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