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                 return -EINVAL;
 668         }
 669         consume(as, hw_argc - 1);
 670
 671         return 0;
 672 }
 673
 674 static int parse_features(struct arg_set *as, struct multipath *m)
 675 {
 676         int r;
 677         unsigned argc;
 678         struct dm_target *ti = m->ti;
 679         const char *param_name;
 680
 681         static struct param _params[] = {
 682                 {0, 3, "invalid number of feature args"},
 683                 {1, 50, "pg_init_retries must be between 1 and 50"},
 684         };
 685
 686         r = read_param(_params, shift(as), &argc, &ti->error);
 687         if (r)
 688                 return -EINVAL;
 689
 690         if (!argc)
 691                 return 0;
 692
 693         do {
 694                 param_name = shift(as);
 695                 argc--;
 696
 697                 if (!strnicmp(param_name, MESG_STR("queue_if_no_path"))) {
 698                         r = queue_if_no_path(m, 1, 0);
 699                         continue;
 700                 }
 701
 702                 if (!strnicmp(param_name, MESG_STR("pg_init_retries")) &&
 703                     (argc >= 1)) {
 704                         r = read_param(_params + 1, shift(as),
 705                                        &m->pg_init_retries, &ti->error);
 706                         argc--;
 707                         continue;
 708                 }
 709
 710                 ti->error = "Unrecognised multipath feature request";
 711                 r = -EINVAL;
 712         } while (argc && !r);
 713
 714         return r;
 715 }
 716
 717 static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 718                          char **argv)
 719 {
 720         /* target parameters */
 721         static struct param _params[] = {
 722                 {1, 1024, "invalid number of priority groups"},
 723                 {1, 1024, "invalid initial priority group number"},
 724         };
 725
 726         int r;
 727         struct multipath *m;
 728         struct arg_set as;
 729         unsigned pg_count = 0;
 730         unsigned next_pg_num;
 731
 732         as.argc = argc;
 733         as.argv = argv;
 734
 735         m = alloc_multipath(ti);
 736         if (!m) {
 737                 ti->error = "can't allocate multipath";
 738                 return -EINVAL;
 739         }
 740
 741         r = parse_features(&as, m);
 742         if (r)
 743                 goto bad;
 744
 745         r = parse_hw_handler(&as, m);
 746         if (r)
 747                 goto bad;
 748
 749         r = read_param(_params, shift(&as), &m->nr_priority_groups, &ti->error);
 750         if (r)
 751                 goto bad;
 752
 753         r = read_param(_params + 1, shift(&as), &next_pg_num, &ti->error);
 754         if (r)
 755                 goto bad;
 756
 757         /* parse the priority groups */
 758         while (as.argc) {
 759                 struct priority_group *pg;
 760
 761                 pg = parse_priority_group(&as, m);
 762                 if (!pg) {
 763                         r = -EINVAL;
 764                         goto bad;
 765                 }
 766
 767                 m->nr_valid_paths += pg->nr_pgpaths;
 768                 list_add_tail(&pg->list, &m->priority_groups);
 769                 pg_count++;
 770                 pg->pg_num = pg_count;
 771                 if (!--next_pg_num)
 772                         m->next_pg = pg;
 773         }
 774
 775         if (pg_count != m->nr_priority_groups) {
 776                 ti->error = "priority group count mismatch";
 777                 r = -EINVAL;
 778                 goto bad;
 779         }
 780
 781         return 0;
 782
 783  bad:
 784         free_multipath(m);
 785         return r;
 786 }
 787
 788 static void multipath_dtr(struct dm_target *ti)
 789 {
 790         struct multipath *m = (struct multipath *) ti->private;
 791
 792         flush_workqueue(kmpath_handlerd);
 793         flush_workqueue(kmultipathd);
 794         free_multipath(m);
 795 }
 796
 797 /*
 798  * Map bios, recording original fields for later in case we have to resubmit
 799  */
 800 static int multipath_map(struct dm_target *ti, struct bio *bio,
 801                          union map_info *map_context)
 802 {
 803         int r;
 804         struct dm_mpath_io *mpio;
 805         struct multipath *m = (struct multipath *) ti->private;
 806
 807         mpio = mempool_alloc(m->mpio_pool, GFP_NOIO);
 808         dm_bio_record(&mpio->details, bio);
 809
 810         map_context->ptr = mpio;
 811         bio->bi_rw |= (1 << BIO_RW_FAILFAST);
 812         r = map_io(m, bio, mpio, 0);
 813         if (r < 0 || r == DM_MAPIO_REQUEUE)
 814                 mempool_free(mpio, m->mpio_pool);
 815
 816         return r;
 817 }
 818
 819 /*
 820  * Take a path out of use.
 821  */
 822 static int fail_path(struct pgpath *pgpath)
 823 {
 824         unsigned long flags;
 825         struct multipath *m = pgpath->pg->m;
 826
 827         spin_lock_irqsave(&m->lock, flags);
 828
 829         if (!pgpath->path.is_active)
 830                 goto out;
 831
 832         DMWARN("Failing path %s.", pgpath->path.dev->name);
 833
 834         pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 835         pgpath->path.is_active = 0;
 836         pgpath->fail_count++;
 837
 838         m->nr_valid_paths--;
 839
 840         if (pgpath == m->current_pgpath)
 841                 m->current_pgpath = NULL;
 842
 843         dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 844                       pgpath->path.dev->name, m->nr_valid_paths);
 845
 846         queue_work(kmultipathd, &m->trigger_event);
 847
 848 out:
 849         spin_unlock_irqrestore(&m->lock, flags);
 850
 851         return 0;
 852 }
 853
 854 /*
 855  * Reinstate a previously-failed path
 856  */
 857 static int reinstate_path(struct pgpath *pgpath)
 858 {
 859         int r = 0;
 860         unsigned long flags;
 861         struct multipath *m = pgpath->pg->m;
 862
 863         spin_lock_irqsave(&m->lock, flags);
 864
 865         if (pgpath->path.is_active)
 866                 goto out;
 867
 868         if (!pgpath->pg->ps.type) {
 869                 DMWARN("Reinstate path not supported by path selector %s",
 870                        pgpath->pg->ps.type->name);
 871                 r = -EINVAL;
 872                 goto out;
 873         }
 874
 875         r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 876         if (r)
 877                 goto out;
 878
 879         pgpath->path.is_active = 1;
 880
 881         m->current_pgpath = NULL;
 882         if (!m->nr_valid_paths++ && m->queue_size)
 883                 queue_work(kmultipathd, &m->process_queued_ios);
 884
 885         dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
 886                       pgpath->path.dev->name, m->nr_valid_paths);
 887
 888         queue_work(kmultipathd, &m->trigger_event);
 889
 890 out:
 891         spin_unlock_irqrestore(&m->lock, flags);
 892
 893         return r;
 894 }
 895
 896 /*
 897  * Fail or reinstate all paths that match the provided struct dm_dev.
 898  */
 899 static int action_dev(struct multipath *m, struct dm_dev *dev,
 900                       action_fn action)
 901 {
 902         int r = 0;
 903         struct pgpath *pgpath;
 904         struct priority_group *pg;
 905
 906         list_for_each_entry(pg, &m->priority_groups, list) {
 907                 list_for_each_entry(pgpath, &pg->pgpaths, list) {
 908                         if (pgpath->path.dev == dev)
 909                                 r = action(pgpath);
 910                 }
 911         }
 912
 913         return r;
 914 }
 915
 916 /*
 917  * Temporarily try to avoid having to use the specified PG
 918  */
 919 static void bypass_pg(struct multipath *m, struct priority_group *pg,
 920                       int bypassed)
 921 {
 922         unsigned long flags;
 923
 924         spin_lock_irqsave(&m->lock, flags);
 925
 926         pg->bypassed = bypassed;
 927         m->current_pgpath = NULL;
 928         m->current_pg = NULL;
 929
 930         spin_unlock_irqrestore(&m->lock, flags);
 931
 932         queue_work(kmultipathd, &m->trigger_event);
 933 }
 934
 935 /*
 936  * Switch to using the specified PG from the next I/O that gets mapped
 937  */
 938 static int switch_pg_num(struct multipath *m, const char *pgstr)
 939 {
 940         struct priority_group *pg;
 941         unsigned pgnum;
 942         unsigned long flags;
 943
 944         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 945             (pgnum > m->nr_priority_groups)) {
 946                 DMWARN("invalid PG number supplied to switch_pg_num");
 947                 return -EINVAL;
 948         }
 949
 950         spin_lock_irqsave(&m->lock, flags);
 951         list_for_each_entry(pg, &m->priority_groups, list) {
 952                 pg->bypassed = 0;
 953                 if (--pgnum)
 954                         continue;
 955
 956                 m->current_pgpath = NULL;
 957                 m->current_pg = NULL;
 958                 m->next_pg = pg;
 959         }
 960         spin_unlock_irqrestore(&m->lock, flags);
 961
 962         queue_work(kmultipathd, &m->trigger_event);
 963         return 0;
 964 }
 965
 966 /*
 967  * Set/clear bypassed status of a PG.
 968  * PGs are numbered upwards from 1 in the order they were declared.
 969  */
 970 static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
 971 {
 972         struct priority_group *pg;
 973         unsigned pgnum;
 974
 975         if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
 976             (pgnum > m->nr_priority_groups)) {
 977                 DMWARN("invalid PG number supplied to bypass_pg");
 978                 return -EINVAL;
 979         }
 980
 981         list_for_each_entry(pg, &m->priority_groups, list) {
 982                 if (!--pgnum)
 983                         break;
 984         }
 985
 986         bypass_pg(m, pg, bypassed);
 987         return 0;
 988 }
 989
 990 /*
 991  * Should we retry pg_init immediately?
 992  */
 993 static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
 994 {
 995         unsigned long flags;
 996         int limit_reached = 0;
 997
 998         spin_lock_irqsave(&m->lock, flags);
 999
1000         if (m->pg_init_count <= m->pg_init_retries)
1001                 m->pg_init_required = 1;
1002         else
1003                 limit_reached = 1;
1004
1005         spin_unlock_irqrestore(&m->lock, flags);
1006
1007         return limit_reached;
1008 }
1009
1010 static void pg_init_done(struct dm_path *path, int errors)
1011 {
1012         struct pgpath *pgpath = path_to_pgpath(path);
1013         struct priority_group *pg = pgpath->pg;
1014         struct multipath *m = pg->m;
1015         unsigned long flags;
1016
1017         /* device or driver problems */
1018         switch (errors) {
1019         case SCSI_DH_OK:
1020                 break;
1021         case SCSI_DH_NOSYS:
1022                 if (!m->hw_handler_name) {
1023                         errors = 0;
1024                         break;
1025                 }
1026                 DMERR("Cannot failover device because scsi_dh_%s was not "
1027                       "loaded.", m->hw_handler_name);
1028                 /*
1029                  * Fail path for now, so we do not ping pong
1030                  */
1031                 fail_path(pgpath);
1032                 break;
1033         case SCSI_DH_DEV_TEMP_BUSY:
1034                 /*
1035                  * Probably doing something like FW upgrade on the
1036                  * controller so try the other pg.
1037                  */
1038                 bypass_pg(m, pg, 1);
1039                 break;
1040         /* TODO: For SCSI_DH_RETRY we should wait a couple seconds */
1041         case SCSI_DH_RETRY:
1042         case SCSI_DH_IMM_RETRY:
1043         case SCSI_DH_RES_TEMP_UNAVAIL:
1044                 if (pg_init_limit_reached(m, pgpath))
1045                         fail_path(pgpath);
1046                 errors = 0;
1047                 break;
1048         default:
1049                 /*
1050                  * We probably do not want to fail the path for a device
1051                  * error, but this is what the old dm did. In future
1052                  * patches we can do more advanced handling.
1053                  */
1054                 fail_path(pgpath);
1055         }
1056
1057         spin_lock_irqsave(&m->lock, flags);
1058         if (errors) {
1059                 DMERR("Could not failover device. Error %d.", errors);
1060                 m->current_pgpath = NULL;
1061                 m->current_pg = NULL;
1062         } else if (!m->pg_init_required) {
1063                 m->queue_io = 0;
1064                 pg->bypassed = 0;
1065         }
1066
1067         m->pg_init_in_progress = 0;
1068         queue_work(kmultipathd, &m->process_queued_ios);
1069         spin_unlock_irqrestore(&m->lock, flags);
1070 }
1071
1072 static void activate_path(struct work_struct *work)
1073 {
1074         int ret;
1075         struct multipath *m =
1076                 container_of(work, struct multipath, activate_path);
1077         struct dm_path *path = &m->current_pgpath->path;
1078
1079         ret = scsi_dh_activate(bdev_get_queue(path->dev->bdev));
1080         pg_init_done(path, ret);
1081 }
1082
1083 /*
1084  * end_io handling
1085  */
1086 static int do_end_io(struct multipath *m, struct bio *bio,
1087                      int error, struct dm_mpath_io *mpio)
1088 {
1089         unsigned long flags;
1090
1091         if (!error)
1092                 return 0;       /* I/O complete */
1093
1094         if ((error == -EWOULDBLOCK) && bio_rw_ahead(bio))
1095                 return error;
1096
1097         if (error == -EOPNOTSUPP)
1098                 return error;
1099
1100         spin_lock_irqsave(&m->lock, flags);
1101         if (!m->nr_valid_paths) {
1102                 if (__must_push_back(m)) {
1103                         spin_unlock_irqrestore(&m->lock, flags);
1104                         return DM_ENDIO_REQUEUE;
1105                 } else if (!m->queue_if_no_path) {
1106                         spin_unlock_irqrestore(&m->lock, flags);
1107                         return -EIO;
1108                 } else {
1109                         spin_unlock_irqrestore(&m->lock, flags);
1110                         goto requeue;
1111                 }
1112         }
1113         spin_unlock_irqrestore(&m->lock, flags);
1114
1115         if (mpio->pgpath)
1116                 fail_path(mpio->pgpath);
1117
1118       requeue:
1119         dm_bio_restore(&mpio->details, bio);
1120
1121         /* queue for the daemon to resubmit or fail */
1122         spin_lock_irqsave(&m->lock, flags);
1123         bio_list_add(&m->queued_ios, bio);
1124         m->queue_size++;
1125         if (!m->queue_io)
1126                 queue_work(kmultipathd, &m->process_queued_ios);
1127         spin_unlock_irqrestore(&m->lock, flags);
1128
1129         return DM_ENDIO_INCOMPLETE;     /* io not complete */
1130 }
1131
1132 static int multipath_end_io(struct dm_target *ti, struct bio *bio,
1133                             int error, union map_info *map_context)
1134 {
1135         struct multipath *m = ti->private;
1136         struct dm_mpath_io *mpio = map_context->ptr;
1137         struct pgpath *pgpath = mpio->pgpath;
1138         struct path_selector *ps;
1139         int r;
1140
1141         r  = do_end_io(m, bio, error, mpio);
1142         if (pgpath) {
1143                 ps = &pgpath->pg->ps;
1144                 if (ps->type->end_io)
1145                         ps->type->end_io(ps, &pgpath->path);
1146         }
1147         if (r != DM_ENDIO_INCOMPLETE)
1148                 mempool_free(mpio, m->mpio_pool);
1149
1150         return r;
1151 }
1152
1153 /*
1154  * Suspend can't complete until all the I/O is processed so if
1155  * the last path fails we must error any remaining I/O.
1156  * Note that if the freeze_bdev fails while suspending, the
1157  * queue_if_no_path state is lost - userspace should reset it.
1158  */
1159 static void multipath_presuspend(struct dm_target *ti)
1160 {
1161         struct multipath *m = (struct multipath *) ti->private;
1162
1163         queue_if_no_path(m, 0, 1);
1164 }
1165
1166 /*
1167  * Restore the queue_if_no_path setting.
1168  */
1169 static void multipath_resume(struct dm_target *ti)
1170 {
1171         struct multipath *m = (struct multipath *) ti->private;
1172         unsigned long flags;
1173
1174         spin_lock_irqsave(&m->lock, flags);
1175         m->queue_if_no_path = m->saved_queue_if_no_path;
1176         spin_unlock_irqrestore(&m->lock, flags);
1177 }
1178
1179 /*
1180  * Info output has the following format:
1181  * num_multipath_feature_args [multipath_feature_args]*
1182  * num_handler_status_args [handler_status_args]*
1183  * num_groups init_group_number
1184  *            [A|D|E num_ps_status_args [ps_status_args]*
1185  *             num_paths num_selector_args
1186  *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1187  *
1188  * Table output has the following format (identical to the constructor string):
1189  * num_feature_args [features_args]*
1190  * num_handler_args hw_handler [hw_handler_args]*
1191  * num_groups init_group_number
1192  *     [priority selector-name num_ps_args [ps_args]*
1193  *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1194  */
1195 static int multipath_status(struct dm_target *ti, status_type_t type,
1196                             char *result, unsigned int maxlen)
1197 {
1198         int sz = 0;
1199         unsigned long flags;
1200         struct multipath *m = (struct multipath *) ti->private;
1201         struct priority_group *pg;
1202         struct pgpath *p;
1203         unsigned pg_num;
1204         char state;
1205
1206         spin_lock_irqsave(&m->lock, flags);
1207
1208         /* Features */
1209         if (type == STATUSTYPE_INFO)
1210                 DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
1211         else {
1212                 DMEMIT("%u ", m->queue_if_no_path +
1213                               (m->pg_init_retries > 0) * 2);
1214                 if (m->queue_if_no_path)
1215                         DMEMIT("queue_if_no_path ");
1216                 if (m->pg_init_retries)
1217                         DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1218         }
1219
1220         if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1221                 DMEMIT("0 ");
1222         else
1223                 DMEMIT("1 %s ", m->hw_handler_name);
1224
1225         DMEMIT("%u ", m->nr_priority_groups);
1226
1227         if (m->next_pg)
1228                 pg_num = m->next_pg->pg_num;
1229         else if (m->current_pg)
1230                 pg_num = m->current_pg->pg_num;
1231         else
1232                         pg_num = 1;
1233
1234         DMEMIT("%u ", pg_num);
1235
1236         switch (type) {
1237         case STATUSTYPE_INFO:
1238                 list_for_each_entry(pg, &m->priority_groups, list) {
1239                         if (pg->bypassed)
1240                                 state = 'D';    /* Disabled */
1241                         else if (pg == m->current_pg)
1242                                 state = 'A';    /* Currently Active */
1243                         else
1244                                 state = 'E';    /* Enabled */
1245
1246                         DMEMIT("%c ", state);
1247
1248                         if (pg->ps.type->status)
1249                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1250                                                           result + sz,
1251                                                           maxlen - sz);
1252                         else
1253                                 DMEMIT("0 ");
1254
1255                         DMEMIT("%u %u ", pg->nr_pgpaths,
1256                                pg->ps.type->info_args);
1257
1258                         list_for_each_entry(p, &pg->pgpaths, list) {
1259                                 DMEMIT("%s %s %u ", p->path.dev->name,
1260                                        p->path.is_active ? "A" : "F",
1261                                        p->fail_count);
1262                                 if (pg->ps.type->status)
1263                                         sz += pg->ps.type->status(&pg->ps,
1264                                               &p->path, type, result + sz,
1265                                               maxlen - sz);
1266                         }
1267                 }
1268                 break;
1269
1270         case STATUSTYPE_TABLE:
1271                 list_for_each_entry(pg, &m->priority_groups, list) {
1272                         DMEMIT("%s ", pg->ps.type->name);
1273
1274                         if (pg->ps.type->status)
1275                                 sz += pg->ps.type->status(&pg->ps, NULL, type,
1276                                                           result + sz,
1277                                                           maxlen - sz);
1278                         else
1279                                 DMEMIT("0 ");
1280
1281                         DMEMIT("%u %u ", pg->nr_pgpaths,
1282                                pg->ps.type->table_args);
1283
1284                         list_for_each_entry(p, &pg->pgpaths, list) {
1285                                 DMEMIT("%s ", p->path.dev->name);
1286                                 if (pg->ps.type->status)
1287                                         sz += pg->ps.type->status(&pg->ps,
1288                                               &p->path, type, result + sz,
1289                                               maxlen - sz);
1290                         }
1291                 }
1292                 break;
1293         }
1294
1295         spin_unlock_irqrestore(&m->lock, flags);
1296
1297         return 0;
1298 }
1299
1300 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
1301 {
1302         int r;
1303         struct dm_dev *dev;
1304         struct multipath *m = (struct multipath *) ti->private;
1305         action_fn action;
1306
1307         if (argc == 1) {
1308                 if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
1309                         return queue_if_no_path(m, 1, 0);
1310                 else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
1311                         return queue_if_no_path(m, 0, 0);
1312         }
1313
1314         if (argc != 2)
1315                 goto error;
1316
1317         if (!strnicmp(argv[0], MESG_STR("disable_group")))
1318                 return bypass_pg_num(m, argv[1], 1);
1319         else if (!strnicmp(argv[0], MESG_STR("enable_group")))
1320                 return bypass_pg_num(m, argv[1], 0);
1321         else if (!strnicmp(argv[0], MESG_STR("switch_group")))
1322                 return switch_pg_num(m, argv[1]);
1323         else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
1324                 action = reinstate_path;
1325         else if (!strnicmp(argv[0], MESG_STR("fail_path")))
1326                 action = fail_path;
1327         else
1328                 goto error;
1329
1330         r = dm_get_device(ti, argv[1], ti->begin, ti->len,
1331                           dm_table_get_mode(ti->table), &dev);
1332         if (r) {
1333                 DMWARN("message: error getting device %s",
1334                        argv[1]);
1335                 return -EINVAL;
1336         }
1337
1338         r = action_dev(m, dev, action);
1339
1340         dm_put_device(ti, dev);
1341
1342         return r;
1343
1344 error:
1345         DMWARN("Unrecognised multipath message received.");
1346         return -EINVAL;
1347 }
1348
1349 static int multipath_ioctl(struct dm_target *ti, struct inode *inode,
1350                            struct file *filp, unsigned int cmd,
1351                            unsigned long arg)
1352 {
1353         struct multipath *m = (struct multipath *) ti->private;
1354         struct block_device *bdev = NULL;
1355         unsigned long flags;
1356         struct file fake_file = {};
1357         struct dentry fake_dentry = {};
1358         int r = 0;
1359
1360         fake_file.f_path.dentry = &fake_dentry;
1361
1362         spin_lock_irqsave(&m->lock, flags);
1363
1364         if (!m->current_pgpath)
1365                 __choose_pgpath(m);
1366
1367         if (m->current_pgpath) {
1368                 bdev = m->current_pgpath->path.dev->bdev;
1369                 fake_dentry.d_inode = bdev->bd_inode;
1370                 fake_file.f_mode = m->current_pgpath->path.dev->mode;
1371         }
1372
1373         if (m->queue_io)
1374                 r = -EAGAIN;
1375         else if (!bdev)
1376                 r = -EIO;
1377
1378         spin_unlock_irqrestore(&m->lock, flags);
1379
1380         return r ? : blkdev_driver_ioctl(bdev->bd_inode, &fake_file,
1381                                          bdev->bd_disk, cmd, arg);
1382 }
1383
1384 /*-----------------------------------------------------------------
1385  * Module setup
1386  *---------------------------------------------------------------*/
1387 static struct target_type multipath_target = {
1388         .name = "multipath",
1389         .version = {1, 0, 5},
1390         .module = THIS_MODULE,
1391         .ctr = multipath_ctr,
1392         .dtr = multipath_dtr,
1393         .map = multipath_map,
1394         .end_io = multipath_end_io,
1395         .presuspend = multipath_presuspend,
1396         .resume = multipath_resume,
1397         .status = multipath_status,
1398         .message = multipath_message,
1399         .ioctl  = multipath_ioctl,
1400 };
1401
1402 static int __init dm_multipath_init(void)
1403 {
1404         int r;
1405
1406         /* allocate a slab for the dm_ios */
1407         _mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1408         if (!_mpio_cache)
1409                 return -ENOMEM;
1410
1411         r = dm_register_target(&multipath_target);
1412         if (r < 0) {
1413                 DMERR("register failed %d", r);
1414                 kmem_cache_destroy(_mpio_cache);
1415                 return -EINVAL;
1416         }
1417
1418         kmultipathd = create_workqueue("kmpathd");
1419         if (!kmultipathd) {
1420                 DMERR("failed to create workqueue kmpathd");
1421                 dm_unregister_target(&multipath_target);
1422                 kmem_cache_destroy(_mpio_cache);
1423                 return -ENOMEM;
1424         }
1425
1426         /*
1427          * A separate workqueue is used to handle the device handlers
1428          * to avoid overloading existing workqueue. Overloading the
1429          * old workqueue would also create a bottleneck in the
1430          * path of the storage hardware device activation.
1431          */
1432         kmpath_handlerd = create_singlethread_workqueue("kmpath_handlerd");
1433         if (!kmpath_handlerd) {
1434                 DMERR("failed to create workqueue kmpath_handlerd");
1435                 destroy_workqueue(kmultipathd);
1436                 dm_unregister_target(&multipath_target);
1437                 kmem_cache_destroy(_mpio_cache);
1438                 return -ENOMEM;
1439         }
1440
1441         DMINFO("version %u.%u.%u loaded",
1442                multipath_target.version[0], multipath_target.version[1],
1443                multipath_target.version[2]);
1444
1445         return r;
1446 }
1447
1448 static void __exit dm_multipath_exit(void)
1449 {
1450         int r;
1451
1452         destroy_workqueue(kmpath_handlerd);
1453         destroy_workqueue(kmultipathd);
1454
1455         r = dm_unregister_target(&multipath_target);
1456         if (r < 0)
1457                 DMERR("target unregister failed %d", r);
1458         kmem_cache_destroy(_mpio_cache);
1459 }
1460
1461 module_init(dm_multipath_init);
1462 module_exit(dm_multipath_exit);
1463
1464 MODULE_DESCRIPTION(DM_NAME " multipath target");
1465 MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1466 MODULE_LICENSE("GPL");