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 <asm/atomic.h>
  23
  24 #define DM_MSG_PREFIX "multipath"
  25 #define MESG_STR(x) x, sizeof(x)
  26
  27 /* Path properties */
  28 struct pgpath {
  29         struct list_head list;
  30
  31         struct priority_group *pg;      /* Owning PG */
  32         unsigned is_active;             /* Path status */
  33         unsigned fail_count;            /* Cumulative failure count */
  34
  35         struct dm_path path;
  36         struct work_struct deactivate_path;
  37         struct work_struct 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         unsigned nr_priority_groups;
  69         struct list_head priority_groups;
  70         unsigned pg_init_required;      /* pg_init needs calling? */
  71         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  72         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  73
  74         unsigned nr_valid_paths;        /* Total number of usable paths */
  75         struct pgpath *current_pgpath;
  76         struct priority_group *current_pg;
  77         struct priority_group *next_pg; /* Switch to this PG if set */
  78         unsigned repeat_count;          /* I/Os left before calling PS again */
  79
  80         unsigned queue_io;              /* Must we queue all I/O? */
  81         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  82         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  83         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  84         unsigned pg_init_count;         /* Number of times pg_init called */
  85
  86         struct work_struct process_queued_ios;
  87         struct list_head queued_ios;
  88         unsigned queue_size;
  89
  90         struct work_struct trigger_event;
  91
  92         /*
  93          * We must use a mempool of dm_mpath_io structs so that we
  94          * can resubmit bios on error.
  95          */
  96         mempool_t *mpio_pool;
  97
  98         struct mutex work_mutex;
  99 };
 100
 101 /*
 102  * Context information attached to each bio we process.
 103  */
 104 struct dm_mpath_io {
 105         struct pgpath *pgpath;
 106         size_t nr_bytes;
 107 };
 108
 109 typedef int (*action_fn) (struct pgpath *pgpath);
 110
 111 #define MIN_IOS 256     /* Mempool size */
 112
 113 static struct kmem_cache *_mpio_cache;
 114
 115 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 116 static void process_queued_ios(struct work_struct *work);
 117 static void trigger_event(struct work_struct *work);
 118 static void activate_path(struct work_struct *work);
 119 static void deactivate_path(struct work_struct *work);
 120
 121
 122 /*-----------------------------------------------
 123  * Allocation routines
 124  *-----------------------------------------------*/
 125
 126 static struct pgpath *alloc_pgpath(void)
 127 {
 128         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 129
 130         if (pgpath) {
 131                 pgpath->is_active = 1;
 132                 INIT_WORK(&pgpath->deactivate_path, deactivate_path);
 133                 INIT_WORK(&pgpath->activate_path, activate_path);
 134         }
 135
 136         return pgpath;
 137 }
 138
 139 static void free_pgpath(struct pgpath *pgpath)
 140 {
 141         kfree(pgpath);
 142 }
 143
 144 static void deactivate_path(struct work_struct *work)
 145 {
 146         struct pgpath *pgpath =
 147                 container_of(work, struct pgpath, deactivate_path);
 148
 149         blk_abort_queue(pgpath->path.dev->bdev->bd_disk->queue);
 150 }
 151
 152 static struct priority_group *alloc_priority_group(void)
 153 {
 154         struct priority_group *pg;
 155
 156         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 157
 158         if (pg)
 159                 INIT_LIST_HEAD(&pg->pgpaths);
 160
 161         return pg;
 162 }
 163
 164 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 165 {
 166         struct pgpath *pgpath, *tmp;
 167         struct multipath *m = ti->private;
 168
 169         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 170                 list_del(&pgpath->list);
 171                 if (m->hw_handler_name)
 172                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 173                 dm_put_device(ti, pgpath->path.dev);
 174                 free_pgpath(pgpath);
 175         }
 176 }
 177
 178 static void free_priority_group(struct priority_group *pg,
 179                                 struct dm_target *ti)
 180 {
 181         struct path_selector *ps = &pg->ps;
 182
 183         if (ps->type) {
 184                 ps->type->destroy(ps);
 185                 dm_put_path_selector(ps->type);
 186         }
 187
 188         free_pgpaths(&pg->pgpaths, ti);
 189         kfree(pg);
 190 }
 191
 192 static struct multipath *alloc_multipath(struct dm_target *ti)
 193 {
 194         struct multipath *m;
 195
 196         m = kzalloc(sizeof(*m), GFP_KERNEL);
 197         if (m) {
 198                 INIT_LIST_HEAD(&m->priority_groups);
 199                 INIT_LIST_HEAD(&m->queued_ios);
 200                 spin_lock_init(&m->lock);
 201                 m->queue_io = 1;
 202                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 203                 INIT_WORK(&m->trigger_event, trigger_event);
 204                 init_waitqueue_head(&m->pg_init_wait);
 205                 mutex_init(&m->work_mutex);
 206                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 207                 if (!m->mpio_pool) {
 208                         kfree(m);
 209                         return NULL;
 210                 }
 211                 m->ti = ti;
 212                 ti->private = m;
 213         }
 214
 215         return m;
 216 }
 217
 218 static void free_multipath(struct multipath *m)
 219 {
 220         struct priority_group *pg, *tmp;
 221
 222         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 223                 list_del(&pg->list);
 224                 free_priority_group(pg, m->ti);
 225         }
 226
 227         kfree(m->hw_handler_name);
 228         kfree(m->hw_handler_params);
 229         mempool_destroy(m->mpio_pool);
 230         kfree(m);
 231 }
 232
 233
 234 /*-----------------------------------------------
 235  * Path selection
 236  *-----------------------------------------------*/
 237
 238 static void __pg_init_all_paths(struct multipath *m)
 239 {
 240         struct pgpath *pgpath;
 241
 242         m->pg_init_count++;
 243         m->pg_init_required = 0;
 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_work(kmpath_handlerd, &pgpath->activate_path))
 249                         m->pg_init_in_progress++;
 250         }
 251 }
 252
 253 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 254 {
 255         m->current_pg = pgpath->pg;
 256
 257         /* Must we initialise the PG first, and queue I/O till it's ready? */
 258         if (m->hw_handler_name) {
 259                 m->pg_init_required = 1;
 260                 m->queue_io = 1;
 261         } else {
 262                 m->pg_init_required = 0;
 263                 m->queue_io = 0;
 264         }
 265
 266         m->pg_init_count = 0;
 267 }
 268
 269 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 270                                size_t nr_bytes)
 271 {
 272         struct dm_path *path;
 273
 274         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 275         if (!path)
 276                 return -ENXIO;
 277
 278         m->current_pgpath = path_to_pgpath(path);
 279
 280         if (m->current_pg != pg)
 281                 __switch_pg(m, m->current_pgpath);
 282
 283         return 0;
 284 }
 285
 286 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 287 {
 288         struct priority_group *pg;
 289         unsigned bypassed = 1;
 290
 291         if (!m->nr_valid_paths)
 292                 goto failed;
 293
 294         /* Were we instructed to switch PG? */
 295         if (m->next_pg) {
 296                 pg = m->next_pg;
 297                 m->next_pg = NULL;
 298                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 299                         return;
 300         }
 301
 302         /* Don't change PG until it has no remaining paths */
 303         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 304                 return;
 305
 306         /*
 307          * Loop through priority groups until we find a valid path.
 308          * First time we skip PGs marked 'bypassed'.
 309          * Second time we only try the ones we skipped.
 310          */
 311         do {
 312                 list_for_each_entry(pg, &m->priority_groups, list) {
 313                         if (pg->bypassed == bypassed)
 314                                 continue;
 315                         if (!__choose_path_in_pg(m, pg, nr_bytes))
 316                                 return;
 317                 }
 318         } while (bypassed--);
 319
 320 failed:
 321         m->current_pgpath = NULL;
 322         m->current_pg = NULL;
 323 }
 324
 325 /*
 326  * Check whether bios must be queued in the device-mapper core rather
 327  * than here in the target.
 328  *
 329  * m->lock must be held on entry.
 330  *
 331  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 332  * same value then we are not between multipath_presuspend()
 333  * and multipath_resume() calls and we have no need to check
 334  * for the DMF_NOFLUSH_SUSPENDING flag.
 335  */
 336 static int __must_push_back(struct multipath *m)
 337 {
 338         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 339                 dm_noflush_suspending(m->ti));
 340 }
 341
 342 static int map_io(struct multipath *m, struct request *clone,
 343                   struct dm_mpath_io *mpio, unsigned was_queued)
 344 {
 345         int r = DM_MAPIO_REMAPPED;
 346         size_t nr_bytes = blk_rq_bytes(clone);
 347         unsigned long flags;
 348         struct pgpath *pgpath;
 349         struct block_device *bdev;
 350
 351         spin_lock_irqsave(&m->lock, flags);
 352
 353         /* Do we need to select a new pgpath? */
 354         if (!m->current_pgpath ||
 355             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 356                 __choose_pgpath(m, nr_bytes);
 357
 358         pgpath = m->current_pgpath;
 359
 360         if (was_queued)
 361                 m->queue_size--;
 362
 363         if ((pgpath && m->queue_io) ||
 364             (!pgpath && m->queue_if_no_path)) {
 365                 /* Queue for the daemon to resubmit */
 366                 list_add_tail(&clone->queuelist, &m->queued_ios);
 367                 m->queue_size++;
 368                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 369                     !m->queue_io)
 370                         queue_work(kmultipathd, &m->process_queued_ios);
 371                 pgpath = NULL;
 372                 r = DM_MAPIO_SUBMITTED;
 373         } else if (pgpath) {
 374                 bdev = pgpath->path.dev->bdev;
 375                 clone->q = bdev_get_queue(bdev);
 376                 clone->rq_disk = bdev->bd_disk;
 377         } else if (__must_push_back(m))
 378                 r = DM_MAPIO_REQUEUE;
 379         else
 380                 r = -EIO;       /* Failed */
 381
 382         mpio->pgpath = pgpath;
 383         mpio->nr_bytes = nr_bytes;
 384
 385         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 386                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 387                                               nr_bytes);
 388
 389         spin_unlock_irqrestore(&m->lock, flags);
 390
 391         return r;
 392 }
 393
 394 /*
 395  * If we run out of usable paths, should we queue I/O or error it?
 396  */
 397 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 398                             unsigned save_old_value)
 399 {
 400         unsigned long flags;
 401
 402         spin_lock_irqsave(&m->lock, flags);
 403
 404         if (save_old_value)
 405                 m->saved_queue_if_no_path = m->queue_if_no_path;
 406         else
 407                 m->saved_queue_if_no_path = queue_if_no_path;
 408         m->queue_if_no_path = queue_if_no_path;
 409         if (!m->queue_if_no_path && m->queue_size)
 410                 queue_work(kmultipathd, &m->process_queued_ios);
 411
 412         spin_unlock_irqrestore(&m->lock, flags);
 413
 414         return 0;
 415 }
 416
 417 /*-----------------------------------------------------------------
 418  * The multipath daemon is responsible for resubmitting queued ios.
 419  *---------------------------------------------------------------*/
 420
 421 static void dispatch_queued_ios(struct multipath *m)
 422 {
 423         int r;
 424         unsigned long flags;
 425         struct dm_mpath_io *mpio;
 426         union map_info *info;
 427         struct request *clone, *n;
 428         LIST_HEAD(cl);
 429
 430         spin_lock_irqsave(&m->lock, flags);
 431         list_splice_init(&m->queued_ios, &cl);
 432         spin_unlock_irqrestore(&m->lock, flags);
 433
 434         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 435                 list_del_init(&clone->queuelist);
 436
 437                 info = dm_get_rq_mapinfo(clone);
 438                 mpio = info->ptr;
 439
 440                 r = map_io(m, clone, mpio, 1);
 441                 if (r < 0) {
 442                         mempool_free(mpio, m->mpio_pool);
 443                         dm_kill_unmapped_request(clone, r);
 444                 } else if (r == DM_MAPIO_REMAPPED)
 445                         dm_dispatch_request(clone);
 446                 else if (r == DM_MAPIO_REQUEUE) {
 447                         mempool_free(mpio, m->mpio_pool);
 448                         dm_requeue_unmapped_request(clone);
 449                 }
 450         }
 451 }
 452
 453 static void process_queued_ios(struct work_struct *work)
 454 {
 455         struct multipath *m =
 456                 container_of(work, struct multipath, process_queued_ios);
 457         struct pgpath *pgpath = NULL;
 458         unsigned must_queue = 1;
 459         unsigned long flags;
 460
 461         spin_lock_irqsave(&m->lock, flags);
 462
 463         if (!m->queue_size)
 464                 goto out;
 465
 466         if (!m->current_pgpath)
 467                 __choose_pgpath(m, 0);
 468
 469         pgpath = m->current_pgpath;
 470
 471         if ((pgpath && !m->queue_io) ||
 472             (!pgpath && !m->queue_if_no_path))
 473                 must_queue = 0;
 474
 475         if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
 476                 __pg_init_all_paths(m);
 477
 478 out:
 479         spin_unlock_irqrestore(&m->lock, flags);
 480         if (!must_queue)
 481                 dispatch_queued_ios(m);
 482 }
 483
 484 /*
 485  * An event is triggered whenever a path is taken out of use.
 486  * Includes path failure and PG bypass.
 487  */
 488 static void trigger_event(struct work_struct *work)
 489 {
 490         struct multipath *m =
 491                 container_of(work, struct multipath, trigger_event);
 492
 493         dm_table_event(m->ti->table);
 494 }
 495
 496 /*-----------------------------------------------------------------
 497  * Constructor/argument parsing:
 498  * <#multipath feature args> [<arg>]*
 499  * <#hw_handler args> [hw_handler [<arg>]*]
 500  * <#priority groups>
 501  * <initial priority group>
 502  *     [<selector> <#selector args> [<arg>]*
 503  *      <#paths> <#per-path selector args>
 504  *         [<path> [<arg>]* ]+ ]+
 505  *---------------------------------------------------------------*/
 506 struct param {
 507         unsigned min;
 508         unsigned max;
 509         char *error;
 510 };
 511
 512 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 513 {
 514         if (!str ||
 515             (sscanf(str, "%u", v) != 1) ||
 516             (*v < param->min) ||
 517             (*v > param->max)) {
 518                 *error = param->error;
 519                 return -EINVAL;
 520         }
 521
 522         return 0;
 523 }
 524
 525 struct arg_set {
 526         unsigned argc;
 527         char **argv;
 528 };
 529
 530 static char *shift(struct arg_set *as)
 531 {
 532         char *r;
 533
 534         if (as->argc) {
 535                 as->argc--;
 536                 r = *as->argv;
 537                 as->argv++;
 538                 return r;
 539         }
 540
 541         return NULL;
 542 }
 543
 544 static void consume(struct arg_set *as, unsigned n)
 545 {
 546         BUG_ON (as->argc < n);
 547         as->argc -= n;
 548         as->argv += n;
 549 }
 550
 551 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 552                                struct dm_target *ti)
 553 {
 554         int r;
 555         struct path_selector_type *pst;
 556         unsigned ps_argc;
 557
 558         static struct param _params[] = {
 559                 {0, 1024, "invalid number of path selector args"},
 560         };
 561
 562         pst = dm_get_path_selector(shift(as));
 563         if (!pst) {
 564                 ti->error = "unknown path selector type";
 565                 return -EINVAL;
 566         }
 567
 568         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 569         if (r) {
 570                 dm_put_path_selector(pst);
 571                 return -EINVAL;
 572         }
 573
 574         if (ps_argc > as->argc) {
 575                 dm_put_path_selector(pst);
 576                 ti->error = "not enough arguments for path selector";
 577                 return -EINVAL;
 578         }
 579
 580         r = pst->create(&pg->ps, ps_argc, as->argv);
 581         if (r) {
 582                 dm_put_path_selector(pst);
 583                 ti->error = "path selector constructor failed";
 584                 return r;
 585         }
 586
 587         pg->ps.type = pst;
 588         consume(as, ps_argc);
 589
 590         return 0;
 591 }
 592
 593 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 594                                struct dm_target *ti)
 595 {
 596         int r;
 597         struct pgpath *p;
 598         struct multipath *m = ti->private;
 599
 600         /* we need at least a path arg */
 601         if (as->argc < 1) {
 602                 ti->error = "no device given";
 603                 return ERR_PTR(-EINVAL);
 604         }
 605
 606         p = alloc_pgpath();
 607         if (!p)
 608                 return ERR_PTR(-ENOMEM);
 609
 610         r = dm_get_device(ti, shift(as), dm_table_get_mode(ti->table),
 611                           &p->path.dev);
 612         if (r) {
 613                 ti->error = "error getting device";
 614                 goto bad;
 615         }
 616
 617         if (m->hw_handler_name) {
 618                 struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 619
 620                 r = scsi_dh_attach(q, m->hw_handler_name);
 621                 if (r == -EBUSY) {
 622                         /*
 623                          * Already attached to different hw_handler,
 624                          * try to reattach with correct one.
 625                          */
 626                         scsi_dh_detach(q);
 627                         r = scsi_dh_attach(q, m->hw_handler_name);
 628                 }
 629
 630                 if (r < 0) {
 631                         ti->error = "error attaching hardware handler";
 632                         dm_put_device(ti, p->path.dev);
 633                         goto bad;
 634                 }
 635
 636                 if (m->hw_handler_params) {
 637                         r = scsi_dh_set_params(q, m->hw_handler_params);
 638                         if (r < 0) {
 639                                 ti->error = "unable to set hardware "
 640                                                         "handler parameters";
 641                                 scsi_dh_detach(q);
 642                                 dm_put_device(ti, p->path.dev);
 643                                 goto bad;
 644                         }
 645                 }
 646         }
 647
 648         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 649         if (r) {
 650                 dm_put_device(ti, p->path.dev);
 651                 goto bad;
 652         }
 653
 654         return p;
 655
 656  bad:
 657         free_pgpath(p);
 658         return ERR_PTR(r);
 659 }
 660
 661 static struct priority_group *parse_priority_group(struct arg_set *as,
 662                                                    struct multipath *m)
 663 {
 664         static struct param _params[] = {
 665                 {1, 1024, "invalid number of paths"},
 666                 {0, 1024, "invalid number of selector args"}
 667         };
 668
 669         int r;
 670         unsigned i, nr_selector_args, nr_params;
 671         struct priority_group *pg;
 672         struct dm_target *ti = m->ti;
 673
 674         if (as->argc < 2) {
 675                 as->argc = 0;
 676                 ti->error = "not enough priority group arguments";
 677                 return ERR_PTR(-EINVAL);
 678         }
 679
 680         pg = alloc_priority_group();
 681         if (!pg) {
 682                 ti->error = "couldn't allocate priority group";
 683                 return ERR_PTR(-ENOMEM);
 684         }
 685         pg->m = m;
 686
 687         r = parse_path_selector(as, pg, ti);
 688         if (r)
 689                 goto bad;
 690
 691         /*
 692          * read the paths
 693          */
 694         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 695         if (r)
 696                 goto bad;
 697
 698         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 699         if (r)
 700                 goto bad;
 701
 702         nr_params = 1 + nr_selector_args;
 703         for (i = 0; i < pg->nr_pgpaths; i++) {
 704                 struct pgpath *pgpath;
 705                 struct arg_set path_args;
 706
 707                 if (as->argc < nr_params) {
 708                         ti->error = "not enough path parameters";
 709                         r = -EINVAL;
 710                         goto bad;
 711                 }
 712
 713                 path_args.argc = nr_params;
 714                 path_args.argv = as->argv;
 715
 716                 pgpath = parse_path(&path_args, &pg->ps, ti);
 717                 if (IS_ERR(pgpath)) {
 718                         r = PTR_ERR(pgpath);
 719                         goto bad;
 720                 }
 721
 722                 pgpath->pg = pg;
 723                 list_add_tail(&pgpath->list, &pg->pgpaths);
 724                 consume(as, nr_params);
 725         }
 726
 727         return pg;
 728
 729  bad:
 730         free_priority_group(pg, ti);
 731         return ERR_PTR(r);
 732 }
 733
 734 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 735 {
 736         unsigned hw_argc;
 737         int ret;
 738         struct dm_target *ti = m->ti;
 739
 740         static struct param _params[] = {
 741                 {0, 1024, "invalid number of hardware handler args"},
 742         };
 743
 744         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 745                 return -EINVAL;
 746
 747         if (!hw_argc)
 748                 return 0;
 749
 750         if (hw_argc > as->argc) {
 751                 ti->error = "not enough arguments for hardware handler";
 752                 return -EINVAL;
 753         }
 754
 755         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 756         request_module("scsi_dh_%s", m->hw_handler_name);
 757         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 758                 ti->error = "unknown hardware handler type";
 759                 ret = -EINVAL;
 760                 goto fail;
 761         }
 762
 763         if (hw_argc > 1) {
 764                 char *p;
 765                 int i, j, len = 4;
 766
 767                 for (i = 0; i <= hw_argc - 2; i++)
 768                         len += strlen(as->argv[i]) + 1;
 769                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 770                 if (!p) {
 771                         ti->error = "memory allocation failed";
 772                         ret = -ENOMEM;
 773                         goto fail;
 774                 }
 775                 j = sprintf(p, "%d", hw_argc - 1);
 776                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 777                         j = sprintf(p, "%s", as->argv[i]);
 778         }
 779         consume(as, hw_argc - 1);
 780
 781         return 0;
 782 fail:
 783         kfree(m->hw_handler_name);
 784         m->hw_handler_name = NULL;
 785         return ret;
 786 }
 787
 788 static int parse_features(struct arg_set *as, struct multipath *m)
 789 {
 790         int r;
 791         unsigned argc;
 792         struct dm_target *ti = m->ti;
 793         const char *param_name;
 794
 795         static struct param _params[] = {
 796                 {0, 3, "invalid number of feature args"},
 797                 {1, 50, "pg_init_retries must be between 1 and 50"},
 798         };
 799
 800         r = read_param(_params, shift(as), &argc, &ti->error);
 801         if (r)
 802                 return -EINVAL;
 803
 804         if (!argc)
 805                 return 0;
 806
 807         do {
 808                 param_name = shift(as);
 809                 argc--;
 810
 811                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 812                         r = queue_if_no_path(m, 1, 0);
 813                         continue;
 814                 }
 815
 816                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 817                     (argc >= 1)) {
 818                         r = read_param(_params + 1, shift(as),
 819                                        &m->pg_init_retries, &ti->error);
 820                         argc--;
 821                         continue;
 822                 }
 823
 824                 ti->error = "Unrecognised multipath feature request";
 825                 r = -EINVAL;
 826         } while (argc && !r);
 827
 828         return r;
 829 }
 830
 831 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 832                          char **argv)
 833 {
 834         /* target parameters */
 835         static struct param _params[] = {
 836                 {1, 1024, "invalid number of priority groups"},
 837                 {1, 1024, "invalid initial priority group number"},
 838         };
 839
 840         int r;
 841         struct multipath *m;
 842         struct arg_set as;
 843         unsigned pg_count = 0;
 844         unsigned next_pg_num;
 845
 846         as.argc = argc;
 847         as.argv = argv;
 848
 849         m = alloc_multipath(ti);
 850         if (!m) {
 851                 ti->error = "can't allocate multipath";
 852                 return -EINVAL;
 853         }
 854
 855         r = parse_features(&as, m);
 856         if (r)
 857                 goto bad;
 858
 859         r = parse_hw_handler(&as, m);
 860         if (r)
 861                 goto bad;
 862
 863         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 864         if (r)
 865                 goto bad;
 866
 867         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 868         if (r)
 869                 goto bad;
 870
 871         /* parse the priority groups */
 872         while (as.argc) {
 873                 struct priority_group *pg;
 874
 875                 pg = parse_priority_group(&as, m);
 876                 if (IS_ERR(pg)) {
 877                         r = PTR_ERR(pg);
 878                         goto bad;
 879                 }
 880
 881                 m->nr_valid_paths += pg->nr_pgpaths;
 882                 list_add_tail(&pg->list, &m->priority_groups);
 883                 pg_count++;
 884                 pg->pg_num = pg_count;
 885                 if (!--next_pg_num)
 886                         m->next_pg = pg;
 887         }
 888
 889         if (pg_count != m->nr_priority_groups) {
 890                 ti->error = "priority group count mismatch";
 891                 r = -EINVAL;
 892                 goto bad;
 893         }
 894
 895         ti->num_flush_requests = 1;
 896
 897         return 0;
 898
 899  bad:
 900         free_multipath(m);
 901         return r;
 902 }
 903
 904 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 905 {
 906         DECLARE_WAITQUEUE(wait, current);
 907         unsigned long flags;
 908
 909         add_wait_queue(&m->pg_init_wait, &wait);
 910
 911         while (1) {
 912                 set_current_state(TASK_UNINTERRUPTIBLE);
 913
 914                 spin_lock_irqsave(&m->lock, flags);
 915                 if (!m->pg_init_in_progress) {
 916                         spin_unlock_irqrestore(&m->lock, flags);
 917                         break;
 918                 }
 919                 spin_unlock_irqrestore(&m->lock, flags);
 920
 921                 io_schedule();
 922         }
 923         set_current_state(TASK_RUNNING);
 924
 925         remove_wait_queue(&m->pg_init_wait, &wait);
 926 }
 927
 928 static void flush_multipath_work(struct multipath *m)
 929 {
 930         flush_workqueue(kmpath_handlerd);
 931         multipath_wait_for_pg_init_completion(m);
 932         flush_workqueue(kmultipathd);
 933         flush_scheduled_work();
 934 }
 935
 936 static void multipath_dtr(struct dm_target *ti)
 937 {
 938         struct multipath *m = ti->private;
 939
 940         flush_multipath_work(m);
 941         free_multipath(m);
 942 }
 943
 944 /*
 945  * Map cloned requests
 946  */
 947 static int multipath_map(struct dm_target *ti, struct request *clone,
 948                          union map_info *map_context)
 949 {
 950         int r;
 951         struct dm_mpath_io *mpio;
 952         struct multipath *m = (struct multipath *) ti->private;
 953
 954         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 955         if (!mpio)
 956                 /* ENOMEM, requeue */
 957                 return DM_MAPIO_REQUEUE;
 958         memset(mpio, 0, sizeof(*mpio));
 959
 960         map_context->ptr = mpio;
 961         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 962         r = map_io(m, clone, mpio, 0);
 963         if (r < 0 || r == DM_MAPIO_REQUEUE)
 964                 mempool_free(mpio, m->mpio_pool);
 965
 966         return r;
 967 }
 968
 969 /*
 970  * Take a path out of use.
 971  */
 972 static int fail_path(struct pgpath *pgpath)
 973 {
 974         unsigned long flags;
 975         struct multipath *m = pgpath->pg->m;
 976
 977         spin_lock_irqsave(&m->lock, flags);
 978
 979         if (!pgpath->is_active)
 980                 goto out;
 981
 982         DMWARN("Failing path %s.", pgpath->path.dev->name);
 983
 984         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 985         pgpath->is_active = 0;
 986         pgpath->fail_count++;
 987
 988         m->nr_valid_paths--;
 989
 990         if (pgpath == m->current_pgpath)
 991                 m->current_pgpath = NULL;
 992
 993         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 994                       pgpath->path.dev->name, m->nr_valid_paths);
 995
 996         schedule_work(&m->trigger_event);
 997         queue_work(kmultipathd, &pgpath->deactivate_path);
 998
 999 out:
1000         spin_unlock_irqrestore(&m->lock, flags);
1001
1002         return 0;
1003 }
1004
1005 /*
1006  * Reinstate a previously-failed path
1007  */
1008 static int reinstate_path(struct pgpath *pgpath)
1009 {
1010         int r = 0;
1011         unsigned long flags;
1012         struct multipath *m = pgpath->pg->m;
1013
1014         spin_lock_irqsave(&m->lock, flags);
1015
1016         if (pgpath->is_active)
1017                 goto out;
1018
1019         if (!pgpath->pg->ps.type->reinstate_path) {
1020                 DMWARN("Reinstate path not supported by path selector %s",
1021                        pgpath->pg->ps.type->name);
1022                 r = -EINVAL;
1023                 goto out;
1024         }
1025
1026         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1027         if (r)
1028                 goto out;
1029
1030         pgpath->is_active = 1;
1031
1032         if (!m->nr_valid_paths++ && m->queue_size) {
1033                 m->current_pgpath = NULL;
1034                 queue_work(kmultipathd, &m->process_queued_ios);
1035         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1036                 if (queue_work(kmpath_handlerd, &pgpath->activate_path))
1037                         m->pg_init_in_progress++;
1038         }
1039
1040         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1041                       pgpath->path.dev->name, m->nr_valid_paths);
1042
1043         schedule_work(&m->trigger_event);
1044
1045 out:
1046         spin_unlock_irqrestore(&m->lock, flags);
1047
1048         return r;
1049 }
1050
1051 /*
1052  * Fail or reinstate all paths that match the provided struct dm_dev.
1053  */
1054 static int action_dev(struct multipath *m, struct dm_dev *dev,
1055                       action_fn action)
1056 {
1057         int r = 0;
1058         struct pgpath *pgpath;
1059         struct priority_group *pg;
1060
1061         list_for_each_entry(pg, &m->priority_groups, list) {
1062                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1063                         if (pgpath->path.dev == dev)
1064                                 r = action(pgpath);
1065                 }
1066         }
1067
1068         return r;
1069 }
1070
1071 /*
1072  * Temporarily try to avoid having to use the specified PG
1073  */
1074 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1075                       int bypassed)
1076 {
1077         unsigned long flags;
1078
1079         spin_lock_irqsave(&m->lock, flags);
1080
1081         pg->bypassed = bypassed;
1082         m->current_pgpath = NULL;
1083         m->current_pg = NULL;
1084
1085         spin_unlock_irqrestore(&m->lock, flags);
1086
1087         schedule_work(&m->trigger_event);
1088 }
1089
1090 /*
1091  * Switch to using the specified PG from the next I/O that gets mapped
1092  */
1093 static int switch_pg_num(struct multipath *m, const char *pgstr)
1094 {
1095         struct priority_group *pg;
1096         unsigned pgnum;
1097         unsigned long flags;
1098
1099         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1100             (pgnum > m->nr_priority_groups)) {
1101                 DMWARN("invalid PG number supplied to switch_pg_num");
1102                 return -EINVAL;
1103         }
1104
1105         spin_lock_irqsave(&m->lock, flags);
1106         list_for_each_entry(pg, &m->priority_groups, list) {
1107                 pg->bypassed = 0;
1108                 if (--pgnum)
1109                         continue;
1110
1111                 m->current_pgpath = NULL;
1112                 m->current_pg = NULL;
1113                 m->next_pg = pg;
1114         }
1115         spin_unlock_irqrestore(&m->lock, flags);
1116
1117         schedule_work(&m->trigger_event);
1118         return 0;
1119 }
1120
1121 /*
1122  * Set/clear bypassed status of a PG.
1123  * PGs are numbered upwards from 1 in the order they were declared.
1124  */
1125 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1126 {
1127         struct priority_group *pg;
1128         unsigned pgnum;
1129
1130         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1131             (pgnum > m->nr_priority_groups)) {
1132                 DMWARN("invalid PG number supplied to bypass_pg");
1133                 return -EINVAL;
1134         }
1135
1136         list_for_each_entry(pg, &m->priority_groups, list) {
1137                 if (!--pgnum)
1138                         break;
1139         }
1140
1141         bypass_pg(m, pg, bypassed);
1142         return 0;
1143 }
1144
1145 /*
1146  * Should we retry pg_init immediately?
1147  */
1148 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1149 {
1150         unsigned long flags;
1151         int limit_reached = 0;
1152
1153         spin_lock_irqsave(&m->lock, flags);
1154
1155         if (m->pg_init_count <= m->pg_init_retries)
1156                 m->pg_init_required = 1;
1157         else
1158                 limit_reached = 1;
1159
1160         spin_unlock_irqrestore(&m->lock, flags);
1161
1162         return limit_reached;
1163 }
1164
1165 static void pg_init_done(void *data, int errors)
1166 {
1167         struct pgpath *pgpath = data;
1168         struct priority_group *pg = pgpath->pg;
1169         struct multipath *m = pg->m;
1170         unsigned long flags;
1171
1172         /* device or driver problems */
1173         switch (errors) {
1174         case SCSI_DH_OK:
1175                 break;
1176         case SCSI_DH_NOSYS:
1177                 if (!m->hw_handler_name) {
1178                         errors = 0;
1179                         break;
1180                 }
1181                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1182                       "Error %d.", m->hw_handler_name, errors);
1183                 /*
1184                  * Fail path for now, so we do not ping pong
1185                  */
1186                 fail_path(pgpath);
1187                 break;
1188         case SCSI_DH_DEV_TEMP_BUSY:
1189                 /*
1190                  * Probably doing something like FW upgrade on the
1191                  * controller so try the other pg.
1192                  */
1193                 bypass_pg(m, pg, 1);
1194                 break;
1195         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1196         case SCSI_DH_RETRY:
1197         case SCSI_DH_IMM_RETRY:
1198         case SCSI_DH_RES_TEMP_UNAVAIL:
1199                 if (pg_init_limit_reached(m, pgpath))
1200                         fail_path(pgpath);
1201                 errors = 0;
1202                 break;
1203         default:
1204                 /*
1205                  * We probably do not want to fail the path for a device
1206                  * error, but this is what the old dm did. In future
1207                  * patches we can do more advanced handling.
1208                  */
1209                 fail_path(pgpath);
1210         }
1211
1212         spin_lock_irqsave(&m->lock, flags);
1213         if (errors) {
1214                 if (pgpath == m->current_pgpath) {
1215                         DMERR("Could not failover device. Error %d.", errors);
1216                         m->current_pgpath = NULL;
1217                         m->current_pg = NULL;
1218                 }
1219         } else if (!m->pg_init_required)
1220                 pg->bypassed = 0;
1221
1222         if (--m->pg_init_in_progress)
1223                 /* Activations of other paths are still on going */
1224                 goto out;
1225
1226         if (!m->pg_init_required)
1227                 m->queue_io = 0;
1228
1229         queue_work(kmultipathd, &m->process_queued_ios);
1230
1231         /*
1232          * Wake up any thread waiting to suspend.
1233          */
1234         wake_up(&m->pg_init_wait);
1235
1236 out:
1237         spin_unlock_irqrestore(&m->lock, flags);
1238 }
1239
1240 static void activate_path(struct work_struct *work)
1241 {
1242         struct pgpath *pgpath =
1243                 container_of(work, struct pgpath, activate_path);
1244
1245         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1246                                 pg_init_done, pgpath);
1247 }
1248
1249 /*
1250  * end_io handling
1251  */
1252 static int do_end_io(struct multipath *m, struct request *clone,
1253                      int error, struct dm_mpath_io *mpio)
1254 {
1255         /*
1256          * We don't queue any clone request inside the multipath target
1257          * during end I/O handling, since those clone requests don't have
1258          * bio clones.  If we queue them inside the multipath target,
1259          * we need to make bio clones, that requires memory allocation.
1260          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1261          *  don't have bio clones.)
1262          * Instead of queueing the clone request here, we queue the original
1263          * request into dm core, which will remake a clone request and
1264          * clone bios for it and resubmit it later.
1265          */
1266         int r = DM_ENDIO_REQUEUE;
1267         unsigned long flags;
1268
1269         if (!error && !clone->errors)
1270                 return 0;       /* I/O complete */
1271
1272         if (error == -EOPNOTSUPP)
1273                 return error;
1274
1275         if (mpio->pgpath)
1276                 fail_path(mpio->pgpath);
1277
1278         spin_lock_irqsave(&m->lock, flags);
1279         if (!m->nr_valid_paths && !m->queue_if_no_path && !__must_push_back(m))
1280                 r = -EIO;
1281         spin_unlock_irqrestore(&m->lock, flags);
1282
1283         return r;
1284 }
1285
1286 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1287                             int error, union map_info *map_context)
1288 {
1289         struct multipath *m = ti->private;
1290         struct dm_mpath_io *mpio = map_context->ptr;
1291         struct pgpath *pgpath = mpio->pgpath;
1292         struct path_selector *ps;
1293         int r;
1294
1295         r  = do_end_io(m, clone, error, mpio);
1296         if (pgpath) {
1297                 ps = &pgpath->pg->ps;
1298                 if (ps->type->end_io)
1299                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1300         }
1301         mempool_free(mpio, m->mpio_pool);
1302
1303         return r;
1304 }
1305
1306 /*
1307  * Suspend can't complete until all the I/O is processed so if
1308  * the last path fails we must error any remaining I/O.
1309  * Note that if the freeze_bdev fails while suspending, the
1310  * queue_if_no_path state is lost - userspace should reset it.
1311  */
1312 static void multipath_presuspend(struct dm_target *ti)
1313 {
1314         struct multipath *m = (struct multipath *) ti->private;
1315
1316         queue_if_no_path(m, 0, 1);
1317 }
1318
1319 static void multipath_postsuspend(struct dm_target *ti)
1320 {
1321         struct multipath *m = ti->private;
1322
1323         mutex_lock(&m->work_mutex);
1324         flush_multipath_work(m);
1325         mutex_unlock(&m->work_mutex);
1326 }
1327
1328 /*
1329  * Restore the queue_if_no_path setting.
1330  */
1331 static void multipath_resume(struct dm_target *ti)
1332 {
1333         struct multipath *m = (struct multipath *) ti->private;
1334         unsigned long flags;
1335
1336         spin_lock_irqsave(&m->lock, flags);
1337         m->queue_if_no_path = m->saved_queue_if_no_path;
1338         spin_unlock_irqrestore(&m->lock, flags);
1339 }
1340
1341 /*
1342  * Info output has the following format:
1343  * num_multipath_feature_args [multipath_feature_args]*
1344  * num_handler_status_args [handler_status_args]*
1345  * num_groups init_group_number
1346  *            [A|D|E num_ps_status_args [ps_status_args]*
1347  *             num_paths num_selector_args
1348  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1349  *
1350  * Table output has the following format (identical to the constructor string):
1351  * num_feature_args [features_args]*
1352  * num_handler_args hw_handler [hw_handler_args]*
1353  * num_groups init_group_number
1354  *     [priority selector-name num_ps_args [ps_args]*
1355  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1356  */
1357 static int multipath_status(struct dm_target *ti, status_type_t type,
1358                             char *result, unsigned int maxlen)
1359 {
1360         int sz = 0;
1361         unsigned long flags;
1362         struct multipath *m = (struct multipath *) ti->private;
1363         struct priority_group *pg;
1364         struct pgpath *p;
1365         unsigned pg_num;
1366         char state;
1367
1368         spin_lock_irqsave(&m->lock, flags);
1369
1370         /* Features */
1371         if (type == STATUSTYPE_INFO)
1372                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1373         else {
1374                 DMEMIT("%u ", m->queue_if_no_path +
1375                               (m->pg_init_retries > 0) * 2);
1376                 if (m->queue_if_no_path)
1377                         DMEMIT("queue_if_no_path ");
1378                 if (m->pg_init_retries)
1379                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1380         }
1381
1382         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1383                 DMEMIT("0 ");
1384         else
1385                 DMEMIT("1 %s ", m->hw_handler_name);
1386
1387         DMEMIT("%u ", m->nr_priority_groups);
1388
1389         if (m->next_pg)
1390                 pg_num = m->next_pg->pg_num;
1391         else if (m->current_pg)
1392                 pg_num = m->current_pg->pg_num;
1393         else
1394                         pg_num = 1;
1395
1396         DMEMIT("%u ", pg_num);
1397
1398         switch (type) {
1399         case STATUSTYPE_INFO:
1400                 list_for_each_entry(pg, &m->priority_groups, list) {
1401                         if (pg->bypassed)
1402                                 state = 'D';    /* Disabled */
1403                         else if (pg == m->current_pg)
1404                                 state = 'A';    /* Currently Active */
1405                         else
1406                                 state = 'E';    /* Enabled */
1407
1408                         DMEMIT("%c ", state);
1409
1410                         if (pg->ps.type->status)
1411                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1412                                                           result + sz,
1413                                                           maxlen - sz);
1414                         else
1415                                 DMEMIT("0 ");
1416
1417                         DMEMIT("%u %u ", pg->nr_pgpaths,
1418                                pg->ps.type->info_args);
1419
1420                         list_for_each_entry(p, &pg->pgpaths, list) {
1421                                 DMEMIT("%s %s %u ", p->path.dev->name,
1422                                        p->is_active ? "A" : "F",
1423                                        p->fail_count);
1424                                 if (pg->ps.type->status)
1425                                         sz += pg->ps.type->status(&pg->ps,
1426                                               &p->path, type, result + sz,
1427                                               maxlen - sz);
1428                         }
1429                 }
1430                 break;
1431
1432         case STATUSTYPE_TABLE:
1433                 list_for_each_entry(pg, &m->priority_groups, list) {
1434                         DMEMIT("%s ", pg->ps.type->name);
1435
1436                         if (pg->ps.type->status)
1437                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1438                                                           result + sz,
1439                                                           maxlen - sz);
1440                         else
1441                                 DMEMIT("0 ");
1442
1443                         DMEMIT("%u %u ", pg->nr_pgpaths,
1444                                pg->ps.type->table_args);
1445
1446                         list_for_each_entry(p, &pg->pgpaths, list) {
1447                                 DMEMIT("%s ", p->path.dev->name);
1448                                 if (pg->ps.type->status)
1449                                         sz += pg->ps.type->status(&pg->ps,
1450                                               &p->path, type, result + sz,
1451                                               maxlen - sz);
1452                         }
1453                 }
1454                 break;
1455         }
1456
1457         spin_unlock_irqrestore(&m->lock, flags);
1458
1459         return 0;
1460 }
1461
1462 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1463 {
1464         int r = -EINVAL;
1465         struct dm_dev *dev;
1466         struct multipath *m = (struct multipath *) ti->private;
1467         action_fn action;
1468
1469         mutex_lock(&m->work_mutex);
1470
1471         if (dm_suspended(ti)) {
1472                 r = -EBUSY;
1473                 goto out;
1474         }
1475
1476         if (argc == 1) {
1477                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
1478                         r = queue_if_no_path(m, 1, 0);
1479                         goto out;
1480                 } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
1481                         r = queue_if_no_path(m, 0, 0);
1482                         goto out;
1483                 }
1484         }
1485
1486         if (argc != 2) {
1487                 DMWARN("Unrecognised multipath message received.");
1488                 goto out;
1489         }
1490
1491         if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
1492                 r = bypass_pg_num(m, argv[1], 1);
1493                 goto out;
1494         } else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
1495                 r = bypass_pg_num(m, argv[1], 0);
1496                 goto out;
1497         } else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
1498                 r = switch_pg_num(m, argv[1]);
1499                 goto out;
1500         } else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1501                 action = reinstate_path;
1502         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1503                 action = fail_path;
1504         else {
1505                 DMWARN("Unrecognised multipath message received.");
1506                 goto out;
1507         }
1508
1509         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1510         if (r) {
1511                 DMWARN("message: error getting device %s",
1512                        argv[1]);
1513                 goto out;
1514         }
1515
1516         r = action_dev(m, dev, action);
1517
1518         dm_put_device(ti, dev);
1519
1520 out:
1521         mutex_unlock(&m->work_mutex);
1522         return r;
1523 }
1524
1525 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1526                            unsigned long arg)
1527 {
1528         struct multipath *m = (struct multipath *) ti->private;
1529         struct block_device *bdev = NULL;
1530         fmode_t mode = 0;
1531         unsigned long flags;
1532         int r = 0;
1533
1534         spin_lock_irqsave(&m->lock, flags);
1535
1536         if (!m->current_pgpath)
1537                 __choose_pgpath(m, 0);
1538
1539         if (m->current_pgpath) {
1540                 bdev = m->current_pgpath->path.dev->bdev;
1541                 mode = m->current_pgpath->path.dev->mode;
1542         }
1543
1544         if (m->queue_io)
1545                 r = -EAGAIN;
1546         else if (!bdev)
1547                 r = -EIO;
1548
1549         spin_unlock_irqrestore(&m->lock, flags);
1550
1551         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1552 }
1553
1554 static int multipath_iterate_devices(struct dm_target *ti,
1555                                      iterate_devices_callout_fn fn, void *data)
1556 {
1557         struct multipath *m = ti->private;
1558         struct priority_group *pg;
1559         struct pgpath *p;
1560         int ret = 0;
1561
1562         list_for_each_entry(pg, &m->priority_groups, list) {
1563                 list_for_each_entry(p, &pg->pgpaths, list) {
1564                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1565                         if (ret)
1566                                 goto out;
1567                 }
1568         }
1569
1570 out:
1571         return ret;
1572 }
1573
1574 static int __pgpath_busy(struct pgpath *pgpath)
1575 {
1576         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1577
1578         return dm_underlying_device_busy(q);
1579 }
1580
1581 /*
1582  * We return "busy", only when we can map I/Os but underlying devices
1583  * are busy (so even if we map I/Os now, the I/Os will wait on
1584  * the underlying queue).
1585  * In other words, if we want to kill I/Os or queue them inside us
1586  * due to map unavailability, we don't return "busy".  Otherwise,
1587  * dm core won't give us the I/Os and we can't do what we want.
1588  */
1589 static int multipath_busy(struct dm_target *ti)
1590 {
1591         int busy = 0, has_active = 0;
1592         struct multipath *m = ti->private;
1593         struct priority_group *pg;
1594         struct pgpath *pgpath;
1595         unsigned long flags;
1596
1597         spin_lock_irqsave(&m->lock, flags);
1598
1599         /* Guess which priority_group will be used at next mapping time */
1600         if (unlikely(!m->current_pgpath && m->next_pg))
1601                 pg = m->next_pg;
1602         else if (likely(m->current_pg))
1603                 pg = m->current_pg;
1604         else
1605                 /*
1606                  * We don't know which pg will be used at next mapping time.
1607                  * We don't call __choose_pgpath() here to avoid to trigger
1608                  * pg_init just by busy checking.
1609                  * So we don't know whether underlying devices we will be using
1610                  * at next mapping time are busy or not. Just try mapping.
1611                  */
1612                 goto out;
1613
1614         /*
1615          * If there is one non-busy active path at least, the path selector
1616          * will be able to select it. So we consider such a pg as not busy.
1617          */
1618         busy = 1;
1619         list_for_each_entry(pgpath, &pg->pgpaths, list)
1620                 if (pgpath->is_active) {
1621                         has_active = 1;
1622
1623                         if (!__pgpath_busy(pgpath)) {
1624                                 busy = 0;
1625                                 break;
1626                         }
1627                 }
1628
1629         if (!has_active)
1630                 /*
1631                  * No active path in this pg, so this pg won't be used and
1632                  * the current_pg will be changed at next mapping time.
1633                  * We need to try mapping to determine it.
1634                  */
1635                 busy = 0;
1636
1637 out:
1638         spin_unlock_irqrestore(&m->lock, flags);
1639
1640         return busy;
1641 }
1642
1643 /*-----------------------------------------------------------------
1644  * Module setup
1645  *---------------------------------------------------------------*/
1646 static struct target_type multipath_target = {
1647         .name = "multipath",
1648         .version = {1, 1, 1},
1649         .module = THIS_MODULE,
1650         .ctr = multipath_ctr,
1651         .dtr = multipath_dtr,
1652         .map_rq = multipath_map,
1653         .rq_end_io = multipath_end_io,
1654         .presuspend = multipath_presuspend,
1655         .postsuspend = multipath_postsuspend,
1656         .resume = multipath_resume,
1657         .status = multipath_status,
1658         .message = multipath_message,
1659         .ioctl  = multipath_ioctl,
1660         .iterate_devices = multipath_iterate_devices,
1661         .busy = multipath_busy,
1662 };
1663
1664 static int __init dm_multipath_init(void)
1665 {
1666         int r;
1667
1668         /* allocate a slab for the dm_ios */
1669         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1670         if (!_mpio_cache)
1671                 return -ENOMEM;
1672
1673         r = dm_register_target(&multipath_target);
1674         if (r < 0) {
1675                 DMERR("register failed %d", r);
1676                 kmem_cache_destroy(_mpio_cache);
1677                 return -EINVAL;
1678         }
1679
1680         kmultipathd = create_workqueue("kmpathd");
1681         if (!kmultipathd) {
1682                 DMERR("failed to create workqueue kmpathd");
1683                 dm_unregister_target(&multipath_target);
1684                 kmem_cache_destroy(_mpio_cache);
1685                 return -ENOMEM;
1686         }
1687
1688         /*
1689          * A separate workqueue is used to handle the device handlers
1690          * to avoid overloading existing workqueue. Overloading the
1691          * old workqueue would also create a bottleneck in the
1692          * path of the storage hardware device activation.
1693          */
1694         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1695         if (!kmpath_handlerd) {
1696                 DMERR("failed to create workqueue kmpath_handlerd");
1697                 destroy_workqueue(kmultipathd);
1698                 dm_unregister_target(&multipath_target);
1699                 kmem_cache_destroy(_mpio_cache);
1700                 return -ENOMEM;
1701         }
1702
1703         DMINFO("version %u.%u.%u loaded",
1704                multipath_target.version[0], multipath_target.version[1],
1705                multipath_target.version[2]);
1706
1707         return r;
1708 }
1709
1710 static void __exit dm_multipath_exit(void)
1711 {
1712         destroy_workqueue(kmpath_handlerd);
1713         destroy_workqueue(kmultipathd);
1714
1715         dm_unregister_target(&multipath_target);
1716         kmem_cache_destroy(_mpio_cache);
1717 }
1718
1719 module_init(dm_multipath_init);
1720 module_exit(dm_multipath_exit);
1721
1722 MODULE_DESCRIPTION(DM_NAME " multipath target");
1723 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1724 MODULE_LICENSE("GPL");