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                 return -EINVAL;
 530
 531         r = pst->create(&pg->ps, ps_argc, as->argv);
 532         if (r) {
 533                 dm_put_path_selector(pst);
 534                 ti->error = "path selector constructor failed";
 535                 return r;
 536         }
 537
 538         pg->ps.type = pst;
 539         consume(as, ps_argc);
 540
 541         return 0;
 542 }
 543
 544 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 545                                struct dm_target *ti)
 546 {
 547         int r;
 548         struct pgpath *p;
 549
 550         /* we need at least a path arg */
 551         if (as->argc < 1) {
 552                 ti->error = "no device given";
 553                 return NULL;
 554         }
 555
 556         p = alloc_pgpath();
 557         if (!p)
 558                 return NULL;
 559
 560         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 561                           dm_table_get_mode(ti->table), &p->path.dev);
 562         if (r) {
 563                 ti->error = "error getting device";
 564                 goto bad;
 565         }
 566
 567         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 568         if (r) {
 569                 dm_put_device(ti, p->path.dev);
 570                 goto bad;
 571         }
 572
 573         return p;
 574
 575  bad:
 576         free_pgpath(p);
 577         return NULL;
 578 }
 579
 580 static struct priority_group *parse_priority_group(struct arg_set *as,
 581                                                    struct multipath *m)
 582 {
 583         static struct param _params[] = {
 584                 {1, 1024, "invalid number of paths"},
 585                 {0, 1024, "invalid number of selector args"}
 586         };
 587
 588         int r;
 589         unsigned i, nr_selector_args, nr_params;
 590         struct priority_group *pg;
 591         struct dm_target *ti = m->ti;
 592
 593         if (as->argc < 2) {
 594                 as->argc = 0;
 595                 ti->error = "not enough priority group aruments";
 596                 return NULL;
 597         }
 598
 599         pg = alloc_priority_group();
 600         if (!pg) {
 601                 ti->error = "couldn't allocate priority group";
 602                 return NULL;
 603         }
 604         pg->m = m;
 605
 606         r = parse_path_selector(as, pg, ti);
 607         if (r)
 608                 goto bad;
 609
 610         /*
 611          * read the paths
 612          */
 613         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 614         if (r)
 615                 goto bad;
 616
 617         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 618         if (r)
 619                 goto bad;
 620
 621         nr_params = 1 + nr_selector_args;
 622         for (i = 0; i < pg->nr_pgpaths; i++) {
 623                 struct pgpath *pgpath;
 624                 struct arg_set path_args;
 625
 626                 if (as->argc < nr_params)
 627                         goto bad;
 628
 629                 path_args.argc = nr_params;
 630                 path_args.argv = as->argv;
 631
 632                 pgpath = parse_path(&path_args, &pg->ps, ti);
 633                 if (!pgpath)
 634                         goto bad;
 635
 636                 pgpath->pg = pg;
 637                 list_add_tail(&pgpath->list, &pg->pgpaths);
 638                 consume(as, nr_params);
 639         }
 640
 641         return pg;
 642
 643  bad:
 644         free_priority_group(pg, ti);
 645         return NULL;
 646 }
 647
 648 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 649 {
 650         unsigned hw_argc;
 651         struct dm_target *ti = m->ti;
 652
 653         static struct param _params[] = {
 654                 {0, 1024, "invalid number of hardware handler args"},
 655         };
 656
 657         if (read_param(_params, shift(as), &hw_argc, &ti->error))
 658                 return -EINVAL;
 659
 660         if (!hw_argc)
 661                 return 0;
 662
 663         m->hw_handler_name = kstrdup(shift(as), GFP_KERNEL);
 664         request_module("scsi_dh_%s", m->hw_handler_name);
 665         if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 666                 ti->error = "unknown hardware handler type";
 667                 kfree(m->hw_handler_name);
 668                 m->hw_handler_name = NULL;
 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(kmpath_handlerd);
 795         flush_workqueue(kmultipathd);
 796         free_multipath(m);
 797 }
 798
 799 /*
 800  * Map bios, recording original fields for later in case we have to resubmit
 801  */
 802 static int multipath_map(struct dm_target *ti, struct bio *bio,
 803                          union map_info *map_context)
 804 {
 805         int r;
 806         struct dm_mpath_io *mpio;
 807         struct multipath *m = (struct multipath *) ti->private;
 808
 809         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 810         dm_bio_record(&mpio->details, bio);
 811
 812         map_context->ptr = mpio;
 813         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 814         r = map_io(m, bio, mpio, 0);
 815         if (r < 0 || r == DM_MAPIO_REQUEUE)
 816                 mempool_free(mpio, m->mpio_pool);
 817
 818         return r;
 819 }
 820
 821 /*
 822  * Take a path out of use.
 823  */
 824 static int fail_path(struct pgpath *pgpath)
 825 {
 826         unsigned long flags;
 827         struct multipath *m = pgpath->pg->m;
 828
 829         spin_lock_irqsave(&m->lock, flags);
 830
 831         if (!pgpath->path.is_active)
 832                 goto out;
 833
 834         DMWARN("Failing path %s.", pgpath->path.dev->name);
 835
 836         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 837         pgpath->path.is_active = 0;
 838         pgpath->fail_count++;
 839
 840         m->nr_valid_paths--;
 841
 842         if (pgpath == m->current_pgpath)
 843                 m->current_pgpath = NULL;
 844
 845         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 846                       pgpath->path.dev->name, m->nr_valid_paths);
 847
 848         queue_work(kmultipathd, &m->trigger_event);
 849
 850 out:
 851         spin_unlock_irqrestore(&m->lock, flags);
 852
 853         return 0;
 854 }
 855
 856 /*
 857  * Reinstate a previously-failed path
 858  */
 859 static int reinstate_path(struct pgpath *pgpath)
 860 {
 861         int r = 0;
 862         unsigned long flags;
 863         struct multipath *m = pgpath->pg->m;
 864
 865         spin_lock_irqsave(&m->lock, flags);
 866
 867         if (pgpath->path.is_active)
 868                 goto out;
 869
 870         if (!pgpath->pg->ps.type) {
 871                 DMWARN("Reinstate path not supported by path selector %s",
 872                        pgpath->pg->ps.type->name);
 873                 r = -EINVAL;
 874                 goto out;
 875         }
 876
 877         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 878         if (r)
 879                 goto out;
 880
 881         pgpath->path.is_active = 1;
 882
 883         m->current_pgpath = NULL;
 884         if (!m->nr_valid_paths++ && m->queue_size)
 885                 queue_work(kmultipathd, &m->process_queued_ios);
 886
 887         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 888                       pgpath->path.dev->name, m->nr_valid_paths);
 889
 890         queue_work(kmultipathd, &m->trigger_event);
 891
 892 out:
 893         spin_unlock_irqrestore(&m->lock, flags);
 894
 895         return r;
 896 }
 897
 898 /*
 899  * Fail or reinstate all paths that match the provided struct dm_dev.
 900  */
 901 static int action_dev(struct multipath *m, struct dm_dev *dev,
 902                       action_fn action)
 903 {
 904         int r = 0;
 905         struct pgpath *pgpath;
 906         struct priority_group *pg;
 907
 908         list_for_each_entry(pg, &m->priority_groups, list) {
 909                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 910                         if (pgpath->path.dev == dev)
 911                                 r = action(pgpath);
 912                 }
 913         }
 914
 915         return r;
 916 }
 917
 918 /*
 919  * Temporarily try to avoid having to use the specified PG
 920  */
 921 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 922                       int bypassed)
 923 {
 924         unsigned long flags;
 925
 926         spin_lock_irqsave(&m->lock, flags);
 927
 928         pg->bypassed = bypassed;
 929         m->current_pgpath = NULL;
 930         m->current_pg = NULL;
 931
 932         spin_unlock_irqrestore(&m->lock, flags);
 933
 934         queue_work(kmultipathd, &m->trigger_event);
 935 }
 936
 937 /*
 938  * Switch to using the specified PG from the next I/O that gets mapped
 939  */
 940 static int switch_pg_num(struct multipath *m, const char *pgstr)
 941 {
 942         struct priority_group *pg;
 943         unsigned pgnum;
 944         unsigned long flags;
 945
 946         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 947             (pgnum > m->nr_priority_groups)) {
 948                 DMWARN("invalid PG number supplied to switch_pg_num");
 949                 return -EINVAL;
 950         }
 951
 952         spin_lock_irqsave(&m->lock, flags);
 953         list_for_each_entry(pg, &m->priority_groups, list) {
 954                 pg->bypassed = 0;
 955                 if (--pgnum)
 956                         continue;
 957
 958                 m->current_pgpath = NULL;
 959                 m->current_pg = NULL;
 960                 m->next_pg = pg;
 961         }
 962         spin_unlock_irqrestore(&m->lock, flags);
 963
 964         queue_work(kmultipathd, &m->trigger_event);
 965         return 0;
 966 }
 967
 968 /*
 969  * Set/clear bypassed status of a PG.
 970  * PGs are numbered upwards from 1 in the order they were declared.
 971  */
 972 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 973 {
 974         struct priority_group *pg;
 975         unsigned pgnum;
 976
 977         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 978             (pgnum > m->nr_priority_groups)) {
 979                 DMWARN("invalid PG number supplied to bypass_pg");
 980                 return -EINVAL;
 981         }
 982
 983         list_for_each_entry(pg, &m->priority_groups, list) {
 984                 if (!--pgnum)
 985                         break;
 986         }
 987
 988         bypass_pg(m, pg, bypassed);
 989         return 0;
 990 }
 991
 992 /*
 993  * Should we retry pg_init immediately?
 994  */
 995 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
 996 {
 997         unsigned long flags;
 998         int limit_reached = 0;
 999
1000         spin_lock_irqsave(&m->lock, flags);
1001
1002         if (m->pg_init_count <= m->pg_init_retries)
1003                 m->pg_init_required = 1;
1004         else
1005                 limit_reached = 1;
1006
1007         spin_unlock_irqrestore(&m->lock, flags);
1008
1009         return limit_reached;
1010 }
1011
1012 static void pg_init_done(struct dm_path *path, int errors)
1013 {
1014         struct pgpath *pgpath = path_to_pgpath(path);
1015         struct priority_group *pg = pgpath->pg;
1016         struct multipath *m = pg->m;
1017         unsigned long flags;
1018
1019         /* device or driver problems */
1020         switch (errors) {
1021         case SCSI_DH_OK:
1022                 break;
1023         case SCSI_DH_NOSYS:
1024                 if (!m->hw_handler_name) {
1025                         errors = 0;
1026                         break;
1027                 }
1028                 DMERR("Cannot failover device because scsi_dh_%s was not "
1029                       "loaded.", m->hw_handler_name);
1030                 /*
1031                  * Fail path for now, so we do not ping pong
1032                  */
1033                 fail_path(pgpath);
1034                 break;
1035         case SCSI_DH_DEV_TEMP_BUSY:
1036                 /*
1037                  * Probably doing something like FW upgrade on the
1038                  * controller so try the other pg.
1039                  */
1040                 bypass_pg(m, pg, 1);
1041                 break;
1042         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1043         case SCSI_DH_RETRY:
1044         case SCSI_DH_IMM_RETRY:
1045         case SCSI_DH_RES_TEMP_UNAVAIL:
1046                 if (pg_init_limit_reached(m, pgpath))
1047                         fail_path(pgpath);
1048                 errors = 0;
1049                 break;
1050         default:
1051                 /*
1052                  * We probably do not want to fail the path for a device
1053                  * error, but this is what the old dm did. In future
1054                  * patches we can do more advanced handling.
1055                  */
1056                 fail_path(pgpath);
1057         }
1058
1059         spin_lock_irqsave(&m->lock, flags);
1060         if (errors) {
1061                 DMERR("Could not failover device. Error %d.", errors);
1062                 m->current_pgpath = NULL;
1063                 m->current_pg = NULL;
1064         } else if (!m->pg_init_required) {
1065                 m->queue_io = 0;
1066                 pg->bypassed = 0;
1067         }
1068
1069         m->pg_init_in_progress = 0;
1070         queue_work(kmultipathd, &m->process_queued_ios);
1071         spin_unlock_irqrestore(&m->lock, flags);
1072 }
1073
1074 static void activate_path(struct work_struct *work)
1075 {
1076         int ret;
1077         struct multipath *m =
1078                 container_of(work, struct multipath, activate_path);
1079         struct dm_path *path = &m->current_pgpath->path;
1080
1081         ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
1082         pg_init_done(path, ret);
1083 }
1084
1085 /*
1086  * end_io handling
1087  */
1088 static int do_end_io(struct multipath *m, struct bio *bio,
1089                      int error, struct dm_mpath_io *mpio)
1090 {
1091         unsigned long flags;
1092
1093         if (!error)
1094                 return 0;       /* I/O complete */
1095
1096         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1097                 return error;
1098
1099         if (error == -EOPNOTSUPP)
1100                 return error;
1101
1102         spin_lock_irqsave(&m->lock, flags);
1103         if (!m->nr_valid_paths) {
1104                 if (__must_push_back(m)) {
1105                         spin_unlock_irqrestore(&m->lock, flags);
1106                         return DM_ENDIO_REQUEUE;
1107                 } else if (!m->queue_if_no_path) {
1108                         spin_unlock_irqrestore(&m->lock, flags);
1109                         return -EIO;
1110                 } else {
1111                         spin_unlock_irqrestore(&m->lock, flags);
1112                         goto requeue;
1113                 }
1114         }
1115         spin_unlock_irqrestore(&m->lock, flags);
1116
1117         if (mpio->pgpath)
1118                 fail_path(mpio->pgpath);
1119
1120       requeue:
1121         dm_bio_restore(&mpio->details, bio);
1122
1123         /* queue for the daemon to resubmit or fail */
1124         spin_lock_irqsave(&m->lock, flags);
1125         bio_list_add(&m->queued_ios, bio);
1126         m->queue_size++;
1127         if (!m->queue_io)
1128                 queue_work(kmultipathd, &m->process_queued_ios);
1129         spin_unlock_irqrestore(&m->lock, flags);
1130
1131         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1132 }
1133
1134 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1135                             int error, union map_info *map_context)
1136 {
1137         struct multipath *m = ti->private;
1138         struct dm_mpath_io *mpio = map_context->ptr;
1139         struct pgpath *pgpath = mpio->pgpath;
1140         struct path_selector *ps;
1141         int r;
1142
1143         r  = do_end_io(m, bio, error, mpio);
1144         if (pgpath) {
1145                 ps = &pgpath->pg->ps;
1146                 if (ps->type->end_io)
1147                         ps->type->end_io(ps, &pgpath->path);
1148         }
1149         if (r != DM_ENDIO_INCOMPLETE)
1150                 mempool_free(mpio, m->mpio_pool);
1151
1152         return r;
1153 }
1154
1155 /*
1156  * Suspend can't complete until all the I/O is processed so if
1157  * the last path fails we must error any remaining I/O.
1158  * Note that if the freeze_bdev fails while suspending, the
1159  * queue_if_no_path state is lost - userspace should reset it.
1160  */
1161 static void multipath_presuspend(struct dm_target *ti)
1162 {
1163         struct multipath *m = (struct multipath *) ti->private;
1164
1165         queue_if_no_path(m, 0, 1);
1166 }
1167
1168 /*
1169  * Restore the queue_if_no_path setting.
1170  */
1171 static void multipath_resume(struct dm_target *ti)
1172 {
1173         struct multipath *m = (struct multipath *) ti->private;
1174         unsigned long flags;
1175
1176         spin_lock_irqsave(&m->lock, flags);
1177         m->queue_if_no_path = m->saved_queue_if_no_path;
1178         spin_unlock_irqrestore(&m->lock, flags);
1179 }
1180
1181 /*
1182  * Info output has the following format:
1183  * num_multipath_feature_args [multipath_feature_args]*
1184  * num_handler_status_args [handler_status_args]*
1185  * num_groups init_group_number
1186  *            [A|D|E num_ps_status_args [ps_status_args]*
1187  *             num_paths num_selector_args
1188  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1189  *
1190  * Table output has the following format (identical to the constructor string):
1191  * num_feature_args [features_args]*
1192  * num_handler_args hw_handler [hw_handler_args]*
1193  * num_groups init_group_number
1194  *     [priority selector-name num_ps_args [ps_args]*
1195  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1196  */
1197 static int multipath_status(struct dm_target *ti, status_type_t type,
1198                             char *result, unsigned int maxlen)
1199 {
1200         int sz = 0;
1201         unsigned long flags;
1202         struct multipath *m = (struct multipath *) ti->private;
1203         struct priority_group *pg;
1204         struct pgpath *p;
1205         unsigned pg_num;
1206         char state;
1207
1208         spin_lock_irqsave(&m->lock, flags);
1209
1210         /* Features */
1211         if (type == STATUSTYPE_INFO)
1212                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1213         else {
1214                 DMEMIT("%u ", m->queue_if_no_path +
1215                               (m->pg_init_retries > 0) * 2);
1216                 if (m->queue_if_no_path)
1217                         DMEMIT("queue_if_no_path ");
1218                 if (m->pg_init_retries)
1219                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1220         }
1221
1222         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1223                 DMEMIT("0 ");
1224         else
1225                 DMEMIT("1 %s ", m->hw_handler_name);
1226
1227         DMEMIT("%u ", m->nr_priority_groups);
1228
1229         if (m->next_pg)
1230                 pg_num = m->next_pg->pg_num;
1231         else if (m->current_pg)
1232                 pg_num = m->current_pg->pg_num;
1233         else
1234                         pg_num = 1;
1235
1236         DMEMIT("%u ", pg_num);
1237
1238         switch (type) {
1239         case STATUSTYPE_INFO:
1240                 list_for_each_entry(pg, &m->priority_groups, list) {
1241                         if (pg->bypassed)
1242                                 state = 'D';    /* Disabled */
1243                         else if (pg == m->current_pg)
1244                                 state = 'A';    /* Currently Active */
1245                         else
1246                                 state = 'E';    /* Enabled */
1247
1248                         DMEMIT("%c ", state);
1249
1250                         if (pg->ps.type->status)
1251                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1252                                                           result + sz,
1253                                                           maxlen - sz);
1254                         else
1255                                 DMEMIT("0 ");
1256
1257                         DMEMIT("%u %u ", pg->nr_pgpaths,
1258                                pg->ps.type->info_args);
1259
1260                         list_for_each_entry(p, &pg->pgpaths, list) {
1261                                 DMEMIT("%s %s %u ", p->path.dev->name,
1262                                        p->path.is_active ? "A" : "F",
1263                                        p->fail_count);
1264                                 if (pg->ps.type->status)
1265                                         sz += pg->ps.type->status(&pg->ps,
1266                                               &p->path, type, result + sz,
1267                                               maxlen - sz);
1268                         }
1269                 }
1270                 break;
1271
1272         case STATUSTYPE_TABLE:
1273                 list_for_each_entry(pg, &m->priority_groups, list) {
1274                         DMEMIT("%s ", pg->ps.type->name);
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->table_args);
1285
1286                         list_for_each_entry(p, &pg->pgpaths, list) {
1287                                 DMEMIT("%s ", p->path.dev->name);
1288                                 if (pg->ps.type->status)
1289                                         sz += pg->ps.type->status(&pg->ps,
1290                                               &p->path, type, result + sz,
1291                                               maxlen - sz);
1292                         }
1293                 }
1294                 break;
1295         }
1296
1297         spin_unlock_irqrestore(&m->lock, flags);
1298
1299         return 0;
1300 }
1301
1302 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1303 {
1304         int r;
1305         struct dm_dev *dev;
1306         struct multipath *m = (struct multipath *) ti->private;
1307         action_fn action;
1308
1309         if (argc == 1) {
1310                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1311                         return queue_if_no_path(m, 1, 0);
1312                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1313                         return queue_if_no_path(m, 0, 0);
1314         }
1315
1316         if (argc != 2)
1317                 goto error;
1318
1319         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1320                 return bypass_pg_num(m, argv[1], 1);
1321         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1322                 return bypass_pg_num(m, argv[1], 0);
1323         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1324                 return switch_pg_num(m, argv[1]);
1325         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1326                 action = reinstate_path;
1327         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1328                 action = fail_path;
1329         else
1330                 goto error;
1331
1332         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1333                           dm_table_get_mode(ti->table), &dev);
1334         if (r) {
1335                 DMWARN("message: error getting device %s",
1336                        argv[1]);
1337                 return -EINVAL;
1338         }
1339
1340         r = action_dev(m, dev, action);
1341
1342         dm_put_device(ti, dev);
1343
1344         return r;
1345
1346 error:
1347         DMWARN("Unrecognised multipath message received.");
1348         return -EINVAL;
1349 }
1350
1351 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1352                            struct file *filp, unsigned int cmd,
1353                            unsigned long arg)
1354 {
1355         struct multipath *m = (struct multipath *) ti->private;
1356         struct block_device *bdev = NULL;
1357         unsigned long flags;
1358         struct file fake_file = {};
1359         struct dentry fake_dentry = {};
1360         int r = 0;
1361
1362         fake_file.f_path.dentry = &fake_dentry;
1363
1364         spin_lock_irqsave(&m->lock, flags);
1365
1366         if (!m->current_pgpath)
1367                 __choose_pgpath(m);
1368
1369         if (m->current_pgpath) {
1370                 bdev = m->current_pgpath->path.dev->bdev;
1371                 fake_dentry.d_inode = bdev->bd_inode;
1372                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1373         }
1374
1375         if (m->queue_io)
1376                 r = -EAGAIN;
1377         else if (!bdev)
1378                 r = -EIO;
1379
1380         spin_unlock_irqrestore(&m->lock, flags);
1381
1382         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1383                                          bdev->bd_disk, cmd, arg);
1384 }
1385
1386 /*-----------------------------------------------------------------
1387  * Module setup
1388  *---------------------------------------------------------------*/
1389 static struct target_type multipath_target = {
1390         .name = "multipath",
1391         .version = {1, 0, 5},
1392         .module = THIS_MODULE,
1393         .ctr = multipath_ctr,
1394         .dtr = multipath_dtr,
1395         .map = multipath_map,
1396         .end_io = multipath_end_io,
1397         .presuspend = multipath_presuspend,
1398         .resume = multipath_resume,
1399         .status = multipath_status,
1400         .message = multipath_message,
1401         .ioctl  = multipath_ioctl,
1402 };
1403
1404 static int __init dm_multipath_init(void)
1405 {
1406         int r;
1407
1408         /* allocate a slab for the dm_ios */
1409         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1410         if (!_mpio_cache)
1411                 return -ENOMEM;
1412
1413         r = dm_register_target(&multipath_target);
1414         if (r < 0) {
1415                 DMERR("register failed %d", r);
1416                 kmem_cache_destroy(_mpio_cache);
1417                 return -EINVAL;
1418         }
1419
1420         kmultipathd = create_workqueue("kmpathd");
1421         if (!kmultipathd) {
1422                 DMERR("failed to create workqueue kmpathd");
1423                 dm_unregister_target(&multipath_target);
1424                 kmem_cache_destroy(_mpio_cache);
1425                 return -ENOMEM;
1426         }
1427
1428         /*
1429          * A separate workqueue is used to handle the device handlers
1430          * to avoid overloading existing workqueue. Overloading the
1431          * old workqueue would also create a bottleneck in the
1432          * path of the storage hardware device activation.
1433          */
1434         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1435         if (!kmpath_handlerd) {
1436                 DMERR("failed to create workqueue kmpath_handlerd");
1437                 destroy_workqueue(kmultipathd);
1438                 dm_unregister_target(&multipath_target);
1439                 kmem_cache_destroy(_mpio_cache);
1440                 return -ENOMEM;
1441         }
1442
1443         DMINFO("version %u.%u.%u loaded",
1444                multipath_target.version[0], multipath_target.version[1],
1445                multipath_target.version[2]);
1446
1447         return r;
1448 }
1449
1450 static void __exit dm_multipath_exit(void)
1451 {
1452         int r;
1453
1454         destroy_workqueue(kmpath_handlerd);
1455         destroy_workqueue(kmultipathd);
1456
1457         r = dm_unregister_target(&multipath_target);
1458         if (r < 0)
1459                 DMERR("target unregister failed %d", r);
1460         kmem_cache_destroy(_mpio_cache);
1461 }
1462
1463 module_init(dm_multipath_init);
1464 module_exit(dm_multipath_exit);
1465
1466 MODULE_DESCRIPTION(DM_NAME " multipath target");
1467 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1468 MODULE_LICENSE("GPL");