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