drivers/md/dm-raid.c

   1 /*
   2  * Copyright (C) 2010-2011 Neil Brown
   3  * Copyright (C) 2010-2015 Red Hat, Inc. All rights reserved.
   4  *
   5  * This file is released under the GPL.
   6  */
   7
   8 #include <linux/slab.h>
   9 #include <linux/module.h>
  10
  11 #include "md.h"
  12 #include "raid1.h"
  13 #include "raid5.h"
  14 #include "raid10.h"
  15 #include "bitmap.h"
  16
  17 #include <linux/device-mapper.h>
  18
  19 #define DM_MSG_PREFIX "raid"
  20 #define MAX_RAID_DEVICES        253 /* raid4/5/6 limit */
  21
  22 static bool devices_handle_discard_safely = false;
  23
  24 /*
  25  * The following flags are used by dm-raid.c to set up the array state.
  26  * They must be cleared before md_run is called.
  27  */
  28 #define FirstUse 10             /* rdev flag */
  29
  30 struct raid_dev {
  31         /*
  32          * Two DM devices, one to hold metadata and one to hold the
  33          * actual data/parity.  The reason for this is to not confuse
  34          * ti->len and give more flexibility in altering size and
  35          * characteristics.
  36          *
  37          * While it is possible for this device to be associated
  38          * with a different physical device than the data_dev, it
  39          * is intended for it to be the same.
  40          *    |--------- Physical Device ---------|
  41          *    |- meta_dev -|------ data_dev ------|
  42          */
  43         struct dm_dev *meta_dev;
  44         struct dm_dev *data_dev;
  45         struct md_rdev rdev;
  46 };
  47
  48 /*
  49  * Flags for rs->ctr_flags field.
  50  */
  51 #define CTR_FLAG_SYNC              0x1
  52 #define CTR_FLAG_NOSYNC            0x2
  53 #define CTR_FLAG_REBUILD           0x4
  54 #define CTR_FLAG_DAEMON_SLEEP      0x8
  55 #define CTR_FLAG_MIN_RECOVERY_RATE 0x10
  56 #define CTR_FLAG_MAX_RECOVERY_RATE 0x20
  57 #define CTR_FLAG_MAX_WRITE_BEHIND  0x40
  58 #define CTR_FLAG_STRIPE_CACHE      0x80
  59 #define CTR_FLAG_REGION_SIZE       0x100
  60 #define CTR_FLAG_RAID10_COPIES     0x200
  61 #define CTR_FLAG_RAID10_FORMAT     0x400
  62
  63 struct raid_set {
  64         struct dm_target *ti;
  65
  66         uint32_t bitmap_loaded;
  67         uint32_t ctr_flags;
  68
  69         struct mddev md;
  70         struct raid_type *raid_type;
  71         struct dm_target_callbacks callbacks;
  72
  73         struct raid_dev dev[0];
  74 };
  75
  76 /* Supported raid types and properties. */
  77 static struct raid_type {
  78         const char *name;               /* RAID algorithm. */
  79         const char *descr;              /* Descriptor text for logging. */
  80         const unsigned parity_devs;     /* # of parity devices. */
  81         const unsigned minimal_devs;    /* minimal # of devices in set. */
  82         const unsigned level;           /* RAID level. */
  83         const unsigned algorithm;       /* RAID algorithm. */
  84 } raid_types[] = {
  85         {"raid0",    "RAID0 (striping)",                0, 2, 0, 0 /* NONE */},
  86         {"raid1",    "RAID1 (mirroring)",               0, 2, 1, 0 /* NONE */},
  87         {"raid10",   "RAID10 (striped mirrors)",        0, 2, 10, UINT_MAX /* Varies */},
  88         {"raid4",    "RAID4 (dedicated parity disk)",   1, 2, 5, ALGORITHM_PARITY_0},
  89         {"raid5_la", "RAID5 (left asymmetric)",         1, 2, 5, ALGORITHM_LEFT_ASYMMETRIC},
  90         {"raid5_ra", "RAID5 (right asymmetric)",        1, 2, 5, ALGORITHM_RIGHT_ASYMMETRIC},
  91         {"raid5_ls", "RAID5 (left symmetric)",          1, 2, 5, ALGORITHM_LEFT_SYMMETRIC},
  92         {"raid5_rs", "RAID5 (right symmetric)",         1, 2, 5, ALGORITHM_RIGHT_SYMMETRIC},
  93         {"raid6_zr", "RAID6 (zero restart)",            2, 4, 6, ALGORITHM_ROTATING_ZERO_RESTART},
  94         {"raid6_nr", "RAID6 (N restart)",               2, 4, 6, ALGORITHM_ROTATING_N_RESTART},
  95         {"raid6_nc", "RAID6 (N continue)",              2, 4, 6, ALGORITHM_ROTATING_N_CONTINUE}
  96 };
  97
  98 static char *raid10_md_layout_to_format(int layout)
  99 {
 100         /*
 101          * Bit 16 and 17 stand for "offset" and "use_far_sets"
 102          * Refer to MD's raid10.c for details
 103          */
 104         if ((layout & 0x10000) && (layout & 0x20000))
 105                 return "offset";
 106
 107         if ((layout & 0xFF) > 1)
 108                 return "near";
 109
 110         return "far";
 111 }
 112
 113 static unsigned raid10_md_layout_to_copies(int layout)
 114 {
 115         if ((layout & 0xFF) > 1)
 116                 return layout & 0xFF;
 117         return (layout >> 8) & 0xFF;
 118 }
 119
 120 static int raid10_format_to_md_layout(char *format, unsigned copies)
 121 {
 122         unsigned n = 1, f = 1;
 123
 124         if (!strcasecmp("near", format))
 125                 n = copies;
 126         else
 127                 f = copies;
 128
 129         if (!strcasecmp("offset", format))
 130                 return 0x30000 | (f << 8) | n;
 131
 132         if (!strcasecmp("far", format))
 133                 return 0x20000 | (f << 8) | n;
 134
 135         return (f << 8) | n;
 136 }
 137
 138 static struct raid_type *get_raid_type(char *name)
 139 {
 140         int i;
 141
 142         for (i = 0; i < ARRAY_SIZE(raid_types); i++)
 143                 if (!strcmp(raid_types[i].name, name))
 144                         return &raid_types[i];
 145
 146         return NULL;
 147 }
 148
 149 static struct raid_set *context_alloc(struct dm_target *ti, struct raid_type *raid_type, unsigned raid_devs)
 150 {
 151         unsigned i;
 152         struct raid_set *rs;
 153
 154         if (raid_devs <= raid_type->parity_devs) {
 155                 ti->error = "Insufficient number of devices";
 156                 return ERR_PTR(-EINVAL);
 157         }
 158
 159         rs = kzalloc(sizeof(*rs) + raid_devs * sizeof(rs->dev[0]), GFP_KERNEL);
 160         if (!rs) {
 161                 ti->error = "Cannot allocate raid context";
 162                 return ERR_PTR(-ENOMEM);
 163         }
 164
 165         mddev_init(&rs->md);
 166
 167         rs->ti = ti;
 168         rs->raid_type = raid_type;
 169         rs->md.raid_disks = raid_devs;
 170         rs->md.level = raid_type->level;
 171         rs->md.new_level = rs->md.level;
 172         rs->md.layout = raid_type->algorithm;
 173         rs->md.new_layout = rs->md.layout;
 174         rs->md.delta_disks = 0;
 175         rs->md.recovery_cp = 0;
 176
 177         for (i = 0; i < raid_devs; i++)
 178                 md_rdev_init(&rs->dev[i].rdev);
 179
 180         /*
 181          * Remaining items to be initialized by further RAID params:
 182          *  rs->md.persistent
 183          *  rs->md.external
 184          *  rs->md.chunk_sectors
 185          *  rs->md.new_chunk_sectors
 186          *  rs->md.dev_sectors
 187          */
 188
 189         return rs;
 190 }
 191
 192 static void context_free(struct raid_set *rs)
 193 {
 194         int i;
 195
 196         for (i = 0; i < rs->md.raid_disks; i++) {
 197                 if (rs->dev[i].meta_dev)
 198                         dm_put_device(rs->ti, rs->dev[i].meta_dev);
 199                 md_rdev_clear(&rs->dev[i].rdev);
 200                 if (rs->dev[i].data_dev)
 201                         dm_put_device(rs->ti, rs->dev[i].data_dev);
 202         }
 203
 204         kfree(rs);
 205 }
 206
 207 /*
 208  * For every device we have two words
 209  *  <meta_dev>: meta device name or '-' if missing
 210  *  <data_dev>: data device name or '-' if missing
 211  *
 212  * The following are permitted:
 213  *    - -
 214  *    - <data_dev>
 215  *    <meta_dev> <data_dev>
 216  *
 217  * The following is not allowed:
 218  *    <meta_dev> -
 219  *
 220  * This code parses those words.  If there is a failure,
 221  * the caller must use context_free to unwind the operations.
 222  */
 223 static int dev_parms(struct raid_set *rs, char **argv)
 224 {
 225         int i;
 226         int rebuild = 0;
 227         int metadata_available = 0;
 228         int ret = 0;
 229
 230         for (i = 0; i < rs->md.raid_disks; i++, argv += 2) {
 231                 rs->dev[i].rdev.raid_disk = i;
 232
 233                 rs->dev[i].meta_dev = NULL;
 234                 rs->dev[i].data_dev = NULL;
 235
 236                 /*
 237                  * There are no offsets, since there is a separate device
 238                  * for data and metadata.
 239                  */
 240                 rs->dev[i].rdev.data_offset = 0;
 241                 rs->dev[i].rdev.mddev = &rs->md;
 242
 243                 if (strcmp(argv[0], "-")) {
 244                         ret = dm_get_device(rs->ti, argv[0],
 245                                             dm_table_get_mode(rs->ti->table),
 246                                             &rs->dev[i].meta_dev);
 247                         rs->ti->error = "RAID metadata device lookup failure";
 248                         if (ret)
 249                                 return ret;
 250
 251                         rs->dev[i].rdev.sb_page = alloc_page(GFP_KERNEL);
 252                         if (!rs->dev[i].rdev.sb_page)
 253                                 return -ENOMEM;
 254                 }
 255
 256                 if (!strcmp(argv[1], "-")) {
 257                         if (!test_bit(In_sync, &rs->dev[i].rdev.flags) &&
 258                             (!rs->dev[i].rdev.recovery_offset)) {
 259                                 rs->ti->error = "Drive designated for rebuild not specified";
 260                                 return -EINVAL;
 261                         }
 262
 263                         rs->ti->error = "No data device supplied with metadata device";
 264                         if (rs->dev[i].meta_dev)
 265                                 return -EINVAL;
 266
 267                         continue;
 268                 }
 269
 270                 ret = dm_get_device(rs->ti, argv[1],
 271                                     dm_table_get_mode(rs->ti->table),
 272                                     &rs->dev[i].data_dev);
 273                 if (ret) {
 274                         rs->ti->error = "RAID device lookup failure";
 275                         return ret;
 276                 }
 277
 278                 if (rs->dev[i].meta_dev) {
 279                         metadata_available = 1;
 280                         rs->dev[i].rdev.meta_bdev = rs->dev[i].meta_dev->bdev;
 281                 }
 282                 rs->dev[i].rdev.bdev = rs->dev[i].data_dev->bdev;
 283                 list_add(&rs->dev[i].rdev.same_set, &rs->md.disks);
 284                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
 285                         rebuild++;
 286         }
 287
 288         if (metadata_available) {
 289                 rs->md.external = 0;
 290                 rs->md.persistent = 1;
 291                 rs->md.major_version = 2;
 292         } else if (rebuild && !rs->md.recovery_cp) {
 293                 /*
 294                  * Without metadata, we will not be able to tell if the array
 295                  * is in-sync or not - we must assume it is not.  Therefore,
 296                  * it is impossible to rebuild a drive.
 297                  *
 298                  * Even if there is metadata, the on-disk information may
 299                  * indicate that the array is not in-sync and it will then
 300                  * fail at that time.
 301                  *
 302                  * User could specify 'nosync' option if desperate.
 303                  */
 304                 DMERR("Unable to rebuild drive while array is not in-sync");
 305                 rs->ti->error = "RAID device lookup failure";
 306                 return -EINVAL;
 307         }
 308
 309         return 0;
 310 }
 311
 312 /*
 313  * validate_region_size
 314  * @rs
 315  * @region_size:  region size in sectors.  If 0, pick a size (4MiB default).
 316  *
 317  * Set rs->md.bitmap_info.chunksize (which really refers to 'region size').
 318  * Ensure that (ti->len/region_size < 2^21) - required by MD bitmap.
 319  *
 320  * Returns: 0 on success, -EINVAL on failure.
 321  */
 322 static int validate_region_size(struct raid_set *rs, unsigned long region_size)
 323 {
 324         unsigned long min_region_size = rs->ti->len / (1 << 21);
 325
 326         if (!region_size) {
 327                 /*
 328                  * Choose a reasonable default.  All figures in sectors.
 329                  */
 330                 if (min_region_size > (1 << 13)) {
 331                         /* If not a power of 2, make it the next power of 2 */
 332                         region_size = roundup_pow_of_two(min_region_size);
 333                         DMINFO("Choosing default region size of %lu sectors",
 334                                region_size);
 335                 } else {
 336                         DMINFO("Choosing default region size of 4MiB");
 337                         region_size = 1 << 13; /* sectors */
 338                 }
 339         } else {
 340                 /*
 341                  * Validate user-supplied value.
 342                  */
 343                 if (region_size > rs->ti->len) {
 344                         rs->ti->error = "Supplied region size is too large";
 345                         return -EINVAL;
 346                 }
 347
 348                 if (region_size < min_region_size) {
 349                         DMERR("Supplied region_size (%lu sectors) below minimum (%lu)",
 350                               region_size, min_region_size);
 351                         rs->ti->error = "Supplied region size is too small";
 352                         return -EINVAL;
 353                 }
 354
 355                 if (!is_power_of_2(region_size)) {
 356                         rs->ti->error = "Region size is not a power of 2";
 357                         return -EINVAL;
 358                 }
 359
 360                 if (region_size < rs->md.chunk_sectors) {
 361                         rs->ti->error = "Region size is smaller than the chunk size";
 362                         return -EINVAL;
 363                 }
 364         }
 365
 366         /*
 367          * Convert sectors to bytes.
 368          */
 369         rs->md.bitmap_info.chunksize = (region_size << 9);
 370
 371         return 0;
 372 }
 373
 374 /*
 375  * validate_raid_redundancy
 376  * @rs
 377  *
 378  * Determine if there are enough devices in the array that haven't
 379  * failed (or are being rebuilt) to form a usable array.
 380  *
 381  * Returns: 0 on success, -EINVAL on failure.
 382  */
 383 static int validate_raid_redundancy(struct raid_set *rs)
 384 {
 385         unsigned i, rebuild_cnt = 0;
 386         unsigned rebuilds_per_group = 0, copies, d;
 387         unsigned group_size, last_group_start;
 388
 389         for (i = 0; i < rs->md.raid_disks; i++)
 390                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags) ||
 391                     !rs->dev[i].rdev.sb_page)
 392                         rebuild_cnt++;
 393
 394         switch (rs->raid_type->level) {
 395         case 1:
 396                 if (rebuild_cnt >= rs->md.raid_disks)
 397                         goto too_many;
 398                 break;
 399         case 4:
 400         case 5:
 401         case 6:
 402                 if (rebuild_cnt > rs->raid_type->parity_devs)
 403                         goto too_many;
 404                 break;
 405         case 10:
 406                 copies = raid10_md_layout_to_copies(rs->md.layout);
 407                 if (rebuild_cnt < copies)
 408                         break;
 409
 410                 /*
 411                  * It is possible to have a higher rebuild count for RAID10,
 412                  * as long as the failed devices occur in different mirror
 413                  * groups (i.e. different stripes).
 414                  *
 415                  * When checking "near" format, make sure no adjacent devices
 416                  * have failed beyond what can be handled.  In addition to the
 417                  * simple case where the number of devices is a multiple of the
 418                  * number of copies, we must also handle cases where the number
 419                  * of devices is not a multiple of the number of copies.
 420                  * E.g.    dev1 dev2 dev3 dev4 dev5
 421                  *          A    A    B    B    C
 422                  *          C    D    D    E    E
 423                  */
 424                 if (!strcmp("near", raid10_md_layout_to_format(rs->md.layout))) {
 425                         for (i = 0; i < rs->md.raid_disks * copies; i++) {
 426                                 if (!(i % copies))
 427                                         rebuilds_per_group = 0;
 428                                 d = i % rs->md.raid_disks;
 429                                 if ((!rs->dev[d].rdev.sb_page ||
 430                                      !test_bit(In_sync, &rs->dev[d].rdev.flags)) &&
 431                                     (++rebuilds_per_group >= copies))
 432                                         goto too_many;
 433                         }
 434                         break;
 435                 }
 436
 437                 /*
 438                  * When checking "far" and "offset" formats, we need to ensure
 439                  * that the device that holds its copy is not also dead or
 440                  * being rebuilt.  (Note that "far" and "offset" formats only
 441                  * support two copies right now.  These formats also only ever
 442                  * use the 'use_far_sets' variant.)
 443                  *
 444                  * This check is somewhat complicated by the need to account
 445                  * for arrays that are not a multiple of (far) copies.  This
 446                  * results in the need to treat the last (potentially larger)
 447                  * set differently.
 448                  */
 449                 group_size = (rs->md.raid_disks / copies);
 450                 last_group_start = (rs->md.raid_disks / group_size) - 1;
 451                 last_group_start *= group_size;
 452                 for (i = 0; i < rs->md.raid_disks; i++) {
 453                         if (!(i % copies) && !(i > last_group_start))
 454                                 rebuilds_per_group = 0;
 455                         if ((!rs->dev[i].rdev.sb_page ||
 456                              !test_bit(In_sync, &rs->dev[i].rdev.flags)) &&
 457                             (++rebuilds_per_group >= copies))
 458                                         goto too_many;
 459                 }
 460                 break;
 461         default:
 462                 if (rebuild_cnt)
 463                         return -EINVAL;
 464         }
 465
 466         return 0;
 467
 468 too_many:
 469         return -EINVAL;
 470 }
 471
 472 /*
 473  * Possible arguments are...
 474  *      <chunk_size> [optional_args]
 475  *
 476  * Argument definitions
 477  *    <chunk_size>                      The number of sectors per disk that
 478  *                                      will form the "stripe"
 479  *    [[no]sync]                        Force or prevent recovery of the
 480  *                                      entire array
 481  *    [rebuild <idx>]                   Rebuild the drive indicated by the index
 482  *    [daemon_sleep <ms>]               Time between bitmap daemon work to
 483  *                                      clear bits
 484  *    [min_recovery_rate <kB/sec/disk>] Throttle RAID initialization
 485  *    [max_recovery_rate <kB/sec/disk>] Throttle RAID initialization
 486  *    [write_mostly <idx>]              Indicate a write mostly drive via index
 487  *    [max_write_behind <sectors>]      See '-write-behind=' (man mdadm)
 488  *    [stripe_cache <sectors>]          Stripe cache size for higher RAIDs
 489  *    [region_size <sectors>]           Defines granularity of bitmap
 490  *
 491  * RAID10-only options:
 492  *    [raid10_copies <# copies>]        Number of copies.  (Default: 2)
 493  *    [raid10_format <near|far|offset>] Layout algorithm.  (Default: near)
 494  */
 495 static int parse_raid_params(struct raid_set *rs, char **argv,
 496                              unsigned num_raid_params)
 497 {
 498         char *raid10_format = "near";
 499         unsigned raid10_copies = 2;
 500         unsigned i;
 501         unsigned long value, region_size = 0;
 502         sector_t sectors_per_dev = rs->ti->len;
 503         sector_t max_io_len;
 504         char *key;
 505
 506         /*
 507          * First, parse the in-order required arguments
 508          * "chunk_size" is the only argument of this type.
 509          */
 510         if ((kstrtoul(argv[0], 10, &value) < 0)) {
 511                 rs->ti->error = "Bad chunk size";
 512                 return -EINVAL;
 513         } else if (rs->raid_type->level == 1) {
 514                 if (value)
 515                         DMERR("Ignoring chunk size parameter for RAID 1");
 516                 value = 0;
 517         } else if (!is_power_of_2(value)) {
 518                 rs->ti->error = "Chunk size must be a power of 2";
 519                 return -EINVAL;
 520         } else if (value < 8) {
 521                 rs->ti->error = "Chunk size value is too small";
 522                 return -EINVAL;
 523         }
 524
 525         rs->md.new_chunk_sectors = rs->md.chunk_sectors = value;
 526         argv++;
 527         num_raid_params--;
 528
 529         /*
 530          * We set each individual device as In_sync with a completed
 531          * 'recovery_offset'.  If there has been a device failure or
 532          * replacement then one of the following cases applies:
 533          *
 534          *   1) User specifies 'rebuild'.
 535          *      - Device is reset when param is read.
 536          *   2) A new device is supplied.
 537          *      - No matching superblock found, resets device.
 538          *   3) Device failure was transient and returns on reload.
 539          *      - Failure noticed, resets device for bitmap replay.
 540          *   4) Device hadn't completed recovery after previous failure.
 541          *      - Superblock is read and overrides recovery_offset.
 542          *
 543          * What is found in the superblocks of the devices is always
 544          * authoritative, unless 'rebuild' or '[no]sync' was specified.
 545          */
 546         for (i = 0; i < rs->md.raid_disks; i++) {
 547                 set_bit(In_sync, &rs->dev[i].rdev.flags);
 548                 rs->dev[i].rdev.recovery_offset = MaxSector;
 549         }
 550
 551         /*
 552          * Second, parse the unordered optional arguments
 553          */
 554         for (i = 0; i < num_raid_params; i++) {
 555                 if (!strcasecmp(argv[i], "nosync")) {
 556                         rs->md.recovery_cp = MaxSector;
 557                         rs->ctr_flags |= CTR_FLAG_NOSYNC;
 558                         continue;
 559                 }
 560                 if (!strcasecmp(argv[i], "sync")) {
 561                         rs->md.recovery_cp = 0;
 562                         rs->ctr_flags |= CTR_FLAG_SYNC;
 563                         continue;
 564                 }
 565
 566                 /* The rest of the optional arguments come in key/value pairs */
 567                 if ((i + 1) >= num_raid_params) {
 568                         rs->ti->error = "Wrong number of raid parameters given";
 569                         return -EINVAL;
 570                 }
 571
 572                 key = argv[i++];
 573
 574                 /* Parameters that take a string value are checked here. */
 575                 if (!strcasecmp(key, "raid10_format")) {
 576                         if (rs->raid_type->level != 10) {
 577                                 rs->ti->error = "'raid10_format' is an invalid parameter for this RAID type";
 578                                 return -EINVAL;
 579                         }
 580                         if (strcmp("near", argv[i]) &&
 581                             strcmp("far", argv[i]) &&
 582                             strcmp("offset", argv[i])) {
 583                                 rs->ti->error = "Invalid 'raid10_format' value given";
 584                                 return -EINVAL;
 585                         }
 586                         raid10_format = argv[i];
 587                         rs->ctr_flags |= CTR_FLAG_RAID10_FORMAT;
 588                         continue;
 589                 }
 590
 591                 if (kstrtoul(argv[i], 10, &value) < 0) {
 592                         rs->ti->error = "Bad numerical argument given in raid params";
 593                         return -EINVAL;
 594                 }
 595
 596                 /* Parameters that take a numeric value are checked here */
 597                 if (!strcasecmp(key, "rebuild")) {
 598                         if (value >= rs->md.raid_disks) {
 599                                 rs->ti->error = "Invalid rebuild index given";
 600                                 return -EINVAL;
 601                         }
 602                         clear_bit(In_sync, &rs->dev[value].rdev.flags);
 603                         rs->dev[value].rdev.recovery_offset = 0;
 604                         rs->ctr_flags |= CTR_FLAG_REBUILD;
 605                 } else if (!strcasecmp(key, "write_mostly")) {
 606                         if (rs->raid_type->level != 1) {
 607                                 rs->ti->error = "write_mostly option is only valid for RAID1";
 608                                 return -EINVAL;
 609                         }
 610                         if (value >= rs->md.raid_disks) {
 611                                 rs->ti->error = "Invalid write_mostly drive index given";
 612                                 return -EINVAL;
 613                         }
 614                         set_bit(WriteMostly, &rs->dev[value].rdev.flags);
 615                 } else if (!strcasecmp(key, "max_write_behind")) {
 616                         if (rs->raid_type->level != 1) {
 617                                 rs->ti->error = "max_write_behind option is only valid for RAID1";
 618                                 return -EINVAL;
 619                         }
 620                         rs->ctr_flags |= CTR_FLAG_MAX_WRITE_BEHIND;
 621
 622                         /*
 623                          * In device-mapper, we specify things in sectors, but
 624                          * MD records this value in kB
 625                          */
 626                         value /= 2;
 627                         if (value > COUNTER_MAX) {
 628                                 rs->ti->error = "Max write-behind limit out of range";
 629                                 return -EINVAL;
 630                         }
 631                         rs->md.bitmap_info.max_write_behind = value;
 632                 } else if (!strcasecmp(key, "daemon_sleep")) {
 633                         rs->ctr_flags |= CTR_FLAG_DAEMON_SLEEP;
 634                         if (!value || (value > MAX_SCHEDULE_TIMEOUT)) {
 635                                 rs->ti->error = "daemon sleep period out of range";
 636                                 return -EINVAL;
 637                         }
 638                         rs->md.bitmap_info.daemon_sleep = value;
 639                 } else if (!strcasecmp(key, "stripe_cache")) {
 640                         rs->ctr_flags |= CTR_FLAG_STRIPE_CACHE;
 641
 642                         /*
 643                          * In device-mapper, we specify things in sectors, but
 644                          * MD records this value in kB
 645                          */
 646                         value /= 2;
 647
 648                         if ((rs->raid_type->level != 5) &&
 649                             (rs->raid_type->level != 6)) {
 650                                 rs->ti->error = "Inappropriate argument: stripe_cache";
 651                                 return -EINVAL;
 652                         }
 653                         if (raid5_set_cache_size(&rs->md, (int)value)) {
 654                                 rs->ti->error = "Bad stripe_cache size";
 655                                 return -EINVAL;
 656                         }
 657                 } else if (!strcasecmp(key, "min_recovery_rate")) {
 658                         rs->ctr_flags |= CTR_FLAG_MIN_RECOVERY_RATE;
 659                         if (value > INT_MAX) {
 660                                 rs->ti->error = "min_recovery_rate out of range";
 661                                 return -EINVAL;
 662                         }
 663                         rs->md.sync_speed_min = (int)value;
 664                 } else if (!strcasecmp(key, "max_recovery_rate")) {
 665                         rs->ctr_flags |= CTR_FLAG_MAX_RECOVERY_RATE;
 666                         if (value > INT_MAX) {
 667                                 rs->ti->error = "max_recovery_rate out of range";
 668                                 return -EINVAL;
 669                         }
 670                         rs->md.sync_speed_max = (int)value;
 671                 } else if (!strcasecmp(key, "region_size")) {
 672                         rs->ctr_flags |= CTR_FLAG_REGION_SIZE;
 673                         region_size = value;
 674                 } else if (!strcasecmp(key, "raid10_copies") &&
 675                            (rs->raid_type->level == 10)) {
 676                         if ((value < 2) || (value > 0xFF)) {
 677                                 rs->ti->error = "Bad value for 'raid10_copies'";
 678                                 return -EINVAL;
 679                         }
 680                         rs->ctr_flags |= CTR_FLAG_RAID10_COPIES;
 681                         raid10_copies = value;
 682                 } else {
 683                         DMERR("Unable to parse RAID parameter: %s", key);
 684                         rs->ti->error = "Unable to parse RAID parameters";
 685                         return -EINVAL;
 686                 }
 687         }
 688
 689         if (validate_region_size(rs, region_size))
 690                 return -EINVAL;
 691
 692         if (rs->md.chunk_sectors)
 693                 max_io_len = rs->md.chunk_sectors;
 694         else
 695                 max_io_len = region_size;
 696
 697         if (dm_set_target_max_io_len(rs->ti, max_io_len))
 698                 return -EINVAL;
 699
 700         if (rs->raid_type->level == 10) {
 701                 if (raid10_copies > rs->md.raid_disks) {
 702                         rs->ti->error = "Not enough devices to satisfy specification";
 703                         return -EINVAL;
 704                 }
 705
 706                 /*
 707                  * If the format is not "near", we only support
 708                  * two copies at the moment.
 709                  */
 710                 if (strcmp("near", raid10_format) && (raid10_copies > 2)) {
 711                         rs->ti->error = "Too many copies for given RAID10 format.";
 712                         return -EINVAL;
 713                 }
 714
 715                 /* (Len * #mirrors) / #devices */
 716                 sectors_per_dev = rs->ti->len * raid10_copies;
 717                 sector_div(sectors_per_dev, rs->md.raid_disks);
 718
 719                 rs->md.layout = raid10_format_to_md_layout(raid10_format,
 720                                                            raid10_copies);
 721                 rs->md.new_layout = rs->md.layout;
 722         } else if ((!rs->raid_type->level || rs->raid_type->level > 1) &&
 723                    sector_div(sectors_per_dev,
 724                               (rs->md.raid_disks - rs->raid_type->parity_devs))) {
 725                 rs->ti->error = "Target length not divisible by number of data devices";
 726                 return -EINVAL;
 727         }
 728         rs->md.dev_sectors = sectors_per_dev;
 729
 730         /* Assume there are no metadata devices until the drives are parsed */
 731         rs->md.persistent = 0;
 732         rs->md.external = 1;
 733
 734         return 0;
 735 }
 736
 737 static void do_table_event(struct work_struct *ws)
 738 {
 739         struct raid_set *rs = container_of(ws, struct raid_set, md.event_work);
 740
 741         dm_table_event(rs->ti->table);
 742 }
 743
 744 static int raid_is_congested(struct dm_target_callbacks *cb, int bits)
 745 {
 746         struct raid_set *rs = container_of(cb, struct raid_set, callbacks);
 747
 748         return mddev_congested(&rs->md, bits);
 749 }
 750
 751 /*
 752  * This structure is never routinely used by userspace, unlike md superblocks.
 753  * Devices with this superblock should only ever be accessed via device-mapper.
 754  */
 755 #define DM_RAID_MAGIC 0x64526D44
 756 struct dm_raid_superblock {
 757         __le32 magic;           /* "DmRd" */
 758         __le32 features;        /* Used to indicate possible future changes */
 759
 760         __le32 num_devices;     /* Number of devices in this array. (Max 64) */
 761         __le32 array_position;  /* The position of this drive in the array */
 762
 763         __le64 events;          /* Incremented by md when superblock updated */
 764         __le64 failed_devices;  /* Bit field of devices to indicate failures */
 765
 766         /*
 767          * This offset tracks the progress of the repair or replacement of
 768          * an individual drive.
 769          */
 770         __le64 disk_recovery_offset;
 771
 772         /*
 773          * This offset tracks the progress of the initial array
 774          * synchronisation/parity calculation.
 775          */
 776         __le64 array_resync_offset;
 777
 778         /*
 779          * RAID characteristics
 780          */
 781         __le32 level;
 782         __le32 layout;
 783         __le32 stripe_sectors;
 784
 785         /* Remainder of a logical block is zero-filled when writing (see super_sync()). */
 786 } __packed;
 787
 788 static int read_disk_sb(struct md_rdev *rdev, int size)
 789 {
 790         BUG_ON(!rdev->sb_page);
 791
 792         if (rdev->sb_loaded)
 793                 return 0;
 794
 795         if (!sync_page_io(rdev, 0, size, rdev->sb_page, REQ_OP_READ, 0, 1)) {
 796                 DMERR("Failed to read superblock of device at position %d",
 797                       rdev->raid_disk);
 798                 md_error(rdev->mddev, rdev);
 799                 return -EINVAL;
 800         }
 801
 802         rdev->sb_loaded = 1;
 803
 804         return 0;
 805 }
 806
 807 static void super_sync(struct mddev *mddev, struct md_rdev *rdev)
 808 {
 809         int i;
 810         uint64_t failed_devices;
 811         struct dm_raid_superblock *sb;
 812         struct raid_set *rs = container_of(mddev, struct raid_set, md);
 813
 814         sb = page_address(rdev->sb_page);
 815         failed_devices = le64_to_cpu(sb->failed_devices);
 816
 817         for (i = 0; i < mddev->raid_disks; i++)
 818                 if (!rs->dev[i].data_dev ||
 819                     test_bit(Faulty, &(rs->dev[i].rdev.flags)))
 820                         failed_devices |= (1ULL << i);
 821
 822         memset(sb + 1, 0, rdev->sb_size - sizeof(*sb));
 823
 824         sb->magic = cpu_to_le32(DM_RAID_MAGIC);
 825         sb->features = cpu_to_le32(0);  /* No features yet */
 826
 827         sb->num_devices = cpu_to_le32(mddev->raid_disks);
 828         sb->array_position = cpu_to_le32(rdev->raid_disk);
 829
 830         sb->events = cpu_to_le64(mddev->events);
 831         sb->failed_devices = cpu_to_le64(failed_devices);
 832
 833         sb->disk_recovery_offset = cpu_to_le64(rdev->recovery_offset);
 834         sb->array_resync_offset = cpu_to_le64(mddev->recovery_cp);
 835
 836         sb->level = cpu_to_le32(mddev->level);
 837         sb->layout = cpu_to_le32(mddev->layout);
 838         sb->stripe_sectors = cpu_to_le32(mddev->chunk_sectors);
 839 }
 840
 841 /*
 842  * super_load
 843  *
 844  * This function creates a superblock if one is not found on the device
 845  * and will decide which superblock to use if there's a choice.
 846  *
 847  * Return: 1 if use rdev, 0 if use refdev, -Exxx otherwise
 848  */
 849 static int super_load(struct md_rdev *rdev, struct md_rdev *refdev)
 850 {
 851         int ret;
 852         struct dm_raid_superblock *sb;
 853         struct dm_raid_superblock *refsb;
 854         uint64_t events_sb, events_refsb;
 855
 856         rdev->sb_start = 0;
 857         rdev->sb_size = bdev_logical_block_size(rdev->meta_bdev);
 858         if (rdev->sb_size < sizeof(*sb) || rdev->sb_size > PAGE_SIZE) {
 859                 DMERR("superblock size of a logical block is no longer valid");
 860                 return -EINVAL;
 861         }
 862
 863         ret = read_disk_sb(rdev, rdev->sb_size);
 864         if (ret)
 865                 return ret;
 866
 867         sb = page_address(rdev->sb_page);
 868
 869         /*
 870          * Two cases that we want to write new superblocks and rebuild:
 871          * 1) New device (no matching magic number)
 872          * 2) Device specified for rebuild (!In_sync w/ offset == 0)
 873          */
 874         if ((sb->magic != cpu_to_le32(DM_RAID_MAGIC)) ||
 875             (!test_bit(In_sync, &rdev->flags) && !rdev->recovery_offset)) {
 876                 super_sync(rdev->mddev, rdev);
 877
 878                 set_bit(FirstUse, &rdev->flags);
 879
 880                 /* Force writing of superblocks to disk */
 881                 set_bit(MD_CHANGE_DEVS, &rdev->mddev->flags);
 882
 883                 /* Any superblock is better than none, choose that if given */
 884                 return refdev ? 0 : 1;
 885         }
 886
 887         if (!refdev)
 888                 return 1;
 889
 890         events_sb = le64_to_cpu(sb->events);
 891
 892         refsb = page_address(refdev->sb_page);
 893         events_refsb = le64_to_cpu(refsb->events);
 894
 895         return (events_sb > events_refsb) ? 1 : 0;
 896 }
 897
 898 static int super_init_validation(struct mddev *mddev, struct md_rdev *rdev)
 899 {
 900         int role;
 901         struct raid_set *rs = container_of(mddev, struct raid_set, md);
 902         uint64_t events_sb;
 903         uint64_t failed_devices;
 904         struct dm_raid_superblock *sb;
 905         uint32_t new_devs = 0;
 906         uint32_t rebuilds = 0;
 907         struct md_rdev *r;
 908         struct dm_raid_superblock *sb2;
 909
 910         sb = page_address(rdev->sb_page);
 911         events_sb = le64_to_cpu(sb->events);
 912         failed_devices = le64_to_cpu(sb->failed_devices);
 913
 914         /*
 915          * Initialise to 1 if this is a new superblock.
 916          */
 917         mddev->events = events_sb ? : 1;
 918
 919         /*
 920          * Reshaping is not currently allowed
 921          */
 922         if (le32_to_cpu(sb->level) != mddev->level) {
 923                 DMERR("Reshaping arrays not yet supported. (RAID level change)");
 924                 return -EINVAL;
 925         }
 926         if (le32_to_cpu(sb->layout) != mddev->layout) {
 927                 DMERR("Reshaping arrays not yet supported. (RAID layout change)");
 928                 DMERR("  0x%X vs 0x%X", le32_to_cpu(sb->layout), mddev->layout);
 929                 DMERR("  Old layout: %s w/ %d copies",
 930                       raid10_md_layout_to_format(le32_to_cpu(sb->layout)),
 931                       raid10_md_layout_to_copies(le32_to_cpu(sb->layout)));
 932                 DMERR("  New layout: %s w/ %d copies",
 933                       raid10_md_layout_to_format(mddev->layout),
 934                       raid10_md_layout_to_copies(mddev->layout));
 935                 return -EINVAL;
 936         }
 937         if (le32_to_cpu(sb->stripe_sectors) != mddev->chunk_sectors) {
 938                 DMERR("Reshaping arrays not yet supported. (stripe sectors change)");
 939                 return -EINVAL;
 940         }
 941
 942         /* We can only change the number of devices in RAID1 right now */
 943         if ((rs->raid_type->level != 1) &&
 944             (le32_to_cpu(sb->num_devices) != mddev->raid_disks)) {
 945                 DMERR("Reshaping arrays not yet supported. (device count change)");
 946                 return -EINVAL;
 947         }
 948
 949         if (!(rs->ctr_flags & (CTR_FLAG_SYNC | CTR_FLAG_NOSYNC)))
 950                 mddev->recovery_cp = le64_to_cpu(sb->array_resync_offset);
 951
 952         /*
 953          * During load, we set FirstUse if a new superblock was written.
 954          * There are two reasons we might not have a superblock:
 955          * 1) The array is brand new - in which case, all of the
 956          *    devices must have their In_sync bit set.  Also,
 957          *    recovery_cp must be 0, unless forced.
 958          * 2) This is a new device being added to an old array
 959          *    and the new device needs to be rebuilt - in which
 960          *    case the In_sync bit will /not/ be set and
 961          *    recovery_cp must be MaxSector.
 962          */
 963         rdev_for_each(r, mddev) {
 964                 if (!test_bit(In_sync, &r->flags)) {
 965                         DMINFO("Device %d specified for rebuild: "
 966                                "Clearing superblock", r->raid_disk);
 967                         rebuilds++;
 968                 } else if (test_bit(FirstUse, &r->flags))
 969                         new_devs++;
 970         }
 971
 972         if (!rebuilds) {
 973                 if (new_devs == mddev->raid_disks) {
 974                         DMINFO("Superblocks created for new array");
 975                         set_bit(MD_ARRAY_FIRST_USE, &mddev->flags);
 976                 } else if (new_devs) {
 977                         DMERR("New device injected "
 978                               "into existing array without 'rebuild' "
 979                               "parameter specified");
 980                         return -EINVAL;
 981                 }
 982         } else if (new_devs) {
 983                 DMERR("'rebuild' devices cannot be "
 984                       "injected into an array with other first-time devices");
 985                 return -EINVAL;
 986         } else if (mddev->recovery_cp != MaxSector) {
 987                 DMERR("'rebuild' specified while array is not in-sync");
 988                 return -EINVAL;
 989         }
 990
 991         /*
 992          * Now we set the Faulty bit for those devices that are
 993          * recorded in the superblock as failed.
 994          */
 995         rdev_for_each(r, mddev) {
 996                 if (!r->sb_page)
 997                         continue;
 998                 sb2 = page_address(r->sb_page);
 999                 sb2->failed_devices = 0;
1000
1001                 /*
1002                  * Check for any device re-ordering.
1003                  */
1004                 if (!test_bit(FirstUse, &r->flags) && (r->raid_disk >= 0)) {
1005                         role = le32_to_cpu(sb2->array_position);
1006                         if (role != r->raid_disk) {
1007                                 if (rs->raid_type->level != 1) {
1008                                         rs->ti->error = "Cannot change device "
1009                                                 "positions in RAID array";
1010                                         return -EINVAL;
1011                                 }
1012                                 DMINFO("RAID1 device #%d now at position #%d",
1013                                        role, r->raid_disk);
1014                         }
1015
1016                         /*
1017                          * Partial recovery is performed on
1018                          * returning failed devices.
1019                          */
1020                         if (failed_devices & (1 << role))
1021                                 set_bit(Faulty, &r->flags);
1022                 }
1023         }
1024
1025         return 0;
1026 }
1027
1028 static int super_validate(struct raid_set *rs, struct md_rdev *rdev)
1029 {
1030         struct mddev *mddev = &rs->md;
1031         struct dm_raid_superblock *sb = page_address(rdev->sb_page);
1032
1033         /*
1034          * If mddev->events is not set, we know we have not yet initialized
1035          * the array.
1036          */
1037         if (!mddev->events && super_init_validation(mddev, rdev))
1038                 return -EINVAL;
1039
1040         if (le32_to_cpu(sb->features)) {
1041                 rs->ti->error = "Unable to assemble array: No feature flags supported yet";
1042                 return -EINVAL;
1043         }
1044
1045         /* Enable bitmap creation for RAID levels != 0 */
1046         mddev->bitmap_info.offset = (rs->raid_type->level) ? to_sector(4096) : 0;
1047         rdev->mddev->bitmap_info.default_offset = mddev->bitmap_info.offset;
1048
1049         if (!test_bit(FirstUse, &rdev->flags)) {
1050                 rdev->recovery_offset = le64_to_cpu(sb->disk_recovery_offset);
1051                 if (rdev->recovery_offset != MaxSector)
1052                         clear_bit(In_sync, &rdev->flags);
1053         }
1054
1055         /*
1056          * If a device comes back, set it as not In_sync and no longer faulty.
1057          */
1058         if (test_bit(Faulty, &rdev->flags)) {
1059                 clear_bit(Faulty, &rdev->flags);
1060                 clear_bit(In_sync, &rdev->flags);
1061                 rdev->saved_raid_disk = rdev->raid_disk;
1062                 rdev->recovery_offset = 0;
1063         }
1064
1065         clear_bit(FirstUse, &rdev->flags);
1066
1067         return 0;
1068 }
1069
1070 /*
1071  * Analyse superblocks and select the freshest.
1072  */
1073 static int analyse_superblocks(struct dm_target *ti, struct raid_set *rs)
1074 {
1075         int ret;
1076         struct raid_dev *dev;
1077         struct md_rdev *rdev, *tmp, *freshest;
1078         struct mddev *mddev = &rs->md;
1079
1080         freshest = NULL;
1081         rdev_for_each_safe(rdev, tmp, mddev) {
1082                 /*
1083                  * Skipping super_load due to CTR_FLAG_SYNC will cause
1084                  * the array to undergo initialization again as
1085                  * though it were new.  This is the intended effect
1086                  * of the "sync" directive.
1087                  *
1088                  * When reshaping capability is added, we must ensure
1089                  * that the "sync" directive is disallowed during the
1090                  * reshape.
1091                  */
1092                 rdev->sectors = to_sector(i_size_read(rdev->bdev->bd_inode));
1093
1094                 if (rs->ctr_flags & CTR_FLAG_SYNC)
1095                         continue;
1096
1097                 if (!rdev->meta_bdev)
1098                         continue;
1099
1100                 ret = super_load(rdev, freshest);
1101
1102                 switch (ret) {
1103                 case 1:
1104                         freshest = rdev;
1105                         break;
1106                 case 0:
1107                         break;
1108                 default:
1109                         dev = container_of(rdev, struct raid_dev, rdev);
1110                         if (dev->meta_dev)
1111                                 dm_put_device(ti, dev->meta_dev);
1112
1113                         dev->meta_dev = NULL;
1114                         rdev->meta_bdev = NULL;
1115
1116                         if (rdev->sb_page)
1117                                 put_page(rdev->sb_page);
1118
1119                         rdev->sb_page = NULL;
1120
1121                         rdev->sb_loaded = 0;
1122
1123                         /*
1124                          * We might be able to salvage the data device
1125                          * even though the meta device has failed.  For
1126                          * now, we behave as though '- -' had been
1127                          * set for this device in the table.
1128                          */
1129                         if (dev->data_dev)
1130                                 dm_put_device(ti, dev->data_dev);
1131
1132                         dev->data_dev = NULL;
1133                         rdev->bdev = NULL;
1134
1135                         list_del(&rdev->same_set);
1136                 }
1137         }
1138
1139         if (!freshest)
1140                 return 0;
1141
1142         if (validate_raid_redundancy(rs)) {
1143                 rs->ti->error = "Insufficient redundancy to activate array";
1144                 return -EINVAL;
1145         }
1146
1147         /*
1148          * Validation of the freshest device provides the source of
1149          * validation for the remaining devices.
1150          */
1151         ti->error = "Unable to assemble array: Invalid superblocks";
1152         if (super_validate(rs, freshest))
1153                 return -EINVAL;
1154
1155         rdev_for_each(rdev, mddev)
1156                 if ((rdev != freshest) && super_validate(rs, rdev))
1157                         return -EINVAL;
1158
1159         return 0;
1160 }
1161
1162 /*
1163  * Enable/disable discard support on RAID set depending on
1164  * RAID level and discard properties of underlying RAID members.
1165  */
1166 static void configure_discard_support(struct dm_target *ti, struct raid_set *rs)
1167 {
1168         int i;
1169         bool raid456;
1170
1171         /* Assume discards not supported until after checks below. */
1172         ti->discards_supported = false;
1173
1174         /* RAID level 4,5,6 require discard_zeroes_data for data integrity! */
1175         raid456 = (rs->md.level == 4 || rs->md.level == 5 || rs->md.level == 6);
1176
1177         for (i = 0; i < rs->md.raid_disks; i++) {
1178                 struct request_queue *q;
1179
1180                 if (!rs->dev[i].rdev.bdev)
1181                         continue;
1182
1183                 q = bdev_get_queue(rs->dev[i].rdev.bdev);
1184                 if (!q || !blk_queue_discard(q))
1185                         return;
1186
1187                 if (raid456) {
1188                         if (!q->limits.discard_zeroes_data)
1189                                 return;
1190                         if (!devices_handle_discard_safely) {
1191                                 DMERR("raid456 discard support disabled due to discard_zeroes_data uncertainty.");
1192                                 DMERR("Set dm-raid.devices_handle_discard_safely=Y to override.");
1193                                 return;
1194                         }
1195                 }
1196         }
1197
1198         /* All RAID members properly support discards */
1199         ti->discards_supported = true;
1200
1201         /*
1202          * RAID1 and RAID10 personalities require bio splitting,
1203          * RAID0/4/5/6 don't and process large discard bios properly.
1204          */
1205         ti->split_discard_bios = !!(rs->md.level == 1 || rs->md.level == 10);
1206         ti->num_discard_bios = 1;
1207 }
1208
1209 /*
1210  * Construct a RAID4/5/6 mapping:
1211  * Args:
1212  *      <raid_type> <#raid_params> <raid_params>                \
1213  *      <#raid_devs> { <meta_dev1> <dev1> .. <meta_devN> <devN> }
1214  *
1215  * <raid_params> varies by <raid_type>.  See 'parse_raid_params' for
1216  * details on possible <raid_params>.
1217  */
1218 static int raid_ctr(struct dm_target *ti, unsigned argc, char **argv)
1219 {
1220         int ret;
1221         struct raid_type *rt;
1222         unsigned long num_raid_params, num_raid_devs;
1223         struct raid_set *rs = NULL;
1224
1225         /* Must have at least <raid_type> <#raid_params> */
1226         if (argc < 2) {
1227                 ti->error = "Too few arguments";
1228                 return -EINVAL;
1229         }
1230
1231         /* raid type */
1232         rt = get_raid_type(argv[0]);
1233         if (!rt) {
1234                 ti->error = "Unrecognised raid_type";
1235                 return -EINVAL;
1236         }
1237         argc--;
1238         argv++;
1239
1240         /* number of RAID parameters */
1241         if (kstrtoul(argv[0], 10, &num_raid_params) < 0) {
1242                 ti->error = "Cannot understand number of RAID parameters";
1243                 return -EINVAL;
1244         }
1245         argc--;
1246         argv++;
1247
1248         /* Skip over RAID params for now and find out # of devices */
1249         if (num_raid_params >= argc) {
1250                 ti->error = "Arguments do not agree with counts given";
1251                 return -EINVAL;
1252         }
1253
1254         if ((kstrtoul(argv[num_raid_params], 10, &num_raid_devs) < 0) ||
1255             (num_raid_devs > MAX_RAID_DEVICES)) {
1256                 ti->error = "Cannot understand number of raid devices";
1257                 return -EINVAL;
1258         }
1259
1260         argc -= num_raid_params + 1; /* +1: we already have num_raid_devs */
1261         if (argc != (num_raid_devs * 2)) {
1262                 ti->error = "Supplied RAID devices does not match the count given";
1263                 return -EINVAL;
1264         }
1265
1266         rs = context_alloc(ti, rt, (unsigned)num_raid_devs);
1267         if (IS_ERR(rs))
1268                 return PTR_ERR(rs);
1269
1270         ret = parse_raid_params(rs, argv, (unsigned)num_raid_params);
1271         if (ret)
1272                 goto bad;
1273
1274         argv += num_raid_params + 1;
1275
1276         ret = dev_parms(rs, argv);
1277         if (ret)
1278                 goto bad;
1279
1280         rs->md.sync_super = super_sync;
1281         ret = analyse_superblocks(ti, rs);
1282         if (ret)
1283                 goto bad;
1284
1285         INIT_WORK(&rs->md.event_work, do_table_event);
1286         ti->private = rs;
1287         ti->num_flush_bios = 1;
1288
1289         /*
1290          * Disable/enable discard support on RAID set.
1291          */
1292         configure_discard_support(ti, rs);
1293
1294         /* Has to be held on running the array */
1295         mddev_lock_nointr(&rs->md);
1296         ret = md_run(&rs->md);
1297         rs->md.in_sync = 0; /* Assume already marked dirty */
1298         mddev_unlock(&rs->md);
1299
1300         if (ret) {
1301                 ti->error = "Fail to run raid array";
1302                 goto bad;
1303         }
1304
1305         if (ti->len != rs->md.array_sectors) {
1306                 ti->error = "Array size does not match requested target length";
1307                 ret = -EINVAL;
1308                 goto size_mismatch;
1309         }
1310         rs->callbacks.congested_fn = raid_is_congested;
1311         dm_table_add_target_callbacks(ti->table, &rs->callbacks);
1312
1313         mddev_suspend(&rs->md);
1314         return 0;
1315
1316 size_mismatch:
1317         md_stop(&rs->md);
1318 bad:
1319         context_free(rs);
1320
1321         return ret;
1322 }
1323
1324 static void raid_dtr(struct dm_target *ti)
1325 {
1326         struct raid_set *rs = ti->private;
1327
1328         list_del_init(&rs->callbacks.list);
1329         md_stop(&rs->md);
1330         context_free(rs);
1331 }
1332
1333 static int raid_map(struct dm_target *ti, struct bio *bio)
1334 {
1335         struct raid_set *rs = ti->private;
1336         struct mddev *mddev = &rs->md;
1337
1338         mddev->pers->make_request(mddev, bio);
1339
1340         return DM_MAPIO_SUBMITTED;
1341 }
1342
1343 static const char *decipher_sync_action(struct mddev *mddev)
1344 {
1345         if (test_bit(MD_RECOVERY_FROZEN, &mddev->recovery))
1346                 return "frozen";
1347
1348         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
1349             (!mddev->ro && test_bit(MD_RECOVERY_NEEDED, &mddev->recovery))) {
1350                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1351                         return "reshape";
1352
1353                 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1354                         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1355                                 return "resync";
1356                         else if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1357                                 return "check";
1358                         return "repair";
1359                 }
1360
1361                 if (test_bit(MD_RECOVERY_RECOVER, &mddev->recovery))
1362                         return "recover";
1363         }
1364
1365         return "idle";
1366 }
1367
1368 static void raid_status(struct dm_target *ti, status_type_t type,
1369                         unsigned status_flags, char *result, unsigned maxlen)
1370 {
1371         struct raid_set *rs = ti->private;
1372         unsigned raid_param_cnt = 1; /* at least 1 for chunksize */
1373         unsigned sz = 0;
1374         int i, array_in_sync = 0;
1375         sector_t sync;
1376
1377         switch (type) {
1378         case STATUSTYPE_INFO:
1379                 DMEMIT("%s %d ", rs->raid_type->name, rs->md.raid_disks);
1380
1381                 if (rs->raid_type->level) {
1382                         if (test_bit(MD_RECOVERY_RUNNING, &rs->md.recovery))
1383                                 sync = rs->md.curr_resync_completed;
1384                         else
1385                                 sync = rs->md.recovery_cp;
1386
1387                         if (sync >= rs->md.resync_max_sectors) {
1388                                 /*
1389                                  * Sync complete.
1390                                  */
1391                                 array_in_sync = 1;
1392                                 sync = rs->md.resync_max_sectors;
1393                         } else if (test_bit(MD_RECOVERY_REQUESTED, &rs->md.recovery)) {
1394                                 /*
1395                                  * If "check" or "repair" is occurring, the array has
1396                                  * undergone and initial sync and the health characters
1397                                  * should not be 'a' anymore.
1398                                  */
1399                                 array_in_sync = 1;
1400                         } else {
1401                                 /*
1402                                  * The array may be doing an initial sync, or it may
1403                                  * be rebuilding individual components.  If all the
1404                                  * devices are In_sync, then it is the array that is
1405                                  * being initialized.
1406                                  */
1407                                 for (i = 0; i < rs->md.raid_disks; i++)
1408                                         if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
1409                                                 array_in_sync = 1;
1410                         }
1411                 } else {
1412                         /* RAID0 */
1413                         array_in_sync = 1;
1414                         sync = rs->md.resync_max_sectors;
1415                 }
1416
1417                 /*
1418                  * Status characters:
1419                  *  'D' = Dead/Failed device
1420                  *  'a' = Alive but not in-sync
1421                  *  'A' = Alive and in-sync
1422                  */
1423                 for (i = 0; i < rs->md.raid_disks; i++) {
1424                         if (test_bit(Faulty, &rs->dev[i].rdev.flags))
1425                                 DMEMIT("D");
1426                         else if (!array_in_sync ||
1427                                  !test_bit(In_sync, &rs->dev[i].rdev.flags))
1428                                 DMEMIT("a");
1429                         else
1430                                 DMEMIT("A");
1431                 }
1432
1433                 /*
1434                  * In-sync ratio:
1435                  *  The in-sync ratio shows the progress of:
1436                  *   - Initializing the array
1437                  *   - Rebuilding a subset of devices of the array
1438                  *  The user can distinguish between the two by referring
1439                  *  to the status characters.
1440                  */
1441                 DMEMIT(" %llu/%llu",
1442                        (unsigned long long) sync,
1443                        (unsigned long long) rs->md.resync_max_sectors);
1444
1445                 /*
1446                  * Sync action:
1447                  *   See Documentation/device-mapper/dm-raid.c for
1448                  *   information on each of these states.
1449                  */
1450                 DMEMIT(" %s", decipher_sync_action(&rs->md));
1451
1452                 /*
1453                  * resync_mismatches/mismatch_cnt
1454                  *   This field shows the number of discrepancies found when
1455                  *   performing a "check" of the array.
1456                  */
1457                 DMEMIT(" %llu",
1458                        (strcmp(rs->md.last_sync_action, "check")) ? 0 :
1459                        (unsigned long long)
1460                        atomic64_read(&rs->md.resync_mismatches));
1461                 break;
1462         case STATUSTYPE_TABLE:
1463                 /* The string you would use to construct this array */
1464                 for (i = 0; i < rs->md.raid_disks; i++) {
1465                         if ((rs->ctr_flags & CTR_FLAG_REBUILD) &&
1466                             rs->dev[i].data_dev &&
1467                             !test_bit(In_sync, &rs->dev[i].rdev.flags))
1468                                 raid_param_cnt += 2; /* for rebuilds */
1469                         if (rs->dev[i].data_dev &&
1470                             test_bit(WriteMostly, &rs->dev[i].rdev.flags))
1471                                 raid_param_cnt += 2;
1472                 }
1473
1474                 raid_param_cnt += (hweight32(rs->ctr_flags & ~CTR_FLAG_REBUILD) * 2);
1475                 if (rs->ctr_flags & (CTR_FLAG_SYNC | CTR_FLAG_NOSYNC))
1476                         raid_param_cnt--;
1477
1478                 DMEMIT("%s %u %u", rs->raid_type->name,
1479                        raid_param_cnt, rs->md.chunk_sectors);
1480
1481                 if ((rs->ctr_flags & CTR_FLAG_SYNC) &&
1482                     (rs->md.recovery_cp == MaxSector))
1483                         DMEMIT(" sync");
1484                 if (rs->ctr_flags & CTR_FLAG_NOSYNC)
1485                         DMEMIT(" nosync");
1486
1487                 for (i = 0; i < rs->md.raid_disks; i++)
1488                         if ((rs->ctr_flags & CTR_FLAG_REBUILD) &&
1489                             rs->dev[i].data_dev &&
1490                             !test_bit(In_sync, &rs->dev[i].rdev.flags))
1491                                 DMEMIT(" rebuild %u", i);
1492
1493                 if (rs->ctr_flags & CTR_FLAG_DAEMON_SLEEP)
1494                         DMEMIT(" daemon_sleep %lu",
1495                                rs->md.bitmap_info.daemon_sleep);
1496
1497                 if (rs->ctr_flags & CTR_FLAG_MIN_RECOVERY_RATE)
1498                         DMEMIT(" min_recovery_rate %d", rs->md.sync_speed_min);
1499
1500                 if (rs->ctr_flags & CTR_FLAG_MAX_RECOVERY_RATE)
1501                         DMEMIT(" max_recovery_rate %d", rs->md.sync_speed_max);
1502
1503                 for (i = 0; i < rs->md.raid_disks; i++)
1504                         if (rs->dev[i].data_dev &&
1505                             test_bit(WriteMostly, &rs->dev[i].rdev.flags))
1506                                 DMEMIT(" write_mostly %u", i);
1507
1508                 if (rs->ctr_flags & CTR_FLAG_MAX_WRITE_BEHIND)
1509                         DMEMIT(" max_write_behind %lu",
1510                                rs->md.bitmap_info.max_write_behind);
1511
1512                 if (rs->ctr_flags & CTR_FLAG_STRIPE_CACHE) {
1513                         struct r5conf *conf = rs->md.private;
1514
1515                         /* convert from kiB to sectors */
1516                         DMEMIT(" stripe_cache %d",
1517                                conf ? conf->max_nr_stripes * 2 : 0);
1518                 }
1519
1520                 if (rs->ctr_flags & CTR_FLAG_REGION_SIZE)
1521                         DMEMIT(" region_size %lu",
1522                                rs->md.bitmap_info.chunksize >> 9);
1523
1524                 if (rs->ctr_flags & CTR_FLAG_RAID10_COPIES)
1525                         DMEMIT(" raid10_copies %u",
1526                                raid10_md_layout_to_copies(rs->md.layout));
1527
1528                 if (rs->ctr_flags & CTR_FLAG_RAID10_FORMAT)
1529                         DMEMIT(" raid10_format %s",
1530                                raid10_md_layout_to_format(rs->md.layout));
1531
1532                 DMEMIT(" %d", rs->md.raid_disks);
1533                 for (i = 0; i < rs->md.raid_disks; i++) {
1534                         if (rs->dev[i].meta_dev)
1535                                 DMEMIT(" %s", rs->dev[i].meta_dev->name);
1536                         else
1537                                 DMEMIT(" -");
1538
1539                         if (rs->dev[i].data_dev)
1540                                 DMEMIT(" %s", rs->dev[i].data_dev->name);
1541                         else
1542                                 DMEMIT(" -");
1543                 }
1544         }
1545 }
1546
1547 static int raid_message(struct dm_target *ti, unsigned argc, char **argv)
1548 {
1549         struct raid_set *rs = ti->private;
1550         struct mddev *mddev = &rs->md;
1551
1552         if (!strcasecmp(argv[0], "reshape")) {
1553                 DMERR("Reshape not supported.");
1554                 return -EINVAL;
1555         }
1556
1557         if (!mddev->pers || !mddev->pers->sync_request)
1558                 return -EINVAL;
1559
1560         if (!strcasecmp(argv[0], "frozen"))
1561                 set_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
1562         else
1563                 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
1564
1565         if (!strcasecmp(argv[0], "idle") || !strcasecmp(argv[0], "frozen")) {
1566                 if (mddev->sync_thread) {
1567                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1568                         md_reap_sync_thread(mddev);
1569                 }
1570         } else if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
1571                    test_bit(MD_RECOVERY_NEEDED, &mddev->recovery))
1572                 return -EBUSY;
1573         else if (!strcasecmp(argv[0], "resync"))
1574                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1575         else if (!strcasecmp(argv[0], "recover")) {
1576                 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
1577                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1578         } else {
1579                 if (!strcasecmp(argv[0], "check"))
1580                         set_bit(MD_RECOVERY_CHECK, &mddev->recovery);
1581                 else if (!!strcasecmp(argv[0], "repair"))
1582                         return -EINVAL;
1583                 set_bit(MD_RECOVERY_REQUESTED, &mddev->recovery);
1584                 set_bit(MD_RECOVERY_SYNC, &mddev->recovery);
1585         }
1586         if (mddev->ro == 2) {
1587                 /* A write to sync_action is enough to justify
1588                  * canceling read-auto mode
1589                  */
1590                 mddev->ro = 0;
1591                 if (!mddev->suspended)
1592                         md_wakeup_thread(mddev->sync_thread);
1593         }
1594         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1595         if (!mddev->suspended)
1596                 md_wakeup_thread(mddev->thread);
1597
1598         return 0;
1599 }
1600
1601 static int raid_iterate_devices(struct dm_target *ti,
1602                                 iterate_devices_callout_fn fn, void *data)
1603 {
1604         struct raid_set *rs = ti->private;
1605         unsigned i;
1606         int ret = 0;
1607
1608         for (i = 0; !ret && i < rs->md.raid_disks; i++)
1609                 if (rs->dev[i].data_dev)
1610                         ret = fn(ti,
1611                                  rs->dev[i].data_dev,
1612                                  0, /* No offset on data devs */
1613                                  rs->md.dev_sectors,
1614                                  data);
1615
1616         return ret;
1617 }
1618
1619 static void raid_io_hints(struct dm_target *ti, struct queue_limits *limits)
1620 {
1621         struct raid_set *rs = ti->private;
1622         unsigned chunk_size = rs->md.chunk_sectors << 9;
1623         struct r5conf *conf = rs->md.private;
1624
1625         blk_limits_io_min(limits, chunk_size);
1626         blk_limits_io_opt(limits, chunk_size * (conf->raid_disks - conf->max_degraded));
1627 }
1628
1629 static void raid_presuspend(struct dm_target *ti)
1630 {
1631         struct raid_set *rs = ti->private;
1632
1633         md_stop_writes(&rs->md);
1634 }
1635
1636 static void raid_postsuspend(struct dm_target *ti)
1637 {
1638         struct raid_set *rs = ti->private;
1639
1640         mddev_suspend(&rs->md);
1641 }
1642
1643 static void attempt_restore_of_faulty_devices(struct raid_set *rs)
1644 {
1645         int i;
1646         uint64_t failed_devices, cleared_failed_devices = 0;
1647         unsigned long flags;
1648         struct dm_raid_superblock *sb;
1649         struct md_rdev *r;
1650
1651         for (i = 0; i < rs->md.raid_disks; i++) {
1652                 r = &rs->dev[i].rdev;
1653                 if (test_bit(Faulty, &r->flags) && r->sb_page &&
1654                     sync_page_io(r, 0, r->sb_size, r->sb_page, REQ_OP_READ, 0,
1655                                  1)) {
1656                         DMINFO("Faulty %s device #%d has readable super block."
1657                                "  Attempting to revive it.",
1658                                rs->raid_type->name, i);
1659
1660                         /*
1661                          * Faulty bit may be set, but sometimes the array can
1662                          * be suspended before the personalities can respond
1663                          * by removing the device from the array (i.e. calling
1664                          * 'hot_remove_disk').  If they haven't yet removed
1665                          * the failed device, its 'raid_disk' number will be
1666                          * '>= 0' - meaning we must call this function
1667                          * ourselves.
1668                          */
1669                         if ((r->raid_disk >= 0) &&
1670                             (r->mddev->pers->hot_remove_disk(r->mddev, r) != 0))
1671                                 /* Failed to revive this device, try next */
1672                                 continue;
1673
1674                         r->raid_disk = i;
1675                         r->saved_raid_disk = i;
1676                         flags = r->flags;
1677                         clear_bit(Faulty, &r->flags);
1678                         clear_bit(WriteErrorSeen, &r->flags);
1679                         clear_bit(In_sync, &r->flags);
1680                         if (r->mddev->pers->hot_add_disk(r->mddev, r)) {
1681                                 r->raid_disk = -1;
1682                                 r->saved_raid_disk = -1;
1683                                 r->flags = flags;
1684                         } else {
1685                                 r->recovery_offset = 0;
1686                                 cleared_failed_devices |= 1 << i;
1687                         }
1688                 }
1689         }
1690         if (cleared_failed_devices) {
1691                 rdev_for_each(r, &rs->md) {
1692                         sb = page_address(r->sb_page);
1693                         failed_devices = le64_to_cpu(sb->failed_devices);
1694                         failed_devices &= ~cleared_failed_devices;
1695                         sb->failed_devices = cpu_to_le64(failed_devices);
1696                 }
1697         }
1698 }
1699
1700 static void raid_resume(struct dm_target *ti)
1701 {
1702         struct raid_set *rs = ti->private;
1703
1704         if (rs->raid_type->level) {
1705                 set_bit(MD_CHANGE_DEVS, &rs->md.flags);
1706
1707                 if (!rs->bitmap_loaded) {
1708                         bitmap_load(&rs->md);
1709                         rs->bitmap_loaded = 1;
1710                 } else {
1711                         /*
1712                          * A secondary resume while the device is active.
1713                          * Take this opportunity to check whether any failed
1714                          * devices are reachable again.
1715                          */
1716                         attempt_restore_of_faulty_devices(rs);
1717                 }
1718
1719                 clear_bit(MD_RECOVERY_FROZEN, &rs->md.recovery);
1720         }
1721
1722         mddev_resume(&rs->md);
1723 }
1724
1725 static struct target_type raid_target = {
1726         .name = "raid",
1727         .version = {1, 8, 0},
1728         .module = THIS_MODULE,
1729         .ctr = raid_ctr,
1730         .dtr = raid_dtr,
1731         .map = raid_map,
1732         .status = raid_status,
1733         .message = raid_message,
1734         .iterate_devices = raid_iterate_devices,
1735         .io_hints = raid_io_hints,
1736         .presuspend = raid_presuspend,
1737         .postsuspend = raid_postsuspend,
1738         .resume = raid_resume,
1739 };
1740
1741 static int __init dm_raid_init(void)
1742 {
1743         DMINFO("Loading target version %u.%u.%u",
1744                raid_target.version[0],
1745                raid_target.version[1],
1746                raid_target.version[2]);
1747         return dm_register_target(&raid_target);
1748 }
1749
1750 static void __exit dm_raid_exit(void)
1751 {
1752         dm_unregister_target(&raid_target);
1753 }
1754
1755 module_init(dm_raid_init);
1756 module_exit(dm_raid_exit);
1757
1758 module_param(devices_handle_discard_safely, bool, 0644);
1759 MODULE_PARM_DESC(devices_handle_discard_safely,
1760                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
1761
1762 MODULE_DESCRIPTION(DM_NAME " raid4/5/6 target");
1763 MODULE_ALIAS("dm-raid1");
1764 MODULE_ALIAS("dm-raid10");
1765 MODULE_ALIAS("dm-raid4");
1766 MODULE_ALIAS("dm-raid5");
1767 MODULE_ALIAS("dm-raid6");
1768 MODULE_AUTHOR("Neil Brown <dm-devel@redhat.com>");
1769 MODULE_LICENSE("GPL");