drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/device-mapper.h>
   9
  10 #include "dm.h"
  11 #include "dm-path-selector.h"
  12 #include "dm-uevent.h"
  13
  14 #include <linux/ctype.h>
  15 #include <linux/init.h>
  16 #include <linux/mempool.h>
  17 #include <linux/module.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/slab.h>
  20 #include <linux/time.h>
  21 #include <linux/workqueue.h>
  22 #include <linux/delay.h>
  23 #include <scsi/scsi_dh.h>
  24 #include <linux/atomic.h>
  25
  26 #define DM_MSG_PREFIX "multipath"
  27 #define DM_PG_INIT_DELAY_MSECS 2000
  28 #define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  29
  30 /* Path properties */
  31 struct pgpath {
  32         struct list_head list;
  33
  34         struct priority_group *pg;      /* Owning PG */
  35         unsigned is_active;             /* Path status */
  36         unsigned fail_count;            /* Cumulative failure count */
  37
  38         struct dm_path path;
  39         struct delayed_work activate_path;
  40 };
  41
  42 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  43
  44 /*
  45  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  46  * Each has a path selector which controls which path gets used.
  47  */
  48 struct priority_group {
  49         struct list_head list;
  50
  51         struct multipath *m;            /* Owning multipath instance */
  52         struct path_selector ps;
  53
  54         unsigned pg_num;                /* Reference number */
  55         unsigned bypassed;              /* Temporarily bypass this PG? */
  56
  57         unsigned nr_pgpaths;            /* Number of paths in PG */
  58         struct list_head pgpaths;
  59 };
  60
  61 /* Multipath context */
  62 struct multipath {
  63         struct list_head list;
  64         struct dm_target *ti;
  65
  66         const char *hw_handler_name;
  67         char *hw_handler_params;
  68
  69         spinlock_t lock;
  70
  71         unsigned nr_priority_groups;
  72         struct list_head priority_groups;
  73
  74         wait_queue_head_t pg_init_wait; /* Wait for pg_init completion */
  75
  76         unsigned pg_init_required;      /* pg_init needs calling? */
  77         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  78         unsigned pg_init_delay_retry;   /* Delay pg_init retry? */
  79
  80         unsigned nr_valid_paths;        /* Total number of usable paths */
  81         struct pgpath *current_pgpath;
  82         struct priority_group *current_pg;
  83         struct priority_group *next_pg; /* Switch to this PG if set */
  84         unsigned repeat_count;          /* I/Os left before calling PS again */
  85
  86         unsigned queue_io:1;            /* Must we queue all I/O? */
  87         unsigned queue_if_no_path:1;    /* Queue I/O if last path fails? */
  88         unsigned saved_queue_if_no_path:1; /* Saved state during suspension */
  89         unsigned retain_attached_hw_handler:1; /* If there's already a hw_handler present, don't change it. */
  90         unsigned pg_init_disabled:1;    /* pg_init is not currently allowed */
  91
  92         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  93         unsigned pg_init_count;         /* Number of times pg_init called */
  94         unsigned pg_init_delay_msecs;   /* Number of msecs before pg_init retry */
  95
  96         struct work_struct trigger_event;
  97
  98         /*
  99          * We must use a mempool of dm_mpath_io structs so that we
 100          * can resubmit bios on error.
 101          */
 102         mempool_t *mpio_pool;
 103
 104         struct mutex work_mutex;
 105 };
 106
 107 /*
 108  * Context information attached to each bio we process.
 109  */
 110 struct dm_mpath_io {
 111         struct pgpath *pgpath;
 112         size_t nr_bytes;
 113 };
 114
 115 typedef int (*action_fn) (struct pgpath *pgpath);
 116
 117 static struct kmem_cache *_mpio_cache;
 118
 119 static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 120 static void trigger_event(struct work_struct *work);
 121 static void activate_path(struct work_struct *work);
 122 static int __pgpath_busy(struct pgpath *pgpath);
 123
 124
 125 /*-----------------------------------------------
 126  * Allocation routines
 127  *-----------------------------------------------*/
 128
 129 static struct pgpath *alloc_pgpath(void)
 130 {
 131         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 132
 133         if (pgpath) {
 134                 pgpath->is_active = 1;
 135                 INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 136         }
 137
 138         return pgpath;
 139 }
 140
 141 static void free_pgpath(struct pgpath *pgpath)
 142 {
 143         kfree(pgpath);
 144 }
 145
 146 static struct priority_group *alloc_priority_group(void)
 147 {
 148         struct priority_group *pg;
 149
 150         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 151
 152         if (pg)
 153                 INIT_LIST_HEAD(&pg->pgpaths);
 154
 155         return pg;
 156 }
 157
 158 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 159 {
 160         struct pgpath *pgpath, *tmp;
 161         struct multipath *m = ti->private;
 162
 163         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 164                 list_del(&pgpath->list);
 165                 if (m->hw_handler_name)
 166                         scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 167                 dm_put_device(ti, pgpath->path.dev);
 168                 free_pgpath(pgpath);
 169         }
 170 }
 171
 172 static void free_priority_group(struct priority_group *pg,
 173                                 struct dm_target *ti)
 174 {
 175         struct path_selector *ps = &pg->ps;
 176
 177         if (ps->type) {
 178                 ps->type->destroy(ps);
 179                 dm_put_path_selector(ps->type);
 180         }
 181
 182         free_pgpaths(&pg->pgpaths, ti);
 183         kfree(pg);
 184 }
 185
 186 static struct multipath *alloc_multipath(struct dm_target *ti)
 187 {
 188         struct multipath *m;
 189         unsigned min_ios = dm_get_reserved_rq_based_ios();
 190
 191         m = kzalloc(sizeof(*m), GFP_KERNEL);
 192         if (m) {
 193                 INIT_LIST_HEAD(&m->priority_groups);
 194                 spin_lock_init(&m->lock);
 195                 m->queue_io = 1;
 196                 m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 197                 INIT_WORK(&m->trigger_event, trigger_event);
 198                 init_waitqueue_head(&m->pg_init_wait);
 199                 mutex_init(&m->work_mutex);
 200                 m->mpio_pool = mempool_create_slab_pool(min_ios, _mpio_cache);
 201                 if (!m->mpio_pool) {
 202                         kfree(m);
 203                         return NULL;
 204                 }
 205                 m->ti = ti;
 206                 ti->private = m;
 207         }
 208
 209         return m;
 210 }
 211
 212 static void free_multipath(struct multipath *m)
 213 {
 214         struct priority_group *pg, *tmp;
 215
 216         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 217                 list_del(&pg->list);
 218                 free_priority_group(pg, m->ti);
 219         }
 220
 221         kfree(m->hw_handler_name);
 222         kfree(m->hw_handler_params);
 223         mempool_destroy(m->mpio_pool);
 224         kfree(m);
 225 }
 226
 227 static int set_mapinfo(struct multipath *m, union map_info *info)
 228 {
 229         struct dm_mpath_io *mpio;
 230
 231         mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 232         if (!mpio)
 233                 return -ENOMEM;
 234
 235         memset(mpio, 0, sizeof(*mpio));
 236         info->ptr = mpio;
 237
 238         return 0;
 239 }
 240
 241 static void clear_mapinfo(struct multipath *m, union map_info *info)
 242 {
 243         struct dm_mpath_io *mpio = info->ptr;
 244
 245         info->ptr = NULL;
 246         mempool_free(mpio, m->mpio_pool);
 247 }
 248
 249 /*-----------------------------------------------
 250  * Path selection
 251  *-----------------------------------------------*/
 252
 253 static int __pg_init_all_paths(struct multipath *m)
 254 {
 255         struct pgpath *pgpath;
 256         unsigned long pg_init_delay = 0;
 257
 258         if (m->pg_init_in_progress || m->pg_init_disabled)
 259                 return 0;
 260
 261         m->pg_init_count++;
 262         m->pg_init_required = 0;
 263
 264         /* Check here to reset pg_init_required */
 265         if (!m->current_pg)
 266                 return 0;
 267
 268         if (m->pg_init_delay_retry)
 269                 pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 270                                                  m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 271         list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 272                 /* Skip failed paths */
 273                 if (!pgpath->is_active)
 274                         continue;
 275                 if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 276                                        pg_init_delay))
 277                         m->pg_init_in_progress++;
 278         }
 279         return m->pg_init_in_progress;
 280 }
 281
 282 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 283 {
 284         m->current_pg = pgpath->pg;
 285
 286         /* Must we initialise the PG first, and queue I/O till it's ready? */
 287         if (m->hw_handler_name) {
 288                 m->pg_init_required = 1;
 289                 m->queue_io = 1;
 290         } else {
 291                 m->pg_init_required = 0;
 292                 m->queue_io = 0;
 293         }
 294
 295         m->pg_init_count = 0;
 296 }
 297
 298 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 299                                size_t nr_bytes)
 300 {
 301         struct dm_path *path;
 302
 303         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 304         if (!path)
 305                 return -ENXIO;
 306
 307         m->current_pgpath = path_to_pgpath(path);
 308
 309         if (m->current_pg != pg)
 310                 __switch_pg(m, m->current_pgpath);
 311
 312         return 0;
 313 }
 314
 315 static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 316 {
 317         struct priority_group *pg;
 318         unsigned bypassed = 1;
 319
 320         if (!m->nr_valid_paths)
 321                 goto failed;
 322
 323         /* Were we instructed to switch PG? */
 324         if (m->next_pg) {
 325                 pg = m->next_pg;
 326                 m->next_pg = NULL;
 327                 if (!__choose_path_in_pg(m, pg, nr_bytes))
 328                         return;
 329         }
 330
 331         /* Don't change PG until it has no remaining paths */
 332         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 333                 return;
 334
 335         /*
 336          * Loop through priority groups until we find a valid path.
 337          * First time we skip PGs marked 'bypassed'.
 338          * Second time we only try the ones we skipped, but set
 339          * pg_init_delay_retry so we do not hammer controllers.
 340          */
 341         do {
 342                 list_for_each_entry(pg, &m->priority_groups, list) {
 343                         if (pg->bypassed == bypassed)
 344                                 continue;
 345                         if (!__choose_path_in_pg(m, pg, nr_bytes)) {
 346                                 if (!bypassed)
 347                                         m->pg_init_delay_retry = 1;
 348                                 return;
 349                         }
 350                 }
 351         } while (bypassed--);
 352
 353 failed:
 354         m->current_pgpath = NULL;
 355         m->current_pg = NULL;
 356 }
 357
 358 /*
 359  * Check whether bios must be queued in the device-mapper core rather
 360  * than here in the target.
 361  *
 362  * m->lock must be held on entry.
 363  *
 364  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 365  * same value then we are not between multipath_presuspend()
 366  * and multipath_resume() calls and we have no need to check
 367  * for the DMF_NOFLUSH_SUSPENDING flag.
 368  */
 369 static int __must_push_back(struct multipath *m)
 370 {
 371         return (m->queue_if_no_path ||
 372                 (m->queue_if_no_path != m->saved_queue_if_no_path &&
 373                  dm_noflush_suspending(m->ti)));
 374 }
 375
 376 #define pg_ready(m) (!(m)->queue_io && !(m)->pg_init_required)
 377
 378 static int map_io(struct multipath *m, struct request *clone,
 379                   union map_info *map_context)
 380 {
 381         int r = DM_MAPIO_REMAPPED;
 382         size_t nr_bytes = blk_rq_bytes(clone);
 383         unsigned long flags;
 384         struct pgpath *pgpath;
 385         struct block_device *bdev;
 386         struct dm_mpath_io *mpio = map_context->ptr;
 387
 388         spin_lock_irqsave(&m->lock, flags);
 389
 390         /* Do we need to select a new pgpath? */
 391         if (!m->current_pgpath ||
 392             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 393                 __choose_pgpath(m, nr_bytes);
 394
 395         pgpath = m->current_pgpath;
 396
 397         if (pgpath) {
 398                 if (pg_ready(m)) {
 399                         bdev = pgpath->path.dev->bdev;
 400                         clone->q = bdev_get_queue(bdev);
 401                         clone->rq_disk = bdev->bd_disk;
 402                         mpio->pgpath = pgpath;
 403                         mpio->nr_bytes = nr_bytes;
 404                         if (pgpath->pg->ps.type->start_io)
 405                                 pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 406                                                               &pgpath->path,
 407                                                               nr_bytes);
 408                 } else {
 409                         __pg_init_all_paths(m);
 410                         r = DM_MAPIO_REQUEUE;
 411                 }
 412         } else {
 413                 /* No path */
 414                 if (__must_push_back(m))
 415                         r = DM_MAPIO_REQUEUE;
 416                 else
 417                         r = -EIO;       /* Failed */
 418         }
 419
 420         spin_unlock_irqrestore(&m->lock, flags);
 421
 422         return r;
 423 }
 424
 425 /*
 426  * If we run out of usable paths, should we queue I/O or error it?
 427  */
 428 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 429                             unsigned save_old_value)
 430 {
 431         unsigned long flags;
 432
 433         spin_lock_irqsave(&m->lock, flags);
 434
 435         if (save_old_value)
 436                 m->saved_queue_if_no_path = m->queue_if_no_path;
 437         else
 438                 m->saved_queue_if_no_path = queue_if_no_path;
 439         m->queue_if_no_path = queue_if_no_path;
 440         if (!m->queue_if_no_path)
 441                 dm_table_run_md_queue_async(m->ti->table);
 442
 443         spin_unlock_irqrestore(&m->lock, flags);
 444
 445         return 0;
 446 }
 447
 448 /*
 449  * An event is triggered whenever a path is taken out of use.
 450  * Includes path failure and PG bypass.
 451  */
 452 static void trigger_event(struct work_struct *work)
 453 {
 454         struct multipath *m =
 455                 container_of(work, struct multipath, trigger_event);
 456
 457         dm_table_event(m->ti->table);
 458 }
 459
 460 /*-----------------------------------------------------------------
 461  * Constructor/argument parsing:
 462  * <#multipath feature args> [<arg>]*
 463  * <#hw_handler args> [hw_handler [<arg>]*]
 464  * <#priority groups>
 465  * <initial priority group>
 466  *     [<selector> <#selector args> [<arg>]*
 467  *      <#paths> <#per-path selector args>
 468  *         [<path> [<arg>]* ]+ ]+
 469  *---------------------------------------------------------------*/
 470 static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 471                                struct dm_target *ti)
 472 {
 473         int r;
 474         struct path_selector_type *pst;
 475         unsigned ps_argc;
 476
 477         static struct dm_arg _args[] = {
 478                 {0, 1024, "invalid number of path selector args"},
 479         };
 480
 481         pst = dm_get_path_selector(dm_shift_arg(as));
 482         if (!pst) {
 483                 ti->error = "unknown path selector type";
 484                 return -EINVAL;
 485         }
 486
 487         r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 488         if (r) {
 489                 dm_put_path_selector(pst);
 490                 return -EINVAL;
 491         }
 492
 493         r = pst->create(&pg->ps, ps_argc, as->argv);
 494         if (r) {
 495                 dm_put_path_selector(pst);
 496                 ti->error = "path selector constructor failed";
 497                 return r;
 498         }
 499
 500         pg->ps.type = pst;
 501         dm_consume_args(as, ps_argc);
 502
 503         return 0;
 504 }
 505
 506 static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 507                                struct dm_target *ti)
 508 {
 509         int r;
 510         struct pgpath *p;
 511         struct multipath *m = ti->private;
 512         struct request_queue *q = NULL;
 513         const char *attached_handler_name;
 514
 515         /* we need at least a path arg */
 516         if (as->argc < 1) {
 517                 ti->error = "no device given";
 518                 return ERR_PTR(-EINVAL);
 519         }
 520
 521         p = alloc_pgpath();
 522         if (!p)
 523                 return ERR_PTR(-ENOMEM);
 524
 525         r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 526                           &p->path.dev);
 527         if (r) {
 528                 ti->error = "error getting device";
 529                 goto bad;
 530         }
 531
 532         if (m->retain_attached_hw_handler || m->hw_handler_name)
 533                 q = bdev_get_queue(p->path.dev->bdev);
 534
 535         if (m->retain_attached_hw_handler) {
 536                 attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 537                 if (attached_handler_name) {
 538                         /*
 539                          * Reset hw_handler_name to match the attached handler
 540                          * and clear any hw_handler_params associated with the
 541                          * ignored handler.
 542                          *
 543                          * NB. This modifies the table line to show the actual
 544                          * handler instead of the original table passed in.
 545                          */
 546                         kfree(m->hw_handler_name);
 547                         m->hw_handler_name = attached_handler_name;
 548
 549                         kfree(m->hw_handler_params);
 550                         m->hw_handler_params = NULL;
 551                 }
 552         }
 553
 554         if (m->hw_handler_name) {
 555                 /*
 556                  * Increments scsi_dh reference, even when using an
 557                  * already-attached handler.
 558                  */
 559                 r = scsi_dh_attach(q, m->hw_handler_name);
 560                 if (r == -EBUSY) {
 561                         /*
 562                          * Already attached to different hw_handler:
 563                          * try to reattach with correct one.
 564                          */
 565                         scsi_dh_detach(q);
 566                         r = scsi_dh_attach(q, m->hw_handler_name);
 567                 }
 568
 569                 if (r < 0) {
 570                         ti->error = "error attaching hardware handler";
 571                         dm_put_device(ti, p->path.dev);
 572                         goto bad;
 573                 }
 574
 575                 if (m->hw_handler_params) {
 576                         r = scsi_dh_set_params(q, m->hw_handler_params);
 577                         if (r < 0) {
 578                                 ti->error = "unable to set hardware "
 579                                                         "handler parameters";
 580                                 scsi_dh_detach(q);
 581                                 dm_put_device(ti, p->path.dev);
 582                                 goto bad;
 583                         }
 584                 }
 585         }
 586
 587         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 588         if (r) {
 589                 dm_put_device(ti, p->path.dev);
 590                 goto bad;
 591         }
 592
 593         return p;
 594
 595  bad:
 596         free_pgpath(p);
 597         return ERR_PTR(r);
 598 }
 599
 600 static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 601                                                    struct multipath *m)
 602 {
 603         static struct dm_arg _args[] = {
 604                 {1, 1024, "invalid number of paths"},
 605                 {0, 1024, "invalid number of selector args"}
 606         };
 607
 608         int r;
 609         unsigned i, nr_selector_args, nr_args;
 610         struct priority_group *pg;
 611         struct dm_target *ti = m->ti;
 612
 613         if (as->argc < 2) {
 614                 as->argc = 0;
 615                 ti->error = "not enough priority group arguments";
 616                 return ERR_PTR(-EINVAL);
 617         }
 618
 619         pg = alloc_priority_group();
 620         if (!pg) {
 621                 ti->error = "couldn't allocate priority group";
 622                 return ERR_PTR(-ENOMEM);
 623         }
 624         pg->m = m;
 625
 626         r = parse_path_selector(as, pg, ti);
 627         if (r)
 628                 goto bad;
 629
 630         /*
 631          * read the paths
 632          */
 633         r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 634         if (r)
 635                 goto bad;
 636
 637         r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 638         if (r)
 639                 goto bad;
 640
 641         nr_args = 1 + nr_selector_args;
 642         for (i = 0; i < pg->nr_pgpaths; i++) {
 643                 struct pgpath *pgpath;
 644                 struct dm_arg_set path_args;
 645
 646                 if (as->argc < nr_args) {
 647                         ti->error = "not enough path parameters";
 648                         r = -EINVAL;
 649                         goto bad;
 650                 }
 651
 652                 path_args.argc = nr_args;
 653                 path_args.argv = as->argv;
 654
 655                 pgpath = parse_path(&path_args, &pg->ps, ti);
 656                 if (IS_ERR(pgpath)) {
 657                         r = PTR_ERR(pgpath);
 658                         goto bad;
 659                 }
 660
 661                 pgpath->pg = pg;
 662                 list_add_tail(&pgpath->list, &pg->pgpaths);
 663                 dm_consume_args(as, nr_args);
 664         }
 665
 666         return pg;
 667
 668  bad:
 669         free_priority_group(pg, ti);
 670         return ERR_PTR(r);
 671 }
 672
 673 static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 674 {
 675         unsigned hw_argc;
 676         int ret;
 677         struct dm_target *ti = m->ti;
 678
 679         static struct dm_arg _args[] = {
 680                 {0, 1024, "invalid number of hardware handler args"},
 681         };
 682
 683         if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 684                 return -EINVAL;
 685
 686         if (!hw_argc)
 687                 return 0;
 688
 689         m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 690         if (!try_then_request_module(scsi_dh_handler_exist(m->hw_handler_name),
 691                                      "scsi_dh_%s", m->hw_handler_name)) {
 692                 ti->error = "unknown hardware handler type";
 693                 ret = -EINVAL;
 694                 goto fail;
 695         }
 696
 697         if (hw_argc > 1) {
 698                 char *p;
 699                 int i, j, len = 4;
 700
 701                 for (i = 0; i <= hw_argc - 2; i++)
 702                         len += strlen(as->argv[i]) + 1;
 703                 p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 704                 if (!p) {
 705                         ti->error = "memory allocation failed";
 706                         ret = -ENOMEM;
 707                         goto fail;
 708                 }
 709                 j = sprintf(p, "%d", hw_argc - 1);
 710                 for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 711                         j = sprintf(p, "%s", as->argv[i]);
 712         }
 713         dm_consume_args(as, hw_argc - 1);
 714
 715         return 0;
 716 fail:
 717         kfree(m->hw_handler_name);
 718         m->hw_handler_name = NULL;
 719         return ret;
 720 }
 721
 722 static int parse_features(struct dm_arg_set *as, struct multipath *m)
 723 {
 724         int r;
 725         unsigned argc;
 726         struct dm_target *ti = m->ti;
 727         const char *arg_name;
 728
 729         static struct dm_arg _args[] = {
 730                 {0, 6, "invalid number of feature args"},
 731                 {1, 50, "pg_init_retries must be between 1 and 50"},
 732                 {0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 733         };
 734
 735         r = dm_read_arg_group(_args, as, &argc, &ti->error);
 736         if (r)
 737                 return -EINVAL;
 738
 739         if (!argc)
 740                 return 0;
 741
 742         do {
 743                 arg_name = dm_shift_arg(as);
 744                 argc--;
 745
 746                 if (!strcasecmp(arg_name, "queue_if_no_path")) {
 747                         r = queue_if_no_path(m, 1, 0);
 748                         continue;
 749                 }
 750
 751                 if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
 752                         m->retain_attached_hw_handler = 1;
 753                         continue;
 754                 }
 755
 756                 if (!strcasecmp(arg_name, "pg_init_retries") &&
 757                     (argc >= 1)) {
 758                         r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 759                         argc--;
 760                         continue;
 761                 }
 762
 763                 if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 764                     (argc >= 1)) {
 765                         r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 766                         argc--;
 767                         continue;
 768                 }
 769
 770                 ti->error = "Unrecognised multipath feature request";
 771                 r = -EINVAL;
 772         } while (argc && !r);
 773
 774         return r;
 775 }
 776
 777 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 778                          char **argv)
 779 {
 780         /* target arguments */
 781         static struct dm_arg _args[] = {
 782                 {0, 1024, "invalid number of priority groups"},
 783                 {0, 1024, "invalid initial priority group number"},
 784         };
 785
 786         int r;
 787         struct multipath *m;
 788         struct dm_arg_set as;
 789         unsigned pg_count = 0;
 790         unsigned next_pg_num;
 791
 792         as.argc = argc;
 793         as.argv = argv;
 794
 795         m = alloc_multipath(ti);
 796         if (!m) {
 797                 ti->error = "can't allocate multipath";
 798                 return -EINVAL;
 799         }
 800
 801         r = parse_features(&as, m);
 802         if (r)
 803                 goto bad;
 804
 805         r = parse_hw_handler(&as, m);
 806         if (r)
 807                 goto bad;
 808
 809         r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 810         if (r)
 811                 goto bad;
 812
 813         r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 814         if (r)
 815                 goto bad;
 816
 817         if ((!m->nr_priority_groups && next_pg_num) ||
 818             (m->nr_priority_groups && !next_pg_num)) {
 819                 ti->error = "invalid initial priority group";
 820                 r = -EINVAL;
 821                 goto bad;
 822         }
 823
 824         /* parse the priority groups */
 825         while (as.argc) {
 826                 struct priority_group *pg;
 827
 828                 pg = parse_priority_group(&as, m);
 829                 if (IS_ERR(pg)) {
 830                         r = PTR_ERR(pg);
 831                         goto bad;
 832                 }
 833
 834                 m->nr_valid_paths += pg->nr_pgpaths;
 835                 list_add_tail(&pg->list, &m->priority_groups);
 836                 pg_count++;
 837                 pg->pg_num = pg_count;
 838                 if (!--next_pg_num)
 839                         m->next_pg = pg;
 840         }
 841
 842         if (pg_count != m->nr_priority_groups) {
 843                 ti->error = "priority group count mismatch";
 844                 r = -EINVAL;
 845                 goto bad;
 846         }
 847
 848         ti->num_flush_bios = 1;
 849         ti->num_discard_bios = 1;
 850         ti->num_write_same_bios = 1;
 851
 852         return 0;
 853
 854  bad:
 855         free_multipath(m);
 856         return r;
 857 }
 858
 859 static void multipath_wait_for_pg_init_completion(struct multipath *m)
 860 {
 861         DECLARE_WAITQUEUE(wait, current);
 862         unsigned long flags;
 863
 864         add_wait_queue(&m->pg_init_wait, &wait);
 865
 866         while (1) {
 867                 set_current_state(TASK_UNINTERRUPTIBLE);
 868
 869                 spin_lock_irqsave(&m->lock, flags);
 870                 if (!m->pg_init_in_progress) {
 871                         spin_unlock_irqrestore(&m->lock, flags);
 872                         break;
 873                 }
 874                 spin_unlock_irqrestore(&m->lock, flags);
 875
 876                 io_schedule();
 877         }
 878         set_current_state(TASK_RUNNING);
 879
 880         remove_wait_queue(&m->pg_init_wait, &wait);
 881 }
 882
 883 static void flush_multipath_work(struct multipath *m)
 884 {
 885         unsigned long flags;
 886
 887         spin_lock_irqsave(&m->lock, flags);
 888         m->pg_init_disabled = 1;
 889         spin_unlock_irqrestore(&m->lock, flags);
 890
 891         flush_workqueue(kmpath_handlerd);
 892         multipath_wait_for_pg_init_completion(m);
 893         flush_workqueue(kmultipathd);
 894         flush_work(&m->trigger_event);
 895
 896         spin_lock_irqsave(&m->lock, flags);
 897         m->pg_init_disabled = 0;
 898         spin_unlock_irqrestore(&m->lock, flags);
 899 }
 900
 901 static void multipath_dtr(struct dm_target *ti)
 902 {
 903         struct multipath *m = ti->private;
 904
 905         flush_multipath_work(m);
 906         free_multipath(m);
 907 }
 908
 909 /*
 910  * Map cloned requests
 911  */
 912 static int multipath_map(struct dm_target *ti, struct request *clone,
 913                          union map_info *map_context)
 914 {
 915         int r;
 916         struct multipath *m = (struct multipath *) ti->private;
 917
 918         if (set_mapinfo(m, map_context) < 0)
 919                 /* ENOMEM, requeue */
 920                 return DM_MAPIO_REQUEUE;
 921
 922         clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 923         r = map_io(m, clone, map_context);
 924         if (r < 0 || r == DM_MAPIO_REQUEUE)
 925                 clear_mapinfo(m, map_context);
 926
 927         return r;
 928 }
 929
 930 /*
 931  * Take a path out of use.
 932  */
 933 static int fail_path(struct pgpath *pgpath)
 934 {
 935         unsigned long flags;
 936         struct multipath *m = pgpath->pg->m;
 937
 938         spin_lock_irqsave(&m->lock, flags);
 939
 940         if (!pgpath->is_active)
 941                 goto out;
 942
 943         DMWARN("Failing path %s.", pgpath->path.dev->name);
 944
 945         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 946         pgpath->is_active = 0;
 947         pgpath->fail_count++;
 948
 949         m->nr_valid_paths--;
 950
 951         if (pgpath == m->current_pgpath)
 952                 m->current_pgpath = NULL;
 953
 954         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 955                       pgpath->path.dev->name, m->nr_valid_paths);
 956
 957         schedule_work(&m->trigger_event);
 958
 959 out:
 960         spin_unlock_irqrestore(&m->lock, flags);
 961
 962         return 0;
 963 }
 964
 965 /*
 966  * Reinstate a previously-failed path
 967  */
 968 static int reinstate_path(struct pgpath *pgpath)
 969 {
 970         int r = 0;
 971         unsigned long flags;
 972         struct multipath *m = pgpath->pg->m;
 973
 974         spin_lock_irqsave(&m->lock, flags);
 975
 976         if (pgpath->is_active)
 977                 goto out;
 978
 979         if (!pgpath->pg->ps.type->reinstate_path) {
 980                 DMWARN("Reinstate path not supported by path selector %s",
 981                        pgpath->pg->ps.type->name);
 982                 r = -EINVAL;
 983                 goto out;
 984         }
 985
 986         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 987         if (r)
 988                 goto out;
 989
 990         pgpath->is_active = 1;
 991
 992         if (!m->nr_valid_paths++) {
 993                 m->current_pgpath = NULL;
 994                 dm_table_run_md_queue_async(m->ti->table);
 995         } else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
 996                 if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
 997                         m->pg_init_in_progress++;
 998         }
 999
1000         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1001                       pgpath->path.dev->name, m->nr_valid_paths);
1002
1003         schedule_work(&m->trigger_event);
1004
1005 out:
1006         spin_unlock_irqrestore(&m->lock, flags);
1007
1008         return r;
1009 }
1010
1011 /*
1012  * Fail or reinstate all paths that match the provided struct dm_dev.
1013  */
1014 static int action_dev(struct multipath *m, struct dm_dev *dev,
1015                       action_fn action)
1016 {
1017         int r = -EINVAL;
1018         struct pgpath *pgpath;
1019         struct priority_group *pg;
1020
1021         list_for_each_entry(pg, &m->priority_groups, list) {
1022                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
1023                         if (pgpath->path.dev == dev)
1024                                 r = action(pgpath);
1025                 }
1026         }
1027
1028         return r;
1029 }
1030
1031 /*
1032  * Temporarily try to avoid having to use the specified PG
1033  */
1034 static void bypass_pg(struct multipath *m, struct priority_group *pg,
1035                       int bypassed)
1036 {
1037         unsigned long flags;
1038
1039         spin_lock_irqsave(&m->lock, flags);
1040
1041         pg->bypassed = bypassed;
1042         m->current_pgpath = NULL;
1043         m->current_pg = NULL;
1044
1045         spin_unlock_irqrestore(&m->lock, flags);
1046
1047         schedule_work(&m->trigger_event);
1048 }
1049
1050 /*
1051  * Switch to using the specified PG from the next I/O that gets mapped
1052  */
1053 static int switch_pg_num(struct multipath *m, const char *pgstr)
1054 {
1055         struct priority_group *pg;
1056         unsigned pgnum;
1057         unsigned long flags;
1058         char dummy;
1059
1060         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1061             (pgnum > m->nr_priority_groups)) {
1062                 DMWARN("invalid PG number supplied to switch_pg_num");
1063                 return -EINVAL;
1064         }
1065
1066         spin_lock_irqsave(&m->lock, flags);
1067         list_for_each_entry(pg, &m->priority_groups, list) {
1068                 pg->bypassed = 0;
1069                 if (--pgnum)
1070                         continue;
1071
1072                 m->current_pgpath = NULL;
1073                 m->current_pg = NULL;
1074                 m->next_pg = pg;
1075         }
1076         spin_unlock_irqrestore(&m->lock, flags);
1077
1078         schedule_work(&m->trigger_event);
1079         return 0;
1080 }
1081
1082 /*
1083  * Set/clear bypassed status of a PG.
1084  * PGs are numbered upwards from 1 in the order they were declared.
1085  */
1086 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1087 {
1088         struct priority_group *pg;
1089         unsigned pgnum;
1090         char dummy;
1091
1092         if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1093             (pgnum > m->nr_priority_groups)) {
1094                 DMWARN("invalid PG number supplied to bypass_pg");
1095                 return -EINVAL;
1096         }
1097
1098         list_for_each_entry(pg, &m->priority_groups, list) {
1099                 if (!--pgnum)
1100                         break;
1101         }
1102
1103         bypass_pg(m, pg, bypassed);
1104         return 0;
1105 }
1106
1107 /*
1108  * Should we retry pg_init immediately?
1109  */
1110 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1111 {
1112         unsigned long flags;
1113         int limit_reached = 0;
1114
1115         spin_lock_irqsave(&m->lock, flags);
1116
1117         if (m->pg_init_count <= m->pg_init_retries && !m->pg_init_disabled)
1118                 m->pg_init_required = 1;
1119         else
1120                 limit_reached = 1;
1121
1122         spin_unlock_irqrestore(&m->lock, flags);
1123
1124         return limit_reached;
1125 }
1126
1127 static void pg_init_done(void *data, int errors)
1128 {
1129         struct pgpath *pgpath = data;
1130         struct priority_group *pg = pgpath->pg;
1131         struct multipath *m = pg->m;
1132         unsigned long flags;
1133         unsigned delay_retry = 0;
1134
1135         /* device or driver problems */
1136         switch (errors) {
1137         case SCSI_DH_OK:
1138                 break;
1139         case SCSI_DH_NOSYS:
1140                 if (!m->hw_handler_name) {
1141                         errors = 0;
1142                         break;
1143                 }
1144                 DMERR("Could not failover the device: Handler scsi_dh_%s "
1145                       "Error %d.", m->hw_handler_name, errors);
1146                 /*
1147                  * Fail path for now, so we do not ping pong
1148                  */
1149                 fail_path(pgpath);
1150                 break;
1151         case SCSI_DH_DEV_TEMP_BUSY:
1152                 /*
1153                  * Probably doing something like FW upgrade on the
1154                  * controller so try the other pg.
1155                  */
1156                 bypass_pg(m, pg, 1);
1157                 break;
1158         case SCSI_DH_RETRY:
1159                 /* Wait before retrying. */
1160                 delay_retry = 1;
1161         case SCSI_DH_IMM_RETRY:
1162         case SCSI_DH_RES_TEMP_UNAVAIL:
1163                 if (pg_init_limit_reached(m, pgpath))
1164                         fail_path(pgpath);
1165                 errors = 0;
1166                 break;
1167         default:
1168                 /*
1169                  * We probably do not want to fail the path for a device
1170                  * error, but this is what the old dm did. In future
1171                  * patches we can do more advanced handling.
1172                  */
1173                 fail_path(pgpath);
1174         }
1175
1176         spin_lock_irqsave(&m->lock, flags);
1177         if (errors) {
1178                 if (pgpath == m->current_pgpath) {
1179                         DMERR("Could not failover device. Error %d.", errors);
1180                         m->current_pgpath = NULL;
1181                         m->current_pg = NULL;
1182                 }
1183         } else if (!m->pg_init_required)
1184                 pg->bypassed = 0;
1185
1186         if (--m->pg_init_in_progress)
1187                 /* Activations of other paths are still on going */
1188                 goto out;
1189
1190         if (m->pg_init_required) {
1191                 m->pg_init_delay_retry = delay_retry;
1192                 if (__pg_init_all_paths(m))
1193                         goto out;
1194         }
1195         m->queue_io = 0;
1196
1197         /*
1198          * Wake up any thread waiting to suspend.
1199          */
1200         wake_up(&m->pg_init_wait);
1201
1202 out:
1203         spin_unlock_irqrestore(&m->lock, flags);
1204 }
1205
1206 static void activate_path(struct work_struct *work)
1207 {
1208         struct pgpath *pgpath =
1209                 container_of(work, struct pgpath, activate_path.work);
1210
1211         scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1212                                 pg_init_done, pgpath);
1213 }
1214
1215 static int noretry_error(int error)
1216 {
1217         switch (error) {
1218         case -EOPNOTSUPP:
1219         case -EREMOTEIO:
1220         case -EILSEQ:
1221         case -ENODATA:
1222         case -ENOSPC:
1223                 return 1;
1224         }
1225
1226         /* Anything else could be a path failure, so should be retried */
1227         return 0;
1228 }
1229
1230 /*
1231  * end_io handling
1232  */
1233 static int do_end_io(struct multipath *m, struct request *clone,
1234                      int error, struct dm_mpath_io *mpio)
1235 {
1236         /*
1237          * We don't queue any clone request inside the multipath target
1238          * during end I/O handling, since those clone requests don't have
1239          * bio clones.  If we queue them inside the multipath target,
1240          * we need to make bio clones, that requires memory allocation.
1241          * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1242          *  don't have bio clones.)
1243          * Instead of queueing the clone request here, we queue the original
1244          * request into dm core, which will remake a clone request and
1245          * clone bios for it and resubmit it later.
1246          */
1247         int r = DM_ENDIO_REQUEUE;
1248         unsigned long flags;
1249
1250         if (!error && !clone->errors)
1251                 return 0;       /* I/O complete */
1252
1253         if (noretry_error(error)) {
1254                 if ((clone->cmd_flags & REQ_WRITE_SAME) &&
1255                     !clone->q->limits.max_write_same_sectors) {
1256                         struct queue_limits *limits;
1257
1258                         /* device doesn't really support WRITE SAME, disable it */
1259                         limits = dm_get_queue_limits(dm_table_get_md(m->ti->table));
1260                         limits->max_write_same_sectors = 0;
1261                 }
1262                 return error;
1263         }
1264
1265         if (mpio->pgpath)
1266                 fail_path(mpio->pgpath);
1267
1268         spin_lock_irqsave(&m->lock, flags);
1269         if (!m->nr_valid_paths) {
1270                 if (!m->queue_if_no_path) {
1271                         if (!__must_push_back(m))
1272                                 r = -EIO;
1273                 } else {
1274                         if (error == -EBADE)
1275                                 r = error;
1276                 }
1277         }
1278         spin_unlock_irqrestore(&m->lock, flags);
1279
1280         return r;
1281 }
1282
1283 static int multipath_end_io(struct dm_target *ti, struct request *clone,
1284                             int error, union map_info *map_context)
1285 {
1286         struct multipath *m = ti->private;
1287         struct dm_mpath_io *mpio = map_context->ptr;
1288         struct pgpath *pgpath;
1289         struct path_selector *ps;
1290         int r;
1291
1292         BUG_ON(!mpio);
1293
1294         r  = do_end_io(m, clone, error, mpio);
1295         pgpath = mpio->pgpath;
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         clear_mapinfo(m, map_context);
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 void multipath_status(struct dm_target *ti, status_type_t type,
1358                              unsigned status_flags, char *result, unsigned 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_io, m->pg_init_count);
1373         else {
1374                 DMEMIT("%u ", m->queue_if_no_path +
1375                               (m->pg_init_retries > 0) * 2 +
1376                               (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1377                               m->retain_attached_hw_handler);
1378                 if (m->queue_if_no_path)
1379                         DMEMIT("queue_if_no_path ");
1380                 if (m->pg_init_retries)
1381                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1382                 if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1383                         DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1384                 if (m->retain_attached_hw_handler)
1385                         DMEMIT("retain_attached_hw_handler ");
1386         }
1387
1388         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1389                 DMEMIT("0 ");
1390         else
1391                 DMEMIT("1 %s ", m->hw_handler_name);
1392
1393         DMEMIT("%u ", m->nr_priority_groups);
1394
1395         if (m->next_pg)
1396                 pg_num = m->next_pg->pg_num;
1397         else if (m->current_pg)
1398                 pg_num = m->current_pg->pg_num;
1399         else
1400                 pg_num = (m->nr_priority_groups ? 1 : 0);
1401
1402         DMEMIT("%u ", pg_num);
1403
1404         switch (type) {
1405         case STATUSTYPE_INFO:
1406                 list_for_each_entry(pg, &m->priority_groups, list) {
1407                         if (pg->bypassed)
1408                                 state = 'D';    /* Disabled */
1409                         else if (pg == m->current_pg)
1410                                 state = 'A';    /* Currently Active */
1411                         else
1412                                 state = 'E';    /* Enabled */
1413
1414                         DMEMIT("%c ", state);
1415
1416                         if (pg->ps.type->status)
1417                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1418                                                           result + sz,
1419                                                           maxlen - sz);
1420                         else
1421                                 DMEMIT("0 ");
1422
1423                         DMEMIT("%u %u ", pg->nr_pgpaths,
1424                                pg->ps.type->info_args);
1425
1426                         list_for_each_entry(p, &pg->pgpaths, list) {
1427                                 DMEMIT("%s %s %u ", p->path.dev->name,
1428                                        p->is_active ? "A" : "F",
1429                                        p->fail_count);
1430                                 if (pg->ps.type->status)
1431                                         sz += pg->ps.type->status(&pg->ps,
1432                                               &p->path, type, result + sz,
1433                                               maxlen - sz);
1434                         }
1435                 }
1436                 break;
1437
1438         case STATUSTYPE_TABLE:
1439                 list_for_each_entry(pg, &m->priority_groups, list) {
1440                         DMEMIT("%s ", pg->ps.type->name);
1441
1442                         if (pg->ps.type->status)
1443                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1444                                                           result + sz,
1445                                                           maxlen - sz);
1446                         else
1447                                 DMEMIT("0 ");
1448
1449                         DMEMIT("%u %u ", pg->nr_pgpaths,
1450                                pg->ps.type->table_args);
1451
1452                         list_for_each_entry(p, &pg->pgpaths, list) {
1453                                 DMEMIT("%s ", p->path.dev->name);
1454                                 if (pg->ps.type->status)
1455                                         sz += pg->ps.type->status(&pg->ps,
1456                                               &p->path, type, result + sz,
1457                                               maxlen - sz);
1458                         }
1459                 }
1460                 break;
1461         }
1462
1463         spin_unlock_irqrestore(&m->lock, flags);
1464 }
1465
1466 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1467 {
1468         int r = -EINVAL;
1469         struct dm_dev *dev;
1470         struct multipath *m = (struct multipath *) ti->private;
1471         action_fn action;
1472
1473         mutex_lock(&m->work_mutex);
1474
1475         if (dm_suspended(ti)) {
1476                 r = -EBUSY;
1477                 goto out;
1478         }
1479
1480         if (argc == 1) {
1481                 if (!strcasecmp(argv[0], "queue_if_no_path")) {
1482                         r = queue_if_no_path(m, 1, 0);
1483                         goto out;
1484                 } else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1485                         r = queue_if_no_path(m, 0, 0);
1486                         goto out;
1487                 }
1488         }
1489
1490         if (argc != 2) {
1491                 DMWARN("Unrecognised multipath message received.");
1492                 goto out;
1493         }
1494
1495         if (!strcasecmp(argv[0], "disable_group")) {
1496                 r = bypass_pg_num(m, argv[1], 1);
1497                 goto out;
1498         } else if (!strcasecmp(argv[0], "enable_group")) {
1499                 r = bypass_pg_num(m, argv[1], 0);
1500                 goto out;
1501         } else if (!strcasecmp(argv[0], "switch_group")) {
1502                 r = switch_pg_num(m, argv[1]);
1503                 goto out;
1504         } else if (!strcasecmp(argv[0], "reinstate_path"))
1505                 action = reinstate_path;
1506         else if (!strcasecmp(argv[0], "fail_path"))
1507                 action = fail_path;
1508         else {
1509                 DMWARN("Unrecognised multipath message received.");
1510                 goto out;
1511         }
1512
1513         r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1514         if (r) {
1515                 DMWARN("message: error getting device %s",
1516                        argv[1]);
1517                 goto out;
1518         }
1519
1520         r = action_dev(m, dev, action);
1521
1522         dm_put_device(ti, dev);
1523
1524 out:
1525         mutex_unlock(&m->work_mutex);
1526         return r;
1527 }
1528
1529 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1530                            unsigned long arg)
1531 {
1532         struct multipath *m = ti->private;
1533         struct pgpath *pgpath;
1534         struct block_device *bdev;
1535         fmode_t mode;
1536         unsigned long flags;
1537         int r;
1538
1539         bdev = NULL;
1540         mode = 0;
1541         r = 0;
1542
1543         spin_lock_irqsave(&m->lock, flags);
1544
1545         if (!m->current_pgpath)
1546                 __choose_pgpath(m, 0);
1547
1548         pgpath = m->current_pgpath;
1549
1550         if (pgpath) {
1551                 bdev = pgpath->path.dev->bdev;
1552                 mode = pgpath->path.dev->mode;
1553         }
1554
1555         if ((pgpath && m->queue_io) || (!pgpath && m->queue_if_no_path))
1556                 r = -ENOTCONN;
1557         else if (!bdev)
1558                 r = -EIO;
1559
1560         spin_unlock_irqrestore(&m->lock, flags);
1561
1562         /*
1563          * Only pass ioctls through if the device sizes match exactly.
1564          */
1565         if (!bdev || ti->len != i_size_read(bdev->bd_inode) >> SECTOR_SHIFT) {
1566                 int err = scsi_verify_blk_ioctl(NULL, cmd);
1567                 if (err)
1568                         r = err;
1569         }
1570
1571         if (r == -ENOTCONN && !fatal_signal_pending(current)) {
1572                 spin_lock_irqsave(&m->lock, flags);
1573                 if (!m->current_pg) {
1574                         /* Path status changed, redo selection */
1575                         __choose_pgpath(m, 0);
1576                 }
1577                 if (m->pg_init_required)
1578                         __pg_init_all_paths(m);
1579                 spin_unlock_irqrestore(&m->lock, flags);
1580                 dm_table_run_md_queue_async(m->ti->table);
1581         }
1582
1583         return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
1584 }
1585
1586 static int multipath_iterate_devices(struct dm_target *ti,
1587                                      iterate_devices_callout_fn fn, void *data)
1588 {
1589         struct multipath *m = ti->private;
1590         struct priority_group *pg;
1591         struct pgpath *p;
1592         int ret = 0;
1593
1594         list_for_each_entry(pg, &m->priority_groups, list) {
1595                 list_for_each_entry(p, &pg->pgpaths, list) {
1596                         ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1597                         if (ret)
1598                                 goto out;
1599                 }
1600         }
1601
1602 out:
1603         return ret;
1604 }
1605
1606 static int __pgpath_busy(struct pgpath *pgpath)
1607 {
1608         struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1609
1610         return dm_underlying_device_busy(q);
1611 }
1612
1613 /*
1614  * We return "busy", only when we can map I/Os but underlying devices
1615  * are busy (so even if we map I/Os now, the I/Os will wait on
1616  * the underlying queue).
1617  * In other words, if we want to kill I/Os or queue them inside us
1618  * due to map unavailability, we don't return "busy".  Otherwise,
1619  * dm core won't give us the I/Os and we can't do what we want.
1620  */
1621 static int multipath_busy(struct dm_target *ti)
1622 {
1623         int busy = 0, has_active = 0;
1624         struct multipath *m = ti->private;
1625         struct priority_group *pg;
1626         struct pgpath *pgpath;
1627         unsigned long flags;
1628
1629         spin_lock_irqsave(&m->lock, flags);
1630
1631         /* pg_init in progress, requeue until done */
1632         if (!pg_ready(m)) {
1633                 busy = 1;
1634                 goto out;
1635         }
1636         /* Guess which priority_group will be used at next mapping time */
1637         if (unlikely(!m->current_pgpath && m->next_pg))
1638                 pg = m->next_pg;
1639         else if (likely(m->current_pg))
1640                 pg = m->current_pg;
1641         else
1642                 /*
1643                  * We don't know which pg will be used at next mapping time.
1644                  * We don't call __choose_pgpath() here to avoid to trigger
1645                  * pg_init just by busy checking.
1646                  * So we don't know whether underlying devices we will be using
1647                  * at next mapping time are busy or not. Just try mapping.
1648                  */
1649                 goto out;
1650
1651         /*
1652          * If there is one non-busy active path at least, the path selector
1653          * will be able to select it. So we consider such a pg as not busy.
1654          */
1655         busy = 1;
1656         list_for_each_entry(pgpath, &pg->pgpaths, list)
1657                 if (pgpath->is_active) {
1658                         has_active = 1;
1659
1660                         if (!__pgpath_busy(pgpath)) {
1661                                 busy = 0;
1662                                 break;
1663                         }
1664                 }
1665
1666         if (!has_active)
1667                 /*
1668                  * No active path in this pg, so this pg won't be used and
1669                  * the current_pg will be changed at next mapping time.
1670                  * We need to try mapping to determine it.
1671                  */
1672                 busy = 0;
1673
1674 out:
1675         spin_unlock_irqrestore(&m->lock, flags);
1676
1677         return busy;
1678 }
1679
1680 /*-----------------------------------------------------------------
1681  * Module setup
1682  *---------------------------------------------------------------*/
1683 static struct target_type multipath_target = {
1684         .name = "multipath",
1685         .version = {1, 7, 0},
1686         .module = THIS_MODULE,
1687         .ctr = multipath_ctr,
1688         .dtr = multipath_dtr,
1689         .map_rq = multipath_map,
1690         .rq_end_io = multipath_end_io,
1691         .presuspend = multipath_presuspend,
1692         .postsuspend = multipath_postsuspend,
1693         .resume = multipath_resume,
1694         .status = multipath_status,
1695         .message = multipath_message,
1696         .ioctl  = multipath_ioctl,
1697         .iterate_devices = multipath_iterate_devices,
1698         .busy = multipath_busy,
1699 };
1700
1701 static int __init dm_multipath_init(void)
1702 {
1703         int r;
1704
1705         /* allocate a slab for the dm_ios */
1706         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1707         if (!_mpio_cache)
1708                 return -ENOMEM;
1709
1710         r = dm_register_target(&multipath_target);
1711         if (r < 0) {
1712                 DMERR("register failed %d", r);
1713                 kmem_cache_destroy(_mpio_cache);
1714                 return -EINVAL;
1715         }
1716
1717         kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1718         if (!kmultipathd) {
1719                 DMERR("failed to create workqueue kmpathd");
1720                 dm_unregister_target(&multipath_target);
1721                 kmem_cache_destroy(_mpio_cache);
1722                 return -ENOMEM;
1723         }
1724
1725         /*
1726          * A separate workqueue is used to handle the device handlers
1727          * to avoid overloading existing workqueue. Overloading the
1728          * old workqueue would also create a bottleneck in the
1729          * path of the storage hardware device activation.
1730          */
1731         kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1732                                                   WQ_MEM_RECLAIM);
1733         if (!kmpath_handlerd) {
1734                 DMERR("failed to create workqueue kmpath_handlerd");
1735                 destroy_workqueue(kmultipathd);
1736                 dm_unregister_target(&multipath_target);
1737                 kmem_cache_destroy(_mpio_cache);
1738                 return -ENOMEM;
1739         }
1740
1741         DMINFO("version %u.%u.%u loaded",
1742                multipath_target.version[0], multipath_target.version[1],
1743                multipath_target.version[2]);
1744
1745         return r;
1746 }
1747
1748 static void __exit dm_multipath_exit(void)
1749 {
1750         destroy_workqueue(kmpath_handlerd);
1751         destroy_workqueue(kmultipathd);
1752
1753         dm_unregister_target(&multipath_target);
1754         kmem_cache_destroy(_mpio_cache);
1755 }
1756
1757 module_init(dm_multipath_init);
1758 module_exit(dm_multipath_exit);
1759
1760 MODULE_DESCRIPTION(DM_NAME " multipath target");
1761 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1762 MODULE_LICENSE("GPL");