drivers/md/dm-raid.c

   1 /*
   2  * Copyright (C) 2010-2011 Neil Brown
   3  * Copyright (C) 2010-2014 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
  21 static bool devices_handle_discard_safely = false;
  22
  23 /*
  24  * The following flags are used by dm-raid.c to set up the array state.
  25  * They must be cleared before md_run is called.
  26  */
  27 #define FirstUse 10             /* rdev flag */
  28
  29 struct raid_dev {
  30         /*
  31          * Two DM devices, one to hold metadata and one to hold the
  32          * actual data/parity.  The reason for this is to not confuse
  33          * ti->len and give more flexibility in altering size and
  34          * characteristics.
  35          *
  36          * While it is possible for this device to be associated
  37          * with a different physical device than the data_dev, it
  38          * is intended for it to be the same.
  39          *    |--------- Physical Device ---------|
  40          *    |- meta_dev -|------ data_dev ------|
  41          */
  42         struct dm_dev *meta_dev;
  43         struct dm_dev *data_dev;
  44         struct md_rdev rdev;
  45 };
  46
  47 /*
  48  * Flags for rs->print_flags field.
  49  */
  50 #define DMPF_SYNC              0x1
  51 #define DMPF_NOSYNC            0x2
  52 #define DMPF_REBUILD           0x4
  53 #define DMPF_DAEMON_SLEEP      0x8
  54 #define DMPF_MIN_RECOVERY_RATE 0x10
  55 #define DMPF_MAX_RECOVERY_RATE 0x20
  56 #define DMPF_MAX_WRITE_BEHIND  0x40
  57 #define DMPF_STRIPE_CACHE      0x80
  58 #define DMPF_REGION_SIZE       0x100
  59 #define DMPF_RAID10_COPIES     0x200
  60 #define DMPF_RAID10_FORMAT     0x400
  61
  62 struct raid_set {
  63         struct dm_target *ti;
  64
  65         uint32_t bitmap_loaded;
  66         uint32_t print_flags;
  67
  68         struct mddev md;
  69         struct raid_type *raid_type;
  70         struct dm_target_callbacks callbacks;
  71
  72         struct raid_dev dev[0];
  73 };
  74
  75 /* Supported raid types and properties. */
  76 static struct raid_type {
  77         const char *name;               /* RAID algorithm. */
  78         const char *descr;              /* Descriptor text for logging. */
  79         const unsigned parity_devs;     /* # of parity devices. */
  80         const unsigned minimal_devs;    /* minimal # of devices in set. */
  81         const unsigned level;           /* RAID level. */
  82         const unsigned algorithm;       /* RAID algorithm. */
  83 } raid_types[] = {
  84         {"raid1",    "RAID1 (mirroring)",               0, 2, 1, 0 /* NONE */},
  85         {"raid10",   "RAID10 (striped mirrors)",        0, 2, 10, UINT_MAX /* Varies */},
  86         {"raid4",    "RAID4 (dedicated parity disk)",   1, 2, 5, ALGORITHM_PARITY_0},
  87         {"raid5_la", "RAID5 (left asymmetric)",         1, 2, 5, ALGORITHM_LEFT_ASYMMETRIC},
  88         {"raid5_ra", "RAID5 (right asymmetric)",        1, 2, 5, ALGORITHM_RIGHT_ASYMMETRIC},
  89         {"raid5_ls", "RAID5 (left symmetric)",          1, 2, 5, ALGORITHM_LEFT_SYMMETRIC},
  90         {"raid5_rs", "RAID5 (right symmetric)",         1, 2, 5, ALGORITHM_RIGHT_SYMMETRIC},
  91         {"raid6_zr", "RAID6 (zero restart)",            2, 4, 6, ALGORITHM_ROTATING_ZERO_RESTART},
  92         {"raid6_nr", "RAID6 (N restart)",               2, 4, 6, ALGORITHM_ROTATING_N_RESTART},
  93         {"raid6_nc", "RAID6 (N continue)",              2, 4, 6, ALGORITHM_ROTATING_N_CONTINUE}
  94 };
  95
  96 static char *raid10_md_layout_to_format(int layout)
  97 {
  98         /*
  99          * Bit 16 and 17 stand for "offset" and "use_far_sets"
 100          * Refer to MD's raid10.c for details
 101          */
 102         if ((layout & 0x10000) && (layout & 0x20000))
 103                 return "offset";
 104
 105         if ((layout & 0xFF) > 1)
 106                 return "near";
 107
 108         return "far";
 109 }
 110
 111 static unsigned raid10_md_layout_to_copies(int layout)
 112 {
 113         if ((layout & 0xFF) > 1)
 114                 return layout & 0xFF;
 115         return (layout >> 8) & 0xFF;
 116 }
 117
 118 static int raid10_format_to_md_layout(char *format, unsigned copies)
 119 {
 120         unsigned n = 1, f = 1;
 121
 122         if (!strcmp("near", format))
 123                 n = copies;
 124         else
 125                 f = copies;
 126
 127         if (!strcmp("offset", format))
 128                 return 0x30000 | (f << 8) | n;
 129
 130         if (!strcmp("far", format))
 131                 return 0x20000 | (f << 8) | n;
 132
 133         return (f << 8) | n;
 134 }
 135
 136 static struct raid_type *get_raid_type(char *name)
 137 {
 138         int i;
 139
 140         for (i = 0; i < ARRAY_SIZE(raid_types); i++)
 141                 if (!strcmp(raid_types[i].name, name))
 142                         return &raid_types[i];
 143
 144         return NULL;
 145 }
 146
 147 static struct raid_set *context_alloc(struct dm_target *ti, struct raid_type *raid_type, unsigned raid_devs)
 148 {
 149         unsigned i;
 150         struct raid_set *rs;
 151
 152         if (raid_devs <= raid_type->parity_devs) {
 153                 ti->error = "Insufficient number of devices";
 154                 return ERR_PTR(-EINVAL);
 155         }
 156
 157         rs = kzalloc(sizeof(*rs) + raid_devs * sizeof(rs->dev[0]), GFP_KERNEL);
 158         if (!rs) {
 159                 ti->error = "Cannot allocate raid context";
 160                 return ERR_PTR(-ENOMEM);
 161         }
 162
 163         mddev_init(&rs->md);
 164
 165         rs->ti = ti;
 166         rs->raid_type = raid_type;
 167         rs->md.raid_disks = raid_devs;
 168         rs->md.level = raid_type->level;
 169         rs->md.new_level = rs->md.level;
 170         rs->md.layout = raid_type->algorithm;
 171         rs->md.new_layout = rs->md.layout;
 172         rs->md.delta_disks = 0;
 173         rs->md.recovery_cp = 0;
 174
 175         for (i = 0; i < raid_devs; i++)
 176                 md_rdev_init(&rs->dev[i].rdev);
 177
 178         /*
 179          * Remaining items to be initialized by further RAID params:
 180          *  rs->md.persistent
 181          *  rs->md.external
 182          *  rs->md.chunk_sectors
 183          *  rs->md.new_chunk_sectors
 184          *  rs->md.dev_sectors
 185          */
 186
 187         return rs;
 188 }
 189
 190 static void context_free(struct raid_set *rs)
 191 {
 192         int i;
 193
 194         for (i = 0; i < rs->md.raid_disks; i++) {
 195                 if (rs->dev[i].meta_dev)
 196                         dm_put_device(rs->ti, rs->dev[i].meta_dev);
 197                 md_rdev_clear(&rs->dev[i].rdev);
 198                 if (rs->dev[i].data_dev)
 199                         dm_put_device(rs->ti, rs->dev[i].data_dev);
 200         }
 201
 202         kfree(rs);
 203 }
 204
 205 /*
 206  * For every device we have two words
 207  *  <meta_dev>: meta device name or '-' if missing
 208  *  <data_dev>: data device name or '-' if missing
 209  *
 210  * The following are permitted:
 211  *    - -
 212  *    - <data_dev>
 213  *    <meta_dev> <data_dev>
 214  *
 215  * The following is not allowed:
 216  *    <meta_dev> -
 217  *
 218  * This code parses those words.  If there is a failure,
 219  * the caller must use context_free to unwind the operations.
 220  */
 221 static int dev_parms(struct raid_set *rs, char **argv)
 222 {
 223         int i;
 224         int rebuild = 0;
 225         int metadata_available = 0;
 226         int ret = 0;
 227
 228         for (i = 0; i < rs->md.raid_disks; i++, argv += 2) {
 229                 rs->dev[i].rdev.raid_disk = i;
 230
 231                 rs->dev[i].meta_dev = NULL;
 232                 rs->dev[i].data_dev = NULL;
 233
 234                 /*
 235                  * There are no offsets, since there is a separate device
 236                  * for data and metadata.
 237                  */
 238                 rs->dev[i].rdev.data_offset = 0;
 239                 rs->dev[i].rdev.mddev = &rs->md;
 240
 241                 if (strcmp(argv[0], "-")) {
 242                         ret = dm_get_device(rs->ti, argv[0],
 243                                             dm_table_get_mode(rs->ti->table),
 244                                             &rs->dev[i].meta_dev);
 245                         rs->ti->error = "RAID metadata device lookup failure";
 246                         if (ret)
 247                                 return ret;
 248
 249                         rs->dev[i].rdev.sb_page = alloc_page(GFP_KERNEL);
 250                         if (!rs->dev[i].rdev.sb_page)
 251                                 return -ENOMEM;
 252                 }
 253
 254                 if (!strcmp(argv[1], "-")) {
 255                         if (!test_bit(In_sync, &rs->dev[i].rdev.flags) &&
 256                             (!rs->dev[i].rdev.recovery_offset)) {
 257                                 rs->ti->error = "Drive designated for rebuild not specified";
 258                                 return -EINVAL;
 259                         }
 260
 261                         rs->ti->error = "No data device supplied with metadata device";
 262                         if (rs->dev[i].meta_dev)
 263                                 return -EINVAL;
 264
 265                         continue;
 266                 }
 267
 268                 ret = dm_get_device(rs->ti, argv[1],
 269                                     dm_table_get_mode(rs->ti->table),
 270                                     &rs->dev[i].data_dev);
 271                 if (ret) {
 272                         rs->ti->error = "RAID device lookup failure";
 273                         return ret;
 274                 }
 275
 276                 if (rs->dev[i].meta_dev) {
 277                         metadata_available = 1;
 278                         rs->dev[i].rdev.meta_bdev = rs->dev[i].meta_dev->bdev;
 279                 }
 280                 rs->dev[i].rdev.bdev = rs->dev[i].data_dev->bdev;
 281                 list_add(&rs->dev[i].rdev.same_set, &rs->md.disks);
 282                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
 283                         rebuild++;
 284         }
 285
 286         if (metadata_available) {
 287                 rs->md.external = 0;
 288                 rs->md.persistent = 1;
 289                 rs->md.major_version = 2;
 290         } else if (rebuild && !rs->md.recovery_cp) {
 291                 /*
 292                  * Without metadata, we will not be able to tell if the array
 293                  * is in-sync or not - we must assume it is not.  Therefore,
 294                  * it is impossible to rebuild a drive.
 295                  *
 296                  * Even if there is metadata, the on-disk information may
 297                  * indicate that the array is not in-sync and it will then
 298                  * fail at that time.
 299                  *
 300                  * User could specify 'nosync' option if desperate.
 301                  */
 302                 DMERR("Unable to rebuild drive while array is not in-sync");
 303                 rs->ti->error = "RAID device lookup failure";
 304                 return -EINVAL;
 305         }
 306
 307         return 0;
 308 }
 309
 310 /*
 311  * validate_region_size
 312  * @rs
 313  * @region_size:  region size in sectors.  If 0, pick a size (4MiB default).
 314  *
 315  * Set rs->md.bitmap_info.chunksize (which really refers to 'region size').
 316  * Ensure that (ti->len/region_size < 2^21) - required by MD bitmap.
 317  *
 318  * Returns: 0 on success, -EINVAL on failure.
 319  */
 320 static int validate_region_size(struct raid_set *rs, unsigned long region_size)
 321 {
 322         unsigned long min_region_size = rs->ti->len / (1 << 21);
 323
 324         if (!region_size) {
 325                 /*
 326                  * Choose a reasonable default.  All figures in sectors.
 327                  */
 328                 if (min_region_size > (1 << 13)) {
 329                         /* If not a power of 2, make it the next power of 2 */
 330                         if (min_region_size & (min_region_size - 1))
 331                                 region_size = 1 << fls(region_size);
 332                         DMINFO("Choosing default region size of %lu sectors",
 333                                region_size);
 334                 } else {
 335                         DMINFO("Choosing default region size of 4MiB");
 336                         region_size = 1 << 13; /* sectors */
 337                 }
 338         } else {
 339                 /*
 340                  * Validate user-supplied value.
 341                  */
 342                 if (region_size > rs->ti->len) {
 343                         rs->ti->error = "Supplied region size is too large";
 344                         return -EINVAL;
 345                 }
 346
 347                 if (region_size < min_region_size) {
 348                         DMERR("Supplied region_size (%lu sectors) below minimum (%lu)",
 349                               region_size, min_region_size);
 350                         rs->ti->error = "Supplied region size is too small";
 351                         return -EINVAL;
 352                 }
 353
 354                 if (!is_power_of_2(region_size)) {
 355                         rs->ti->error = "Region size is not a power of 2";
 356                         return -EINVAL;
 357                 }
 358
 359                 if (region_size < rs->md.chunk_sectors) {
 360                         rs->ti->error = "Region size is smaller than the chunk size";
 361                         return -EINVAL;
 362                 }
 363         }
 364
 365         /*
 366          * Convert sectors to bytes.
 367          */
 368         rs->md.bitmap_info.chunksize = (region_size << 9);
 369
 370         return 0;
 371 }
 372
 373 /*
 374  * validate_raid_redundancy
 375  * @rs
 376  *
 377  * Determine if there are enough devices in the array that haven't
 378  * failed (or are being rebuilt) to form a usable array.
 379  *
 380  * Returns: 0 on success, -EINVAL on failure.
 381  */
 382 static int validate_raid_redundancy(struct raid_set *rs)
 383 {
 384         unsigned i, rebuild_cnt = 0;
 385         unsigned rebuilds_per_group = 0, copies, d;
 386         unsigned group_size, last_group_start;
 387
 388         for (i = 0; i < rs->md.raid_disks; i++)
 389                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags) ||
 390                     !rs->dev[i].rdev.sb_page)
 391                         rebuild_cnt++;
 392
 393         switch (rs->raid_type->level) {
 394         case 1:
 395                 if (rebuild_cnt >= rs->md.raid_disks)
 396                         goto too_many;
 397                 break;
 398         case 4:
 399         case 5:
 400         case 6:
 401                 if (rebuild_cnt > rs->raid_type->parity_devs)
 402                         goto too_many;
 403                 break;
 404         case 10:
 405                 copies = raid10_md_layout_to_copies(rs->md.layout);
 406                 if (rebuild_cnt < copies)
 407                         break;
 408
 409                 /*
 410                  * It is possible to have a higher rebuild count for RAID10,
 411                  * as long as the failed devices occur in different mirror
 412                  * groups (i.e. different stripes).
 413                  *
 414                  * When checking "near" format, make sure no adjacent devices
 415                  * have failed beyond what can be handled.  In addition to the
 416                  * simple case where the number of devices is a multiple of the
 417                  * number of copies, we must also handle cases where the number
 418                  * of devices is not a multiple of the number of copies.
 419                  * E.g.    dev1 dev2 dev3 dev4 dev5
 420                  *          A    A    B    B    C
 421                  *          C    D    D    E    E
 422                  */
 423                 if (!strcmp("near", raid10_md_layout_to_format(rs->md.layout))) {
 424                         for (i = 0; i < rs->md.raid_disks * copies; i++) {
 425                                 if (!(i % copies))
 426                                         rebuilds_per_group = 0;
 427                                 d = i % rs->md.raid_disks;
 428                                 if ((!rs->dev[d].rdev.sb_page ||
 429                                      !test_bit(In_sync, &rs->dev[d].rdev.flags)) &&
 430                                     (++rebuilds_per_group >= copies))
 431                                         goto too_many;
 432                         }
 433                         break;
 434                 }
 435
 436                 /*
 437                  * When checking "far" and "offset" formats, we need to ensure
 438                  * that the device that holds its copy is not also dead or
 439                  * being rebuilt.  (Note that "far" and "offset" formats only
 440                  * support two copies right now.  These formats also only ever
 441                  * use the 'use_far_sets' variant.)
 442                  *
 443                  * This check is somewhat complicated by the need to account
 444                  * for arrays that are not a multiple of (far) copies.  This
 445                  * results in the need to treat the last (potentially larger)
 446                  * set differently.
 447                  */
 448                 group_size = (rs->md.raid_disks / copies);
 449                 last_group_start = (rs->md.raid_disks / group_size) - 1;
 450                 last_group_start *= group_size;
 451                 for (i = 0; i < rs->md.raid_disks; i++) {
 452                         if (!(i % copies) && !(i > last_group_start))
 453                                 rebuilds_per_group = 0;
 454                         if ((!rs->dev[i].rdev.sb_page ||
 455                              !test_bit(In_sync, &rs->dev[i].rdev.flags)) &&
 456                             (++rebuilds_per_group >= copies))
 457                                         goto too_many;
 458                 }
 459                 break;
 460         default:
 461                 if (rebuild_cnt)
 462                         return -EINVAL;
 463         }
 464
 465         return 0;
 466
 467 too_many:
 468         return -EINVAL;
 469 }
 470
 471 /*
 472  * Possible arguments are...
 473  *      <chunk_size> [optional_args]
 474  *
 475  * Argument definitions
 476  *    <chunk_size>                      The number of sectors per disk that
 477  *                                      will form the "stripe"
 478  *    [[no]sync]                        Force or prevent recovery of the
 479  *                                      entire array
 480  *    [devices_handle_discard_safely]   Allow discards on RAID4/5/6; useful if RAID
 481  *                                      member device(s) properly support TRIM/UNMAP
 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->print_flags |= DMPF_NOSYNC;
 559                         continue;
 560                 }
 561                 if (!strcasecmp(argv[i], "sync")) {
 562                         rs->md.recovery_cp = 0;
 563                         rs->print_flags |= DMPF_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->print_flags |= DMPF_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->print_flags |= DMPF_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->print_flags |= DMPF_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->print_flags |= DMPF_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->print_flags |= DMPF_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->print_flags |= DMPF_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->print_flags |= DMPF_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->print_flags |= DMPF_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->print_flags |= DMPF_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 > 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         if (rs->raid_type->level == 1)
 750                 return md_raid1_congested(&rs->md, bits);
 751
 752         if (rs->raid_type->level == 10)
 753                 return md_raid10_congested(&rs->md, bits);
 754
 755         return md_raid5_congested(&rs->md, bits);
 756 }
 757
 758 /*
 759  * This structure is never routinely used by userspace, unlike md superblocks.
 760  * Devices with this superblock should only ever be accessed via device-mapper.
 761  */
 762 #define DM_RAID_MAGIC 0x64526D44
 763 struct dm_raid_superblock {
 764         __le32 magic;           /* "DmRd" */
 765         __le32 features;        /* Used to indicate possible future changes */
 766
 767         __le32 num_devices;     /* Number of devices in this array. (Max 64) */
 768         __le32 array_position;  /* The position of this drive in the array */
 769
 770         __le64 events;          /* Incremented by md when superblock updated */
 771         __le64 failed_devices;  /* Bit field of devices to indicate failures */
 772
 773         /*
 774          * This offset tracks the progress of the repair or replacement of
 775          * an individual drive.
 776          */
 777         __le64 disk_recovery_offset;
 778
 779         /*
 780          * This offset tracks the progress of the initial array
 781          * synchronisation/parity calculation.
 782          */
 783         __le64 array_resync_offset;
 784
 785         /*
 786          * RAID characteristics
 787          */
 788         __le32 level;
 789         __le32 layout;
 790         __le32 stripe_sectors;
 791
 792         __u8 pad[452];          /* Round struct to 512 bytes. */
 793                                 /* Always set to 0 when writing. */
 794 } __packed;
 795
 796 static int read_disk_sb(struct md_rdev *rdev, int size)
 797 {
 798         BUG_ON(!rdev->sb_page);
 799
 800         if (rdev->sb_loaded)
 801                 return 0;
 802
 803         if (!sync_page_io(rdev, 0, size, rdev->sb_page, READ, 1)) {
 804                 DMERR("Failed to read superblock of device at position %d",
 805                       rdev->raid_disk);
 806                 md_error(rdev->mddev, rdev);
 807                 return -EINVAL;
 808         }
 809
 810         rdev->sb_loaded = 1;
 811
 812         return 0;
 813 }
 814
 815 static void super_sync(struct mddev *mddev, struct md_rdev *rdev)
 816 {
 817         int i;
 818         uint64_t failed_devices;
 819         struct dm_raid_superblock *sb;
 820         struct raid_set *rs = container_of(mddev, struct raid_set, md);
 821
 822         sb = page_address(rdev->sb_page);
 823         failed_devices = le64_to_cpu(sb->failed_devices);
 824
 825         for (i = 0; i < mddev->raid_disks; i++)
 826                 if (!rs->dev[i].data_dev ||
 827                     test_bit(Faulty, &(rs->dev[i].rdev.flags)))
 828                         failed_devices |= (1ULL << i);
 829
 830         memset(sb, 0, sizeof(*sb));
 831
 832         sb->magic = cpu_to_le32(DM_RAID_MAGIC);
 833         sb->features = cpu_to_le32(0);  /* No features yet */
 834
 835         sb->num_devices = cpu_to_le32(mddev->raid_disks);
 836         sb->array_position = cpu_to_le32(rdev->raid_disk);
 837
 838         sb->events = cpu_to_le64(mddev->events);
 839         sb->failed_devices = cpu_to_le64(failed_devices);
 840
 841         sb->disk_recovery_offset = cpu_to_le64(rdev->recovery_offset);
 842         sb->array_resync_offset = cpu_to_le64(mddev->recovery_cp);
 843
 844         sb->level = cpu_to_le32(mddev->level);
 845         sb->layout = cpu_to_le32(mddev->layout);
 846         sb->stripe_sectors = cpu_to_le32(mddev->chunk_sectors);
 847 }
 848
 849 /*
 850  * super_load
 851  *
 852  * This function creates a superblock if one is not found on the device
 853  * and will decide which superblock to use if there's a choice.
 854  *
 855  * Return: 1 if use rdev, 0 if use refdev, -Exxx otherwise
 856  */
 857 static int super_load(struct md_rdev *rdev, struct md_rdev *refdev)
 858 {
 859         int ret;
 860         struct dm_raid_superblock *sb;
 861         struct dm_raid_superblock *refsb;
 862         uint64_t events_sb, events_refsb;
 863
 864         rdev->sb_start = 0;
 865         rdev->sb_size = sizeof(*sb);
 866
 867         ret = read_disk_sb(rdev, rdev->sb_size);
 868         if (ret)
 869                 return ret;
 870
 871         sb = page_address(rdev->sb_page);
 872
 873         /*
 874          * Two cases that we want to write new superblocks and rebuild:
 875          * 1) New device (no matching magic number)
 876          * 2) Device specified for rebuild (!In_sync w/ offset == 0)
 877          */
 878         if ((sb->magic != cpu_to_le32(DM_RAID_MAGIC)) ||
 879             (!test_bit(In_sync, &rdev->flags) && !rdev->recovery_offset)) {
 880                 super_sync(rdev->mddev, rdev);
 881
 882                 set_bit(FirstUse, &rdev->flags);
 883
 884                 /* Force writing of superblocks to disk */
 885                 set_bit(MD_CHANGE_DEVS, &rdev->mddev->flags);
 886
 887                 /* Any superblock is better than none, choose that if given */
 888                 return refdev ? 0 : 1;
 889         }
 890
 891         if (!refdev)
 892                 return 1;
 893
 894         events_sb = le64_to_cpu(sb->events);
 895
 896         refsb = page_address(refdev->sb_page);
 897         events_refsb = le64_to_cpu(refsb->events);
 898
 899         return (events_sb > events_refsb) ? 1 : 0;
 900 }
 901
 902 static int super_init_validation(struct mddev *mddev, struct md_rdev *rdev)
 903 {
 904         int role;
 905         struct raid_set *rs = container_of(mddev, struct raid_set, md);
 906         uint64_t events_sb;
 907         uint64_t failed_devices;
 908         struct dm_raid_superblock *sb;
 909         uint32_t new_devs = 0;
 910         uint32_t rebuilds = 0;
 911         struct md_rdev *r;
 912         struct dm_raid_superblock *sb2;
 913
 914         sb = page_address(rdev->sb_page);
 915         events_sb = le64_to_cpu(sb->events);
 916         failed_devices = le64_to_cpu(sb->failed_devices);
 917
 918         /*
 919          * Initialise to 1 if this is a new superblock.
 920          */
 921         mddev->events = events_sb ? : 1;
 922
 923         /*
 924          * Reshaping is not currently allowed
 925          */
 926         if (le32_to_cpu(sb->level) != mddev->level) {
 927                 DMERR("Reshaping arrays not yet supported. (RAID level change)");
 928                 return -EINVAL;
 929         }
 930         if (le32_to_cpu(sb->layout) != mddev->layout) {
 931                 DMERR("Reshaping arrays not yet supported. (RAID layout change)");
 932                 DMERR("  0x%X vs 0x%X", le32_to_cpu(sb->layout), mddev->layout);
 933                 DMERR("  Old layout: %s w/ %d copies",
 934                       raid10_md_layout_to_format(le32_to_cpu(sb->layout)),
 935                       raid10_md_layout_to_copies(le32_to_cpu(sb->layout)));
 936                 DMERR("  New layout: %s w/ %d copies",
 937                       raid10_md_layout_to_format(mddev->layout),
 938                       raid10_md_layout_to_copies(mddev->layout));
 939                 return -EINVAL;
 940         }
 941         if (le32_to_cpu(sb->stripe_sectors) != mddev->chunk_sectors) {
 942                 DMERR("Reshaping arrays not yet supported. (stripe sectors change)");
 943                 return -EINVAL;
 944         }
 945
 946         /* We can only change the number of devices in RAID1 right now */
 947         if ((rs->raid_type->level != 1) &&
 948             (le32_to_cpu(sb->num_devices) != mddev->raid_disks)) {
 949                 DMERR("Reshaping arrays not yet supported. (device count change)");
 950                 return -EINVAL;
 951         }
 952
 953         if (!(rs->print_flags & (DMPF_SYNC | DMPF_NOSYNC)))
 954                 mddev->recovery_cp = le64_to_cpu(sb->array_resync_offset);
 955
 956         /*
 957          * During load, we set FirstUse if a new superblock was written.
 958          * There are two reasons we might not have a superblock:
 959          * 1) The array is brand new - in which case, all of the
 960          *    devices must have their In_sync bit set.  Also,
 961          *    recovery_cp must be 0, unless forced.
 962          * 2) This is a new device being added to an old array
 963          *    and the new device needs to be rebuilt - in which
 964          *    case the In_sync bit will /not/ be set and
 965          *    recovery_cp must be MaxSector.
 966          */
 967         rdev_for_each(r, mddev) {
 968                 if (!test_bit(In_sync, &r->flags)) {
 969                         DMINFO("Device %d specified for rebuild: "
 970                                "Clearing superblock", r->raid_disk);
 971                         rebuilds++;
 972                 } else if (test_bit(FirstUse, &r->flags))
 973                         new_devs++;
 974         }
 975
 976         if (!rebuilds) {
 977                 if (new_devs == mddev->raid_disks) {
 978                         DMINFO("Superblocks created for new array");
 979                         set_bit(MD_ARRAY_FIRST_USE, &mddev->flags);
 980                 } else if (new_devs) {
 981                         DMERR("New device injected "
 982                               "into existing array without 'rebuild' "
 983                               "parameter specified");
 984                         return -EINVAL;
 985                 }
 986         } else if (new_devs) {
 987                 DMERR("'rebuild' devices cannot be "
 988                       "injected into an array with other first-time devices");
 989                 return -EINVAL;
 990         } else if (mddev->recovery_cp != MaxSector) {
 991                 DMERR("'rebuild' specified while array is not in-sync");
 992                 return -EINVAL;
 993         }
 994
 995         /*
 996          * Now we set the Faulty bit for those devices that are
 997          * recorded in the superblock as failed.
 998          */
 999         rdev_for_each(r, mddev) {
1000                 if (!r->sb_page)
1001                         continue;
1002                 sb2 = page_address(r->sb_page);
1003                 sb2->failed_devices = 0;
1004
1005                 /*
1006                  * Check for any device re-ordering.
1007                  */
1008                 if (!test_bit(FirstUse, &r->flags) && (r->raid_disk >= 0)) {
1009                         role = le32_to_cpu(sb2->array_position);
1010                         if (role != r->raid_disk) {
1011                                 if (rs->raid_type->level != 1) {
1012                                         rs->ti->error = "Cannot change device "
1013                                                 "positions in RAID array";
1014                                         return -EINVAL;
1015                                 }
1016                                 DMINFO("RAID1 device #%d now at position #%d",
1017                                        role, r->raid_disk);
1018                         }
1019
1020                         /*
1021                          * Partial recovery is performed on
1022                          * returning failed devices.
1023                          */
1024                         if (failed_devices & (1 << role))
1025                                 set_bit(Faulty, &r->flags);
1026                 }
1027         }
1028
1029         return 0;
1030 }
1031
1032 static int super_validate(struct mddev *mddev, struct md_rdev *rdev)
1033 {
1034         struct dm_raid_superblock *sb = page_address(rdev->sb_page);
1035
1036         /*
1037          * If mddev->events is not set, we know we have not yet initialized
1038          * the array.
1039          */
1040         if (!mddev->events && super_init_validation(mddev, rdev))
1041                 return -EINVAL;
1042
1043         mddev->bitmap_info.offset = 4096 >> 9; /* Enable bitmap creation */
1044         rdev->mddev->bitmap_info.default_offset = 4096 >> 9;
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 DMPF_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                 if (rs->print_flags & DMPF_SYNC)
1089                         continue;
1090
1091                 if (!rdev->meta_bdev)
1092                         continue;
1093
1094                 ret = super_load(rdev, freshest);
1095
1096                 switch (ret) {
1097                 case 1:
1098                         freshest = rdev;
1099                         break;
1100                 case 0:
1101                         break;
1102                 default:
1103                         dev = container_of(rdev, struct raid_dev, rdev);
1104                         if (dev->meta_dev)
1105                                 dm_put_device(ti, dev->meta_dev);
1106
1107                         dev->meta_dev = NULL;
1108                         rdev->meta_bdev = NULL;
1109
1110                         if (rdev->sb_page)
1111                                 put_page(rdev->sb_page);
1112
1113                         rdev->sb_page = NULL;
1114
1115                         rdev->sb_loaded = 0;
1116
1117                         /*
1118                          * We might be able to salvage the data device
1119                          * even though the meta device has failed.  For
1120                          * now, we behave as though '- -' had been
1121                          * set for this device in the table.
1122                          */
1123                         if (dev->data_dev)
1124                                 dm_put_device(ti, dev->data_dev);
1125
1126                         dev->data_dev = NULL;
1127                         rdev->bdev = NULL;
1128
1129                         list_del(&rdev->same_set);
1130                 }
1131         }
1132
1133         if (!freshest)
1134                 return 0;
1135
1136         if (validate_raid_redundancy(rs)) {
1137                 rs->ti->error = "Insufficient redundancy to activate array";
1138                 return -EINVAL;
1139         }
1140
1141         /*
1142          * Validation of the freshest device provides the source of
1143          * validation for the remaining devices.
1144          */
1145         ti->error = "Unable to assemble array: Invalid superblocks";
1146         if (super_validate(mddev, freshest))
1147                 return -EINVAL;
1148
1149         rdev_for_each(rdev, mddev)
1150                 if ((rdev != freshest) && super_validate(mddev, rdev))
1151                         return -EINVAL;
1152
1153         return 0;
1154 }
1155
1156 /*
1157  * Enable/disable discard support on RAID set depending on
1158  * RAID level and discard properties of underlying RAID members.
1159  */
1160 static void configure_discard_support(struct dm_target *ti, struct raid_set *rs)
1161 {
1162         int i;
1163         bool raid456;
1164
1165         /* Assume discards not supported until after checks below. */
1166         ti->discards_supported = false;
1167
1168         /* RAID level 4,5,6 require discard_zeroes_data for data integrity! */
1169         raid456 = (rs->md.level == 4 || rs->md.level == 5 || rs->md.level == 6);
1170
1171         for (i = 0; i < rs->md.raid_disks; i++) {
1172                 struct request_queue *q = bdev_get_queue(rs->dev[i].rdev.bdev);
1173
1174                 if (!q || !blk_queue_discard(q))
1175                         return;
1176
1177                 if (raid456) {
1178                         if (!q->limits.discard_zeroes_data)
1179                                 return;
1180                         if (!devices_handle_discard_safely) {
1181                                 DMERR("raid456 discard support disabled due to discard_zeroes_data uncertainty.");
1182                                 DMERR("Set dm-raid.devices_handle_discard_safely=Y to override.");
1183                                 return;
1184                         }
1185                 }
1186         }
1187
1188         /* All RAID members properly support discards */
1189         ti->discards_supported = true;
1190
1191         /*
1192          * RAID1 and RAID10 personalities require bio splitting,
1193          * RAID0/4/5/6 don't and process large discard bios properly.
1194          */
1195         ti->split_discard_bios = !!(rs->md.level == 1 || rs->md.level == 10);
1196         ti->num_discard_bios = 1;
1197 }
1198
1199 /*
1200  * Construct a RAID4/5/6 mapping:
1201  * Args:
1202  *      <raid_type> <#raid_params> <raid_params>                \
1203  *      <#raid_devs> { <meta_dev1> <dev1> .. <meta_devN> <devN> }
1204  *
1205  * <raid_params> varies by <raid_type>.  See 'parse_raid_params' for
1206  * details on possible <raid_params>.
1207  */
1208 static int raid_ctr(struct dm_target *ti, unsigned argc, char **argv)
1209 {
1210         int ret;
1211         struct raid_type *rt;
1212         unsigned long num_raid_params, num_raid_devs;
1213         struct raid_set *rs = NULL;
1214
1215         /* Must have at least <raid_type> <#raid_params> */
1216         if (argc < 2) {
1217                 ti->error = "Too few arguments";
1218                 return -EINVAL;
1219         }
1220
1221         /* raid type */
1222         rt = get_raid_type(argv[0]);
1223         if (!rt) {
1224                 ti->error = "Unrecognised raid_type";
1225                 return -EINVAL;
1226         }
1227         argc--;
1228         argv++;
1229
1230         /* number of RAID parameters */
1231         if (kstrtoul(argv[0], 10, &num_raid_params) < 0) {
1232                 ti->error = "Cannot understand number of RAID parameters";
1233                 return -EINVAL;
1234         }
1235         argc--;
1236         argv++;
1237
1238         /* Skip over RAID params for now and find out # of devices */
1239         if (num_raid_params + 1 > argc) {
1240                 ti->error = "Arguments do not agree with counts given";
1241                 return -EINVAL;
1242         }
1243
1244         if ((kstrtoul(argv[num_raid_params], 10, &num_raid_devs) < 0) ||
1245             (num_raid_devs >= INT_MAX)) {
1246                 ti->error = "Cannot understand number of raid devices";
1247                 return -EINVAL;
1248         }
1249
1250         rs = context_alloc(ti, rt, (unsigned)num_raid_devs);
1251         if (IS_ERR(rs))
1252                 return PTR_ERR(rs);
1253
1254         ret = parse_raid_params(rs, argv, (unsigned)num_raid_params);
1255         if (ret)
1256                 goto bad;
1257
1258         ret = -EINVAL;
1259
1260         argc -= num_raid_params + 1; /* +1: we already have num_raid_devs */
1261         argv += num_raid_params + 1;
1262
1263         if (argc != (num_raid_devs * 2)) {
1264                 ti->error = "Supplied RAID devices does not match the count given";
1265                 goto bad;
1266         }
1267
1268         ret = dev_parms(rs, argv);
1269         if (ret)
1270                 goto bad;
1271
1272         rs->md.sync_super = super_sync;
1273         ret = analyse_superblocks(ti, rs);
1274         if (ret)
1275                 goto bad;
1276
1277         INIT_WORK(&rs->md.event_work, do_table_event);
1278         ti->private = rs;
1279         ti->num_flush_bios = 1;
1280
1281         /*
1282          * Disable/enable discard support on RAID set.
1283          */
1284         configure_discard_support(ti, rs);
1285
1286         mutex_lock(&rs->md.reconfig_mutex);
1287         ret = md_run(&rs->md);
1288         rs->md.in_sync = 0; /* Assume already marked dirty */
1289         mutex_unlock(&rs->md.reconfig_mutex);
1290
1291         if (ret) {
1292                 ti->error = "Fail to run raid array";
1293                 goto bad;
1294         }
1295
1296         if (ti->len != rs->md.array_sectors) {
1297                 ti->error = "Array size does not match requested target length";
1298                 ret = -EINVAL;
1299                 goto size_mismatch;
1300         }
1301         rs->callbacks.congested_fn = raid_is_congested;
1302         dm_table_add_target_callbacks(ti->table, &rs->callbacks);
1303
1304         mddev_suspend(&rs->md);
1305         return 0;
1306
1307 size_mismatch:
1308         md_stop(&rs->md);
1309 bad:
1310         context_free(rs);
1311
1312         return ret;
1313 }
1314
1315 static void raid_dtr(struct dm_target *ti)
1316 {
1317         struct raid_set *rs = ti->private;
1318
1319         list_del_init(&rs->callbacks.list);
1320         md_stop(&rs->md);
1321         context_free(rs);
1322 }
1323
1324 static int raid_map(struct dm_target *ti, struct bio *bio)
1325 {
1326         struct raid_set *rs = ti->private;
1327         struct mddev *mddev = &rs->md;
1328
1329         mddev->pers->make_request(mddev, bio);
1330
1331         return DM_MAPIO_SUBMITTED;
1332 }
1333
1334 static const char *decipher_sync_action(struct mddev *mddev)
1335 {
1336         if (test_bit(MD_RECOVERY_FROZEN, &mddev->recovery))
1337                 return "frozen";
1338
1339         if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
1340             (!mddev->ro && test_bit(MD_RECOVERY_NEEDED, &mddev->recovery))) {
1341                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1342                         return "reshape";
1343
1344                 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1345                         if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1346                                 return "resync";
1347                         else if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1348                                 return "check";
1349                         return "repair";
1350                 }
1351
1352                 if (test_bit(MD_RECOVERY_RECOVER, &mddev->recovery))
1353                         return "recover";
1354         }
1355
1356         return "idle";
1357 }
1358
1359 static void raid_status(struct dm_target *ti, status_type_t type,
1360                         unsigned status_flags, char *result, unsigned maxlen)
1361 {
1362         struct raid_set *rs = ti->private;
1363         unsigned raid_param_cnt = 1; /* at least 1 for chunksize */
1364         unsigned sz = 0;
1365         int i, array_in_sync = 0;
1366         sector_t sync;
1367
1368         switch (type) {
1369         case STATUSTYPE_INFO:
1370                 DMEMIT("%s %d ", rs->raid_type->name, rs->md.raid_disks);
1371
1372                 if (test_bit(MD_RECOVERY_RUNNING, &rs->md.recovery))
1373                         sync = rs->md.curr_resync_completed;
1374                 else
1375                         sync = rs->md.recovery_cp;
1376
1377                 if (sync >= rs->md.resync_max_sectors) {
1378                         /*
1379                          * Sync complete.
1380                          */
1381                         array_in_sync = 1;
1382                         sync = rs->md.resync_max_sectors;
1383                 } else if (test_bit(MD_RECOVERY_REQUESTED, &rs->md.recovery)) {
1384                         /*
1385                          * If "check" or "repair" is occurring, the array has
1386                          * undergone and initial sync and the health characters
1387                          * should not be 'a' anymore.
1388                          */
1389                         array_in_sync = 1;
1390                 } else {
1391                         /*
1392                          * The array may be doing an initial sync, or it may
1393                          * be rebuilding individual components.  If all the
1394                          * devices are In_sync, then it is the array that is
1395                          * being initialized.
1396                          */
1397                         for (i = 0; i < rs->md.raid_disks; i++)
1398                                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
1399                                         array_in_sync = 1;
1400                 }
1401
1402                 /*
1403                  * Status characters:
1404                  *  'D' = Dead/Failed device
1405                  *  'a' = Alive but not in-sync
1406                  *  'A' = Alive and in-sync
1407                  */
1408                 for (i = 0; i < rs->md.raid_disks; i++) {
1409                         if (test_bit(Faulty, &rs->dev[i].rdev.flags))
1410                                 DMEMIT("D");
1411                         else if (!array_in_sync ||
1412                                  !test_bit(In_sync, &rs->dev[i].rdev.flags))
1413                                 DMEMIT("a");
1414                         else
1415                                 DMEMIT("A");
1416                 }
1417
1418                 /*
1419                  * In-sync ratio:
1420                  *  The in-sync ratio shows the progress of:
1421                  *   - Initializing the array
1422                  *   - Rebuilding a subset of devices of the array
1423                  *  The user can distinguish between the two by referring
1424                  *  to the status characters.
1425                  */
1426                 DMEMIT(" %llu/%llu",
1427                        (unsigned long long) sync,
1428                        (unsigned long long) rs->md.resync_max_sectors);
1429
1430                 /*
1431                  * Sync action:
1432                  *   See Documentation/device-mapper/dm-raid.c for
1433                  *   information on each of these states.
1434                  */
1435                 DMEMIT(" %s", decipher_sync_action(&rs->md));
1436
1437                 /*
1438                  * resync_mismatches/mismatch_cnt
1439                  *   This field shows the number of discrepancies found when
1440                  *   performing a "check" of the array.
1441                  */
1442                 DMEMIT(" %llu",
1443                        (strcmp(rs->md.last_sync_action, "check")) ? 0 :
1444                        (unsigned long long)
1445                        atomic64_read(&rs->md.resync_mismatches));
1446                 break;
1447         case STATUSTYPE_TABLE:
1448                 /* The string you would use to construct this array */
1449                 for (i = 0; i < rs->md.raid_disks; i++) {
1450                         if ((rs->print_flags & DMPF_REBUILD) &&
1451                             rs->dev[i].data_dev &&
1452                             !test_bit(In_sync, &rs->dev[i].rdev.flags))
1453                                 raid_param_cnt += 2; /* for rebuilds */
1454                         if (rs->dev[i].data_dev &&
1455                             test_bit(WriteMostly, &rs->dev[i].rdev.flags))
1456                                 raid_param_cnt += 2;
1457                 }
1458
1459                 raid_param_cnt += (hweight32(rs->print_flags & ~DMPF_REBUILD) * 2);
1460                 if (rs->print_flags & (DMPF_SYNC | DMPF_NOSYNC))
1461                         raid_param_cnt--;
1462
1463                 DMEMIT("%s %u %u", rs->raid_type->name,
1464                        raid_param_cnt, rs->md.chunk_sectors);
1465
1466                 if ((rs->print_flags & DMPF_SYNC) &&
1467                     (rs->md.recovery_cp == MaxSector))
1468                         DMEMIT(" sync");
1469                 if (rs->print_flags & DMPF_NOSYNC)
1470                         DMEMIT(" nosync");
1471
1472                 for (i = 0; i < rs->md.raid_disks; i++)
1473                         if ((rs->print_flags & DMPF_REBUILD) &&
1474                             rs->dev[i].data_dev &&
1475                             !test_bit(In_sync, &rs->dev[i].rdev.flags))
1476                                 DMEMIT(" rebuild %u", i);
1477
1478                 if (rs->print_flags & DMPF_DAEMON_SLEEP)
1479                         DMEMIT(" daemon_sleep %lu",
1480                                rs->md.bitmap_info.daemon_sleep);
1481
1482                 if (rs->print_flags & DMPF_MIN_RECOVERY_RATE)
1483                         DMEMIT(" min_recovery_rate %d", rs->md.sync_speed_min);
1484
1485                 if (rs->print_flags & DMPF_MAX_RECOVERY_RATE)
1486                         DMEMIT(" max_recovery_rate %d", rs->md.sync_speed_max);
1487
1488                 for (i = 0; i < rs->md.raid_disks; i++)
1489                         if (rs->dev[i].data_dev &&
1490                             test_bit(WriteMostly, &rs->dev[i].rdev.flags))
1491                                 DMEMIT(" write_mostly %u", i);
1492
1493                 if (rs->print_flags & DMPF_MAX_WRITE_BEHIND)
1494                         DMEMIT(" max_write_behind %lu",
1495                                rs->md.bitmap_info.max_write_behind);
1496
1497                 if (rs->print_flags & DMPF_STRIPE_CACHE) {
1498                         struct r5conf *conf = rs->md.private;
1499
1500                         /* convert from kiB to sectors */
1501                         DMEMIT(" stripe_cache %d",
1502                                conf ? conf->max_nr_stripes * 2 : 0);
1503                 }
1504
1505                 if (rs->print_flags & DMPF_REGION_SIZE)
1506                         DMEMIT(" region_size %lu",
1507                                rs->md.bitmap_info.chunksize >> 9);
1508
1509                 if (rs->print_flags & DMPF_RAID10_COPIES)
1510                         DMEMIT(" raid10_copies %u",
1511                                raid10_md_layout_to_copies(rs->md.layout));
1512
1513                 if (rs->print_flags & DMPF_RAID10_FORMAT)
1514                         DMEMIT(" raid10_format %s",
1515                                raid10_md_layout_to_format(rs->md.layout));
1516
1517                 DMEMIT(" %d", rs->md.raid_disks);
1518                 for (i = 0; i < rs->md.raid_disks; i++) {
1519                         if (rs->dev[i].meta_dev)
1520                                 DMEMIT(" %s", rs->dev[i].meta_dev->name);
1521                         else
1522                                 DMEMIT(" -");
1523
1524                         if (rs->dev[i].data_dev)
1525                                 DMEMIT(" %s", rs->dev[i].data_dev->name);
1526                         else
1527                                 DMEMIT(" -");
1528                 }
1529         }
1530 }
1531
1532 static int raid_message(struct dm_target *ti, unsigned argc, char **argv)
1533 {
1534         struct raid_set *rs = ti->private;
1535         struct mddev *mddev = &rs->md;
1536
1537         if (!strcasecmp(argv[0], "reshape")) {
1538                 DMERR("Reshape not supported.");
1539                 return -EINVAL;
1540         }
1541
1542         if (!mddev->pers || !mddev->pers->sync_request)
1543                 return -EINVAL;
1544
1545         if (!strcasecmp(argv[0], "frozen"))
1546                 set_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
1547         else
1548                 clear_bit(MD_RECOVERY_FROZEN, &mddev->recovery);
1549
1550         if (!strcasecmp(argv[0], "idle") || !strcasecmp(argv[0], "frozen")) {
1551                 if (mddev->sync_thread) {
1552                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1553                         md_reap_sync_thread(mddev);
1554                 }
1555         } else if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery) ||
1556                    test_bit(MD_RECOVERY_NEEDED, &mddev->recovery))
1557                 return -EBUSY;
1558         else if (!strcasecmp(argv[0], "resync"))
1559                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1560         else if (!strcasecmp(argv[0], "recover")) {
1561                 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
1562                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1563         } else {
1564                 if (!strcasecmp(argv[0], "check"))
1565                         set_bit(MD_RECOVERY_CHECK, &mddev->recovery);
1566                 else if (!!strcasecmp(argv[0], "repair"))
1567                         return -EINVAL;
1568                 set_bit(MD_RECOVERY_REQUESTED, &mddev->recovery);
1569                 set_bit(MD_RECOVERY_SYNC, &mddev->recovery);
1570         }
1571         if (mddev->ro == 2) {
1572                 /* A write to sync_action is enough to justify
1573                  * canceling read-auto mode
1574                  */
1575                 mddev->ro = 0;
1576                 if (!mddev->suspended)
1577                         md_wakeup_thread(mddev->sync_thread);
1578         }
1579         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
1580         if (!mddev->suspended)
1581                 md_wakeup_thread(mddev->thread);
1582
1583         return 0;
1584 }
1585
1586 static int raid_iterate_devices(struct dm_target *ti,
1587                                 iterate_devices_callout_fn fn, void *data)
1588 {
1589         struct raid_set *rs = ti->private;
1590         unsigned i;
1591         int ret = 0;
1592
1593         for (i = 0; !ret && i < rs->md.raid_disks; i++)
1594                 if (rs->dev[i].data_dev)
1595                         ret = fn(ti,
1596                                  rs->dev[i].data_dev,
1597                                  0, /* No offset on data devs */
1598                                  rs->md.dev_sectors,
1599                                  data);
1600
1601         return ret;
1602 }
1603
1604 static void raid_io_hints(struct dm_target *ti, struct queue_limits *limits)
1605 {
1606         struct raid_set *rs = ti->private;
1607         unsigned chunk_size = rs->md.chunk_sectors << 9;
1608         struct r5conf *conf = rs->md.private;
1609
1610         blk_limits_io_min(limits, chunk_size);
1611         blk_limits_io_opt(limits, chunk_size * (conf->raid_disks - conf->max_degraded));
1612 }
1613
1614 static void raid_presuspend(struct dm_target *ti)
1615 {
1616         struct raid_set *rs = ti->private;
1617
1618         md_stop_writes(&rs->md);
1619 }
1620
1621 static void raid_postsuspend(struct dm_target *ti)
1622 {
1623         struct raid_set *rs = ti->private;
1624
1625         mddev_suspend(&rs->md);
1626 }
1627
1628 static void attempt_restore_of_faulty_devices(struct raid_set *rs)
1629 {
1630         int i;
1631         uint64_t failed_devices, cleared_failed_devices = 0;
1632         unsigned long flags;
1633         struct dm_raid_superblock *sb;
1634         struct md_rdev *r;
1635
1636         for (i = 0; i < rs->md.raid_disks; i++) {
1637                 r = &rs->dev[i].rdev;
1638                 if (test_bit(Faulty, &r->flags) && r->sb_page &&
1639                     sync_page_io(r, 0, r->sb_size, r->sb_page, READ, 1)) {
1640                         DMINFO("Faulty %s device #%d has readable super block."
1641                                "  Attempting to revive it.",
1642                                rs->raid_type->name, i);
1643
1644                         /*
1645                          * Faulty bit may be set, but sometimes the array can
1646                          * be suspended before the personalities can respond
1647                          * by removing the device from the array (i.e. calling
1648                          * 'hot_remove_disk').  If they haven't yet removed
1649                          * the failed device, its 'raid_disk' number will be
1650                          * '>= 0' - meaning we must call this function
1651                          * ourselves.
1652                          */
1653                         if ((r->raid_disk >= 0) &&
1654                             (r->mddev->pers->hot_remove_disk(r->mddev, r) != 0))
1655                                 /* Failed to revive this device, try next */
1656                                 continue;
1657
1658                         r->raid_disk = i;
1659                         r->saved_raid_disk = i;
1660                         flags = r->flags;
1661                         clear_bit(Faulty, &r->flags);
1662                         clear_bit(WriteErrorSeen, &r->flags);
1663                         clear_bit(In_sync, &r->flags);
1664                         if (r->mddev->pers->hot_add_disk(r->mddev, r)) {
1665                                 r->raid_disk = -1;
1666                                 r->saved_raid_disk = -1;
1667                                 r->flags = flags;
1668                         } else {
1669                                 r->recovery_offset = 0;
1670                                 cleared_failed_devices |= 1 << i;
1671                         }
1672                 }
1673         }
1674         if (cleared_failed_devices) {
1675                 rdev_for_each(r, &rs->md) {
1676                         sb = page_address(r->sb_page);
1677                         failed_devices = le64_to_cpu(sb->failed_devices);
1678                         failed_devices &= ~cleared_failed_devices;
1679                         sb->failed_devices = cpu_to_le64(failed_devices);
1680                 }
1681         }
1682 }
1683
1684 static void raid_resume(struct dm_target *ti)
1685 {
1686         struct raid_set *rs = ti->private;
1687
1688         set_bit(MD_CHANGE_DEVS, &rs->md.flags);
1689         if (!rs->bitmap_loaded) {
1690                 bitmap_load(&rs->md);
1691                 rs->bitmap_loaded = 1;
1692         } else {
1693                 /*
1694                  * A secondary resume while the device is active.
1695                  * Take this opportunity to check whether any failed
1696                  * devices are reachable again.
1697                  */
1698                 attempt_restore_of_faulty_devices(rs);
1699         }
1700
1701         clear_bit(MD_RECOVERY_FROZEN, &rs->md.recovery);
1702         mddev_resume(&rs->md);
1703 }
1704
1705 static struct target_type raid_target = {
1706         .name = "raid",
1707         .version = {1, 6, 0},
1708         .module = THIS_MODULE,
1709         .ctr = raid_ctr,
1710         .dtr = raid_dtr,
1711         .map = raid_map,
1712         .status = raid_status,
1713         .message = raid_message,
1714         .iterate_devices = raid_iterate_devices,
1715         .io_hints = raid_io_hints,
1716         .presuspend = raid_presuspend,
1717         .postsuspend = raid_postsuspend,
1718         .resume = raid_resume,
1719 };
1720
1721 static int __init dm_raid_init(void)
1722 {
1723         DMINFO("Loading target version %u.%u.%u",
1724                raid_target.version[0],
1725                raid_target.version[1],
1726                raid_target.version[2]);
1727         return dm_register_target(&raid_target);
1728 }
1729
1730 static void __exit dm_raid_exit(void)
1731 {
1732         dm_unregister_target(&raid_target);
1733 }
1734
1735 module_init(dm_raid_init);
1736 module_exit(dm_raid_exit);
1737
1738 module_param(devices_handle_discard_safely, bool, 0644);
1739 MODULE_PARM_DESC(devices_handle_discard_safely,
1740                  "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
1741
1742 MODULE_DESCRIPTION(DM_NAME " raid4/5/6 target");
1743 MODULE_ALIAS("dm-raid1");
1744 MODULE_ALIAS("dm-raid10");
1745 MODULE_ALIAS("dm-raid4");
1746 MODULE_ALIAS("dm-raid5");
1747 MODULE_ALIAS("dm-raid6");
1748 MODULE_AUTHOR("Neil Brown <dm-devel@redhat.com>");
1749 MODULE_LICENSE("GPL");