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         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  67
  68         unsigned nr_valid_paths;        /* Total number of usable paths */
  69         struct pgpath *current_pgpath;
  70         struct priority_group *current_pg;
  71         struct priority_group *next_pg; /* Switch to this PG if set */
  72         unsigned repeat_count;          /* I/Os left before calling PS again */
  73
  74         unsigned queue_io;              /* Must we queue all I/O? */
  75         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  76         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  77
  78         struct work_struct process_queued_ios;
  79         struct bio_list queued_ios;
  80         unsigned queue_size;
  81
  82         struct work_struct trigger_event;
  83
  84         /*
  85          * We must use a mempool of mpath_io structs so that we
  86          * can resubmit bios on error.
  87          */
  88         mempool_t *mpio_pool;
  89 };
  90
  91 /*
  92  * Context information attached to each bio we process.
  93  */
  94 struct mpath_io {
  95         struct pgpath *pgpath;
  96         struct dm_bio_details details;
  97 };
  98
  99 typedef int (*action_fn) (struct pgpath *pgpath);
 100
 101 #define MIN_IOS 256     /* Mempool size */
 102
 103 static kmem_cache_t *_mpio_cache;
 104
 105 struct workqueue_struct *kmultipathd;
 106 static void process_queued_ios(void *data);
 107 static void trigger_event(void *data);
 108
 109
 110 /*-----------------------------------------------
 111  * Allocation routines
 112  *-----------------------------------------------*/
 113
 114 static struct pgpath *alloc_pgpath(void)
 115 {
 116         struct pgpath *pgpath = kmalloc(sizeof(*pgpath), GFP_KERNEL);
 117
 118         if (pgpath) {
 119                 memset(pgpath, 0, sizeof(*pgpath));
 120                 pgpath->path.is_active = 1;
 121         }
 122
 123         return pgpath;
 124 }
 125
 126 static inline void free_pgpath(struct pgpath *pgpath)
 127 {
 128         kfree(pgpath);
 129 }
 130
 131 static struct priority_group *alloc_priority_group(void)
 132 {
 133         struct priority_group *pg;
 134
 135         pg = kmalloc(sizeof(*pg), GFP_KERNEL);
 136         if (!pg)
 137                 return NULL;
 138
 139         memset(pg, 0, sizeof(*pg));
 140         INIT_LIST_HEAD(&pg->pgpaths);
 141
 142         return pg;
 143 }
 144
 145 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 146 {
 147         struct pgpath *pgpath, *tmp;
 148
 149         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 150                 list_del(&pgpath->list);
 151                 dm_put_device(ti, pgpath->path.dev);
 152                 free_pgpath(pgpath);
 153         }
 154 }
 155
 156 static void free_priority_group(struct priority_group *pg,
 157                                 struct dm_target *ti)
 158 {
 159         struct path_selector *ps = &pg->ps;
 160
 161         if (ps->type) {
 162                 ps->type->destroy(ps);
 163                 dm_put_path_selector(ps->type);
 164         }
 165
 166         free_pgpaths(&pg->pgpaths, ti);
 167         kfree(pg);
 168 }
 169
 170 static struct multipath *alloc_multipath(void)
 171 {
 172         struct multipath *m;
 173
 174         m = kmalloc(sizeof(*m), GFP_KERNEL);
 175         if (m) {
 176                 memset(m, 0, sizeof(*m));
 177                 INIT_LIST_HEAD(&m->priority_groups);
 178                 spin_lock_init(&m->lock);
 179                 m->queue_io = 1;
 180                 INIT_WORK(&m->process_queued_ios, process_queued_ios, m);
 181                 INIT_WORK(&m->trigger_event, trigger_event, m);
 182                 m->mpio_pool = mempool_create(MIN_IOS, mempool_alloc_slab,
 183                                               mempool_free_slab, _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 #define ESTR(s) ("dm-multipath: " s)
 451
 452 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 453 {
 454         if (!str ||
 455             (sscanf(str, "%u", v) != 1) ||
 456             (*v < param->min) ||
 457             (*v > param->max)) {
 458                 *error = param->error;
 459                 return -EINVAL;
 460         }
 461
 462         return 0;
 463 }
 464
 465 struct arg_set {
 466         unsigned argc;
 467         char **argv;
 468 };
 469
 470 static char *shift(struct arg_set *as)
 471 {
 472         char *r;
 473
 474         if (as->argc) {
 475                 as->argc--;
 476                 r = *as->argv;
 477                 as->argv++;
 478                 return r;
 479         }
 480
 481         return NULL;
 482 }
 483
 484 static void consume(struct arg_set *as, unsigned n)
 485 {
 486         BUG_ON (as->argc < n);
 487         as->argc -= n;
 488         as->argv += n;
 489 }
 490
 491 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 492                                struct dm_target *ti)
 493 {
 494         int r;
 495         struct path_selector_type *pst;
 496         unsigned ps_argc;
 497
 498         static struct param _params[] = {
 499                 {0, 1024, ESTR("invalid number of path selector args")},
 500         };
 501
 502         pst = dm_get_path_selector(shift(as));
 503         if (!pst) {
 504                 ti->error = ESTR("unknown path selector type");
 505                 return -EINVAL;
 506         }
 507
 508         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 509         if (r)
 510                 return -EINVAL;
 511
 512         r = pst->create(&pg->ps, ps_argc, as->argv);
 513         if (r) {
 514                 dm_put_path_selector(pst);
 515                 ti->error = ESTR("path selector constructor failed");
 516                 return r;
 517         }
 518
 519         pg->ps.type = pst;
 520         consume(as, ps_argc);
 521
 522         return 0;
 523 }
 524
 525 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 526                                struct dm_target *ti)
 527 {
 528         int r;
 529         struct pgpath *p;
 530
 531         /* we need at least a path arg */
 532         if (as->argc < 1) {
 533                 ti->error = ESTR("no device given");
 534                 return NULL;
 535         }
 536
 537         p = alloc_pgpath();
 538         if (!p)
 539                 return NULL;
 540
 541         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 542                           dm_table_get_mode(ti->table), &p->path.dev);
 543         if (r) {
 544                 ti->error = ESTR("error getting device");
 545                 goto bad;
 546         }
 547
 548         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 549         if (r) {
 550                 dm_put_device(ti, p->path.dev);
 551                 goto bad;
 552         }
 553
 554         return p;
 555
 556  bad:
 557         free_pgpath(p);
 558         return NULL;
 559 }
 560
 561 static struct priority_group *parse_priority_group(struct arg_set *as,
 562                                                    struct multipath *m,
 563                                                    struct dm_target *ti)
 564 {
 565         static struct param _params[] = {
 566                 {1, 1024, ESTR("invalid number of paths")},
 567                 {0, 1024, ESTR("invalid number of selector args")}
 568         };
 569
 570         int r;
 571         unsigned i, nr_selector_args, nr_params;
 572         struct priority_group *pg;
 573
 574         if (as->argc < 2) {
 575                 as->argc = 0;
 576                 ti->error = ESTR("not enough priority group aruments");
 577                 return NULL;
 578         }
 579
 580         pg = alloc_priority_group();
 581         if (!pg) {
 582                 ti->error = ESTR("couldn't allocate priority group");
 583                 return NULL;
 584         }
 585         pg->m = m;
 586
 587         r = parse_path_selector(as, pg, ti);
 588         if (r)
 589                 goto bad;
 590
 591         /*
 592          * read the paths
 593          */
 594         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 595         if (r)
 596                 goto bad;
 597
 598         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 599         if (r)
 600                 goto bad;
 601
 602         nr_params = 1 + nr_selector_args;
 603         for (i = 0; i < pg->nr_pgpaths; i++) {
 604                 struct pgpath *pgpath;
 605                 struct arg_set path_args;
 606
 607                 if (as->argc < nr_params)
 608                         goto bad;
 609
 610                 path_args.argc = nr_params;
 611                 path_args.argv = as->argv;
 612
 613                 pgpath = parse_path(&path_args, &pg->ps, ti);
 614                 if (!pgpath)
 615                         goto bad;
 616
 617                 pgpath->pg = pg;
 618                 list_add_tail(&pgpath->list, &pg->pgpaths);
 619                 consume(as, nr_params);
 620         }
 621
 622         return pg;
 623
 624  bad:
 625         free_priority_group(pg, ti);
 626         return NULL;
 627 }
 628
 629 static int parse_hw_handler(struct arg_set *as, struct multipath *m,
 630                             struct dm_target *ti)
 631 {
 632         int r;
 633         struct hw_handler_type *hwht;
 634         unsigned hw_argc;
 635
 636         static struct param _params[] = {
 637                 {0, 1024, ESTR("invalid number of hardware handler args")},
 638         };
 639
 640         r = read_param(_params, shift(as), &hw_argc, &ti->error);
 641         if (r)
 642                 return -EINVAL;
 643
 644         if (!hw_argc)
 645                 return 0;
 646
 647         hwht = dm_get_hw_handler(shift(as));
 648         if (!hwht) {
 649                 ti->error = ESTR("unknown hardware handler type");
 650                 return -EINVAL;
 651         }
 652
 653         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
 654         if (r) {
 655                 dm_put_hw_handler(hwht);
 656                 ti->error = ESTR("hardware handler constructor failed");
 657                 return r;
 658         }
 659
 660         m->hw_handler.type = hwht;
 661         consume(as, hw_argc - 1);
 662
 663         return 0;
 664 }
 665
 666 static int parse_features(struct arg_set *as, struct multipath *m,
 667                           struct dm_target *ti)
 668 {
 669         int r;
 670         unsigned argc;
 671
 672         static struct param _params[] = {
 673                 {0, 1, ESTR("invalid number of feature args")},
 674         };
 675
 676         r = read_param(_params, shift(as), &argc, &ti->error);
 677         if (r)
 678                 return -EINVAL;
 679
 680         if (!argc)
 681                 return 0;
 682
 683         if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
 684                 return queue_if_no_path(m, 1, 0);
 685         else {
 686                 ti->error = "Unrecognised multipath feature request";
 687                 return -EINVAL;
 688         }
 689 }
 690
 691 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 692                          char **argv)
 693 {
 694         /* target parameters */
 695         static struct param _params[] = {
 696                 {1, 1024, ESTR("invalid number of priority groups")},
 697                 {1, 1024, ESTR("invalid initial priority group number")},
 698         };
 699
 700         int r;
 701         struct multipath *m;
 702         struct arg_set as;
 703         unsigned pg_count = 0;
 704         unsigned next_pg_num;
 705
 706         as.argc = argc;
 707         as.argv = argv;
 708
 709         m = alloc_multipath();
 710         if (!m) {
 711                 ti->error = ESTR("can't allocate multipath");
 712                 return -EINVAL;
 713         }
 714
 715         m->ti = ti;
 716
 717         r = parse_features(&as, m, ti);
 718         if (r)
 719                 goto bad;
 720
 721         r = parse_hw_handler(&as, m, ti);
 722         if (r)
 723                 goto bad;
 724
 725         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 726         if (r)
 727                 goto bad;
 728
 729         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 730         if (r)
 731                 goto bad;
 732
 733         /* parse the priority groups */
 734         while (as.argc) {
 735                 struct priority_group *pg;
 736
 737                 pg = parse_priority_group(&as, m, ti);
 738                 if (!pg) {
 739                         r = -EINVAL;
 740                         goto bad;
 741                 }
 742
 743                 m->nr_valid_paths += pg->nr_pgpaths;
 744                 list_add_tail(&pg->list, &m->priority_groups);
 745                 pg_count++;
 746                 pg->pg_num = pg_count;
 747                 if (!--next_pg_num)
 748                         m->next_pg = pg;
 749         }
 750
 751         if (pg_count != m->nr_priority_groups) {
 752                 ti->error = ESTR("priority group count mismatch");
 753                 r = -EINVAL;
 754                 goto bad;
 755         }
 756
 757         ti->private = m;
 758
 759         return 0;
 760
 761  bad:
 762         free_multipath(m);
 763         return r;
 764 }
 765
 766 static void multipath_dtr(struct dm_target *ti)
 767 {
 768         struct multipath *m = (struct multipath *) ti->private;
 769
 770         flush_workqueue(kmultipathd);
 771         free_multipath(m);
 772 }
 773
 774 /*
 775  * Map bios, recording original fields for later in case we have to resubmit
 776  */
 777 static int multipath_map(struct dm_target *ti, struct bio *bio,
 778                          union map_info *map_context)
 779 {
 780         int r;
 781         struct mpath_io *mpio;
 782         struct multipath *m = (struct multipath *) ti->private;
 783
 784         if (bio_barrier(bio))
 785                 return -EOPNOTSUPP;
 786
 787         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 788         dm_bio_record(&mpio->details, bio);
 789
 790         map_context->ptr = mpio;
 791         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 792         r = map_io(m, bio, mpio, 0);
 793         if (r < 0)
 794                 mempool_free(mpio, m->mpio_pool);
 795
 796         return r;
 797 }
 798
 799 /*
 800  * Take a path out of use.
 801  */
 802 static int fail_path(struct pgpath *pgpath)
 803 {
 804         unsigned long flags;
 805         struct multipath *m = pgpath->pg->m;
 806
 807         spin_lock_irqsave(&m->lock, flags);
 808
 809         if (!pgpath->path.is_active)
 810                 goto out;
 811
 812         DMWARN("dm-multipath: Failing path %s.", pgpath->path.dev->name);
 813
 814         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 815         pgpath->path.is_active = 0;
 816         pgpath->fail_count++;
 817
 818         m->nr_valid_paths--;
 819
 820         if (pgpath == m->current_pgpath)
 821                 m->current_pgpath = NULL;
 822
 823         queue_work(kmultipathd, &m->trigger_event);
 824
 825 out:
 826         spin_unlock_irqrestore(&m->lock, flags);
 827
 828         return 0;
 829 }
 830
 831 /*
 832  * Reinstate a previously-failed path
 833  */
 834 static int reinstate_path(struct pgpath *pgpath)
 835 {
 836         int r = 0;
 837         unsigned long flags;
 838         struct multipath *m = pgpath->pg->m;
 839
 840         spin_lock_irqsave(&m->lock, flags);
 841
 842         if (pgpath->path.is_active)
 843                 goto out;
 844
 845         if (!pgpath->pg->ps.type) {
 846                 DMWARN("Reinstate path not supported by path selector %s",
 847                        pgpath->pg->ps.type->name);
 848                 r = -EINVAL;
 849                 goto out;
 850         }
 851
 852         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 853         if (r)
 854                 goto out;
 855
 856         pgpath->path.is_active = 1;
 857
 858         m->current_pgpath = NULL;
 859         if (!m->nr_valid_paths++ && m->queue_size)
 860                 queue_work(kmultipathd, &m->process_queued_ios);
 861
 862         queue_work(kmultipathd, &m->trigger_event);
 863
 864 out:
 865         spin_unlock_irqrestore(&m->lock, flags);
 866
 867         return r;
 868 }
 869
 870 /*
 871  * Fail or reinstate all paths that match the provided struct dm_dev.
 872  */
 873 static int action_dev(struct multipath *m, struct dm_dev *dev,
 874                       action_fn action)
 875 {
 876         int r = 0;
 877         struct pgpath *pgpath;
 878         struct priority_group *pg;
 879
 880         list_for_each_entry(pg, &m->priority_groups, list) {
 881                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 882                         if (pgpath->path.dev == dev)
 883                                 r = action(pgpath);
 884                 }
 885         }
 886
 887         return r;
 888 }
 889
 890 /*
 891  * Temporarily try to avoid having to use the specified PG
 892  */
 893 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 894                       int bypassed)
 895 {
 896         unsigned long flags;
 897
 898         spin_lock_irqsave(&m->lock, flags);
 899
 900         pg->bypassed = bypassed;
 901         m->current_pgpath = NULL;
 902         m->current_pg = NULL;
 903
 904         spin_unlock_irqrestore(&m->lock, flags);
 905
 906         queue_work(kmultipathd, &m->trigger_event);
 907 }
 908
 909 /*
 910  * Switch to using the specified PG from the next I/O that gets mapped
 911  */
 912 static int switch_pg_num(struct multipath *m, const char *pgstr)
 913 {
 914         struct priority_group *pg;
 915         unsigned pgnum;
 916         unsigned long flags;
 917
 918         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 919             (pgnum > m->nr_priority_groups)) {
 920                 DMWARN("invalid PG number supplied to switch_pg_num");
 921                 return -EINVAL;
 922         }
 923
 924         spin_lock_irqsave(&m->lock, flags);
 925         list_for_each_entry(pg, &m->priority_groups, list) {
 926                 pg->bypassed = 0;
 927                 if (--pgnum)
 928                         continue;
 929
 930                 m->current_pgpath = NULL;
 931                 m->current_pg = NULL;
 932                 m->next_pg = pg;
 933         }
 934         spin_unlock_irqrestore(&m->lock, flags);
 935
 936         queue_work(kmultipathd, &m->trigger_event);
 937         return 0;
 938 }
 939
 940 /*
 941  * Set/clear bypassed status of a PG.
 942  * PGs are numbered upwards from 1 in the order they were declared.
 943  */
 944 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 945 {
 946         struct priority_group *pg;
 947         unsigned pgnum;
 948
 949         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 950             (pgnum > m->nr_priority_groups)) {
 951                 DMWARN("invalid PG number supplied to bypass_pg");
 952                 return -EINVAL;
 953         }
 954
 955         list_for_each_entry(pg, &m->priority_groups, list) {
 956                 if (!--pgnum)
 957                         break;
 958         }
 959
 960         bypass_pg(m, pg, bypassed);
 961         return 0;
 962 }
 963
 964 /*
 965  * pg_init must call this when it has completed its initialisation
 966  */
 967 void dm_pg_init_complete(struct path *path, unsigned err_flags)
 968 {
 969         struct pgpath *pgpath = path_to_pgpath(path);
 970         struct priority_group *pg = pgpath->pg;
 971         struct multipath *m = pg->m;
 972         unsigned long flags;
 973
 974         /* We insist on failing the path if the PG is already bypassed. */
 975         if (err_flags && pg->bypassed)
 976                 err_flags |= MP_FAIL_PATH;
 977
 978         if (err_flags & MP_FAIL_PATH)
 979                 fail_path(pgpath);
 980
 981         if (err_flags & MP_BYPASS_PG)
 982                 bypass_pg(m, pg, 1);
 983
 984         spin_lock_irqsave(&m->lock, flags);
 985         if (err_flags) {
 986                 m->current_pgpath = NULL;
 987                 m->current_pg = NULL;
 988         } else if (!m->pg_init_required)
 989                 m->queue_io = 0;
 990
 991         m->pg_init_in_progress = 0;
 992         queue_work(kmultipathd, &m->process_queued_ios);
 993         spin_unlock_irqrestore(&m->lock, flags);
 994 }
 995
 996 /*
 997  * end_io handling
 998  */
 999 static int do_end_io(struct multipath *m, struct bio *bio,
1000                      int error, struct mpath_io *mpio)
1001 {
1002         struct hw_handler *hwh = &m->hw_handler;
1003         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
1004         unsigned long flags;
1005
1006         if (!error)
1007                 return 0;       /* I/O complete */
1008
1009         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1010                 return error;
1011
1012         if (error == -EOPNOTSUPP)
1013                 return error;
1014
1015         spin_lock_irqsave(&m->lock, flags);
1016         if (!m->nr_valid_paths) {
1017                 if (!m->queue_if_no_path) {
1018                         spin_unlock_irqrestore(&m->lock, flags);
1019                         return -EIO;
1020                 } else {
1021                         spin_unlock_irqrestore(&m->lock, flags);
1022                         goto requeue;
1023                 }
1024         }
1025         spin_unlock_irqrestore(&m->lock, flags);
1026
1027         if (hwh->type && hwh->type->error)
1028                 err_flags = hwh->type->error(hwh, bio);
1029
1030         if (mpio->pgpath) {
1031                 if (err_flags & MP_FAIL_PATH)
1032                         fail_path(mpio->pgpath);
1033
1034                 if (err_flags & MP_BYPASS_PG)
1035                         bypass_pg(m, mpio->pgpath->pg, 1);
1036         }
1037
1038         if (err_flags & MP_ERROR_IO)
1039                 return -EIO;
1040
1041       requeue:
1042         dm_bio_restore(&mpio->details, bio);
1043
1044         /* queue for the daemon to resubmit or fail */
1045         spin_lock_irqsave(&m->lock, flags);
1046         bio_list_add(&m->queued_ios, bio);
1047         m->queue_size++;
1048         if (!m->queue_io)
1049                 queue_work(kmultipathd, &m->process_queued_ios);
1050         spin_unlock_irqrestore(&m->lock, flags);
1051
1052         return 1;       /* io not complete */
1053 }
1054
1055 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1056                             int error, union map_info *map_context)
1057 {
1058         struct multipath *m = (struct multipath *) ti->private;
1059         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1060         struct pgpath *pgpath = mpio->pgpath;
1061         struct path_selector *ps;
1062         int r;
1063
1064         r  = do_end_io(m, bio, error, mpio);
1065         if (pgpath) {
1066                 ps = &pgpath->pg->ps;
1067                 if (ps->type->end_io)
1068                         ps->type->end_io(ps, &pgpath->path);
1069         }
1070         if (r <= 0)
1071                 mempool_free(mpio, m->mpio_pool);
1072
1073         return r;
1074 }
1075
1076 /*
1077  * Suspend can't complete until all the I/O is processed so if
1078  * the last path fails we must error any remaining I/O.
1079  * Note that if the freeze_bdev fails while suspending, the
1080  * queue_if_no_path state is lost - userspace should reset it.
1081  */
1082 static void multipath_presuspend(struct dm_target *ti)
1083 {
1084         struct multipath *m = (struct multipath *) ti->private;
1085
1086         queue_if_no_path(m, 0, 1);
1087 }
1088
1089 /*
1090  * Restore the queue_if_no_path setting.
1091  */
1092 static void multipath_resume(struct dm_target *ti)
1093 {
1094         struct multipath *m = (struct multipath *) ti->private;
1095         unsigned long flags;
1096
1097         spin_lock_irqsave(&m->lock, flags);
1098         m->queue_if_no_path = m->saved_queue_if_no_path;
1099         spin_unlock_irqrestore(&m->lock, flags);
1100 }
1101
1102 /*
1103  * Info output has the following format:
1104  * num_multipath_feature_args [multipath_feature_args]*
1105  * num_handler_status_args [handler_status_args]*
1106  * num_groups init_group_number
1107  *            [A|D|E num_ps_status_args [ps_status_args]*
1108  *             num_paths num_selector_args
1109  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1110  *
1111  * Table output has the following format (identical to the constructor string):
1112  * num_feature_args [features_args]*
1113  * num_handler_args hw_handler [hw_handler_args]*
1114  * num_groups init_group_number
1115  *     [priority selector-name num_ps_args [ps_args]*
1116  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1117  */
1118 static int multipath_status(struct dm_target *ti, status_type_t type,
1119                             char *result, unsigned int maxlen)
1120 {
1121         int sz = 0;
1122         unsigned long flags;
1123         struct multipath *m = (struct multipath *) ti->private;
1124         struct hw_handler *hwh = &m->hw_handler;
1125         struct priority_group *pg;
1126         struct pgpath *p;
1127         unsigned pg_num;
1128         char state;
1129
1130         spin_lock_irqsave(&m->lock, flags);
1131
1132         /* Features */
1133         if (type == STATUSTYPE_INFO)
1134                 DMEMIT("1 %u ", m->queue_size);
1135         else if (m->queue_if_no_path)
1136                 DMEMIT("1 queue_if_no_path ");
1137         else
1138                 DMEMIT("0 ");
1139
1140         if (hwh->type && hwh->type->status)
1141                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1142         else if (!hwh->type || type == STATUSTYPE_INFO)
1143                 DMEMIT("0 ");
1144         else
1145                 DMEMIT("1 %s ", hwh->type->name);
1146
1147         DMEMIT("%u ", m->nr_priority_groups);
1148
1149         if (m->next_pg)
1150                 pg_num = m->next_pg->pg_num;
1151         else if (m->current_pg)
1152                 pg_num = m->current_pg->pg_num;
1153         else
1154                         pg_num = 1;
1155
1156         DMEMIT("%u ", pg_num);
1157
1158         switch (type) {
1159         case STATUSTYPE_INFO:
1160                 list_for_each_entry(pg, &m->priority_groups, list) {
1161                         if (pg->bypassed)
1162                                 state = 'D';    /* Disabled */
1163                         else if (pg == m->current_pg)
1164                                 state = 'A';    /* Currently Active */
1165                         else
1166                                 state = 'E';    /* Enabled */
1167
1168                         DMEMIT("%c ", state);
1169
1170                         if (pg->ps.type->status)
1171                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1172                                                           result + sz,
1173                                                           maxlen - sz);
1174                         else
1175                                 DMEMIT("0 ");
1176
1177                         DMEMIT("%u %u ", pg->nr_pgpaths,
1178                                pg->ps.type->info_args);
1179
1180                         list_for_each_entry(p, &pg->pgpaths, list) {
1181                                 DMEMIT("%s %s %u ", p->path.dev->name,
1182                                        p->path.is_active ? "A" : "F",
1183                                        p->fail_count);
1184                                 if (pg->ps.type->status)
1185                                         sz += pg->ps.type->status(&pg->ps,
1186                                               &p->path, type, result + sz,
1187                                               maxlen - sz);
1188                         }
1189                 }
1190                 break;
1191
1192         case STATUSTYPE_TABLE:
1193                 list_for_each_entry(pg, &m->priority_groups, list) {
1194                         DMEMIT("%s ", pg->ps.type->name);
1195
1196                         if (pg->ps.type->status)
1197                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1198                                                           result + sz,
1199                                                           maxlen - sz);
1200                         else
1201                                 DMEMIT("0 ");
1202
1203                         DMEMIT("%u %u ", pg->nr_pgpaths,
1204                                pg->ps.type->table_args);
1205
1206                         list_for_each_entry(p, &pg->pgpaths, list) {
1207                                 DMEMIT("%s ", p->path.dev->name);
1208                                 if (pg->ps.type->status)
1209                                         sz += pg->ps.type->status(&pg->ps,
1210                                               &p->path, type, result + sz,
1211                                               maxlen - sz);
1212                         }
1213                 }
1214                 break;
1215         }
1216
1217         spin_unlock_irqrestore(&m->lock, flags);
1218
1219         return 0;
1220 }
1221
1222 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1223 {
1224         int r;
1225         struct dm_dev *dev;
1226         struct multipath *m = (struct multipath *) ti->private;
1227         action_fn action;
1228
1229         if (argc == 1) {
1230                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1231                         return queue_if_no_path(m, 1, 0);
1232                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1233                         return queue_if_no_path(m, 0, 0);
1234         }
1235
1236         if (argc != 2)
1237                 goto error;
1238
1239         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1240                 return bypass_pg_num(m, argv[1], 1);
1241         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1242                 return bypass_pg_num(m, argv[1], 0);
1243         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1244                 return switch_pg_num(m, argv[1]);
1245         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1246                 action = reinstate_path;
1247         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1248                 action = fail_path;
1249         else
1250                 goto error;
1251
1252         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1253                           dm_table_get_mode(ti->table), &dev);
1254         if (r) {
1255                 DMWARN("dm-multipath message: error getting device %s",
1256                        argv[1]);
1257                 return -EINVAL;
1258         }
1259
1260         r = action_dev(m, dev, action);
1261
1262         dm_put_device(ti, dev);
1263
1264         return r;
1265
1266 error:
1267         DMWARN("Unrecognised multipath message received.");
1268         return -EINVAL;
1269 }
1270
1271 /*-----------------------------------------------------------------
1272  * Module setup
1273  *---------------------------------------------------------------*/
1274 static struct target_type multipath_target = {
1275         .name = "multipath",
1276         .version = {1, 0, 4},
1277         .module = THIS_MODULE,
1278         .ctr = multipath_ctr,
1279         .dtr = multipath_dtr,
1280         .map = multipath_map,
1281         .end_io = multipath_end_io,
1282         .presuspend = multipath_presuspend,
1283         .resume = multipath_resume,
1284         .status = multipath_status,
1285         .message = multipath_message,
1286 };
1287
1288 static int __init dm_multipath_init(void)
1289 {
1290         int r;
1291
1292         /* allocate a slab for the dm_ios */
1293         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1294                                         0, 0, NULL, NULL);
1295         if (!_mpio_cache)
1296                 return -ENOMEM;
1297
1298         r = dm_register_target(&multipath_target);
1299         if (r < 0) {
1300                 DMERR("%s: register failed %d", multipath_target.name, r);
1301                 kmem_cache_destroy(_mpio_cache);
1302                 return -EINVAL;
1303         }
1304
1305         kmultipathd = create_workqueue("kmpathd");
1306         if (!kmultipathd) {
1307                 DMERR("%s: failed to create workqueue kmpathd",
1308                                 multipath_target.name);
1309                 dm_unregister_target(&multipath_target);
1310                 kmem_cache_destroy(_mpio_cache);
1311                 return -ENOMEM;
1312         }
1313
1314         DMINFO("dm-multipath version %u.%u.%u loaded",
1315                multipath_target.version[0], multipath_target.version[1],
1316                multipath_target.version[2]);
1317
1318         return r;
1319 }
1320
1321 static void __exit dm_multipath_exit(void)
1322 {
1323         int r;
1324
1325         destroy_workqueue(kmultipathd);
1326
1327         r = dm_unregister_target(&multipath_target);
1328         if (r < 0)
1329                 DMERR("%s: target unregister failed %d",
1330                       multipath_target.name, r);
1331         kmem_cache_destroy(_mpio_cache);
1332 }
1333
1334 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1335
1336 module_init(dm_multipath_init);
1337 module_exit(dm_multipath_exit);
1338
1339 MODULE_DESCRIPTION(DM_NAME " multipath target");
1340 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1341 MODULE_LICENSE("GPL");