drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include "dm.h"
   9 #include "dm-path-selector.h"
  10 #include "dm-hw-handler.h"
  11 #include "dm-bio-list.h"
  12 #include "dm-bio-record.h"
  13
  14 #include <linux/ctype.h>
  15 #include <linux/init.h>
  16 #include <linux/mempool.h>
  17 #include <linux/module.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/slab.h>
  20 #include <linux/time.h>
  21 #include <linux/workqueue.h>
  22 #include <asm/atomic.h>
  23
  24 #define DM_MSG_PREFIX "multipath"
  25 #define MESG_STR(x) x, sizeof(x)
  26
  27 /* Path properties */
  28 struct pgpath {
  29         struct list_head list;
  30
  31         struct priority_group *pg;      /* Owning PG */
  32         unsigned fail_count;            /* Cumulative failure count */
  33
  34         struct path path;
  35 };
  36
  37 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  38
  39 /*
  40  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  41  * Each has a path selector which controls which path gets used.
  42  */
  43 struct priority_group {
  44         struct list_head list;
  45
  46         struct multipath *m;            /* Owning multipath instance */
  47         struct path_selector ps;
  48
  49         unsigned pg_num;                /* Reference number */
  50         unsigned bypassed;              /* Temporarily bypass this PG? */
  51
  52         unsigned nr_pgpaths;            /* Number of paths in PG */
  53         struct list_head pgpaths;
  54 };
  55
  56 /* Multipath context */
  57 struct multipath {
  58         struct list_head list;
  59         struct dm_target *ti;
  60
  61         spinlock_t lock;
  62
  63         struct hw_handler hw_handler;
  64         unsigned nr_priority_groups;
  65         struct list_head priority_groups;
  66         unsigned pg_init_required;      /* pg_init needs calling? */
  67         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  68
  69         unsigned nr_valid_paths;        /* Total number of usable paths */
  70         struct pgpath *current_pgpath;
  71         struct priority_group *current_pg;
  72         struct priority_group *next_pg; /* Switch to this PG if set */
  73         unsigned repeat_count;          /* I/Os left before calling PS again */
  74
  75         unsigned queue_io;              /* Must we queue all I/O? */
  76         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  77         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  78
  79         struct work_struct process_queued_ios;
  80         struct bio_list queued_ios;
  81         unsigned queue_size;
  82
  83         struct work_struct trigger_event;
  84
  85         /*
  86          * We must use a mempool of mpath_io structs so that we
  87          * can resubmit bios on error.
  88          */
  89         mempool_t *mpio_pool;
  90 };
  91
  92 /*
  93  * Context information attached to each bio we process.
  94  */
  95 struct mpath_io {
  96         struct pgpath *pgpath;
  97         struct dm_bio_details details;
  98 };
  99
 100 typedef int (*action_fn) (struct pgpath *pgpath);
 101
 102 #define MIN_IOS 256     /* Mempool size */
 103
 104 static kmem_cache_t *_mpio_cache;
 105
 106 struct workqueue_struct *kmultipathd;
 107 static void process_queued_ios(void *data);
 108 static void trigger_event(void *data);
 109
 110
 111 /*-----------------------------------------------
 112  * Allocation routines
 113  *-----------------------------------------------*/
 114
 115 static struct pgpath *alloc_pgpath(void)
 116 {
 117         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 118
 119         if (pgpath)
 120                 pgpath->path.is_active = 1;
 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 = kzalloc(sizeof(*pg), GFP_KERNEL);
 135
 136         if (pg)
 137                 INIT_LIST_HEAD(&pg->pgpaths);
 138
 139         return pg;
 140 }
 141
 142 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 143 {
 144         struct pgpath *pgpath, *tmp;
 145
 146         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 147                 list_del(&pgpath->list);
 148                 dm_put_device(ti, pgpath->path.dev);
 149                 free_pgpath(pgpath);
 150         }
 151 }
 152
 153 static void free_priority_group(struct priority_group *pg,
 154                                 struct dm_target *ti)
 155 {
 156         struct path_selector *ps = &pg->ps;
 157
 158         if (ps->type) {
 159                 ps->type->destroy(ps);
 160                 dm_put_path_selector(ps->type);
 161         }
 162
 163         free_pgpaths(&pg->pgpaths, ti);
 164         kfree(pg);
 165 }
 166
 167 static struct multipath *alloc_multipath(struct dm_target *ti)
 168 {
 169         struct multipath *m;
 170
 171         m = kzalloc(sizeof(*m), GFP_KERNEL);
 172         if (m) {
 173                 INIT_LIST_HEAD(&m->priority_groups);
 174                 spin_lock_init(&m->lock);
 175                 m->queue_io = 1;
 176                 INIT_WORK(&m->process_queued_ios, process_queued_ios, m);
 177                 INIT_WORK(&m->trigger_event, trigger_event, m);
 178                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 179                 if (!m->mpio_pool) {
 180                         kfree(m);
 181                         return NULL;
 182                 }
 183                 m->ti = ti;
 184                 ti->private = m;
 185         }
 186
 187         return m;
 188 }
 189
 190 static void free_multipath(struct multipath *m)
 191 {
 192         struct priority_group *pg, *tmp;
 193         struct hw_handler *hwh = &m->hw_handler;
 194
 195         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 196                 list_del(&pg->list);
 197                 free_priority_group(pg, m->ti);
 198         }
 199
 200         if (hwh->type) {
 201                 hwh->type->destroy(hwh);
 202                 dm_put_hw_handler(hwh->type);
 203         }
 204
 205         mempool_destroy(m->mpio_pool);
 206         kfree(m);
 207 }
 208
 209
 210 /*-----------------------------------------------
 211  * Path selection
 212  *-----------------------------------------------*/
 213
 214 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 215 {
 216         struct hw_handler *hwh = &m->hw_handler;
 217
 218         m->current_pg = pgpath->pg;
 219
 220         /* Must we initialise the PG first, and queue I/O till it's ready? */
 221         if (hwh->type && hwh->type->pg_init) {
 222                 m->pg_init_required = 1;
 223                 m->queue_io = 1;
 224         } else {
 225                 m->pg_init_required = 0;
 226                 m->queue_io = 0;
 227         }
 228 }
 229
 230 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
 231 {
 232         struct path *path;
 233
 234         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
 235         if (!path)
 236                 return -ENXIO;
 237
 238         m->current_pgpath = path_to_pgpath(path);
 239
 240         if (m->current_pg != pg)
 241                 __switch_pg(m, m->current_pgpath);
 242
 243         return 0;
 244 }
 245
 246 static void __choose_pgpath(struct multipath *m)
 247 {
 248         struct priority_group *pg;
 249         unsigned bypassed = 1;
 250
 251         if (!m->nr_valid_paths)
 252                 goto failed;
 253
 254         /* Were we instructed to switch PG? */
 255         if (m->next_pg) {
 256                 pg = m->next_pg;
 257                 m->next_pg = NULL;
 258                 if (!__choose_path_in_pg(m, pg))
 259                         return;
 260         }
 261
 262         /* Don't change PG until it has no remaining paths */
 263         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
 264                 return;
 265
 266         /*
 267          * Loop through priority groups until we find a valid path.
 268          * First time we skip PGs marked 'bypassed'.
 269          * Second time we only try the ones we skipped.
 270          */
 271         do {
 272                 list_for_each_entry(pg, &m->priority_groups, list) {
 273                         if (pg->bypassed == bypassed)
 274                                 continue;
 275                         if (!__choose_path_in_pg(m, pg))
 276                                 return;
 277                 }
 278         } while (bypassed--);
 279
 280 failed:
 281         m->current_pgpath = NULL;
 282         m->current_pg = NULL;
 283 }
 284
 285 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
 286                   unsigned was_queued)
 287 {
 288         int r = 1;
 289         unsigned long flags;
 290         struct pgpath *pgpath;
 291
 292         spin_lock_irqsave(&m->lock, flags);
 293
 294         /* Do we need to select a new pgpath? */
 295         if (!m->current_pgpath ||
 296             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 297                 __choose_pgpath(m);
 298
 299         pgpath = m->current_pgpath;
 300
 301         if (was_queued)
 302                 m->queue_size--;
 303
 304         if ((pgpath && m->queue_io) ||
 305             (!pgpath && m->queue_if_no_path)) {
 306                 /* Queue for the daemon to resubmit */
 307                 bio_list_add(&m->queued_ios, bio);
 308                 m->queue_size++;
 309                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 310                     !m->queue_io)
 311                         queue_work(kmultipathd, &m->process_queued_ios);
 312                 pgpath = NULL;
 313                 r = 0;
 314         } else if (!pgpath)
 315                 r = -EIO;               /* Failed */
 316         else
 317                 bio->bi_bdev = pgpath->path.dev->bdev;
 318
 319         mpio->pgpath = pgpath;
 320
 321         spin_unlock_irqrestore(&m->lock, flags);
 322
 323         return r;
 324 }
 325
 326 /*
 327  * If we run out of usable paths, should we queue I/O or error it?
 328  */
 329 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 330                             unsigned save_old_value)
 331 {
 332         unsigned long flags;
 333
 334         spin_lock_irqsave(&m->lock, flags);
 335
 336         if (save_old_value)
 337                 m->saved_queue_if_no_path = m->queue_if_no_path;
 338         else
 339                 m->saved_queue_if_no_path = queue_if_no_path;
 340         m->queue_if_no_path = queue_if_no_path;
 341         if (!m->queue_if_no_path && m->queue_size)
 342                 queue_work(kmultipathd, &m->process_queued_ios);
 343
 344         spin_unlock_irqrestore(&m->lock, flags);
 345
 346         return 0;
 347 }
 348
 349 /*-----------------------------------------------------------------
 350  * The multipath daemon is responsible for resubmitting queued ios.
 351  *---------------------------------------------------------------*/
 352
 353 static void dispatch_queued_ios(struct multipath *m)
 354 {
 355         int r;
 356         unsigned long flags;
 357         struct bio *bio = NULL, *next;
 358         struct mpath_io *mpio;
 359         union map_info *info;
 360
 361         spin_lock_irqsave(&m->lock, flags);
 362         bio = bio_list_get(&m->queued_ios);
 363         spin_unlock_irqrestore(&m->lock, flags);
 364
 365         while (bio) {
 366                 next = bio->bi_next;
 367                 bio->bi_next = NULL;
 368
 369                 info = dm_get_mapinfo(bio);
 370                 mpio = info->ptr;
 371
 372                 r = map_io(m, bio, mpio, 1);
 373                 if (r < 0)
 374                         bio_endio(bio, bio->bi_size, r);
 375                 else if (r == 1)
 376                         generic_make_request(bio);
 377
 378                 bio = next;
 379         }
 380 }
 381
 382 static void process_queued_ios(void *data)
 383 {
 384         struct multipath *m = (struct multipath *) data;
 385         struct hw_handler *hwh = &m->hw_handler;
 386         struct pgpath *pgpath = NULL;
 387         unsigned init_required = 0, must_queue = 1;
 388         unsigned long flags;
 389
 390         spin_lock_irqsave(&m->lock, flags);
 391
 392         if (!m->queue_size)
 393                 goto out;
 394
 395         if (!m->current_pgpath)
 396                 __choose_pgpath(m);
 397
 398         pgpath = m->current_pgpath;
 399
 400         if ((pgpath && !m->queue_io) ||
 401             (!pgpath && !m->queue_if_no_path))
 402                 must_queue = 0;
 403
 404         if (m->pg_init_required && !m->pg_init_in_progress) {
 405                 m->pg_init_required = 0;
 406                 m->pg_init_in_progress = 1;
 407                 init_required = 1;
 408         }
 409
 410 out:
 411         spin_unlock_irqrestore(&m->lock, flags);
 412
 413         if (init_required)
 414                 hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
 415
 416         if (!must_queue)
 417                 dispatch_queued_ios(m);
 418 }
 419
 420 /*
 421  * An event is triggered whenever a path is taken out of use.
 422  * Includes path failure and PG bypass.
 423  */
 424 static void trigger_event(void *data)
 425 {
 426         struct multipath *m = (struct multipath *) data;
 427
 428         dm_table_event(m->ti->table);
 429 }
 430
 431 /*-----------------------------------------------------------------
 432  * Constructor/argument parsing:
 433  * <#multipath feature args> [<arg>]*
 434  * <#hw_handler args> [hw_handler [<arg>]*]
 435  * <#priority groups>
 436  * <initial priority group>
 437  *     [<selector> <#selector args> [<arg>]*
 438  *      <#paths> <#per-path selector args>
 439  *         [<path> [<arg>]* ]+ ]+
 440  *---------------------------------------------------------------*/
 441 struct param {
 442         unsigned min;
 443         unsigned max;
 444         char *error;
 445 };
 446
 447 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 448 {
 449         if (!str ||
 450             (sscanf(str, "%u", v) != 1) ||
 451             (*v < param->min) ||
 452             (*v > param->max)) {
 453                 *error = param->error;
 454                 return -EINVAL;
 455         }
 456
 457         return 0;
 458 }
 459
 460 struct arg_set {
 461         unsigned argc;
 462         char **argv;
 463 };
 464
 465 static char *shift(struct arg_set *as)
 466 {
 467         char *r;
 468
 469         if (as->argc) {
 470                 as->argc--;
 471                 r = *as->argv;
 472                 as->argv++;
 473                 return r;
 474         }
 475
 476         return NULL;
 477 }
 478
 479 static void consume(struct arg_set *as, unsigned n)
 480 {
 481         BUG_ON (as->argc < n);
 482         as->argc -= n;
 483         as->argv += n;
 484 }
 485
 486 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 487                                struct dm_target *ti)
 488 {
 489         int r;
 490         struct path_selector_type *pst;
 491         unsigned ps_argc;
 492
 493         static struct param _params[] = {
 494                 {0, 1024, "invalid number of path selector args"},
 495         };
 496
 497         pst = dm_get_path_selector(shift(as));
 498         if (!pst) {
 499                 ti->error = "unknown path selector type";
 500                 return -EINVAL;
 501         }
 502
 503         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 504         if (r)
 505                 return -EINVAL;
 506
 507         r = pst->create(&pg->ps, ps_argc, as->argv);
 508         if (r) {
 509                 dm_put_path_selector(pst);
 510                 ti->error = "path selector constructor failed";
 511                 return r;
 512         }
 513
 514         pg->ps.type = pst;
 515         consume(as, ps_argc);
 516
 517         return 0;
 518 }
 519
 520 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 521                                struct dm_target *ti)
 522 {
 523         int r;
 524         struct pgpath *p;
 525
 526         /* we need at least a path arg */
 527         if (as->argc < 1) {
 528                 ti->error = "no device given";
 529                 return NULL;
 530         }
 531
 532         p = alloc_pgpath();
 533         if (!p)
 534                 return NULL;
 535
 536         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 537                           dm_table_get_mode(ti->table), &p->path.dev);
 538         if (r) {
 539                 ti->error = "error getting device";
 540                 goto bad;
 541         }
 542
 543         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 544         if (r) {
 545                 dm_put_device(ti, p->path.dev);
 546                 goto bad;
 547         }
 548
 549         return p;
 550
 551  bad:
 552         free_pgpath(p);
 553         return NULL;
 554 }
 555
 556 static struct priority_group *parse_priority_group(struct arg_set *as,
 557                                                    struct multipath *m)
 558 {
 559         static struct param _params[] = {
 560                 {1, 1024, "invalid number of paths"},
 561                 {0, 1024, "invalid number of selector args"}
 562         };
 563
 564         int r;
 565         unsigned i, nr_selector_args, nr_params;
 566         struct priority_group *pg;
 567         struct dm_target *ti = m->ti;
 568
 569         if (as->argc < 2) {
 570                 as->argc = 0;
 571                 ti->error = "not enough priority group aruments";
 572                 return NULL;
 573         }
 574
 575         pg = alloc_priority_group();
 576         if (!pg) {
 577                 ti->error = "couldn't allocate priority group";
 578                 return NULL;
 579         }
 580         pg->m = m;
 581
 582         r = parse_path_selector(as, pg, ti);
 583         if (r)
 584                 goto bad;
 585
 586         /*
 587          * read the paths
 588          */
 589         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 590         if (r)
 591                 goto bad;
 592
 593         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 594         if (r)
 595                 goto bad;
 596
 597         nr_params = 1 + nr_selector_args;
 598         for (i = 0; i < pg->nr_pgpaths; i++) {
 599                 struct pgpath *pgpath;
 600                 struct arg_set path_args;
 601
 602                 if (as->argc < nr_params)
 603                         goto bad;
 604
 605                 path_args.argc = nr_params;
 606                 path_args.argv = as->argv;
 607
 608                 pgpath = parse_path(&path_args, &pg->ps, ti);
 609                 if (!pgpath)
 610                         goto bad;
 611
 612                 pgpath->pg = pg;
 613                 list_add_tail(&pgpath->list, &pg->pgpaths);
 614                 consume(as, nr_params);
 615         }
 616
 617         return pg;
 618
 619  bad:
 620         free_priority_group(pg, ti);
 621         return NULL;
 622 }
 623
 624 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 625 {
 626         int r;
 627         struct hw_handler_type *hwht;
 628         unsigned hw_argc;
 629         struct dm_target *ti = m->ti;
 630
 631         static struct param _params[] = {
 632                 {0, 1024, "invalid number of hardware handler args"},
 633         };
 634
 635         r = read_param(_params, shift(as), &hw_argc, &ti->error);
 636         if (r)
 637                 return -EINVAL;
 638
 639         if (!hw_argc)
 640                 return 0;
 641
 642         hwht = dm_get_hw_handler(shift(as));
 643         if (!hwht) {
 644                 ti->error = "unknown hardware handler type";
 645                 return -EINVAL;
 646         }
 647
 648         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
 649         if (r) {
 650                 dm_put_hw_handler(hwht);
 651                 ti->error = "hardware handler constructor failed";
 652                 return r;
 653         }
 654
 655         m->hw_handler.type = hwht;
 656         consume(as, hw_argc - 1);
 657
 658         return 0;
 659 }
 660
 661 static int parse_features(struct arg_set *as, struct multipath *m)
 662 {
 663         int r;
 664         unsigned argc;
 665         struct dm_target *ti = m->ti;
 666
 667         static struct param _params[] = {
 668                 {0, 1, "invalid number of feature args"},
 669         };
 670
 671         r = read_param(_params, shift(as), &argc, &ti->error);
 672         if (r)
 673                 return -EINVAL;
 674
 675         if (!argc)
 676                 return 0;
 677
 678         if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
 679                 return queue_if_no_path(m, 1, 0);
 680         else {
 681                 ti->error = "Unrecognised multipath feature request";
 682                 return -EINVAL;
 683         }
 684 }
 685
 686 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 687                          char **argv)
 688 {
 689         /* target parameters */
 690         static struct param _params[] = {
 691                 {1, 1024, "invalid number of priority groups"},
 692                 {1, 1024, "invalid initial priority group number"},
 693         };
 694
 695         int r;
 696         struct multipath *m;
 697         struct arg_set as;
 698         unsigned pg_count = 0;
 699         unsigned next_pg_num;
 700
 701         as.argc = argc;
 702         as.argv = argv;
 703
 704         m = alloc_multipath(ti);
 705         if (!m) {
 706                 ti->error = "can't allocate multipath";
 707                 return -EINVAL;
 708         }
 709
 710         r = parse_features(&as, m);
 711         if (r)
 712                 goto bad;
 713
 714         r = parse_hw_handler(&as, m);
 715         if (r)
 716                 goto bad;
 717
 718         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 719         if (r)
 720                 goto bad;
 721
 722         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 723         if (r)
 724                 goto bad;
 725
 726         /* parse the priority groups */
 727         while (as.argc) {
 728                 struct priority_group *pg;
 729
 730                 pg = parse_priority_group(&as, m);
 731                 if (!pg) {
 732                         r = -EINVAL;
 733                         goto bad;
 734                 }
 735
 736                 m->nr_valid_paths += pg->nr_pgpaths;
 737                 list_add_tail(&pg->list, &m->priority_groups);
 738                 pg_count++;
 739                 pg->pg_num = pg_count;
 740                 if (!--next_pg_num)
 741                         m->next_pg = pg;
 742         }
 743
 744         if (pg_count != m->nr_priority_groups) {
 745                 ti->error = "priority group count mismatch";
 746                 r = -EINVAL;
 747                 goto bad;
 748         }
 749
 750         return 0;
 751
 752  bad:
 753         free_multipath(m);
 754         return r;
 755 }
 756
 757 static void multipath_dtr(struct dm_target *ti)
 758 {
 759         struct multipath *m = (struct multipath *) ti->private;
 760
 761         flush_workqueue(kmultipathd);
 762         free_multipath(m);
 763 }
 764
 765 /*
 766  * Map bios, recording original fields for later in case we have to resubmit
 767  */
 768 static int multipath_map(struct dm_target *ti, struct bio *bio,
 769                          union map_info *map_context)
 770 {
 771         int r;
 772         struct mpath_io *mpio;
 773         struct multipath *m = (struct multipath *) ti->private;
 774
 775         if (bio_barrier(bio))
 776                 return -EOPNOTSUPP;
 777
 778         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 779         dm_bio_record(&mpio->details, bio);
 780
 781         map_context->ptr = mpio;
 782         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 783         r = map_io(m, bio, mpio, 0);
 784         if (r < 0)
 785                 mempool_free(mpio, m->mpio_pool);
 786
 787         return r;
 788 }
 789
 790 /*
 791  * Take a path out of use.
 792  */
 793 static int fail_path(struct pgpath *pgpath)
 794 {
 795         unsigned long flags;
 796         struct multipath *m = pgpath->pg->m;
 797
 798         spin_lock_irqsave(&m->lock, flags);
 799
 800         if (!pgpath->path.is_active)
 801                 goto out;
 802
 803         DMWARN("Failing path %s.", pgpath->path.dev->name);
 804
 805         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 806         pgpath->path.is_active = 0;
 807         pgpath->fail_count++;
 808
 809         m->nr_valid_paths--;
 810
 811         if (pgpath == m->current_pgpath)
 812                 m->current_pgpath = NULL;
 813
 814         queue_work(kmultipathd, &m->trigger_event);
 815
 816 out:
 817         spin_unlock_irqrestore(&m->lock, flags);
 818
 819         return 0;
 820 }
 821
 822 /*
 823  * Reinstate a previously-failed path
 824  */
 825 static int reinstate_path(struct pgpath *pgpath)
 826 {
 827         int r = 0;
 828         unsigned long flags;
 829         struct multipath *m = pgpath->pg->m;
 830
 831         spin_lock_irqsave(&m->lock, flags);
 832
 833         if (pgpath->path.is_active)
 834                 goto out;
 835
 836         if (!pgpath->pg->ps.type) {
 837                 DMWARN("Reinstate path not supported by path selector %s",
 838                        pgpath->pg->ps.type->name);
 839                 r = -EINVAL;
 840                 goto out;
 841         }
 842
 843         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 844         if (r)
 845                 goto out;
 846
 847         pgpath->path.is_active = 1;
 848
 849         m->current_pgpath = NULL;
 850         if (!m->nr_valid_paths++ && m->queue_size)
 851                 queue_work(kmultipathd, &m->process_queued_ios);
 852
 853         queue_work(kmultipathd, &m->trigger_event);
 854
 855 out:
 856         spin_unlock_irqrestore(&m->lock, flags);
 857
 858         return r;
 859 }
 860
 861 /*
 862  * Fail or reinstate all paths that match the provided struct dm_dev.
 863  */
 864 static int action_dev(struct multipath *m, struct dm_dev *dev,
 865                       action_fn action)
 866 {
 867         int r = 0;
 868         struct pgpath *pgpath;
 869         struct priority_group *pg;
 870
 871         list_for_each_entry(pg, &m->priority_groups, list) {
 872                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 873                         if (pgpath->path.dev == dev)
 874                                 r = action(pgpath);
 875                 }
 876         }
 877
 878         return r;
 879 }
 880
 881 /*
 882  * Temporarily try to avoid having to use the specified PG
 883  */
 884 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 885                       int bypassed)
 886 {
 887         unsigned long flags;
 888
 889         spin_lock_irqsave(&m->lock, flags);
 890
 891         pg->bypassed = bypassed;
 892         m->current_pgpath = NULL;
 893         m->current_pg = NULL;
 894
 895         spin_unlock_irqrestore(&m->lock, flags);
 896
 897         queue_work(kmultipathd, &m->trigger_event);
 898 }
 899
 900 /*
 901  * Switch to using the specified PG from the next I/O that gets mapped
 902  */
 903 static int switch_pg_num(struct multipath *m, const char *pgstr)
 904 {
 905         struct priority_group *pg;
 906         unsigned pgnum;
 907         unsigned long flags;
 908
 909         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 910             (pgnum > m->nr_priority_groups)) {
 911                 DMWARN("invalid PG number supplied to switch_pg_num");
 912                 return -EINVAL;
 913         }
 914
 915         spin_lock_irqsave(&m->lock, flags);
 916         list_for_each_entry(pg, &m->priority_groups, list) {
 917                 pg->bypassed = 0;
 918                 if (--pgnum)
 919                         continue;
 920
 921                 m->current_pgpath = NULL;
 922                 m->current_pg = NULL;
 923                 m->next_pg = pg;
 924         }
 925         spin_unlock_irqrestore(&m->lock, flags);
 926
 927         queue_work(kmultipathd, &m->trigger_event);
 928         return 0;
 929 }
 930
 931 /*
 932  * Set/clear bypassed status of a PG.
 933  * PGs are numbered upwards from 1 in the order they were declared.
 934  */
 935 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 936 {
 937         struct priority_group *pg;
 938         unsigned pgnum;
 939
 940         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 941             (pgnum > m->nr_priority_groups)) {
 942                 DMWARN("invalid PG number supplied to bypass_pg");
 943                 return -EINVAL;
 944         }
 945
 946         list_for_each_entry(pg, &m->priority_groups, list) {
 947                 if (!--pgnum)
 948                         break;
 949         }
 950
 951         bypass_pg(m, pg, bypassed);
 952         return 0;
 953 }
 954
 955 /*
 956  * pg_init must call this when it has completed its initialisation
 957  */
 958 void dm_pg_init_complete(struct path *path, unsigned err_flags)
 959 {
 960         struct pgpath *pgpath = path_to_pgpath(path);
 961         struct priority_group *pg = pgpath->pg;
 962         struct multipath *m = pg->m;
 963         unsigned long flags;
 964
 965         /* We insist on failing the path if the PG is already bypassed. */
 966         if (err_flags && pg->bypassed)
 967                 err_flags |= MP_FAIL_PATH;
 968
 969         if (err_flags & MP_FAIL_PATH)
 970                 fail_path(pgpath);
 971
 972         if (err_flags & MP_BYPASS_PG)
 973                 bypass_pg(m, pg, 1);
 974
 975         spin_lock_irqsave(&m->lock, flags);
 976         if (err_flags) {
 977                 m->current_pgpath = NULL;
 978                 m->current_pg = NULL;
 979         } else if (!m->pg_init_required)
 980                 m->queue_io = 0;
 981
 982         m->pg_init_in_progress = 0;
 983         queue_work(kmultipathd, &m->process_queued_ios);
 984         spin_unlock_irqrestore(&m->lock, flags);
 985 }
 986
 987 /*
 988  * end_io handling
 989  */
 990 static int do_end_io(struct multipath *m, struct bio *bio,
 991                      int error, struct mpath_io *mpio)
 992 {
 993         struct hw_handler *hwh = &m->hw_handler;
 994         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
 995         unsigned long flags;
 996
 997         if (!error)
 998                 return 0;       /* I/O complete */
 999
1000         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1001                 return error;
1002
1003         if (error == -EOPNOTSUPP)
1004                 return error;
1005
1006         spin_lock_irqsave(&m->lock, flags);
1007         if (!m->nr_valid_paths) {
1008                 if (!m->queue_if_no_path) {
1009                         spin_unlock_irqrestore(&m->lock, flags);
1010                         return -EIO;
1011                 } else {
1012                         spin_unlock_irqrestore(&m->lock, flags);
1013                         goto requeue;
1014                 }
1015         }
1016         spin_unlock_irqrestore(&m->lock, flags);
1017
1018         if (hwh->type && hwh->type->error)
1019                 err_flags = hwh->type->error(hwh, bio);
1020
1021         if (mpio->pgpath) {
1022                 if (err_flags & MP_FAIL_PATH)
1023                         fail_path(mpio->pgpath);
1024
1025                 if (err_flags & MP_BYPASS_PG)
1026                         bypass_pg(m, mpio->pgpath->pg, 1);
1027         }
1028
1029         if (err_flags & MP_ERROR_IO)
1030                 return -EIO;
1031
1032       requeue:
1033         dm_bio_restore(&mpio->details, bio);
1034
1035         /* queue for the daemon to resubmit or fail */
1036         spin_lock_irqsave(&m->lock, flags);
1037         bio_list_add(&m->queued_ios, bio);
1038         m->queue_size++;
1039         if (!m->queue_io)
1040                 queue_work(kmultipathd, &m->process_queued_ios);
1041         spin_unlock_irqrestore(&m->lock, flags);
1042
1043         return 1;       /* io not complete */
1044 }
1045
1046 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1047                             int error, union map_info *map_context)
1048 {
1049         struct multipath *m = (struct multipath *) ti->private;
1050         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1051         struct pgpath *pgpath = mpio->pgpath;
1052         struct path_selector *ps;
1053         int r;
1054
1055         r  = do_end_io(m, bio, error, mpio);
1056         if (pgpath) {
1057                 ps = &pgpath->pg->ps;
1058                 if (ps->type->end_io)
1059                         ps->type->end_io(ps, &pgpath->path);
1060         }
1061         if (r <= 0)
1062                 mempool_free(mpio, m->mpio_pool);
1063
1064         return r;
1065 }
1066
1067 /*
1068  * Suspend can't complete until all the I/O is processed so if
1069  * the last path fails we must error any remaining I/O.
1070  * Note that if the freeze_bdev fails while suspending, the
1071  * queue_if_no_path state is lost - userspace should reset it.
1072  */
1073 static void multipath_presuspend(struct dm_target *ti)
1074 {
1075         struct multipath *m = (struct multipath *) ti->private;
1076
1077         queue_if_no_path(m, 0, 1);
1078 }
1079
1080 /*
1081  * Restore the queue_if_no_path setting.
1082  */
1083 static void multipath_resume(struct dm_target *ti)
1084 {
1085         struct multipath *m = (struct multipath *) ti->private;
1086         unsigned long flags;
1087
1088         spin_lock_irqsave(&m->lock, flags);
1089         m->queue_if_no_path = m->saved_queue_if_no_path;
1090         spin_unlock_irqrestore(&m->lock, flags);
1091 }
1092
1093 /*
1094  * Info output has the following format:
1095  * num_multipath_feature_args [multipath_feature_args]*
1096  * num_handler_status_args [handler_status_args]*
1097  * num_groups init_group_number
1098  *            [A|D|E num_ps_status_args [ps_status_args]*
1099  *             num_paths num_selector_args
1100  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1101  *
1102  * Table output has the following format (identical to the constructor string):
1103  * num_feature_args [features_args]*
1104  * num_handler_args hw_handler [hw_handler_args]*
1105  * num_groups init_group_number
1106  *     [priority selector-name num_ps_args [ps_args]*
1107  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1108  */
1109 static int multipath_status(struct dm_target *ti, status_type_t type,
1110                             char *result, unsigned int maxlen)
1111 {
1112         int sz = 0;
1113         unsigned long flags;
1114         struct multipath *m = (struct multipath *) ti->private;
1115         struct hw_handler *hwh = &m->hw_handler;
1116         struct priority_group *pg;
1117         struct pgpath *p;
1118         unsigned pg_num;
1119         char state;
1120
1121         spin_lock_irqsave(&m->lock, flags);
1122
1123         /* Features */
1124         if (type == STATUSTYPE_INFO)
1125                 DMEMIT("1 %u ", m->queue_size);
1126         else if (m->queue_if_no_path)
1127                 DMEMIT("1 queue_if_no_path ");
1128         else
1129                 DMEMIT("0 ");
1130
1131         if (hwh->type && hwh->type->status)
1132                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1133         else if (!hwh->type || type == STATUSTYPE_INFO)
1134                 DMEMIT("0 ");
1135         else
1136                 DMEMIT("1 %s ", hwh->type->name);
1137
1138         DMEMIT("%u ", m->nr_priority_groups);
1139
1140         if (m->next_pg)
1141                 pg_num = m->next_pg->pg_num;
1142         else if (m->current_pg)
1143                 pg_num = m->current_pg->pg_num;
1144         else
1145                         pg_num = 1;
1146
1147         DMEMIT("%u ", pg_num);
1148
1149         switch (type) {
1150         case STATUSTYPE_INFO:
1151                 list_for_each_entry(pg, &m->priority_groups, list) {
1152                         if (pg->bypassed)
1153                                 state = 'D';    /* Disabled */
1154                         else if (pg == m->current_pg)
1155                                 state = 'A';    /* Currently Active */
1156                         else
1157                                 state = 'E';    /* Enabled */
1158
1159                         DMEMIT("%c ", state);
1160
1161                         if (pg->ps.type->status)
1162                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1163                                                           result + sz,
1164                                                           maxlen - sz);
1165                         else
1166                                 DMEMIT("0 ");
1167
1168                         DMEMIT("%u %u ", pg->nr_pgpaths,
1169                                pg->ps.type->info_args);
1170
1171                         list_for_each_entry(p, &pg->pgpaths, list) {
1172                                 DMEMIT("%s %s %u ", p->path.dev->name,
1173                                        p->path.is_active ? "A" : "F",
1174                                        p->fail_count);
1175                                 if (pg->ps.type->status)
1176                                         sz += pg->ps.type->status(&pg->ps,
1177                                               &p->path, type, result + sz,
1178                                               maxlen - sz);
1179                         }
1180                 }
1181                 break;
1182
1183         case STATUSTYPE_TABLE:
1184                 list_for_each_entry(pg, &m->priority_groups, list) {
1185                         DMEMIT("%s ", pg->ps.type->name);
1186
1187                         if (pg->ps.type->status)
1188                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1189                                                           result + sz,
1190                                                           maxlen - sz);
1191                         else
1192                                 DMEMIT("0 ");
1193
1194                         DMEMIT("%u %u ", pg->nr_pgpaths,
1195                                pg->ps.type->table_args);
1196
1197                         list_for_each_entry(p, &pg->pgpaths, list) {
1198                                 DMEMIT("%s ", p->path.dev->name);
1199                                 if (pg->ps.type->status)
1200                                         sz += pg->ps.type->status(&pg->ps,
1201                                               &p->path, type, result + sz,
1202                                               maxlen - sz);
1203                         }
1204                 }
1205                 break;
1206         }
1207
1208         spin_unlock_irqrestore(&m->lock, flags);
1209
1210         return 0;
1211 }
1212
1213 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1214 {
1215         int r;
1216         struct dm_dev *dev;
1217         struct multipath *m = (struct multipath *) ti->private;
1218         action_fn action;
1219
1220         if (argc == 1) {
1221                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1222                         return queue_if_no_path(m, 1, 0);
1223                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1224                         return queue_if_no_path(m, 0, 0);
1225         }
1226
1227         if (argc != 2)
1228                 goto error;
1229
1230         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1231                 return bypass_pg_num(m, argv[1], 1);
1232         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1233                 return bypass_pg_num(m, argv[1], 0);
1234         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1235                 return switch_pg_num(m, argv[1]);
1236         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1237                 action = reinstate_path;
1238         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1239                 action = fail_path;
1240         else
1241                 goto error;
1242
1243         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1244                           dm_table_get_mode(ti->table), &dev);
1245         if (r) {
1246                 DMWARN("message: error getting device %s",
1247                        argv[1]);
1248                 return -EINVAL;
1249         }
1250
1251         r = action_dev(m, dev, action);
1252
1253         dm_put_device(ti, dev);
1254
1255         return r;
1256
1257 error:
1258         DMWARN("Unrecognised multipath message received.");
1259         return -EINVAL;
1260 }
1261
1262 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1263                            struct file *filp, unsigned int cmd,
1264                            unsigned long arg)
1265 {
1266         struct multipath *m = (struct multipath *) ti->private;
1267         struct block_device *bdev = NULL;
1268         unsigned long flags;
1269         struct file fake_file = {};
1270         struct dentry fake_dentry = {};
1271         int r = 0;
1272
1273         fake_file.f_dentry = &fake_dentry;
1274
1275         spin_lock_irqsave(&m->lock, flags);
1276
1277         if (!m->current_pgpath)
1278                 __choose_pgpath(m);
1279
1280         if (m->current_pgpath) {
1281                 bdev = m->current_pgpath->path.dev->bdev;
1282                 fake_dentry.d_inode = bdev->bd_inode;
1283                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1284         }
1285
1286         if (m->queue_io)
1287                 r = -EAGAIN;
1288         else if (!bdev)
1289                 r = -EIO;
1290
1291         spin_unlock_irqrestore(&m->lock, flags);
1292
1293         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1294                                          bdev->bd_disk, cmd, arg);
1295 }
1296
1297 /*-----------------------------------------------------------------
1298  * Module setup
1299  *---------------------------------------------------------------*/
1300 static struct target_type multipath_target = {
1301         .name = "multipath",
1302         .version = {1, 0, 5},
1303         .module = THIS_MODULE,
1304         .ctr = multipath_ctr,
1305         .dtr = multipath_dtr,
1306         .map = multipath_map,
1307         .end_io = multipath_end_io,
1308         .presuspend = multipath_presuspend,
1309         .resume = multipath_resume,
1310         .status = multipath_status,
1311         .message = multipath_message,
1312         .ioctl  = multipath_ioctl,
1313 };
1314
1315 static int __init dm_multipath_init(void)
1316 {
1317         int r;
1318
1319         /* allocate a slab for the dm_ios */
1320         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1321                                         0, 0, NULL, NULL);
1322         if (!_mpio_cache)
1323                 return -ENOMEM;
1324
1325         r = dm_register_target(&multipath_target);
1326         if (r < 0) {
1327                 DMERR("%s: register failed %d", multipath_target.name, r);
1328                 kmem_cache_destroy(_mpio_cache);
1329                 return -EINVAL;
1330         }
1331
1332         kmultipathd = create_workqueue("kmpathd");
1333         if (!kmultipathd) {
1334                 DMERR("%s: failed to create workqueue kmpathd",
1335                                 multipath_target.name);
1336                 dm_unregister_target(&multipath_target);
1337                 kmem_cache_destroy(_mpio_cache);
1338                 return -ENOMEM;
1339         }
1340
1341         DMINFO("version %u.%u.%u loaded",
1342                multipath_target.version[0], multipath_target.version[1],
1343                multipath_target.version[2]);
1344
1345         return r;
1346 }
1347
1348 static void __exit dm_multipath_exit(void)
1349 {
1350         int r;
1351
1352         destroy_workqueue(kmultipathd);
1353
1354         r = dm_unregister_target(&multipath_target);
1355         if (r < 0)
1356                 DMERR("%s: target unregister failed %d",
1357                       multipath_target.name, r);
1358         kmem_cache_destroy(_mpio_cache);
1359 }
1360
1361 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1362
1363 module_init(dm_multipath_init);
1364 module_exit(dm_multipath_exit);
1365
1366 MODULE_DESCRIPTION(DM_NAME " multipath target");
1367 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1368 MODULE_LICENSE("GPL");