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