drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9
  10 #include "dm-path-selector.h"
  11 #include "dm-uevent.h"
  12
  13 #include <linux/ctype.h>
  14 #include <linux/init.h>
  15 #include <linux/mempool.h>
  16 #include <linux/module.h>
  17 #include <linux/pagemap.h>
  18 #include <linux/slab.h>
  19 #include <linux/time.h>
  20 #include <linux/workqueue.h>
  21 #include <scsi/scsi_dh.h>
  22 #include <linux/atomic.h>
  23
  24 #define DM_MSG_PREFIX "multipath"
  25 #define DM_PG_INIT_DELAY_MSECS 2000
  26 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  27
  28 /* Path properties */
  29 struct pgpath {
  30         struct list_head list;
  31
  32         struct priority_group *pg;      /* Owning PG */
  33         unsigned is_active;             /* Path status */
  34         unsigned fail_count;            /* Cumulative failure count */
  35
  36         struct dm_path path;
  37         struct delayed_work activate_path;
  38 };
  39
  40 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  41
  42 /*
  43  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  44  * Each has a path selector which controls which path gets used.
  45  */
  46 struct priority_group {
  47         struct list_head list;
  48
  49         struct multipath *m;            /* Owning multipath instance */
  50         struct path_selector ps;
  51
  52         unsigned pg_num;                /* Reference number */
  53         unsigned bypassed;              /* Temporarily bypass this PG? */
  54
  55         unsigned nr_pgpaths;            /* Number of paths in PG */
  56         struct list_head pgpaths;
  57 };
  58
  59 /* Multipath context */
  60 struct multipath {
  61         struct list_head list;
  62         struct dm_target *ti;
  63
  64         spinlock_t lock;
  65
  66         const char *hw_handler_name;
  67         char *hw_handler_params;
  68
  69         unsigned nr_priority_groups;
  70         struct list_head priority_groups;
  71
  72         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  73
  74         unsigned pg_init_required;      /* pg_init needs calling? */
  75         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  76         unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
  77
  78         unsigned nr_valid_paths;        /* Total number of usable paths */
  79         struct pgpath *current_pgpath;
  80         struct priority_group *current_pg;
  81         struct priority_group *next_pg; /* Switch to this PG if set */
  82         unsigned repeat_count;          /* I/Os left before calling PS again */
  83
  84         unsigned queue_io;              /* Must we queue all I/O? */
  85         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  86         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  87         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  88         unsigned pg_init_count;         /* Number of times pg_init called */
  89         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  90
  91         struct work_struct process_queued_ios;
  92         struct list_head queued_ios;
  93         unsigned queue_size;
  94
  95         struct work_struct trigger_event;
  96
  97         /*
  98          * We must use a mempool of dm_mpath_io structs so that we
  99          * can resubmit bios on error.
 100          */
 101         mempool_t *mpio_pool;
 102
 103         struct mutex work_mutex;
 104 };
 105
 106 /*
 107  * Context information attached to each bio we process.
 108  */
 109 struct dm_mpath_io {
 110         struct pgpath *pgpath;
 111         size_t nr_bytes;
 112 };
 113
 114 typedef int (*action_fn) (struct pgpath *pgpath);
 115
 116 #define MIN_IOS 256     /* Mempool size */
 117
 118 static struct kmem_cache *_mpio_cache;
 119
 120 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 121 static void process_queued_ios(struct work_struct *work);
 122 static void trigger_event(struct work_struct *work);
 123 static void activate_path(struct work_struct *work);
 124
 125
 126 /*-----------------------------------------------
 127  * Allocation routines
 128  *-----------------------------------------------*/
 129
 130 static struct pgpath *alloc_pgpath(void)
 131 {
 132         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 133
 134         if (pgpath) {
 135                 pgpath->is_active = 1;
 136                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 137         }
 138
 139         return pgpath;
 140 }
 141
 142 static void free_pgpath(struct pgpath *pgpath)
 143 {
 144         kfree(pgpath);
 145 }
 146
 147 static struct priority_group *alloc_priority_group(void)
 148 {
 149         struct priority_group *pg;
 150
 151         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 152
 153         if (pg)
 154                 INIT_LIST_HEAD(&pg->pgpaths);
 155
 156         return pg;
 157 }
 158
 159 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 160 {
 161         struct pgpath *pgpath, *tmp;
 162         struct multipath *m = ti->private;
 163
 164         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 165                 list_del(&pgpath->list);
 166                 if (m->hw_handler_name)
 167                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 168                 dm_put_device(ti, pgpath->path.dev);
 169                 free_pgpath(pgpath);
 170         }
 171 }
 172
 173 static void free_priority_group(struct priority_group *pg,
 174                                 struct dm_target *ti)
 175 {
 176         struct path_selector *ps = &pg->ps;
 177
 178         if (ps->type) {
 179                 ps->type->destroy(ps);
 180                 dm_put_path_selector(ps->type);
 181         }
 182
 183         free_pgpaths(&pg->pgpaths, ti);
 184         kfree(pg);
 185 }
 186
 187 static struct multipath *alloc_multipath(struct dm_target *ti)
 188 {
 189         struct multipath *m;
 190
 191         m = kzalloc(sizeof(*m), GFP_KERNEL);
 192         if (m) {
 193                 INIT_LIST_HEAD(&m->priority_groups);
 194                 INIT_LIST_HEAD(&m->queued_ios);
 195                 spin_lock_init(&m->lock);
 196                 m->queue_io = 1;
 197                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 198                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 199                 INIT_WORK(&m->trigger_event, trigger_event);
 200                 init_waitqueue_head(&m->pg_init_wait);
 201                 mutex_init(&m->work_mutex);
 202                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 203                 if (!m->mpio_pool) {
 204                         kfree(m);
 205                         return NULL;
 206                 }
 207                 m->ti = ti;
 208                 ti->private = m;
 209         }
 210
 211         return m;
 212 }
 213
 214 static void free_multipath(struct multipath *m)
 215 {
 216         struct priority_group *pg, *tmp;
 217
 218         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 219                 list_del(&pg->list);
 220                 free_priority_group(pg, m->ti);
 221         }
 222
 223         kfree(m->hw_handler_name);
 224         kfree(m->hw_handler_params);
 225         mempool_destroy(m->mpio_pool);
 226         kfree(m);
 227 }
 228
 229
 230 /*-----------------------------------------------
 231  * Path selection
 232  *-----------------------------------------------*/
 233
 234 static void __pg_init_all_paths(struct multipath *m)
 235 {
 236         struct pgpath *pgpath;
 237         unsigned long pg_init_delay = 0;
 238
 239         m->pg_init_count++;
 240         m->pg_init_required = 0;
 241         if (m->pg_init_delay_retry)
 242                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 243                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 244         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 245                 /* Skip failed paths */
 246                 if (!pgpath->is_active)
 247                         continue;
 248                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 249                                        pg_init_delay))
 250                         m->pg_init_in_progress++;
 251         }
 252 }
 253
 254 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 255 {
 256         m->current_pg = pgpath->pg;
 257
 258         /* Must we initialise the PG first, and queue I/O till it's ready? */
 259         if (m->hw_handler_name) {
 260                 m->pg_init_required = 1;
 261                 m->queue_io = 1;
 262         } else {
 263                 m->pg_init_required = 0;
 264                 m->queue_io = 0;
 265         }
 266
 267         m->pg_init_count = 0;
 268 }
 269
 270 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 271                                size_t nr_bytes)
 272 {
 273         struct dm_path *path;
 274
 275         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 276         if (!path)
 277                 return -ENXIO;
 278
 279         m->current_pgpath = path_to_pgpath(path);
 280
 281         if (m->current_pg != pg)
 282                 __switch_pg(m, m->current_pgpath);
 283
 284         return 0;
 285 }
 286
 287 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 288 {
 289         struct priority_group *pg;
 290         unsigned bypassed = 1;
 291
 292         if (!m->nr_valid_paths)
 293                 goto failed;
 294
 295         /* Were we instructed to switch PG? */
 296         if (m->next_pg) {
 297                 pg = m->next_pg;
 298                 m->next_pg = NULL;
 299                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 300                         return;
 301         }
 302
 303         /* Don't change PG until it has no remaining paths */
 304         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 305                 return;
 306
 307         /*
 308          * Loop through priority groups until we find a valid path.
 309          * First time we skip PGs marked 'bypassed'.
 310          * Second time we only try the ones we skipped.
 311          */
 312         do {
 313                 list_for_each_entry(pg, &m->priority_groups, list) {
 314                         if (pg->bypassed == bypassed)
 315                                 continue;
 316                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 317                                 return;
 318                 }
 319         } while (bypassed--);
 320
 321 failed:
 322         m->current_pgpath = NULL;
 323         m->current_pg = NULL;
 324 }
 325
 326 /*
 327  * Check whether bios must be queued in the device-mapper core rather
 328  * than here in the target.
 329  *
 330  * m->lock must be held on entry.
 331  *
 332  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 333  * same value then we are not between multipath_presuspend()
 334  * and multipath_resume() calls and we have no need to check
 335  * for the DMF_NOFLUSH_SUSPENDING flag.
 336  */
 337 static int __must_push_back(struct multipath *m)
 338 {
 339         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 340                 dm_noflush_suspending(m->ti));
 341 }
 342
 343 static int map_io(struct multipath *m, struct request *clone,
 344                   struct dm_mpath_io *mpio, unsigned was_queued)
 345 {
 346         int r = DM_MAPIO_REMAPPED;
 347         size_t nr_bytes = blk_rq_bytes(clone);
 348         unsigned long flags;
 349         struct pgpath *pgpath;
 350         struct block_device *bdev;
 351
 352         spin_lock_irqsave(&m->lock, flags);
 353
 354         /* Do we need to select a new pgpath? */
 355         if (!m->current_pgpath ||
 356             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 357                 __choose_pgpath(m, nr_bytes);
 358
 359         pgpath = m->current_pgpath;
 360
 361         if (was_queued)
 362                 m->queue_size--;
 363
 364         if ((pgpath && m->queue_io) ||
 365             (!pgpath && m->queue_if_no_path)) {
 366                 /* Queue for the daemon to resubmit */
 367                 list_add_tail(&clone->queuelist, &m->queued_ios);
 368                 m->queue_size++;
 369                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 370                     !m->queue_io)
 371                         queue_work(kmultipathd, &m->process_queued_ios);
 372                 pgpath = NULL;
 373                 r = DM_MAPIO_SUBMITTED;
 374         } else if (pgpath) {
 375                 bdev = pgpath->path.dev->bdev;
 376                 clone->q = bdev_get_queue(bdev);
 377                 clone->rq_disk = bdev->bd_disk;
 378         } else if (__must_push_back(m))
 379                 r = DM_MAPIO_REQUEUE;
 380         else
 381                 r = -EIO;       /* Failed */
 382
 383         mpio->pgpath = pgpath;
 384         mpio->nr_bytes = nr_bytes;
 385
 386         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 387                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 388                                               nr_bytes);
 389
 390         spin_unlock_irqrestore(&m->lock, flags);
 391
 392         return r;
 393 }
 394
 395 /*
 396  * If we run out of usable paths, should we queue I/O or error it?
 397  */
 398 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 399                             unsigned save_old_value)
 400 {
 401         unsigned long flags;
 402
 403         spin_lock_irqsave(&m->lock, flags);
 404
 405         if (save_old_value)
 406                 m->saved_queue_if_no_path = m->queue_if_no_path;
 407         else
 408                 m->saved_queue_if_no_path = queue_if_no_path;
 409         m->queue_if_no_path = queue_if_no_path;
 410         if (!m->queue_if_no_path && m->queue_size)
 411                 queue_work(kmultipathd, &m->process_queued_ios);
 412
 413         spin_unlock_irqrestore(&m->lock, flags);
 414
 415         return 0;
 416 }
 417
 418 /*-----------------------------------------------------------------
 419  * The multipath daemon is responsible for resubmitting queued ios.
 420  *---------------------------------------------------------------*/
 421
 422 static void dispatch_queued_ios(struct multipath *m)
 423 {
 424         int r;
 425         unsigned long flags;
 426         struct dm_mpath_io *mpio;
 427         union map_info *info;
 428         struct request *clone, *n;
 429         LIST_HEAD(cl);
 430
 431         spin_lock_irqsave(&m->lock, flags);
 432         list_splice_init(&m->queued_ios, &cl);
 433         spin_unlock_irqrestore(&m->lock, flags);
 434
 435         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 436                 list_del_init(&clone->queuelist);
 437
 438                 info = dm_get_rq_mapinfo(clone);
 439                 mpio = info->ptr;
 440
 441                 r = map_io(m, clone, mpio, 1);
 442                 if (r < 0) {
 443                         mempool_free(mpio, m->mpio_pool);
 444                         dm_kill_unmapped_request(clone, r);
 445                 } else if (r == DM_MAPIO_REMAPPED)
 446                         dm_dispatch_request(clone);
 447                 else if (r == DM_MAPIO_REQUEUE) {
 448                         mempool_free(mpio, m->mpio_pool);
 449                         dm_requeue_unmapped_request(clone);
 450                 }
 451         }
 452 }
 453
 454 static void process_queued_ios(struct work_struct *work)
 455 {
 456         struct multipath *m =
 457                 container_of(work, struct multipath, process_queued_ios);
 458         struct pgpath *pgpath = NULL;
 459         unsigned must_queue = 1;
 460         unsigned long flags;
 461
 462         spin_lock_irqsave(&m->lock, flags);
 463
 464         if (!m->queue_size)
 465                 goto out;
 466
 467         if (!m->current_pgpath)
 468                 __choose_pgpath(m, 0);
 469
 470         pgpath = m->current_pgpath;
 471
 472         if ((pgpath && !m->queue_io) ||
 473             (!pgpath && !m->queue_if_no_path))
 474                 must_queue = 0;
 475
 476         if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
 477                 __pg_init_all_paths(m);
 478
 479 out:
 480         spin_unlock_irqrestore(&m->lock, flags);
 481         if (!must_queue)
 482                 dispatch_queued_ios(m);
 483 }
 484
 485 /*
 486  * An event is triggered whenever a path is taken out of use.
 487  * Includes path failure and PG bypass.
 488  */
 489 static void trigger_event(struct work_struct *work)
 490 {
 491         struct multipath *m =
 492                 container_of(work, struct multipath, trigger_event);
 493
 494         dm_table_event(m->ti->table);
 495 }
 496
 497 /*-----------------------------------------------------------------
 498  * Constructor/argument parsing:
 499  * <#multipath feature args> [<arg>]*
 500  * <#hw_handler args> [hw_handler [<arg>]*]
 501  * <#priority groups>
 502  * <initial priority group>
 503  *     [<selector> <#selector args> [<arg>]*
 504  *      <#paths> <#per-path selector args>
 505  *         [<path> [<arg>]* ]+ ]+
 506  *---------------------------------------------------------------*/
 507 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 508                                struct dm_target *ti)
 509 {
 510         int r;
 511         struct path_selector_type *pst;
 512         unsigned ps_argc;
 513
 514         static struct dm_arg _args[] = {
 515                 {0, 1024, "invalid number of path selector args"},
 516         };
 517
 518         pst = dm_get_path_selector(dm_shift_arg(as));
 519         if (!pst) {
 520                 ti->error = "unknown path selector type";
 521                 return -EINVAL;
 522         }
 523
 524         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 525         if (r) {
 526                 dm_put_path_selector(pst);
 527                 return -EINVAL;
 528         }
 529
 530         r = pst->create(&pg->ps, ps_argc, as->argv);
 531         if (r) {
 532                 dm_put_path_selector(pst);
 533                 ti->error = "path selector constructor failed";
 534                 return r;
 535         }
 536
 537         pg->ps.type = pst;
 538         dm_consume_args(as, ps_argc);
 539
 540         return 0;
 541 }
 542
 543 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 544                                struct dm_target *ti)
 545 {
 546         int r;
 547         struct pgpath *p;
 548         struct multipath *m = ti->private;
 549
 550         /* we need at least a path arg */
 551         if (as->argc < 1) {
 552                 ti->error = "no device given";
 553                 return ERR_PTR(-EINVAL);
 554         }
 555
 556         p = alloc_pgpath();
 557         if (!p)
 558                 return ERR_PTR(-ENOMEM);
 559
 560         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 561                           &p->path.dev);
 562         if (r) {
 563                 ti->error = "error getting device";
 564                 goto bad;
 565         }
 566
 567         if (m->hw_handler_name) {
 568                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 569
 570                 r = scsi_dh_attach(q, m->hw_handler_name);
 571                 if (r == -EBUSY) {
 572                         /*
 573                          * Already attached to different hw_handler,
 574                          * try to reattach with correct one.
 575                          */
 576                         scsi_dh_detach(q);
 577                         r = scsi_dh_attach(q, m->hw_handler_name);
 578                 }
 579
 580                 if (r < 0) {
 581                         ti->error = "error attaching hardware handler";
 582                         dm_put_device(ti, p->path.dev);
 583                         goto bad;
 584                 }
 585
 586                 if (m->hw_handler_params) {
 587                         r = scsi_dh_set_params(q, m->hw_handler_params);
 588                         if (r < 0) {
 589                                 ti->error = "unable to set hardware "
 590                                                         "handler parameters";
 591                                 scsi_dh_detach(q);
 592                                 dm_put_device(ti, p->path.dev);
 593                                 goto bad;
 594                         }
 595                 }
 596         }
 597
 598         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 599         if (r) {
 600                 dm_put_device(ti, p->path.dev);
 601                 goto bad;
 602         }
 603
 604         return p;
 605
 606  bad:
 607         free_pgpath(p);
 608         return ERR_PTR(r);
 609 }
 610
 611 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 612                                                    struct multipath *m)
 613 {
 614         static struct dm_arg _args[] = {
 615                 {1, 1024, "invalid number of paths"},
 616                 {0, 1024, "invalid number of selector args"}
 617         };
 618
 619         int r;
 620         unsigned i, nr_selector_args, nr_args;
 621         struct priority_group *pg;
 622         struct dm_target *ti = m->ti;
 623
 624         if (as->argc < 2) {
 625                 as->argc = 0;
 626                 ti->error = "not enough priority group arguments";
 627                 return ERR_PTR(-EINVAL);
 628         }
 629
 630         pg = alloc_priority_group();
 631         if (!pg) {
 632                 ti->error = "couldn't allocate priority group";
 633                 return ERR_PTR(-ENOMEM);
 634         }
 635         pg->m = m;
 636
 637         r = parse_path_selector(as, pg, ti);
 638         if (r)
 639                 goto bad;
 640
 641         /*
 642          * read the paths
 643          */
 644         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 645         if (r)
 646                 goto bad;
 647
 648         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 649         if (r)
 650                 goto bad;
 651
 652         nr_args = 1 + nr_selector_args;
 653         for (i = 0; i < pg->nr_pgpaths; i++) {
 654                 struct pgpath *pgpath;
 655                 struct dm_arg_set path_args;
 656
 657                 if (as->argc < nr_args) {
 658                         ti->error = "not enough path parameters";
 659                         r = -EINVAL;
 660                         goto bad;
 661                 }
 662
 663                 path_args.argc = nr_args;
 664                 path_args.argv = as->argv;
 665
 666                 pgpath = parse_path(&path_args, &pg->ps, ti);
 667                 if (IS_ERR(pgpath)) {
 668                         r = PTR_ERR(pgpath);
 669                         goto bad;
 670                 }
 671
 672                 pgpath->pg = pg;
 673                 list_add_tail(&pgpath->list, &pg->pgpaths);
 674                 dm_consume_args(as, nr_args);
 675         }
 676
 677         return pg;
 678
 679  bad:
 680         free_priority_group(pg, ti);
 681         return ERR_PTR(r);
 682 }
 683
 684 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 685 {
 686         unsigned hw_argc;
 687         int ret;
 688         struct dm_target *ti = m->ti;
 689
 690         static struct dm_arg _args[] = {
 691                 {0, 1024, "invalid number of hardware handler args"},
 692         };
 693
 694         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 695                 return -EINVAL;
 696
 697         if (!hw_argc)
 698                 return 0;
 699
 700         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 701         request_module("scsi_dh_%s", m->hw_handler_name);
 702         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 703                 ti->error = "unknown hardware handler type";
 704                 ret = -EINVAL;
 705                 goto fail;
 706         }
 707
 708         if (hw_argc > 1) {
 709                 char *p;
 710                 int i, j, len = 4;
 711
 712                 for (i = 0; i <= hw_argc - 2; i++)
 713                         len += strlen(as->argv[i]) + 1;
 714                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 715                 if (!p) {
 716                         ti->error = "memory allocation failed";
 717                         ret = -ENOMEM;
 718                         goto fail;
 719                 }
 720                 j = sprintf(p, "%d", hw_argc - 1);
 721                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 722                         j = sprintf(p, "%s", as->argv[i]);
 723         }
 724         dm_consume_args(as, hw_argc - 1);
 725
 726         return 0;
 727 fail:
 728         kfree(m->hw_handler_name);
 729         m->hw_handler_name = NULL;
 730         return ret;
 731 }
 732
 733 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 734 {
 735         int r;
 736         unsigned argc;
 737         struct dm_target *ti = m->ti;
 738         const char *arg_name;
 739
 740         static struct dm_arg _args[] = {
 741                 {0, 5, "invalid number of feature args"},
 742                 {1, 50, "pg_init_retries must be between 1 and 50"},
 743                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 744         };
 745
 746         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 747         if (r)
 748                 return -EINVAL;
 749
 750         if (!argc)
 751                 return 0;
 752
 753         do {
 754                 arg_name = dm_shift_arg(as);
 755                 argc--;
 756
 757                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 758                         r = queue_if_no_path(m, 1, 0);
 759                         continue;
 760                 }
 761
 762                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 763                     (argc >= 1)) {
 764                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 765                         argc--;
 766                         continue;
 767                 }
 768
 769                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 770                     (argc >= 1)) {
 771                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 772                         argc--;
 773                         continue;
 774                 }
 775
 776                 ti->error = "Unrecognised multipath feature request";
 777                 r = -EINVAL;
 778         } while (argc && !r);
 779
 780         return r;
 781 }
 782
 783 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 784                          char **argv)
 785 {
 786         /* target arguments */
 787         static struct dm_arg _args[] = {
 788                 {0, 1024, "invalid number of priority groups"},
 789                 {0, 1024, "invalid initial priority group number"},
 790         };
 791
 792         int r;
 793         struct multipath *m;
 794         struct dm_arg_set as;
 795         unsigned pg_count = 0;
 796         unsigned next_pg_num;
 797
 798         as.argc = argc;
 799         as.argv = argv;
 800
 801         m = alloc_multipath(ti);
 802         if (!m) {
 803                 ti->error = "can't allocate multipath";
 804                 return -EINVAL;
 805         }
 806
 807         r = parse_features(&as, m);
 808         if (r)
 809                 goto bad;
 810
 811         r = parse_hw_handler(&as, m);
 812         if (r)
 813                 goto bad;
 814
 815         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 816         if (r)
 817                 goto bad;
 818
 819         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 820         if (r)
 821                 goto bad;
 822
 823         if ((!m->nr_priority_groups && next_pg_num) ||
 824             (m->nr_priority_groups && !next_pg_num)) {
 825                 ti->error = "invalid initial priority group";
 826                 r = -EINVAL;
 827                 goto bad;
 828         }
 829
 830         /* parse the priority groups */
 831         while (as.argc) {
 832                 struct priority_group *pg;
 833
 834                 pg = parse_priority_group(&as, m);
 835                 if (IS_ERR(pg)) {
 836                         r = PTR_ERR(pg);
 837                         goto bad;
 838                 }
 839
 840                 m->nr_valid_paths += pg->nr_pgpaths;
 841                 list_add_tail(&pg->list, &m->priority_groups);
 842                 pg_count++;
 843                 pg->pg_num = pg_count;
 844                 if (!--next_pg_num)
 845                         m->next_pg = pg;
 846         }
 847
 848         if (pg_count != m->nr_priority_groups) {
 849                 ti->error = "priority group count mismatch";
 850                 r = -EINVAL;
 851                 goto bad;
 852         }
 853
 854         ti->num_flush_requests = 1;
 855         ti->num_discard_requests = 1;
 856
 857         return 0;
 858
 859  bad:
 860         free_multipath(m);
 861         return r;
 862 }
 863
 864 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 865 {
 866         DECLARE_WAITQUEUE(wait, current);
 867         unsigned long flags;
 868
 869         add_wait_queue(&m->pg_init_wait, &wait);
 870
 871         while (1) {
 872                 set_current_state(TASK_UNINTERRUPTIBLE);
 873
 874                 spin_lock_irqsave(&m->lock, flags);
 875                 if (!m->pg_init_in_progress) {
 876                         spin_unlock_irqrestore(&m->lock, flags);
 877                         break;
 878                 }
 879                 spin_unlock_irqrestore(&m->lock, flags);
 880
 881                 io_schedule();
 882         }
 883         set_current_state(TASK_RUNNING);
 884
 885         remove_wait_queue(&m->pg_init_wait, &wait);
 886 }
 887
 888 static void flush_multipath_work(struct multipath *m)
 889 {
 890         flush_workqueue(kmpath_handlerd);
 891         multipath_wait_for_pg_init_completion(m);
 892         flush_workqueue(kmultipathd);
 893         flush_work_sync(&m->trigger_event);
 894 }
 895
 896 static void multipath_dtr(struct dm_target *ti)
 897 {
 898         struct multipath *m = ti->private;
 899
 900         flush_multipath_work(m);
 901         free_multipath(m);
 902 }
 903
 904 /*
 905  * Map cloned requests
 906  */
 907 static int multipath_map(struct dm_target *ti, struct request *clone,
 908                          union map_info *map_context)
 909 {
 910         int r;
 911         struct dm_mpath_io *mpio;
 912         struct multipath *m = (struct multipath *) ti->private;
 913
 914         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 915         if (!mpio)
 916                 /* ENOMEM, requeue */
 917                 return DM_MAPIO_REQUEUE;
 918         memset(mpio, 0, sizeof(*mpio));
 919
 920         map_context->ptr = mpio;
 921         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 922         r = map_io(m, clone, mpio, 0);
 923         if (r < 0 || r == DM_MAPIO_REQUEUE)
 924                 mempool_free(mpio, m->mpio_pool);
 925
 926         return r;
 927 }
 928
 929 /*
 930  * Take a path out of use.
 931  */
 932 static int fail_path(struct pgpath *pgpath)
 933 {
 934         unsigned long flags;
 935         struct multipath *m = pgpath->pg->m;
 936
 937         spin_lock_irqsave(&m->lock, flags);
 938
 939         if (!pgpath->is_active)
 940                 goto out;
 941
 942         DMWARN("Failing path %s.", pgpath->path.dev->name);
 943
 944         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 945         pgpath->is_active = 0;
 946         pgpath->fail_count++;
 947
 948         m->nr_valid_paths--;
 949
 950         if (pgpath == m->current_pgpath)
 951                 m->current_pgpath = NULL;
 952
 953         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 954                       pgpath->path.dev->name, m->nr_valid_paths);
 955
 956         schedule_work(&m->trigger_event);
 957
 958 out:
 959         spin_unlock_irqrestore(&m->lock, flags);
 960
 961         return 0;
 962 }
 963
 964 /*
 965  * Reinstate a previously-failed path
 966  */
 967 static int reinstate_path(struct pgpath *pgpath)
 968 {
 969         int r = 0;
 970         unsigned long flags;
 971         struct multipath *m = pgpath->pg->m;
 972
 973         spin_lock_irqsave(&m->lock, flags);
 974
 975         if (pgpath->is_active)
 976                 goto out;
 977
 978         if (!pgpath->pg->ps.type->reinstate_path) {
 979                 DMWARN("Reinstate path not supported by path selector %s",
 980                        pgpath->pg->ps.type->name);
 981                 r = -EINVAL;
 982                 goto out;
 983         }
 984
 985         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 986         if (r)
 987                 goto out;
 988
 989         pgpath->is_active = 1;
 990
 991         if (!m->nr_valid_paths++ && m->queue_size) {
 992                 m->current_pgpath = NULL;
 993                 queue_work(kmultipathd, &m->process_queued_ios);
 994         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
 995                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
 996                         m->pg_init_in_progress++;
 997         }
 998
 999         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1000                       pgpath->path.dev->name, m->nr_valid_paths);
1001
1002         schedule_work(&m->trigger_event);
1003
1004 out:
1005         spin_unlock_irqrestore(&m->lock, flags);
1006
1007         return r;
1008 }
1009
1010 /*
1011  * Fail or reinstate all paths that match the provided struct dm_dev.
1012  */
1013 static int action_dev(struct multipath *m, struct dm_dev *dev,
1014                       action_fn action)
1015 {
1016         int r = -EINVAL;
1017         struct pgpath *pgpath;
1018         struct priority_group *pg;
1019
1020         list_for_each_entry(pg, &m->priority_groups, list) {
1021                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1022                         if (pgpath->path.dev == dev)
1023                                 r = action(pgpath);
1024                 }
1025         }
1026
1027         return r;
1028 }
1029
1030 /*
1031  * Temporarily try to avoid having to use the specified PG
1032  */
1033 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1034                       int bypassed)
1035 {
1036         unsigned long flags;
1037
1038         spin_lock_irqsave(&m->lock, flags);
1039
1040         pg->bypassed = bypassed;
1041         m->current_pgpath = NULL;
1042         m->current_pg = NULL;
1043
1044         spin_unlock_irqrestore(&m->lock, flags);
1045
1046         schedule_work(&m->trigger_event);
1047 }
1048
1049 /*
1050  * Switch to using the specified PG from the next I/O that gets mapped
1051  */
1052 static int switch_pg_num(struct multipath *m, const char *pgstr)
1053 {
1054         struct priority_group *pg;
1055         unsigned pgnum;
1056         unsigned long flags;
1057
1058         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1059             (pgnum > m->nr_priority_groups)) {
1060                 DMWARN("invalid PG number supplied to switch_pg_num");
1061                 return -EINVAL;
1062         }
1063
1064         spin_lock_irqsave(&m->lock, flags);
1065         list_for_each_entry(pg, &m->priority_groups, list) {
1066                 pg->bypassed = 0;
1067                 if (--pgnum)
1068                         continue;
1069
1070                 m->current_pgpath = NULL;
1071                 m->current_pg = NULL;
1072                 m->next_pg = pg;
1073         }
1074         spin_unlock_irqrestore(&m->lock, flags);
1075
1076         schedule_work(&m->trigger_event);
1077         return 0;
1078 }
1079
1080 /*
1081  * Set/clear bypassed status of a PG.
1082  * PGs are numbered upwards from 1 in the order they were declared.
1083  */
1084 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1085 {
1086         struct priority_group *pg;
1087         unsigned pgnum;
1088
1089         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1090             (pgnum > m->nr_priority_groups)) {
1091                 DMWARN("invalid PG number supplied to bypass_pg");
1092                 return -EINVAL;
1093         }
1094
1095         list_for_each_entry(pg, &m->priority_groups, list) {
1096                 if (!--pgnum)
1097                         break;
1098         }
1099
1100         bypass_pg(m, pg, bypassed);
1101         return 0;
1102 }
1103
1104 /*
1105  * Should we retry pg_init immediately?
1106  */
1107 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1108 {
1109         unsigned long flags;
1110         int limit_reached = 0;
1111
1112         spin_lock_irqsave(&m->lock, flags);
1113
1114         if (m->pg_init_count <= m->pg_init_retries)
1115                 m->pg_init_required = 1;
1116         else
1117                 limit_reached = 1;
1118
1119         spin_unlock_irqrestore(&m->lock, flags);
1120
1121         return limit_reached;
1122 }
1123
1124 static void pg_init_done(void *data, int errors)
1125 {
1126         struct pgpath *pgpath = data;
1127         struct priority_group *pg = pgpath->pg;
1128         struct multipath *m = pg->m;
1129         unsigned long flags;
1130         unsigned delay_retry = 0;
1131
1132         /* device or driver problems */
1133         switch (errors) {
1134         case SCSI_DH_OK:
1135                 break;
1136         case SCSI_DH_NOSYS:
1137                 if (!m->hw_handler_name) {
1138                         errors = 0;
1139                         break;
1140                 }
1141                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1142                       "Error %d.", m->hw_handler_name, errors);
1143                 /*
1144                  * Fail path for now, so we do not ping pong
1145                  */
1146                 fail_path(pgpath);
1147                 break;
1148         case SCSI_DH_DEV_TEMP_BUSY:
1149                 /*
1150                  * Probably doing something like FW upgrade on the
1151                  * controller so try the other pg.
1152                  */
1153                 bypass_pg(m, pg, 1);
1154                 break;
1155         case SCSI_DH_RETRY:
1156                 /* Wait before retrying. */
1157                 delay_retry = 1;
1158         case SCSI_DH_IMM_RETRY:
1159         case SCSI_DH_RES_TEMP_UNAVAIL:
1160                 if (pg_init_limit_reached(m, pgpath))
1161                         fail_path(pgpath);
1162                 errors = 0;
1163                 break;
1164         default:
1165                 /*
1166                  * We probably do not want to fail the path for a device
1167                  * error, but this is what the old dm did. In future
1168                  * patches we can do more advanced handling.
1169                  */
1170                 fail_path(pgpath);
1171         }
1172
1173         spin_lock_irqsave(&m->lock, flags);
1174         if (errors) {
1175                 if (pgpath == m->current_pgpath) {
1176                         DMERR("Could not failover device. Error %d.", errors);
1177                         m->current_pgpath = NULL;
1178                         m->current_pg = NULL;
1179                 }
1180         } else if (!m->pg_init_required)
1181                 pg->bypassed = 0;
1182
1183         if (--m->pg_init_in_progress)
1184                 /* Activations of other paths are still on going */
1185                 goto out;
1186
1187         if (!m->pg_init_required)
1188                 m->queue_io = 0;
1189
1190         m->pg_init_delay_retry = delay_retry;
1191         queue_work(kmultipathd, &m->process_queued_ios);
1192
1193         /*
1194          * Wake up any thread waiting to suspend.
1195          */
1196         wake_up(&m->pg_init_wait);
1197
1198 out:
1199         spin_unlock_irqrestore(&m->lock, flags);
1200 }
1201
1202 static void activate_path(struct work_struct *work)
1203 {
1204         struct pgpath *pgpath =
1205                 container_of(work, struct pgpath, activate_path.work);
1206
1207         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1208                                 pg_init_done, pgpath);
1209 }
1210
1211 /*
1212  * end_io handling
1213  */
1214 static int do_end_io(struct multipath *m, struct request *clone,
1215                      int error, struct dm_mpath_io *mpio)
1216 {
1217         /*
1218          * We don't queue any clone request inside the multipath target
1219          * during end I/O handling, since those clone requests don't have
1220          * bio clones.  If we queue them inside the multipath target,
1221          * we need to make bio clones, that requires memory allocation.
1222          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1223          *  don't have bio clones.)
1224          * Instead of queueing the clone request here, we queue the original
1225          * request into dm core, which will remake a clone request and
1226          * clone bios for it and resubmit it later.
1227          */
1228         int r = DM_ENDIO_REQUEUE;
1229         unsigned long flags;
1230
1231         if (!error && !clone->errors)
1232                 return 0;       /* I/O complete */
1233
1234         if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1235                 return error;
1236
1237         if (mpio->pgpath)
1238                 fail_path(mpio->pgpath);
1239
1240         spin_lock_irqsave(&m->lock, flags);
1241         if (!m->nr_valid_paths) {
1242                 if (!m->queue_if_no_path) {
1243                         if (!__must_push_back(m))
1244                                 r = -EIO;
1245                 } else {
1246                         if (error == -EBADE)
1247                                 r = error;
1248                 }
1249         }
1250         spin_unlock_irqrestore(&m->lock, flags);
1251
1252         return r;
1253 }
1254
1255 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1256                             int error, union map_info *map_context)
1257 {
1258         struct multipath *m = ti->private;
1259         struct dm_mpath_io *mpio = map_context->ptr;
1260         struct pgpath *pgpath = mpio->pgpath;
1261         struct path_selector *ps;
1262         int r;
1263
1264         r  = do_end_io(m, clone, error, mpio);
1265         if (pgpath) {
1266                 ps = &pgpath->pg->ps;
1267                 if (ps->type->end_io)
1268                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1269         }
1270         mempool_free(mpio, m->mpio_pool);
1271
1272         return r;
1273 }
1274
1275 /*
1276  * Suspend can't complete until all the I/O is processed so if
1277  * the last path fails we must error any remaining I/O.
1278  * Note that if the freeze_bdev fails while suspending, the
1279  * queue_if_no_path state is lost - userspace should reset it.
1280  */
1281 static void multipath_presuspend(struct dm_target *ti)
1282 {
1283         struct multipath *m = (struct multipath *) ti->private;
1284
1285         queue_if_no_path(m, 0, 1);
1286 }
1287
1288 static void multipath_postsuspend(struct dm_target *ti)
1289 {
1290         struct multipath *m = ti->private;
1291
1292         mutex_lock(&m->work_mutex);
1293         flush_multipath_work(m);
1294         mutex_unlock(&m->work_mutex);
1295 }
1296
1297 /*
1298  * Restore the queue_if_no_path setting.
1299  */
1300 static void multipath_resume(struct dm_target *ti)
1301 {
1302         struct multipath *m = (struct multipath *) ti->private;
1303         unsigned long flags;
1304
1305         spin_lock_irqsave(&m->lock, flags);
1306         m->queue_if_no_path = m->saved_queue_if_no_path;
1307         spin_unlock_irqrestore(&m->lock, flags);
1308 }
1309
1310 /*
1311  * Info output has the following format:
1312  * num_multipath_feature_args [multipath_feature_args]*
1313  * num_handler_status_args [handler_status_args]*
1314  * num_groups init_group_number
1315  *            [A|D|E num_ps_status_args [ps_status_args]*
1316  *             num_paths num_selector_args
1317  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1318  *
1319  * Table output has the following format (identical to the constructor string):
1320  * num_feature_args [features_args]*
1321  * num_handler_args hw_handler [hw_handler_args]*
1322  * num_groups init_group_number
1323  *     [priority selector-name num_ps_args [ps_args]*
1324  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1325  */
1326 static int multipath_status(struct dm_target *ti, status_type_t type,
1327                             char *result, unsigned int maxlen)
1328 {
1329         int sz = 0;
1330         unsigned long flags;
1331         struct multipath *m = (struct multipath *) ti->private;
1332         struct priority_group *pg;
1333         struct pgpath *p;
1334         unsigned pg_num;
1335         char state;
1336
1337         spin_lock_irqsave(&m->lock, flags);
1338
1339         /* Features */
1340         if (type == STATUSTYPE_INFO)
1341                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1342         else {
1343                 DMEMIT("%u ", m->queue_if_no_path +
1344                               (m->pg_init_retries > 0) * 2 +
1345                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2);
1346                 if (m->queue_if_no_path)
1347                         DMEMIT("queue_if_no_path ");
1348                 if (m->pg_init_retries)
1349                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1350                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1351                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1352         }
1353
1354         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1355                 DMEMIT("0 ");
1356         else
1357                 DMEMIT("1 %s ", m->hw_handler_name);
1358
1359         DMEMIT("%u ", m->nr_priority_groups);
1360
1361         if (m->next_pg)
1362                 pg_num = m->next_pg->pg_num;
1363         else if (m->current_pg)
1364                 pg_num = m->current_pg->pg_num;
1365         else
1366                 pg_num = (m->nr_priority_groups ? 1 : 0);
1367
1368         DMEMIT("%u ", pg_num);
1369
1370         switch (type) {
1371         case STATUSTYPE_INFO:
1372                 list_for_each_entry(pg, &m->priority_groups, list) {
1373                         if (pg->bypassed)
1374                                 state = 'D';    /* Disabled */
1375                         else if (pg == m->current_pg)
1376                                 state = 'A';    /* Currently Active */
1377                         else
1378                                 state = 'E';    /* Enabled */
1379
1380                         DMEMIT("%c ", state);
1381
1382                         if (pg->ps.type->status)
1383                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1384                                                           result + sz,
1385                                                           maxlen - sz);
1386                         else
1387                                 DMEMIT("0 ");
1388
1389                         DMEMIT("%u %u ", pg->nr_pgpaths,
1390                                pg->ps.type->info_args);
1391
1392                         list_for_each_entry(p, &pg->pgpaths, list) {
1393                                 DMEMIT("%s %s %u ", p->path.dev->name,
1394                                        p->is_active ? "A" : "F",
1395                                        p->fail_count);
1396                                 if (pg->ps.type->status)
1397                                         sz += pg->ps.type->status(&pg->ps,
1398                                               &p->path, type, result + sz,
1399                                               maxlen - sz);
1400                         }
1401                 }
1402                 break;
1403
1404         case STATUSTYPE_TABLE:
1405                 list_for_each_entry(pg, &m->priority_groups, list) {
1406                         DMEMIT("%s ", pg->ps.type->name);
1407
1408                         if (pg->ps.type->status)
1409                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1410                                                           result + sz,
1411                                                           maxlen - sz);
1412                         else
1413                                 DMEMIT("0 ");
1414
1415                         DMEMIT("%u %u ", pg->nr_pgpaths,
1416                                pg->ps.type->table_args);
1417
1418                         list_for_each_entry(p, &pg->pgpaths, list) {
1419                                 DMEMIT("%s ", p->path.dev->name);
1420                                 if (pg->ps.type->status)
1421                                         sz += pg->ps.type->status(&pg->ps,
1422                                               &p->path, type, result + sz,
1423                                               maxlen - sz);
1424                         }
1425                 }
1426                 break;
1427         }
1428
1429         spin_unlock_irqrestore(&m->lock, flags);
1430
1431         return 0;
1432 }
1433
1434 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1435 {
1436         int r = -EINVAL;
1437         struct dm_dev *dev;
1438         struct multipath *m = (struct multipath *) ti->private;
1439         action_fn action;
1440
1441         mutex_lock(&m->work_mutex);
1442
1443         if (dm_suspended(ti)) {
1444                 r = -EBUSY;
1445                 goto out;
1446         }
1447
1448         if (argc == 1) {
1449                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1450                         r = queue_if_no_path(m, 1, 0);
1451                         goto out;
1452                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1453                         r = queue_if_no_path(m, 0, 0);
1454                         goto out;
1455                 }
1456         }
1457
1458         if (argc != 2) {
1459                 DMWARN("Unrecognised multipath message received.");
1460                 goto out;
1461         }
1462
1463         if (!strcasecmp(argv[0], "disable_group")) {
1464                 r = bypass_pg_num(m, argv[1], 1);
1465                 goto out;
1466         } else if (!strcasecmp(argv[0], "enable_group")) {
1467                 r = bypass_pg_num(m, argv[1], 0);
1468                 goto out;
1469         } else if (!strcasecmp(argv[0], "switch_group")) {
1470                 r = switch_pg_num(m, argv[1]);
1471                 goto out;
1472         } else if (!strcasecmp(argv[0], "reinstate_path"))
1473                 action = reinstate_path;
1474         else if (!strcasecmp(argv[0], "fail_path"))
1475                 action = fail_path;
1476         else {
1477                 DMWARN("Unrecognised multipath message received.");
1478                 goto out;
1479         }
1480
1481         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1482         if (r) {
1483                 DMWARN("message: error getting device %s",
1484                        argv[1]);
1485                 goto out;
1486         }
1487
1488         r = action_dev(m, dev, action);
1489
1490         dm_put_device(ti, dev);
1491
1492 out:
1493         mutex_unlock(&m->work_mutex);
1494         return r;
1495 }
1496
1497 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1498                            unsigned long arg)
1499 {
1500         struct multipath *m = (struct multipath *) ti->private;
1501         struct block_device *bdev = NULL;
1502         fmode_t mode = 0;
1503         unsigned long flags;
1504         int r = 0;
1505
1506         spin_lock_irqsave(&m->lock, flags);
1507
1508         if (!m->current_pgpath)
1509                 __choose_pgpath(m, 0);
1510
1511         if (m->current_pgpath) {
1512                 bdev = m->current_pgpath->path.dev->bdev;
1513                 mode = m->current_pgpath->path.dev->mode;
1514         }
1515
1516         if (m->queue_io)
1517                 r = -EAGAIN;
1518         else if (!bdev)
1519                 r = -EIO;
1520
1521         spin_unlock_irqrestore(&m->lock, flags);
1522
1523         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1524 }
1525
1526 static int multipath_iterate_devices(struct dm_target *ti,
1527                                      iterate_devices_callout_fn fn, void *data)
1528 {
1529         struct multipath *m = ti->private;
1530         struct priority_group *pg;
1531         struct pgpath *p;
1532         int ret = 0;
1533
1534         list_for_each_entry(pg, &m->priority_groups, list) {
1535                 list_for_each_entry(p, &pg->pgpaths, list) {
1536                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1537                         if (ret)
1538                                 goto out;
1539                 }
1540         }
1541
1542 out:
1543         return ret;
1544 }
1545
1546 static int __pgpath_busy(struct pgpath *pgpath)
1547 {
1548         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1549
1550         return dm_underlying_device_busy(q);
1551 }
1552
1553 /*
1554  * We return "busy", only when we can map I/Os but underlying devices
1555  * are busy (so even if we map I/Os now, the I/Os will wait on
1556  * the underlying queue).
1557  * In other words, if we want to kill I/Os or queue them inside us
1558  * due to map unavailability, we don't return "busy".  Otherwise,
1559  * dm core won't give us the I/Os and we can't do what we want.
1560  */
1561 static int multipath_busy(struct dm_target *ti)
1562 {
1563         int busy = 0, has_active = 0;
1564         struct multipath *m = ti->private;
1565         struct priority_group *pg;
1566         struct pgpath *pgpath;
1567         unsigned long flags;
1568
1569         spin_lock_irqsave(&m->lock, flags);
1570
1571         /* Guess which priority_group will be used at next mapping time */
1572         if (unlikely(!m->current_pgpath && m->next_pg))
1573                 pg = m->next_pg;
1574         else if (likely(m->current_pg))
1575                 pg = m->current_pg;
1576         else
1577                 /*
1578                  * We don't know which pg will be used at next mapping time.
1579                  * We don't call __choose_pgpath() here to avoid to trigger
1580                  * pg_init just by busy checking.
1581                  * So we don't know whether underlying devices we will be using
1582                  * at next mapping time are busy or not. Just try mapping.
1583                  */
1584                 goto out;
1585
1586         /*
1587          * If there is one non-busy active path at least, the path selector
1588          * will be able to select it. So we consider such a pg as not busy.
1589          */
1590         busy = 1;
1591         list_for_each_entry(pgpath, &pg->pgpaths, list)
1592                 if (pgpath->is_active) {
1593                         has_active = 1;
1594
1595                         if (!__pgpath_busy(pgpath)) {
1596                                 busy = 0;
1597                                 break;
1598                         }
1599                 }
1600
1601         if (!has_active)
1602                 /*
1603                  * No active path in this pg, so this pg won't be used and
1604                  * the current_pg will be changed at next mapping time.
1605                  * We need to try mapping to determine it.
1606                  */
1607                 busy = 0;
1608
1609 out:
1610         spin_unlock_irqrestore(&m->lock, flags);
1611
1612         return busy;
1613 }
1614
1615 /*-----------------------------------------------------------------
1616  * Module setup
1617  *---------------------------------------------------------------*/
1618 static struct target_type multipath_target = {
1619         .name = "multipath",
1620         .version = {1, 3, 0},
1621         .module = THIS_MODULE,
1622         .ctr = multipath_ctr,
1623         .dtr = multipath_dtr,
1624         .map_rq = multipath_map,
1625         .rq_end_io = multipath_end_io,
1626         .presuspend = multipath_presuspend,
1627         .postsuspend = multipath_postsuspend,
1628         .resume = multipath_resume,
1629         .status = multipath_status,
1630         .message = multipath_message,
1631         .ioctl  = multipath_ioctl,
1632         .iterate_devices = multipath_iterate_devices,
1633         .busy = multipath_busy,
1634 };
1635
1636 static int __init dm_multipath_init(void)
1637 {
1638         int r;
1639
1640         /* allocate a slab for the dm_ios */
1641         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1642         if (!_mpio_cache)
1643                 return -ENOMEM;
1644
1645         r = dm_register_target(&multipath_target);
1646         if (r < 0) {
1647                 DMERR("register failed %d", r);
1648                 kmem_cache_destroy(_mpio_cache);
1649                 return -EINVAL;
1650         }
1651
1652         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1653         if (!kmultipathd) {
1654                 DMERR("failed to create workqueue kmpathd");
1655                 dm_unregister_target(&multipath_target);
1656                 kmem_cache_destroy(_mpio_cache);
1657                 return -ENOMEM;
1658         }
1659
1660         /*
1661          * A separate workqueue is used to handle the device handlers
1662          * to avoid overloading existing workqueue. Overloading the
1663          * old workqueue would also create a bottleneck in the
1664          * path of the storage hardware device activation.
1665          */
1666         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1667                                                   WQ_MEM_RECLAIM);
1668         if (!kmpath_handlerd) {
1669                 DMERR("failed to create workqueue kmpath_handlerd");
1670                 destroy_workqueue(kmultipathd);
1671                 dm_unregister_target(&multipath_target);
1672                 kmem_cache_destroy(_mpio_cache);
1673                 return -ENOMEM;
1674         }
1675
1676         DMINFO("version %u.%u.%u loaded",
1677                multipath_target.version[0], multipath_target.version[1],
1678                multipath_target.version[2]);
1679
1680         return r;
1681 }
1682
1683 static void __exit dm_multipath_exit(void)
1684 {
1685         destroy_workqueue(kmpath_handlerd);
1686         destroy_workqueue(kmultipathd);
1687
1688         dm_unregister_target(&multipath_target);
1689         kmem_cache_destroy(_mpio_cache);
1690 }
1691
1692 module_init(dm_multipath_init);
1693 module_exit(dm_multipath_exit);
1694
1695 MODULE_DESCRIPTION(DM_NAME " multipath target");
1696 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1697 MODULE_LICENSE("GPL");