drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9
  10 #include "dm.h"
  11 #include "dm-path-selector.h"
  12 #include "dm-uevent.h"
  13
  14 #include <linux/ctype.h>
  15 #include <linux/init.h>
  16 #include <linux/mempool.h>
  17 #include <linux/module.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/slab.h>
  20 #include <linux/time.h>
  21 #include <linux/workqueue.h>
  22 #include <linux/delay.h>
  23 #include <scsi/scsi_dh.h>
  24 #include <linux/atomic.h>
  25
  26 #define DM_MSG_PREFIX "multipath"
  27 #define DM_PG_INIT_DELAY_MSECS 2000
  28 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  29
  30 /* Path properties */
  31 struct pgpath {
  32         struct list_head list;
  33
  34         struct priority_group *pg;      /* Owning PG */
  35         unsigned is_active;             /* Path status */
  36         unsigned fail_count;            /* Cumulative failure count */
  37
  38         struct dm_path path;
  39         struct delayed_work activate_path;
  40 };
  41
  42 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  43
  44 /*
  45  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  46  * Each has a path selector which controls which path gets used.
  47  */
  48 struct priority_group {
  49         struct list_head list;
  50
  51         struct multipath *m;            /* Owning multipath instance */
  52         struct path_selector ps;
  53
  54         unsigned pg_num;                /* Reference number */
  55         unsigned bypassed;              /* Temporarily bypass this PG? */
  56
  57         unsigned nr_pgpaths;            /* Number of paths in PG */
  58         struct list_head pgpaths;
  59 };
  60
  61 /* Multipath context */
  62 struct multipath {
  63         struct list_head list;
  64         struct dm_target *ti;
  65
  66         const char *hw_handler_name;
  67         char *hw_handler_params;
  68
  69         spinlock_t lock;
  70
  71         unsigned nr_priority_groups;
  72         struct list_head priority_groups;
  73
  74         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  75
  76         unsigned pg_init_required;      /* pg_init needs calling? */
  77         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  78         unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
  79
  80         unsigned nr_valid_paths;        /* Total number of usable paths */
  81         struct pgpath *current_pgpath;
  82         struct priority_group *current_pg;
  83         struct priority_group *next_pg; /* Switch to this PG if set */
  84         unsigned repeat_count;          /* I/Os left before calling PS again */
  85
  86         unsigned queue_io:1;            /* Must we queue all I/O? */
  87         unsigned queue_if_no_path:1;    /* Queue I/O if last path fails? */
  88         unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
  89         unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
  90         unsigned pg_init_disabled:1;    /* pg_init is not currently allowed */
  91
  92         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  93         unsigned pg_init_count;         /* Number of times pg_init called */
  94         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  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 static struct kmem_cache *_mpio_cache;
 118
 119 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 120 static void trigger_event(struct work_struct *work);
 121 static void activate_path(struct work_struct *work);
 122 static int __pgpath_busy(struct pgpath *pgpath);
 123
 124
 125 /*-----------------------------------------------
 126  * Allocation routines
 127  *-----------------------------------------------*/
 128
 129 static struct pgpath *alloc_pgpath(void)
 130 {
 131         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 132
 133         if (pgpath) {
 134                 pgpath->is_active = 1;
 135                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 136         }
 137
 138         return pgpath;
 139 }
 140
 141 static void free_pgpath(struct pgpath *pgpath)
 142 {
 143         kfree(pgpath);
 144 }
 145
 146 static struct priority_group *alloc_priority_group(void)
 147 {
 148         struct priority_group *pg;
 149
 150         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 151
 152         if (pg)
 153                 INIT_LIST_HEAD(&pg->pgpaths);
 154
 155         return pg;
 156 }
 157
 158 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 159 {
 160         struct pgpath *pgpath, *tmp;
 161         struct multipath *m = ti->private;
 162
 163         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 164                 list_del(&pgpath->list);
 165                 if (m->hw_handler_name)
 166                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 167                 dm_put_device(ti, pgpath->path.dev);
 168                 free_pgpath(pgpath);
 169         }
 170 }
 171
 172 static void free_priority_group(struct priority_group *pg,
 173                                 struct dm_target *ti)
 174 {
 175         struct path_selector *ps = &pg->ps;
 176
 177         if (ps->type) {
 178                 ps->type->destroy(ps);
 179                 dm_put_path_selector(ps->type);
 180         }
 181
 182         free_pgpaths(&pg->pgpaths, ti);
 183         kfree(pg);
 184 }
 185
 186 static struct multipath *alloc_multipath(struct dm_target *ti)
 187 {
 188         struct multipath *m;
 189         unsigned min_ios = dm_get_reserved_rq_based_ios();
 190
 191         m = kzalloc(sizeof(*m), GFP_KERNEL);
 192         if (m) {
 193                 INIT_LIST_HEAD(&m->priority_groups);
 194                 spin_lock_init(&m->lock);
 195                 m->queue_io = 1;
 196                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 197                 INIT_WORK(&m->trigger_event, trigger_event);
 198                 init_waitqueue_head(&m->pg_init_wait);
 199                 mutex_init(&m->work_mutex);
 200                 m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
 201                 if (!m->mpio_pool) {
 202                         kfree(m);
 203                         return NULL;
 204                 }
 205                 m->ti = ti;
 206                 ti->private = m;
 207         }
 208
 209         return m;
 210 }
 211
 212 static void free_multipath(struct multipath *m)
 213 {
 214         struct priority_group *pg, *tmp;
 215
 216         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 217                 list_del(&pg->list);
 218                 free_priority_group(pg, m->ti);
 219         }
 220
 221         kfree(m->hw_handler_name);
 222         kfree(m->hw_handler_params);
 223         mempool_destroy(m->mpio_pool);
 224         kfree(m);
 225 }
 226
 227 static int set_mapinfo(struct multipath *m, union map_info *info)
 228 {
 229         struct dm_mpath_io *mpio;
 230
 231         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 232         if (!mpio)
 233                 return -ENOMEM;
 234
 235         memset(mpio, 0, sizeof(*mpio));
 236         info->ptr = mpio;
 237
 238         return 0;
 239 }
 240
 241 static void clear_mapinfo(struct multipath *m, union map_info *info)
 242 {
 243         struct dm_mpath_io *mpio = info->ptr;
 244
 245         info->ptr = NULL;
 246         mempool_free(mpio, m->mpio_pool);
 247 }
 248
 249 /*-----------------------------------------------
 250  * Path selection
 251  *-----------------------------------------------*/
 252
 253 static int __pg_init_all_paths(struct multipath *m)
 254 {
 255         struct pgpath *pgpath;
 256         unsigned long pg_init_delay = 0;
 257
 258         if (m->pg_init_in_progress || m->pg_init_disabled)
 259                 return 0;
 260
 261         m->pg_init_count++;
 262         m->pg_init_required = 0;
 263
 264         /* Check here to reset pg_init_required */
 265         if (!m->current_pg)
 266                 return 0;
 267
 268         if (m->pg_init_delay_retry)
 269                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 270                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 271         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 272                 /* Skip failed paths */
 273                 if (!pgpath->is_active)
 274                         continue;
 275                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 276                                        pg_init_delay))
 277                         m->pg_init_in_progress++;
 278         }
 279         return m->pg_init_in_progress;
 280 }
 281
 282 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 283 {
 284         m->current_pg = pgpath->pg;
 285
 286         /* Must we initialise the PG first, and queue I/O till it's ready? */
 287         if (m->hw_handler_name) {
 288                 m->pg_init_required = 1;
 289                 m->queue_io = 1;
 290         } else {
 291                 m->pg_init_required = 0;
 292                 m->queue_io = 0;
 293         }
 294
 295         m->pg_init_count = 0;
 296 }
 297
 298 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 299                                size_t nr_bytes)
 300 {
 301         struct dm_path *path;
 302
 303         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 304         if (!path)
 305                 return -ENXIO;
 306
 307         m->current_pgpath = path_to_pgpath(path);
 308
 309         if (m->current_pg != pg)
 310                 __switch_pg(m, m->current_pgpath);
 311
 312         return 0;
 313 }
 314
 315 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 316 {
 317         struct priority_group *pg;
 318         unsigned bypassed = 1;
 319
 320         if (!m->nr_valid_paths)
 321                 goto failed;
 322
 323         /* Were we instructed to switch PG? */
 324         if (m->next_pg) {
 325                 pg = m->next_pg;
 326                 m->next_pg = NULL;
 327                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 328                         return;
 329         }
 330
 331         /* Don't change PG until it has no remaining paths */
 332         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 333                 return;
 334
 335         /*
 336          * Loop through priority groups until we find a valid path.
 337          * First time we skip PGs marked 'bypassed'.
 338          * Second time we only try the ones we skipped, but set
 339          * pg_init_delay_retry so we do not hammer controllers.
 340          */
 341         do {
 342                 list_for_each_entry(pg, &m->priority_groups, list) {
 343                         if (pg->bypassed == bypassed)
 344                                 continue;
 345                         if (!__choose_path_in_pg(m, pg, nr_bytes)) {
 346                                 if (!bypassed)
 347                                         m->pg_init_delay_retry = 1;
 348                                 return;
 349                         }
 350                 }
 351         } while (bypassed--);
 352
 353 failed:
 354         m->current_pgpath = NULL;
 355         m->current_pg = NULL;
 356 }
 357
 358 /*
 359  * Check whether bios must be queued in the device-mapper core rather
 360  * than here in the target.
 361  *
 362  * m->lock must be held on entry.
 363  *
 364  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 365  * same value then we are not between multipath_presuspend()
 366  * and multipath_resume() calls and we have no need to check
 367  * for the DMF_NOFLUSH_SUSPENDING flag.
 368  */
 369 static int __must_push_back(struct multipath *m)
 370 {
 371         return (m->queue_if_no_path ||
 372                 (m->queue_if_no_path != m->saved_queue_if_no_path &&
 373                  dm_noflush_suspending(m->ti)));
 374 }
 375
 376 #define pg_ready(m) (!(m)->queue_io && !(m)->pg_init_required)
 377
 378 /*
 379  * Map cloned requests
 380  */
 381 static int multipath_map(struct dm_target *ti, struct request *clone,
 382                          union map_info *map_context)
 383 {
 384         struct multipath *m = (struct multipath *) ti->private;
 385         int r = DM_MAPIO_REQUEUE;
 386         size_t nr_bytes = blk_rq_bytes(clone);
 387         unsigned long flags;
 388         struct pgpath *pgpath;
 389         struct block_device *bdev;
 390         struct dm_mpath_io *mpio;
 391
 392         spin_lock_irqsave(&m->lock, flags);
 393
 394         /* Do we need to select a new pgpath? */
 395         if (!m->current_pgpath ||
 396             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 397                 __choose_pgpath(m, nr_bytes);
 398
 399         pgpath = m->current_pgpath;
 400
 401         if (!pgpath) {
 402                 if (!__must_push_back(m))
 403                         r = -EIO;       /* Failed */
 404                 goto out_unlock;
 405         }
 406         if (!pg_ready(m)) {
 407                 __pg_init_all_paths(m);
 408                 goto out_unlock;
 409         }
 410         if (set_mapinfo(m, map_context) < 0)
 411                 /* ENOMEM, requeue */
 412                 goto out_unlock;
 413
 414         bdev = pgpath->path.dev->bdev;
 415         clone->q = bdev_get_queue(bdev);
 416         clone->rq_disk = bdev->bd_disk;
 417         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 418         mpio = map_context->ptr;
 419         mpio->pgpath = pgpath;
 420         mpio->nr_bytes = nr_bytes;
 421         if (pgpath->pg->ps.type->start_io)
 422                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 423                                               &pgpath->path,
 424                                               nr_bytes);
 425         r = DM_MAPIO_REMAPPED;
 426
 427 out_unlock:
 428         spin_unlock_irqrestore(&m->lock, flags);
 429
 430         return r;
 431 }
 432
 433 /*
 434  * If we run out of usable paths, should we queue I/O or error it?
 435  */
 436 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 437                             unsigned save_old_value)
 438 {
 439         unsigned long flags;
 440
 441         spin_lock_irqsave(&m->lock, flags);
 442
 443         if (save_old_value)
 444                 m->saved_queue_if_no_path = m->queue_if_no_path;
 445         else
 446                 m->saved_queue_if_no_path = queue_if_no_path;
 447         m->queue_if_no_path = queue_if_no_path;
 448         spin_unlock_irqrestore(&m->lock, flags);
 449
 450         if (!queue_if_no_path)
 451                 dm_table_run_md_queue_async(m->ti->table);
 452
 453         return 0;
 454 }
 455
 456 /*
 457  * An event is triggered whenever a path is taken out of use.
 458  * Includes path failure and PG bypass.
 459  */
 460 static void trigger_event(struct work_struct *work)
 461 {
 462         struct multipath *m =
 463                 container_of(work, struct multipath, trigger_event);
 464
 465         dm_table_event(m->ti->table);
 466 }
 467
 468 /*-----------------------------------------------------------------
 469  * Constructor/argument parsing:
 470  * <#multipath feature args> [<arg>]*
 471  * <#hw_handler args> [hw_handler [<arg>]*]
 472  * <#priority groups>
 473  * <initial priority group>
 474  *     [<selector> <#selector args> [<arg>]*
 475  *      <#paths> <#per-path selector args>
 476  *         [<path> [<arg>]* ]+ ]+
 477  *---------------------------------------------------------------*/
 478 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 479                                struct dm_target *ti)
 480 {
 481         int r;
 482         struct path_selector_type *pst;
 483         unsigned ps_argc;
 484
 485         static struct dm_arg _args[] = {
 486                 {0, 1024, "invalid number of path selector args"},
 487         };
 488
 489         pst = dm_get_path_selector(dm_shift_arg(as));
 490         if (!pst) {
 491                 ti->error = "unknown path selector type";
 492                 return -EINVAL;
 493         }
 494
 495         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 496         if (r) {
 497                 dm_put_path_selector(pst);
 498                 return -EINVAL;
 499         }
 500
 501         r = pst->create(&pg->ps, ps_argc, as->argv);
 502         if (r) {
 503                 dm_put_path_selector(pst);
 504                 ti->error = "path selector constructor failed";
 505                 return r;
 506         }
 507
 508         pg->ps.type = pst;
 509         dm_consume_args(as, ps_argc);
 510
 511         return 0;
 512 }
 513
 514 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 515                                struct dm_target *ti)
 516 {
 517         int r;
 518         struct pgpath *p;
 519         struct multipath *m = ti->private;
 520         struct request_queue *q = NULL;
 521         const char *attached_handler_name;
 522
 523         /* we need at least a path arg */
 524         if (as->argc < 1) {
 525                 ti->error = "no device given";
 526                 return ERR_PTR(-EINVAL);
 527         }
 528
 529         p = alloc_pgpath();
 530         if (!p)
 531                 return ERR_PTR(-ENOMEM);
 532
 533         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 534                           &p->path.dev);
 535         if (r) {
 536                 ti->error = "error getting device";
 537                 goto bad;
 538         }
 539
 540         if (m->retain_attached_hw_handler || m->hw_handler_name)
 541                 q = bdev_get_queue(p->path.dev->bdev);
 542
 543         if (m->retain_attached_hw_handler) {
 544                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 545                 if (attached_handler_name) {
 546                         /*
 547                          * Reset hw_handler_name to match the attached handler
 548                          * and clear any hw_handler_params associated with the
 549                          * ignored handler.
 550                          *
 551                          * NB. This modifies the table line to show the actual
 552                          * handler instead of the original table passed in.
 553                          */
 554                         kfree(m->hw_handler_name);
 555                         m->hw_handler_name = attached_handler_name;
 556
 557                         kfree(m->hw_handler_params);
 558                         m->hw_handler_params = NULL;
 559                 }
 560         }
 561
 562         if (m->hw_handler_name) {
 563                 /*
 564                  * Increments scsi_dh reference, even when using an
 565                  * already-attached handler.
 566                  */
 567                 r = scsi_dh_attach(q, m->hw_handler_name);
 568                 if (r == -EBUSY) {
 569                         /*
 570                          * Already attached to different hw_handler:
 571                          * try to reattach with correct one.
 572                          */
 573                         scsi_dh_detach(q);
 574                         r = scsi_dh_attach(q, m->hw_handler_name);
 575                 }
 576
 577                 if (r < 0) {
 578                         ti->error = "error attaching hardware handler";
 579                         dm_put_device(ti, p->path.dev);
 580                         goto bad;
 581                 }
 582
 583                 if (m->hw_handler_params) {
 584                         r = scsi_dh_set_params(q, m->hw_handler_params);
 585                         if (r < 0) {
 586                                 ti->error = "unable to set hardware "
 587                                                         "handler parameters";
 588                                 scsi_dh_detach(q);
 589                                 dm_put_device(ti, p->path.dev);
 590                                 goto bad;
 591                         }
 592                 }
 593         }
 594
 595         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 596         if (r) {
 597                 dm_put_device(ti, p->path.dev);
 598                 goto bad;
 599         }
 600
 601         return p;
 602
 603  bad:
 604         free_pgpath(p);
 605         return ERR_PTR(r);
 606 }
 607
 608 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 609                                                    struct multipath *m)
 610 {
 611         static struct dm_arg _args[] = {
 612                 {1, 1024, "invalid number of paths"},
 613                 {0, 1024, "invalid number of selector args"}
 614         };
 615
 616         int r;
 617         unsigned i, nr_selector_args, nr_args;
 618         struct priority_group *pg;
 619         struct dm_target *ti = m->ti;
 620
 621         if (as->argc < 2) {
 622                 as->argc = 0;
 623                 ti->error = "not enough priority group arguments";
 624                 return ERR_PTR(-EINVAL);
 625         }
 626
 627         pg = alloc_priority_group();
 628         if (!pg) {
 629                 ti->error = "couldn't allocate priority group";
 630                 return ERR_PTR(-ENOMEM);
 631         }
 632         pg->m = m;
 633
 634         r = parse_path_selector(as, pg, ti);
 635         if (r)
 636                 goto bad;
 637
 638         /*
 639          * read the paths
 640          */
 641         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 642         if (r)
 643                 goto bad;
 644
 645         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 646         if (r)
 647                 goto bad;
 648
 649         nr_args = 1 + nr_selector_args;
 650         for (i = 0; i < pg->nr_pgpaths; i++) {
 651                 struct pgpath *pgpath;
 652                 struct dm_arg_set path_args;
 653
 654                 if (as->argc < nr_args) {
 655                         ti->error = "not enough path parameters";
 656                         r = -EINVAL;
 657                         goto bad;
 658                 }
 659
 660                 path_args.argc = nr_args;
 661                 path_args.argv = as->argv;
 662
 663                 pgpath = parse_path(&path_args, &pg->ps, ti);
 664                 if (IS_ERR(pgpath)) {
 665                         r = PTR_ERR(pgpath);
 666                         goto bad;
 667                 }
 668
 669                 pgpath->pg = pg;
 670                 list_add_tail(&pgpath->list, &pg->pgpaths);
 671                 dm_consume_args(as, nr_args);
 672         }
 673
 674         return pg;
 675
 676  bad:
 677         free_priority_group(pg, ti);
 678         return ERR_PTR(r);
 679 }
 680
 681 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 682 {
 683         unsigned hw_argc;
 684         int ret;
 685         struct dm_target *ti = m->ti;
 686
 687         static struct dm_arg _args[] = {
 688                 {0, 1024, "invalid number of hardware handler args"},
 689         };
 690
 691         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 692                 return -EINVAL;
 693
 694         if (!hw_argc)
 695                 return 0;
 696
 697         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 698         if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
 699                                      "scsi_dh_%s", m->hw_handler_name)) {
 700                 ti->error = "unknown hardware handler type";
 701                 ret = -EINVAL;
 702                 goto fail;
 703         }
 704
 705         if (hw_argc > 1) {
 706                 char *p;
 707                 int i, j, len = 4;
 708
 709                 for (i = 0; i <= hw_argc - 2; i++)
 710                         len += strlen(as->argv[i]) + 1;
 711                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 712                 if (!p) {
 713                         ti->error = "memory allocation failed";
 714                         ret = -ENOMEM;
 715                         goto fail;
 716                 }
 717                 j = sprintf(p, "%d", hw_argc - 1);
 718                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 719                         j = sprintf(p, "%s", as->argv[i]);
 720         }
 721         dm_consume_args(as, hw_argc - 1);
 722
 723         return 0;
 724 fail:
 725         kfree(m->hw_handler_name);
 726         m->hw_handler_name = NULL;
 727         return ret;
 728 }
 729
 730 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 731 {
 732         int r;
 733         unsigned argc;
 734         struct dm_target *ti = m->ti;
 735         const char *arg_name;
 736
 737         static struct dm_arg _args[] = {
 738                 {0, 6, "invalid number of feature args"},
 739                 {1, 50, "pg_init_retries must be between 1 and 50"},
 740                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 741         };
 742
 743         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 744         if (r)
 745                 return -EINVAL;
 746
 747         if (!argc)
 748                 return 0;
 749
 750         do {
 751                 arg_name = dm_shift_arg(as);
 752                 argc--;
 753
 754                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 755                         r = queue_if_no_path(m, 1, 0);
 756                         continue;
 757                 }
 758
 759                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
 760                         m->retain_attached_hw_handler = 1;
 761                         continue;
 762                 }
 763
 764                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 765                     (argc >= 1)) {
 766                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 767                         argc--;
 768                         continue;
 769                 }
 770
 771                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 772                     (argc >= 1)) {
 773                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 774                         argc--;
 775                         continue;
 776                 }
 777
 778                 ti->error = "Unrecognised multipath feature request";
 779                 r = -EINVAL;
 780         } while (argc && !r);
 781
 782         return r;
 783 }
 784
 785 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 786                          char **argv)
 787 {
 788         /* target arguments */
 789         static struct dm_arg _args[] = {
 790                 {0, 1024, "invalid number of priority groups"},
 791                 {0, 1024, "invalid initial priority group number"},
 792         };
 793
 794         int r;
 795         struct multipath *m;
 796         struct dm_arg_set as;
 797         unsigned pg_count = 0;
 798         unsigned next_pg_num;
 799
 800         as.argc = argc;
 801         as.argv = argv;
 802
 803         m = alloc_multipath(ti);
 804         if (!m) {
 805                 ti->error = "can't allocate multipath";
 806                 return -EINVAL;
 807         }
 808
 809         r = parse_features(&as, m);
 810         if (r)
 811                 goto bad;
 812
 813         r = parse_hw_handler(&as, m);
 814         if (r)
 815                 goto bad;
 816
 817         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 818         if (r)
 819                 goto bad;
 820
 821         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 822         if (r)
 823                 goto bad;
 824
 825         if ((!m->nr_priority_groups && next_pg_num) ||
 826             (m->nr_priority_groups && !next_pg_num)) {
 827                 ti->error = "invalid initial priority group";
 828                 r = -EINVAL;
 829                 goto bad;
 830         }
 831
 832         /* parse the priority groups */
 833         while (as.argc) {
 834                 struct priority_group *pg;
 835
 836                 pg = parse_priority_group(&as, m);
 837                 if (IS_ERR(pg)) {
 838                         r = PTR_ERR(pg);
 839                         goto bad;
 840                 }
 841
 842                 m->nr_valid_paths += pg->nr_pgpaths;
 843                 list_add_tail(&pg->list, &m->priority_groups);
 844                 pg_count++;
 845                 pg->pg_num = pg_count;
 846                 if (!--next_pg_num)
 847                         m->next_pg = pg;
 848         }
 849
 850         if (pg_count != m->nr_priority_groups) {
 851                 ti->error = "priority group count mismatch";
 852                 r = -EINVAL;
 853                 goto bad;
 854         }
 855
 856         ti->num_flush_bios = 1;
 857         ti->num_discard_bios = 1;
 858         ti->num_write_same_bios = 1;
 859
 860         return 0;
 861
 862  bad:
 863         free_multipath(m);
 864         return r;
 865 }
 866
 867 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 868 {
 869         DECLARE_WAITQUEUE(wait, current);
 870         unsigned long flags;
 871
 872         add_wait_queue(&m->pg_init_wait, &wait);
 873
 874         while (1) {
 875                 set_current_state(TASK_UNINTERRUPTIBLE);
 876
 877                 spin_lock_irqsave(&m->lock, flags);
 878                 if (!m->pg_init_in_progress) {
 879                         spin_unlock_irqrestore(&m->lock, flags);
 880                         break;
 881                 }
 882                 spin_unlock_irqrestore(&m->lock, flags);
 883
 884                 io_schedule();
 885         }
 886         set_current_state(TASK_RUNNING);
 887
 888         remove_wait_queue(&m->pg_init_wait, &wait);
 889 }
 890
 891 static void flush_multipath_work(struct multipath *m)
 892 {
 893         unsigned long flags;
 894
 895         spin_lock_irqsave(&m->lock, flags);
 896         m->pg_init_disabled = 1;
 897         spin_unlock_irqrestore(&m->lock, flags);
 898
 899         flush_workqueue(kmpath_handlerd);
 900         multipath_wait_for_pg_init_completion(m);
 901         flush_workqueue(kmultipathd);
 902         flush_work(&m->trigger_event);
 903
 904         spin_lock_irqsave(&m->lock, flags);
 905         m->pg_init_disabled = 0;
 906         spin_unlock_irqrestore(&m->lock, flags);
 907 }
 908
 909 static void multipath_dtr(struct dm_target *ti)
 910 {
 911         struct multipath *m = ti->private;
 912
 913         flush_multipath_work(m);
 914         free_multipath(m);
 915 }
 916
 917 /*
 918  * Take a path out of use.
 919  */
 920 static int fail_path(struct pgpath *pgpath)
 921 {
 922         unsigned long flags;
 923         struct multipath *m = pgpath->pg->m;
 924
 925         spin_lock_irqsave(&m->lock, flags);
 926
 927         if (!pgpath->is_active)
 928                 goto out;
 929
 930         DMWARN("Failing path %s.", pgpath->path.dev->name);
 931
 932         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 933         pgpath->is_active = 0;
 934         pgpath->fail_count++;
 935
 936         m->nr_valid_paths--;
 937
 938         if (pgpath == m->current_pgpath)
 939                 m->current_pgpath = NULL;
 940
 941         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 942                       pgpath->path.dev->name, m->nr_valid_paths);
 943
 944         schedule_work(&m->trigger_event);
 945
 946 out:
 947         spin_unlock_irqrestore(&m->lock, flags);
 948
 949         return 0;
 950 }
 951
 952 /*
 953  * Reinstate a previously-failed path
 954  */
 955 static int reinstate_path(struct pgpath *pgpath)
 956 {
 957         int r = 0, run_queue = 0;
 958         unsigned long flags;
 959         struct multipath *m = pgpath->pg->m;
 960
 961         spin_lock_irqsave(&m->lock, flags);
 962
 963         if (pgpath->is_active)
 964                 goto out;
 965
 966         if (!pgpath->pg->ps.type->reinstate_path) {
 967                 DMWARN("Reinstate path not supported by path selector %s",
 968                        pgpath->pg->ps.type->name);
 969                 r = -EINVAL;
 970                 goto out;
 971         }
 972
 973         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 974         if (r)
 975                 goto out;
 976
 977         pgpath->is_active = 1;
 978
 979         if (!m->nr_valid_paths++) {
 980                 m->current_pgpath = NULL;
 981                 run_queue = 1;
 982         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
 983                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
 984                         m->pg_init_in_progress++;
 985         }
 986
 987         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 988                       pgpath->path.dev->name, m->nr_valid_paths);
 989
 990         schedule_work(&m->trigger_event);
 991
 992 out:
 993         spin_unlock_irqrestore(&m->lock, flags);
 994         if (run_queue)
 995                 dm_table_run_md_queue_async(m->ti->table);
 996
 997         return r;
 998 }
 999
1000 /*
1001  * Fail or reinstate all paths that match the provided struct dm_dev.
1002  */
1003 static int action_dev(struct multipath *m, struct dm_dev *dev,
1004                       action_fn action)
1005 {
1006         int r = -EINVAL;
1007         struct pgpath *pgpath;
1008         struct priority_group *pg;
1009
1010         list_for_each_entry(pg, &m->priority_groups, list) {
1011                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1012                         if (pgpath->path.dev == dev)
1013                                 r = action(pgpath);
1014                 }
1015         }
1016
1017         return r;
1018 }
1019
1020 /*
1021  * Temporarily try to avoid having to use the specified PG
1022  */
1023 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1024                       int bypassed)
1025 {
1026         unsigned long flags;
1027
1028         spin_lock_irqsave(&m->lock, flags);
1029
1030         pg->bypassed = bypassed;
1031         m->current_pgpath = NULL;
1032         m->current_pg = NULL;
1033
1034         spin_unlock_irqrestore(&m->lock, flags);
1035
1036         schedule_work(&m->trigger_event);
1037 }
1038
1039 /*
1040  * Switch to using the specified PG from the next I/O that gets mapped
1041  */
1042 static int switch_pg_num(struct multipath *m, const char *pgstr)
1043 {
1044         struct priority_group *pg;
1045         unsigned pgnum;
1046         unsigned long flags;
1047         char dummy;
1048
1049         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1050             (pgnum > m->nr_priority_groups)) {
1051                 DMWARN("invalid PG number supplied to switch_pg_num");
1052                 return -EINVAL;
1053         }
1054
1055         spin_lock_irqsave(&m->lock, flags);
1056         list_for_each_entry(pg, &m->priority_groups, list) {
1057                 pg->bypassed = 0;
1058                 if (--pgnum)
1059                         continue;
1060
1061                 m->current_pgpath = NULL;
1062                 m->current_pg = NULL;
1063                 m->next_pg = pg;
1064         }
1065         spin_unlock_irqrestore(&m->lock, flags);
1066
1067         schedule_work(&m->trigger_event);
1068         return 0;
1069 }
1070
1071 /*
1072  * Set/clear bypassed status of a PG.
1073  * PGs are numbered upwards from 1 in the order they were declared.
1074  */
1075 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1076 {
1077         struct priority_group *pg;
1078         unsigned pgnum;
1079         char dummy;
1080
1081         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1082             (pgnum > m->nr_priority_groups)) {
1083                 DMWARN("invalid PG number supplied to bypass_pg");
1084                 return -EINVAL;
1085         }
1086
1087         list_for_each_entry(pg, &m->priority_groups, list) {
1088                 if (!--pgnum)
1089                         break;
1090         }
1091
1092         bypass_pg(m, pg, bypassed);
1093         return 0;
1094 }
1095
1096 /*
1097  * Should we retry pg_init immediately?
1098  */
1099 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1100 {
1101         unsigned long flags;
1102         int limit_reached = 0;
1103
1104         spin_lock_irqsave(&m->lock, flags);
1105
1106         if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1107                 m->pg_init_required = 1;
1108         else
1109                 limit_reached = 1;
1110
1111         spin_unlock_irqrestore(&m->lock, flags);
1112
1113         return limit_reached;
1114 }
1115
1116 static void pg_init_done(void *data, int errors)
1117 {
1118         struct pgpath *pgpath = data;
1119         struct priority_group *pg = pgpath->pg;
1120         struct multipath *m = pg->m;
1121         unsigned long flags;
1122         unsigned delay_retry = 0;
1123
1124         /* device or driver problems */
1125         switch (errors) {
1126         case SCSI_DH_OK:
1127                 break;
1128         case SCSI_DH_NOSYS:
1129                 if (!m->hw_handler_name) {
1130                         errors = 0;
1131                         break;
1132                 }
1133                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1134                       "Error %d.", m->hw_handler_name, errors);
1135                 /*
1136                  * Fail path for now, so we do not ping pong
1137                  */
1138                 fail_path(pgpath);
1139                 break;
1140         case SCSI_DH_DEV_TEMP_BUSY:
1141                 /*
1142                  * Probably doing something like FW upgrade on the
1143                  * controller so try the other pg.
1144                  */
1145                 bypass_pg(m, pg, 1);
1146                 break;
1147         case SCSI_DH_RETRY:
1148                 /* Wait before retrying. */
1149                 delay_retry = 1;
1150         case SCSI_DH_IMM_RETRY:
1151         case SCSI_DH_RES_TEMP_UNAVAIL:
1152                 if (pg_init_limit_reached(m, pgpath))
1153                         fail_path(pgpath);
1154                 errors = 0;
1155                 break;
1156         default:
1157                 /*
1158                  * We probably do not want to fail the path for a device
1159                  * error, but this is what the old dm did. In future
1160                  * patches we can do more advanced handling.
1161                  */
1162                 fail_path(pgpath);
1163         }
1164
1165         spin_lock_irqsave(&m->lock, flags);
1166         if (errors) {
1167                 if (pgpath == m->current_pgpath) {
1168                         DMERR("Could not failover device. Error %d.", errors);
1169                         m->current_pgpath = NULL;
1170                         m->current_pg = NULL;
1171                 }
1172         } else if (!m->pg_init_required)
1173                 pg->bypassed = 0;
1174
1175         if (--m->pg_init_in_progress)
1176                 /* Activations of other paths are still on going */
1177                 goto out;
1178
1179         if (m->pg_init_required) {
1180                 m->pg_init_delay_retry = delay_retry;
1181                 if (__pg_init_all_paths(m))
1182                         goto out;
1183         }
1184         m->queue_io = 0;
1185
1186         /*
1187          * Wake up any thread waiting to suspend.
1188          */
1189         wake_up(&m->pg_init_wait);
1190
1191 out:
1192         spin_unlock_irqrestore(&m->lock, flags);
1193 }
1194
1195 static void activate_path(struct work_struct *work)
1196 {
1197         struct pgpath *pgpath =
1198                 container_of(work, struct pgpath, activate_path.work);
1199
1200         if (pgpath->is_active)
1201                 scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1202                                  pg_init_done, pgpath);
1203         else
1204                 pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1205 }
1206
1207 static int noretry_error(int error)
1208 {
1209         switch (error) {
1210         case -EOPNOTSUPP:
1211         case -EREMOTEIO:
1212         case -EILSEQ:
1213         case -ENODATA:
1214         case -ENOSPC:
1215                 return 1;
1216         }
1217
1218         /* Anything else could be a path failure, so should be retried */
1219         return 0;
1220 }
1221
1222 /*
1223  * end_io handling
1224  */
1225 static int do_end_io(struct multipath *m, struct request *clone,
1226                      int error, struct dm_mpath_io *mpio)
1227 {
1228         /*
1229          * We don't queue any clone request inside the multipath target
1230          * during end I/O handling, since those clone requests don't have
1231          * bio clones.  If we queue them inside the multipath target,
1232          * we need to make bio clones, that requires memory allocation.
1233          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1234          *  don't have bio clones.)
1235          * Instead of queueing the clone request here, we queue the original
1236          * request into dm core, which will remake a clone request and
1237          * clone bios for it and resubmit it later.
1238          */
1239         int r = DM_ENDIO_REQUEUE;
1240         unsigned long flags;
1241
1242         if (!error && !clone->errors)
1243                 return 0;       /* I/O complete */
1244
1245         if (noretry_error(error))
1246                 return error;
1247
1248         if (mpio->pgpath)
1249                 fail_path(mpio->pgpath);
1250
1251         spin_lock_irqsave(&m->lock, flags);
1252         if (!m->nr_valid_paths) {
1253                 if (!m->queue_if_no_path) {
1254                         if (!__must_push_back(m))
1255                                 r = -EIO;
1256                 } else {
1257                         if (error == -EBADE)
1258                                 r = error;
1259                 }
1260         }
1261         spin_unlock_irqrestore(&m->lock, flags);
1262
1263         return r;
1264 }
1265
1266 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1267                             int error, union map_info *map_context)
1268 {
1269         struct multipath *m = ti->private;
1270         struct dm_mpath_io *mpio = map_context->ptr;
1271         struct pgpath *pgpath;
1272         struct path_selector *ps;
1273         int r;
1274
1275         BUG_ON(!mpio);
1276
1277         r  = do_end_io(m, clone, error, mpio);
1278         pgpath = mpio->pgpath;
1279         if (pgpath) {
1280                 ps = &pgpath->pg->ps;
1281                 if (ps->type->end_io)
1282                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1283         }
1284         clear_mapinfo(m, map_context);
1285
1286         return r;
1287 }
1288
1289 /*
1290  * Suspend can't complete until all the I/O is processed so if
1291  * the last path fails we must error any remaining I/O.
1292  * Note that if the freeze_bdev fails while suspending, the
1293  * queue_if_no_path state is lost - userspace should reset it.
1294  */
1295 static void multipath_presuspend(struct dm_target *ti)
1296 {
1297         struct multipath *m = (struct multipath *) ti->private;
1298
1299         queue_if_no_path(m, 0, 1);
1300 }
1301
1302 static void multipath_postsuspend(struct dm_target *ti)
1303 {
1304         struct multipath *m = ti->private;
1305
1306         mutex_lock(&m->work_mutex);
1307         flush_multipath_work(m);
1308         mutex_unlock(&m->work_mutex);
1309 }
1310
1311 /*
1312  * Restore the queue_if_no_path setting.
1313  */
1314 static void multipath_resume(struct dm_target *ti)
1315 {
1316         struct multipath *m = (struct multipath *) ti->private;
1317         unsigned long flags;
1318
1319         spin_lock_irqsave(&m->lock, flags);
1320         m->queue_if_no_path = m->saved_queue_if_no_path;
1321         spin_unlock_irqrestore(&m->lock, flags);
1322 }
1323
1324 /*
1325  * Info output has the following format:
1326  * num_multipath_feature_args [multipath_feature_args]*
1327  * num_handler_status_args [handler_status_args]*
1328  * num_groups init_group_number
1329  *            [A|D|E num_ps_status_args [ps_status_args]*
1330  *             num_paths num_selector_args
1331  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1332  *
1333  * Table output has the following format (identical to the constructor string):
1334  * num_feature_args [features_args]*
1335  * num_handler_args hw_handler [hw_handler_args]*
1336  * num_groups init_group_number
1337  *     [priority selector-name num_ps_args [ps_args]*
1338  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1339  */
1340 static void multipath_status(struct dm_target *ti, status_type_t type,
1341                              unsigned status_flags, char *result, unsigned maxlen)
1342 {
1343         int sz = 0;
1344         unsigned long flags;
1345         struct multipath *m = (struct multipath *) ti->private;
1346         struct priority_group *pg;
1347         struct pgpath *p;
1348         unsigned pg_num;
1349         char state;
1350
1351         spin_lock_irqsave(&m->lock, flags);
1352
1353         /* Features */
1354         if (type == STATUSTYPE_INFO)
1355                 DMEMIT("2 %u %u ", m->queue_io, m->pg_init_count);
1356         else {
1357                 DMEMIT("%u ", m->queue_if_no_path +
1358                               (m->pg_init_retries > 0) * 2 +
1359                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1360                               m->retain_attached_hw_handler);
1361                 if (m->queue_if_no_path)
1362                         DMEMIT("queue_if_no_path ");
1363                 if (m->pg_init_retries)
1364                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1365                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1366                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1367                 if (m->retain_attached_hw_handler)
1368                         DMEMIT("retain_attached_hw_handler ");
1369         }
1370
1371         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1372                 DMEMIT("0 ");
1373         else
1374                 DMEMIT("1 %s ", m->hw_handler_name);
1375
1376         DMEMIT("%u ", m->nr_priority_groups);
1377
1378         if (m->next_pg)
1379                 pg_num = m->next_pg->pg_num;
1380         else if (m->current_pg)
1381                 pg_num = m->current_pg->pg_num;
1382         else
1383                 pg_num = (m->nr_priority_groups ? 1 : 0);
1384
1385         DMEMIT("%u ", pg_num);
1386
1387         switch (type) {
1388         case STATUSTYPE_INFO:
1389                 list_for_each_entry(pg, &m->priority_groups, list) {
1390                         if (pg->bypassed)
1391                                 state = 'D';    /* Disabled */
1392                         else if (pg == m->current_pg)
1393                                 state = 'A';    /* Currently Active */
1394                         else
1395                                 state = 'E';    /* Enabled */
1396
1397                         DMEMIT("%c ", state);
1398
1399                         if (pg->ps.type->status)
1400                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1401                                                           result + sz,
1402                                                           maxlen - sz);
1403                         else
1404                                 DMEMIT("0 ");
1405
1406                         DMEMIT("%u %u ", pg->nr_pgpaths,
1407                                pg->ps.type->info_args);
1408
1409                         list_for_each_entry(p, &pg->pgpaths, list) {
1410                                 DMEMIT("%s %s %u ", p->path.dev->name,
1411                                        p->is_active ? "A" : "F",
1412                                        p->fail_count);
1413                                 if (pg->ps.type->status)
1414                                         sz += pg->ps.type->status(&pg->ps,
1415                                               &p->path, type, result + sz,
1416                                               maxlen - sz);
1417                         }
1418                 }
1419                 break;
1420
1421         case STATUSTYPE_TABLE:
1422                 list_for_each_entry(pg, &m->priority_groups, list) {
1423                         DMEMIT("%s ", pg->ps.type->name);
1424
1425                         if (pg->ps.type->status)
1426                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1427                                                           result + sz,
1428                                                           maxlen - sz);
1429                         else
1430                                 DMEMIT("0 ");
1431
1432                         DMEMIT("%u %u ", pg->nr_pgpaths,
1433                                pg->ps.type->table_args);
1434
1435                         list_for_each_entry(p, &pg->pgpaths, list) {
1436                                 DMEMIT("%s ", p->path.dev->name);
1437                                 if (pg->ps.type->status)
1438                                         sz += pg->ps.type->status(&pg->ps,
1439                                               &p->path, type, result + sz,
1440                                               maxlen - sz);
1441                         }
1442                 }
1443                 break;
1444         }
1445
1446         spin_unlock_irqrestore(&m->lock, flags);
1447 }
1448
1449 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1450 {
1451         int r = -EINVAL;
1452         struct dm_dev *dev;
1453         struct multipath *m = (struct multipath *) ti->private;
1454         action_fn action;
1455
1456         mutex_lock(&m->work_mutex);
1457
1458         if (dm_suspended(ti)) {
1459                 r = -EBUSY;
1460                 goto out;
1461         }
1462
1463         if (argc == 1) {
1464                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1465                         r = queue_if_no_path(m, 1, 0);
1466                         goto out;
1467                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1468                         r = queue_if_no_path(m, 0, 0);
1469                         goto out;
1470                 }
1471         }
1472
1473         if (argc != 2) {
1474                 DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1475                 goto out;
1476         }
1477
1478         if (!strcasecmp(argv[0], "disable_group")) {
1479                 r = bypass_pg_num(m, argv[1], 1);
1480                 goto out;
1481         } else if (!strcasecmp(argv[0], "enable_group")) {
1482                 r = bypass_pg_num(m, argv[1], 0);
1483                 goto out;
1484         } else if (!strcasecmp(argv[0], "switch_group")) {
1485                 r = switch_pg_num(m, argv[1]);
1486                 goto out;
1487         } else if (!strcasecmp(argv[0], "reinstate_path"))
1488                 action = reinstate_path;
1489         else if (!strcasecmp(argv[0], "fail_path"))
1490                 action = fail_path;
1491         else {
1492                 DMWARN("Unrecognised multipath message received: %s", argv[0]);
1493                 goto out;
1494         }
1495
1496         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1497         if (r) {
1498                 DMWARN("message: error getting device %s",
1499                        argv[1]);
1500                 goto out;
1501         }
1502
1503         r = action_dev(m, dev, action);
1504
1505         dm_put_device(ti, dev);
1506
1507 out:
1508         mutex_unlock(&m->work_mutex);
1509         return r;
1510 }
1511
1512 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1513                            unsigned long arg)
1514 {
1515         struct multipath *m = ti->private;
1516         struct pgpath *pgpath;
1517         struct block_device *bdev;
1518         fmode_t mode;
1519         unsigned long flags;
1520         int r;
1521
1522         bdev = NULL;
1523         mode = 0;
1524         r = 0;
1525
1526         spin_lock_irqsave(&m->lock, flags);
1527
1528         if (!m->current_pgpath)
1529                 __choose_pgpath(m, 0);
1530
1531         pgpath = m->current_pgpath;
1532
1533         if (pgpath) {
1534                 bdev = pgpath->path.dev->bdev;
1535                 mode = pgpath->path.dev->mode;
1536         }
1537
1538         if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1539                 r = -ENOTCONN;
1540         else if (!bdev)
1541                 r = -EIO;
1542
1543         spin_unlock_irqrestore(&m->lock, flags);
1544
1545         /*
1546          * Only pass ioctls through if the device sizes match exactly.
1547          */
1548         if (!bdev || ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) {
1549                 int err = scsi_verify_blk_ioctl(NULL, cmd);
1550                 if (err)
1551                         r = err;
1552         }
1553
1554         if (r == -ENOTCONN && !fatal_signal_pending(current)) {
1555                 spin_lock_irqsave(&m->lock, flags);
1556                 if (!m->current_pg) {
1557                         /* Path status changed, redo selection */
1558                         __choose_pgpath(m, 0);
1559                 }
1560                 if (m->pg_init_required)
1561                         __pg_init_all_paths(m);
1562                 spin_unlock_irqrestore(&m->lock, flags);
1563                 dm_table_run_md_queue_async(m->ti->table);
1564         }
1565
1566         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1567 }
1568
1569 static int multipath_iterate_devices(struct dm_target *ti,
1570                                      iterate_devices_callout_fn fn, void *data)
1571 {
1572         struct multipath *m = ti->private;
1573         struct priority_group *pg;
1574         struct pgpath *p;
1575         int ret = 0;
1576
1577         list_for_each_entry(pg, &m->priority_groups, list) {
1578                 list_for_each_entry(p, &pg->pgpaths, list) {
1579                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1580                         if (ret)
1581                                 goto out;
1582                 }
1583         }
1584
1585 out:
1586         return ret;
1587 }
1588
1589 static int __pgpath_busy(struct pgpath *pgpath)
1590 {
1591         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1592
1593         return dm_underlying_device_busy(q);
1594 }
1595
1596 /*
1597  * We return "busy", only when we can map I/Os but underlying devices
1598  * are busy (so even if we map I/Os now, the I/Os will wait on
1599  * the underlying queue).
1600  * In other words, if we want to kill I/Os or queue them inside us
1601  * due to map unavailability, we don't return "busy".  Otherwise,
1602  * dm core won't give us the I/Os and we can't do what we want.
1603  */
1604 static int multipath_busy(struct dm_target *ti)
1605 {
1606         int busy = 0, has_active = 0;
1607         struct multipath *m = ti->private;
1608         struct priority_group *pg;
1609         struct pgpath *pgpath;
1610         unsigned long flags;
1611
1612         spin_lock_irqsave(&m->lock, flags);
1613
1614         /* pg_init in progress or no paths available */
1615         if (m->pg_init_in_progress ||
1616             (!m->nr_valid_paths && m->queue_if_no_path)) {
1617                 busy = 1;
1618                 goto out;
1619         }
1620         /* Guess which priority_group will be used at next mapping time */
1621         if (unlikely(!m->current_pgpath && m->next_pg))
1622                 pg = m->next_pg;
1623         else if (likely(m->current_pg))
1624                 pg = m->current_pg;
1625         else
1626                 /*
1627                  * We don't know which pg will be used at next mapping time.
1628                  * We don't call __choose_pgpath() here to avoid to trigger
1629                  * pg_init just by busy checking.
1630                  * So we don't know whether underlying devices we will be using
1631                  * at next mapping time are busy or not. Just try mapping.
1632                  */
1633                 goto out;
1634
1635         /*
1636          * If there is one non-busy active path at least, the path selector
1637          * will be able to select it. So we consider such a pg as not busy.
1638          */
1639         busy = 1;
1640         list_for_each_entry(pgpath, &pg->pgpaths, list)
1641                 if (pgpath->is_active) {
1642                         has_active = 1;
1643
1644                         if (!__pgpath_busy(pgpath)) {
1645                                 busy = 0;
1646                                 break;
1647                         }
1648                 }
1649
1650         if (!has_active)
1651                 /*
1652                  * No active path in this pg, so this pg won't be used and
1653                  * the current_pg will be changed at next mapping time.
1654                  * We need to try mapping to determine it.
1655                  */
1656                 busy = 0;
1657
1658 out:
1659         spin_unlock_irqrestore(&m->lock, flags);
1660
1661         return busy;
1662 }
1663
1664 /*-----------------------------------------------------------------
1665  * Module setup
1666  *---------------------------------------------------------------*/
1667 static struct target_type multipath_target = {
1668         .name = "multipath",
1669         .version = {1, 7, 0},
1670         .module = THIS_MODULE,
1671         .ctr = multipath_ctr,
1672         .dtr = multipath_dtr,
1673         .map_rq = multipath_map,
1674         .rq_end_io = multipath_end_io,
1675         .presuspend = multipath_presuspend,
1676         .postsuspend = multipath_postsuspend,
1677         .resume = multipath_resume,
1678         .status = multipath_status,
1679         .message = multipath_message,
1680         .ioctl  = multipath_ioctl,
1681         .iterate_devices = multipath_iterate_devices,
1682         .busy = multipath_busy,
1683 };
1684
1685 static int __init dm_multipath_init(void)
1686 {
1687         int r;
1688
1689         /* allocate a slab for the dm_ios */
1690         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1691         if (!_mpio_cache)
1692                 return -ENOMEM;
1693
1694         r = dm_register_target(&multipath_target);
1695         if (r < 0) {
1696                 DMERR("register failed %d", r);
1697                 kmem_cache_destroy(_mpio_cache);
1698                 return -EINVAL;
1699         }
1700
1701         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1702         if (!kmultipathd) {
1703                 DMERR("failed to create workqueue kmpathd");
1704                 dm_unregister_target(&multipath_target);
1705                 kmem_cache_destroy(_mpio_cache);
1706                 return -ENOMEM;
1707         }
1708
1709         /*
1710          * A separate workqueue is used to handle the device handlers
1711          * to avoid overloading existing workqueue. Overloading the
1712          * old workqueue would also create a bottleneck in the
1713          * path of the storage hardware device activation.
1714          */
1715         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1716                                                   WQ_MEM_RECLAIM);
1717         if (!kmpath_handlerd) {
1718                 DMERR("failed to create workqueue kmpath_handlerd");
1719                 destroy_workqueue(kmultipathd);
1720                 dm_unregister_target(&multipath_target);
1721                 kmem_cache_destroy(_mpio_cache);
1722                 return -ENOMEM;
1723         }
1724
1725         DMINFO("version %u.%u.%u loaded",
1726                multipath_target.version[0], multipath_target.version[1],
1727                multipath_target.version[2]);
1728
1729         return r;
1730 }
1731
1732 static void __exit dm_multipath_exit(void)
1733 {
1734         destroy_workqueue(kmpath_handlerd);
1735         destroy_workqueue(kmultipathd);
1736
1737         dm_unregister_target(&multipath_target);
1738         kmem_cache_destroy(_mpio_cache);
1739 }
1740
1741 module_init(dm_multipath_init);
1742 module_exit(dm_multipath_exit);
1743
1744 MODULE_DESCRIPTION(DM_NAME " multipath target");
1745 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1746 MODULE_LICENSE("GPL");