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