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 #include "dm-uevent.h"
  14
  15 #include <linux/ctype.h>
  16 #include <linux/init.h>
  17 #include <linux/mempool.h>
  18 #include <linux/module.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/slab.h>
  21 #include <linux/time.h>
  22 #include <linux/workqueue.h>
  23 #include <asm/atomic.h>
  24
  25 #define DM_MSG_PREFIX "multipath"
  26 #define MESG_STR(x) x, sizeof(x)
  27
  28 /* Path properties */
  29 struct pgpath {
  30         struct list_head list;
  31
  32         struct priority_group *pg;      /* Owning PG */
  33         unsigned fail_count;            /* Cumulative failure count */
  34
  35         struct dm_path path;
  36 };
  37
  38 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  39
  40 /*
  41  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  42  * Each has a path selector which controls which path gets used.
  43  */
  44 struct priority_group {
  45         struct list_head list;
  46
  47         struct multipath *m;            /* Owning multipath instance */
  48         struct path_selector ps;
  49
  50         unsigned pg_num;                /* Reference number */
  51         unsigned bypassed;              /* Temporarily bypass this PG? */
  52
  53         unsigned nr_pgpaths;            /* Number of paths in PG */
  54         struct list_head pgpaths;
  55 };
  56
  57 /* Multipath context */
  58 struct multipath {
  59         struct list_head list;
  60         struct dm_target *ti;
  61
  62         spinlock_t lock;
  63
  64         struct hw_handler hw_handler;
  65         unsigned nr_priority_groups;
  66         struct list_head priority_groups;
  67         unsigned pg_init_required;      /* pg_init needs calling? */
  68         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  69
  70         unsigned nr_valid_paths;        /* Total number of usable paths */
  71         struct pgpath *current_pgpath;
  72         struct priority_group *current_pg;
  73         struct priority_group *next_pg; /* Switch to this PG if set */
  74         unsigned repeat_count;          /* I/Os left before calling PS again */
  75
  76         unsigned queue_io;              /* Must we queue all I/O? */
  77         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  78         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  79         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  80         unsigned pg_init_count;         /* Number of times pg_init called */
  81
  82         struct work_struct process_queued_ios;
  83         struct bio_list queued_ios;
  84         unsigned queue_size;
  85
  86         struct work_struct trigger_event;
  87
  88         /*
  89          * We must use a mempool of dm_mpath_io structs so that we
  90          * can resubmit bios on error.
  91          */
  92         mempool_t *mpio_pool;
  93 };
  94
  95 /*
  96  * Context information attached to each bio we process.
  97  */
  98 struct dm_mpath_io {
  99         struct pgpath *pgpath;
 100         struct dm_bio_details details;
 101 };
 102
 103 typedef int (*action_fn) (struct pgpath *pgpath);
 104
 105 #define MIN_IOS 256     /* Mempool size */
 106
 107 static struct kmem_cache *_mpio_cache;
 108
 109 static struct workqueue_struct *kmultipathd;
 110 static void process_queued_ios(struct work_struct *work);
 111 static void trigger_event(struct work_struct *work);
 112
 113
 114 /*-----------------------------------------------
 115  * Allocation routines
 116  *-----------------------------------------------*/
 117
 118 static struct pgpath *alloc_pgpath(void)
 119 {
 120         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 121
 122         if (pgpath)
 123                 pgpath->path.is_active = 1;
 124
 125         return pgpath;
 126 }
 127
 128 static void free_pgpath(struct pgpath *pgpath)
 129 {
 130         kfree(pgpath);
 131 }
 132
 133 static struct priority_group *alloc_priority_group(void)
 134 {
 135         struct priority_group *pg;
 136
 137         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 138
 139         if (pg)
 140                 INIT_LIST_HEAD(&pg->pgpaths);
 141
 142         return pg;
 143 }
 144
 145 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 146 {
 147         struct pgpath *pgpath, *tmp;
 148
 149         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 150                 list_del(&pgpath->list);
 151                 dm_put_device(ti, pgpath->path.dev);
 152                 free_pgpath(pgpath);
 153         }
 154 }
 155
 156 static void free_priority_group(struct priority_group *pg,
 157                                 struct dm_target *ti)
 158 {
 159         struct path_selector *ps = &pg->ps;
 160
 161         if (ps->type) {
 162                 ps->type->destroy(ps);
 163                 dm_put_path_selector(ps->type);
 164         }
 165
 166         free_pgpaths(&pg->pgpaths, ti);
 167         kfree(pg);
 168 }
 169
 170 static struct multipath *alloc_multipath(struct dm_target *ti)
 171 {
 172         struct multipath *m;
 173
 174         m = kzalloc(sizeof(*m), GFP_KERNEL);
 175         if (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);
 180                 INIT_WORK(&m->trigger_event, trigger_event);
 181                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 182                 if (!m->mpio_pool) {
 183                         kfree(m);
 184                         return NULL;
 185                 }
 186                 m->ti = ti;
 187                 ti->private = m;
 188         }
 189
 190         return m;
 191 }
 192
 193 static void free_multipath(struct multipath *m)
 194 {
 195         struct priority_group *pg, *tmp;
 196         struct hw_handler *hwh = &m->hw_handler;
 197
 198         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 199                 list_del(&pg->list);
 200                 free_priority_group(pg, m->ti);
 201         }
 202
 203         if (hwh->type) {
 204                 hwh->type->destroy(hwh);
 205                 dm_put_hw_handler(hwh->type);
 206         }
 207
 208         mempool_destroy(m->mpio_pool);
 209         kfree(m);
 210 }
 211
 212
 213 /*-----------------------------------------------
 214  * Path selection
 215  *-----------------------------------------------*/
 216
 217 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 218 {
 219         struct hw_handler *hwh = &m->hw_handler;
 220
 221         m->current_pg = pgpath->pg;
 222
 223         /* Must we initialise the PG first, and queue I/O till it's ready? */
 224         if (hwh->type && hwh->type->pg_init) {
 225                 m->pg_init_required = 1;
 226                 m->queue_io = 1;
 227         } else {
 228                 m->pg_init_required = 0;
 229                 m->queue_io = 0;
 230         }
 231
 232         m->pg_init_count = 0;
 233 }
 234
 235 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
 236 {
 237         struct dm_path *path;
 238
 239         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
 240         if (!path)
 241                 return -ENXIO;
 242
 243         m->current_pgpath = path_to_pgpath(path);
 244
 245         if (m->current_pg != pg)
 246                 __switch_pg(m, m->current_pgpath);
 247
 248         return 0;
 249 }
 250
 251 static void __choose_pgpath(struct multipath *m)
 252 {
 253         struct priority_group *pg;
 254         unsigned bypassed = 1;
 255
 256         if (!m->nr_valid_paths)
 257                 goto failed;
 258
 259         /* Were we instructed to switch PG? */
 260         if (m->next_pg) {
 261                 pg = m->next_pg;
 262                 m->next_pg = NULL;
 263                 if (!__choose_path_in_pg(m, pg))
 264                         return;
 265         }
 266
 267         /* Don't change PG until it has no remaining paths */
 268         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
 269                 return;
 270
 271         /*
 272          * Loop through priority groups until we find a valid path.
 273          * First time we skip PGs marked 'bypassed'.
 274          * Second time we only try the ones we skipped.
 275          */
 276         do {
 277                 list_for_each_entry(pg, &m->priority_groups, list) {
 278                         if (pg->bypassed == bypassed)
 279                                 continue;
 280                         if (!__choose_path_in_pg(m, pg))
 281                                 return;
 282                 }
 283         } while (bypassed--);
 284
 285 failed:
 286         m->current_pgpath = NULL;
 287         m->current_pg = NULL;
 288 }
 289
 290 /*
 291  * Check whether bios must be queued in the device-mapper core rather
 292  * than here in the target.
 293  *
 294  * m->lock must be held on entry.
 295  *
 296  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 297  * same value then we are not between multipath_presuspend()
 298  * and multipath_resume() calls and we have no need to check
 299  * for the DMF_NOFLUSH_SUSPENDING flag.
 300  */
 301 static int __must_push_back(struct multipath *m)
 302 {
 303         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 304                 dm_noflush_suspending(m->ti));
 305 }
 306
 307 static int map_io(struct multipath *m, struct bio *bio,
 308                   struct dm_mpath_io *mpio, unsigned was_queued)
 309 {
 310         int r = DM_MAPIO_REMAPPED;
 311         unsigned long flags;
 312         struct pgpath *pgpath;
 313
 314         spin_lock_irqsave(&m->lock, flags);
 315
 316         /* Do we need to select a new pgpath? */
 317         if (!m->current_pgpath ||
 318             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 319                 __choose_pgpath(m);
 320
 321         pgpath = m->current_pgpath;
 322
 323         if (was_queued)
 324                 m->queue_size--;
 325
 326         if ((pgpath && m->queue_io) ||
 327             (!pgpath && m->queue_if_no_path)) {
 328                 /* Queue for the daemon to resubmit */
 329                 bio_list_add(&m->queued_ios, bio);
 330                 m->queue_size++;
 331                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 332                     !m->queue_io)
 333                         queue_work(kmultipathd, &m->process_queued_ios);
 334                 pgpath = NULL;
 335                 r = DM_MAPIO_SUBMITTED;
 336         } else if (pgpath)
 337                 bio->bi_bdev = pgpath->path.dev->bdev;
 338         else if (__must_push_back(m))
 339                 r = DM_MAPIO_REQUEUE;
 340         else
 341                 r = -EIO;       /* Failed */
 342
 343         mpio->pgpath = pgpath;
 344
 345         spin_unlock_irqrestore(&m->lock, flags);
 346
 347         return r;
 348 }
 349
 350 /*
 351  * If we run out of usable paths, should we queue I/O or error it?
 352  */
 353 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 354                             unsigned save_old_value)
 355 {
 356         unsigned long flags;
 357
 358         spin_lock_irqsave(&m->lock, flags);
 359
 360         if (save_old_value)
 361                 m->saved_queue_if_no_path = m->queue_if_no_path;
 362         else
 363                 m->saved_queue_if_no_path = queue_if_no_path;
 364         m->queue_if_no_path = queue_if_no_path;
 365         if (!m->queue_if_no_path && m->queue_size)
 366                 queue_work(kmultipathd, &m->process_queued_ios);
 367
 368         spin_unlock_irqrestore(&m->lock, flags);
 369
 370         return 0;
 371 }
 372
 373 /*-----------------------------------------------------------------
 374  * The multipath daemon is responsible for resubmitting queued ios.
 375  *---------------------------------------------------------------*/
 376
 377 static void dispatch_queued_ios(struct multipath *m)
 378 {
 379         int r;
 380         unsigned long flags;
 381         struct bio *bio = NULL, *next;
 382         struct dm_mpath_io *mpio;
 383         union map_info *info;
 384
 385         spin_lock_irqsave(&m->lock, flags);
 386         bio = bio_list_get(&m->queued_ios);
 387         spin_unlock_irqrestore(&m->lock, flags);
 388
 389         while (bio) {
 390                 next = bio->bi_next;
 391                 bio->bi_next = NULL;
 392
 393                 info = dm_get_mapinfo(bio);
 394                 mpio = info->ptr;
 395
 396                 r = map_io(m, bio, mpio, 1);
 397                 if (r < 0)
 398                         bio_endio(bio, r);
 399                 else if (r == DM_MAPIO_REMAPPED)
 400                         generic_make_request(bio);
 401                 else if (r == DM_MAPIO_REQUEUE)
 402                         bio_endio(bio, -EIO);
 403
 404                 bio = next;
 405         }
 406 }
 407
 408 static void process_queued_ios(struct work_struct *work)
 409 {
 410         struct multipath *m =
 411                 container_of(work, struct multipath, process_queued_ios);
 412         struct hw_handler *hwh = &m->hw_handler;
 413         struct pgpath *pgpath = NULL;
 414         unsigned init_required = 0, must_queue = 1;
 415         unsigned long flags;
 416
 417         spin_lock_irqsave(&m->lock, flags);
 418
 419         if (!m->queue_size)
 420                 goto out;
 421
 422         if (!m->current_pgpath)
 423                 __choose_pgpath(m);
 424
 425         pgpath = m->current_pgpath;
 426
 427         if ((pgpath && !m->queue_io) ||
 428             (!pgpath && !m->queue_if_no_path))
 429                 must_queue = 0;
 430
 431         if (m->pg_init_required && !m->pg_init_in_progress) {
 432                 m->pg_init_count++;
 433                 m->pg_init_required = 0;
 434                 m->pg_init_in_progress = 1;
 435                 init_required = 1;
 436         }
 437
 438 out:
 439         spin_unlock_irqrestore(&m->lock, flags);
 440
 441         if (init_required)
 442                 hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
 443
 444         if (!must_queue)
 445                 dispatch_queued_ios(m);
 446 }
 447
 448 /*
 449  * An event is triggered whenever a path is taken out of use.
 450  * Includes path failure and PG bypass.
 451  */
 452 static void trigger_event(struct work_struct *work)
 453 {
 454         struct multipath *m =
 455                 container_of(work, struct multipath, trigger_event);
 456
 457         dm_table_event(m->ti->table);
 458 }
 459
 460 /*-----------------------------------------------------------------
 461  * Constructor/argument parsing:
 462  * <#multipath feature args> [<arg>]*
 463  * <#hw_handler args> [hw_handler [<arg>]*]
 464  * <#priority groups>
 465  * <initial priority group>
 466  *     [<selector> <#selector args> [<arg>]*
 467  *      <#paths> <#per-path selector args>
 468  *         [<path> [<arg>]* ]+ ]+
 469  *---------------------------------------------------------------*/
 470 struct param {
 471         unsigned min;
 472         unsigned max;
 473         char *error;
 474 };
 475
 476 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 477 {
 478         if (!str ||
 479             (sscanf(str, "%u", v) != 1) ||
 480             (*v < param->min) ||
 481             (*v > param->max)) {
 482                 *error = param->error;
 483                 return -EINVAL;
 484         }
 485
 486         return 0;
 487 }
 488
 489 struct arg_set {
 490         unsigned argc;
 491         char **argv;
 492 };
 493
 494 static char *shift(struct arg_set *as)
 495 {
 496         char *r;
 497
 498         if (as->argc) {
 499                 as->argc--;
 500                 r = *as->argv;
 501                 as->argv++;
 502                 return r;
 503         }
 504
 505         return NULL;
 506 }
 507
 508 static void consume(struct arg_set *as, unsigned n)
 509 {
 510         BUG_ON (as->argc < n);
 511         as->argc -= n;
 512         as->argv += n;
 513 }
 514
 515 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 516                                struct dm_target *ti)
 517 {
 518         int r;
 519         struct path_selector_type *pst;
 520         unsigned ps_argc;
 521
 522         static struct param _params[] = {
 523                 {0, 1024, "invalid number of path selector args"},
 524         };
 525
 526         pst = dm_get_path_selector(shift(as));
 527         if (!pst) {
 528                 ti->error = "unknown path selector type";
 529                 return -EINVAL;
 530         }
 531
 532         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 533         if (r)
 534                 return -EINVAL;
 535
 536         r = pst->create(&pg->ps, ps_argc, as->argv);
 537         if (r) {
 538                 dm_put_path_selector(pst);
 539                 ti->error = "path selector constructor failed";
 540                 return r;
 541         }
 542
 543         pg->ps.type = pst;
 544         consume(as, ps_argc);
 545
 546         return 0;
 547 }
 548
 549 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 550                                struct dm_target *ti)
 551 {
 552         int r;
 553         struct pgpath *p;
 554
 555         /* we need at least a path arg */
 556         if (as->argc < 1) {
 557                 ti->error = "no device given";
 558                 return NULL;
 559         }
 560
 561         p = alloc_pgpath();
 562         if (!p)
 563                 return NULL;
 564
 565         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 566                           dm_table_get_mode(ti->table), &p->path.dev);
 567         if (r) {
 568                 ti->error = "error getting device";
 569                 goto bad;
 570         }
 571
 572         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 573         if (r) {
 574                 dm_put_device(ti, p->path.dev);
 575                 goto bad;
 576         }
 577
 578         return p;
 579
 580  bad:
 581         free_pgpath(p);
 582         return NULL;
 583 }
 584
 585 static struct priority_group *parse_priority_group(struct arg_set *as,
 586                                                    struct multipath *m)
 587 {
 588         static struct param _params[] = {
 589                 {1, 1024, "invalid number of paths"},
 590                 {0, 1024, "invalid number of selector args"}
 591         };
 592
 593         int r;
 594         unsigned i, nr_selector_args, nr_params;
 595         struct priority_group *pg;
 596         struct dm_target *ti = m->ti;
 597
 598         if (as->argc < 2) {
 599                 as->argc = 0;
 600                 ti->error = "not enough priority group aruments";
 601                 return NULL;
 602         }
 603
 604         pg = alloc_priority_group();
 605         if (!pg) {
 606                 ti->error = "couldn't allocate priority group";
 607                 return NULL;
 608         }
 609         pg->m = m;
 610
 611         r = parse_path_selector(as, pg, ti);
 612         if (r)
 613                 goto bad;
 614
 615         /*
 616          * read the paths
 617          */
 618         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 619         if (r)
 620                 goto bad;
 621
 622         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 623         if (r)
 624                 goto bad;
 625
 626         nr_params = 1 + nr_selector_args;
 627         for (i = 0; i < pg->nr_pgpaths; i++) {
 628                 struct pgpath *pgpath;
 629                 struct arg_set path_args;
 630
 631                 if (as->argc < nr_params)
 632                         goto bad;
 633
 634                 path_args.argc = nr_params;
 635                 path_args.argv = as->argv;
 636
 637                 pgpath = parse_path(&path_args, &pg->ps, ti);
 638                 if (!pgpath)
 639                         goto bad;
 640
 641                 pgpath->pg = pg;
 642                 list_add_tail(&pgpath->list, &pg->pgpaths);
 643                 consume(as, nr_params);
 644         }
 645
 646         return pg;
 647
 648  bad:
 649         free_priority_group(pg, ti);
 650         return NULL;
 651 }
 652
 653 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 654 {
 655         int r;
 656         struct hw_handler_type *hwht;
 657         unsigned hw_argc;
 658         struct dm_target *ti = m->ti;
 659
 660         static struct param _params[] = {
 661                 {0, 1024, "invalid number of hardware handler args"},
 662         };
 663
 664         r = read_param(_params, shift(as), &hw_argc, &ti->error);
 665         if (r)
 666                 return -EINVAL;
 667
 668         if (!hw_argc)
 669                 return 0;
 670
 671         hwht = dm_get_hw_handler(shift(as));
 672         if (!hwht) {
 673                 ti->error = "unknown hardware handler type";
 674                 return -EINVAL;
 675         }
 676
 677         m->hw_handler.md = dm_table_get_md(ti->table);
 678         dm_put(m->hw_handler.md);
 679
 680         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
 681         if (r) {
 682                 dm_put_hw_handler(hwht);
 683                 ti->error = "hardware handler constructor failed";
 684                 return r;
 685         }
 686
 687         m->hw_handler.type = hwht;
 688         consume(as, hw_argc - 1);
 689
 690         return 0;
 691 }
 692
 693 static int parse_features(struct arg_set *as, struct multipath *m)
 694 {
 695         int r;
 696         unsigned argc;
 697         struct dm_target *ti = m->ti;
 698         const char *param_name;
 699
 700         static struct param _params[] = {
 701                 {0, 3, "invalid number of feature args"},
 702                 {1, 50, "pg_init_retries must be between 1 and 50"},
 703         };
 704
 705         r = read_param(_params, shift(as), &argc, &ti->error);
 706         if (r)
 707                 return -EINVAL;
 708
 709         if (!argc)
 710                 return 0;
 711
 712         do {
 713                 param_name = shift(as);
 714                 argc--;
 715
 716                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 717                         r = queue_if_no_path(m, 1, 0);
 718                         continue;
 719                 }
 720
 721                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 722                     (argc >= 1)) {
 723                         r = read_param(_params + 1, shift(as),
 724                                        &m->pg_init_retries, &ti->error);
 725                         argc--;
 726                         continue;
 727                 }
 728
 729                 ti->error = "Unrecognised multipath feature request";
 730                 r = -EINVAL;
 731         } while (argc && !r);
 732
 733         return r;
 734 }
 735
 736 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 737                          char **argv)
 738 {
 739         /* target parameters */
 740         static struct param _params[] = {
 741                 {1, 1024, "invalid number of priority groups"},
 742                 {1, 1024, "invalid initial priority group number"},
 743         };
 744
 745         int r;
 746         struct multipath *m;
 747         struct arg_set as;
 748         unsigned pg_count = 0;
 749         unsigned next_pg_num;
 750
 751         as.argc = argc;
 752         as.argv = argv;
 753
 754         m = alloc_multipath(ti);
 755         if (!m) {
 756                 ti->error = "can't allocate multipath";
 757                 return -EINVAL;
 758         }
 759
 760         r = parse_features(&as, m);
 761         if (r)
 762                 goto bad;
 763
 764         r = parse_hw_handler(&as, m);
 765         if (r)
 766                 goto bad;
 767
 768         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 769         if (r)
 770                 goto bad;
 771
 772         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 773         if (r)
 774                 goto bad;
 775
 776         /* parse the priority groups */
 777         while (as.argc) {
 778                 struct priority_group *pg;
 779
 780                 pg = parse_priority_group(&as, m);
 781                 if (!pg) {
 782                         r = -EINVAL;
 783                         goto bad;
 784                 }
 785
 786                 m->nr_valid_paths += pg->nr_pgpaths;
 787                 list_add_tail(&pg->list, &m->priority_groups);
 788                 pg_count++;
 789                 pg->pg_num = pg_count;
 790                 if (!--next_pg_num)
 791                         m->next_pg = pg;
 792         }
 793
 794         if (pg_count != m->nr_priority_groups) {
 795                 ti->error = "priority group count mismatch";
 796                 r = -EINVAL;
 797                 goto bad;
 798         }
 799
 800         return 0;
 801
 802  bad:
 803         free_multipath(m);
 804         return r;
 805 }
 806
 807 static void multipath_dtr(struct dm_target *ti)
 808 {
 809         struct multipath *m = (struct multipath *) ti->private;
 810
 811         flush_workqueue(kmultipathd);
 812         free_multipath(m);
 813 }
 814
 815 /*
 816  * Map bios, recording original fields for later in case we have to resubmit
 817  */
 818 static int multipath_map(struct dm_target *ti, struct bio *bio,
 819                          union map_info *map_context)
 820 {
 821         int r;
 822         struct dm_mpath_io *mpio;
 823         struct multipath *m = (struct multipath *) ti->private;
 824
 825         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 826         dm_bio_record(&mpio->details, bio);
 827
 828         map_context->ptr = mpio;
 829         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 830         r = map_io(m, bio, mpio, 0);
 831         if (r < 0 || r == DM_MAPIO_REQUEUE)
 832                 mempool_free(mpio, m->mpio_pool);
 833
 834         return r;
 835 }
 836
 837 /*
 838  * Take a path out of use.
 839  */
 840 static int fail_path(struct pgpath *pgpath)
 841 {
 842         unsigned long flags;
 843         struct multipath *m = pgpath->pg->m;
 844
 845         spin_lock_irqsave(&m->lock, flags);
 846
 847         if (!pgpath->path.is_active)
 848                 goto out;
 849
 850         DMWARN("Failing path %s.", pgpath->path.dev->name);
 851
 852         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 853         pgpath->path.is_active = 0;
 854         pgpath->fail_count++;
 855
 856         m->nr_valid_paths--;
 857
 858         if (pgpath == m->current_pgpath)
 859                 m->current_pgpath = NULL;
 860
 861         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 862                       pgpath->path.dev->name, m->nr_valid_paths);
 863
 864         queue_work(kmultipathd, &m->trigger_event);
 865
 866 out:
 867         spin_unlock_irqrestore(&m->lock, flags);
 868
 869         return 0;
 870 }
 871
 872 /*
 873  * Reinstate a previously-failed path
 874  */
 875 static int reinstate_path(struct pgpath *pgpath)
 876 {
 877         int r = 0;
 878         unsigned long flags;
 879         struct multipath *m = pgpath->pg->m;
 880
 881         spin_lock_irqsave(&m->lock, flags);
 882
 883         if (pgpath->path.is_active)
 884                 goto out;
 885
 886         if (!pgpath->pg->ps.type) {
 887                 DMWARN("Reinstate path not supported by path selector %s",
 888                        pgpath->pg->ps.type->name);
 889                 r = -EINVAL;
 890                 goto out;
 891         }
 892
 893         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 894         if (r)
 895                 goto out;
 896
 897         pgpath->path.is_active = 1;
 898
 899         m->current_pgpath = NULL;
 900         if (!m->nr_valid_paths++ && m->queue_size)
 901                 queue_work(kmultipathd, &m->process_queued_ios);
 902
 903         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 904                       pgpath->path.dev->name, m->nr_valid_paths);
 905
 906         queue_work(kmultipathd, &m->trigger_event);
 907
 908 out:
 909         spin_unlock_irqrestore(&m->lock, flags);
 910
 911         return r;
 912 }
 913
 914 /*
 915  * Fail or reinstate all paths that match the provided struct dm_dev.
 916  */
 917 static int action_dev(struct multipath *m, struct dm_dev *dev,
 918                       action_fn action)
 919 {
 920         int r = 0;
 921         struct pgpath *pgpath;
 922         struct priority_group *pg;
 923
 924         list_for_each_entry(pg, &m->priority_groups, list) {
 925                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 926                         if (pgpath->path.dev == dev)
 927                                 r = action(pgpath);
 928                 }
 929         }
 930
 931         return r;
 932 }
 933
 934 /*
 935  * Temporarily try to avoid having to use the specified PG
 936  */
 937 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 938                       int bypassed)
 939 {
 940         unsigned long flags;
 941
 942         spin_lock_irqsave(&m->lock, flags);
 943
 944         pg->bypassed = bypassed;
 945         m->current_pgpath = NULL;
 946         m->current_pg = NULL;
 947
 948         spin_unlock_irqrestore(&m->lock, flags);
 949
 950         queue_work(kmultipathd, &m->trigger_event);
 951 }
 952
 953 /*
 954  * Switch to using the specified PG from the next I/O that gets mapped
 955  */
 956 static int switch_pg_num(struct multipath *m, const char *pgstr)
 957 {
 958         struct priority_group *pg;
 959         unsigned pgnum;
 960         unsigned long flags;
 961
 962         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 963             (pgnum > m->nr_priority_groups)) {
 964                 DMWARN("invalid PG number supplied to switch_pg_num");
 965                 return -EINVAL;
 966         }
 967
 968         spin_lock_irqsave(&m->lock, flags);
 969         list_for_each_entry(pg, &m->priority_groups, list) {
 970                 pg->bypassed = 0;
 971                 if (--pgnum)
 972                         continue;
 973
 974                 m->current_pgpath = NULL;
 975                 m->current_pg = NULL;
 976                 m->next_pg = pg;
 977         }
 978         spin_unlock_irqrestore(&m->lock, flags);
 979
 980         queue_work(kmultipathd, &m->trigger_event);
 981         return 0;
 982 }
 983
 984 /*
 985  * Set/clear bypassed status of a PG.
 986  * PGs are numbered upwards from 1 in the order they were declared.
 987  */
 988 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 989 {
 990         struct priority_group *pg;
 991         unsigned pgnum;
 992
 993         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 994             (pgnum > m->nr_priority_groups)) {
 995                 DMWARN("invalid PG number supplied to bypass_pg");
 996                 return -EINVAL;
 997         }
 998
 999         list_for_each_entry(pg, &m->priority_groups, list) {
1000                 if (!--pgnum)
1001                         break;
1002         }
1003
1004         bypass_pg(m, pg, bypassed);
1005         return 0;
1006 }
1007
1008 /*
1009  * Should we retry pg_init immediately?
1010  */
1011 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1012 {
1013         unsigned long flags;
1014         int limit_reached = 0;
1015
1016         spin_lock_irqsave(&m->lock, flags);
1017
1018         if (m->pg_init_count <= m->pg_init_retries)
1019                 m->pg_init_required = 1;
1020         else
1021                 limit_reached = 1;
1022
1023         spin_unlock_irqrestore(&m->lock, flags);
1024
1025         return limit_reached;
1026 }
1027
1028 /*
1029  * pg_init must call this when it has completed its initialisation
1030  */
1031 void dm_pg_init_complete(struct dm_path *path, unsigned err_flags)
1032 {
1033         struct pgpath *pgpath = path_to_pgpath(path);
1034         struct priority_group *pg = pgpath->pg;
1035         struct multipath *m = pg->m;
1036         unsigned long flags;
1037
1038         /*
1039          * If requested, retry pg_init until maximum number of retries exceeded.
1040          * If retry not requested and PG already bypassed, always fail the path.
1041          */
1042         if (err_flags & MP_RETRY) {
1043                 if (pg_init_limit_reached(m, pgpath))
1044                         err_flags |= MP_FAIL_PATH;
1045         } else if (err_flags && pg->bypassed)
1046                 err_flags |= MP_FAIL_PATH;
1047
1048         if (err_flags & MP_FAIL_PATH)
1049                 fail_path(pgpath);
1050
1051         if (err_flags & MP_BYPASS_PG)
1052                 bypass_pg(m, pg, 1);
1053
1054         spin_lock_irqsave(&m->lock, flags);
1055         if (err_flags & ~MP_RETRY) {
1056                 m->current_pgpath = NULL;
1057                 m->current_pg = NULL;
1058         } else if (!m->pg_init_required)
1059                 m->queue_io = 0;
1060
1061         m->pg_init_in_progress = 0;
1062         queue_work(kmultipathd, &m->process_queued_ios);
1063         spin_unlock_irqrestore(&m->lock, flags);
1064 }
1065
1066 /*
1067  * end_io handling
1068  */
1069 static int do_end_io(struct multipath *m, struct bio *bio,
1070                      int error, struct dm_mpath_io *mpio)
1071 {
1072         struct hw_handler *hwh = &m->hw_handler;
1073         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
1074         unsigned long flags;
1075
1076         if (!error)
1077                 return 0;       /* I/O complete */
1078
1079         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1080                 return error;
1081
1082         if (error == -EOPNOTSUPP)
1083                 return error;
1084
1085         spin_lock_irqsave(&m->lock, flags);
1086         if (!m->nr_valid_paths) {
1087                 if (__must_push_back(m)) {
1088                         spin_unlock_irqrestore(&m->lock, flags);
1089                         return DM_ENDIO_REQUEUE;
1090                 } else if (!m->queue_if_no_path) {
1091                         spin_unlock_irqrestore(&m->lock, flags);
1092                         return -EIO;
1093                 } else {
1094                         spin_unlock_irqrestore(&m->lock, flags);
1095                         goto requeue;
1096                 }
1097         }
1098         spin_unlock_irqrestore(&m->lock, flags);
1099
1100         if (hwh->type && hwh->type->error)
1101                 err_flags = hwh->type->error(hwh, bio);
1102
1103         if (mpio->pgpath) {
1104                 if (err_flags & MP_FAIL_PATH)
1105                         fail_path(mpio->pgpath);
1106
1107                 if (err_flags & MP_BYPASS_PG)
1108                         bypass_pg(m, mpio->pgpath->pg, 1);
1109         }
1110
1111         if (err_flags & MP_ERROR_IO)
1112                 return -EIO;
1113
1114       requeue:
1115         dm_bio_restore(&mpio->details, bio);
1116
1117         /* queue for the daemon to resubmit or fail */
1118         spin_lock_irqsave(&m->lock, flags);
1119         bio_list_add(&m->queued_ios, bio);
1120         m->queue_size++;
1121         if (!m->queue_io)
1122                 queue_work(kmultipathd, &m->process_queued_ios);
1123         spin_unlock_irqrestore(&m->lock, flags);
1124
1125         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1126 }
1127
1128 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1129                             int error, union map_info *map_context)
1130 {
1131         struct multipath *m = ti->private;
1132         struct dm_mpath_io *mpio = map_context->ptr;
1133         struct pgpath *pgpath = mpio->pgpath;
1134         struct path_selector *ps;
1135         int r;
1136
1137         r  = do_end_io(m, bio, error, mpio);
1138         if (pgpath) {
1139                 ps = &pgpath->pg->ps;
1140                 if (ps->type->end_io)
1141                         ps->type->end_io(ps, &pgpath->path);
1142         }
1143         if (r != DM_ENDIO_INCOMPLETE)
1144                 mempool_free(mpio, m->mpio_pool);
1145
1146         return r;
1147 }
1148
1149 /*
1150  * Suspend can't complete until all the I/O is processed so if
1151  * the last path fails we must error any remaining I/O.
1152  * Note that if the freeze_bdev fails while suspending, the
1153  * queue_if_no_path state is lost - userspace should reset it.
1154  */
1155 static void multipath_presuspend(struct dm_target *ti)
1156 {
1157         struct multipath *m = (struct multipath *) ti->private;
1158
1159         queue_if_no_path(m, 0, 1);
1160 }
1161
1162 /*
1163  * Restore the queue_if_no_path setting.
1164  */
1165 static void multipath_resume(struct dm_target *ti)
1166 {
1167         struct multipath *m = (struct multipath *) ti->private;
1168         unsigned long flags;
1169
1170         spin_lock_irqsave(&m->lock, flags);
1171         m->queue_if_no_path = m->saved_queue_if_no_path;
1172         spin_unlock_irqrestore(&m->lock, flags);
1173 }
1174
1175 /*
1176  * Info output has the following format:
1177  * num_multipath_feature_args [multipath_feature_args]*
1178  * num_handler_status_args [handler_status_args]*
1179  * num_groups init_group_number
1180  *            [A|D|E num_ps_status_args [ps_status_args]*
1181  *             num_paths num_selector_args
1182  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1183  *
1184  * Table output has the following format (identical to the constructor string):
1185  * num_feature_args [features_args]*
1186  * num_handler_args hw_handler [hw_handler_args]*
1187  * num_groups init_group_number
1188  *     [priority selector-name num_ps_args [ps_args]*
1189  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1190  */
1191 static int multipath_status(struct dm_target *ti, status_type_t type,
1192                             char *result, unsigned int maxlen)
1193 {
1194         int sz = 0;
1195         unsigned long flags;
1196         struct multipath *m = (struct multipath *) ti->private;
1197         struct hw_handler *hwh = &m->hw_handler;
1198         struct priority_group *pg;
1199         struct pgpath *p;
1200         unsigned pg_num;
1201         char state;
1202
1203         spin_lock_irqsave(&m->lock, flags);
1204
1205         /* Features */
1206         if (type == STATUSTYPE_INFO)
1207                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1208         else {
1209                 DMEMIT("%u ", m->queue_if_no_path +
1210                               (m->pg_init_retries > 0) * 2);
1211                 if (m->queue_if_no_path)
1212                         DMEMIT("queue_if_no_path ");
1213                 if (m->pg_init_retries)
1214                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1215         }
1216
1217         if (hwh->type && hwh->type->status)
1218                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1219         else if (!hwh->type || type == STATUSTYPE_INFO)
1220                 DMEMIT("0 ");
1221         else
1222                 DMEMIT("1 %s ", hwh->type->name);
1223
1224         DMEMIT("%u ", m->nr_priority_groups);
1225
1226         if (m->next_pg)
1227                 pg_num = m->next_pg->pg_num;
1228         else if (m->current_pg)
1229                 pg_num = m->current_pg->pg_num;
1230         else
1231                         pg_num = 1;
1232
1233         DMEMIT("%u ", pg_num);
1234
1235         switch (type) {
1236         case STATUSTYPE_INFO:
1237                 list_for_each_entry(pg, &m->priority_groups, list) {
1238                         if (pg->bypassed)
1239                                 state = 'D';    /* Disabled */
1240                         else if (pg == m->current_pg)
1241                                 state = 'A';    /* Currently Active */
1242                         else
1243                                 state = 'E';    /* Enabled */
1244
1245                         DMEMIT("%c ", state);
1246
1247                         if (pg->ps.type->status)
1248                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1249                                                           result + sz,
1250                                                           maxlen - sz);
1251                         else
1252                                 DMEMIT("0 ");
1253
1254                         DMEMIT("%u %u ", pg->nr_pgpaths,
1255                                pg->ps.type->info_args);
1256
1257                         list_for_each_entry(p, &pg->pgpaths, list) {
1258                                 DMEMIT("%s %s %u ", p->path.dev->name,
1259                                        p->path.is_active ? "A" : "F",
1260                                        p->fail_count);
1261                                 if (pg->ps.type->status)
1262                                         sz += pg->ps.type->status(&pg->ps,
1263                                               &p->path, type, result + sz,
1264                                               maxlen - sz);
1265                         }
1266                 }
1267                 break;
1268
1269         case STATUSTYPE_TABLE:
1270                 list_for_each_entry(pg, &m->priority_groups, list) {
1271                         DMEMIT("%s ", pg->ps.type->name);
1272
1273                         if (pg->ps.type->status)
1274                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1275                                                           result + sz,
1276                                                           maxlen - sz);
1277                         else
1278                                 DMEMIT("0 ");
1279
1280                         DMEMIT("%u %u ", pg->nr_pgpaths,
1281                                pg->ps.type->table_args);
1282
1283                         list_for_each_entry(p, &pg->pgpaths, list) {
1284                                 DMEMIT("%s ", p->path.dev->name);
1285                                 if (pg->ps.type->status)
1286                                         sz += pg->ps.type->status(&pg->ps,
1287                                               &p->path, type, result + sz,
1288                                               maxlen - sz);
1289                         }
1290                 }
1291                 break;
1292         }
1293
1294         spin_unlock_irqrestore(&m->lock, flags);
1295
1296         return 0;
1297 }
1298
1299 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1300 {
1301         int r;
1302         struct dm_dev *dev;
1303         struct multipath *m = (struct multipath *) ti->private;
1304         action_fn action;
1305
1306         if (argc == 1) {
1307                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1308                         return queue_if_no_path(m, 1, 0);
1309                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1310                         return queue_if_no_path(m, 0, 0);
1311         }
1312
1313         if (argc != 2)
1314                 goto error;
1315
1316         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1317                 return bypass_pg_num(m, argv[1], 1);
1318         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1319                 return bypass_pg_num(m, argv[1], 0);
1320         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1321                 return switch_pg_num(m, argv[1]);
1322         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1323                 action = reinstate_path;
1324         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1325                 action = fail_path;
1326         else
1327                 goto error;
1328
1329         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1330                           dm_table_get_mode(ti->table), &dev);
1331         if (r) {
1332                 DMWARN("message: error getting device %s",
1333                        argv[1]);
1334                 return -EINVAL;
1335         }
1336
1337         r = action_dev(m, dev, action);
1338
1339         dm_put_device(ti, dev);
1340
1341         return r;
1342
1343 error:
1344         DMWARN("Unrecognised multipath message received.");
1345         return -EINVAL;
1346 }
1347
1348 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1349                            struct file *filp, unsigned int cmd,
1350                            unsigned long arg)
1351 {
1352         struct multipath *m = (struct multipath *) ti->private;
1353         struct block_device *bdev = NULL;
1354         unsigned long flags;
1355         struct file fake_file = {};
1356         struct dentry fake_dentry = {};
1357         int r = 0;
1358
1359         fake_file.f_path.dentry = &fake_dentry;
1360
1361         spin_lock_irqsave(&m->lock, flags);
1362
1363         if (!m->current_pgpath)
1364                 __choose_pgpath(m);
1365
1366         if (m->current_pgpath) {
1367                 bdev = m->current_pgpath->path.dev->bdev;
1368                 fake_dentry.d_inode = bdev->bd_inode;
1369                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1370         }
1371
1372         if (m->queue_io)
1373                 r = -EAGAIN;
1374         else if (!bdev)
1375                 r = -EIO;
1376
1377         spin_unlock_irqrestore(&m->lock, flags);
1378
1379         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1380                                          bdev->bd_disk, cmd, arg);
1381 }
1382
1383 /*-----------------------------------------------------------------
1384  * Module setup
1385  *---------------------------------------------------------------*/
1386 static struct target_type multipath_target = {
1387         .name = "multipath",
1388         .version = {1, 0, 5},
1389         .module = THIS_MODULE,
1390         .ctr = multipath_ctr,
1391         .dtr = multipath_dtr,
1392         .map = multipath_map,
1393         .end_io = multipath_end_io,
1394         .presuspend = multipath_presuspend,
1395         .resume = multipath_resume,
1396         .status = multipath_status,
1397         .message = multipath_message,
1398         .ioctl  = multipath_ioctl,
1399 };
1400
1401 static int __init dm_multipath_init(void)
1402 {
1403         int r;
1404
1405         /* allocate a slab for the dm_ios */
1406         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1407         if (!_mpio_cache)
1408                 return -ENOMEM;
1409
1410         r = dm_register_target(&multipath_target);
1411         if (r < 0) {
1412                 DMERR("register failed %d", r);
1413                 kmem_cache_destroy(_mpio_cache);
1414                 return -EINVAL;
1415         }
1416
1417         kmultipathd = create_workqueue("kmpathd");
1418         if (!kmultipathd) {
1419                 DMERR("failed to create workqueue kmpathd");
1420                 dm_unregister_target(&multipath_target);
1421                 kmem_cache_destroy(_mpio_cache);
1422                 return -ENOMEM;
1423         }
1424
1425         DMINFO("version %u.%u.%u loaded",
1426                multipath_target.version[0], multipath_target.version[1],
1427                multipath_target.version[2]);
1428
1429         return r;
1430 }
1431
1432 static void __exit dm_multipath_exit(void)
1433 {
1434         int r;
1435
1436         destroy_workqueue(kmultipathd);
1437
1438         r = dm_unregister_target(&multipath_target);
1439         if (r < 0)
1440                 DMERR("target unregister failed %d", r);
1441         kmem_cache_destroy(_mpio_cache);
1442 }
1443
1444 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1445
1446 module_init(dm_multipath_init);
1447 module_exit(dm_multipath_exit);
1448
1449 MODULE_DESCRIPTION(DM_NAME " multipath target");
1450 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1451 MODULE_LICENSE("GPL");