drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9
  10 #include "dm.h"
  11 #include "dm-path-selector.h"
  12 #include "dm-uevent.h"
  13
  14 #include <linux/ctype.h>
  15 #include <linux/init.h>
  16 #include <linux/mempool.h>
  17 #include <linux/module.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/slab.h>
  20 #include <linux/time.h>
  21 #include <linux/workqueue.h>
  22 #include <linux/delay.h>
  23 #include <scsi/scsi_dh.h>
  24 #include <linux/atomic.h>
  25
  26 #define DM_MSG_PREFIX "multipath"
  27 #define DM_PG_INIT_DELAY_MSECS 2000
  28 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  29
  30 /* Path properties */
  31 struct pgpath {
  32         struct list_head list;
  33
  34         struct priority_group *pg;      /* Owning PG */
  35         unsigned is_active;             /* Path status */
  36         unsigned fail_count;            /* Cumulative failure count */
  37
  38         struct dm_path path;
  39         struct delayed_work activate_path;
  40 };
  41
  42 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  43
  44 /*
  45  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  46  * Each has a path selector which controls which path gets used.
  47  */
  48 struct priority_group {
  49         struct list_head list;
  50
  51         struct multipath *m;            /* Owning multipath instance */
  52         struct path_selector ps;
  53
  54         unsigned pg_num;                /* Reference number */
  55         unsigned bypassed;              /* Temporarily bypass this PG? */
  56
  57         unsigned nr_pgpaths;            /* Number of paths in PG */
  58         struct list_head pgpaths;
  59 };
  60
  61 /* Multipath context */
  62 struct multipath {
  63         struct list_head list;
  64         struct dm_target *ti;
  65
  66         const char *hw_handler_name;
  67         char *hw_handler_params;
  68
  69         spinlock_t lock;
  70
  71         unsigned nr_priority_groups;
  72         struct list_head priority_groups;
  73
  74         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  75
  76         unsigned pg_init_required;      /* pg_init needs calling? */
  77         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  78         unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
  79
  80         unsigned nr_valid_paths;        /* Total number of usable paths */
  81         struct pgpath *current_pgpath;
  82         struct priority_group *current_pg;
  83         struct priority_group *next_pg; /* Switch to this PG if set */
  84         unsigned repeat_count;          /* I/Os left before calling PS again */
  85
  86         unsigned queue_io:1;            /* Must we queue all I/O? */
  87         unsigned queue_if_no_path:1;    /* Queue I/O if last path fails? */
  88         unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
  89         unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
  90         unsigned pg_init_disabled:1;    /* pg_init is not currently allowed */
  91
  92         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  93         unsigned pg_init_count;         /* Number of times pg_init called */
  94         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  95
  96         unsigned queue_size;
  97         struct work_struct process_queued_ios;
  98         struct list_head queued_ios;
  99
 100         struct work_struct trigger_event;
 101
 102         /*
 103          * We must use a mempool of dm_mpath_io structs so that we
 104          * can resubmit bios on error.
 105          */
 106         mempool_t *mpio_pool;
 107
 108         struct mutex work_mutex;
 109 };
 110
 111 /*
 112  * Context information attached to each bio we process.
 113  */
 114 struct dm_mpath_io {
 115         struct pgpath *pgpath;
 116         size_t nr_bytes;
 117 };
 118
 119 typedef int (*action_fn) (struct pgpath *pgpath);
 120
 121 static struct kmem_cache *_mpio_cache;
 122
 123 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 124 static void process_queued_ios(struct work_struct *work);
 125 static void trigger_event(struct work_struct *work);
 126 static void activate_path(struct work_struct *work);
 127
 128
 129 /*-----------------------------------------------
 130  * Allocation routines
 131  *-----------------------------------------------*/
 132
 133 static struct pgpath *alloc_pgpath(void)
 134 {
 135         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 136
 137         if (pgpath) {
 138                 pgpath->is_active = 1;
 139                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 140         }
 141
 142         return pgpath;
 143 }
 144
 145 static void free_pgpath(struct pgpath *pgpath)
 146 {
 147         kfree(pgpath);
 148 }
 149
 150 static struct priority_group *alloc_priority_group(void)
 151 {
 152         struct priority_group *pg;
 153
 154         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 155
 156         if (pg)
 157                 INIT_LIST_HEAD(&pg->pgpaths);
 158
 159         return pg;
 160 }
 161
 162 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 163 {
 164         struct pgpath *pgpath, *tmp;
 165         struct multipath *m = ti->private;
 166
 167         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 168                 list_del(&pgpath->list);
 169                 if (m->hw_handler_name)
 170                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 171                 dm_put_device(ti, pgpath->path.dev);
 172                 free_pgpath(pgpath);
 173         }
 174 }
 175
 176 static void free_priority_group(struct priority_group *pg,
 177                                 struct dm_target *ti)
 178 {
 179         struct path_selector *ps = &pg->ps;
 180
 181         if (ps->type) {
 182                 ps->type->destroy(ps);
 183                 dm_put_path_selector(ps->type);
 184         }
 185
 186         free_pgpaths(&pg->pgpaths, ti);
 187         kfree(pg);
 188 }
 189
 190 static struct multipath *alloc_multipath(struct dm_target *ti)
 191 {
 192         struct multipath *m;
 193         unsigned min_ios = dm_get_reserved_rq_based_ios();
 194
 195         m = kzalloc(sizeof(*m), GFP_KERNEL);
 196         if (m) {
 197                 INIT_LIST_HEAD(&m->priority_groups);
 198                 INIT_LIST_HEAD(&m->queued_ios);
 199                 spin_lock_init(&m->lock);
 200                 m->queue_io = 1;
 201                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 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 static int set_mapinfo(struct multipath *m, union map_info *info)
 234 {
 235         struct dm_mpath_io *mpio;
 236
 237         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 238         if (!mpio)
 239                 return -ENOMEM;
 240
 241         memset(mpio, 0, sizeof(*mpio));
 242         info->ptr = mpio;
 243
 244         return 0;
 245 }
 246
 247 static void clear_mapinfo(struct multipath *m, union map_info *info)
 248 {
 249         struct dm_mpath_io *mpio = info->ptr;
 250
 251         info->ptr = NULL;
 252         mempool_free(mpio, m->mpio_pool);
 253 }
 254
 255 /*-----------------------------------------------
 256  * Path selection
 257  *-----------------------------------------------*/
 258
 259 static void __pg_init_all_paths(struct multipath *m)
 260 {
 261         struct pgpath *pgpath;
 262         unsigned long pg_init_delay = 0;
 263
 264         m->pg_init_count++;
 265         m->pg_init_required = 0;
 266         if (m->pg_init_delay_retry)
 267                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 268                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 269         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 270                 /* Skip failed paths */
 271                 if (!pgpath->is_active)
 272                         continue;
 273                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 274                                        pg_init_delay))
 275                         m->pg_init_in_progress++;
 276         }
 277 }
 278
 279 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 280 {
 281         m->current_pg = pgpath->pg;
 282
 283         /* Must we initialise the PG first, and queue I/O till it's ready? */
 284         if (m->hw_handler_name) {
 285                 m->pg_init_required = 1;
 286                 m->queue_io = 1;
 287         } else {
 288                 m->pg_init_required = 0;
 289                 m->queue_io = 0;
 290         }
 291
 292         m->pg_init_count = 0;
 293 }
 294
 295 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 296                                size_t nr_bytes)
 297 {
 298         struct dm_path *path;
 299
 300         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 301         if (!path)
 302                 return -ENXIO;
 303
 304         m->current_pgpath = path_to_pgpath(path);
 305
 306         if (m->current_pg != pg)
 307                 __switch_pg(m, m->current_pgpath);
 308
 309         return 0;
 310 }
 311
 312 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 313 {
 314         struct priority_group *pg;
 315         unsigned bypassed = 1;
 316
 317         if (!m->nr_valid_paths)
 318                 goto failed;
 319
 320         /* Were we instructed to switch PG? */
 321         if (m->next_pg) {
 322                 pg = m->next_pg;
 323                 m->next_pg = NULL;
 324                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 325                         return;
 326         }
 327
 328         /* Don't change PG until it has no remaining paths */
 329         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 330                 return;
 331
 332         /*
 333          * Loop through priority groups until we find a valid path.
 334          * First time we skip PGs marked 'bypassed'.
 335          * Second time we only try the ones we skipped, but set
 336          * pg_init_delay_retry so we do not hammer controllers.
 337          */
 338         do {
 339                 list_for_each_entry(pg, &m->priority_groups, list) {
 340                         if (pg->bypassed == bypassed)
 341                                 continue;
 342                         if (!__choose_path_in_pg(m, pg, nr_bytes)) {
 343                                 if (!bypassed)
 344                                         m->pg_init_delay_retry = 1;
 345                                 return;
 346                         }
 347                 }
 348         } while (bypassed--);
 349
 350 failed:
 351         m->current_pgpath = NULL;
 352         m->current_pg = NULL;
 353 }
 354
 355 /*
 356  * Check whether bios must be queued in the device-mapper core rather
 357  * than here in the target.
 358  *
 359  * m->lock must be held on entry.
 360  *
 361  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 362  * same value then we are not between multipath_presuspend()
 363  * and multipath_resume() calls and we have no need to check
 364  * for the DMF_NOFLUSH_SUSPENDING flag.
 365  */
 366 static int __must_push_back(struct multipath *m)
 367 {
 368         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 369                 dm_noflush_suspending(m->ti));
 370 }
 371
 372 static int map_io(struct multipath *m, struct request *clone,
 373                   union map_info *map_context, unsigned was_queued)
 374 {
 375         int r = DM_MAPIO_REMAPPED;
 376         size_t nr_bytes = blk_rq_bytes(clone);
 377         unsigned long flags;
 378         struct pgpath *pgpath;
 379         struct block_device *bdev;
 380         struct dm_mpath_io *mpio = map_context->ptr;
 381
 382         spin_lock_irqsave(&m->lock, flags);
 383
 384         /* Do we need to select a new pgpath? */
 385         if (!m->current_pgpath ||
 386             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 387                 __choose_pgpath(m, nr_bytes);
 388
 389         pgpath = m->current_pgpath;
 390
 391         if (was_queued)
 392                 m->queue_size--;
 393
 394         if (m->pg_init_required) {
 395                 if (!m->pg_init_in_progress)
 396                         queue_work(kmultipathd, &m->process_queued_ios);
 397                 r = DM_MAPIO_REQUEUE;
 398         } else if ((pgpath && m->queue_io) ||
 399                    (!pgpath && m->queue_if_no_path)) {
 400                 /* Queue for the daemon to resubmit */
 401                 list_add_tail(&clone->queuelist, &m->queued_ios);
 402                 m->queue_size++;
 403                 if (!m->queue_io)
 404                         queue_work(kmultipathd, &m->process_queued_ios);
 405                 pgpath = NULL;
 406                 r = DM_MAPIO_SUBMITTED;
 407         } else if (pgpath) {
 408                 bdev = pgpath->path.dev->bdev;
 409                 clone->q = bdev_get_queue(bdev);
 410                 clone->rq_disk = bdev->bd_disk;
 411         } else if (__must_push_back(m))
 412                 r = DM_MAPIO_REQUEUE;
 413         else
 414                 r = -EIO;       /* Failed */
 415
 416         mpio->pgpath = pgpath;
 417         mpio->nr_bytes = nr_bytes;
 418
 419         if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 420                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 421                                               nr_bytes);
 422
 423         spin_unlock_irqrestore(&m->lock, flags);
 424
 425         return r;
 426 }
 427
 428 /*
 429  * If we run out of usable paths, should we queue I/O or error it?
 430  */
 431 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 432                             unsigned save_old_value)
 433 {
 434         unsigned long flags;
 435
 436         spin_lock_irqsave(&m->lock, flags);
 437
 438         if (save_old_value)
 439                 m->saved_queue_if_no_path = m->queue_if_no_path;
 440         else
 441                 m->saved_queue_if_no_path = queue_if_no_path;
 442         m->queue_if_no_path = queue_if_no_path;
 443         if (!m->queue_if_no_path && m->queue_size)
 444                 queue_work(kmultipathd, &m->process_queued_ios);
 445
 446         spin_unlock_irqrestore(&m->lock, flags);
 447
 448         return 0;
 449 }
 450
 451 /*-----------------------------------------------------------------
 452  * The multipath daemon is responsible for resubmitting queued ios.
 453  *---------------------------------------------------------------*/
 454
 455 static void dispatch_queued_ios(struct multipath *m)
 456 {
 457         int r;
 458         unsigned long flags;
 459         union map_info *info;
 460         struct request *clone, *n;
 461         LIST_HEAD(cl);
 462
 463         spin_lock_irqsave(&m->lock, flags);
 464         list_splice_init(&m->queued_ios, &cl);
 465         spin_unlock_irqrestore(&m->lock, flags);
 466
 467         list_for_each_entry_safe(clone, n, &cl, queuelist) {
 468                 list_del_init(&clone->queuelist);
 469
 470                 info = dm_get_rq_mapinfo(clone);
 471
 472                 r = map_io(m, clone, info, 1);
 473                 if (r < 0) {
 474                         clear_mapinfo(m, info);
 475                         dm_kill_unmapped_request(clone, r);
 476                 } else if (r == DM_MAPIO_REMAPPED)
 477                         dm_dispatch_request(clone);
 478                 else if (r == DM_MAPIO_REQUEUE) {
 479                         clear_mapinfo(m, info);
 480                         dm_requeue_unmapped_request(clone);
 481                 }
 482         }
 483 }
 484
 485 static void process_queued_ios(struct work_struct *work)
 486 {
 487         struct multipath *m =
 488                 container_of(work, struct multipath, process_queued_ios);
 489         struct pgpath *pgpath = NULL;
 490         unsigned must_queue = 1;
 491         unsigned long flags;
 492
 493         spin_lock_irqsave(&m->lock, flags);
 494
 495         if (!m->current_pgpath)
 496                 __choose_pgpath(m, 0);
 497
 498         pgpath = m->current_pgpath;
 499
 500         if ((pgpath && !m->queue_io) ||
 501             (!pgpath && !m->queue_if_no_path))
 502                 must_queue = 0;
 503
 504         if (m->pg_init_required && !m->pg_init_in_progress && pgpath &&
 505             !m->pg_init_disabled)
 506                 __pg_init_all_paths(m);
 507
 508         spin_unlock_irqrestore(&m->lock, flags);
 509         if (!must_queue)
 510                 dispatch_queued_ios(m);
 511 }
 512
 513 /*
 514  * An event is triggered whenever a path is taken out of use.
 515  * Includes path failure and PG bypass.
 516  */
 517 static void trigger_event(struct work_struct *work)
 518 {
 519         struct multipath *m =
 520                 container_of(work, struct multipath, trigger_event);
 521
 522         dm_table_event(m->ti->table);
 523 }
 524
 525 /*-----------------------------------------------------------------
 526  * Constructor/argument parsing:
 527  * <#multipath feature args> [<arg>]*
 528  * <#hw_handler args> [hw_handler [<arg>]*]
 529  * <#priority groups>
 530  * <initial priority group>
 531  *     [<selector> <#selector args> [<arg>]*
 532  *      <#paths> <#per-path selector args>
 533  *         [<path> [<arg>]* ]+ ]+
 534  *---------------------------------------------------------------*/
 535 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 536                                struct dm_target *ti)
 537 {
 538         int r;
 539         struct path_selector_type *pst;
 540         unsigned ps_argc;
 541
 542         static struct dm_arg _args[] = {
 543                 {0, 1024, "invalid number of path selector args"},
 544         };
 545
 546         pst = dm_get_path_selector(dm_shift_arg(as));
 547         if (!pst) {
 548                 ti->error = "unknown path selector type";
 549                 return -EINVAL;
 550         }
 551
 552         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 553         if (r) {
 554                 dm_put_path_selector(pst);
 555                 return -EINVAL;
 556         }
 557
 558         r = pst->create(&pg->ps, ps_argc, as->argv);
 559         if (r) {
 560                 dm_put_path_selector(pst);
 561                 ti->error = "path selector constructor failed";
 562                 return r;
 563         }
 564
 565         pg->ps.type = pst;
 566         dm_consume_args(as, ps_argc);
 567
 568         return 0;
 569 }
 570
 571 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 572                                struct dm_target *ti)
 573 {
 574         int r;
 575         struct pgpath *p;
 576         struct multipath *m = ti->private;
 577         struct request_queue *q = NULL;
 578         const char *attached_handler_name;
 579
 580         /* we need at least a path arg */
 581         if (as->argc < 1) {
 582                 ti->error = "no device given";
 583                 return ERR_PTR(-EINVAL);
 584         }
 585
 586         p = alloc_pgpath();
 587         if (!p)
 588                 return ERR_PTR(-ENOMEM);
 589
 590         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 591                           &p->path.dev);
 592         if (r) {
 593                 ti->error = "error getting device";
 594                 goto bad;
 595         }
 596
 597         if (m->retain_attached_hw_handler || m->hw_handler_name)
 598                 q = bdev_get_queue(p->path.dev->bdev);
 599
 600         if (m->retain_attached_hw_handler) {
 601                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 602                 if (attached_handler_name) {
 603                         /*
 604                          * Reset hw_handler_name to match the attached handler
 605                          * and clear any hw_handler_params associated with the
 606                          * ignored handler.
 607                          *
 608                          * NB. This modifies the table line to show the actual
 609                          * handler instead of the original table passed in.
 610                          */
 611                         kfree(m->hw_handler_name);
 612                         m->hw_handler_name = attached_handler_name;
 613
 614                         kfree(m->hw_handler_params);
 615                         m->hw_handler_params = NULL;
 616                 }
 617         }
 618
 619         if (m->hw_handler_name) {
 620                 /*
 621                  * Increments scsi_dh reference, even when using an
 622                  * already-attached handler.
 623                  */
 624                 r = scsi_dh_attach(q, m->hw_handler_name);
 625                 if (r == -EBUSY) {
 626                         /*
 627                          * Already attached to different hw_handler:
 628                          * try to reattach with correct one.
 629                          */
 630                         scsi_dh_detach(q);
 631                         r = scsi_dh_attach(q, m->hw_handler_name);
 632                 }
 633
 634                 if (r < 0) {
 635                         ti->error = "error attaching hardware handler";
 636                         dm_put_device(ti, p->path.dev);
 637                         goto bad;
 638                 }
 639
 640                 if (m->hw_handler_params) {
 641                         r = scsi_dh_set_params(q, m->hw_handler_params);
 642                         if (r < 0) {
 643                                 ti->error = "unable to set hardware "
 644                                                         "handler parameters";
 645                                 scsi_dh_detach(q);
 646                                 dm_put_device(ti, p->path.dev);
 647                                 goto bad;
 648                         }
 649                 }
 650         }
 651
 652         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 653         if (r) {
 654                 dm_put_device(ti, p->path.dev);
 655                 goto bad;
 656         }
 657
 658         return p;
 659
 660  bad:
 661         free_pgpath(p);
 662         return ERR_PTR(r);
 663 }
 664
 665 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 666                                                    struct multipath *m)
 667 {
 668         static struct dm_arg _args[] = {
 669                 {1, 1024, "invalid number of paths"},
 670                 {0, 1024, "invalid number of selector args"}
 671         };
 672
 673         int r;
 674         unsigned i, nr_selector_args, nr_args;
 675         struct priority_group *pg;
 676         struct dm_target *ti = m->ti;
 677
 678         if (as->argc < 2) {
 679                 as->argc = 0;
 680                 ti->error = "not enough priority group arguments";
 681                 return ERR_PTR(-EINVAL);
 682         }
 683
 684         pg = alloc_priority_group();
 685         if (!pg) {
 686                 ti->error = "couldn't allocate priority group";
 687                 return ERR_PTR(-ENOMEM);
 688         }
 689         pg->m = m;
 690
 691         r = parse_path_selector(as, pg, ti);
 692         if (r)
 693                 goto bad;
 694
 695         /*
 696          * read the paths
 697          */
 698         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 699         if (r)
 700                 goto bad;
 701
 702         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 703         if (r)
 704                 goto bad;
 705
 706         nr_args = 1 + nr_selector_args;
 707         for (i = 0; i < pg->nr_pgpaths; i++) {
 708                 struct pgpath *pgpath;
 709                 struct dm_arg_set path_args;
 710
 711                 if (as->argc < nr_args) {
 712                         ti->error = "not enough path parameters";
 713                         r = -EINVAL;
 714                         goto bad;
 715                 }
 716
 717                 path_args.argc = nr_args;
 718                 path_args.argv = as->argv;
 719
 720                 pgpath = parse_path(&path_args, &pg->ps, ti);
 721                 if (IS_ERR(pgpath)) {
 722                         r = PTR_ERR(pgpath);
 723                         goto bad;
 724                 }
 725
 726                 pgpath->pg = pg;
 727                 list_add_tail(&pgpath->list, &pg->pgpaths);
 728                 dm_consume_args(as, nr_args);
 729         }
 730
 731         return pg;
 732
 733  bad:
 734         free_priority_group(pg, ti);
 735         return ERR_PTR(r);
 736 }
 737
 738 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 739 {
 740         unsigned hw_argc;
 741         int ret;
 742         struct dm_target *ti = m->ti;
 743
 744         static struct dm_arg _args[] = {
 745                 {0, 1024, "invalid number of hardware handler args"},
 746         };
 747
 748         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 749                 return -EINVAL;
 750
 751         if (!hw_argc)
 752                 return 0;
 753
 754         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 755         if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
 756                                      "scsi_dh_%s", m->hw_handler_name)) {
 757                 ti->error = "unknown hardware handler type";
 758                 ret = -EINVAL;
 759                 goto fail;
 760         }
 761
 762         if (hw_argc > 1) {
 763                 char *p;
 764                 int i, j, len = 4;
 765
 766                 for (i = 0; i <= hw_argc - 2; i++)
 767                         len += strlen(as->argv[i]) + 1;
 768                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 769                 if (!p) {
 770                         ti->error = "memory allocation failed";
 771                         ret = -ENOMEM;
 772                         goto fail;
 773                 }
 774                 j = sprintf(p, "%d", hw_argc - 1);
 775                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 776                         j = sprintf(p, "%s", as->argv[i]);
 777         }
 778         dm_consume_args(as, hw_argc - 1);
 779
 780         return 0;
 781 fail:
 782         kfree(m->hw_handler_name);
 783         m->hw_handler_name = NULL;
 784         return ret;
 785 }
 786
 787 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 788 {
 789         int r;
 790         unsigned argc;
 791         struct dm_target *ti = m->ti;
 792         const char *arg_name;
 793
 794         static struct dm_arg _args[] = {
 795                 {0, 6, "invalid number of feature args"},
 796                 {1, 50, "pg_init_retries must be between 1 and 50"},
 797                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 798         };
 799
 800         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 801         if (r)
 802                 return -EINVAL;
 803
 804         if (!argc)
 805                 return 0;
 806
 807         do {
 808                 arg_name = dm_shift_arg(as);
 809                 argc--;
 810
 811                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 812                         r = queue_if_no_path(m, 1, 0);
 813                         continue;
 814                 }
 815
 816                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
 817                         m->retain_attached_hw_handler = 1;
 818                         continue;
 819                 }
 820
 821                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 822                     (argc >= 1)) {
 823                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 824                         argc--;
 825                         continue;
 826                 }
 827
 828                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 829                     (argc >= 1)) {
 830                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 831                         argc--;
 832                         continue;
 833                 }
 834
 835                 ti->error = "Unrecognised multipath feature request";
 836                 r = -EINVAL;
 837         } while (argc && !r);
 838
 839         return r;
 840 }
 841
 842 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 843                          char **argv)
 844 {
 845         /* target arguments */
 846         static struct dm_arg _args[] = {
 847                 {0, 1024, "invalid number of priority groups"},
 848                 {0, 1024, "invalid initial priority group number"},
 849         };
 850
 851         int r;
 852         struct multipath *m;
 853         struct dm_arg_set as;
 854         unsigned pg_count = 0;
 855         unsigned next_pg_num;
 856
 857         as.argc = argc;
 858         as.argv = argv;
 859
 860         m = alloc_multipath(ti);
 861         if (!m) {
 862                 ti->error = "can't allocate multipath";
 863                 return -EINVAL;
 864         }
 865
 866         r = parse_features(&as, m);
 867         if (r)
 868                 goto bad;
 869
 870         r = parse_hw_handler(&as, m);
 871         if (r)
 872                 goto bad;
 873
 874         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 875         if (r)
 876                 goto bad;
 877
 878         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 879         if (r)
 880                 goto bad;
 881
 882         if ((!m->nr_priority_groups && next_pg_num) ||
 883             (m->nr_priority_groups && !next_pg_num)) {
 884                 ti->error = "invalid initial priority group";
 885                 r = -EINVAL;
 886                 goto bad;
 887         }
 888
 889         /* parse the priority groups */
 890         while (as.argc) {
 891                 struct priority_group *pg;
 892
 893                 pg = parse_priority_group(&as, m);
 894                 if (IS_ERR(pg)) {
 895                         r = PTR_ERR(pg);
 896                         goto bad;
 897                 }
 898
 899                 m->nr_valid_paths += pg->nr_pgpaths;
 900                 list_add_tail(&pg->list, &m->priority_groups);
 901                 pg_count++;
 902                 pg->pg_num = pg_count;
 903                 if (!--next_pg_num)
 904                         m->next_pg = pg;
 905         }
 906
 907         if (pg_count != m->nr_priority_groups) {
 908                 ti->error = "priority group count mismatch";
 909                 r = -EINVAL;
 910                 goto bad;
 911         }
 912
 913         ti->num_flush_bios = 1;
 914         ti->num_discard_bios = 1;
 915         ti->num_write_same_bios = 1;
 916
 917         return 0;
 918
 919  bad:
 920         free_multipath(m);
 921         return r;
 922 }
 923
 924 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 925 {
 926         DECLARE_WAITQUEUE(wait, current);
 927         unsigned long flags;
 928
 929         add_wait_queue(&m->pg_init_wait, &wait);
 930
 931         while (1) {
 932                 set_current_state(TASK_UNINTERRUPTIBLE);
 933
 934                 spin_lock_irqsave(&m->lock, flags);
 935                 if (!m->pg_init_in_progress) {
 936                         spin_unlock_irqrestore(&m->lock, flags);
 937                         break;
 938                 }
 939                 spin_unlock_irqrestore(&m->lock, flags);
 940
 941                 io_schedule();
 942         }
 943         set_current_state(TASK_RUNNING);
 944
 945         remove_wait_queue(&m->pg_init_wait, &wait);
 946 }
 947
 948 static void flush_multipath_work(struct multipath *m)
 949 {
 950         unsigned long flags;
 951
 952         spin_lock_irqsave(&m->lock, flags);
 953         m->pg_init_disabled = 1;
 954         spin_unlock_irqrestore(&m->lock, flags);
 955
 956         flush_workqueue(kmpath_handlerd);
 957         multipath_wait_for_pg_init_completion(m);
 958         flush_workqueue(kmultipathd);
 959         flush_work(&m->trigger_event);
 960
 961         spin_lock_irqsave(&m->lock, flags);
 962         m->pg_init_disabled = 0;
 963         spin_unlock_irqrestore(&m->lock, flags);
 964 }
 965
 966 static void multipath_dtr(struct dm_target *ti)
 967 {
 968         struct multipath *m = ti->private;
 969
 970         flush_multipath_work(m);
 971         free_multipath(m);
 972 }
 973
 974 /*
 975  * Map cloned requests
 976  */
 977 static int multipath_map(struct dm_target *ti, struct request *clone,
 978                          union map_info *map_context)
 979 {
 980         int r;
 981         struct multipath *m = (struct multipath *) ti->private;
 982
 983         if (set_mapinfo(m, map_context) < 0)
 984                 /* ENOMEM, requeue */
 985                 return DM_MAPIO_REQUEUE;
 986
 987         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 988         r = map_io(m, clone, map_context, 0);
 989         if (r < 0 || r == DM_MAPIO_REQUEUE)
 990                 clear_mapinfo(m, map_context);
 991
 992         return r;
 993 }
 994
 995 /*
 996  * Take a path out of use.
 997  */
 998 static int fail_path(struct pgpath *pgpath)
 999 {
1000         unsigned long flags;
1001         struct multipath *m = pgpath->pg->m;
1002
1003         spin_lock_irqsave(&m->lock, flags);
1004
1005         if (!pgpath->is_active)
1006                 goto out;
1007
1008         DMWARN("Failing path %s.", pgpath->path.dev->name);
1009
1010         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1011         pgpath->is_active = 0;
1012         pgpath->fail_count++;
1013
1014         m->nr_valid_paths--;
1015
1016         if (pgpath == m->current_pgpath)
1017                 m->current_pgpath = NULL;
1018
1019         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1020                       pgpath->path.dev->name, m->nr_valid_paths);
1021
1022         schedule_work(&m->trigger_event);
1023
1024 out:
1025         spin_unlock_irqrestore(&m->lock, flags);
1026
1027         return 0;
1028 }
1029
1030 /*
1031  * Reinstate a previously-failed path
1032  */
1033 static int reinstate_path(struct pgpath *pgpath)
1034 {
1035         int r = 0;
1036         unsigned long flags;
1037         struct multipath *m = pgpath->pg->m;
1038
1039         spin_lock_irqsave(&m->lock, flags);
1040
1041         if (pgpath->is_active)
1042                 goto out;
1043
1044         if (!pgpath->pg->ps.type->reinstate_path) {
1045                 DMWARN("Reinstate path not supported by path selector %s",
1046                        pgpath->pg->ps.type->name);
1047                 r = -EINVAL;
1048                 goto out;
1049         }
1050
1051         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1052         if (r)
1053                 goto out;
1054
1055         pgpath->is_active = 1;
1056
1057         if (!m->nr_valid_paths++ && m->queue_size) {
1058                 m->current_pgpath = NULL;
1059                 queue_work(kmultipathd, &m->process_queued_ios);
1060         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1061                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1062                         m->pg_init_in_progress++;
1063         }
1064
1065         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1066                       pgpath->path.dev->name, m->nr_valid_paths);
1067
1068         schedule_work(&m->trigger_event);
1069
1070 out:
1071         spin_unlock_irqrestore(&m->lock, flags);
1072
1073         return r;
1074 }
1075
1076 /*
1077  * Fail or reinstate all paths that match the provided struct dm_dev.
1078  */
1079 static int action_dev(struct multipath *m, struct dm_dev *dev,
1080                       action_fn action)
1081 {
1082         int r = -EINVAL;
1083         struct pgpath *pgpath;
1084         struct priority_group *pg;
1085
1086         list_for_each_entry(pg, &m->priority_groups, list) {
1087                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1088                         if (pgpath->path.dev == dev)
1089                                 r = action(pgpath);
1090                 }
1091         }
1092
1093         return r;
1094 }
1095
1096 /*
1097  * Temporarily try to avoid having to use the specified PG
1098  */
1099 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1100                       int bypassed)
1101 {
1102         unsigned long flags;
1103
1104         spin_lock_irqsave(&m->lock, flags);
1105
1106         pg->bypassed = bypassed;
1107         m->current_pgpath = NULL;
1108         m->current_pg = NULL;
1109
1110         spin_unlock_irqrestore(&m->lock, flags);
1111
1112         schedule_work(&m->trigger_event);
1113 }
1114
1115 /*
1116  * Switch to using the specified PG from the next I/O that gets mapped
1117  */
1118 static int switch_pg_num(struct multipath *m, const char *pgstr)
1119 {
1120         struct priority_group *pg;
1121         unsigned pgnum;
1122         unsigned long flags;
1123         char dummy;
1124
1125         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1126             (pgnum > m->nr_priority_groups)) {
1127                 DMWARN("invalid PG number supplied to switch_pg_num");
1128                 return -EINVAL;
1129         }
1130
1131         spin_lock_irqsave(&m->lock, flags);
1132         list_for_each_entry(pg, &m->priority_groups, list) {
1133                 pg->bypassed = 0;
1134                 if (--pgnum)
1135                         continue;
1136
1137                 m->current_pgpath = NULL;
1138                 m->current_pg = NULL;
1139                 m->next_pg = pg;
1140         }
1141         spin_unlock_irqrestore(&m->lock, flags);
1142
1143         schedule_work(&m->trigger_event);
1144         return 0;
1145 }
1146
1147 /*
1148  * Set/clear bypassed status of a PG.
1149  * PGs are numbered upwards from 1 in the order they were declared.
1150  */
1151 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1152 {
1153         struct priority_group *pg;
1154         unsigned pgnum;
1155         char dummy;
1156
1157         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1158             (pgnum > m->nr_priority_groups)) {
1159                 DMWARN("invalid PG number supplied to bypass_pg");
1160                 return -EINVAL;
1161         }
1162
1163         list_for_each_entry(pg, &m->priority_groups, list) {
1164                 if (!--pgnum)
1165                         break;
1166         }
1167
1168         bypass_pg(m, pg, bypassed);
1169         return 0;
1170 }
1171
1172 /*
1173  * Should we retry pg_init immediately?
1174  */
1175 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1176 {
1177         unsigned long flags;
1178         int limit_reached = 0;
1179
1180         spin_lock_irqsave(&m->lock, flags);
1181
1182         if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1183                 m->pg_init_required = 1;
1184         else
1185                 limit_reached = 1;
1186
1187         spin_unlock_irqrestore(&m->lock, flags);
1188
1189         return limit_reached;
1190 }
1191
1192 static void pg_init_done(void *data, int errors)
1193 {
1194         struct pgpath *pgpath = data;
1195         struct priority_group *pg = pgpath->pg;
1196         struct multipath *m = pg->m;
1197         unsigned long flags;
1198         unsigned delay_retry = 0;
1199
1200         /* device or driver problems */
1201         switch (errors) {
1202         case SCSI_DH_OK:
1203                 break;
1204         case SCSI_DH_NOSYS:
1205                 if (!m->hw_handler_name) {
1206                         errors = 0;
1207                         break;
1208                 }
1209                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1210                       "Error %d.", m->hw_handler_name, errors);
1211                 /*
1212                  * Fail path for now, so we do not ping pong
1213                  */
1214                 fail_path(pgpath);
1215                 break;
1216         case SCSI_DH_DEV_TEMP_BUSY:
1217                 /*
1218                  * Probably doing something like FW upgrade on the
1219                  * controller so try the other pg.
1220                  */
1221                 bypass_pg(m, pg, 1);
1222                 break;
1223         case SCSI_DH_RETRY:
1224                 /* Wait before retrying. */
1225                 delay_retry = 1;
1226         case SCSI_DH_IMM_RETRY:
1227         case SCSI_DH_RES_TEMP_UNAVAIL:
1228                 if (pg_init_limit_reached(m, pgpath))
1229                         fail_path(pgpath);
1230                 errors = 0;
1231                 break;
1232         default:
1233                 /*
1234                  * We probably do not want to fail the path for a device
1235                  * error, but this is what the old dm did. In future
1236                  * patches we can do more advanced handling.
1237                  */
1238                 fail_path(pgpath);
1239         }
1240
1241         spin_lock_irqsave(&m->lock, flags);
1242         if (errors) {
1243                 if (pgpath == m->current_pgpath) {
1244                         DMERR("Could not failover device. Error %d.", errors);
1245                         m->current_pgpath = NULL;
1246                         m->current_pg = NULL;
1247                 }
1248         } else if (!m->pg_init_required)
1249                 pg->bypassed = 0;
1250
1251         if (--m->pg_init_in_progress)
1252                 /* Activations of other paths are still on going */
1253                 goto out;
1254
1255         if (!m->pg_init_required)
1256                 m->queue_io = 0;
1257
1258         m->pg_init_delay_retry = delay_retry;
1259         queue_work(kmultipathd, &m->process_queued_ios);
1260
1261         /*
1262          * Wake up any thread waiting to suspend.
1263          */
1264         wake_up(&m->pg_init_wait);
1265
1266 out:
1267         spin_unlock_irqrestore(&m->lock, flags);
1268 }
1269
1270 static void activate_path(struct work_struct *work)
1271 {
1272         struct pgpath *pgpath =
1273                 container_of(work, struct pgpath, activate_path.work);
1274
1275         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1276                                 pg_init_done, pgpath);
1277 }
1278
1279 static int noretry_error(int error)
1280 {
1281         switch (error) {
1282         case -EOPNOTSUPP:
1283         case -EREMOTEIO:
1284         case -EILSEQ:
1285         case -ENODATA:
1286         case -ENOSPC:
1287                 return 1;
1288         }
1289
1290         /* Anything else could be a path failure, so should be retried */
1291         return 0;
1292 }
1293
1294 /*
1295  * end_io handling
1296  */
1297 static int do_end_io(struct multipath *m, struct request *clone,
1298                      int error, struct dm_mpath_io *mpio)
1299 {
1300         /*
1301          * We don't queue any clone request inside the multipath target
1302          * during end I/O handling, since those clone requests don't have
1303          * bio clones.  If we queue them inside the multipath target,
1304          * we need to make bio clones, that requires memory allocation.
1305          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1306          *  don't have bio clones.)
1307          * Instead of queueing the clone request here, we queue the original
1308          * request into dm core, which will remake a clone request and
1309          * clone bios for it and resubmit it later.
1310          */
1311         int r = DM_ENDIO_REQUEUE;
1312         unsigned long flags;
1313
1314         if (!error && !clone->errors)
1315                 return 0;       /* I/O complete */
1316
1317         if (noretry_error(error)) {
1318                 if ((clone->cmd_flags & REQ_WRITE_SAME) &&
1319                     !clone->q->limits.max_write_same_sectors) {
1320                         struct queue_limits *limits;
1321
1322                         /* device doesn't really support WRITE SAME, disable it */
1323                         limits = dm_get_queue_limits(dm_table_get_md(m->ti->table));
1324                         limits->max_write_same_sectors = 0;
1325                 }
1326                 return error;
1327         }
1328
1329         if (mpio->pgpath)
1330                 fail_path(mpio->pgpath);
1331
1332         spin_lock_irqsave(&m->lock, flags);
1333         if (!m->nr_valid_paths) {
1334                 if (!m->queue_if_no_path) {
1335                         if (!__must_push_back(m))
1336                                 r = -EIO;
1337                 } else {
1338                         if (error == -EBADE)
1339                                 r = error;
1340                 }
1341         }
1342         spin_unlock_irqrestore(&m->lock, flags);
1343
1344         return r;
1345 }
1346
1347 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1348                             int error, union map_info *map_context)
1349 {
1350         struct multipath *m = ti->private;
1351         struct dm_mpath_io *mpio = map_context->ptr;
1352         struct pgpath *pgpath;
1353         struct path_selector *ps;
1354         int r;
1355
1356         BUG_ON(!mpio);
1357
1358         r  = do_end_io(m, clone, error, mpio);
1359         pgpath = mpio->pgpath;
1360         if (pgpath) {
1361                 ps = &pgpath->pg->ps;
1362                 if (ps->type->end_io)
1363                         ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
1364         }
1365         clear_mapinfo(m, map_context);
1366
1367         return r;
1368 }
1369
1370 /*
1371  * Suspend can't complete until all the I/O is processed so if
1372  * the last path fails we must error any remaining I/O.
1373  * Note that if the freeze_bdev fails while suspending, the
1374  * queue_if_no_path state is lost - userspace should reset it.
1375  */
1376 static void multipath_presuspend(struct dm_target *ti)
1377 {
1378         struct multipath *m = (struct multipath *) ti->private;
1379
1380         queue_if_no_path(m, 0, 1);
1381 }
1382
1383 static void multipath_postsuspend(struct dm_target *ti)
1384 {
1385         struct multipath *m = ti->private;
1386
1387         mutex_lock(&m->work_mutex);
1388         flush_multipath_work(m);
1389         mutex_unlock(&m->work_mutex);
1390 }
1391
1392 /*
1393  * Restore the queue_if_no_path setting.
1394  */
1395 static void multipath_resume(struct dm_target *ti)
1396 {
1397         struct multipath *m = (struct multipath *) ti->private;
1398         unsigned long flags;
1399
1400         spin_lock_irqsave(&m->lock, flags);
1401         m->queue_if_no_path = m->saved_queue_if_no_path;
1402         spin_unlock_irqrestore(&m->lock, flags);
1403 }
1404
1405 /*
1406  * Info output has the following format:
1407  * num_multipath_feature_args [multipath_feature_args]*
1408  * num_handler_status_args [handler_status_args]*
1409  * num_groups init_group_number
1410  *            [A|D|E num_ps_status_args [ps_status_args]*
1411  *             num_paths num_selector_args
1412  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1413  *
1414  * Table output has the following format (identical to the constructor string):
1415  * num_feature_args [features_args]*
1416  * num_handler_args hw_handler [hw_handler_args]*
1417  * num_groups init_group_number
1418  *     [priority selector-name num_ps_args [ps_args]*
1419  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1420  */
1421 static void multipath_status(struct dm_target *ti, status_type_t type,
1422                              unsigned status_flags, char *result, unsigned maxlen)
1423 {
1424         int sz = 0;
1425         unsigned long flags;
1426         struct multipath *m = (struct multipath *) ti->private;
1427         struct priority_group *pg;
1428         struct pgpath *p;
1429         unsigned pg_num;
1430         char state;
1431
1432         spin_lock_irqsave(&m->lock, flags);
1433
1434         /* Features */
1435         if (type == STATUSTYPE_INFO)
1436                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1437         else {
1438                 DMEMIT("%u ", m->queue_if_no_path +
1439                               (m->pg_init_retries > 0) * 2 +
1440                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1441                               m->retain_attached_hw_handler);
1442                 if (m->queue_if_no_path)
1443                         DMEMIT("queue_if_no_path ");
1444                 if (m->pg_init_retries)
1445                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1446                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1447                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1448                 if (m->retain_attached_hw_handler)
1449                         DMEMIT("retain_attached_hw_handler ");
1450         }
1451
1452         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1453                 DMEMIT("0 ");
1454         else
1455                 DMEMIT("1 %s ", m->hw_handler_name);
1456
1457         DMEMIT("%u ", m->nr_priority_groups);
1458
1459         if (m->next_pg)
1460                 pg_num = m->next_pg->pg_num;
1461         else if (m->current_pg)
1462                 pg_num = m->current_pg->pg_num;
1463         else
1464                 pg_num = (m->nr_priority_groups ? 1 : 0);
1465
1466         DMEMIT("%u ", pg_num);
1467
1468         switch (type) {
1469         case STATUSTYPE_INFO:
1470                 list_for_each_entry(pg, &m->priority_groups, list) {
1471                         if (pg->bypassed)
1472                                 state = 'D';    /* Disabled */
1473                         else if (pg == m->current_pg)
1474                                 state = 'A';    /* Currently Active */
1475                         else
1476                                 state = 'E';    /* Enabled */
1477
1478                         DMEMIT("%c ", state);
1479
1480                         if (pg->ps.type->status)
1481                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1482                                                           result + sz,
1483                                                           maxlen - sz);
1484                         else
1485                                 DMEMIT("0 ");
1486
1487                         DMEMIT("%u %u ", pg->nr_pgpaths,
1488                                pg->ps.type->info_args);
1489
1490                         list_for_each_entry(p, &pg->pgpaths, list) {
1491                                 DMEMIT("%s %s %u ", p->path.dev->name,
1492                                        p->is_active ? "A" : "F",
1493                                        p->fail_count);
1494                                 if (pg->ps.type->status)
1495                                         sz += pg->ps.type->status(&pg->ps,
1496                                               &p->path, type, result + sz,
1497                                               maxlen - sz);
1498                         }
1499                 }
1500                 break;
1501
1502         case STATUSTYPE_TABLE:
1503                 list_for_each_entry(pg, &m->priority_groups, list) {
1504                         DMEMIT("%s ", pg->ps.type->name);
1505
1506                         if (pg->ps.type->status)
1507                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1508                                                           result + sz,
1509                                                           maxlen - sz);
1510                         else
1511                                 DMEMIT("0 ");
1512
1513                         DMEMIT("%u %u ", pg->nr_pgpaths,
1514                                pg->ps.type->table_args);
1515
1516                         list_for_each_entry(p, &pg->pgpaths, list) {
1517                                 DMEMIT("%s ", p->path.dev->name);
1518                                 if (pg->ps.type->status)
1519                                         sz += pg->ps.type->status(&pg->ps,
1520                                               &p->path, type, result + sz,
1521                                               maxlen - sz);
1522                         }
1523                 }
1524                 break;
1525         }
1526
1527         spin_unlock_irqrestore(&m->lock, flags);
1528 }
1529
1530 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1531 {
1532         int r = -EINVAL;
1533         struct dm_dev *dev;
1534         struct multipath *m = (struct multipath *) ti->private;
1535         action_fn action;
1536
1537         mutex_lock(&m->work_mutex);
1538
1539         if (dm_suspended(ti)) {
1540                 r = -EBUSY;
1541                 goto out;
1542         }
1543
1544         if (argc == 1) {
1545                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1546                         r = queue_if_no_path(m, 1, 0);
1547                         goto out;
1548                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1549                         r = queue_if_no_path(m, 0, 0);
1550                         goto out;
1551                 }
1552         }
1553
1554         if (argc != 2) {
1555                 DMWARN("Unrecognised multipath message received.");
1556                 goto out;
1557         }
1558
1559         if (!strcasecmp(argv[0], "disable_group")) {
1560                 r = bypass_pg_num(m, argv[1], 1);
1561                 goto out;
1562         } else if (!strcasecmp(argv[0], "enable_group")) {
1563                 r = bypass_pg_num(m, argv[1], 0);
1564                 goto out;
1565         } else if (!strcasecmp(argv[0], "switch_group")) {
1566                 r = switch_pg_num(m, argv[1]);
1567                 goto out;
1568         } else if (!strcasecmp(argv[0], "reinstate_path"))
1569                 action = reinstate_path;
1570         else if (!strcasecmp(argv[0], "fail_path"))
1571                 action = fail_path;
1572         else {
1573                 DMWARN("Unrecognised multipath message received.");
1574                 goto out;
1575         }
1576
1577         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1578         if (r) {
1579                 DMWARN("message: error getting device %s",
1580                        argv[1]);
1581                 goto out;
1582         }
1583
1584         r = action_dev(m, dev, action);
1585
1586         dm_put_device(ti, dev);
1587
1588 out:
1589         mutex_unlock(&m->work_mutex);
1590         return r;
1591 }
1592
1593 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1594                            unsigned long arg)
1595 {
1596         struct multipath *m = ti->private;
1597         struct pgpath *pgpath;
1598         struct block_device *bdev;
1599         fmode_t mode;
1600         unsigned long flags;
1601         int r;
1602
1603         bdev = NULL;
1604         mode = 0;
1605         r = 0;
1606
1607         spin_lock_irqsave(&m->lock, flags);
1608
1609         if (!m->current_pgpath)
1610                 __choose_pgpath(m, 0);
1611
1612         pgpath = m->current_pgpath;
1613
1614         if (pgpath) {
1615                 bdev = pgpath->path.dev->bdev;
1616                 mode = pgpath->path.dev->mode;
1617         }
1618
1619         if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1620                 r = -ENOTCONN;
1621         else if (!bdev)
1622                 r = -EIO;
1623
1624         spin_unlock_irqrestore(&m->lock, flags);
1625
1626         /*
1627          * Only pass ioctls through if the device sizes match exactly.
1628          */
1629         if (!bdev || ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) {
1630                 int err = scsi_verify_blk_ioctl(NULL, cmd);
1631                 if (err)
1632                         r = err;
1633         }
1634
1635         if (r == -ENOTCONN && !fatal_signal_pending(current))
1636                 queue_work(kmultipathd, &m->process_queued_ios);
1637
1638         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1639 }
1640
1641 static int multipath_iterate_devices(struct dm_target *ti,
1642                                      iterate_devices_callout_fn fn, void *data)
1643 {
1644         struct multipath *m = ti->private;
1645         struct priority_group *pg;
1646         struct pgpath *p;
1647         int ret = 0;
1648
1649         list_for_each_entry(pg, &m->priority_groups, list) {
1650                 list_for_each_entry(p, &pg->pgpaths, list) {
1651                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1652                         if (ret)
1653                                 goto out;
1654                 }
1655         }
1656
1657 out:
1658         return ret;
1659 }
1660
1661 static int __pgpath_busy(struct pgpath *pgpath)
1662 {
1663         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1664
1665         return dm_underlying_device_busy(q);
1666 }
1667
1668 /*
1669  * We return "busy", only when we can map I/Os but underlying devices
1670  * are busy (so even if we map I/Os now, the I/Os will wait on
1671  * the underlying queue).
1672  * In other words, if we want to kill I/Os or queue them inside us
1673  * due to map unavailability, we don't return "busy".  Otherwise,
1674  * dm core won't give us the I/Os and we can't do what we want.
1675  */
1676 static int multipath_busy(struct dm_target *ti)
1677 {
1678         int busy = 0, has_active = 0;
1679         struct multipath *m = ti->private;
1680         struct priority_group *pg;
1681         struct pgpath *pgpath;
1682         unsigned long flags;
1683
1684         spin_lock_irqsave(&m->lock, flags);
1685
1686         /* pg_init in progress, requeue until done */
1687         if (m->pg_init_in_progress) {
1688                 busy = 1;
1689                 goto out;
1690         }
1691         /* Guess which priority_group will be used at next mapping time */
1692         if (unlikely(!m->current_pgpath && m->next_pg))
1693                 pg = m->next_pg;
1694         else if (likely(m->current_pg))
1695                 pg = m->current_pg;
1696         else
1697                 /*
1698                  * We don't know which pg will be used at next mapping time.
1699                  * We don't call __choose_pgpath() here to avoid to trigger
1700                  * pg_init just by busy checking.
1701                  * So we don't know whether underlying devices we will be using
1702                  * at next mapping time are busy or not. Just try mapping.
1703                  */
1704                 goto out;
1705
1706         /*
1707          * If there is one non-busy active path at least, the path selector
1708          * will be able to select it. So we consider such a pg as not busy.
1709          */
1710         busy = 1;
1711         list_for_each_entry(pgpath, &pg->pgpaths, list)
1712                 if (pgpath->is_active) {
1713                         has_active = 1;
1714
1715                         if (!__pgpath_busy(pgpath)) {
1716                                 busy = 0;
1717                                 break;
1718                         }
1719                 }
1720
1721         if (!has_active)
1722                 /*
1723                  * No active path in this pg, so this pg won't be used and
1724                  * the current_pg will be changed at next mapping time.
1725                  * We need to try mapping to determine it.
1726                  */
1727                 busy = 0;
1728
1729 out:
1730         spin_unlock_irqrestore(&m->lock, flags);
1731
1732         return busy;
1733 }
1734
1735 /*-----------------------------------------------------------------
1736  * Module setup
1737  *---------------------------------------------------------------*/
1738 static struct target_type multipath_target = {
1739         .name = "multipath",
1740         .version = {1, 6, 0},
1741         .module = THIS_MODULE,
1742         .ctr = multipath_ctr,
1743         .dtr = multipath_dtr,
1744         .map_rq = multipath_map,
1745         .rq_end_io = multipath_end_io,
1746         .presuspend = multipath_presuspend,
1747         .postsuspend = multipath_postsuspend,
1748         .resume = multipath_resume,
1749         .status = multipath_status,
1750         .message = multipath_message,
1751         .ioctl  = multipath_ioctl,
1752         .iterate_devices = multipath_iterate_devices,
1753         .busy = multipath_busy,
1754 };
1755
1756 static int __init dm_multipath_init(void)
1757 {
1758         int r;
1759
1760         /* allocate a slab for the dm_ios */
1761         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1762         if (!_mpio_cache)
1763                 return -ENOMEM;
1764
1765         r = dm_register_target(&multipath_target);
1766         if (r < 0) {
1767                 DMERR("register failed %d", r);
1768                 kmem_cache_destroy(_mpio_cache);
1769                 return -EINVAL;
1770         }
1771
1772         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1773         if (!kmultipathd) {
1774                 DMERR("failed to create workqueue kmpathd");
1775                 dm_unregister_target(&multipath_target);
1776                 kmem_cache_destroy(_mpio_cache);
1777                 return -ENOMEM;
1778         }
1779
1780         /*
1781          * A separate workqueue is used to handle the device handlers
1782          * to avoid overloading existing workqueue. Overloading the
1783          * old workqueue would also create a bottleneck in the
1784          * path of the storage hardware device activation.
1785          */
1786         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1787                                                   WQ_MEM_RECLAIM);
1788         if (!kmpath_handlerd) {
1789                 DMERR("failed to create workqueue kmpath_handlerd");
1790                 destroy_workqueue(kmultipathd);
1791                 dm_unregister_target(&multipath_target);
1792                 kmem_cache_destroy(_mpio_cache);
1793                 return -ENOMEM;
1794         }
1795
1796         DMINFO("version %u.%u.%u loaded",
1797                multipath_target.version[0], multipath_target.version[1],
1798                multipath_target.version[2]);
1799
1800         return r;
1801 }
1802
1803 static void __exit dm_multipath_exit(void)
1804 {
1805         destroy_workqueue(kmpath_handlerd);
1806         destroy_workqueue(kmultipathd);
1807
1808         dm_unregister_target(&multipath_target);
1809         kmem_cache_destroy(_mpio_cache);
1810 }
1811
1812 module_init(dm_multipath_init);
1813 module_exit(dm_multipath_exit);
1814
1815 MODULE_DESCRIPTION(DM_NAME " multipath target");
1816 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1817 MODULE_LICENSE("GPL");