.version[1],
1755                multipath_target.version[2]);
1756
1757         return r;
1758 }
1759
1760 static void __exit dm_multipath_exit(void)
1761 {
1762         destroy_workqueue(kmpath_handlerd);
1763         destroy_workqueue(kmultipathd);
1764
1765         dm_unregister_target(&multipath_target);
1766         kmem_cache_destroy(_mpio_cache);
1767 }
1768
1769 module_init(dm_multipath_init);
1770 module_exit(dm_multipath_exit);
1771
1772 MODULE_DESCRIPTION(DM_NAME " multipath target");
1773 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1774 MODULE_LICENSE("GPL");