git.oblomov.eu Git - linux-2.6/blob - drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include "dm.h"
   9 #include "dm-path-selector.h"
  10 #include "dm-hw-handler.h"
  11 #include "dm-bio-list.h"
  12 #include "dm-bio-record.h"
  13
  14 #include <linux/ctype.h>
  15 #include <linux/init.h>
  16 #include <linux/mempool.h>
  17 #include <linux/module.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/slab.h>
  20 #include <linux/time.h>
  21 #include <linux/workqueue.h>
  22 #include <asm/atomic.h>
  23
  24 #define 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 saved_queue_if_no_path;/* Saved state during suspension */
  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 struct workqueue_struct *kmultipathd;
 105 static void process_queued_ios(void *data);
 106 static void trigger_event(void *data);
 107
 108
 109 /*-----------------------------------------------
 110  * Allocation routines
 111  *-----------------------------------------------*/
 112
 113 static struct pgpath *alloc_pgpath(void)
 114 {
 115         struct pgpath *pgpath = kmalloc(sizeof(*pgpath), GFP_KERNEL);
 116
 117         if (pgpath) {
 118                 memset(pgpath, 0, sizeof(*pgpath));
 119                 pgpath->path.is_active = 1;
 120         }
 121
 122         return pgpath;
 123 }
 124
 125 static inline void free_pgpath(struct pgpath *pgpath)
 126 {
 127         kfree(pgpath);
 128 }
 129
 130 static struct priority_group *alloc_priority_group(void)
 131 {
 132         struct priority_group *pg;
 133
 134         pg = kmalloc(sizeof(*pg), GFP_KERNEL);
 135         if (!pg)
 136                 return NULL;
 137
 138         memset(pg, 0, sizeof(*pg));
 139         INIT_LIST_HEAD(&pg->pgpaths);
 140
 141         return pg;
 142 }
 143
 144 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 145 {
 146         struct pgpath *pgpath, *tmp;
 147
 148         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 149                 list_del(&pgpath->list);
 150                 dm_put_device(ti, pgpath->path.dev);
 151                 free_pgpath(pgpath);
 152         }
 153 }
 154
 155 static void free_priority_group(struct priority_group *pg,
 156                                 struct dm_target *ti)
 157 {
 158         struct path_selector *ps = &pg->ps;
 159
 160         if (ps->type) {
 161                 ps->type->destroy(ps);
 162                 dm_put_path_selector(ps->type);
 163         }
 164
 165         free_pgpaths(&pg->pgpaths, ti);
 166         kfree(pg);
 167 }
 168
 169 static struct multipath *alloc_multipath(void)
 170 {
 171         struct multipath *m;
 172
 173         m = kmalloc(sizeof(*m), GFP_KERNEL);
 174         if (m) {
 175                 memset(m, 0, sizeof(*m));
 176                 INIT_LIST_HEAD(&m->priority_groups);
 177                 spin_lock_init(&m->lock);
 178                 m->queue_io = 1;
 179                 INIT_WORK(&m->process_queued_ios, process_queued_ios, m);
 180                 INIT_WORK(&m->trigger_event, trigger_event, m);
 181                 m->mpio_pool = mempool_create(MIN_IOS, mempool_alloc_slab,
 182                                               mempool_free_slab, _mpio_cache);
 183                 if (!m->mpio_pool) {
 184                         kfree(m);
 185                         return NULL;
 186                 }
 187         }
 188
 189         return m;
 190 }
 191
 192 static void free_multipath(struct multipath *m)
 193 {
 194         struct priority_group *pg, *tmp;
 195         struct hw_handler *hwh = &m->hw_handler;
 196
 197         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 198                 list_del(&pg->list);
 199                 free_priority_group(pg, m->ti);
 200         }
 201
 202         if (hwh->type) {
 203                 hwh->type->destroy(hwh);
 204                 dm_put_hw_handler(hwh->type);
 205         }
 206
 207         mempool_destroy(m->mpio_pool);
 208         kfree(m);
 209 }
 210
 211
 212 /*-----------------------------------------------
 213  * Path selection
 214  *-----------------------------------------------*/
 215
 216 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 217 {
 218         struct hw_handler *hwh = &m->hw_handler;
 219
 220         m->current_pg = pgpath->pg;
 221
 222         /* Must we initialise the PG first, and queue I/O till it's ready? */
 223         if (hwh->type && hwh->type->pg_init) {
 224                 m->pg_init_required = 1;
 225                 m->queue_io = 1;
 226         } else {
 227                 m->pg_init_required = 0;
 228                 m->queue_io = 0;
 229         }
 230 }
 231
 232 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
 233 {
 234         struct path *path;
 235
 236         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
 237         if (!path)
 238                 return -ENXIO;
 239
 240         m->current_pgpath = path_to_pgpath(path);
 241
 242         if (m->current_pg != pg)
 243                 __switch_pg(m, m->current_pgpath);
 244
 245         return 0;
 246 }
 247
 248 static void __choose_pgpath(struct multipath *m)
 249 {
 250         struct priority_group *pg;
 251         unsigned bypassed = 1;
 252
 253         if (!m->nr_valid_paths)
 254                 goto failed;
 255
 256         /* Were we instructed to switch PG? */
 257         if (m->next_pg) {
 258                 pg = m->next_pg;
 259                 m->next_pg = NULL;
 260                 if (!__choose_path_in_pg(m, pg))
 261                         return;
 262         }
 263
 264         /* Don't change PG until it has no remaining paths */
 265         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
 266                 return;
 267
 268         /*
 269          * Loop through priority groups until we find a valid path.
 270          * First time we skip PGs marked 'bypassed'.
 271          * Second time we only try the ones we skipped.
 272          */
 273         do {
 274                 list_for_each_entry(pg, &m->priority_groups, list) {
 275                         if (pg->bypassed == bypassed)
 276                                 continue;
 277                         if (!__choose_path_in_pg(m, pg))
 278                                 return;
 279                 }
 280         } while (bypassed--);
 281
 282 failed:
 283         m->current_pgpath = NULL;
 284         m->current_pg = NULL;
 285 }
 286
 287 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
 288                   unsigned was_queued)
 289 {
 290         int r = 1;
 291         unsigned long flags;
 292         struct pgpath *pgpath;
 293
 294         spin_lock_irqsave(&m->lock, flags);
 295
 296         /* Do we need to select a new pgpath? */
 297         if (!m->current_pgpath ||
 298             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 299                 __choose_pgpath(m);
 300
 301         pgpath = m->current_pgpath;
 302
 303         if (was_queued)
 304                 m->queue_size--;
 305
 306         if ((pgpath && m->queue_io) ||
 307             (!pgpath && m->queue_if_no_path)) {
 308                 /* Queue for the daemon to resubmit */
 309                 bio_list_add(&m->queued_ios, bio);
 310                 m->queue_size++;
 311                 if (m->pg_init_required || !m->queue_io)
 312                         queue_work(kmultipathd, &m->process_queued_ios);
 313                 pgpath = NULL;
 314                 r = 0;
 315         } else if (!pgpath)
 316                 r = -EIO;               /* Failed */
 317         else
 318                 bio->bi_bdev = pgpath->path.dev->bdev;
 319
 320         mpio->pgpath = pgpath;
 321
 322         spin_unlock_irqrestore(&m->lock, flags);
 323
 324         return r;
 325 }
 326
 327 /*
 328  * If we run out of usable paths, should we queue I/O or error it?
 329  */
 330 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path)
 331 {
 332         unsigned long flags;
 333
 334         spin_lock_irqsave(&m->lock, flags);
 335
 336         m->saved_queue_if_no_path = m->queue_if_no_path;
 337         m->queue_if_no_path = queue_if_no_path;
 338         if (!m->queue_if_no_path)
 339                 queue_work(kmultipathd, &m->process_queued_ios);
 340
 341         spin_unlock_irqrestore(&m->lock, flags);
 342
 343         return 0;
 344 }
 345
 346 /*-----------------------------------------------------------------
 347  * The multipath daemon is responsible for resubmitting queued ios.
 348  *---------------------------------------------------------------*/
 349
 350 static void dispatch_queued_ios(struct multipath *m)
 351 {
 352         int r;
 353         unsigned long flags;
 354         struct bio *bio = NULL, *next;
 355         struct mpath_io *mpio;
 356         union map_info *info;
 357
 358         spin_lock_irqsave(&m->lock, flags);
 359         bio = bio_list_get(&m->queued_ios);
 360         spin_unlock_irqrestore(&m->lock, flags);
 361
 362         while (bio) {
 363                 next = bio->bi_next;
 364                 bio->bi_next = NULL;
 365
 366                 info = dm_get_mapinfo(bio);
 367                 mpio = info->ptr;
 368
 369                 r = map_io(m, bio, mpio, 1);
 370                 if (r < 0)
 371                         bio_endio(bio, bio->bi_size, r);
 372                 else if (r == 1)
 373                         generic_make_request(bio);
 374
 375                 bio = next;
 376         }
 377 }
 378
 379 static void process_queued_ios(void *data)
 380 {
 381         struct multipath *m = (struct multipath *) data;
 382         struct hw_handler *hwh = &m->hw_handler;
 383         struct pgpath *pgpath;
 384         unsigned init_required, must_queue = 0;
 385         unsigned long flags;
 386
 387         spin_lock_irqsave(&m->lock, flags);
 388
 389         if (!m->current_pgpath)
 390                 __choose_pgpath(m);
 391
 392         pgpath = m->current_pgpath;
 393
 394         if ((pgpath && m->queue_io) ||
 395             (!pgpath && m->queue_if_no_path))
 396                 must_queue = 1;
 397
 398         init_required = m->pg_init_required;
 399         if (init_required)
 400                 m->pg_init_required = 0;
 401
 402         spin_unlock_irqrestore(&m->lock, flags);
 403
 404         if (init_required)
 405                 hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
 406
 407         if (!must_queue)
 408                 dispatch_queued_ios(m);
 409 }
 410
 411 /*
 412  * An event is triggered whenever a path is taken out of use.
 413  * Includes path failure and PG bypass.
 414  */
 415 static void trigger_event(void *data)
 416 {
 417         struct multipath *m = (struct multipath *) data;
 418
 419         dm_table_event(m->ti->table);
 420 }
 421
 422 /*-----------------------------------------------------------------
 423  * Constructor/argument parsing:
 424  * <#multipath feature args> [<arg>]*
 425  * <#hw_handler args> [hw_handler [<arg>]*]
 426  * <#priority groups>
 427  * <initial priority group>
 428  *     [<selector> <#selector args> [<arg>]*
 429  *      <#paths> <#per-path selector args>
 430  *         [<path> [<arg>]* ]+ ]+
 431  *---------------------------------------------------------------*/
 432 struct param {
 433         unsigned min;
 434         unsigned max;
 435         char *error;
 436 };
 437
 438 #define ESTR(s) ("dm-multipath: " s)
 439
 440 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 441 {
 442         if (!str ||
 443             (sscanf(str, "%u", v) != 1) ||
 444             (*v < param->min) ||
 445             (*v > param->max)) {
 446                 *error = param->error;
 447                 return -EINVAL;
 448         }
 449
 450         return 0;
 451 }
 452
 453 struct arg_set {
 454         unsigned argc;
 455         char **argv;
 456 };
 457
 458 static char *shift(struct arg_set *as)
 459 {
 460         char *r;
 461
 462         if (as->argc) {
 463                 as->argc--;
 464                 r = *as->argv;
 465                 as->argv++;
 466                 return r;
 467         }
 468
 469         return NULL;
 470 }
 471
 472 static void consume(struct arg_set *as, unsigned n)
 473 {
 474         BUG_ON (as->argc < n);
 475         as->argc -= n;
 476         as->argv += n;
 477 }
 478
 479 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 480                                struct dm_target *ti)
 481 {
 482         int r;
 483         struct path_selector_type *pst;
 484         unsigned ps_argc;
 485
 486         static struct param _params[] = {
 487                 {0, 1024, ESTR("invalid number of path selector args")},
 488         };
 489
 490         pst = dm_get_path_selector(shift(as));
 491         if (!pst) {
 492                 ti->error = ESTR("unknown path selector type");
 493                 return -EINVAL;
 494         }
 495
 496         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 497         if (r)
 498                 return -EINVAL;
 499
 500         r = pst->create(&pg->ps, ps_argc, as->argv);
 501         if (r) {
 502                 dm_put_path_selector(pst);
 503                 ti->error = ESTR("path selector constructor failed");
 504                 return r;
 505         }
 506
 507         pg->ps.type = pst;
 508         consume(as, ps_argc);
 509
 510         return 0;
 511 }
 512
 513 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 514                                struct dm_target *ti)
 515 {
 516         int r;
 517         struct pgpath *p;
 518
 519         /* we need at least a path arg */
 520         if (as->argc < 1) {
 521                 ti->error = ESTR("no device given");
 522                 return NULL;
 523         }
 524
 525         p = alloc_pgpath();
 526         if (!p)
 527                 return NULL;
 528
 529         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 530                           dm_table_get_mode(ti->table), &p->path.dev);
 531         if (r) {
 532                 ti->error = ESTR("error getting device");
 533                 goto bad;
 534         }
 535
 536         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 537         if (r) {
 538                 dm_put_device(ti, p->path.dev);
 539                 goto bad;
 540         }
 541
 542         return p;
 543
 544  bad:
 545         free_pgpath(p);
 546         return NULL;
 547 }
 548
 549 static struct priority_group *parse_priority_group(struct arg_set *as,
 550                                                    struct multipath *m,
 551                                                    struct dm_target *ti)
 552 {
 553         static struct param _params[] = {
 554                 {1, 1024, ESTR("invalid number of paths")},
 555                 {0, 1024, ESTR("invalid number of selector args")}
 556         };
 557
 558         int r;
 559         unsigned i, nr_selector_args, nr_params;
 560         struct priority_group *pg;
 561
 562         if (as->argc < 2) {
 563                 as->argc = 0;
 564                 ti->error = ESTR("not enough priority group aruments");
 565                 return NULL;
 566         }
 567
 568         pg = alloc_priority_group();
 569         if (!pg) {
 570                 ti->error = ESTR("couldn't allocate priority group");
 571                 return NULL;
 572         }
 573         pg->m = m;
 574
 575         r = parse_path_selector(as, pg, ti);
 576         if (r)
 577                 goto bad;
 578
 579         /*
 580          * read the paths
 581          */
 582         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 583         if (r)
 584                 goto bad;
 585
 586         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 587         if (r)
 588                 goto bad;
 589
 590         nr_params = 1 + nr_selector_args;
 591         for (i = 0; i < pg->nr_pgpaths; i++) {
 592                 struct pgpath *pgpath;
 593                 struct arg_set path_args;
 594
 595                 if (as->argc < nr_params)
 596                         goto bad;
 597
 598                 path_args.argc = nr_params;
 599                 path_args.argv = as->argv;
 600
 601                 pgpath = parse_path(&path_args, &pg->ps, ti);
 602                 if (!pgpath)
 603                         goto bad;
 604
 605                 pgpath->pg = pg;
 606                 list_add_tail(&pgpath->list, &pg->pgpaths);
 607                 consume(as, nr_params);
 608         }
 609
 610         return pg;
 611
 612  bad:
 613         free_priority_group(pg, ti);
 614         return NULL;
 615 }
 616
 617 static int parse_hw_handler(struct arg_set *as, struct multipath *m,
 618                             struct dm_target *ti)
 619 {
 620         int r;
 621         struct hw_handler_type *hwht;
 622         unsigned hw_argc;
 623
 624         static struct param _params[] = {
 625                 {0, 1024, ESTR("invalid number of hardware handler args")},
 626         };
 627
 628         r = read_param(_params, shift(as), &hw_argc, &ti->error);
 629         if (r)
 630                 return -EINVAL;
 631
 632         if (!hw_argc)
 633                 return 0;
 634
 635         hwht = dm_get_hw_handler(shift(as));
 636         if (!hwht) {
 637                 ti->error = ESTR("unknown hardware handler type");
 638                 return -EINVAL;
 639         }
 640
 641         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
 642         if (r) {
 643                 dm_put_hw_handler(hwht);
 644                 ti->error = ESTR("hardware handler constructor failed");
 645                 return r;
 646         }
 647
 648         m->hw_handler.type = hwht;
 649         consume(as, hw_argc - 1);
 650
 651         return 0;
 652 }
 653
 654 static int parse_features(struct arg_set *as, struct multipath *m,
 655                           struct dm_target *ti)
 656 {
 657         int r;
 658         unsigned argc;
 659
 660         static struct param _params[] = {
 661                 {0, 1, ESTR("invalid number of feature args")},
 662         };
 663
 664         r = read_param(_params, shift(as), &argc, &ti->error);
 665         if (r)
 666                 return -EINVAL;
 667
 668         if (!argc)
 669                 return 0;
 670
 671         if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
 672                 return queue_if_no_path(m, 1);
 673         else {
 674                 ti->error = "Unrecognised multipath feature request";
 675                 return -EINVAL;
 676         }
 677 }
 678
 679 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 680                          char **argv)
 681 {
 682         /* target parameters */
 683         static struct param _params[] = {
 684                 {1, 1024, ESTR("invalid number of priority groups")},
 685                 {1, 1024, ESTR("invalid initial priority group number")},
 686         };
 687
 688         int r;
 689         struct multipath *m;
 690         struct arg_set as;
 691         unsigned pg_count = 0;
 692         unsigned next_pg_num;
 693
 694         as.argc = argc;
 695         as.argv = argv;
 696
 697         m = alloc_multipath();
 698         if (!m) {
 699                 ti->error = ESTR("can't allocate multipath");
 700                 return -EINVAL;
 701         }
 702
 703         r = parse_features(&as, m, ti);
 704         if (r)
 705                 goto bad;
 706
 707         r = parse_hw_handler(&as, m, ti);
 708         if (r)
 709                 goto bad;
 710
 711         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 712         if (r)
 713                 goto bad;
 714
 715         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 716         if (r)
 717                 goto bad;
 718
 719         /* parse the priority groups */
 720         while (as.argc) {
 721                 struct priority_group *pg;
 722
 723                 pg = parse_priority_group(&as, m, ti);
 724                 if (!pg) {
 725                         r = -EINVAL;
 726                         goto bad;
 727                 }
 728
 729                 m->nr_valid_paths += pg->nr_pgpaths;
 730                 list_add_tail(&pg->list, &m->priority_groups);
 731                 pg_count++;
 732                 pg->pg_num = pg_count;
 733                 if (!--next_pg_num)
 734                         m->next_pg = pg;
 735         }
 736
 737         if (pg_count != m->nr_priority_groups) {
 738                 ti->error = ESTR("priority group count mismatch");
 739                 r = -EINVAL;
 740                 goto bad;
 741         }
 742
 743         ti->private = m;
 744         m->ti = ti;
 745
 746         return 0;
 747
 748  bad:
 749         free_multipath(m);
 750         return r;
 751 }
 752
 753 static void multipath_dtr(struct dm_target *ti)
 754 {
 755         struct multipath *m = (struct multipath *) ti->private;
 756
 757         flush_workqueue(kmultipathd);
 758         free_multipath(m);
 759 }
 760
 761 /*
 762  * Map bios, recording original fields for later in case we have to resubmit
 763  */
 764 static int multipath_map(struct dm_target *ti, struct bio *bio,
 765                          union map_info *map_context)
 766 {
 767         int r;
 768         struct mpath_io *mpio;
 769         struct multipath *m = (struct multipath *) ti->private;
 770
 771         if (bio_barrier(bio))
 772                 return -EOPNOTSUPP;
 773
 774         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 775         dm_bio_record(&mpio->details, bio);
 776
 777         map_context->ptr = mpio;
 778         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 779         r = map_io(m, bio, mpio, 0);
 780         if (r < 0)
 781                 mempool_free(mpio, m->mpio_pool);
 782
 783         return r;
 784 }
 785
 786 /*
 787  * Take a path out of use.
 788  */
 789 static int fail_path(struct pgpath *pgpath)
 790 {
 791         unsigned long flags;
 792         struct multipath *m = pgpath->pg->m;
 793
 794         spin_lock_irqsave(&m->lock, flags);
 795
 796         if (!pgpath->path.is_active)
 797                 goto out;
 798
 799         DMWARN("dm-multipath: Failing path %s.", pgpath->path.dev->name);
 800
 801         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 802         pgpath->path.is_active = 0;
 803         pgpath->fail_count++;
 804
 805         m->nr_valid_paths--;
 806
 807         if (pgpath == m->current_pgpath)
 808                 m->current_pgpath = NULL;
 809
 810         queue_work(kmultipathd, &m->trigger_event);
 811
 812 out:
 813         spin_unlock_irqrestore(&m->lock, flags);
 814
 815         return 0;
 816 }
 817
 818 /*
 819  * Reinstate a previously-failed path
 820  */
 821 static int reinstate_path(struct pgpath *pgpath)
 822 {
 823         int r = 0;
 824         unsigned long flags;
 825         struct multipath *m = pgpath->pg->m;
 826
 827         spin_lock_irqsave(&m->lock, flags);
 828
 829         if (pgpath->path.is_active)
 830                 goto out;
 831
 832         if (!pgpath->pg->ps.type) {
 833                 DMWARN("Reinstate path not supported by path selector %s",
 834                        pgpath->pg->ps.type->name);
 835                 r = -EINVAL;
 836                 goto out;
 837         }
 838
 839         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 840         if (r)
 841                 goto out;
 842
 843         pgpath->path.is_active = 1;
 844
 845         m->current_pgpath = NULL;
 846         if (!m->nr_valid_paths++)
 847                 queue_work(kmultipathd, &m->process_queued_ios);
 848
 849         queue_work(kmultipathd, &m->trigger_event);
 850
 851 out:
 852         spin_unlock_irqrestore(&m->lock, flags);
 853
 854         return r;
 855 }
 856
 857 /*
 858  * Fail or reinstate all paths that match the provided struct dm_dev.
 859  */
 860 static int action_dev(struct multipath *m, struct dm_dev *dev,
 861                       action_fn action)
 862 {
 863         int r = 0;
 864         struct pgpath *pgpath;
 865         struct priority_group *pg;
 866
 867         list_for_each_entry(pg, &m->priority_groups, list) {
 868                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 869                         if (pgpath->path.dev == dev)
 870                                 r = action(pgpath);
 871                 }
 872         }
 873
 874         return r;
 875 }
 876
 877 /*
 878  * Temporarily try to avoid having to use the specified PG
 879  */
 880 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 881                       int bypassed)
 882 {
 883         unsigned long flags;
 884
 885         spin_lock_irqsave(&m->lock, flags);
 886
 887         pg->bypassed = bypassed;
 888         m->current_pgpath = NULL;
 889         m->current_pg = NULL;
 890
 891         spin_unlock_irqrestore(&m->lock, flags);
 892
 893         queue_work(kmultipathd, &m->trigger_event);
 894 }
 895
 896 /*
 897  * Switch to using the specified PG from the next I/O that gets mapped
 898  */
 899 static int switch_pg_num(struct multipath *m, const char *pgstr)
 900 {
 901         struct priority_group *pg;
 902         unsigned pgnum;
 903         unsigned long flags;
 904
 905         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 906             (pgnum > m->nr_priority_groups)) {
 907                 DMWARN("invalid PG number supplied to switch_pg_num");
 908                 return -EINVAL;
 909         }
 910
 911         spin_lock_irqsave(&m->lock, flags);
 912         list_for_each_entry(pg, &m->priority_groups, list) {
 913                 pg->bypassed = 0;
 914                 if (--pgnum)
 915                         continue;
 916
 917                 m->current_pgpath = NULL;
 918                 m->current_pg = NULL;
 919                 m->next_pg = pg;
 920         }
 921         spin_unlock_irqrestore(&m->lock, flags);
 922
 923         queue_work(kmultipathd, &m->trigger_event);
 924         return 0;
 925 }
 926
 927 /*
 928  * Set/clear bypassed status of a PG.
 929  * PGs are numbered upwards from 1 in the order they were declared.
 930  */
 931 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 932 {
 933         struct priority_group *pg;
 934         unsigned pgnum;
 935
 936         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 937             (pgnum > m->nr_priority_groups)) {
 938                 DMWARN("invalid PG number supplied to bypass_pg");
 939                 return -EINVAL;
 940         }
 941
 942         list_for_each_entry(pg, &m->priority_groups, list) {
 943                 if (!--pgnum)
 944                         break;
 945         }
 946
 947         bypass_pg(m, pg, bypassed);
 948         return 0;
 949 }
 950
 951 /*
 952  * pg_init must call this when it has completed its initialisation
 953  */
 954 void dm_pg_init_complete(struct path *path, unsigned err_flags)
 955 {
 956         struct pgpath *pgpath = path_to_pgpath(path);
 957         struct priority_group *pg = pgpath->pg;
 958         struct multipath *m = pg->m;
 959         unsigned long flags;
 960
 961         /* We insist on failing the path if the PG is already bypassed. */
 962         if (err_flags && pg->bypassed)
 963                 err_flags |= MP_FAIL_PATH;
 964
 965         if (err_flags & MP_FAIL_PATH)
 966                 fail_path(pgpath);
 967
 968         if (err_flags & MP_BYPASS_PG)
 969                 bypass_pg(m, pg, 1);
 970
 971         spin_lock_irqsave(&m->lock, flags);
 972         if (!err_flags)
 973                 m->queue_io = 0;
 974         else {
 975                 m->current_pgpath = NULL;
 976                 m->current_pg = NULL;
 977         }
 978         queue_work(kmultipathd, &m->process_queued_ios);
 979         spin_unlock_irqrestore(&m->lock, flags);
 980 }
 981
 982 /*
 983  * end_io handling
 984  */
 985 static int do_end_io(struct multipath *m, struct bio *bio,
 986                      int error, struct mpath_io *mpio)
 987 {
 988         struct hw_handler *hwh = &m->hw_handler;
 989         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
 990
 991         if (!error)
 992                 return 0;       /* I/O complete */
 993
 994         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
 995                 return error;
 996
 997         if (error == -EOPNOTSUPP)
 998                 return error;
 999
1000         spin_lock(&m->lock);
1001         if (!m->nr_valid_paths) {
1002                 if (!m->queue_if_no_path) {
1003                         spin_unlock(&m->lock);
1004                         return -EIO;
1005                 } else {
1006                         spin_unlock(&m->lock);
1007                         goto requeue;
1008                 }
1009         }
1010         spin_unlock(&m->lock);
1011
1012         if (hwh->type && hwh->type->error)
1013                 err_flags = hwh->type->error(hwh, bio);
1014
1015         if (mpio->pgpath) {
1016                 if (err_flags & MP_FAIL_PATH)
1017                         fail_path(mpio->pgpath);
1018
1019                 if (err_flags & MP_BYPASS_PG)
1020                         bypass_pg(m, mpio->pgpath->pg, 1);
1021         }
1022
1023         if (err_flags & MP_ERROR_IO)
1024                 return -EIO;
1025
1026       requeue:
1027         dm_bio_restore(&mpio->details, bio);
1028
1029         /* queue for the daemon to resubmit or fail */
1030         spin_lock(&m->lock);
1031         bio_list_add(&m->queued_ios, bio);
1032         m->queue_size++;
1033         if (!m->queue_io)
1034                 queue_work(kmultipathd, &m->process_queued_ios);
1035         spin_unlock(&m->lock);
1036
1037         return 1;       /* io not complete */
1038 }
1039
1040 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1041                             int error, union map_info *map_context)
1042 {
1043         struct multipath *m = (struct multipath *) ti->private;
1044         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1045         struct pgpath *pgpath = mpio->pgpath;
1046         struct path_selector *ps;
1047         int r;
1048
1049         r  = do_end_io(m, bio, error, mpio);
1050         if (pgpath) {
1051                 ps = &pgpath->pg->ps;
1052                 if (ps->type->end_io)
1053                         ps->type->end_io(ps, &pgpath->path);
1054         }
1055         if (r <= 0)
1056                 mempool_free(mpio, m->mpio_pool);
1057
1058         return r;
1059 }
1060
1061 /*
1062  * Suspend can't complete until all the I/O is processed so if
1063  * the last path fails we must error any remaining I/O.
1064  * Note that if the freeze_bdev fails while suspending, the
1065  * queue_if_no_path state is lost - userspace should reset it.
1066  */
1067 static void multipath_presuspend(struct dm_target *ti)
1068 {
1069         struct multipath *m = (struct multipath *) ti->private;
1070
1071         queue_if_no_path(m, 0);
1072 }
1073
1074 /*
1075  * Restore the queue_if_no_path setting.
1076  */
1077 static void multipath_resume(struct dm_target *ti)
1078 {
1079         struct multipath *m = (struct multipath *) ti->private;
1080         unsigned long flags;
1081
1082         spin_lock_irqsave(&m->lock, flags);
1083         m->queue_if_no_path = m->saved_queue_if_no_path;
1084         spin_unlock_irqrestore(&m->lock, flags);
1085 }
1086
1087 /*
1088  * Info output has the following format:
1089  * num_multipath_feature_args [multipath_feature_args]*
1090  * num_handler_status_args [handler_status_args]*
1091  * num_groups init_group_number
1092  *            [A|D|E num_ps_status_args [ps_status_args]*
1093  *             num_paths num_selector_args
1094  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1095  *
1096  * Table output has the following format (identical to the constructor string):
1097  * num_feature_args [features_args]*
1098  * num_handler_args hw_handler [hw_handler_args]*
1099  * num_groups init_group_number
1100  *     [priority selector-name num_ps_args [ps_args]*
1101  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1102  */
1103 static int multipath_status(struct dm_target *ti, status_type_t type,
1104                             char *result, unsigned int maxlen)
1105 {
1106         int sz = 0;
1107         unsigned long flags;
1108         struct multipath *m = (struct multipath *) ti->private;
1109         struct hw_handler *hwh = &m->hw_handler;
1110         struct priority_group *pg;
1111         struct pgpath *p;
1112         unsigned pg_num;
1113         char state;
1114
1115         spin_lock_irqsave(&m->lock, flags);
1116
1117         /* Features */
1118         if (type == STATUSTYPE_INFO)
1119                 DMEMIT("1 %u ", m->queue_size);
1120         else if (m->queue_if_no_path)
1121                 DMEMIT("1 queue_if_no_path ");
1122         else
1123                 DMEMIT("0 ");
1124
1125         if (hwh->type && hwh->type->status)
1126                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1127         else if (!hwh->type || type == STATUSTYPE_INFO)
1128                 DMEMIT("0 ");
1129         else
1130                 DMEMIT("1 %s ", hwh->type->name);
1131
1132         DMEMIT("%u ", m->nr_priority_groups);
1133
1134         if (m->next_pg)
1135                 pg_num = m->next_pg->pg_num;
1136         else if (m->current_pg)
1137                 pg_num = m->current_pg->pg_num;
1138         else
1139                         pg_num = 1;
1140
1141         DMEMIT("%u ", pg_num);
1142
1143         switch (type) {
1144         case STATUSTYPE_INFO:
1145                 list_for_each_entry(pg, &m->priority_groups, list) {
1146                         if (pg->bypassed)
1147                                 state = 'D';    /* Disabled */
1148                         else if (pg == m->current_pg)
1149                                 state = 'A';    /* Currently Active */
1150                         else
1151                                 state = 'E';    /* Enabled */
1152
1153                         DMEMIT("%c ", state);
1154
1155                         if (pg->ps.type->status)
1156                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1157                                                           result + sz,
1158                                                           maxlen - sz);
1159                         else
1160                                 DMEMIT("0 ");
1161
1162                         DMEMIT("%u %u ", pg->nr_pgpaths,
1163                                pg->ps.type->info_args);
1164
1165                         list_for_each_entry(p, &pg->pgpaths, list) {
1166                                 DMEMIT("%s %s %u ", p->path.dev->name,
1167                                        p->path.is_active ? "A" : "F",
1168                                        p->fail_count);
1169                                 if (pg->ps.type->status)
1170                                         sz += pg->ps.type->status(&pg->ps,
1171                                               &p->path, type, result + sz,
1172                                               maxlen - sz);
1173                         }
1174                 }
1175                 break;
1176
1177         case STATUSTYPE_TABLE:
1178                 list_for_each_entry(pg, &m->priority_groups, list) {
1179                         DMEMIT("%s ", pg->ps.type->name);
1180
1181                         if (pg->ps.type->status)
1182                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1183                                                           result + sz,
1184                                                           maxlen - sz);
1185                         else
1186                                 DMEMIT("0 ");
1187
1188                         DMEMIT("%u %u ", pg->nr_pgpaths,
1189                                pg->ps.type->table_args);
1190
1191                         list_for_each_entry(p, &pg->pgpaths, list) {
1192                                 DMEMIT("%s ", p->path.dev->name);
1193                                 if (pg->ps.type->status)
1194                                         sz += pg->ps.type->status(&pg->ps,
1195                                               &p->path, type, result + sz,
1196                                               maxlen - sz);
1197                         }
1198                 }
1199                 break;
1200         }
1201
1202         spin_unlock_irqrestore(&m->lock, flags);
1203
1204         return 0;
1205 }
1206
1207 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1208 {
1209         int r;
1210         struct dm_dev *dev;
1211         struct multipath *m = (struct multipath *) ti->private;
1212         action_fn action;
1213
1214         if (argc == 1) {
1215                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1216                         return queue_if_no_path(m, 1);
1217                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1218                         return queue_if_no_path(m, 0);
1219         }
1220
1221         if (argc != 2)
1222                 goto error;
1223
1224         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1225                 return bypass_pg_num(m, argv[1], 1);
1226         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1227                 return bypass_pg_num(m, argv[1], 0);
1228         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1229                 return switch_pg_num(m, argv[1]);
1230         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1231                 action = reinstate_path;
1232         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1233                 action = fail_path;
1234         else
1235                 goto error;
1236
1237         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1238                           dm_table_get_mode(ti->table), &dev);
1239         if (r) {
1240                 DMWARN("dm-multipath message: error getting device %s",
1241                        argv[1]);
1242                 return -EINVAL;
1243         }
1244
1245         r = action_dev(m, dev, action);
1246
1247         dm_put_device(ti, dev);
1248
1249         return r;
1250
1251 error:
1252         DMWARN("Unrecognised multipath message received.");
1253         return -EINVAL;
1254 }
1255
1256 /*-----------------------------------------------------------------
1257  * Module setup
1258  *---------------------------------------------------------------*/
1259 static struct target_type multipath_target = {
1260         .name = "multipath",
1261         .version = {1, 0, 4},
1262         .module = THIS_MODULE,
1263         .ctr = multipath_ctr,
1264         .dtr = multipath_dtr,
1265         .map = multipath_map,
1266         .end_io = multipath_end_io,
1267         .presuspend = multipath_presuspend,
1268         .resume = multipath_resume,
1269         .status = multipath_status,
1270         .message = multipath_message,
1271 };
1272
1273 static int __init dm_multipath_init(void)
1274 {
1275         int r;
1276
1277         /* allocate a slab for the dm_ios */
1278         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1279                                         0, 0, NULL, NULL);
1280         if (!_mpio_cache)
1281                 return -ENOMEM;
1282
1283         r = dm_register_target(&multipath_target);
1284         if (r < 0) {
1285                 DMERR("%s: register failed %d", multipath_target.name, r);
1286                 kmem_cache_destroy(_mpio_cache);
1287                 return -EINVAL;
1288         }
1289
1290         kmultipathd = create_workqueue("kmpathd");
1291         if (!kmultipathd) {
1292                 DMERR("%s: failed to create workqueue kmpathd",
1293                                 multipath_target.name);
1294                 dm_unregister_target(&multipath_target);
1295                 kmem_cache_destroy(_mpio_cache);
1296                 return -ENOMEM;
1297         }
1298
1299         DMINFO("dm-multipath version %u.%u.%u loaded",
1300                multipath_target.version[0], multipath_target.version[1],
1301                multipath_target.version[2]);
1302
1303         return r;
1304 }
1305
1306 static void __exit dm_multipath_exit(void)
1307 {
1308         int r;
1309
1310         destroy_workqueue(kmultipathd);
1311
1312         r = dm_unregister_target(&multipath_target);
1313         if (r < 0)
1314                 DMERR("%s: target unregister failed %d",
1315                       multipath_target.name, r);
1316         kmem_cache_destroy(_mpio_cache);
1317 }
1318
1319 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1320
1321 module_init(dm_multipath_init);
1322 module_exit(dm_multipath_exit);
1323
1324 MODULE_DESCRIPTION(DM_NAME " multipath target");
1325 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1326 MODULE_LICENSE("GPL");