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 <scsi/scsi_dh.h>
  24 #include <asm/atomic.h>
  25
  26 #define DM_MSG_PREFIX "multipath"
  27 #define MESG_STR(x) x, sizeof(x)
  28
  29 /* Path properties */
  30 struct pgpath {
  31         struct list_head list;
  32
  33         struct priority_group *pg;      /* Owning PG */
  34         unsigned fail_count;            /* Cumulative failure count */
  35
  36         struct dm_path path;
  37 };
  38
  39 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  40
  41 /*
  42  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  43  * Each has a path selector which controls which path gets used.
  44  */
  45 struct priority_group {
  46         struct list_head list;
  47
  48         struct multipath *m;            /* Owning multipath instance */
  49         struct path_selector ps;
  50
  51         unsigned pg_num;                /* Reference number */
  52         unsigned bypassed;              /* Temporarily bypass this PG? */
  53
  54         unsigned nr_pgpaths;            /* Number of paths in PG */
  55         struct list_head pgpaths;
  56 };
  57
  58 /* Multipath context */
  59 struct multipath {
  60         struct list_head list;
  61         struct dm_target *ti;
  62
  63         spinlock_t lock;
  64
  65         const char *hw_handler_name;
  66         unsigned nr_priority_groups;
  67         struct list_head priority_groups;
  68         unsigned pg_init_required;      /* pg_init needs calling? */
  69         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  70
  71         unsigned nr_valid_paths;        /* Total number of usable paths */
  72         struct pgpath *current_pgpath;
  73         struct priority_group *current_pg;
  74         struct priority_group *next_pg; /* Switch to this PG if set */
  75         unsigned repeat_count;          /* I/Os left before calling PS again */
  76
  77         unsigned queue_io;              /* Must we queue all I/O? */
  78         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  79         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  80         unsigned pg_init_retries;       /* Number of times to retry pg_init */
  81         unsigned pg_init_count;         /* Number of times pg_init called */
  82
  83         struct work_struct process_queued_ios;
  84         struct bio_list queued_ios;
  85         unsigned queue_size;
  86
  87         struct work_struct trigger_event;
  88
  89         /*
  90          * We must use a mempool of dm_mpath_io structs so that we
  91          * can resubmit bios on error.
  92          */
  93         mempool_t *mpio_pool;
  94 };
  95
  96 /*
  97  * Context information attached to each bio we process.
  98  */
  99 struct dm_mpath_io {
 100         struct pgpath *pgpath;
 101         struct dm_bio_details details;
 102 };
 103
 104 typedef int (*action_fn) (struct pgpath *pgpath);
 105
 106 #define MIN_IOS 256     /* Mempool size */
 107
 108 static struct kmem_cache *_mpio_cache;
 109
 110 static struct workqueue_struct *kmultipathd;
 111 static void process_queued_ios(struct work_struct *work);
 112 static void trigger_event(struct work_struct *work);
 113 static void pg_init_done(struct dm_path *, int);
 114
 115
 116 /*-----------------------------------------------
 117  * Allocation routines
 118  *-----------------------------------------------*/
 119
 120 static struct pgpath *alloc_pgpath(void)
 121 {
 122         struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 123
 124         if (pgpath)
 125                 pgpath->path.is_active = 1;
 126
 127         return pgpath;
 128 }
 129
 130 static void free_pgpath(struct pgpath *pgpath)
 131 {
 132         kfree(pgpath);
 133 }
 134
 135 static struct priority_group *alloc_priority_group(void)
 136 {
 137         struct priority_group *pg;
 138
 139         pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 140
 141         if (pg)
 142                 INIT_LIST_HEAD(&pg->pgpaths);
 143
 144         return pg;
 145 }
 146
 147 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 148 {
 149         struct pgpath *pgpath, *tmp;
 150
 151         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 152                 list_del(&pgpath->list);
 153                 dm_put_device(ti, pgpath->path.dev);
 154                 free_pgpath(pgpath);
 155         }
 156 }
 157
 158 static void free_priority_group(struct priority_group *pg,
 159                                 struct dm_target *ti)
 160 {
 161         struct path_selector *ps = &pg->ps;
 162
 163         if (ps->type) {
 164                 ps->type->destroy(ps);
 165                 dm_put_path_selector(ps->type);
 166         }
 167
 168         free_pgpaths(&pg->pgpaths, ti);
 169         kfree(pg);
 170 }
 171
 172 static struct multipath *alloc_multipath(struct dm_target *ti)
 173 {
 174         struct multipath *m;
 175
 176         m = kzalloc(sizeof(*m), GFP_KERNEL);
 177         if (m) {
 178                 INIT_LIST_HEAD(&m->priority_groups);
 179                 spin_lock_init(&m->lock);
 180                 m->queue_io = 1;
 181                 INIT_WORK(&m->process_queued_ios, process_queued_ios);
 182                 INIT_WORK(&m->trigger_event, trigger_event);
 183                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 184                 if (!m->mpio_pool) {
 185                         kfree(m);
 186                         return NULL;
 187                 }
 188                 m->ti = ti;
 189                 ti->private = m;
 190         }
 191
 192         return m;
 193 }
 194
 195 static void free_multipath(struct multipath *m)
 196 {
 197         struct priority_group *pg, *tmp;
 198
 199         list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 200                 list_del(&pg->list);
 201                 free_priority_group(pg, m->ti);
 202         }
 203
 204         kfree(m->hw_handler_name);
 205         mempool_destroy(m->mpio_pool);
 206         kfree(m);
 207 }
 208
 209
 210 /*-----------------------------------------------
 211  * Path selection
 212  *-----------------------------------------------*/
 213
 214 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 215 {
 216         m->current_pg = pgpath->pg;
 217
 218         /* Must we initialise the PG first, and queue I/O till it's ready? */
 219         if (m->hw_handler_name) {
 220                 m->pg_init_required = 1;
 221                 m->queue_io = 1;
 222         } else {
 223                 m->pg_init_required = 0;
 224                 m->queue_io = 0;
 225         }
 226
 227         m->pg_init_count = 0;
 228 }
 229
 230 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
 231 {
 232         struct dm_path *path;
 233
 234         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
 235         if (!path)
 236                 return -ENXIO;
 237
 238         m->current_pgpath = path_to_pgpath(path);
 239
 240         if (m->current_pg != pg)
 241                 __switch_pg(m, m->current_pgpath);
 242
 243         return 0;
 244 }
 245
 246 static void __choose_pgpath(struct multipath *m)
 247 {
 248         struct priority_group *pg;
 249         unsigned bypassed = 1;
 250
 251         if (!m->nr_valid_paths)
 252                 goto failed;
 253
 254         /* Were we instructed to switch PG? */
 255         if (m->next_pg) {
 256                 pg = m->next_pg;
 257                 m->next_pg = NULL;
 258                 if (!__choose_path_in_pg(m, pg))
 259                         return;
 260         }
 261
 262         /* Don't change PG until it has no remaining paths */
 263         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
 264                 return;
 265
 266         /*
 267          * Loop through priority groups until we find a valid path.
 268          * First time we skip PGs marked 'bypassed'.
 269          * Second time we only try the ones we skipped.
 270          */
 271         do {
 272                 list_for_each_entry(pg, &m->priority_groups, list) {
 273                         if (pg->bypassed == bypassed)
 274                                 continue;
 275                         if (!__choose_path_in_pg(m, pg))
 276                                 return;
 277                 }
 278         } while (bypassed--);
 279
 280 failed:
 281         m->current_pgpath = NULL;
 282         m->current_pg = NULL;
 283 }
 284
 285 /*
 286  * Check whether bios must be queued in the device-mapper core rather
 287  * than here in the target.
 288  *
 289  * m->lock must be held on entry.
 290  *
 291  * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 292  * same value then we are not between multipath_presuspend()
 293  * and multipath_resume() calls and we have no need to check
 294  * for the DMF_NOFLUSH_SUSPENDING flag.
 295  */
 296 static int __must_push_back(struct multipath *m)
 297 {
 298         return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 299                 dm_noflush_suspending(m->ti));
 300 }
 301
 302 static int map_io(struct multipath *m, struct bio *bio,
 303                   struct dm_mpath_io *mpio, unsigned was_queued)
 304 {
 305         int r = DM_MAPIO_REMAPPED;
 306         unsigned long flags;
 307         struct pgpath *pgpath;
 308
 309         spin_lock_irqsave(&m->lock, flags);
 310
 311         /* Do we need to select a new pgpath? */
 312         if (!m->current_pgpath ||
 313             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 314                 __choose_pgpath(m);
 315
 316         pgpath = m->current_pgpath;
 317
 318         if (was_queued)
 319                 m->queue_size--;
 320
 321         if ((pgpath && m->queue_io) ||
 322             (!pgpath && m->queue_if_no_path)) {
 323                 /* Queue for the daemon to resubmit */
 324                 bio_list_add(&m->queued_ios, bio);
 325                 m->queue_size++;
 326                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 327                     !m->queue_io)
 328                         queue_work(kmultipathd, &m->process_queued_ios);
 329                 pgpath = NULL;
 330                 r = DM_MAPIO_SUBMITTED;
 331         } else if (pgpath)
 332                 bio->bi_bdev = pgpath->path.dev->bdev;
 333         else if (__must_push_back(m))
 334                 r = DM_MAPIO_REQUEUE;
 335         else
 336                 r = -EIO;       /* Failed */
 337
 338         mpio->pgpath = pgpath;
 339
 340         spin_unlock_irqrestore(&m->lock, flags);
 341
 342         return r;
 343 }
 344
 345 /*
 346  * If we run out of usable paths, should we queue I/O or error it?
 347  */
 348 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 349                             unsigned save_old_value)
 350 {
 351         unsigned long flags;
 352
 353         spin_lock_irqsave(&m->lock, flags);
 354
 355         if (save_old_value)
 356                 m->saved_queue_if_no_path = m->queue_if_no_path;
 357         else
 358                 m->saved_queue_if_no_path = queue_if_no_path;
 359         m->queue_if_no_path = queue_if_no_path;
 360         if (!m->queue_if_no_path && m->queue_size)
 361                 queue_work(kmultipathd, &m->process_queued_ios);
 362
 363         spin_unlock_irqrestore(&m->lock, flags);
 364
 365         return 0;
 366 }
 367
 368 /*-----------------------------------------------------------------
 369  * The multipath daemon is responsible for resubmitting queued ios.
 370  *---------------------------------------------------------------*/
 371
 372 static void dispatch_queued_ios(struct multipath *m)
 373 {
 374         int r;
 375         unsigned long flags;
 376         struct bio *bio = NULL, *next;
 377         struct dm_mpath_io *mpio;
 378         union map_info *info;
 379
 380         spin_lock_irqsave(&m->lock, flags);
 381         bio = bio_list_get(&m->queued_ios);
 382         spin_unlock_irqrestore(&m->lock, flags);
 383
 384         while (bio) {
 385                 next = bio->bi_next;
 386                 bio->bi_next = NULL;
 387
 388                 info = dm_get_mapinfo(bio);
 389                 mpio = info->ptr;
 390
 391                 r = map_io(m, bio, mpio, 1);
 392                 if (r < 0)
 393                         bio_endio(bio, r);
 394                 else if (r == DM_MAPIO_REMAPPED)
 395                         generic_make_request(bio);
 396                 else if (r == DM_MAPIO_REQUEUE)
 397                         bio_endio(bio, -EIO);
 398
 399                 bio = next;
 400         }
 401 }
 402
 403 static void process_queued_ios(struct work_struct *work)
 404 {
 405         struct multipath *m =
 406                 container_of(work, struct multipath, process_queued_ios);
 407         struct pgpath *pgpath = NULL;
 408         unsigned init_required = 0, must_queue = 1;
 409         unsigned long flags;
 410
 411         spin_lock_irqsave(&m->lock, flags);
 412
 413         if (!m->queue_size)
 414                 goto out;
 415
 416         if (!m->current_pgpath)
 417                 __choose_pgpath(m);
 418
 419         pgpath = m->current_pgpath;
 420
 421         if ((pgpath && !m->queue_io) ||
 422             (!pgpath && !m->queue_if_no_path))
 423                 must_queue = 0;
 424
 425         if (m->pg_init_required && !m->pg_init_in_progress) {
 426                 m->pg_init_count++;
 427                 m->pg_init_required = 0;
 428                 m->pg_init_in_progress = 1;
 429                 init_required = 1;
 430         }
 431
 432 out:
 433         spin_unlock_irqrestore(&m->lock, flags);
 434
 435         if (init_required) {
 436                 struct dm_path *path = &pgpath->path;
 437                 int ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
 438                 pg_init_done(path, ret);
 439         }
 440
 441         if (!must_queue)
 442                 dispatch_queued_ios(m);
 443 }
 444
 445 /*
 446  * An event is triggered whenever a path is taken out of use.
 447  * Includes path failure and PG bypass.
 448  */
 449 static void trigger_event(struct work_struct *work)
 450 {
 451         struct multipath *m =
 452                 container_of(work, struct multipath, trigger_event);
 453
 454         dm_table_event(m->ti->table);
 455 }
 456
 457 /*-----------------------------------------------------------------
 458  * Constructor/argument parsing:
 459  * <#multipath feature args> [<arg>]*
 460  * <#hw_handler args> [hw_handler [<arg>]*]
 461  * <#priority groups>
 462  * <initial priority group>
 463  *     [<selector> <#selector args> [<arg>]*
 464  *      <#paths> <#per-path selector args>
 465  *         [<path> [<arg>]* ]+ ]+
 466  *---------------------------------------------------------------*/
 467 struct param {
 468         unsigned min;
 469         unsigned max;
 470         char *error;
 471 };
 472
 473 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 474 {
 475         if (!str ||
 476             (sscanf(str, "%u", v) != 1) ||
 477             (*v < param->min) ||
 478             (*v > param->max)) {
 479                 *error = param->error;
 480                 return -EINVAL;
 481         }
 482
 483         return 0;
 484 }
 485
 486 struct arg_set {
 487         unsigned argc;
 488         char **argv;
 489 };
 490
 491 static char *shift(struct arg_set *as)
 492 {
 493         char *r;
 494
 495         if (as->argc) {
 496                 as->argc--;
 497                 r = *as->argv;
 498                 as->argv++;
 499                 return r;
 500         }
 501
 502         return NULL;
 503 }
 504
 505 static void consume(struct arg_set *as, unsigned n)
 506 {
 507         BUG_ON (as->argc < n);
 508         as->argc -= n;
 509         as->argv += n;
 510 }
 511
 512 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 513                                struct dm_target *ti)
 514 {
 515         int r;
 516         struct path_selector_type *pst;
 517         unsigned ps_argc;
 518
 519         static struct param _params[] = {
 520                 {0, 1024, "invalid number of path selector args"},
 521         };
 522
 523         pst = dm_get_path_selector(shift(as));
 524         if (!pst) {
 525                 ti->error = "unknown path selector type";
 526                 return -EINVAL;
 527         }
 528
 529         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 530         if (r)
 531                 return -EINVAL;
 532
 533         r = pst->create(&pg->ps, ps_argc, as->argv);
 534         if (r) {
 535                 dm_put_path_selector(pst);
 536                 ti->error = "path selector constructor failed";
 537                 return r;
 538         }
 539
 540         pg->ps.type = pst;
 541         consume(as, ps_argc);
 542
 543         return 0;
 544 }
 545
 546 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 547                                struct dm_target *ti)
 548 {
 549         int r;
 550         struct pgpath *p;
 551
 552         /* we need at least a path arg */
 553         if (as->argc < 1) {
 554                 ti->error = "no device given";
 555                 return NULL;
 556         }
 557
 558         p = alloc_pgpath();
 559         if (!p)
 560                 return NULL;
 561
 562         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 563                           dm_table_get_mode(ti->table), &p->path.dev);
 564         if (r) {
 565                 ti->error = "error getting device";
 566                 goto bad;
 567         }
 568
 569         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 570         if (r) {
 571                 dm_put_device(ti, p->path.dev);
 572                 goto bad;
 573         }
 574
 575         return p;
 576
 577  bad:
 578         free_pgpath(p);
 579         return NULL;
 580 }
 581
 582 static struct priority_group *parse_priority_group(struct arg_set *as,
 583                                                    struct multipath *m)
 584 {
 585         static struct param _params[] = {
 586                 {1, 1024, "invalid number of paths"},
 587                 {0, 1024, "invalid number of selector args"}
 588         };
 589
 590         int r;
 591         unsigned i, nr_selector_args, nr_params;
 592         struct priority_group *pg;
 593         struct dm_target *ti = m->ti;
 594
 595         if (as->argc < 2) {
 596                 as->argc = 0;
 597                 ti->error = "not enough priority group aruments";
 598                 return NULL;
 599         }
 600
 601         pg = alloc_priority_group();
 602         if (!pg) {
 603                 ti->error = "couldn't allocate priority group";
 604                 return NULL;
 605         }
 606         pg->m = m;
 607
 608         r = parse_path_selector(as, pg, ti);
 609         if (r)
 610                 goto bad;
 611
 612         /*
 613          * read the paths
 614          */
 615         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 616         if (r)
 617                 goto bad;
 618
 619         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 620         if (r)
 621                 goto bad;
 622
 623         nr_params = 1 + nr_selector_args;
 624         for (i = 0; i < pg->nr_pgpaths; i++) {
 625                 struct pgpath *pgpath;
 626                 struct arg_set path_args;
 627
 628                 if (as->argc < nr_params)
 629                         goto bad;
 630
 631                 path_args.argc = nr_params;
 632                 path_args.argv = as->argv;
 633
 634                 pgpath = parse_path(&path_args, &pg->ps, ti);
 635                 if (!pgpath)
 636                         goto bad;
 637
 638                 pgpath->pg = pg;
 639                 list_add_tail(&pgpath->list, &pg->pgpaths);
 640                 consume(as, nr_params);
 641         }
 642
 643         return pg;
 644
 645  bad:
 646         free_priority_group(pg, ti);
 647         return NULL;
 648 }
 649
 650 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 651 {
 652         unsigned hw_argc;
 653         struct dm_target *ti = m->ti;
 654
 655         static struct param _params[] = {
 656                 {0, 1024, "invalid number of hardware handler args"},
 657         };
 658
 659         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 660                 return -EINVAL;
 661
 662         if (!hw_argc)
 663                 return 0;
 664
 665         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 666         request_module("scsi_dh_%s", m->hw_handler_name);
 667         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 668                 ti->error = "unknown hardware handler type";
 669                 return -EINVAL;
 670         }
 671         consume(as, hw_argc - 1);
 672
 673         return 0;
 674 }
 675
 676 static int parse_features(struct arg_set *as, struct multipath *m)
 677 {
 678         int r;
 679         unsigned argc;
 680         struct dm_target *ti = m->ti;
 681         const char *param_name;
 682
 683         static struct param _params[] = {
 684                 {0, 3, "invalid number of feature args"},
 685                 {1, 50, "pg_init_retries must be between 1 and 50"},
 686         };
 687
 688         r = read_param(_params, shift(as), &argc, &ti->error);
 689         if (r)
 690                 return -EINVAL;
 691
 692         if (!argc)
 693                 return 0;
 694
 695         do {
 696                 param_name = shift(as);
 697                 argc--;
 698
 699                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 700                         r = queue_if_no_path(m, 1, 0);
 701                         continue;
 702                 }
 703
 704                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 705                     (argc >= 1)) {
 706                         r = read_param(_params + 1, shift(as),
 707                                        &m->pg_init_retries, &ti->error);
 708                         argc--;
 709                         continue;
 710                 }
 711
 712                 ti->error = "Unrecognised multipath feature request";
 713                 r = -EINVAL;
 714         } while (argc && !r);
 715
 716         return r;
 717 }
 718
 719 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 720                          char **argv)
 721 {
 722         /* target parameters */
 723         static struct param _params[] = {
 724                 {1, 1024, "invalid number of priority groups"},
 725                 {1, 1024, "invalid initial priority group number"},
 726         };
 727
 728         int r;
 729         struct multipath *m;
 730         struct arg_set as;
 731         unsigned pg_count = 0;
 732         unsigned next_pg_num;
 733
 734         as.argc = argc;
 735         as.argv = argv;
 736
 737         m = alloc_multipath(ti);
 738         if (!m) {
 739                 ti->error = "can't allocate multipath";
 740                 return -EINVAL;
 741         }
 742
 743         r = parse_features(&as, m);
 744         if (r)
 745                 goto bad;
 746
 747         r = parse_hw_handler(&as, m);
 748         if (r)
 749                 goto bad;
 750
 751         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 752         if (r)
 753                 goto bad;
 754
 755         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 756         if (r)
 757                 goto bad;
 758
 759         /* parse the priority groups */
 760         while (as.argc) {
 761                 struct priority_group *pg;
 762
 763                 pg = parse_priority_group(&as, m);
 764                 if (!pg) {
 765                         r = -EINVAL;
 766                         goto bad;
 767                 }
 768
 769                 m->nr_valid_paths += pg->nr_pgpaths;
 770                 list_add_tail(&pg->list, &m->priority_groups);
 771                 pg_count++;
 772                 pg->pg_num = pg_count;
 773                 if (!--next_pg_num)
 774                         m->next_pg = pg;
 775         }
 776
 777         if (pg_count != m->nr_priority_groups) {
 778                 ti->error = "priority group count mismatch";
 779                 r = -EINVAL;
 780                 goto bad;
 781         }
 782
 783         return 0;
 784
 785  bad:
 786         free_multipath(m);
 787         return r;
 788 }
 789
 790 static void multipath_dtr(struct dm_target *ti)
 791 {
 792         struct multipath *m = (struct multipath *) ti->private;
 793
 794         flush_workqueue(kmultipathd);
 795         free_multipath(m);
 796 }
 797
 798 /*
 799  * Map bios, recording original fields for later in case we have to resubmit
 800  */
 801 static int multipath_map(struct dm_target *ti, struct bio *bio,
 802                          union map_info *map_context)
 803 {
 804         int r;
 805         struct dm_mpath_io *mpio;
 806         struct multipath *m = (struct multipath *) ti->private;
 807
 808         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 809         dm_bio_record(&mpio->details, bio);
 810
 811         map_context->ptr = mpio;
 812         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 813         r = map_io(m, bio, mpio, 0);
 814         if (r < 0 || r == DM_MAPIO_REQUEUE)
 815                 mempool_free(mpio, m->mpio_pool);
 816
 817         return r;
 818 }
 819
 820 /*
 821  * Take a path out of use.
 822  */
 823 static int fail_path(struct pgpath *pgpath)
 824 {
 825         unsigned long flags;
 826         struct multipath *m = pgpath->pg->m;
 827
 828         spin_lock_irqsave(&m->lock, flags);
 829
 830         if (!pgpath->path.is_active)
 831                 goto out;
 832
 833         DMWARN("Failing path %s.", pgpath->path.dev->name);
 834
 835         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 836         pgpath->path.is_active = 0;
 837         pgpath->fail_count++;
 838
 839         m->nr_valid_paths--;
 840
 841         if (pgpath == m->current_pgpath)
 842                 m->current_pgpath = NULL;
 843
 844         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 845                       pgpath->path.dev->name, m->nr_valid_paths);
 846
 847         queue_work(kmultipathd, &m->trigger_event);
 848
 849 out:
 850         spin_unlock_irqrestore(&m->lock, flags);
 851
 852         return 0;
 853 }
 854
 855 /*
 856  * Reinstate a previously-failed path
 857  */
 858 static int reinstate_path(struct pgpath *pgpath)
 859 {
 860         int r = 0;
 861         unsigned long flags;
 862         struct multipath *m = pgpath->pg->m;
 863
 864         spin_lock_irqsave(&m->lock, flags);
 865
 866         if (pgpath->path.is_active)
 867                 goto out;
 868
 869         if (!pgpath->pg->ps.type) {
 870                 DMWARN("Reinstate path not supported by path selector %s",
 871                        pgpath->pg->ps.type->name);
 872                 r = -EINVAL;
 873                 goto out;
 874         }
 875
 876         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 877         if (r)
 878                 goto out;
 879
 880         pgpath->path.is_active = 1;
 881
 882         m->current_pgpath = NULL;
 883         if (!m->nr_valid_paths++ && m->queue_size)
 884                 queue_work(kmultipathd, &m->process_queued_ios);
 885
 886         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 887                       pgpath->path.dev->name, m->nr_valid_paths);
 888
 889         queue_work(kmultipathd, &m->trigger_event);
 890
 891 out:
 892         spin_unlock_irqrestore(&m->lock, flags);
 893
 894         return r;
 895 }
 896
 897 /*
 898  * Fail or reinstate all paths that match the provided struct dm_dev.
 899  */
 900 static int action_dev(struct multipath *m, struct dm_dev *dev,
 901                       action_fn action)
 902 {
 903         int r = 0;
 904         struct pgpath *pgpath;
 905         struct priority_group *pg;
 906
 907         list_for_each_entry(pg, &m->priority_groups, list) {
 908                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 909                         if (pgpath->path.dev == dev)
 910                                 r = action(pgpath);
 911                 }
 912         }
 913
 914         return r;
 915 }
 916
 917 /*
 918  * Temporarily try to avoid having to use the specified PG
 919  */
 920 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 921                       int bypassed)
 922 {
 923         unsigned long flags;
 924
 925         spin_lock_irqsave(&m->lock, flags);
 926
 927         pg->bypassed = bypassed;
 928         m->current_pgpath = NULL;
 929         m->current_pg = NULL;
 930
 931         spin_unlock_irqrestore(&m->lock, flags);
 932
 933         queue_work(kmultipathd, &m->trigger_event);
 934 }
 935
 936 /*
 937  * Switch to using the specified PG from the next I/O that gets mapped
 938  */
 939 static int switch_pg_num(struct multipath *m, const char *pgstr)
 940 {
 941         struct priority_group *pg;
 942         unsigned pgnum;
 943         unsigned long flags;
 944
 945         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 946             (pgnum > m->nr_priority_groups)) {
 947                 DMWARN("invalid PG number supplied to switch_pg_num");
 948                 return -EINVAL;
 949         }
 950
 951         spin_lock_irqsave(&m->lock, flags);
 952         list_for_each_entry(pg, &m->priority_groups, list) {
 953                 pg->bypassed = 0;
 954                 if (--pgnum)
 955                         continue;
 956
 957                 m->current_pgpath = NULL;
 958                 m->current_pg = NULL;
 959                 m->next_pg = pg;
 960         }
 961         spin_unlock_irqrestore(&m->lock, flags);
 962
 963         queue_work(kmultipathd, &m->trigger_event);
 964         return 0;
 965 }
 966
 967 /*
 968  * Set/clear bypassed status of a PG.
 969  * PGs are numbered upwards from 1 in the order they were declared.
 970  */
 971 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 972 {
 973         struct priority_group *pg;
 974         unsigned pgnum;
 975
 976         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 977             (pgnum > m->nr_priority_groups)) {
 978                 DMWARN("invalid PG number supplied to bypass_pg");
 979                 return -EINVAL;
 980         }
 981
 982         list_for_each_entry(pg, &m->priority_groups, list) {
 983                 if (!--pgnum)
 984                         break;
 985         }
 986
 987         bypass_pg(m, pg, bypassed);
 988         return 0;
 989 }
 990
 991 /*
 992  * Should we retry pg_init immediately?
 993  */
 994 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
 995 {
 996         unsigned long flags;
 997         int limit_reached = 0;
 998
 999         spin_lock_irqsave(&m->lock, flags);
1000
1001         if (m->pg_init_count <= m->pg_init_retries)
1002                 m->pg_init_required = 1;
1003         else
1004                 limit_reached = 1;
1005
1006         spin_unlock_irqrestore(&m->lock, flags);
1007
1008         return limit_reached;
1009 }
1010
1011 /*
1012  * pg_init must call this when it has completed its initialisation
1013  */
1014 void dm_pg_init_complete(struct dm_path *path, unsigned err_flags)
1015 {
1016         struct pgpath *pgpath = path_to_pgpath(path);
1017         struct priority_group *pg = pgpath->pg;
1018         struct multipath *m = pg->m;
1019         unsigned long flags;
1020
1021         /*
1022          * If requested, retry pg_init until maximum number of retries exceeded.
1023          * If retry not requested and PG already bypassed, always fail the path.
1024          */
1025         if (err_flags & MP_RETRY) {
1026                 if (pg_init_limit_reached(m, pgpath))
1027                         err_flags |= MP_FAIL_PATH;
1028         } else if (err_flags && pg->bypassed)
1029                 err_flags |= MP_FAIL_PATH;
1030
1031         if (err_flags & MP_FAIL_PATH)
1032                 fail_path(pgpath);
1033
1034         if (err_flags & MP_BYPASS_PG)
1035                 bypass_pg(m, pg, 1);
1036
1037         spin_lock_irqsave(&m->lock, flags);
1038         if (err_flags & ~MP_RETRY) {
1039                 m->current_pgpath = NULL;
1040                 m->current_pg = NULL;
1041         } else if (!m->pg_init_required)
1042                 m->queue_io = 0;
1043
1044         m->pg_init_in_progress = 0;
1045         queue_work(kmultipathd, &m->process_queued_ios);
1046         spin_unlock_irqrestore(&m->lock, flags);
1047 }
1048
1049 static void pg_init_done(struct dm_path *path, int errors)
1050 {
1051         struct pgpath *pgpath = path_to_pgpath(path);
1052         struct priority_group *pg = pgpath->pg;
1053         struct multipath *m = pg->m;
1054         unsigned long flags;
1055
1056         /* device or driver problems */
1057         switch (errors) {
1058         case SCSI_DH_OK:
1059                 break;
1060         case SCSI_DH_NOSYS:
1061                 if (!m->hw_handler_name) {
1062                         errors = 0;
1063                         break;
1064                 }
1065                 DMERR("Cannot failover device because scsi_dh_%s was not "
1066                       "loaded.", m->hw_handler_name);
1067                 /*
1068                  * Fail path for now, so we do not ping pong
1069                  */
1070                 fail_path(pgpath);
1071                 break;
1072         case SCSI_DH_DEV_TEMP_BUSY:
1073                 /*
1074                  * Probably doing something like FW upgrade on the
1075                  * controller so try the other pg.
1076                  */
1077                 bypass_pg(m, pg, 1);
1078                 break;
1079         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1080         case SCSI_DH_RETRY:
1081         case SCSI_DH_IMM_RETRY:
1082         case SCSI_DH_RES_TEMP_UNAVAIL:
1083                 if (pg_init_limit_reached(m, pgpath))
1084                         fail_path(pgpath);
1085                 errors = 0;
1086                 break;
1087         default:
1088                 /*
1089                  * We probably do not want to fail the path for a device
1090                  * error, but this is what the old dm did. In future
1091                  * patches we can do more advanced handling.
1092                  */
1093                 fail_path(pgpath);
1094         }
1095
1096         spin_lock_irqsave(&m->lock, flags);
1097         if (errors) {
1098                 DMERR("Could not failover device. Error %d.", errors);
1099                 m->current_pgpath = NULL;
1100                 m->current_pg = NULL;
1101         } else if (!m->pg_init_required) {
1102                 m->queue_io = 0;
1103                 pg->bypassed = 0;
1104         }
1105
1106         m->pg_init_in_progress = 0;
1107         queue_work(kmultipathd, &m->process_queued_ios);
1108         spin_unlock_irqrestore(&m->lock, flags);
1109 }
1110
1111 /*
1112  * end_io handling
1113  */
1114 static int do_end_io(struct multipath *m, struct bio *bio,
1115                      int error, struct dm_mpath_io *mpio)
1116 {
1117         unsigned long flags;
1118
1119         if (!error)
1120                 return 0;       /* I/O complete */
1121
1122         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1123                 return error;
1124
1125         if (error == -EOPNOTSUPP)
1126                 return error;
1127
1128         spin_lock_irqsave(&m->lock, flags);
1129         if (!m->nr_valid_paths) {
1130                 if (__must_push_back(m)) {
1131                         spin_unlock_irqrestore(&m->lock, flags);
1132                         return DM_ENDIO_REQUEUE;
1133                 } else if (!m->queue_if_no_path) {
1134                         spin_unlock_irqrestore(&m->lock, flags);
1135                         return -EIO;
1136                 } else {
1137                         spin_unlock_irqrestore(&m->lock, flags);
1138                         goto requeue;
1139                 }
1140         }
1141         spin_unlock_irqrestore(&m->lock, flags);
1142
1143         if (mpio->pgpath)
1144                 fail_path(mpio->pgpath);
1145
1146       requeue:
1147         dm_bio_restore(&mpio->details, bio);
1148
1149         /* queue for the daemon to resubmit or fail */
1150         spin_lock_irqsave(&m->lock, flags);
1151         bio_list_add(&m->queued_ios, bio);
1152         m->queue_size++;
1153         if (!m->queue_io)
1154                 queue_work(kmultipathd, &m->process_queued_ios);
1155         spin_unlock_irqrestore(&m->lock, flags);
1156
1157         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1158 }
1159
1160 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1161                             int error, union map_info *map_context)
1162 {
1163         struct multipath *m = ti->private;
1164         struct dm_mpath_io *mpio = map_context->ptr;
1165         struct pgpath *pgpath = mpio->pgpath;
1166         struct path_selector *ps;
1167         int r;
1168
1169         r  = do_end_io(m, bio, error, mpio);
1170         if (pgpath) {
1171                 ps = &pgpath->pg->ps;
1172                 if (ps->type->end_io)
1173                         ps->type->end_io(ps, &pgpath->path);
1174         }
1175         if (r != DM_ENDIO_INCOMPLETE)
1176                 mempool_free(mpio, m->mpio_pool);
1177
1178         return r;
1179 }
1180
1181 /*
1182  * Suspend can't complete until all the I/O is processed so if
1183  * the last path fails we must error any remaining I/O.
1184  * Note that if the freeze_bdev fails while suspending, the
1185  * queue_if_no_path state is lost - userspace should reset it.
1186  */
1187 static void multipath_presuspend(struct dm_target *ti)
1188 {
1189         struct multipath *m = (struct multipath *) ti->private;
1190
1191         queue_if_no_path(m, 0, 1);
1192 }
1193
1194 /*
1195  * Restore the queue_if_no_path setting.
1196  */
1197 static void multipath_resume(struct dm_target *ti)
1198 {
1199         struct multipath *m = (struct multipath *) ti->private;
1200         unsigned long flags;
1201
1202         spin_lock_irqsave(&m->lock, flags);
1203         m->queue_if_no_path = m->saved_queue_if_no_path;
1204         spin_unlock_irqrestore(&m->lock, flags);
1205 }
1206
1207 /*
1208  * Info output has the following format:
1209  * num_multipath_feature_args [multipath_feature_args]*
1210  * num_handler_status_args [handler_status_args]*
1211  * num_groups init_group_number
1212  *            [A|D|E num_ps_status_args [ps_status_args]*
1213  *             num_paths num_selector_args
1214  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1215  *
1216  * Table output has the following format (identical to the constructor string):
1217  * num_feature_args [features_args]*
1218  * num_handler_args hw_handler [hw_handler_args]*
1219  * num_groups init_group_number
1220  *     [priority selector-name num_ps_args [ps_args]*
1221  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1222  */
1223 static int multipath_status(struct dm_target *ti, status_type_t type,
1224                             char *result, unsigned int maxlen)
1225 {
1226         int sz = 0;
1227         unsigned long flags;
1228         struct multipath *m = (struct multipath *) ti->private;
1229         struct priority_group *pg;
1230         struct pgpath *p;
1231         unsigned pg_num;
1232         char state;
1233
1234         spin_lock_irqsave(&m->lock, flags);
1235
1236         /* Features */
1237         if (type == STATUSTYPE_INFO)
1238                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1239         else {
1240                 DMEMIT("%u ", m->queue_if_no_path +
1241                               (m->pg_init_retries > 0) * 2);
1242                 if (m->queue_if_no_path)
1243                         DMEMIT("queue_if_no_path ");
1244                 if (m->pg_init_retries)
1245                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1246         }
1247
1248         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1249                 DMEMIT("0 ");
1250         else
1251                 DMEMIT("1 %s ", m->hw_handler_name);
1252
1253         DMEMIT("%u ", m->nr_priority_groups);
1254
1255         if (m->next_pg)
1256                 pg_num = m->next_pg->pg_num;
1257         else if (m->current_pg)
1258                 pg_num = m->current_pg->pg_num;
1259         else
1260                         pg_num = 1;
1261
1262         DMEMIT("%u ", pg_num);
1263
1264         switch (type) {
1265         case STATUSTYPE_INFO:
1266                 list_for_each_entry(pg, &m->priority_groups, list) {
1267                         if (pg->bypassed)
1268                                 state = 'D';    /* Disabled */
1269                         else if (pg == m->current_pg)
1270                                 state = 'A';    /* Currently Active */
1271                         else
1272                                 state = 'E';    /* Enabled */
1273
1274                         DMEMIT("%c ", state);
1275
1276                         if (pg->ps.type->status)
1277                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1278                                                           result + sz,
1279                                                           maxlen - sz);
1280                         else
1281                                 DMEMIT("0 ");
1282
1283                         DMEMIT("%u %u ", pg->nr_pgpaths,
1284                                pg->ps.type->info_args);
1285
1286                         list_for_each_entry(p, &pg->pgpaths, list) {
1287                                 DMEMIT("%s %s %u ", p->path.dev->name,
1288                                        p->path.is_active ? "A" : "F",
1289                                        p->fail_count);
1290                                 if (pg->ps.type->status)
1291                                         sz += pg->ps.type->status(&pg->ps,
1292                                               &p->path, type, result + sz,
1293                                               maxlen - sz);
1294                         }
1295                 }
1296                 break;
1297
1298         case STATUSTYPE_TABLE:
1299                 list_for_each_entry(pg, &m->priority_groups, list) {
1300                         DMEMIT("%s ", pg->ps.type->name);
1301
1302                         if (pg->ps.type->status)
1303                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1304                                                           result + sz,
1305                                                           maxlen - sz);
1306                         else
1307                                 DMEMIT("0 ");
1308
1309                         DMEMIT("%u %u ", pg->nr_pgpaths,
1310                                pg->ps.type->table_args);
1311
1312                         list_for_each_entry(p, &pg->pgpaths, list) {
1313                                 DMEMIT("%s ", p->path.dev->name);
1314                                 if (pg->ps.type->status)
1315                                         sz += pg->ps.type->status(&pg->ps,
1316                                               &p->path, type, result + sz,
1317                                               maxlen - sz);
1318                         }
1319                 }
1320                 break;
1321         }
1322
1323         spin_unlock_irqrestore(&m->lock, flags);
1324
1325         return 0;
1326 }
1327
1328 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1329 {
1330         int r;
1331         struct dm_dev *dev;
1332         struct multipath *m = (struct multipath *) ti->private;
1333         action_fn action;
1334
1335         if (argc == 1) {
1336                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1337                         return queue_if_no_path(m, 1, 0);
1338                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1339                         return queue_if_no_path(m, 0, 0);
1340         }
1341
1342         if (argc != 2)
1343                 goto error;
1344
1345         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1346                 return bypass_pg_num(m, argv[1], 1);
1347         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1348                 return bypass_pg_num(m, argv[1], 0);
1349         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1350                 return switch_pg_num(m, argv[1]);
1351         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1352                 action = reinstate_path;
1353         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1354                 action = fail_path;
1355         else
1356                 goto error;
1357
1358         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1359                           dm_table_get_mode(ti->table), &dev);
1360         if (r) {
1361                 DMWARN("message: error getting device %s",
1362                        argv[1]);
1363                 return -EINVAL;
1364         }
1365
1366         r = action_dev(m, dev, action);
1367
1368         dm_put_device(ti, dev);
1369
1370         return r;
1371
1372 error:
1373         DMWARN("Unrecognised multipath message received.");
1374         return -EINVAL;
1375 }
1376
1377 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1378                            struct file *filp, unsigned int cmd,
1379                            unsigned long arg)
1380 {
1381         struct multipath *m = (struct multipath *) ti->private;
1382         struct block_device *bdev = NULL;
1383         unsigned long flags;
1384         struct file fake_file = {};
1385         struct dentry fake_dentry = {};
1386         int r = 0;
1387
1388         fake_file.f_path.dentry = &fake_dentry;
1389
1390         spin_lock_irqsave(&m->lock, flags);
1391
1392         if (!m->current_pgpath)
1393                 __choose_pgpath(m);
1394
1395         if (m->current_pgpath) {
1396                 bdev = m->current_pgpath->path.dev->bdev;
1397                 fake_dentry.d_inode = bdev->bd_inode;
1398                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1399         }
1400
1401         if (m->queue_io)
1402                 r = -EAGAIN;
1403         else if (!bdev)
1404                 r = -EIO;
1405
1406         spin_unlock_irqrestore(&m->lock, flags);
1407
1408         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1409                                          bdev->bd_disk, cmd, arg);
1410 }
1411
1412 /*-----------------------------------------------------------------
1413  * Module setup
1414  *---------------------------------------------------------------*/
1415 static struct target_type multipath_target = {
1416         .name = "multipath",
1417         .version = {1, 0, 5},
1418         .module = THIS_MODULE,
1419         .ctr = multipath_ctr,
1420         .dtr = multipath_dtr,
1421         .map = multipath_map,
1422         .end_io = multipath_end_io,
1423         .presuspend = multipath_presuspend,
1424         .resume = multipath_resume,
1425         .status = multipath_status,
1426         .message = multipath_message,
1427         .ioctl  = multipath_ioctl,
1428 };
1429
1430 static int __init dm_multipath_init(void)
1431 {
1432         int r;
1433
1434         /* allocate a slab for the dm_ios */
1435         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1436         if (!_mpio_cache)
1437                 return -ENOMEM;
1438
1439         r = dm_register_target(&multipath_target);
1440         if (r < 0) {
1441                 DMERR("register failed %d", r);
1442                 kmem_cache_destroy(_mpio_cache);
1443                 return -EINVAL;
1444         }
1445
1446         kmultipathd = create_workqueue("kmpathd");
1447         if (!kmultipathd) {
1448                 DMERR("failed to create workqueue kmpathd");
1449                 dm_unregister_target(&multipath_target);
1450                 kmem_cache_destroy(_mpio_cache);
1451                 return -ENOMEM;
1452         }
1453
1454         DMINFO("version %u.%u.%u loaded",
1455                multipath_target.version[0], multipath_target.version[1],
1456                multipath_target.version[2]);
1457
1458         return r;
1459 }
1460
1461 static void __exit dm_multipath_exit(void)
1462 {
1463         int r;
1464
1465         destroy_workqueue(kmultipathd);
1466
1467         r = dm_unregister_target(&multipath_target);
1468         if (r < 0)
1469                 DMERR("target unregister failed %d", r);
1470         kmem_cache_destroy(_mpio_cache);
1471 }
1472
1473 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1474
1475 module_init(dm_multipath_init);
1476 module_exit(dm_multipath_exit);
1477
1478 MODULE_DESCRIPTION(DM_NAME " multipath target");
1479 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1480 MODULE_LICENSE("GPL");