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         r = parse_features(&as, m, ti);
 714         if (r)
 715                 goto bad;
 716
 717         r = parse_hw_handler(&as, m, ti);
 718         if (r)
 719                 goto bad;
 720
 721         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 722         if (r)
 723                 goto bad;
 724
 725         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 726         if (r)
 727                 goto bad;
 728
 729         /* parse the priority groups */
 730         while (as.argc) {
 731                 struct priority_group *pg;
 732
 733                 pg = parse_priority_group(&as, m, ti);
 734                 if (!pg) {
 735                         r = -EINVAL;
 736                         goto bad;
 737                 }
 738
 739                 m->nr_valid_paths += pg->nr_pgpaths;
 740                 list_add_tail(&pg->list, &m->priority_groups);
 741                 pg_count++;
 742                 pg->pg_num = pg_count;
 743                 if (!--next_pg_num)
 744                         m->next_pg = pg;
 745         }
 746
 747         if (pg_count != m->nr_priority_groups) {
 748                 ti->error = "priority group count mismatch";
 749                 r = -EINVAL;
 750                 goto bad;
 751         }
 752
 753         ti->private = m;
 754         m->ti = ti;
 755
 756         return 0;
 757
 758  bad:
 759         free_multipath(m);
 760         return r;
 761 }
 762
 763 static void multipath_dtr(struct dm_target *ti)
 764 {
 765         struct multipath *m = (struct multipath *) ti->private;
 766
 767         flush_workqueue(kmultipathd);
 768         free_multipath(m);
 769 }
 770
 771 /*
 772  * Map bios, recording original fields for later in case we have to resubmit
 773  */
 774 static int multipath_map(struct dm_target *ti, struct bio *bio,
 775                          union map_info *map_context)
 776 {
 777         int r;
 778         struct mpath_io *mpio;
 779         struct multipath *m = (struct multipath *) ti->private;
 780
 781         if (bio_barrier(bio))
 782                 return -EOPNOTSUPP;
 783
 784         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 785         dm_bio_record(&mpio->details, bio);
 786
 787         map_context->ptr = mpio;
 788         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 789         r = map_io(m, bio, mpio, 0);
 790         if (r < 0)
 791                 mempool_free(mpio, m->mpio_pool);
 792
 793         return r;
 794 }
 795
 796 /*
 797  * Take a path out of use.
 798  */
 799 static int fail_path(struct pgpath *pgpath)
 800 {
 801         unsigned long flags;
 802         struct multipath *m = pgpath->pg->m;
 803
 804         spin_lock_irqsave(&m->lock, flags);
 805
 806         if (!pgpath->path.is_active)
 807                 goto out;
 808
 809         DMWARN("Failing path %s.", pgpath->path.dev->name);
 810
 811         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 812         pgpath->path.is_active = 0;
 813         pgpath->fail_count++;
 814
 815         m->nr_valid_paths--;
 816
 817         if (pgpath == m->current_pgpath)
 818                 m->current_pgpath = NULL;
 819
 820         queue_work(kmultipathd, &m->trigger_event);
 821
 822 out:
 823         spin_unlock_irqrestore(&m->lock, flags);
 824
 825         return 0;
 826 }
 827
 828 /*
 829  * Reinstate a previously-failed path
 830  */
 831 static int reinstate_path(struct pgpath *pgpath)
 832 {
 833         int r = 0;
 834         unsigned long flags;
 835         struct multipath *m = pgpath->pg->m;
 836
 837         spin_lock_irqsave(&m->lock, flags);
 838
 839         if (pgpath->path.is_active)
 840                 goto out;
 841
 842         if (!pgpath->pg->ps.type) {
 843                 DMWARN("Reinstate path not supported by path selector %s",
 844                        pgpath->pg->ps.type->name);
 845                 r = -EINVAL;
 846                 goto out;
 847         }
 848
 849         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 850         if (r)
 851                 goto out;
 852
 853         pgpath->path.is_active = 1;
 854
 855         m->current_pgpath = NULL;
 856         if (!m->nr_valid_paths++ && m->queue_size)
 857                 queue_work(kmultipathd, &m->process_queued_ios);
 858
 859         queue_work(kmultipathd, &m->trigger_event);
 860
 861 out:
 862         spin_unlock_irqrestore(&m->lock, flags);
 863
 864         return r;
 865 }
 866
 867 /*
 868  * Fail or reinstate all paths that match the provided struct dm_dev.
 869  */
 870 static int action_dev(struct multipath *m, struct dm_dev *dev,
 871                       action_fn action)
 872 {
 873         int r = 0;
 874         struct pgpath *pgpath;
 875         struct priority_group *pg;
 876
 877         list_for_each_entry(pg, &m->priority_groups, list) {
 878                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 879                         if (pgpath->path.dev == dev)
 880                                 r = action(pgpath);
 881                 }
 882         }
 883
 884         return r;
 885 }
 886
 887 /*
 888  * Temporarily try to avoid having to use the specified PG
 889  */
 890 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 891                       int bypassed)
 892 {
 893         unsigned long flags;
 894
 895         spin_lock_irqsave(&m->lock, flags);
 896
 897         pg->bypassed = bypassed;
 898         m->current_pgpath = NULL;
 899         m->current_pg = NULL;
 900
 901         spin_unlock_irqrestore(&m->lock, flags);
 902
 903         queue_work(kmultipathd, &m->trigger_event);
 904 }
 905
 906 /*
 907  * Switch to using the specified PG from the next I/O that gets mapped
 908  */
 909 static int switch_pg_num(struct multipath *m, const char *pgstr)
 910 {
 911         struct priority_group *pg;
 912         unsigned pgnum;
 913         unsigned long flags;
 914
 915         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 916             (pgnum > m->nr_priority_groups)) {
 917                 DMWARN("invalid PG number supplied to switch_pg_num");
 918                 return -EINVAL;
 919         }
 920
 921         spin_lock_irqsave(&m->lock, flags);
 922         list_for_each_entry(pg, &m->priority_groups, list) {
 923                 pg->bypassed = 0;
 924                 if (--pgnum)
 925                         continue;
 926
 927                 m->current_pgpath = NULL;
 928                 m->current_pg = NULL;
 929                 m->next_pg = pg;
 930         }
 931         spin_unlock_irqrestore(&m->lock, flags);
 932
 933         queue_work(kmultipathd, &m->trigger_event);
 934         return 0;
 935 }
 936
 937 /*
 938  * Set/clear bypassed status of a PG.
 939  * PGs are numbered upwards from 1 in the order they were declared.
 940  */
 941 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 942 {
 943         struct priority_group *pg;
 944         unsigned pgnum;
 945
 946         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 947             (pgnum > m->nr_priority_groups)) {
 948                 DMWARN("invalid PG number supplied to bypass_pg");
 949                 return -EINVAL;
 950         }
 951
 952         list_for_each_entry(pg, &m->priority_groups, list) {
 953                 if (!--pgnum)
 954                         break;
 955         }
 956
 957         bypass_pg(m, pg, bypassed);
 958         return 0;
 959 }
 960
 961 /*
 962  * pg_init must call this when it has completed its initialisation
 963  */
 964 void dm_pg_init_complete(struct path *path, unsigned err_flags)
 965 {
 966         struct pgpath *pgpath = path_to_pgpath(path);
 967         struct priority_group *pg = pgpath->pg;
 968         struct multipath *m = pg->m;
 969         unsigned long flags;
 970
 971         /* We insist on failing the path if the PG is already bypassed. */
 972         if (err_flags && pg->bypassed)
 973                 err_flags |= MP_FAIL_PATH;
 974
 975         if (err_flags & MP_FAIL_PATH)
 976                 fail_path(pgpath);
 977
 978         if (err_flags & MP_BYPASS_PG)
 979                 bypass_pg(m, pg, 1);
 980
 981         spin_lock_irqsave(&m->lock, flags);
 982         if (err_flags) {
 983                 m->current_pgpath = NULL;
 984                 m->current_pg = NULL;
 985         } else if (!m->pg_init_required)
 986                 m->queue_io = 0;
 987
 988         m->pg_init_in_progress = 0;
 989         queue_work(kmultipathd, &m->process_queued_ios);
 990         spin_unlock_irqrestore(&m->lock, flags);
 991 }
 992
 993 /*
 994  * end_io handling
 995  */
 996 static int do_end_io(struct multipath *m, struct bio *bio,
 997                      int error, struct mpath_io *mpio)
 998 {
 999         struct hw_handler *hwh = &m->hw_handler;
1000         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
1001         unsigned long flags;
1002
1003         if (!error)
1004                 return 0;       /* I/O complete */
1005
1006         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1007                 return error;
1008
1009         if (error == -EOPNOTSUPP)
1010                 return error;
1011
1012         spin_lock_irqsave(&m->lock, flags);
1013         if (!m->nr_valid_paths) {
1014                 if (!m->queue_if_no_path) {
1015                         spin_unlock_irqrestore(&m->lock, flags);
1016                         return -EIO;
1017                 } else {
1018                         spin_unlock_irqrestore(&m->lock, flags);
1019                         goto requeue;
1020                 }
1021         }
1022         spin_unlock_irqrestore(&m->lock, flags);
1023
1024         if (hwh->type && hwh->type->error)
1025                 err_flags = hwh->type->error(hwh, bio);
1026
1027         if (mpio->pgpath) {
1028                 if (err_flags & MP_FAIL_PATH)
1029                         fail_path(mpio->pgpath);
1030
1031                 if (err_flags & MP_BYPASS_PG)
1032                         bypass_pg(m, mpio->pgpath->pg, 1);
1033         }
1034
1035         if (err_flags & MP_ERROR_IO)
1036                 return -EIO;
1037
1038       requeue:
1039         dm_bio_restore(&mpio->details, bio);
1040
1041         /* queue for the daemon to resubmit or fail */
1042         spin_lock_irqsave(&m->lock, flags);
1043         bio_list_add(&m->queued_ios, bio);
1044         m->queue_size++;
1045         if (!m->queue_io)
1046                 queue_work(kmultipathd, &m->process_queued_ios);
1047         spin_unlock_irqrestore(&m->lock, flags);
1048
1049         return 1;       /* io not complete */
1050 }
1051
1052 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1053                             int error, union map_info *map_context)
1054 {
1055         struct multipath *m = (struct multipath *) ti->private;
1056         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1057         struct pgpath *pgpath = mpio->pgpath;
1058         struct path_selector *ps;
1059         int r;
1060
1061         r  = do_end_io(m, bio, error, mpio);
1062         if (pgpath) {
1063                 ps = &pgpath->pg->ps;
1064                 if (ps->type->end_io)
1065                         ps->type->end_io(ps, &pgpath->path);
1066         }
1067         if (r <= 0)
1068                 mempool_free(mpio, m->mpio_pool);
1069
1070         return r;
1071 }
1072
1073 /*
1074  * Suspend can't complete until all the I/O is processed so if
1075  * the last path fails we must error any remaining I/O.
1076  * Note that if the freeze_bdev fails while suspending, the
1077  * queue_if_no_path state is lost - userspace should reset it.
1078  */
1079 static void multipath_presuspend(struct dm_target *ti)
1080 {
1081         struct multipath *m = (struct multipath *) ti->private;
1082
1083         queue_if_no_path(m, 0, 1);
1084 }
1085
1086 /*
1087  * Restore the queue_if_no_path setting.
1088  */
1089 static void multipath_resume(struct dm_target *ti)
1090 {
1091         struct multipath *m = (struct multipath *) ti->private;
1092         unsigned long flags;
1093
1094         spin_lock_irqsave(&m->lock, flags);
1095         m->queue_if_no_path = m->saved_queue_if_no_path;
1096         spin_unlock_irqrestore(&m->lock, flags);
1097 }
1098
1099 /*
1100  * Info output has the following format:
1101  * num_multipath_feature_args [multipath_feature_args]*
1102  * num_handler_status_args [handler_status_args]*
1103  * num_groups init_group_number
1104  *            [A|D|E num_ps_status_args [ps_status_args]*
1105  *             num_paths num_selector_args
1106  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1107  *
1108  * Table output has the following format (identical to the constructor string):
1109  * num_feature_args [features_args]*
1110  * num_handler_args hw_handler [hw_handler_args]*
1111  * num_groups init_group_number
1112  *     [priority selector-name num_ps_args [ps_args]*
1113  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1114  */
1115 static int multipath_status(struct dm_target *ti, status_type_t type,
1116                             char *result, unsigned int maxlen)
1117 {
1118         int sz = 0;
1119         unsigned long flags;
1120         struct multipath *m = (struct multipath *) ti->private;
1121         struct hw_handler *hwh = &m->hw_handler;
1122         struct priority_group *pg;
1123         struct pgpath *p;
1124         unsigned pg_num;
1125         char state;
1126
1127         spin_lock_irqsave(&m->lock, flags);
1128
1129         /* Features */
1130         if (type == STATUSTYPE_INFO)
1131                 DMEMIT("1 %u ", m->queue_size);
1132         else if (m->queue_if_no_path)
1133                 DMEMIT("1 queue_if_no_path ");
1134         else
1135                 DMEMIT("0 ");
1136
1137         if (hwh->type && hwh->type->status)
1138                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1139         else if (!hwh->type || type == STATUSTYPE_INFO)
1140                 DMEMIT("0 ");
1141         else
1142                 DMEMIT("1 %s ", hwh->type->name);
1143
1144         DMEMIT("%u ", m->nr_priority_groups);
1145
1146         if (m->next_pg)
1147                 pg_num = m->next_pg->pg_num;
1148         else if (m->current_pg)
1149                 pg_num = m->current_pg->pg_num;
1150         else
1151                         pg_num = 1;
1152
1153         DMEMIT("%u ", pg_num);
1154
1155         switch (type) {
1156         case STATUSTYPE_INFO:
1157                 list_for_each_entry(pg, &m->priority_groups, list) {
1158                         if (pg->bypassed)
1159                                 state = 'D';    /* Disabled */
1160                         else if (pg == m->current_pg)
1161                                 state = 'A';    /* Currently Active */
1162                         else
1163                                 state = 'E';    /* Enabled */
1164
1165                         DMEMIT("%c ", state);
1166
1167                         if (pg->ps.type->status)
1168                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1169                                                           result + sz,
1170                                                           maxlen - sz);
1171                         else
1172                                 DMEMIT("0 ");
1173
1174                         DMEMIT("%u %u ", pg->nr_pgpaths,
1175                                pg->ps.type->info_args);
1176
1177                         list_for_each_entry(p, &pg->pgpaths, list) {
1178                                 DMEMIT("%s %s %u ", p->path.dev->name,
1179                                        p->path.is_active ? "A" : "F",
1180                                        p->fail_count);
1181                                 if (pg->ps.type->status)
1182                                         sz += pg->ps.type->status(&pg->ps,
1183                                               &p->path, type, result + sz,
1184                                               maxlen - sz);
1185                         }
1186                 }
1187                 break;
1188
1189         case STATUSTYPE_TABLE:
1190                 list_for_each_entry(pg, &m->priority_groups, list) {
1191                         DMEMIT("%s ", pg->ps.type->name);
1192
1193                         if (pg->ps.type->status)
1194                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1195                                                           result + sz,
1196                                                           maxlen - sz);
1197                         else
1198                                 DMEMIT("0 ");
1199
1200                         DMEMIT("%u %u ", pg->nr_pgpaths,
1201                                pg->ps.type->table_args);
1202
1203                         list_for_each_entry(p, &pg->pgpaths, list) {
1204                                 DMEMIT("%s ", p->path.dev->name);
1205                                 if (pg->ps.type->status)
1206                                         sz += pg->ps.type->status(&pg->ps,
1207                                               &p->path, type, result + sz,
1208                                               maxlen - sz);
1209                         }
1210                 }
1211                 break;
1212         }
1213
1214         spin_unlock_irqrestore(&m->lock, flags);
1215
1216         return 0;
1217 }
1218
1219 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1220 {
1221         int r;
1222         struct dm_dev *dev;
1223         struct multipath *m = (struct multipath *) ti->private;
1224         action_fn action;
1225
1226         if (argc == 1) {
1227                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1228                         return queue_if_no_path(m, 1, 0);
1229                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1230                         return queue_if_no_path(m, 0, 0);
1231         }
1232
1233         if (argc != 2)
1234                 goto error;
1235
1236         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1237                 return bypass_pg_num(m, argv[1], 1);
1238         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1239                 return bypass_pg_num(m, argv[1], 0);
1240         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1241                 return switch_pg_num(m, argv[1]);
1242         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1243                 action = reinstate_path;
1244         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1245                 action = fail_path;
1246         else
1247                 goto error;
1248
1249         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1250                           dm_table_get_mode(ti->table), &dev);
1251         if (r) {
1252                 DMWARN("message: error getting device %s",
1253                        argv[1]);
1254                 return -EINVAL;
1255         }
1256
1257         r = action_dev(m, dev, action);
1258
1259         dm_put_device(ti, dev);
1260
1261         return r;
1262
1263 error:
1264         DMWARN("Unrecognised multipath message received.");
1265         return -EINVAL;
1266 }
1267
1268 /*-----------------------------------------------------------------
1269  * Module setup
1270  *---------------------------------------------------------------*/
1271 static struct target_type multipath_target = {
1272         .name = "multipath",
1273         .version = {1, 0, 4},
1274         .module = THIS_MODULE,
1275         .ctr = multipath_ctr,
1276         .dtr = multipath_dtr,
1277         .map = multipath_map,
1278         .end_io = multipath_end_io,
1279         .presuspend = multipath_presuspend,
1280         .resume = multipath_resume,
1281         .status = multipath_status,
1282         .message = multipath_message,
1283 };
1284
1285 static int __init dm_multipath_init(void)
1286 {
1287         int r;
1288
1289         /* allocate a slab for the dm_ios */
1290         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1291                                         0, 0, NULL, NULL);
1292         if (!_mpio_cache)
1293                 return -ENOMEM;
1294
1295         r = dm_register_target(&multipath_target);
1296         if (r < 0) {
1297                 DMERR("%s: register failed %d", multipath_target.name, r);
1298                 kmem_cache_destroy(_mpio_cache);
1299                 return -EINVAL;
1300         }
1301
1302         kmultipathd = create_workqueue("kmpathd");
1303         if (!kmultipathd) {
1304                 DMERR("%s: failed to create workqueue kmpathd",
1305                                 multipath_target.name);
1306                 dm_unregister_target(&multipath_target);
1307                 kmem_cache_destroy(_mpio_cache);
1308                 return -ENOMEM;
1309         }
1310
1311         DMINFO("version %u.%u.%u loaded",
1312                multipath_target.version[0], multipath_target.version[1],
1313                multipath_target.version[2]);
1314
1315         return r;
1316 }
1317
1318 static void __exit dm_multipath_exit(void)
1319 {
1320         int r;
1321
1322         destroy_workqueue(kmultipathd);
1323
1324         r = dm_unregister_target(&multipath_target);
1325         if (r < 0)
1326                 DMERR("%s: target unregister failed %d",
1327                       multipath_target.name, r);
1328         kmem_cache_destroy(_mpio_cache);
1329 }
1330
1331 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1332
1333 module_init(dm_multipath_init);
1334 module_exit(dm_multipath_exit);
1335
1336 MODULE_DESCRIPTION(DM_NAME " multipath target");
1337 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1338 MODULE_LICENSE("GPL");