git.oblomov.eu Git - linux-2.6/blob - 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 "dm.h"
   9 #include "dm-path-selector.h"
  10 #include "dm-hw-handler.h"
  11 #include "dm-bio-list.h"
  12 #include "dm-bio-record.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 <asm/atomic.h>
  23
  24 #define DM_MSG_PREFIX "multipath"
  25 #define MESG_STR(x) x, sizeof(x)
  26
  27 /* Path properties */
  28 struct pgpath {
  29         struct list_head list;
  30
  31         struct priority_group *pg;      /* Owning PG */
  32         unsigned fail_count;            /* Cumulative failure count */
  33
  34         struct path path;
  35 };
  36
  37 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  38
  39 /*
  40  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  41  * Each has a path selector which controls which path gets used.
  42  */
  43 struct priority_group {
  44         struct list_head list;
  45
  46         struct multipath *m;            /* Owning multipath instance */
  47         struct path_selector ps;
  48
  49         unsigned pg_num;                /* Reference number */
  50         unsigned bypassed;              /* Temporarily bypass this PG? */
  51
  52         unsigned nr_pgpaths;            /* Number of paths in PG */
  53         struct list_head pgpaths;
  54 };
  55
  56 /* Multipath context */
  57 struct multipath {
  58         struct list_head list;
  59         struct dm_target *ti;
  60
  61         spinlock_t lock;
  62
  63         struct hw_handler hw_handler;
  64         unsigned nr_priority_groups;
  65         struct list_head priority_groups;
  66         unsigned pg_init_required;      /* pg_init needs calling? */
  67         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  68
  69         unsigned nr_valid_paths;        /* Total number of usable paths */
  70         struct pgpath *current_pgpath;
  71         struct priority_group *current_pg;
  72         struct priority_group *next_pg; /* Switch to this PG if set */
  73         unsigned repeat_count;          /* I/Os left before calling PS again */
  74
  75         unsigned queue_io;              /* Must we queue all I/O? */
  76         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  77         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  78
  79         struct work_struct process_queued_ios;
  80         struct bio_list queued_ios;
  81         unsigned queue_size;
  82
  83         struct work_struct trigger_event;
  84
  85         /*
  86          * We must use a mempool of mpath_io structs so that we
  87          * can resubmit bios on error.
  88          */
  89         mempool_t *mpio_pool;
  90 };
  91
  92 /*
  93  * Context information attached to each bio we process.
  94  */
  95 struct mpath_io {
  96         struct pgpath *pgpath;
  97         struct dm_bio_details details;
  98 };
  99
 100 typedef int (*action_fn) (struct pgpath *pgpath);
 101
 102 #define MIN_IOS 256     /* Mempool size */
 103
 104 static kmem_cache_t *_mpio_cache;
 105
 106 struct workqueue_struct *kmultipathd;
 107 static void process_queued_ios(void *data);
 108 static void trigger_event(void *data);
 109
 110
 111 /*-----------------------------------------------
 112  * Allocation routines
 113  *-----------------------------------------------*/
 114
 115 static struct pgpath *alloc_pgpath(void)
 116 {
 117         struct pgpath *pgpath = kmalloc(sizeof(*pgpath), GFP_KERNEL);
 118
 119         if (pgpath) {
 120                 memset(pgpath, 0, sizeof(*pgpath));
 121                 pgpath->path.is_active = 1;
 122         }
 123
 124         return pgpath;
 125 }
 126
 127 static inline void free_pgpath(struct pgpath *pgpath)
 128 {
 129         kfree(pgpath);
 130 }
 131
 132 static struct priority_group *alloc_priority_group(void)
 133 {
 134         struct priority_group *pg;
 135
 136         pg = kmalloc(sizeof(*pg), GFP_KERNEL);
 137         if (!pg)
 138                 return NULL;
 139
 140         memset(pg, 0, sizeof(*pg));
 141         INIT_LIST_HEAD(&pg->pgpaths);
 142
 143         return pg;
 144 }
 145
 146 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 147 {
 148         struct pgpath *pgpath, *tmp;
 149
 150         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 151                 list_del(&pgpath->list);
 152                 dm_put_device(ti, pgpath->path.dev);
 153                 free_pgpath(pgpath);
 154         }
 155 }
 156
 157 static void free_priority_group(struct priority_group *pg,
 158                                 struct dm_target *ti)
 159 {
 160         struct path_selector *ps = &pg->ps;
 161
 162         if (ps->type) {
 163                 ps->type->destroy(ps);
 164                 dm_put_path_selector(ps->type);
 165         }
 166
 167         free_pgpaths(&pg->pgpaths, ti);
 168         kfree(pg);
 169 }
 170
 171 static struct multipath *alloc_multipath(void)
 172 {
 173         struct multipath *m;
 174
 175         m = kmalloc(sizeof(*m), GFP_KERNEL);
 176         if (m) {
 177                 memset(m, 0, sizeof(*m));
 178                 INIT_LIST_HEAD(&m->priority_groups);
 179                 spin_lock_init(&m->lock);
 180                 m->queue_io = 1;
 181                 INIT_WORK(&m->process_queued_ios, process_queued_ios, m);
 182                 INIT_WORK(&m->trigger_event, trigger_event, m);
 183                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 184                 if (!m->mpio_pool) {
 185                         kfree(m);
 186                         return NULL;
 187                 }
 188         }
 189
 190         return m;
 191 }
 192
 193 static void free_multipath(struct multipath *m)
 194 {
 195         struct priority_group *pg, *tmp;
 196         struct hw_handler *hwh = &m->hw_handler;
 197
 198         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 199                 list_del(&pg->list);
 200                 free_priority_group(pg, m->ti);
 201         }
 202
 203         if (hwh->type) {
 204                 hwh->type->destroy(hwh);
 205                 dm_put_hw_handler(hwh->type);
 206         }
 207
 208         mempool_destroy(m->mpio_pool);
 209         kfree(m);
 210 }
 211
 212
 213 /*-----------------------------------------------
 214  * Path selection
 215  *-----------------------------------------------*/
 216
 217 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 218 {
 219         struct hw_handler *hwh = &m->hw_handler;
 220
 221         m->current_pg = pgpath->pg;
 222
 223         /* Must we initialise the PG first, and queue I/O till it's ready? */
 224         if (hwh->type && hwh->type->pg_init) {
 225                 m->pg_init_required = 1;
 226                 m->queue_io = 1;
 227         } else {
 228                 m->pg_init_required = 0;
 229                 m->queue_io = 0;
 230         }
 231 }
 232
 233 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
 234 {
 235         struct path *path;
 236
 237         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
 238         if (!path)
 239                 return -ENXIO;
 240
 241         m->current_pgpath = path_to_pgpath(path);
 242
 243         if (m->current_pg != pg)
 244                 __switch_pg(m, m->current_pgpath);
 245
 246         return 0;
 247 }
 248
 249 static void __choose_pgpath(struct multipath *m)
 250 {
 251         struct priority_group *pg;
 252         unsigned bypassed = 1;
 253
 254         if (!m->nr_valid_paths)
 255                 goto failed;
 256
 257         /* Were we instructed to switch PG? */
 258         if (m->next_pg) {
 259                 pg = m->next_pg;
 260                 m->next_pg = NULL;
 261                 if (!__choose_path_in_pg(m, pg))
 262                         return;
 263         }
 264
 265         /* Don't change PG until it has no remaining paths */
 266         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
 267                 return;
 268
 269         /*
 270          * Loop through priority groups until we find a valid path.
 271          * First time we skip PGs marked 'bypassed'.
 272          * Second time we only try the ones we skipped.
 273          */
 274         do {
 275                 list_for_each_entry(pg, &m->priority_groups, list) {
 276                         if (pg->bypassed == bypassed)
 277                                 continue;
 278                         if (!__choose_path_in_pg(m, pg))
 279                                 return;
 280                 }
 281         } while (bypassed--);
 282
 283 failed:
 284         m->current_pgpath = NULL;
 285         m->current_pg = NULL;
 286 }
 287
 288 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
 289                   unsigned was_queued)
 290 {
 291         int r = 1;
 292         unsigned long flags;
 293         struct pgpath *pgpath;
 294
 295         spin_lock_irqsave(&m->lock, flags);
 296
 297         /* Do we need to select a new pgpath? */
 298         if (!m->current_pgpath ||
 299             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 300                 __choose_pgpath(m);
 301
 302         pgpath = m->current_pgpath;
 303
 304         if (was_queued)
 305                 m->queue_size--;
 306
 307         if ((pgpath && m->queue_io) ||
 308             (!pgpath && m->queue_if_no_path)) {
 309                 /* Queue for the daemon to resubmit */
 310                 bio_list_add(&m->queued_ios, bio);
 311                 m->queue_size++;
 312                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 313                     !m->queue_io)
 314                         queue_work(kmultipathd, &m->process_queued_ios);
 315                 pgpath = NULL;
 316                 r = 0;
 317         } else if (!pgpath)
 318                 r = -EIO;               /* Failed */
 319         else
 320                 bio->bi_bdev = pgpath->path.dev->bdev;
 321
 322         mpio->pgpath = pgpath;
 323
 324         spin_unlock_irqrestore(&m->lock, flags);
 325
 326         return r;
 327 }
 328
 329 /*
 330  * If we run out of usable paths, should we queue I/O or error it?
 331  */
 332 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 333                             unsigned save_old_value)
 334 {
 335         unsigned long flags;
 336
 337         spin_lock_irqsave(&m->lock, flags);
 338
 339         if (save_old_value)
 340                 m->saved_queue_if_no_path = m->queue_if_no_path;
 341         else
 342                 m->saved_queue_if_no_path = queue_if_no_path;
 343         m->queue_if_no_path = queue_if_no_path;
 344         if (!m->queue_if_no_path && m->queue_size)
 345                 queue_work(kmultipathd, &m->process_queued_ios);
 346
 347         spin_unlock_irqrestore(&m->lock, flags);
 348
 349         return 0;
 350 }
 351
 352 /*-----------------------------------------------------------------
 353  * The multipath daemon is responsible for resubmitting queued ios.
 354  *---------------------------------------------------------------*/
 355
 356 static void dispatch_queued_ios(struct multipath *m)
 357 {
 358         int r;
 359         unsigned long flags;
 360         struct bio *bio = NULL, *next;
 361         struct mpath_io *mpio;
 362         union map_info *info;
 363
 364         spin_lock_irqsave(&m->lock, flags);
 365         bio = bio_list_get(&m->queued_ios);
 366         spin_unlock_irqrestore(&m->lock, flags);
 367
 368         while (bio) {
 369                 next = bio->bi_next;
 370                 bio->bi_next = NULL;
 371
 372                 info = dm_get_mapinfo(bio);
 373                 mpio = info->ptr;
 374
 375                 r = map_io(m, bio, mpio, 1);
 376                 if (r < 0)
 377                         bio_endio(bio, bio->bi_size, r);
 378                 else if (r == 1)
 379                         generic_make_request(bio);
 380
 381                 bio = next;
 382         }
 383 }
 384
 385 static void process_queued_ios(void *data)
 386 {
 387         struct multipath *m = (struct multipath *) data;
 388         struct hw_handler *hwh = &m->hw_handler;
 389         struct pgpath *pgpath = NULL;
 390         unsigned init_required = 0, must_queue = 1;
 391         unsigned long flags;
 392
 393         spin_lock_irqsave(&m->lock, flags);
 394
 395         if (!m->queue_size)
 396                 goto out;
 397
 398         if (!m->current_pgpath)
 399                 __choose_pgpath(m);
 400
 401         pgpath = m->current_pgpath;
 402
 403         if ((pgpath && !m->queue_io) ||
 404             (!pgpath && !m->queue_if_no_path))
 405                 must_queue = 0;
 406
 407         if (m->pg_init_required && !m->pg_init_in_progress) {
 408                 m->pg_init_required = 0;
 409                 m->pg_init_in_progress = 1;
 410                 init_required = 1;
 411         }
 412
 413 out:
 414         spin_unlock_irqrestore(&m->lock, flags);
 415
 416         if (init_required)
 417                 hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
 418
 419         if (!must_queue)
 420                 dispatch_queued_ios(m);
 421 }
 422
 423 /*
 424  * An event is triggered whenever a path is taken out of use.
 425  * Includes path failure and PG bypass.
 426  */
 427 static void trigger_event(void *data)
 428 {
 429         struct multipath *m = (struct multipath *) data;
 430
 431         dm_table_event(m->ti->table);
 432 }
 433
 434 /*-----------------------------------------------------------------
 435  * Constructor/argument parsing:
 436  * <#multipath feature args> [<arg>]*
 437  * <#hw_handler args> [hw_handler [<arg>]*]
 438  * <#priority groups>
 439  * <initial priority group>
 440  *     [<selector> <#selector args> [<arg>]*
 441  *      <#paths> <#per-path selector args>
 442  *         [<path> [<arg>]* ]+ ]+
 443  *---------------------------------------------------------------*/
 444 struct param {
 445         unsigned min;
 446         unsigned max;
 447         char *error;
 448 };
 449
 450 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 451 {
 452         if (!str ||
 453             (sscanf(str, "%u", v) != 1) ||
 454             (*v < param->min) ||
 455             (*v > param->max)) {
 456                 *error = param->error;
 457                 return -EINVAL;
 458         }
 459
 460         return 0;
 461 }
 462
 463 struct arg_set {
 464         unsigned argc;
 465         char **argv;
 466 };
 467
 468 static char *shift(struct arg_set *as)
 469 {
 470         char *r;
 471
 472         if (as->argc) {
 473                 as->argc--;
 474                 r = *as->argv;
 475                 as->argv++;
 476                 return r;
 477         }
 478
 479         return NULL;
 480 }
 481
 482 static void consume(struct arg_set *as, unsigned n)
 483 {
 484         BUG_ON (as->argc < n);
 485         as->argc -= n;
 486         as->argv += n;
 487 }
 488
 489 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 490                                struct dm_target *ti)
 491 {
 492         int r;
 493         struct path_selector_type *pst;
 494         unsigned ps_argc;
 495
 496         static struct param _params[] = {
 497                 {0, 1024, "invalid number of path selector args"},
 498         };
 499
 500         pst = dm_get_path_selector(shift(as));
 501         if (!pst) {
 502                 ti->error = "unknown path selector type";
 503                 return -EINVAL;
 504         }
 505
 506         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 507         if (r)
 508                 return -EINVAL;
 509
 510         r = pst->create(&pg->ps, ps_argc, as->argv);
 511         if (r) {
 512                 dm_put_path_selector(pst);
 513                 ti->error = "path selector constructor failed";
 514                 return r;
 515         }
 516
 517         pg->ps.type = pst;
 518         consume(as, ps_argc);
 519
 520         return 0;
 521 }
 522
 523 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 524                                struct dm_target *ti)
 525 {
 526         int r;
 527         struct pgpath *p;
 528
 529         /* we need at least a path arg */
 530         if (as->argc < 1) {
 531                 ti->error = "no device given";
 532                 return NULL;
 533         }
 534
 535         p = alloc_pgpath();
 536         if (!p)
 537                 return NULL;
 538
 539         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 540                           dm_table_get_mode(ti->table), &p->path.dev);
 541         if (r) {
 542                 ti->error = "error getting device";
 543                 goto bad;
 544         }
 545
 546         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 547         if (r) {
 548                 dm_put_device(ti, p->path.dev);
 549                 goto bad;
 550         }
 551
 552         return p;
 553
 554  bad:
 555         free_pgpath(p);
 556         return NULL;
 557 }
 558
 559 static struct priority_group *parse_priority_group(struct arg_set *as,
 560                                                    struct multipath *m,
 561                                                    struct dm_target *ti)
 562 {
 563         static struct param _params[] = {
 564                 {1, 1024, "invalid number of paths"},
 565                 {0, 1024, "invalid number of selector args"}
 566         };
 567
 568         int r;
 569         unsigned i, nr_selector_args, nr_params;
 570         struct priority_group *pg;
 571
 572         if (as->argc < 2) {
 573                 as->argc = 0;
 574                 ti->error = "not enough priority group aruments";
 575                 return NULL;
 576         }
 577
 578         pg = alloc_priority_group();
 579         if (!pg) {
 580                 ti->error = "couldn't allocate priority group";
 581                 return NULL;
 582         }
 583         pg->m = m;
 584
 585         r = parse_path_selector(as, pg, ti);
 586         if (r)
 587                 goto bad;
 588
 589         /*
 590          * read the paths
 591          */
 592         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 593         if (r)
 594                 goto bad;
 595
 596         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 597         if (r)
 598                 goto bad;
 599
 600         nr_params = 1 + nr_selector_args;
 601         for (i = 0; i < pg->nr_pgpaths; i++) {
 602                 struct pgpath *pgpath;
 603                 struct arg_set path_args;
 604
 605                 if (as->argc < nr_params)
 606                         goto bad;
 607
 608                 path_args.argc = nr_params;
 609                 path_args.argv = as->argv;
 610
 611                 pgpath = parse_path(&path_args, &pg->ps, ti);
 612                 if (!pgpath)
 613                         goto bad;
 614
 615                 pgpath->pg = pg;
 616                 list_add_tail(&pgpath->list, &pg->pgpaths);
 617                 consume(as, nr_params);
 618         }
 619
 620         return pg;
 621
 622  bad:
 623         free_priority_group(pg, ti);
 624         return NULL;
 625 }
 626
 627 static int parse_hw_handler(struct arg_set *as, struct multipath *m,
 628                             struct dm_target *ti)
 629 {
 630         int r;
 631         struct hw_handler_type *hwht;
 632         unsigned hw_argc;
 633
 634         static struct param _params[] = {
 635                 {0, 1024, "invalid number of hardware handler args"},
 636         };
 637
 638         r = read_param(_params, shift(as), &hw_argc, &ti->error);
 639         if (r)
 640                 return -EINVAL;
 641
 642         if (!hw_argc)
 643                 return 0;
 644
 645         hwht = dm_get_hw_handler(shift(as));
 646         if (!hwht) {
 647                 ti->error = "unknown hardware handler type";
 648                 return -EINVAL;
 649         }
 650
 651         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
 652         if (r) {
 653                 dm_put_hw_handler(hwht);
 654                 ti->error = "hardware handler constructor failed";
 655                 return r;
 656         }
 657
 658         m->hw_handler.type = hwht;
 659         consume(as, hw_argc - 1);
 660
 661         return 0;
 662 }
 663
 664 static int parse_features(struct arg_set *as, struct multipath *m,
 665                           struct dm_target *ti)
 666 {
 667         int r;
 668         unsigned argc;
 669
 670         static struct param _params[] = {
 671                 {0, 1, "invalid number of feature args"},
 672         };
 673
 674         r = read_param(_params, shift(as), &argc, &ti->error);
 675         if (r)
 676                 return -EINVAL;
 677
 678         if (!argc)
 679                 return 0;
 680
 681         if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
 682                 return queue_if_no_path(m, 1, 0);
 683         else {
 684                 ti->error = "Unrecognised multipath feature request";
 685                 return -EINVAL;
 686         }
 687 }
 688
 689 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 690                          char **argv)
 691 {
 692         /* target parameters */
 693         static struct param _params[] = {
 694                 {1, 1024, "invalid number of priority groups"},
 695                 {1, 1024, "invalid initial priority group number"},
 696         };
 697
 698         int r;
 699         struct multipath *m;
 700         struct arg_set as;
 701         unsigned pg_count = 0;
 702         unsigned next_pg_num;
 703
 704         as.argc = argc;
 705         as.argv = argv;
 706
 707         m = alloc_multipath();
 708         if (!m) {
 709                 ti->error = "can't allocate multipath";
 710                 return -EINVAL;
 711         }
 712
 713         m->ti = ti;
 714
 715         r = parse_features(&as, m, ti);
 716         if (r)
 717                 goto bad;
 718
 719         r = parse_hw_handler(&as, m, ti);
 720         if (r)
 721                 goto bad;
 722
 723         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 724         if (r)
 725                 goto bad;
 726
 727         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 728         if (r)
 729                 goto bad;
 730
 731         /* parse the priority groups */
 732         while (as.argc) {
 733                 struct priority_group *pg;
 734
 735                 pg = parse_priority_group(&as, m, ti);
 736                 if (!pg) {
 737                         r = -EINVAL;
 738                         goto bad;
 739                 }
 740
 741                 m->nr_valid_paths += pg->nr_pgpaths;
 742                 list_add_tail(&pg->list, &m->priority_groups);
 743                 pg_count++;
 744                 pg->pg_num = pg_count;
 745                 if (!--next_pg_num)
 746                         m->next_pg = pg;
 747         }
 748
 749         if (pg_count != m->nr_priority_groups) {
 750                 ti->error = "priority group count mismatch";
 751                 r = -EINVAL;
 752                 goto bad;
 753         }
 754
 755         ti->private = m;
 756
 757         return 0;
 758
 759  bad:
 760         free_multipath(m);
 761         return r;
 762 }
 763
 764 static void multipath_dtr(struct dm_target *ti)
 765 {
 766         struct multipath *m = (struct multipath *) ti->private;
 767
 768         flush_workqueue(kmultipathd);
 769         free_multipath(m);
 770 }
 771
 772 /*
 773  * Map bios, recording original fields for later in case we have to resubmit
 774  */
 775 static int multipath_map(struct dm_target *ti, struct bio *bio,
 776                          union map_info *map_context)
 777 {
 778         int r;
 779         struct mpath_io *mpio;
 780         struct multipath *m = (struct multipath *) ti->private;
 781
 782         if (bio_barrier(bio))
 783                 return -EOPNOTSUPP;
 784
 785         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 786         dm_bio_record(&mpio->details, bio);
 787
 788         map_context->ptr = mpio;
 789         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 790         r = map_io(m, bio, mpio, 0);
 791         if (r < 0)
 792                 mempool_free(mpio, m->mpio_pool);
 793
 794         return r;
 795 }
 796
 797 /*
 798  * Take a path out of use.
 799  */
 800 static int fail_path(struct pgpath *pgpath)
 801 {
 802         unsigned long flags;
 803         struct multipath *m = pgpath->pg->m;
 804
 805         spin_lock_irqsave(&m->lock, flags);
 806
 807         if (!pgpath->path.is_active)
 808                 goto out;
 809
 810         DMWARN("Failing path %s.", pgpath->path.dev->name);
 811
 812         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 813         pgpath->path.is_active = 0;
 814         pgpath->fail_count++;
 815
 816         m->nr_valid_paths--;
 817
 818         if (pgpath == m->current_pgpath)
 819                 m->current_pgpath = NULL;
 820
 821         queue_work(kmultipathd, &m->trigger_event);
 822
 823 out:
 824         spin_unlock_irqrestore(&m->lock, flags);
 825
 826         return 0;
 827 }
 828
 829 /*
 830  * Reinstate a previously-failed path
 831  */
 832 static int reinstate_path(struct pgpath *pgpath)
 833 {
 834         int r = 0;
 835         unsigned long flags;
 836         struct multipath *m = pgpath->pg->m;
 837
 838         spin_lock_irqsave(&m->lock, flags);
 839
 840         if (pgpath->path.is_active)
 841                 goto out;
 842
 843         if (!pgpath->pg->ps.type) {
 844                 DMWARN("Reinstate path not supported by path selector %s",
 845                        pgpath->pg->ps.type->name);
 846                 r = -EINVAL;
 847                 goto out;
 848         }
 849
 850         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 851         if (r)
 852                 goto out;
 853
 854         pgpath->path.is_active = 1;
 855
 856         m->current_pgpath = NULL;
 857         if (!m->nr_valid_paths++ && m->queue_size)
 858                 queue_work(kmultipathd, &m->process_queued_ios);
 859
 860         queue_work(kmultipathd, &m->trigger_event);
 861
 862 out:
 863         spin_unlock_irqrestore(&m->lock, flags);
 864
 865         return r;
 866 }
 867
 868 /*
 869  * Fail or reinstate all paths that match the provided struct dm_dev.
 870  */
 871 static int action_dev(struct multipath *m, struct dm_dev *dev,
 872                       action_fn action)
 873 {
 874         int r = 0;
 875         struct pgpath *pgpath;
 876         struct priority_group *pg;
 877
 878         list_for_each_entry(pg, &m->priority_groups, list) {
 879                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 880                         if (pgpath->path.dev == dev)
 881                                 r = action(pgpath);
 882                 }
 883         }
 884
 885         return r;
 886 }
 887
 888 /*
 889  * Temporarily try to avoid having to use the specified PG
 890  */
 891 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 892                       int bypassed)
 893 {
 894         unsigned long flags;
 895
 896         spin_lock_irqsave(&m->lock, flags);
 897
 898         pg->bypassed = bypassed;
 899         m->current_pgpath = NULL;
 900         m->current_pg = NULL;
 901
 902         spin_unlock_irqrestore(&m->lock, flags);
 903
 904         queue_work(kmultipathd, &m->trigger_event);
 905 }
 906
 907 /*
 908  * Switch to using the specified PG from the next I/O that gets mapped
 909  */
 910 static int switch_pg_num(struct multipath *m, const char *pgstr)
 911 {
 912         struct priority_group *pg;
 913         unsigned pgnum;
 914         unsigned long flags;
 915
 916         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 917             (pgnum > m->nr_priority_groups)) {
 918                 DMWARN("invalid PG number supplied to switch_pg_num");
 919                 return -EINVAL;
 920         }
 921
 922         spin_lock_irqsave(&m->lock, flags);
 923         list_for_each_entry(pg, &m->priority_groups, list) {
 924                 pg->bypassed = 0;
 925                 if (--pgnum)
 926                         continue;
 927
 928                 m->current_pgpath = NULL;
 929                 m->current_pg = NULL;
 930                 m->next_pg = pg;
 931         }
 932         spin_unlock_irqrestore(&m->lock, flags);
 933
 934         queue_work(kmultipathd, &m->trigger_event);
 935         return 0;
 936 }
 937
 938 /*
 939  * Set/clear bypassed status of a PG.
 940  * PGs are numbered upwards from 1 in the order they were declared.
 941  */
 942 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 943 {
 944         struct priority_group *pg;
 945         unsigned pgnum;
 946
 947         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 948             (pgnum > m->nr_priority_groups)) {
 949                 DMWARN("invalid PG number supplied to bypass_pg");
 950                 return -EINVAL;
 951         }
 952
 953         list_for_each_entry(pg, &m->priority_groups, list) {
 954                 if (!--pgnum)
 955                         break;
 956         }
 957
 958         bypass_pg(m, pg, bypassed);
 959         return 0;
 960 }
 961
 962 /*
 963  * pg_init must call this when it has completed its initialisation
 964  */
 965 void dm_pg_init_complete(struct path *path, unsigned err_flags)
 966 {
 967         struct pgpath *pgpath = path_to_pgpath(path);
 968         struct priority_group *pg = pgpath->pg;
 969         struct multipath *m = pg->m;
 970         unsigned long flags;
 971
 972         /* We insist on failing the path if the PG is already bypassed. */
 973         if (err_flags && pg->bypassed)
 974                 err_flags |= MP_FAIL_PATH;
 975
 976         if (err_flags & MP_FAIL_PATH)
 977                 fail_path(pgpath);
 978
 979         if (err_flags & MP_BYPASS_PG)
 980                 bypass_pg(m, pg, 1);
 981
 982         spin_lock_irqsave(&m->lock, flags);
 983         if (err_flags) {
 984                 m->current_pgpath = NULL;
 985                 m->current_pg = NULL;
 986         } else if (!m->pg_init_required)
 987                 m->queue_io = 0;
 988
 989         m->pg_init_in_progress = 0;
 990         queue_work(kmultipathd, &m->process_queued_ios);
 991         spin_unlock_irqrestore(&m->lock, flags);
 992 }
 993
 994 /*
 995  * end_io handling
 996  */
 997 static int do_end_io(struct multipath *m, struct bio *bio,
 998                      int error, struct mpath_io *mpio)
 999 {
1000         struct hw_handler *hwh = &m->hw_handler;
1001         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
1002         unsigned long flags;
1003
1004         if (!error)
1005                 return 0;       /* I/O complete */
1006
1007         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1008                 return error;
1009
1010         if (error == -EOPNOTSUPP)
1011                 return error;
1012
1013         spin_lock_irqsave(&m->lock, flags);
1014         if (!m->nr_valid_paths) {
1015                 if (!m->queue_if_no_path) {
1016                         spin_unlock_irqrestore(&m->lock, flags);
1017                         return -EIO;
1018                 } else {
1019                         spin_unlock_irqrestore(&m->lock, flags);
1020                         goto requeue;
1021                 }
1022         }
1023         spin_unlock_irqrestore(&m->lock, flags);
1024
1025         if (hwh->type && hwh->type->error)
1026                 err_flags = hwh->type->error(hwh, bio);
1027
1028         if (mpio->pgpath) {
1029                 if (err_flags & MP_FAIL_PATH)
1030                         fail_path(mpio->pgpath);
1031
1032                 if (err_flags & MP_BYPASS_PG)
1033                         bypass_pg(m, mpio->pgpath->pg, 1);
1034         }
1035
1036         if (err_flags & MP_ERROR_IO)
1037                 return -EIO;
1038
1039       requeue:
1040         dm_bio_restore(&mpio->details, bio);
1041
1042         /* queue for the daemon to resubmit or fail */
1043         spin_lock_irqsave(&m->lock, flags);
1044         bio_list_add(&m->queued_ios, bio);
1045         m->queue_size++;
1046         if (!m->queue_io)
1047                 queue_work(kmultipathd, &m->process_queued_ios);
1048         spin_unlock_irqrestore(&m->lock, flags);
1049
1050         return 1;       /* io not complete */
1051 }
1052
1053 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1054                             int error, union map_info *map_context)
1055 {
1056         struct multipath *m = (struct multipath *) ti->private;
1057         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1058         struct pgpath *pgpath = mpio->pgpath;
1059         struct path_selector *ps;
1060         int r;
1061
1062         r  = do_end_io(m, bio, error, mpio);
1063         if (pgpath) {
1064                 ps = &pgpath->pg->ps;
1065                 if (ps->type->end_io)
1066                         ps->type->end_io(ps, &pgpath->path);
1067         }
1068         if (r <= 0)
1069                 mempool_free(mpio, m->mpio_pool);
1070
1071         return r;
1072 }
1073
1074 /*
1075  * Suspend can't complete until all the I/O is processed so if
1076  * the last path fails we must error any remaining I/O.
1077  * Note that if the freeze_bdev fails while suspending, the
1078  * queue_if_no_path state is lost - userspace should reset it.
1079  */
1080 static void multipath_presuspend(struct dm_target *ti)
1081 {
1082         struct multipath *m = (struct multipath *) ti->private;
1083
1084         queue_if_no_path(m, 0, 1);
1085 }
1086
1087 /*
1088  * Restore the queue_if_no_path setting.
1089  */
1090 static void multipath_resume(struct dm_target *ti)
1091 {
1092         struct multipath *m = (struct multipath *) ti->private;
1093         unsigned long flags;
1094
1095         spin_lock_irqsave(&m->lock, flags);
1096         m->queue_if_no_path = m->saved_queue_if_no_path;
1097         spin_unlock_irqrestore(&m->lock, flags);
1098 }
1099
1100 /*
1101  * Info output has the following format:
1102  * num_multipath_feature_args [multipath_feature_args]*
1103  * num_handler_status_args [handler_status_args]*
1104  * num_groups init_group_number
1105  *            [A|D|E num_ps_status_args [ps_status_args]*
1106  *             num_paths num_selector_args
1107  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1108  *
1109  * Table output has the following format (identical to the constructor string):
1110  * num_feature_args [features_args]*
1111  * num_handler_args hw_handler [hw_handler_args]*
1112  * num_groups init_group_number
1113  *     [priority selector-name num_ps_args [ps_args]*
1114  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1115  */
1116 static int multipath_status(struct dm_target *ti, status_type_t type,
1117                             char *result, unsigned int maxlen)
1118 {
1119         int sz = 0;
1120         unsigned long flags;
1121         struct multipath *m = (struct multipath *) ti->private;
1122         struct hw_handler *hwh = &m->hw_handler;
1123         struct priority_group *pg;
1124         struct pgpath *p;
1125         unsigned pg_num;
1126         char state;
1127
1128         spin_lock_irqsave(&m->lock, flags);
1129
1130         /* Features */
1131         if (type == STATUSTYPE_INFO)
1132                 DMEMIT("1 %u ", m->queue_size);
1133         else if (m->queue_if_no_path)
1134                 DMEMIT("1 queue_if_no_path ");
1135         else
1136                 DMEMIT("0 ");
1137
1138         if (hwh->type && hwh->type->status)
1139                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1140         else if (!hwh->type || type == STATUSTYPE_INFO)
1141                 DMEMIT("0 ");
1142         else
1143                 DMEMIT("1 %s ", hwh->type->name);
1144
1145         DMEMIT("%u ", m->nr_priority_groups);
1146
1147         if (m->next_pg)
1148                 pg_num = m->next_pg->pg_num;
1149         else if (m->current_pg)
1150                 pg_num = m->current_pg->pg_num;
1151         else
1152                         pg_num = 1;
1153
1154         DMEMIT("%u ", pg_num);
1155
1156         switch (type) {
1157         case STATUSTYPE_INFO:
1158                 list_for_each_entry(pg, &m->priority_groups, list) {
1159                         if (pg->bypassed)
1160                                 state = 'D';    /* Disabled */
1161                         else if (pg == m->current_pg)
1162                                 state = 'A';    /* Currently Active */
1163                         else
1164                                 state = 'E';    /* Enabled */
1165
1166                         DMEMIT("%c ", state);
1167
1168                         if (pg->ps.type->status)
1169                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1170                                                           result + sz,
1171                                                           maxlen - sz);
1172                         else
1173                                 DMEMIT("0 ");
1174
1175                         DMEMIT("%u %u ", pg->nr_pgpaths,
1176                                pg->ps.type->info_args);
1177
1178                         list_for_each_entry(p, &pg->pgpaths, list) {
1179                                 DMEMIT("%s %s %u ", p->path.dev->name,
1180                                        p->path.is_active ? "A" : "F",
1181                                        p->fail_count);
1182                                 if (pg->ps.type->status)
1183                                         sz += pg->ps.type->status(&pg->ps,
1184                                               &p->path, type, result + sz,
1185                                               maxlen - sz);
1186                         }
1187                 }
1188                 break;
1189
1190         case STATUSTYPE_TABLE:
1191                 list_for_each_entry(pg, &m->priority_groups, list) {
1192                         DMEMIT("%s ", pg->ps.type->name);
1193
1194                         if (pg->ps.type->status)
1195                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1196                                                           result + sz,
1197                                                           maxlen - sz);
1198                         else
1199                                 DMEMIT("0 ");
1200
1201                         DMEMIT("%u %u ", pg->nr_pgpaths,
1202                                pg->ps.type->table_args);
1203
1204                         list_for_each_entry(p, &pg->pgpaths, list) {
1205                                 DMEMIT("%s ", p->path.dev->name);
1206                                 if (pg->ps.type->status)
1207                                         sz += pg->ps.type->status(&pg->ps,
1208                                               &p->path, type, result + sz,
1209                                               maxlen - sz);
1210                         }
1211                 }
1212                 break;
1213         }
1214
1215         spin_unlock_irqrestore(&m->lock, flags);
1216
1217         return 0;
1218 }
1219
1220 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1221 {
1222         int r;
1223         struct dm_dev *dev;
1224         struct multipath *m = (struct multipath *) ti->private;
1225         action_fn action;
1226
1227         if (argc == 1) {
1228                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1229                         return queue_if_no_path(m, 1, 0);
1230                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1231                         return queue_if_no_path(m, 0, 0);
1232         }
1233
1234         if (argc != 2)
1235                 goto error;
1236
1237         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1238                 return bypass_pg_num(m, argv[1], 1);
1239         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1240                 return bypass_pg_num(m, argv[1], 0);
1241         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1242                 return switch_pg_num(m, argv[1]);
1243         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1244                 action = reinstate_path;
1245         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1246                 action = fail_path;
1247         else
1248                 goto error;
1249
1250         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1251                           dm_table_get_mode(ti->table), &dev);
1252         if (r) {
1253                 DMWARN("message: error getting device %s",
1254                        argv[1]);
1255                 return -EINVAL;
1256         }
1257
1258         r = action_dev(m, dev, action);
1259
1260         dm_put_device(ti, dev);
1261
1262         return r;
1263
1264 error:
1265         DMWARN("Unrecognised multipath message received.");
1266         return -EINVAL;
1267 }
1268
1269 /*-----------------------------------------------------------------
1270  * Module setup
1271  *---------------------------------------------------------------*/
1272 static struct target_type multipath_target = {
1273         .name = "multipath",
1274         .version = {1, 0, 4},
1275         .module = THIS_MODULE,
1276         .ctr = multipath_ctr,
1277         .dtr = multipath_dtr,
1278         .map = multipath_map,
1279         .end_io = multipath_end_io,
1280         .presuspend = multipath_presuspend,
1281         .resume = multipath_resume,
1282         .status = multipath_status,
1283         .message = multipath_message,
1284 };
1285
1286 static int __init dm_multipath_init(void)
1287 {
1288         int r;
1289
1290         /* allocate a slab for the dm_ios */
1291         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1292                                         0, 0, NULL, NULL);
1293         if (!_mpio_cache)
1294                 return -ENOMEM;
1295
1296         r = dm_register_target(&multipath_target);
1297         if (r < 0) {
1298                 DMERR("%s: register failed %d", multipath_target.name, r);
1299                 kmem_cache_destroy(_mpio_cache);
1300                 return -EINVAL;
1301         }
1302
1303         kmultipathd = create_workqueue("kmpathd");
1304         if (!kmultipathd) {
1305                 DMERR("%s: failed to create workqueue kmpathd",
1306                                 multipath_target.name);
1307                 dm_unregister_target(&multipath_target);
1308                 kmem_cache_destroy(_mpio_cache);
1309                 return -ENOMEM;
1310         }
1311
1312         DMINFO("version %u.%u.%u loaded",
1313                multipath_target.version[0], multipath_target.version[1],
1314                multipath_target.version[2]);
1315
1316         return r;
1317 }
1318
1319 static void __exit dm_multipath_exit(void)
1320 {
1321         int r;
1322
1323         destroy_workqueue(kmultipathd);
1324
1325         r = dm_unregister_target(&multipath_target);
1326         if (r < 0)
1327                 DMERR("%s: target unregister failed %d",
1328                       multipath_target.name, r);
1329         kmem_cache_destroy(_mpio_cache);
1330 }
1331
1332 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1333
1334 module_init(dm_multipath_init);
1335 module_exit(dm_multipath_exit);
1336
1337 MODULE_DESCRIPTION(DM_NAME " multipath target");
1338 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1339 MODULE_LICENSE("GPL");