drivers/md/dm-mpath.c

   1 /*
   2  * Copyright (C) 2003 Sistina Software Limited.
   3  * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include "dm.h"
   9 #include "dm-path-selector.h"
  10 #include "dm-hw-handler.h"
  11 #include "dm-bio-list.h"
  12 #include "dm-bio-record.h"
  13
  14 #include <linux/ctype.h>
  15 #include <linux/init.h>
  16 #include <linux/mempool.h>
  17 #include <linux/module.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/slab.h>
  20 #include <linux/time.h>
  21 #include <linux/workqueue.h>
  22 #include <asm/atomic.h>
  23
  24 #define DM_MSG_PREFIX "multipath"
  25 #define MESG_STR(x) x, sizeof(x)
  26
  27 /* Path properties */
  28 struct pgpath {
  29         struct list_head list;
  30
  31         struct priority_group *pg;      /* Owning PG */
  32         unsigned fail_count;            /* Cumulative failure count */
  33
  34         struct path path;
  35 };
  36
  37 #define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  38
  39 /*
  40  * Paths are grouped into Priority Groups and numbered from 1 upwards.
  41  * Each has a path selector which controls which path gets used.
  42  */
  43 struct priority_group {
  44         struct list_head list;
  45
  46         struct multipath *m;            /* Owning multipath instance */
  47         struct path_selector ps;
  48
  49         unsigned pg_num;                /* Reference number */
  50         unsigned bypassed;              /* Temporarily bypass this PG? */
  51
  52         unsigned nr_pgpaths;            /* Number of paths in PG */
  53         struct list_head pgpaths;
  54 };
  55
  56 /* Multipath context */
  57 struct multipath {
  58         struct list_head list;
  59         struct dm_target *ti;
  60
  61         spinlock_t lock;
  62
  63         struct hw_handler hw_handler;
  64         unsigned nr_priority_groups;
  65         struct list_head priority_groups;
  66         unsigned pg_init_required;      /* pg_init needs calling? */
  67         unsigned pg_init_in_progress;   /* Only one pg_init allowed at once */
  68
  69         unsigned nr_valid_paths;        /* Total number of usable paths */
  70         struct pgpath *current_pgpath;
  71         struct priority_group *current_pg;
  72         struct priority_group *next_pg; /* Switch to this PG if set */
  73         unsigned repeat_count;          /* I/Os left before calling PS again */
  74
  75         unsigned queue_io;              /* Must we queue all I/O? */
  76         unsigned queue_if_no_path;      /* Queue I/O if last path fails? */
  77         unsigned saved_queue_if_no_path;/* Saved state during suspension */
  78
  79         struct work_struct process_queued_ios;
  80         struct bio_list queued_ios;
  81         unsigned queue_size;
  82
  83         struct work_struct trigger_event;
  84
  85         /*
  86          * We must use a mempool of mpath_io structs so that we
  87          * can resubmit bios on error.
  88          */
  89         mempool_t *mpio_pool;
  90 };
  91
  92 /*
  93  * Context information attached to each bio we process.
  94  */
  95 struct mpath_io {
  96         struct pgpath *pgpath;
  97         struct dm_bio_details details;
  98 };
  99
 100 typedef int (*action_fn) (struct pgpath *pgpath);
 101
 102 #define MIN_IOS 256     /* Mempool size */
 103
 104 static kmem_cache_t *_mpio_cache;
 105
 106 struct workqueue_struct *kmultipathd;
 107 static void process_queued_ios(void *data);
 108 static void trigger_event(void *data);
 109
 110
 111 /*-----------------------------------------------
 112  * Allocation routines
 113  *-----------------------------------------------*/
 114
 115 static struct pgpath *alloc_pgpath(void)
 116 {
 117         struct pgpath *pgpath = kmalloc(sizeof(*pgpath), GFP_KERNEL);
 118
 119         if (pgpath) {
 120                 memset(pgpath, 0, sizeof(*pgpath));
 121                 pgpath->path.is_active = 1;
 122         }
 123
 124         return pgpath;
 125 }
 126
 127 static inline void free_pgpath(struct pgpath *pgpath)
 128 {
 129         kfree(pgpath);
 130 }
 131
 132 static struct priority_group *alloc_priority_group(void)
 133 {
 134         struct priority_group *pg;
 135
 136         pg = kmalloc(sizeof(*pg), GFP_KERNEL);
 137         if (!pg)
 138                 return NULL;
 139
 140         memset(pg, 0, sizeof(*pg));
 141         INIT_LIST_HEAD(&pg->pgpaths);
 142
 143         return pg;
 144 }
 145
 146 static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 147 {
 148         struct pgpath *pgpath, *tmp;
 149
 150         list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 151                 list_del(&pgpath->list);
 152                 dm_put_device(ti, pgpath->path.dev);
 153                 free_pgpath(pgpath);
 154         }
 155 }
 156
 157 static void free_priority_group(struct priority_group *pg,
 158                                 struct dm_target *ti)
 159 {
 160         struct path_selector *ps = &pg->ps;
 161
 162         if (ps->type) {
 163                 ps->type->destroy(ps);
 164                 dm_put_path_selector(ps->type);
 165         }
 166
 167         free_pgpaths(&pg->pgpaths, ti);
 168         kfree(pg);
 169 }
 170
 171 static struct multipath *alloc_multipath(struct dm_target *ti)
 172 {
 173         struct multipath *m;
 174
 175         m = kmalloc(sizeof(*m), GFP_KERNEL);
 176         if (m) {
 177                 memset(m, 0, sizeof(*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, m);
 182                 INIT_WORK(&m->trigger_event, trigger_event, m);
 183                 m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 184                 if (!m->mpio_pool) {
 185                         kfree(m);
 186                         return NULL;
 187                 }
 188                 m->ti = ti;
 189                 ti->private = m;
 190         }
 191
 192         return m;
 193 }
 194
 195 static void free_multipath(struct multipath *m)
 196 {
 197         struct priority_group *pg, *tmp;
 198         struct hw_handler *hwh = &m->hw_handler;
 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         if (hwh->type) {
 206                 hwh->type->destroy(hwh);
 207                 dm_put_hw_handler(hwh->type);
 208         }
 209
 210         mempool_destroy(m->mpio_pool);
 211         kfree(m);
 212 }
 213
 214
 215 /*-----------------------------------------------
 216  * Path selection
 217  *-----------------------------------------------*/
 218
 219 static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 220 {
 221         struct hw_handler *hwh = &m->hw_handler;
 222
 223         m->current_pg = pgpath->pg;
 224
 225         /* Must we initialise the PG first, and queue I/O till it's ready? */
 226         if (hwh->type && hwh->type->pg_init) {
 227                 m->pg_init_required = 1;
 228                 m->queue_io = 1;
 229         } else {
 230                 m->pg_init_required = 0;
 231                 m->queue_io = 0;
 232         }
 233 }
 234
 235 static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg)
 236 {
 237         struct path *path;
 238
 239         path = pg->ps.type->select_path(&pg->ps, &m->repeat_count);
 240         if (!path)
 241                 return -ENXIO;
 242
 243         m->current_pgpath = path_to_pgpath(path);
 244
 245         if (m->current_pg != pg)
 246                 __switch_pg(m, m->current_pgpath);
 247
 248         return 0;
 249 }
 250
 251 static void __choose_pgpath(struct multipath *m)
 252 {
 253         struct priority_group *pg;
 254         unsigned bypassed = 1;
 255
 256         if (!m->nr_valid_paths)
 257                 goto failed;
 258
 259         /* Were we instructed to switch PG? */
 260         if (m->next_pg) {
 261                 pg = m->next_pg;
 262                 m->next_pg = NULL;
 263                 if (!__choose_path_in_pg(m, pg))
 264                         return;
 265         }
 266
 267         /* Don't change PG until it has no remaining paths */
 268         if (m->current_pg && !__choose_path_in_pg(m, m->current_pg))
 269                 return;
 270
 271         /*
 272          * Loop through priority groups until we find a valid path.
 273          * First time we skip PGs marked 'bypassed'.
 274          * Second time we only try the ones we skipped.
 275          */
 276         do {
 277                 list_for_each_entry(pg, &m->priority_groups, list) {
 278                         if (pg->bypassed == bypassed)
 279                                 continue;
 280                         if (!__choose_path_in_pg(m, pg))
 281                                 return;
 282                 }
 283         } while (bypassed--);
 284
 285 failed:
 286         m->current_pgpath = NULL;
 287         m->current_pg = NULL;
 288 }
 289
 290 static int map_io(struct multipath *m, struct bio *bio, struct mpath_io *mpio,
 291                   unsigned was_queued)
 292 {
 293         int r = 1;
 294         unsigned long flags;
 295         struct pgpath *pgpath;
 296
 297         spin_lock_irqsave(&m->lock, flags);
 298
 299         /* Do we need to select a new pgpath? */
 300         if (!m->current_pgpath ||
 301             (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 302                 __choose_pgpath(m);
 303
 304         pgpath = m->current_pgpath;
 305
 306         if (was_queued)
 307                 m->queue_size--;
 308
 309         if ((pgpath && m->queue_io) ||
 310             (!pgpath && m->queue_if_no_path)) {
 311                 /* Queue for the daemon to resubmit */
 312                 bio_list_add(&m->queued_ios, bio);
 313                 m->queue_size++;
 314                 if ((m->pg_init_required && !m->pg_init_in_progress) ||
 315                     !m->queue_io)
 316                         queue_work(kmultipathd, &m->process_queued_ios);
 317                 pgpath = NULL;
 318                 r = 0;
 319         } else if (!pgpath)
 320                 r = -EIO;               /* Failed */
 321         else
 322                 bio->bi_bdev = pgpath->path.dev->bdev;
 323
 324         mpio->pgpath = pgpath;
 325
 326         spin_unlock_irqrestore(&m->lock, flags);
 327
 328         return r;
 329 }
 330
 331 /*
 332  * If we run out of usable paths, should we queue I/O or error it?
 333  */
 334 static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 335                             unsigned save_old_value)
 336 {
 337         unsigned long flags;
 338
 339         spin_lock_irqsave(&m->lock, flags);
 340
 341         if (save_old_value)
 342                 m->saved_queue_if_no_path = m->queue_if_no_path;
 343         else
 344                 m->saved_queue_if_no_path = queue_if_no_path;
 345         m->queue_if_no_path = queue_if_no_path;
 346         if (!m->queue_if_no_path && m->queue_size)
 347                 queue_work(kmultipathd, &m->process_queued_ios);
 348
 349         spin_unlock_irqrestore(&m->lock, flags);
 350
 351         return 0;
 352 }
 353
 354 /*-----------------------------------------------------------------
 355  * The multipath daemon is responsible for resubmitting queued ios.
 356  *---------------------------------------------------------------*/
 357
 358 static void dispatch_queued_ios(struct multipath *m)
 359 {
 360         int r;
 361         unsigned long flags;
 362         struct bio *bio = NULL, *next;
 363         struct mpath_io *mpio;
 364         union map_info *info;
 365
 366         spin_lock_irqsave(&m->lock, flags);
 367         bio = bio_list_get(&m->queued_ios);
 368         spin_unlock_irqrestore(&m->lock, flags);
 369
 370         while (bio) {
 371                 next = bio->bi_next;
 372                 bio->bi_next = NULL;
 373
 374                 info = dm_get_mapinfo(bio);
 375                 mpio = info->ptr;
 376
 377                 r = map_io(m, bio, mpio, 1);
 378                 if (r < 0)
 379                         bio_endio(bio, bio->bi_size, r);
 380                 else if (r == 1)
 381                         generic_make_request(bio);
 382
 383                 bio = next;
 384         }
 385 }
 386
 387 static void process_queued_ios(void *data)
 388 {
 389         struct multipath *m = (struct multipath *) data;
 390         struct hw_handler *hwh = &m->hw_handler;
 391         struct pgpath *pgpath = NULL;
 392         unsigned init_required = 0, must_queue = 1;
 393         unsigned long flags;
 394
 395         spin_lock_irqsave(&m->lock, flags);
 396
 397         if (!m->queue_size)
 398                 goto out;
 399
 400         if (!m->current_pgpath)
 401                 __choose_pgpath(m);
 402
 403         pgpath = m->current_pgpath;
 404
 405         if ((pgpath && !m->queue_io) ||
 406             (!pgpath && !m->queue_if_no_path))
 407                 must_queue = 0;
 408
 409         if (m->pg_init_required && !m->pg_init_in_progress) {
 410                 m->pg_init_required = 0;
 411                 m->pg_init_in_progress = 1;
 412                 init_required = 1;
 413         }
 414
 415 out:
 416         spin_unlock_irqrestore(&m->lock, flags);
 417
 418         if (init_required)
 419                 hwh->type->pg_init(hwh, pgpath->pg->bypassed, &pgpath->path);
 420
 421         if (!must_queue)
 422                 dispatch_queued_ios(m);
 423 }
 424
 425 /*
 426  * An event is triggered whenever a path is taken out of use.
 427  * Includes path failure and PG bypass.
 428  */
 429 static void trigger_event(void *data)
 430 {
 431         struct multipath *m = (struct multipath *) data;
 432
 433         dm_table_event(m->ti->table);
 434 }
 435
 436 /*-----------------------------------------------------------------
 437  * Constructor/argument parsing:
 438  * <#multipath feature args> [<arg>]*
 439  * <#hw_handler args> [hw_handler [<arg>]*]
 440  * <#priority groups>
 441  * <initial priority group>
 442  *     [<selector> <#selector args> [<arg>]*
 443  *      <#paths> <#per-path selector args>
 444  *         [<path> [<arg>]* ]+ ]+
 445  *---------------------------------------------------------------*/
 446 struct param {
 447         unsigned min;
 448         unsigned max;
 449         char *error;
 450 };
 451
 452 static int read_param(struct param *param, char *str, unsigned *v, char **error)
 453 {
 454         if (!str ||
 455             (sscanf(str, "%u", v) != 1) ||
 456             (*v < param->min) ||
 457             (*v > param->max)) {
 458                 *error = param->error;
 459                 return -EINVAL;
 460         }
 461
 462         return 0;
 463 }
 464
 465 struct arg_set {
 466         unsigned argc;
 467         char **argv;
 468 };
 469
 470 static char *shift(struct arg_set *as)
 471 {
 472         char *r;
 473
 474         if (as->argc) {
 475                 as->argc--;
 476                 r = *as->argv;
 477                 as->argv++;
 478                 return r;
 479         }
 480
 481         return NULL;
 482 }
 483
 484 static void consume(struct arg_set *as, unsigned n)
 485 {
 486         BUG_ON (as->argc < n);
 487         as->argc -= n;
 488         as->argv += n;
 489 }
 490
 491 static int parse_path_selector(struct arg_set *as, struct priority_group *pg,
 492                                struct dm_target *ti)
 493 {
 494         int r;
 495         struct path_selector_type *pst;
 496         unsigned ps_argc;
 497
 498         static struct param _params[] = {
 499                 {0, 1024, "invalid number of path selector args"},
 500         };
 501
 502         pst = dm_get_path_selector(shift(as));
 503         if (!pst) {
 504                 ti->error = "unknown path selector type";
 505                 return -EINVAL;
 506         }
 507
 508         r = read_param(_params, shift(as), &ps_argc, &ti->error);
 509         if (r)
 510                 return -EINVAL;
 511
 512         r = pst->create(&pg->ps, ps_argc, as->argv);
 513         if (r) {
 514                 dm_put_path_selector(pst);
 515                 ti->error = "path selector constructor failed";
 516                 return r;
 517         }
 518
 519         pg->ps.type = pst;
 520         consume(as, ps_argc);
 521
 522         return 0;
 523 }
 524
 525 static struct pgpath *parse_path(struct arg_set *as, struct path_selector *ps,
 526                                struct dm_target *ti)
 527 {
 528         int r;
 529         struct pgpath *p;
 530
 531         /* we need at least a path arg */
 532         if (as->argc < 1) {
 533                 ti->error = "no device given";
 534                 return NULL;
 535         }
 536
 537         p = alloc_pgpath();
 538         if (!p)
 539                 return NULL;
 540
 541         r = dm_get_device(ti, shift(as), ti->begin, ti->len,
 542                           dm_table_get_mode(ti->table), &p->path.dev);
 543         if (r) {
 544                 ti->error = "error getting device";
 545                 goto bad;
 546         }
 547
 548         r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 549         if (r) {
 550                 dm_put_device(ti, p->path.dev);
 551                 goto bad;
 552         }
 553
 554         return p;
 555
 556  bad:
 557         free_pgpath(p);
 558         return NULL;
 559 }
 560
 561 static struct priority_group *parse_priority_group(struct arg_set *as,
 562                                                    struct multipath *m)
 563 {
 564         static struct param _params[] = {
 565                 {1, 1024, "invalid number of paths"},
 566                 {0, 1024, "invalid number of selector args"}
 567         };
 568
 569         int r;
 570         unsigned i, nr_selector_args, nr_params;
 571         struct priority_group *pg;
 572         struct dm_target *ti = m->ti;
 573
 574         if (as->argc < 2) {
 575                 as->argc = 0;
 576                 ti->error = "not enough priority group aruments";
 577                 return NULL;
 578         }
 579
 580         pg = alloc_priority_group();
 581         if (!pg) {
 582                 ti->error = "couldn't allocate priority group";
 583                 return NULL;
 584         }
 585         pg->m = m;
 586
 587         r = parse_path_selector(as, pg, ti);
 588         if (r)
 589                 goto bad;
 590
 591         /*
 592          * read the paths
 593          */
 594         r = read_param(_params, shift(as), &pg->nr_pgpaths, &ti->error);
 595         if (r)
 596                 goto bad;
 597
 598         r = read_param(_params + 1, shift(as), &nr_selector_args, &ti->error);
 599         if (r)
 600                 goto bad;
 601
 602         nr_params = 1 + nr_selector_args;
 603         for (i = 0; i < pg->nr_pgpaths; i++) {
 604                 struct pgpath *pgpath;
 605                 struct arg_set path_args;
 606
 607                 if (as->argc < nr_params)
 608                         goto bad;
 609
 610                 path_args.argc = nr_params;
 611                 path_args.argv = as->argv;
 612
 613                 pgpath = parse_path(&path_args, &pg->ps, ti);
 614                 if (!pgpath)
 615                         goto bad;
 616
 617                 pgpath->pg = pg;
 618                 list_add_tail(&pgpath->list, &pg->pgpaths);
 619                 consume(as, nr_params);
 620         }
 621
 622         return pg;
 623
 624  bad:
 625         free_priority_group(pg, ti);
 626         return NULL;
 627 }
 628
 629 static int parse_hw_handler(struct arg_set *as, struct multipath *m)
 630 {
 631         int r;
 632         struct hw_handler_type *hwht;
 633         unsigned hw_argc;
 634         struct dm_target *ti = m->ti;
 635
 636         static struct param _params[] = {
 637                 {0, 1024, "invalid number of hardware handler args"},
 638         };
 639
 640         r = read_param(_params, shift(as), &hw_argc, &ti->error);
 641         if (r)
 642                 return -EINVAL;
 643
 644         if (!hw_argc)
 645                 return 0;
 646
 647         hwht = dm_get_hw_handler(shift(as));
 648         if (!hwht) {
 649                 ti->error = "unknown hardware handler type";
 650                 return -EINVAL;
 651         }
 652
 653         r = hwht->create(&m->hw_handler, hw_argc - 1, as->argv);
 654         if (r) {
 655                 dm_put_hw_handler(hwht);
 656                 ti->error = "hardware handler constructor failed";
 657                 return r;
 658         }
 659
 660         m->hw_handler.type = hwht;
 661         consume(as, hw_argc - 1);
 662
 663         return 0;
 664 }
 665
 666 static int parse_features(struct arg_set *as, struct multipath *m)
 667 {
 668         int r;
 669         unsigned argc;
 670         struct dm_target *ti = m->ti;
 671
 672         static struct param _params[] = {
 673                 {0, 1, "invalid number of feature args"},
 674         };
 675
 676         r = read_param(_params, shift(as), &argc, &ti->error);
 677         if (r)
 678                 return -EINVAL;
 679
 680         if (!argc)
 681                 return 0;
 682
 683         if (!strnicmp(shift(as), MESG_STR("queue_if_no_path")))
 684                 return queue_if_no_path(m, 1, 0);
 685         else {
 686                 ti->error = "Unrecognised multipath feature request";
 687                 return -EINVAL;
 688         }
 689 }
 690
 691 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 692                          char **argv)
 693 {
 694         /* target parameters */
 695         static struct param _params[] = {
 696                 {1, 1024, "invalid number of priority groups"},
 697                 {1, 1024, "invalid initial priority group number"},
 698         };
 699
 700         int r;
 701         struct multipath *m;
 702         struct arg_set as;
 703         unsigned pg_count = 0;
 704         unsigned next_pg_num;
 705
 706         as.argc = argc;
 707         as.argv = argv;
 708
 709         m = alloc_multipath(ti);
 710         if (!m) {
 711                 ti->error = "can't allocate multipath";
 712                 return -EINVAL;
 713         }
 714
 715         r = parse_features(&as, m);
 716         if (r)
 717                 goto bad;
 718
 719         r = parse_hw_handler(&as, m);
 720         if (r)
 721                 goto bad;
 722
 723         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 724         if (r)
 725                 goto bad;
 726
 727         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 728         if (r)
 729                 goto bad;
 730
 731         /* parse the priority groups */
 732         while (as.argc) {
 733                 struct priority_group *pg;
 734
 735                 pg = parse_priority_group(&as, m);
 736                 if (!pg) {
 737                         r = -EINVAL;
 738                         goto bad;
 739                 }
 740
 741                 m->nr_valid_paths += pg->nr_pgpaths;
 742                 list_add_tail(&pg->list, &m->priority_groups);
 743                 pg_count++;
 744                 pg->pg_num = pg_count;
 745                 if (!--next_pg_num)
 746                         m->next_pg = pg;
 747         }
 748
 749         if (pg_count != m->nr_priority_groups) {
 750                 ti->error = "priority group count mismatch";
 751                 r = -EINVAL;
 752                 goto bad;
 753         }
 754
 755         return 0;
 756
 757  bad:
 758         free_multipath(m);
 759         return r;
 760 }
 761
 762 static void multipath_dtr(struct dm_target *ti)
 763 {
 764         struct multipath *m = (struct multipath *) ti->private;
 765
 766         flush_workqueue(kmultipathd);
 767         free_multipath(m);
 768 }
 769
 770 /*
 771  * Map bios, recording original fields for later in case we have to resubmit
 772  */
 773 static int multipath_map(struct dm_target *ti, struct bio *bio,
 774                          union map_info *map_context)
 775 {
 776         int r;
 777         struct mpath_io *mpio;
 778         struct multipath *m = (struct multipath *) ti->private;
 779
 780         if (bio_barrier(bio))
 781                 return -EOPNOTSUPP;
 782
 783         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 784         dm_bio_record(&mpio->details, bio);
 785
 786         map_context->ptr = mpio;
 787         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 788         r = map_io(m, bio, mpio, 0);
 789         if (r < 0)
 790                 mempool_free(mpio, m->mpio_pool);
 791
 792         return r;
 793 }
 794
 795 /*
 796  * Take a path out of use.
 797  */
 798 static int fail_path(struct pgpath *pgpath)
 799 {
 800         unsigned long flags;
 801         struct multipath *m = pgpath->pg->m;
 802
 803         spin_lock_irqsave(&m->lock, flags);
 804
 805         if (!pgpath->path.is_active)
 806                 goto out;
 807
 808         DMWARN("Failing path %s.", pgpath->path.dev->name);
 809
 810         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 811         pgpath->path.is_active = 0;
 812         pgpath->fail_count++;
 813
 814         m->nr_valid_paths--;
 815
 816         if (pgpath == m->current_pgpath)
 817                 m->current_pgpath = NULL;
 818
 819         queue_work(kmultipathd, &m->trigger_event);
 820
 821 out:
 822         spin_unlock_irqrestore(&m->lock, flags);
 823
 824         return 0;
 825 }
 826
 827 /*
 828  * Reinstate a previously-failed path
 829  */
 830 static int reinstate_path(struct pgpath *pgpath)
 831 {
 832         int r = 0;
 833         unsigned long flags;
 834         struct multipath *m = pgpath->pg->m;
 835
 836         spin_lock_irqsave(&m->lock, flags);
 837
 838         if (pgpath->path.is_active)
 839                 goto out;
 840
 841         if (!pgpath->pg->ps.type) {
 842                 DMWARN("Reinstate path not supported by path selector %s",
 843                        pgpath->pg->ps.type->name);
 844                 r = -EINVAL;
 845                 goto out;
 846         }
 847
 848         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 849         if (r)
 850                 goto out;
 851
 852         pgpath->path.is_active = 1;
 853
 854         m->current_pgpath = NULL;
 855         if (!m->nr_valid_paths++ && m->queue_size)
 856                 queue_work(kmultipathd, &m->process_queued_ios);
 857
 858         queue_work(kmultipathd, &m->trigger_event);
 859
 860 out:
 861         spin_unlock_irqrestore(&m->lock, flags);
 862
 863         return r;
 864 }
 865
 866 /*
 867  * Fail or reinstate all paths that match the provided struct dm_dev.
 868  */
 869 static int action_dev(struct multipath *m, struct dm_dev *dev,
 870                       action_fn action)
 871 {
 872         int r = 0;
 873         struct pgpath *pgpath;
 874         struct priority_group *pg;
 875
 876         list_for_each_entry(pg, &m->priority_groups, list) {
 877                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 878                         if (pgpath->path.dev == dev)
 879                                 r = action(pgpath);
 880                 }
 881         }
 882
 883         return r;
 884 }
 885
 886 /*
 887  * Temporarily try to avoid having to use the specified PG
 888  */
 889 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 890                       int bypassed)
 891 {
 892         unsigned long flags;
 893
 894         spin_lock_irqsave(&m->lock, flags);
 895
 896         pg->bypassed = bypassed;
 897         m->current_pgpath = NULL;
 898         m->current_pg = NULL;
 899
 900         spin_unlock_irqrestore(&m->lock, flags);
 901
 902         queue_work(kmultipathd, &m->trigger_event);
 903 }
 904
 905 /*
 906  * Switch to using the specified PG from the next I/O that gets mapped
 907  */
 908 static int switch_pg_num(struct multipath *m, const char *pgstr)
 909 {
 910         struct priority_group *pg;
 911         unsigned pgnum;
 912         unsigned long flags;
 913
 914         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 915             (pgnum > m->nr_priority_groups)) {
 916                 DMWARN("invalid PG number supplied to switch_pg_num");
 917                 return -EINVAL;
 918         }
 919
 920         spin_lock_irqsave(&m->lock, flags);
 921         list_for_each_entry(pg, &m->priority_groups, list) {
 922                 pg->bypassed = 0;
 923                 if (--pgnum)
 924                         continue;
 925
 926                 m->current_pgpath = NULL;
 927                 m->current_pg = NULL;
 928                 m->next_pg = pg;
 929         }
 930         spin_unlock_irqrestore(&m->lock, flags);
 931
 932         queue_work(kmultipathd, &m->trigger_event);
 933         return 0;
 934 }
 935
 936 /*
 937  * Set/clear bypassed status of a PG.
 938  * PGs are numbered upwards from 1 in the order they were declared.
 939  */
 940 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 941 {
 942         struct priority_group *pg;
 943         unsigned pgnum;
 944
 945         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 946             (pgnum > m->nr_priority_groups)) {
 947                 DMWARN("invalid PG number supplied to bypass_pg");
 948                 return -EINVAL;
 949         }
 950
 951         list_for_each_entry(pg, &m->priority_groups, list) {
 952                 if (!--pgnum)
 953                         break;
 954         }
 955
 956         bypass_pg(m, pg, bypassed);
 957         return 0;
 958 }
 959
 960 /*
 961  * pg_init must call this when it has completed its initialisation
 962  */
 963 void dm_pg_init_complete(struct path *path, unsigned err_flags)
 964 {
 965         struct pgpath *pgpath = path_to_pgpath(path);
 966         struct priority_group *pg = pgpath->pg;
 967         struct multipath *m = pg->m;
 968         unsigned long flags;
 969
 970         /* We insist on failing the path if the PG is already bypassed. */
 971         if (err_flags && pg->bypassed)
 972                 err_flags |= MP_FAIL_PATH;
 973
 974         if (err_flags & MP_FAIL_PATH)
 975                 fail_path(pgpath);
 976
 977         if (err_flags & MP_BYPASS_PG)
 978                 bypass_pg(m, pg, 1);
 979
 980         spin_lock_irqsave(&m->lock, flags);
 981         if (err_flags) {
 982                 m->current_pgpath = NULL;
 983                 m->current_pg = NULL;
 984         } else if (!m->pg_init_required)
 985                 m->queue_io = 0;
 986
 987         m->pg_init_in_progress = 0;
 988         queue_work(kmultipathd, &m->process_queued_ios);
 989         spin_unlock_irqrestore(&m->lock, flags);
 990 }
 991
 992 /*
 993  * end_io handling
 994  */
 995 static int do_end_io(struct multipath *m, struct bio *bio,
 996                      int error, struct mpath_io *mpio)
 997 {
 998         struct hw_handler *hwh = &m->hw_handler;
 999         unsigned err_flags = MP_FAIL_PATH;      /* Default behavior */
1000         unsigned long flags;
1001
1002         if (!error)
1003                 return 0;       /* I/O complete */
1004
1005         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1006                 return error;
1007
1008         if (error == -EOPNOTSUPP)
1009                 return error;
1010
1011         spin_lock_irqsave(&m->lock, flags);
1012         if (!m->nr_valid_paths) {
1013                 if (!m->queue_if_no_path) {
1014                         spin_unlock_irqrestore(&m->lock, flags);
1015                         return -EIO;
1016                 } else {
1017                         spin_unlock_irqrestore(&m->lock, flags);
1018                         goto requeue;
1019                 }
1020         }
1021         spin_unlock_irqrestore(&m->lock, flags);
1022
1023         if (hwh->type && hwh->type->error)
1024                 err_flags = hwh->type->error(hwh, bio);
1025
1026         if (mpio->pgpath) {
1027                 if (err_flags & MP_FAIL_PATH)
1028                         fail_path(mpio->pgpath);
1029
1030                 if (err_flags & MP_BYPASS_PG)
1031                         bypass_pg(m, mpio->pgpath->pg, 1);
1032         }
1033
1034         if (err_flags & MP_ERROR_IO)
1035                 return -EIO;
1036
1037       requeue:
1038         dm_bio_restore(&mpio->details, bio);
1039
1040         /* queue for the daemon to resubmit or fail */
1041         spin_lock_irqsave(&m->lock, flags);
1042         bio_list_add(&m->queued_ios, bio);
1043         m->queue_size++;
1044         if (!m->queue_io)
1045                 queue_work(kmultipathd, &m->process_queued_ios);
1046         spin_unlock_irqrestore(&m->lock, flags);
1047
1048         return 1;       /* io not complete */
1049 }
1050
1051 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1052                             int error, union map_info *map_context)
1053 {
1054         struct multipath *m = (struct multipath *) ti->private;
1055         struct mpath_io *mpio = (struct mpath_io *) map_context->ptr;
1056         struct pgpath *pgpath = mpio->pgpath;
1057         struct path_selector *ps;
1058         int r;
1059
1060         r  = do_end_io(m, bio, error, mpio);
1061         if (pgpath) {
1062                 ps = &pgpath->pg->ps;
1063                 if (ps->type->end_io)
1064                         ps->type->end_io(ps, &pgpath->path);
1065         }
1066         if (r <= 0)
1067                 mempool_free(mpio, m->mpio_pool);
1068
1069         return r;
1070 }
1071
1072 /*
1073  * Suspend can't complete until all the I/O is processed so if
1074  * the last path fails we must error any remaining I/O.
1075  * Note that if the freeze_bdev fails while suspending, the
1076  * queue_if_no_path state is lost - userspace should reset it.
1077  */
1078 static void multipath_presuspend(struct dm_target *ti)
1079 {
1080         struct multipath *m = (struct multipath *) ti->private;
1081
1082         queue_if_no_path(m, 0, 1);
1083 }
1084
1085 /*
1086  * Restore the queue_if_no_path setting.
1087  */
1088 static void multipath_resume(struct dm_target *ti)
1089 {
1090         struct multipath *m = (struct multipath *) ti->private;
1091         unsigned long flags;
1092
1093         spin_lock_irqsave(&m->lock, flags);
1094         m->queue_if_no_path = m->saved_queue_if_no_path;
1095         spin_unlock_irqrestore(&m->lock, flags);
1096 }
1097
1098 /*
1099  * Info output has the following format:
1100  * num_multipath_feature_args [multipath_feature_args]*
1101  * num_handler_status_args [handler_status_args]*
1102  * num_groups init_group_number
1103  *            [A|D|E num_ps_status_args [ps_status_args]*
1104  *             num_paths num_selector_args
1105  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1106  *
1107  * Table output has the following format (identical to the constructor string):
1108  * num_feature_args [features_args]*
1109  * num_handler_args hw_handler [hw_handler_args]*
1110  * num_groups init_group_number
1111  *     [priority selector-name num_ps_args [ps_args]*
1112  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1113  */
1114 static int multipath_status(struct dm_target *ti, status_type_t type,
1115                             char *result, unsigned int maxlen)
1116 {
1117         int sz = 0;
1118         unsigned long flags;
1119         struct multipath *m = (struct multipath *) ti->private;
1120         struct hw_handler *hwh = &m->hw_handler;
1121         struct priority_group *pg;
1122         struct pgpath *p;
1123         unsigned pg_num;
1124         char state;
1125
1126         spin_lock_irqsave(&m->lock, flags);
1127
1128         /* Features */
1129         if (type == STATUSTYPE_INFO)
1130                 DMEMIT("1 %u ", m->queue_size);
1131         else if (m->queue_if_no_path)
1132                 DMEMIT("1 queue_if_no_path ");
1133         else
1134                 DMEMIT("0 ");
1135
1136         if (hwh->type && hwh->type->status)
1137                 sz += hwh->type->status(hwh, type, result + sz, maxlen - sz);
1138         else if (!hwh->type || type == STATUSTYPE_INFO)
1139                 DMEMIT("0 ");
1140         else
1141                 DMEMIT("1 %s ", hwh->type->name);
1142
1143         DMEMIT("%u ", m->nr_priority_groups);
1144
1145         if (m->next_pg)
1146                 pg_num = m->next_pg->pg_num;
1147         else if (m->current_pg)
1148                 pg_num = m->current_pg->pg_num;
1149         else
1150                         pg_num = 1;
1151
1152         DMEMIT("%u ", pg_num);
1153
1154         switch (type) {
1155         case STATUSTYPE_INFO:
1156                 list_for_each_entry(pg, &m->priority_groups, list) {
1157                         if (pg->bypassed)
1158                                 state = 'D';    /* Disabled */
1159                         else if (pg == m->current_pg)
1160                                 state = 'A';    /* Currently Active */
1161                         else
1162                                 state = 'E';    /* Enabled */
1163
1164                         DMEMIT("%c ", state);
1165
1166                         if (pg->ps.type->status)
1167                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1168                                                           result + sz,
1169                                                           maxlen - sz);
1170                         else
1171                                 DMEMIT("0 ");
1172
1173                         DMEMIT("%u %u ", pg->nr_pgpaths,
1174                                pg->ps.type->info_args);
1175
1176                         list_for_each_entry(p, &pg->pgpaths, list) {
1177                                 DMEMIT("%s %s %u ", p->path.dev->name,
1178                                        p->path.is_active ? "A" : "F",
1179                                        p->fail_count);
1180                                 if (pg->ps.type->status)
1181                                         sz += pg->ps.type->status(&pg->ps,
1182                                               &p->path, type, result + sz,
1183                                               maxlen - sz);
1184                         }
1185                 }
1186                 break;
1187
1188         case STATUSTYPE_TABLE:
1189                 list_for_each_entry(pg, &m->priority_groups, list) {
1190                         DMEMIT("%s ", pg->ps.type->name);
1191
1192                         if (pg->ps.type->status)
1193                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1194                                                           result + sz,
1195                                                           maxlen - sz);
1196                         else
1197                                 DMEMIT("0 ");
1198
1199                         DMEMIT("%u %u ", pg->nr_pgpaths,
1200                                pg->ps.type->table_args);
1201
1202                         list_for_each_entry(p, &pg->pgpaths, list) {
1203                                 DMEMIT("%s ", p->path.dev->name);
1204                                 if (pg->ps.type->status)
1205                                         sz += pg->ps.type->status(&pg->ps,
1206                                               &p->path, type, result + sz,
1207                                               maxlen - sz);
1208                         }
1209                 }
1210                 break;
1211         }
1212
1213         spin_unlock_irqrestore(&m->lock, flags);
1214
1215         return 0;
1216 }
1217
1218 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1219 {
1220         int r;
1221         struct dm_dev *dev;
1222         struct multipath *m = (struct multipath *) ti->private;
1223         action_fn action;
1224
1225         if (argc == 1) {
1226                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1227                         return queue_if_no_path(m, 1, 0);
1228                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1229                         return queue_if_no_path(m, 0, 0);
1230         }
1231
1232         if (argc != 2)
1233                 goto error;
1234
1235         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1236                 return bypass_pg_num(m, argv[1], 1);
1237         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1238                 return bypass_pg_num(m, argv[1], 0);
1239         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1240                 return switch_pg_num(m, argv[1]);
1241         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1242                 action = reinstate_path;
1243         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1244                 action = fail_path;
1245         else
1246                 goto error;
1247
1248         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1249                           dm_table_get_mode(ti->table), &dev);
1250         if (r) {
1251                 DMWARN("message: error getting device %s",
1252                        argv[1]);
1253                 return -EINVAL;
1254         }
1255
1256         r = action_dev(m, dev, action);
1257
1258         dm_put_device(ti, dev);
1259
1260         return r;
1261
1262 error:
1263         DMWARN("Unrecognised multipath message received.");
1264         return -EINVAL;
1265 }
1266
1267 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1268                            struct file *filp, unsigned int cmd,
1269                            unsigned long arg)
1270 {
1271         struct multipath *m = (struct multipath *) ti->private;
1272         struct block_device *bdev = NULL;
1273         unsigned long flags;
1274         struct file fake_file = {};
1275         struct dentry fake_dentry = {};
1276         int r = 0;
1277
1278         fake_file.f_dentry = &fake_dentry;
1279
1280         spin_lock_irqsave(&m->lock, flags);
1281
1282         if (!m->current_pgpath)
1283                 __choose_pgpath(m);
1284
1285         if (m->current_pgpath) {
1286                 bdev = m->current_pgpath->path.dev->bdev;
1287                 fake_dentry.d_inode = bdev->bd_inode;
1288                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1289         }
1290
1291         if (m->queue_io)
1292                 r = -EAGAIN;
1293         else if (!bdev)
1294                 r = -EIO;
1295
1296         spin_unlock_irqrestore(&m->lock, flags);
1297
1298         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1299                                          bdev->bd_disk, cmd, arg);
1300 }
1301
1302 /*-----------------------------------------------------------------
1303  * Module setup
1304  *---------------------------------------------------------------*/
1305 static struct target_type multipath_target = {
1306         .name = "multipath",
1307         .version = {1, 0, 5},
1308         .module = THIS_MODULE,
1309         .ctr = multipath_ctr,
1310         .dtr = multipath_dtr,
1311         .map = multipath_map,
1312         .end_io = multipath_end_io,
1313         .presuspend = multipath_presuspend,
1314         .resume = multipath_resume,
1315         .status = multipath_status,
1316         .message = multipath_message,
1317         .ioctl  = multipath_ioctl,
1318 };
1319
1320 static int __init dm_multipath_init(void)
1321 {
1322         int r;
1323
1324         /* allocate a slab for the dm_ios */
1325         _mpio_cache = kmem_cache_create("dm_mpath", sizeof(struct mpath_io),
1326                                         0, 0, NULL, NULL);
1327         if (!_mpio_cache)
1328                 return -ENOMEM;
1329
1330         r = dm_register_target(&multipath_target);
1331         if (r < 0) {
1332                 DMERR("%s: register failed %d", multipath_target.name, r);
1333                 kmem_cache_destroy(_mpio_cache);
1334                 return -EINVAL;
1335         }
1336
1337         kmultipathd = create_workqueue("kmpathd");
1338         if (!kmultipathd) {
1339                 DMERR("%s: failed to create workqueue kmpathd",
1340                                 multipath_target.name);
1341                 dm_unregister_target(&multipath_target);
1342                 kmem_cache_destroy(_mpio_cache);
1343                 return -ENOMEM;
1344         }
1345
1346         DMINFO("version %u.%u.%u loaded",
1347                multipath_target.version[0], multipath_target.version[1],
1348                multipath_target.version[2]);
1349
1350         return r;
1351 }
1352
1353 static void __exit dm_multipath_exit(void)
1354 {
1355         int r;
1356
1357         destroy_workqueue(kmultipathd);
1358
1359         r = dm_unregister_target(&multipath_target);
1360         if (r < 0)
1361                 DMERR("%s: target unregister failed %d",
1362                       multipath_target.name, r);
1363         kmem_cache_destroy(_mpio_cache);
1364 }
1365
1366 EXPORT_SYMBOL_GPL(dm_pg_init_complete);
1367
1368 module_init(dm_multipath_init);
1369 module_exit(dm_multipath_exit);
1370
1371 MODULE_DESCRIPTION(DM_NAME " multipath target");
1372 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1373 MODULE_LICENSE("GPL");