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 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 && !m->suspended)) {
 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->queue_if_no_path = queue_if_no_path;
 337         if (!m->queue_if_no_path)
 338                 queue_work(kmultipathd, &m->process_queued_ios);
 339
 340         spin_unlock_irqrestore(&m->lock, flags);
 341
 342         return 0;
 343 }
 344
 345 /*-----------------------------------------------------------------
 346  * The multipath daemon is responsible for resubmitting queued ios.
 347  *---------------------------------------------------------------*/
 348
 349 static void dispatch_queued_ios(struct multipath *m)
 350 {
 351         int r;
 352         unsigned long flags;
 353         struct bio *bio = NULL, *next;
 354         struct mpath_io *mpio;
 355         union map_info *info;
 356
 357         spin_lock_irqsave(&m->lock, flags);
 358         bio = bio_list_get(&m->queued_ios);
 359         spin_unlock_irqrestore(&m->lock, flags);
 360
 361         while (bio) {
 362                 next = bio->bi_next;
 363                 bio->bi_next = NULL;
 364
 365                 info = dm_get_mapinfo(bio);
 366                 mpio = info->ptr;
 367
 368                 r = map_io(m, bio, mpio, 1);
 369                 if (r < 0)
 370                         bio_endio(bio, bio->bi_size, r);
 371                 else if (r == 1)
 372                         generic_make_request(bio);
 373
 374                 bio = next;
 375         }
 376 }
 377
 378 static void process_queued_ios(void *data)
 379 {
 380         struct multipath *m = (struct multipath *) data;
 381         struct hw_handler *hwh = &m->hw_handler;
 382         struct pgpath *pgpath;
 383         unsigned init_required, must_queue = 0;
 384         unsigned long flags;
 385
 386         spin_lock_irqsave(&m->lock, flags);
 387
 388         if (!m->current_pgpath)
 389                 __choose_pgpath(m);
 390
 391         pgpath = m->current_pgpath;
 392
 393         if ((pgpath && m->queue_io) ||
 394             (!pgpath && m->queue_if_no_path && !m->suspended))
 395                 must_queue = 1;
 396
 397         init_required = m->pg_init_required;
 398         if (init_required)
 399                 m->pg_init_required = 0;
 400
 401         spin_unlock_irqrestore(&m->lock, flags);
 402
 403         if (init_required)
 404                 hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
 405
 406         if (!must_queue)
 407                 dispatch_queued_ios(m);
 408 }
 409
 410 /*
 411  * An event is triggered whenever a path is taken out of use.
 412  * Includes path failure and PG bypass.
 413  */
 414 static void trigger_event(void *data)
 415 {
 416         struct multipath *m = (struct multipath *) data;
 417
 418         dm_table_event(m->ti->table);
 419 }
 420
 421 /*-----------------------------------------------------------------
 422  * Constructor/argument parsing:
 423  * <#multipath feature args> [<arg>]*
 424  * <#hw_handler args> [hw_handler [<arg>]*]
 425  * <#priority groups>
 426  * <initial priority group>
 427  *     [<selector> <#selector args> [<arg>]*
 428  *      <#paths> <#per-path selector args>
 429  *         [<path> [<arg>]* ]+ ]+
 430  *---------------------------------------------------------------*/
 431 struct param {
 432         unsigned min;
 433         unsigned max;
 434         char *error;
 435 };
 436
 437 #define ESTR(s) ("dm-multipath: " s)
 438
 439 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 440 {
 441         if (!str ||
 442             (sscanf(str, "%u", v) != 1) ||
 443             (*v < param->min) ||
 444             (*v > param->max)) {
 445                 *error = param->error;
 446                 return -EINVAL;
 447         }
 448
 449         return 0;
 450 }
 451
 452 struct arg_set {
 453         unsigned argc;
 454         char **argv;
 455 };
 456
 457 static char *shift(struct arg_set *as)
 458 {
 459         char *r;
 460
 461         if (as->argc) {
 462                 as->argc--;
 463                 r = *as->argv;
 464                 as->argv++;
 465                 return r;
 466         }
 467
 468         return NULL;
 469 }
 470
 471 static void consume(struct arg_set *as, unsigned n)
 472 {
 473         BUG_ON (as->argc < n);
 474         as->argc -= n;
 475         as->argv += n;
 476 }
 477
 478 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 479                                struct dm_target *ti)
 480 {
 481         int r;
 482         struct path_selector_type *pst;
 483         unsigned ps_argc;
 484
 485         static struct param _params[] = {
 486                 {0, 1024, ESTR("invalid number of path selector args")},
 487         };
 488
 489         pst = dm_get_path_selector(shift(as));
 490         if (!pst) {
 491                 ti->error = ESTR("unknown path selector type");
 492                 return -EINVAL;
 493         }
 494
 495         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 496         if (r)
 497                 return -EINVAL;
 498
 499         r = pst->create(&pg->ps, ps_argc, as->argv);
 500         if (r) {
 501                 dm_put_path_selector(pst);
 502                 ti->error = ESTR("path selector constructor failed");
 503                 return r;
 504         }
 505
 506         pg->ps.type = pst;
 507         consume(as, ps_argc);
 508
 509         return 0;
 510 }
 511
 512 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 513                                struct dm_target *ti)
 514 {
 515         int r;
 516         struct pgpath *p;
 517
 518         /* we need at least a path arg */
 519         if (as->argc < 1) {
 520                 ti->error = ESTR("no device given");
 521                 return NULL;
 522         }
 523
 524         p = alloc_pgpath();
 525         if (!p)
 526                 return NULL;
 527
 528         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 529                           dm_table_get_mode(ti->table), &p->path.dev);
 530         if (r) {
 531                 ti->error = ESTR("error getting device");
 532                 goto bad;
 533         }
 534
 535         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 536         if (r) {
 537                 dm_put_device(ti, p->path.dev);
 538                 goto bad;
 539         }
 540
 541         return p;
 542
 543  bad:
 544         free_pgpath(p);
 545         return NULL;
 546 }
 547
 548 static struct priority_group *parse_priority_group(struct arg_set *as,
 549                                                    struct multipath *m,
 550                                                    struct dm_target *ti)
 551 {
 552         static struct param _params[] = {
 553                 {1, 1024, ESTR("invalid number of paths")},
 554                 {0, 1024, ESTR("invalid number of selector args")}
 555         };
 556
 557         int r;
 558         unsigned i, nr_selector_args, nr_params;
 559         struct priority_group *pg;
 560
 561         if (as->argc < 2) {
 562                 as->argc = 0;
 563                 ti->error = ESTR("not enough priority group aruments");
 564                 return NULL;
 565         }
 566
 567         pg = alloc_priority_group();
 568         if (!pg) {
 569                 ti->error = ESTR("couldn't allocate priority group");
 570                 return NULL;
 571         }
 572         pg->m = m;
 573
 574         r = parse_path_selector(as, pg, ti);
 575         if (r)
 576                 goto bad;
 577
 578         /*
 579          * read the paths
 580          */
 581         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 582         if (r)
 583                 goto bad;
 584
 585         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 586         if (r)
 587                 goto bad;
 588
 589         nr_params = 1 + nr_selector_args;
 590         for (i = 0; i < pg->nr_pgpaths; i++) {
 591                 struct pgpath *pgpath;
 592                 struct arg_set path_args;
 593
 594                 if (as->argc < nr_params)
 595                         goto bad;
 596
 597                 path_args.argc = nr_params;
 598                 path_args.argv = as->argv;
 599
 600                 pgpath = parse_path(&path_args, &pg->ps, ti);
 601                 if (!pgpath)
 602                         goto bad;
 603
 604                 pgpath->pg = pg;
 605                 list_add_tail(&pgpath->list, &pg->pgpaths);
 606                 consume(as, nr_params);
 607         }
 608
 609         return pg;
 610
 611  bad:
 612         free_priority_group(pg, ti);
 613         return NULL;
 614 }
 615
 616 static int parse_hw_handler(struct arg_set *as, struct multipath *m,
 617                             struct dm_target *ti)
 618 {
 619         int r;
 620         struct hw_handler_type *hwht;
 621         unsigned hw_argc;
 622
 623         static struct param _params[] = {
 624                 {0, 1024, ESTR("invalid number of hardware handler args")},
 625         };
 626
 627         r = read_param(_params, shift(as), &hw_argc, &ti->error);
 628         if (r)
 629                 return -EINVAL;
 630
 631         if (!hw_argc)
 632                 return 0;
 633
 634         hwht = dm_get_hw_handler(shift(as));
 635         if (!hwht) {
 636                 ti->error = ESTR("unknown hardware handler type");
 637                 return -EINVAL;
 638         }
 639
 640         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
 641         if (r) {
 642                 dm_put_hw_handler(hwht);
 643                 ti->error = ESTR("hardware handler constructor failed");
 644                 return r;
 645         }
 646
 647         m->hw_handler.type = hwht;
 648         consume(as, hw_argc - 1);
 649
 650         return 0;
 651 }
 652
 653 static int parse_features(struct arg_set *as, struct multipath *m,
 654                           struct dm_target *ti)
 655 {
 656         int r;
 657         unsigned argc;
 658
 659         static struct param _params[] = {
 660                 {0, 1, ESTR("invalid number of feature args")},
 661         };
 662
 663         r = read_param(_params, shift(as), &argc, &ti->error);
 664         if (r)
 665                 return -EINVAL;
 666
 667         if (!argc)
 668                 return 0;
 669
 670         if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
 671                 return queue_if_no_path(m, 1);
 672         else {
 673                 ti->error = "Unrecognised multipath feature request";
 674                 return -EINVAL;
 675         }
 676 }
 677
 678 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 679                          char **argv)
 680 {
 681         /* target parameters */
 682         static struct param _params[] = {
 683                 {1, 1024, ESTR("invalid number of priority groups")},
 684                 {1, 1024, ESTR("invalid initial priority group number")},
 685         };
 686
 687         int r;
 688         struct multipath *m;
 689         struct arg_set as;
 690         unsigned pg_count = 0;
 691         unsigned next_pg_num;
 692
 693         as.argc = argc;
 694         as.argv = argv;
 695
 696         m = alloc_multipath();
 697         if (!m) {
 698                 ti->error = ESTR("can't allocate multipath");
 699                 return -EINVAL;
 700         }
 701
 702         r = parse_features(&as, m, ti);
 703         if (r)
 704                 goto bad;
 705
 706         r = parse_hw_handler(&as, m, ti);
 707         if (r)
 708                 goto bad;
 709
 710         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 711         if (r)
 712                 goto bad;
 713
 714         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 715         if (r)
 716                 goto bad;
 717
 718         /* parse the priority groups */
 719         while (as.argc) {
 720                 struct priority_group *pg;
 721
 722                 pg = parse_priority_group(&as, m, ti);
 723                 if (!pg) {
 724                         r = -EINVAL;
 725                         goto bad;
 726                 }
 727
 728                 m->nr_valid_paths += pg->nr_pgpaths;
 729                 list_add_tail(&pg->list, &m->priority_groups);
 730                 pg_count++;
 731                 pg->pg_num = pg_count;
 732                 if (!--next_pg_num)
 733                         m->next_pg = pg;
 734         }
 735
 736         if (pg_count != m->nr_priority_groups) {
 737                 ti->error = ESTR("priority group count mismatch");
 738                 r = -EINVAL;
 739                 goto bad;
 740         }
 741
 742         ti->private = m;
 743         m->ti = ti;
 744
 745         return 0;
 746
 747  bad:
 748         free_multipath(m);
 749         return r;
 750 }
 751
 752 static void multipath_dtr(struct dm_target *ti)
 753 {
 754         struct multipath *m = (struct multipath *) ti->private;
 755
 756         flush_workqueue(kmultipathd);
 757         free_multipath(m);
 758 }
 759
 760 /*
 761  * Map bios, recording original fields for later in case we have to resubmit
 762  */
 763 static int multipath_map(struct dm_target *ti, struct bio *bio,
 764                          union map_info *map_context)
 765 {
 766         int r;
 767         struct mpath_io *mpio;
 768         struct multipath *m = (struct multipath *) ti->private;
 769
 770         if (bio_barrier(bio))
 771                 return -EOPNOTSUPP;
 772
 773         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 774         dm_bio_record(&mpio->details, bio);
 775
 776         map_context->ptr = mpio;
 777         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 778         r = map_io(m, bio, mpio, 0);
 779         if (r < 0)
 780                 mempool_free(mpio, m->mpio_pool);
 781
 782         return r;
 783 }
 784
 785 /*
 786  * Take a path out of use.
 787  */
 788 static int fail_path(struct pgpath *pgpath)
 789 {
 790         unsigned long flags;
 791         struct multipath *m = pgpath->pg->m;
 792
 793         spin_lock_irqsave(&m->lock, flags);
 794
 795         if (!pgpath->path.is_active)
 796                 goto out;
 797
 798         DMWARN("dm-multipath: Failing path %s.", pgpath->path.dev->name);
 799
 800         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 801         pgpath->path.is_active = 0;
 802         pgpath->fail_count++;
 803
 804         m->nr_valid_paths--;
 805
 806         if (pgpath == m->current_pgpath)
 807                 m->current_pgpath = NULL;
 808
 809         queue_work(kmultipathd, &m->trigger_event);
 810
 811 out:
 812         spin_unlock_irqrestore(&m->lock, flags);
 813
 814         return 0;
 815 }
 816
 817 /*
 818  * Reinstate a previously-failed path
 819  */
 820 static int reinstate_path(struct pgpath *pgpath)
 821 {
 822         int r = 0;
 823         unsigned long flags;
 824         struct multipath *m = pgpath->pg->m;
 825
 826         spin_lock_irqsave(&m->lock, flags);
 827
 828         if (pgpath->path.is_active)
 829                 goto out;
 830
 831         if (!pgpath->pg->ps.type) {
 832                 DMWARN("Reinstate path not supported by path selector %s",
 833                        pgpath->pg->ps.type->name);
 834                 r = -EINVAL;
 835                 goto out;
 836         }
 837
 838         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 839         if (r)
 840                 goto out;
 841
 842         pgpath->path.is_active = 1;
 843
 844         m->current_pgpath = NULL;
 845         if (!m->nr_valid_paths++)
 846                 queue_work(kmultipathd, &m->process_queued_ios);
 847
 848         queue_work(kmultipathd, &m->trigger_event);
 849
 850 out:
 851         spin_unlock_irqrestore(&m->lock, flags);
 852
 853         return r;
 854 }
 855
 856 /*
 857  * Fail or reinstate all paths that match the provided struct dm_dev.
 858  */
 859 static int action_dev(struct multipath *m, struct dm_dev *dev,
 860                       action_fn action)
 861 {
 862         int r = 0;
 863         struct pgpath *pgpath;
 864         struct priority_group *pg;
 865
 866         list_for_each_entry(pg, &m->priority_groups, list) {
 867                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 868                         if (pgpath->path.dev == dev)
 869                                 r = action(pgpath);
 870                 }
 871         }
 872
 873         return r;
 874 }
 875
 876 /*
 877  * Temporarily try to avoid having to use the specified PG
 878  */
 879 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 880                       int bypassed)
 881 {
 882         unsigned long flags;
 883
 884         spin_lock_irqsave(&m->lock, flags);
 885
 886         pg->bypassed = bypassed;
 887         m->current_pgpath = NULL;
 888         m->current_pg = NULL;
 889
 890         spin_unlock_irqrestore(&m->lock, flags);
 891
 892         queue_work(kmultipathd, &m->trigger_event);
 893 }
 894
 895 /*
 896  * Switch to using the specified PG from the next I/O that gets mapped
 897  */
 898 static int switch_pg_num(struct multipath *m, const char *pgstr)
 899 {
 900         struct priority_group *pg;
 901         unsigned pgnum;
 902         unsigned long flags;
 903
 904         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 905             (pgnum > m->nr_priority_groups)) {
 906                 DMWARN("invalid PG number supplied to switch_pg_num");
 907                 return -EINVAL;
 908         }
 909
 910         spin_lock_irqsave(&m->lock, flags);
 911         list_for_each_entry(pg, &m->priority_groups, list) {
 912                 pg->bypassed = 0;
 913                 if (--pgnum)
 914                         continue;
 915
 916                 m->current_pgpath = NULL;
 917                 m->current_pg = NULL;
 918                 m->next_pg = pg;
 919         }
 920         spin_unlock_irqrestore(&m->lock, flags);
 921
 922         queue_work(kmultipathd, &m->trigger_event);
 923         return 0;
 924 }
 925
 926 /*
 927  * Set/clear bypassed status of a PG.
 928  * PGs are numbered upwards from 1 in the order they were declared.
 929  */
 930 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 931 {
 932         struct priority_group *pg;
 933         unsigned pgnum;
 934
 935         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 936             (pgnum > m->nr_priority_groups)) {
 937                 DMWARN("invalid PG number supplied to bypass_pg");
 938                 return -EINVAL;
 939         }
 940
 941         list_for_each_entry(pg, &m->priority_groups, list) {
 942                 if (!--pgnum)
 943                         break;
 944         }
 945
 946         bypass_pg(m, pg, bypassed);
 947         return 0;
 948 }
 949
 950 /*
 951  * pg_init must call this when it has completed its initialisation
 952  */
 953 void dm_pg_init_complete(struct path *path, unsigned err_flags)
 954 {
 955         struct pgpath *pgpath = path_to_pgpath(path);
 956         struct priority_group *pg = pgpath->pg;
 957         struct multipath *m = pg->m;
 958         unsigned long flags;
 959
 960         /* We insist on failing the path if the PG is already bypassed. */
 961         if (err_flags && pg->bypassed)
 962                 err_flags |= MP_FAIL_PATH;
 963
 964         if (err_flags & MP_FAIL_PATH)
 965                 fail_path(pgpath);
 966
 967         if (err_flags & MP_BYPASS_PG)
 968                 bypass_pg(m, pg, 1);
 969
 970         spin_lock_irqsave(&m->lock, flags);
 971         if (!err_flags)
 972                 m->queue_io = 0;
 973         else {
 974                 m->current_pgpath = NULL;
 975                 m->current_pg = NULL;
 976         }
 977         queue_work(kmultipathd, &m->process_queued_ios);
 978         spin_unlock_irqrestore(&m->lock, flags);
 979 }
 980
 981 /*
 982  * end_io handling
 983  */
 984 static int do_end_io(struct multipath *m, struct bio *bio,
 985                      int error, struct mpath_io *mpio)
 986 {
 987         struct hw_handler *hwh = &m->hw_handler;
 988         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
 989
 990         if (!error)
 991                 return 0;       /* I/O complete */
 992
 993         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
 994                 return error;
 995
 996         if (error == -EOPNOTSUPP)
 997                 return error;
 998
 999         spin_lock(&m->lock);
1000         if (!m->nr_valid_paths) {
1001                 if (!m->queue_if_no_path || m->suspended) {
1002                         spin_unlock(&m->lock);
1003                         return -EIO;
1004                 } else {
1005                         spin_unlock(&m->lock);
1006                         goto requeue;
1007                 }
1008         }
1009         spin_unlock(&m->lock);
1010
1011         if (hwh->type && hwh->type->error)
1012                 err_flags = hwh->type->error(hwh, bio);
1013
1014         if (mpio->pgpath) {
1015                 if (err_flags & MP_FAIL_PATH)
1016                         fail_path(mpio->pgpath);
1017
1018                 if (err_flags & MP_BYPASS_PG)
1019                         bypass_pg(m, mpio->pgpath->pg, 1);
1020         }
1021
1022         if (err_flags & MP_ERROR_IO)
1023                 return -EIO;
1024
1025       requeue:
1026         dm_bio_restore(&mpio->details, bio);
1027
1028         /* queue for the daemon to resubmit or fail */
1029         spin_lock(&m->lock);
1030         bio_list_add(&m->queued_ios, bio);
1031         m->queue_size++;
1032         if (!m->queue_io)
1033                 queue_work(kmultipathd, &m->process_queued_ios);
1034         spin_unlock(&m->lock);
1035
1036         return 1;       /* io not complete */
1037 }
1038
1039 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1040                             int error, union map_info *map_context)
1041 {
1042         struct multipath *m = (struct multipath *) ti->private;
1043         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1044         struct pgpath *pgpath = mpio->pgpath;
1045         struct path_selector *ps;
1046         int r;
1047
1048         r  = do_end_io(m, bio, error, mpio);
1049         if (pgpath) {
1050                 ps = &pgpath->pg->ps;
1051                 if (ps->type->end_io)
1052                         ps->type->end_io(ps, &pgpath->path);
1053         }
1054         if (r <= 0)
1055                 mempool_free(mpio, m->mpio_pool);
1056
1057         return r;
1058 }
1059
1060 /*
1061  * Suspend can't complete until all the I/O is processed so if
1062  * the last path failed we will now error any queued I/O.
1063  */
1064 static void multipath_presuspend(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 = 1;
1071         if (m->queue_if_no_path)
1072                 queue_work(kmultipathd, &m->process_queued_ios);
1073         spin_unlock_irqrestore(&m->lock, flags);
1074 }
1075
1076 static void multipath_resume(struct dm_target *ti)
1077 {
1078         struct multipath *m = (struct multipath *) ti->private;
1079         unsigned long flags;
1080
1081         spin_lock_irqsave(&m->lock, flags);
1082         m->suspended = 0;
1083         spin_unlock_irqrestore(&m->lock, flags);
1084 }
1085
1086 /*
1087  * Info output has the following format:
1088  * num_multipath_feature_args [multipath_feature_args]*
1089  * num_handler_status_args [handler_status_args]*
1090  * num_groups init_group_number
1091  *            [A|D|E num_ps_status_args [ps_status_args]*
1092  *             num_paths num_selector_args
1093  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1094  *
1095  * Table output has the following format (identical to the constructor string):
1096  * num_feature_args [features_args]*
1097  * num_handler_args hw_handler [hw_handler_args]*
1098  * num_groups init_group_number
1099  *     [priority selector-name num_ps_args [ps_args]*
1100  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1101  */
1102 static int multipath_status(struct dm_target *ti, status_type_t type,
1103                             char *result, unsigned int maxlen)
1104 {
1105         int sz = 0;
1106         unsigned long flags;
1107         struct multipath *m = (struct multipath *) ti->private;
1108         struct hw_handler *hwh = &m->hw_handler;
1109         struct priority_group *pg;
1110         struct pgpath *p;
1111         unsigned pg_num;
1112         char state;
1113
1114         spin_lock_irqsave(&m->lock, flags);
1115
1116         /* Features */
1117         if (type == STATUSTYPE_INFO)
1118                 DMEMIT("1 %u ", m->queue_size);
1119         else if (m->queue_if_no_path)
1120                 DMEMIT("1 queue_if_no_path ");
1121         else
1122                 DMEMIT("0 ");
1123
1124         if (hwh->type && hwh->type->status)
1125                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1126         else if (!hwh->type || type == STATUSTYPE_INFO)
1127                 DMEMIT("0 ");
1128         else
1129                 DMEMIT("1 %s ", hwh->type->name);
1130
1131         DMEMIT("%u ", m->nr_priority_groups);
1132
1133         if (m->next_pg)
1134                 pg_num = m->next_pg->pg_num;
1135         else if (m->current_pg)
1136                 pg_num = m->current_pg->pg_num;
1137         else
1138                         pg_num = 1;
1139
1140         DMEMIT("%u ", pg_num);
1141
1142         switch (type) {
1143         case STATUSTYPE_INFO:
1144                 list_for_each_entry(pg, &m->priority_groups, list) {
1145                         if (pg->bypassed)
1146                                 state = 'D';    /* Disabled */
1147                         else if (pg == m->current_pg)
1148                                 state = 'A';    /* Currently Active */
1149                         else
1150                                 state = 'E';    /* Enabled */
1151
1152                         DMEMIT("%c ", state);
1153
1154                         if (pg->ps.type->status)
1155                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1156                                                           result + sz,
1157                                                           maxlen - sz);
1158                         else
1159                                 DMEMIT("0 ");
1160
1161                         DMEMIT("%u %u ", pg->nr_pgpaths,
1162                                pg->ps.type->info_args);
1163
1164                         list_for_each_entry(p, &pg->pgpaths, list) {
1165                                 DMEMIT("%s %s %u ", p->path.dev->name,
1166                                        p->path.is_active ? "A" : "F",
1167                                        p->fail_count);
1168                                 if (pg->ps.type->status)
1169                                         sz += pg->ps.type->status(&pg->ps,
1170                                               &p->path, type, result + sz,
1171                                               maxlen - sz);
1172                         }
1173                 }
1174                 break;
1175
1176         case STATUSTYPE_TABLE:
1177                 list_for_each_entry(pg, &m->priority_groups, list) {
1178                         DMEMIT("%s ", pg->ps.type->name);
1179
1180                         if (pg->ps.type->status)
1181                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1182                                                           result + sz,
1183                                                           maxlen - sz);
1184                         else
1185                                 DMEMIT("0 ");
1186
1187                         DMEMIT("%u %u ", pg->nr_pgpaths,
1188                                pg->ps.type->table_args);
1189
1190                         list_for_each_entry(p, &pg->pgpaths, list) {
1191                                 DMEMIT("%s ", p->path.dev->name);
1192                                 if (pg->ps.type->status)
1193                                         sz += pg->ps.type->status(&pg->ps,
1194                                               &p->path, type, result + sz,
1195                                               maxlen - sz);
1196                         }
1197                 }
1198                 break;
1199         }
1200
1201         spin_unlock_irqrestore(&m->lock, flags);
1202
1203         return 0;
1204 }
1205
1206 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1207 {
1208         int r;
1209         struct dm_dev *dev;
1210         struct multipath *m = (struct multipath *) ti->private;
1211         action_fn action;
1212
1213         if (argc == 1) {
1214                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1215                         return queue_if_no_path(m, 1);
1216                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1217                         return queue_if_no_path(m, 0);
1218         }
1219
1220         if (argc != 2)
1221                 goto error;
1222
1223         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1224                 return bypass_pg_num(m, argv[1], 1);
1225         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1226                 return bypass_pg_num(m, argv[1], 0);
1227         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1228                 return switch_pg_num(m, argv[1]);
1229         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1230                 action = reinstate_path;
1231         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1232                 action = fail_path;
1233         else
1234                 goto error;
1235
1236         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1237                           dm_table_get_mode(ti->table), &dev);
1238         if (r) {
1239                 DMWARN("dm-multipath message: error getting device %s",
1240                        argv[1]);
1241                 return -EINVAL;
1242         }
1243
1244         r = action_dev(m, dev, action);
1245
1246         dm_put_device(ti, dev);
1247
1248         return r;
1249
1250 error:
1251         DMWARN("Unrecognised multipath message received.");
1252         return -EINVAL;
1253 }
1254
1255 /*-----------------------------------------------------------------
1256  * Module setup
1257  *---------------------------------------------------------------*/
1258 static struct target_type multipath_target = {
1259         .name = "multipath",
1260         .version = {1, 0, 4},
1261         .module = THIS_MODULE,
1262         .ctr = multipath_ctr,
1263         .dtr = multipath_dtr,
1264         .map = multipath_map,
1265         .end_io = multipath_end_io,
1266         .presuspend = multipath_presuspend,
1267         .resume = multipath_resume,
1268         .status = multipath_status,
1269         .message = multipath_message,
1270 };
1271
1272 static int __init dm_multipath_init(void)
1273 {
1274         int r;
1275
1276         /* allocate a slab for the dm_ios */
1277         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1278                                         0, 0, NULL, NULL);
1279         if (!_mpio_cache)
1280                 return -ENOMEM;
1281
1282         r = dm_register_target(&multipath_target);
1283         if (r < 0) {
1284                 DMERR("%s: register failed %d", multipath_target.name, r);
1285                 kmem_cache_destroy(_mpio_cache);
1286                 return -EINVAL;
1287         }
1288
1289         kmultipathd = create_workqueue("kmpathd");
1290         if (!kmultipathd) {
1291                 DMERR("%s: failed to create workqueue kmpathd",
1292                                 multipath_target.name);
1293                 dm_unregister_target(&multipath_target);
1294                 kmem_cache_destroy(_mpio_cache);
1295                 return -ENOMEM;
1296         }
1297
1298         DMINFO("dm-multipath version %u.%u.%u loaded",
1299                multipath_target.version[0], multipath_target.version[1],
1300                multipath_target.version[2]);
1301
1302         return r;
1303 }
1304
1305 static void __exit dm_multipath_exit(void)
1306 {
1307         int r;
1308
1309         destroy_workqueue(kmultipathd);
1310
1311         r = dm_unregister_target(&multipath_target);
1312         if (r < 0)
1313                 DMERR("%s: target unregister failed %d",
1314                       multipath_target.name, r);
1315         kmem_cache_destroy(_mpio_cache);
1316 }
1317
1318 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1319
1320 module_init(dm_multipath_init);
1321 module_exit(dm_multipath_exit);
1322
1323 MODULE_DESCRIPTION(DM_NAME " multipath target");
1324 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1325 MODULE_LICENSE("GPL");