drivers/md/dm-raid.c

   1 /*
   2  * Copyright (C) 2010-2011 Neil Brown
   3  * Copyright (C) 2010-2011 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 "bitmap.h"
  15
  16 #include <linux/device-mapper.h>
  17
  18 #define DM_MSG_PREFIX "raid"
  19
  20 /*
  21  * The following flags are used by dm-raid.c to set up the array state.
  22  * They must be cleared before md_run is called.
  23  */
  24 #define FirstUse 10             /* rdev flag */
  25
  26 struct raid_dev {
  27         /*
  28          * Two DM devices, one to hold metadata and one to hold the
  29          * actual data/parity.  The reason for this is to not confuse
  30          * ti->len and give more flexibility in altering size and
  31          * characteristics.
  32          *
  33          * While it is possible for this device to be associated
  34          * with a different physical device than the data_dev, it
  35          * is intended for it to be the same.
  36          *    |--------- Physical Device ---------|
  37          *    |- meta_dev -|------ data_dev ------|
  38          */
  39         struct dm_dev *meta_dev;
  40         struct dm_dev *data_dev;
  41         struct md_rdev rdev;
  42 };
  43
  44 /*
  45  * Flags for rs->print_flags field.
  46  */
  47 #define DMPF_SYNC              0x1
  48 #define DMPF_NOSYNC            0x2
  49 #define DMPF_REBUILD           0x4
  50 #define DMPF_DAEMON_SLEEP      0x8
  51 #define DMPF_MIN_RECOVERY_RATE 0x10
  52 #define DMPF_MAX_RECOVERY_RATE 0x20
  53 #define DMPF_MAX_WRITE_BEHIND  0x40
  54 #define DMPF_STRIPE_CACHE      0x80
  55 #define DMPF_REGION_SIZE       0X100
  56 struct raid_set {
  57         struct dm_target *ti;
  58
  59         uint32_t bitmap_loaded;
  60         uint32_t print_flags;
  61
  62         struct mddev md;
  63         struct raid_type *raid_type;
  64         struct dm_target_callbacks callbacks;
  65
  66         struct raid_dev dev[0];
  67 };
  68
  69 /* Supported raid types and properties. */
  70 static struct raid_type {
  71         const char *name;               /* RAID algorithm. */
  72         const char *descr;              /* Descriptor text for logging. */
  73         const unsigned parity_devs;     /* # of parity devices. */
  74         const unsigned minimal_devs;    /* minimal # of devices in set. */
  75         const unsigned level;           /* RAID level. */
  76         const unsigned algorithm;       /* RAID algorithm. */
  77 } raid_types[] = {
  78         {"raid1",    "RAID1 (mirroring)",               0, 2, 1, 0 /* NONE */},
  79         {"raid4",    "RAID4 (dedicated parity disk)",   1, 2, 5, ALGORITHM_PARITY_0},
  80         {"raid5_la", "RAID5 (left asymmetric)",         1, 2, 5, ALGORITHM_LEFT_ASYMMETRIC},
  81         {"raid5_ra", "RAID5 (right asymmetric)",        1, 2, 5, ALGORITHM_RIGHT_ASYMMETRIC},
  82         {"raid5_ls", "RAID5 (left symmetric)",          1, 2, 5, ALGORITHM_LEFT_SYMMETRIC},
  83         {"raid5_rs", "RAID5 (right symmetric)",         1, 2, 5, ALGORITHM_RIGHT_SYMMETRIC},
  84         {"raid6_zr", "RAID6 (zero restart)",            2, 4, 6, ALGORITHM_ROTATING_ZERO_RESTART},
  85         {"raid6_nr", "RAID6 (N restart)",               2, 4, 6, ALGORITHM_ROTATING_N_RESTART},
  86         {"raid6_nc", "RAID6 (N continue)",              2, 4, 6, ALGORITHM_ROTATING_N_CONTINUE}
  87 };
  88
  89 static struct raid_type *get_raid_type(char *name)
  90 {
  91         int i;
  92
  93         for (i = 0; i < ARRAY_SIZE(raid_types); i++)
  94                 if (!strcmp(raid_types[i].name, name))
  95                         return &raid_types[i];
  96
  97         return NULL;
  98 }
  99
 100 static struct raid_set *context_alloc(struct dm_target *ti, struct raid_type *raid_type, unsigned raid_devs)
 101 {
 102         unsigned i;
 103         struct raid_set *rs;
 104         sector_t sectors_per_dev;
 105
 106         if (raid_devs <= raid_type->parity_devs) {
 107                 ti->error = "Insufficient number of devices";
 108                 return ERR_PTR(-EINVAL);
 109         }
 110
 111         sectors_per_dev = ti->len;
 112         if ((raid_type->level > 1) &&
 113             sector_div(sectors_per_dev, (raid_devs - raid_type->parity_devs))) {
 114                 ti->error = "Target length not divisible by number of data devices";
 115                 return ERR_PTR(-EINVAL);
 116         }
 117
 118         rs = kzalloc(sizeof(*rs) + raid_devs * sizeof(rs->dev[0]), GFP_KERNEL);
 119         if (!rs) {
 120                 ti->error = "Cannot allocate raid context";
 121                 return ERR_PTR(-ENOMEM);
 122         }
 123
 124         mddev_init(&rs->md);
 125
 126         rs->ti = ti;
 127         rs->raid_type = raid_type;
 128         rs->md.raid_disks = raid_devs;
 129         rs->md.level = raid_type->level;
 130         rs->md.new_level = rs->md.level;
 131         rs->md.dev_sectors = sectors_per_dev;
 132         rs->md.layout = raid_type->algorithm;
 133         rs->md.new_layout = rs->md.layout;
 134         rs->md.delta_disks = 0;
 135         rs->md.recovery_cp = 0;
 136
 137         for (i = 0; i < raid_devs; i++)
 138                 md_rdev_init(&rs->dev[i].rdev);
 139
 140         /*
 141          * Remaining items to be initialized by further RAID params:
 142          *  rs->md.persistent
 143          *  rs->md.external
 144          *  rs->md.chunk_sectors
 145          *  rs->md.new_chunk_sectors
 146          */
 147
 148         return rs;
 149 }
 150
 151 static void context_free(struct raid_set *rs)
 152 {
 153         int i;
 154
 155         for (i = 0; i < rs->md.raid_disks; i++) {
 156                 if (rs->dev[i].meta_dev)
 157                         dm_put_device(rs->ti, rs->dev[i].meta_dev);
 158                 if (rs->dev[i].rdev.sb_page)
 159                         put_page(rs->dev[i].rdev.sb_page);
 160                 rs->dev[i].rdev.sb_page = NULL;
 161                 rs->dev[i].rdev.sb_loaded = 0;
 162                 if (rs->dev[i].data_dev)
 163                         dm_put_device(rs->ti, rs->dev[i].data_dev);
 164         }
 165
 166         kfree(rs);
 167 }
 168
 169 /*
 170  * For every device we have two words
 171  *  <meta_dev>: meta device name or '-' if missing
 172  *  <data_dev>: data device name or '-' if missing
 173  *
 174  * The following are permitted:
 175  *    - -
 176  *    - <data_dev>
 177  *    <meta_dev> <data_dev>
 178  *
 179  * The following is not allowed:
 180  *    <meta_dev> -
 181  *
 182  * This code parses those words.  If there is a failure,
 183  * the caller must use context_free to unwind the operations.
 184  */
 185 static int dev_parms(struct raid_set *rs, char **argv)
 186 {
 187         int i;
 188         int rebuild = 0;
 189         int metadata_available = 0;
 190         int ret = 0;
 191
 192         for (i = 0; i < rs->md.raid_disks; i++, argv += 2) {
 193                 rs->dev[i].rdev.raid_disk = i;
 194
 195                 rs->dev[i].meta_dev = NULL;
 196                 rs->dev[i].data_dev = NULL;
 197
 198                 /*
 199                  * There are no offsets, since there is a separate device
 200                  * for data and metadata.
 201                  */
 202                 rs->dev[i].rdev.data_offset = 0;
 203                 rs->dev[i].rdev.mddev = &rs->md;
 204
 205                 if (strcmp(argv[0], "-")) {
 206                         ret = dm_get_device(rs->ti, argv[0],
 207                                             dm_table_get_mode(rs->ti->table),
 208                                             &rs->dev[i].meta_dev);
 209                         rs->ti->error = "RAID metadata device lookup failure";
 210                         if (ret)
 211                                 return ret;
 212
 213                         rs->dev[i].rdev.sb_page = alloc_page(GFP_KERNEL);
 214                         if (!rs->dev[i].rdev.sb_page)
 215                                 return -ENOMEM;
 216                 }
 217
 218                 if (!strcmp(argv[1], "-")) {
 219                         if (!test_bit(In_sync, &rs->dev[i].rdev.flags) &&
 220                             (!rs->dev[i].rdev.recovery_offset)) {
 221                                 rs->ti->error = "Drive designated for rebuild not specified";
 222                                 return -EINVAL;
 223                         }
 224
 225                         rs->ti->error = "No data device supplied with metadata device";
 226                         if (rs->dev[i].meta_dev)
 227                                 return -EINVAL;
 228
 229                         continue;
 230                 }
 231
 232                 ret = dm_get_device(rs->ti, argv[1],
 233                                     dm_table_get_mode(rs->ti->table),
 234                                     &rs->dev[i].data_dev);
 235                 if (ret) {
 236                         rs->ti->error = "RAID device lookup failure";
 237                         return ret;
 238                 }
 239
 240                 if (rs->dev[i].meta_dev) {
 241                         metadata_available = 1;
 242                         rs->dev[i].rdev.meta_bdev = rs->dev[i].meta_dev->bdev;
 243                 }
 244                 rs->dev[i].rdev.bdev = rs->dev[i].data_dev->bdev;
 245                 list_add(&rs->dev[i].rdev.same_set, &rs->md.disks);
 246                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
 247                         rebuild++;
 248         }
 249
 250         if (metadata_available) {
 251                 rs->md.external = 0;
 252                 rs->md.persistent = 1;
 253                 rs->md.major_version = 2;
 254         } else if (rebuild && !rs->md.recovery_cp) {
 255                 /*
 256                  * Without metadata, we will not be able to tell if the array
 257                  * is in-sync or not - we must assume it is not.  Therefore,
 258                  * it is impossible to rebuild a drive.
 259                  *
 260                  * Even if there is metadata, the on-disk information may
 261                  * indicate that the array is not in-sync and it will then
 262                  * fail at that time.
 263                  *
 264                  * User could specify 'nosync' option if desperate.
 265                  */
 266                 DMERR("Unable to rebuild drive while array is not in-sync");
 267                 rs->ti->error = "RAID device lookup failure";
 268                 return -EINVAL;
 269         }
 270
 271         return 0;
 272 }
 273
 274 /*
 275  * validate_region_size
 276  * @rs
 277  * @region_size:  region size in sectors.  If 0, pick a size (4MiB default).
 278  *
 279  * Set rs->md.bitmap_info.chunksize (which really refers to 'region size').
 280  * Ensure that (ti->len/region_size < 2^21) - required by MD bitmap.
 281  *
 282  * Returns: 0 on success, -EINVAL on failure.
 283  */
 284 static int validate_region_size(struct raid_set *rs, unsigned long region_size)
 285 {
 286         unsigned long min_region_size = rs->ti->len / (1 << 21);
 287
 288         if (!region_size) {
 289                 /*
 290                  * Choose a reasonable default.  All figures in sectors.
 291                  */
 292                 if (min_region_size > (1 << 13)) {
 293                         DMINFO("Choosing default region size of %lu sectors",
 294                                region_size);
 295                         region_size = min_region_size;
 296                 } else {
 297                         DMINFO("Choosing default region size of 4MiB");
 298                         region_size = 1 << 13; /* sectors */
 299                 }
 300         } else {
 301                 /*
 302                  * Validate user-supplied value.
 303                  */
 304                 if (region_size > rs->ti->len) {
 305                         rs->ti->error = "Supplied region size is too large";
 306                         return -EINVAL;
 307                 }
 308
 309                 if (region_size < min_region_size) {
 310                         DMERR("Supplied region_size (%lu sectors) below minimum (%lu)",
 311                               region_size, min_region_size);
 312                         rs->ti->error = "Supplied region size is too small";
 313                         return -EINVAL;
 314                 }
 315
 316                 if (!is_power_of_2(region_size)) {
 317                         rs->ti->error = "Region size is not a power of 2";
 318                         return -EINVAL;
 319                 }
 320
 321                 if (region_size < rs->md.chunk_sectors) {
 322                         rs->ti->error = "Region size is smaller than the chunk size";
 323                         return -EINVAL;
 324                 }
 325         }
 326
 327         /*
 328          * Convert sectors to bytes.
 329          */
 330         rs->md.bitmap_info.chunksize = (region_size << 9);
 331
 332         return 0;
 333 }
 334
 335 /*
 336  * Possible arguments are...
 337  *      <chunk_size> [optional_args]
 338  *
 339  * Argument definitions
 340  *    <chunk_size>                      The number of sectors per disk that
 341  *                                      will form the "stripe"
 342  *    [[no]sync]                        Force or prevent recovery of the
 343  *                                      entire array
 344  *    [rebuild <idx>]                   Rebuild the drive indicated by the index
 345  *    [daemon_sleep <ms>]               Time between bitmap daemon work to
 346  *                                      clear bits
 347  *    [min_recovery_rate <kB/sec/disk>] Throttle RAID initialization
 348  *    [max_recovery_rate <kB/sec/disk>] Throttle RAID initialization
 349  *    [write_mostly <idx>]              Indicate a write mostly drive via index
 350  *    [max_write_behind <sectors>]      See '-write-behind=' (man mdadm)
 351  *    [stripe_cache <sectors>]          Stripe cache size for higher RAIDs
 352  *    [region_size <sectors>]           Defines granularity of bitmap
 353  */
 354 static int parse_raid_params(struct raid_set *rs, char **argv,
 355                              unsigned num_raid_params)
 356 {
 357         unsigned i, rebuild_cnt = 0;
 358         unsigned long value, region_size = 0;
 359         char *key;
 360
 361         /*
 362          * First, parse the in-order required arguments
 363          * "chunk_size" is the only argument of this type.
 364          */
 365         if ((strict_strtoul(argv[0], 10, &value) < 0)) {
 366                 rs->ti->error = "Bad chunk size";
 367                 return -EINVAL;
 368         } else if (rs->raid_type->level == 1) {
 369                 if (value)
 370                         DMERR("Ignoring chunk size parameter for RAID 1");
 371                 value = 0;
 372         } else if (!is_power_of_2(value)) {
 373                 rs->ti->error = "Chunk size must be a power of 2";
 374                 return -EINVAL;
 375         } else if (value < 8) {
 376                 rs->ti->error = "Chunk size value is too small";
 377                 return -EINVAL;
 378         }
 379
 380         rs->md.new_chunk_sectors = rs->md.chunk_sectors = value;
 381         argv++;
 382         num_raid_params--;
 383
 384         /*
 385          * We set each individual device as In_sync with a completed
 386          * 'recovery_offset'.  If there has been a device failure or
 387          * replacement then one of the following cases applies:
 388          *
 389          *   1) User specifies 'rebuild'.
 390          *      - Device is reset when param is read.
 391          *   2) A new device is supplied.
 392          *      - No matching superblock found, resets device.
 393          *   3) Device failure was transient and returns on reload.
 394          *      - Failure noticed, resets device for bitmap replay.
 395          *   4) Device hadn't completed recovery after previous failure.
 396          *      - Superblock is read and overrides recovery_offset.
 397          *
 398          * What is found in the superblocks of the devices is always
 399          * authoritative, unless 'rebuild' or '[no]sync' was specified.
 400          */
 401         for (i = 0; i < rs->md.raid_disks; i++) {
 402                 set_bit(In_sync, &rs->dev[i].rdev.flags);
 403                 rs->dev[i].rdev.recovery_offset = MaxSector;
 404         }
 405
 406         /*
 407          * Second, parse the unordered optional arguments
 408          */
 409         for (i = 0; i < num_raid_params; i++) {
 410                 if (!strcasecmp(argv[i], "nosync")) {
 411                         rs->md.recovery_cp = MaxSector;
 412                         rs->print_flags |= DMPF_NOSYNC;
 413                         continue;
 414                 }
 415                 if (!strcasecmp(argv[i], "sync")) {
 416                         rs->md.recovery_cp = 0;
 417                         rs->print_flags |= DMPF_SYNC;
 418                         continue;
 419                 }
 420
 421                 /* The rest of the optional arguments come in key/value pairs */
 422                 if ((i + 1) >= num_raid_params) {
 423                         rs->ti->error = "Wrong number of raid parameters given";
 424                         return -EINVAL;
 425                 }
 426
 427                 key = argv[i++];
 428                 if (strict_strtoul(argv[i], 10, &value) < 0) {
 429                         rs->ti->error = "Bad numerical argument given in raid params";
 430                         return -EINVAL;
 431                 }
 432
 433                 if (!strcasecmp(key, "rebuild")) {
 434                         rebuild_cnt++;
 435                         if (((rs->raid_type->level != 1) &&
 436                              (rebuild_cnt > rs->raid_type->parity_devs)) ||
 437                             ((rs->raid_type->level == 1) &&
 438                              (rebuild_cnt > (rs->md.raid_disks - 1)))) {
 439                                 rs->ti->error = "Too many rebuild devices specified for given RAID type";
 440                                 return -EINVAL;
 441                         }
 442                         if (value > rs->md.raid_disks) {
 443                                 rs->ti->error = "Invalid rebuild index given";
 444                                 return -EINVAL;
 445                         }
 446                         clear_bit(In_sync, &rs->dev[value].rdev.flags);
 447                         rs->dev[value].rdev.recovery_offset = 0;
 448                         rs->print_flags |= DMPF_REBUILD;
 449                 } else if (!strcasecmp(key, "write_mostly")) {
 450                         if (rs->raid_type->level != 1) {
 451                                 rs->ti->error = "write_mostly option is only valid for RAID1";
 452                                 return -EINVAL;
 453                         }
 454                         if (value >= rs->md.raid_disks) {
 455                                 rs->ti->error = "Invalid write_mostly drive index given";
 456                                 return -EINVAL;
 457                         }
 458                         set_bit(WriteMostly, &rs->dev[value].rdev.flags);
 459                 } else if (!strcasecmp(key, "max_write_behind")) {
 460                         if (rs->raid_type->level != 1) {
 461                                 rs->ti->error = "max_write_behind option is only valid for RAID1";
 462                                 return -EINVAL;
 463                         }
 464                         rs->print_flags |= DMPF_MAX_WRITE_BEHIND;
 465
 466                         /*
 467                          * In device-mapper, we specify things in sectors, but
 468                          * MD records this value in kB
 469                          */
 470                         value /= 2;
 471                         if (value > COUNTER_MAX) {
 472                                 rs->ti->error = "Max write-behind limit out of range";
 473                                 return -EINVAL;
 474                         }
 475                         rs->md.bitmap_info.max_write_behind = value;
 476                 } else if (!strcasecmp(key, "daemon_sleep")) {
 477                         rs->print_flags |= DMPF_DAEMON_SLEEP;
 478                         if (!value || (value > MAX_SCHEDULE_TIMEOUT)) {
 479                                 rs->ti->error = "daemon sleep period out of range";
 480                                 return -EINVAL;
 481                         }
 482                         rs->md.bitmap_info.daemon_sleep = value;
 483                 } else if (!strcasecmp(key, "stripe_cache")) {
 484                         rs->print_flags |= DMPF_STRIPE_CACHE;
 485
 486                         /*
 487                          * In device-mapper, we specify things in sectors, but
 488                          * MD records this value in kB
 489                          */
 490                         value /= 2;
 491
 492                         if (rs->raid_type->level < 5) {
 493                                 rs->ti->error = "Inappropriate argument: stripe_cache";
 494                                 return -EINVAL;
 495                         }
 496                         if (raid5_set_cache_size(&rs->md, (int)value)) {
 497                                 rs->ti->error = "Bad stripe_cache size";
 498                                 return -EINVAL;
 499                         }
 500                 } else if (!strcasecmp(key, "min_recovery_rate")) {
 501                         rs->print_flags |= DMPF_MIN_RECOVERY_RATE;
 502                         if (value > INT_MAX) {
 503                                 rs->ti->error = "min_recovery_rate out of range";
 504                                 return -EINVAL;
 505                         }
 506                         rs->md.sync_speed_min = (int)value;
 507                 } else if (!strcasecmp(key, "max_recovery_rate")) {
 508                         rs->print_flags |= DMPF_MAX_RECOVERY_RATE;
 509                         if (value > INT_MAX) {
 510                                 rs->ti->error = "max_recovery_rate out of range";
 511                                 return -EINVAL;
 512                         }
 513                         rs->md.sync_speed_max = (int)value;
 514                 } else if (!strcasecmp(key, "region_size")) {
 515                         rs->print_flags |= DMPF_REGION_SIZE;
 516                         region_size = value;
 517                 } else {
 518                         DMERR("Unable to parse RAID parameter: %s", key);
 519                         rs->ti->error = "Unable to parse RAID parameters";
 520                         return -EINVAL;
 521                 }
 522         }
 523
 524         if (validate_region_size(rs, region_size))
 525                 return -EINVAL;
 526
 527         if (rs->md.chunk_sectors)
 528                 rs->ti->split_io = rs->md.chunk_sectors;
 529         else
 530                 rs->ti->split_io = region_size;
 531
 532         if (rs->md.chunk_sectors)
 533                 rs->ti->split_io = rs->md.chunk_sectors;
 534         else
 535                 rs->ti->split_io = region_size;
 536
 537         /* Assume there are no metadata devices until the drives are parsed */
 538         rs->md.persistent = 0;
 539         rs->md.external = 1;
 540
 541         return 0;
 542 }
 543
 544 static void do_table_event(struct work_struct *ws)
 545 {
 546         struct raid_set *rs = container_of(ws, struct raid_set, md.event_work);
 547
 548         dm_table_event(rs->ti->table);
 549 }
 550
 551 static int raid_is_congested(struct dm_target_callbacks *cb, int bits)
 552 {
 553         struct raid_set *rs = container_of(cb, struct raid_set, callbacks);
 554
 555         if (rs->raid_type->level == 1)
 556                 return md_raid1_congested(&rs->md, bits);
 557
 558         return md_raid5_congested(&rs->md, bits);
 559 }
 560
 561 /*
 562  * This structure is never routinely used by userspace, unlike md superblocks.
 563  * Devices with this superblock should only ever be accessed via device-mapper.
 564  */
 565 #define DM_RAID_MAGIC 0x64526D44
 566 struct dm_raid_superblock {
 567         __le32 magic;           /* "DmRd" */
 568         __le32 features;        /* Used to indicate possible future changes */
 569
 570         __le32 num_devices;     /* Number of devices in this array. (Max 64) */
 571         __le32 array_position;  /* The position of this drive in the array */
 572
 573         __le64 events;          /* Incremented by md when superblock updated */
 574         __le64 failed_devices;  /* Bit field of devices to indicate failures */
 575
 576         /*
 577          * This offset tracks the progress of the repair or replacement of
 578          * an individual drive.
 579          */
 580         __le64 disk_recovery_offset;
 581
 582         /*
 583          * This offset tracks the progress of the initial array
 584          * synchronisation/parity calculation.
 585          */
 586         __le64 array_resync_offset;
 587
 588         /*
 589          * RAID characteristics
 590          */
 591         __le32 level;
 592         __le32 layout;
 593         __le32 stripe_sectors;
 594
 595         __u8 pad[452];          /* Round struct to 512 bytes. */
 596                                 /* Always set to 0 when writing. */
 597 } __packed;
 598
 599 static int read_disk_sb(struct md_rdev *rdev, int size)
 600 {
 601         BUG_ON(!rdev->sb_page);
 602
 603         if (rdev->sb_loaded)
 604                 return 0;
 605
 606         if (!sync_page_io(rdev, 0, size, rdev->sb_page, READ, 1)) {
 607                 DMERR("Failed to read device superblock");
 608                 return -EINVAL;
 609         }
 610
 611         rdev->sb_loaded = 1;
 612
 613         return 0;
 614 }
 615
 616 static void super_sync(struct mddev *mddev, struct md_rdev *rdev)
 617 {
 618         struct md_rdev *r, *t;
 619         uint64_t failed_devices;
 620         struct dm_raid_superblock *sb;
 621
 622         sb = page_address(rdev->sb_page);
 623         failed_devices = le64_to_cpu(sb->failed_devices);
 624
 625         rdev_for_each(r, t, mddev)
 626                 if ((r->raid_disk >= 0) && test_bit(Faulty, &r->flags))
 627                         failed_devices |= (1ULL << r->raid_disk);
 628
 629         memset(sb, 0, sizeof(*sb));
 630
 631         sb->magic = cpu_to_le32(DM_RAID_MAGIC);
 632         sb->features = cpu_to_le32(0);  /* No features yet */
 633
 634         sb->num_devices = cpu_to_le32(mddev->raid_disks);
 635         sb->array_position = cpu_to_le32(rdev->raid_disk);
 636
 637         sb->events = cpu_to_le64(mddev->events);
 638         sb->failed_devices = cpu_to_le64(failed_devices);
 639
 640         sb->disk_recovery_offset = cpu_to_le64(rdev->recovery_offset);
 641         sb->array_resync_offset = cpu_to_le64(mddev->recovery_cp);
 642
 643         sb->level = cpu_to_le32(mddev->level);
 644         sb->layout = cpu_to_le32(mddev->layout);
 645         sb->stripe_sectors = cpu_to_le32(mddev->chunk_sectors);
 646 }
 647
 648 /*
 649  * super_load
 650  *
 651  * This function creates a superblock if one is not found on the device
 652  * and will decide which superblock to use if there's a choice.
 653  *
 654  * Return: 1 if use rdev, 0 if use refdev, -Exxx otherwise
 655  */
 656 static int super_load(struct md_rdev *rdev, struct md_rdev *refdev)
 657 {
 658         int ret;
 659         struct dm_raid_superblock *sb;
 660         struct dm_raid_superblock *refsb;
 661         uint64_t events_sb, events_refsb;
 662
 663         rdev->sb_start = 0;
 664         rdev->sb_size = sizeof(*sb);
 665
 666         ret = read_disk_sb(rdev, rdev->sb_size);
 667         if (ret)
 668                 return ret;
 669
 670         sb = page_address(rdev->sb_page);
 671
 672         /*
 673          * Two cases that we want to write new superblocks and rebuild:
 674          * 1) New device (no matching magic number)
 675          * 2) Device specified for rebuild (!In_sync w/ offset == 0)
 676          */
 677         if ((sb->magic != cpu_to_le32(DM_RAID_MAGIC)) ||
 678             (!test_bit(In_sync, &rdev->flags) && !rdev->recovery_offset)) {
 679                 super_sync(rdev->mddev, rdev);
 680
 681                 set_bit(FirstUse, &rdev->flags);
 682
 683                 /* Force writing of superblocks to disk */
 684                 set_bit(MD_CHANGE_DEVS, &rdev->mddev->flags);
 685
 686                 /* Any superblock is better than none, choose that if given */
 687                 return refdev ? 0 : 1;
 688         }
 689
 690         if (!refdev)
 691                 return 1;
 692
 693         events_sb = le64_to_cpu(sb->events);
 694
 695         refsb = page_address(refdev->sb_page);
 696         events_refsb = le64_to_cpu(refsb->events);
 697
 698         return (events_sb > events_refsb) ? 1 : 0;
 699 }
 700
 701 static int super_init_validation(struct mddev *mddev, struct md_rdev *rdev)
 702 {
 703         int role;
 704         struct raid_set *rs = container_of(mddev, struct raid_set, md);
 705         uint64_t events_sb;
 706         uint64_t failed_devices;
 707         struct dm_raid_superblock *sb;
 708         uint32_t new_devs = 0;
 709         uint32_t rebuilds = 0;
 710         struct md_rdev *r, *t;
 711         struct dm_raid_superblock *sb2;
 712
 713         sb = page_address(rdev->sb_page);
 714         events_sb = le64_to_cpu(sb->events);
 715         failed_devices = le64_to_cpu(sb->failed_devices);
 716
 717         /*
 718          * Initialise to 1 if this is a new superblock.
 719          */
 720         mddev->events = events_sb ? : 1;
 721
 722         /*
 723          * Reshaping is not currently allowed
 724          */
 725         if ((le32_to_cpu(sb->level) != mddev->level) ||
 726             (le32_to_cpu(sb->layout) != mddev->layout) ||
 727             (le32_to_cpu(sb->stripe_sectors) != mddev->chunk_sectors)) {
 728                 DMERR("Reshaping arrays not yet supported.");
 729                 return -EINVAL;
 730         }
 731
 732         /* We can only change the number of devices in RAID1 right now */
 733         if ((rs->raid_type->level != 1) &&
 734             (le32_to_cpu(sb->num_devices) != mddev->raid_disks)) {
 735                 DMERR("Reshaping arrays not yet supported.");
 736                 return -EINVAL;
 737         }
 738
 739         if (!(rs->print_flags & (DMPF_SYNC | DMPF_NOSYNC)))
 740                 mddev->recovery_cp = le64_to_cpu(sb->array_resync_offset);
 741
 742         /*
 743          * During load, we set FirstUse if a new superblock was written.
 744          * There are two reasons we might not have a superblock:
 745          * 1) The array is brand new - in which case, all of the
 746          *    devices must have their In_sync bit set.  Also,
 747          *    recovery_cp must be 0, unless forced.
 748          * 2) This is a new device being added to an old array
 749          *    and the new device needs to be rebuilt - in which
 750          *    case the In_sync bit will /not/ be set and
 751          *    recovery_cp must be MaxSector.
 752          */
 753         rdev_for_each(r, t, mddev) {
 754                 if (!test_bit(In_sync, &r->flags)) {
 755                         DMINFO("Device %d specified for rebuild: "
 756                                "Clearing superblock", r->raid_disk);
 757                         rebuilds++;
 758                 } else if (test_bit(FirstUse, &r->flags))
 759                         new_devs++;
 760         }
 761
 762         if (!rebuilds) {
 763                 if (new_devs == mddev->raid_disks) {
 764                         DMINFO("Superblocks created for new array");
 765                         set_bit(MD_ARRAY_FIRST_USE, &mddev->flags);
 766                 } else if (new_devs) {
 767                         DMERR("New device injected "
 768                               "into existing array without 'rebuild' "
 769                               "parameter specified");
 770                         return -EINVAL;
 771                 }
 772         } else if (new_devs) {
 773                 DMERR("'rebuild' devices cannot be "
 774                       "injected into an array with other first-time devices");
 775                 return -EINVAL;
 776         } else if (mddev->recovery_cp != MaxSector) {
 777                 DMERR("'rebuild' specified while array is not in-sync");
 778                 return -EINVAL;
 779         }
 780
 781         /*
 782          * Now we set the Faulty bit for those devices that are
 783          * recorded in the superblock as failed.
 784          */
 785         rdev_for_each(r, t, mddev) {
 786                 if (!r->sb_page)
 787                         continue;
 788                 sb2 = page_address(r->sb_page);
 789                 sb2->failed_devices = 0;
 790
 791                 /*
 792                  * Check for any device re-ordering.
 793                  */
 794                 if (!test_bit(FirstUse, &r->flags) && (r->raid_disk >= 0)) {
 795                         role = le32_to_cpu(sb2->array_position);
 796                         if (role != r->raid_disk) {
 797                                 if (rs->raid_type->level != 1) {
 798                                         rs->ti->error = "Cannot change device "
 799                                                 "positions in RAID array";
 800                                         return -EINVAL;
 801                                 }
 802                                 DMINFO("RAID1 device #%d now at position #%d",
 803                                        role, r->raid_disk);
 804                         }
 805
 806                         /*
 807                          * Partial recovery is performed on
 808                          * returning failed devices.
 809                          */
 810                         if (failed_devices & (1 << role))
 811                                 set_bit(Faulty, &r->flags);
 812                 }
 813         }
 814
 815         return 0;
 816 }
 817
 818 static int super_validate(struct mddev *mddev, struct md_rdev *rdev)
 819 {
 820         struct dm_raid_superblock *sb = page_address(rdev->sb_page);
 821
 822         /*
 823          * If mddev->events is not set, we know we have not yet initialized
 824          * the array.
 825          */
 826         if (!mddev->events && super_init_validation(mddev, rdev))
 827                 return -EINVAL;
 828
 829         mddev->bitmap_info.offset = 4096 >> 9; /* Enable bitmap creation */
 830         rdev->mddev->bitmap_info.default_offset = 4096 >> 9;
 831         if (!test_bit(FirstUse, &rdev->flags)) {
 832                 rdev->recovery_offset = le64_to_cpu(sb->disk_recovery_offset);
 833                 if (rdev->recovery_offset != MaxSector)
 834                         clear_bit(In_sync, &rdev->flags);
 835         }
 836
 837         /*
 838          * If a device comes back, set it as not In_sync and no longer faulty.
 839          */
 840         if (test_bit(Faulty, &rdev->flags)) {
 841                 clear_bit(Faulty, &rdev->flags);
 842                 clear_bit(In_sync, &rdev->flags);
 843                 rdev->saved_raid_disk = rdev->raid_disk;
 844                 rdev->recovery_offset = 0;
 845         }
 846
 847         clear_bit(FirstUse, &rdev->flags);
 848
 849         return 0;
 850 }
 851
 852 /*
 853  * Analyse superblocks and select the freshest.
 854  */
 855 static int analyse_superblocks(struct dm_target *ti, struct raid_set *rs)
 856 {
 857         int ret;
 858         struct md_rdev *rdev, *freshest, *tmp;
 859         struct mddev *mddev = &rs->md;
 860
 861         freshest = NULL;
 862         rdev_for_each(rdev, tmp, mddev) {
 863                 if (!rdev->meta_bdev)
 864                         continue;
 865
 866                 ret = super_load(rdev, freshest);
 867
 868                 switch (ret) {
 869                 case 1:
 870                         freshest = rdev;
 871                         break;
 872                 case 0:
 873                         break;
 874                 default:
 875                         ti->error = "Failed to load superblock";
 876                         return ret;
 877                 }
 878         }
 879
 880         if (!freshest)
 881                 return 0;
 882
 883         /*
 884          * Validation of the freshest device provides the source of
 885          * validation for the remaining devices.
 886          */
 887         ti->error = "Unable to assemble array: Invalid superblocks";
 888         if (super_validate(mddev, freshest))
 889                 return -EINVAL;
 890
 891         rdev_for_each(rdev, tmp, mddev)
 892                 if ((rdev != freshest) && super_validate(mddev, rdev))
 893                         return -EINVAL;
 894
 895         return 0;
 896 }
 897
 898 /*
 899  * Construct a RAID4/5/6 mapping:
 900  * Args:
 901  *      <raid_type> <#raid_params> <raid_params>                \
 902  *      <#raid_devs> { <meta_dev1> <dev1> .. <meta_devN> <devN> }
 903  *
 904  * <raid_params> varies by <raid_type>.  See 'parse_raid_params' for
 905  * details on possible <raid_params>.
 906  */
 907 static int raid_ctr(struct dm_target *ti, unsigned argc, char **argv)
 908 {
 909         int ret;
 910         struct raid_type *rt;
 911         unsigned long num_raid_params, num_raid_devs;
 912         struct raid_set *rs = NULL;
 913
 914         /* Must have at least <raid_type> <#raid_params> */
 915         if (argc < 2) {
 916                 ti->error = "Too few arguments";
 917                 return -EINVAL;
 918         }
 919
 920         /* raid type */
 921         rt = get_raid_type(argv[0]);
 922         if (!rt) {
 923                 ti->error = "Unrecognised raid_type";
 924                 return -EINVAL;
 925         }
 926         argc--;
 927         argv++;
 928
 929         /* number of RAID parameters */
 930         if (strict_strtoul(argv[0], 10, &num_raid_params) < 0) {
 931                 ti->error = "Cannot understand number of RAID parameters";
 932                 return -EINVAL;
 933         }
 934         argc--;
 935         argv++;
 936
 937         /* Skip over RAID params for now and find out # of devices */
 938         if (num_raid_params + 1 > argc) {
 939                 ti->error = "Arguments do not agree with counts given";
 940                 return -EINVAL;
 941         }
 942
 943         if ((strict_strtoul(argv[num_raid_params], 10, &num_raid_devs) < 0) ||
 944             (num_raid_devs >= INT_MAX)) {
 945                 ti->error = "Cannot understand number of raid devices";
 946                 return -EINVAL;
 947         }
 948
 949         rs = context_alloc(ti, rt, (unsigned)num_raid_devs);
 950         if (IS_ERR(rs))
 951                 return PTR_ERR(rs);
 952
 953         ret = parse_raid_params(rs, argv, (unsigned)num_raid_params);
 954         if (ret)
 955                 goto bad;
 956
 957         ret = -EINVAL;
 958
 959         argc -= num_raid_params + 1; /* +1: we already have num_raid_devs */
 960         argv += num_raid_params + 1;
 961
 962         if (argc != (num_raid_devs * 2)) {
 963                 ti->error = "Supplied RAID devices does not match the count given";
 964                 goto bad;
 965         }
 966
 967         ret = dev_parms(rs, argv);
 968         if (ret)
 969                 goto bad;
 970
 971         rs->md.sync_super = super_sync;
 972         ret = analyse_superblocks(ti, rs);
 973         if (ret)
 974                 goto bad;
 975
 976         INIT_WORK(&rs->md.event_work, do_table_event);
 977         ti->private = rs;
 978
 979         mutex_lock(&rs->md.reconfig_mutex);
 980         ret = md_run(&rs->md);
 981         rs->md.in_sync = 0; /* Assume already marked dirty */
 982         mutex_unlock(&rs->md.reconfig_mutex);
 983
 984         if (ret) {
 985                 ti->error = "Fail to run raid array";
 986                 goto bad;
 987         }
 988
 989         rs->callbacks.congested_fn = raid_is_congested;
 990         dm_table_add_target_callbacks(ti->table, &rs->callbacks);
 991
 992         mddev_suspend(&rs->md);
 993         return 0;
 994
 995 bad:
 996         context_free(rs);
 997
 998         return ret;
 999 }
1000
1001 static void raid_dtr(struct dm_target *ti)
1002 {
1003         struct raid_set *rs = ti->private;
1004
1005         list_del_init(&rs->callbacks.list);
1006         md_stop(&rs->md);
1007         context_free(rs);
1008 }
1009
1010 static int raid_map(struct dm_target *ti, struct bio *bio, union map_info *map_context)
1011 {
1012         struct raid_set *rs = ti->private;
1013         struct mddev *mddev = &rs->md;
1014
1015         mddev->pers->make_request(mddev, bio);
1016
1017         return DM_MAPIO_SUBMITTED;
1018 }
1019
1020 static int raid_status(struct dm_target *ti, status_type_t type,
1021                        char *result, unsigned maxlen)
1022 {
1023         struct raid_set *rs = ti->private;
1024         unsigned raid_param_cnt = 1; /* at least 1 for chunksize */
1025         unsigned sz = 0;
1026         int i, array_in_sync = 0;
1027         sector_t sync;
1028
1029         switch (type) {
1030         case STATUSTYPE_INFO:
1031                 DMEMIT("%s %d ", rs->raid_type->name, rs->md.raid_disks);
1032
1033                 if (test_bit(MD_RECOVERY_RUNNING, &rs->md.recovery))
1034                         sync = rs->md.curr_resync_completed;
1035                 else
1036                         sync = rs->md.recovery_cp;
1037
1038                 if (sync >= rs->md.resync_max_sectors) {
1039                         array_in_sync = 1;
1040                         sync = rs->md.resync_max_sectors;
1041                 } else {
1042                         /*
1043                          * The array may be doing an initial sync, or it may
1044                          * be rebuilding individual components.  If all the
1045                          * devices are In_sync, then it is the array that is
1046                          * being initialized.
1047                          */
1048                         for (i = 0; i < rs->md.raid_disks; i++)
1049                                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
1050                                         array_in_sync = 1;
1051                 }
1052                 /*
1053                  * Status characters:
1054                  *  'D' = Dead/Failed device
1055                  *  'a' = Alive but not in-sync
1056                  *  'A' = Alive and in-sync
1057                  */
1058                 for (i = 0; i < rs->md.raid_disks; i++) {
1059                         if (test_bit(Faulty, &rs->dev[i].rdev.flags))
1060                                 DMEMIT("D");
1061                         else if (!array_in_sync ||
1062                                  !test_bit(In_sync, &rs->dev[i].rdev.flags))
1063                                 DMEMIT("a");
1064                         else
1065                                 DMEMIT("A");
1066                 }
1067
1068                 /*
1069                  * In-sync ratio:
1070                  *  The in-sync ratio shows the progress of:
1071                  *   - Initializing the array
1072                  *   - Rebuilding a subset of devices of the array
1073                  *  The user can distinguish between the two by referring
1074                  *  to the status characters.
1075                  */
1076                 DMEMIT(" %llu/%llu",
1077                        (unsigned long long) sync,
1078                        (unsigned long long) rs->md.resync_max_sectors);
1079
1080                 break;
1081         case STATUSTYPE_TABLE:
1082                 /* The string you would use to construct this array */
1083                 for (i = 0; i < rs->md.raid_disks; i++) {
1084                         if ((rs->print_flags & DMPF_REBUILD) &&
1085                             rs->dev[i].data_dev &&
1086                             !test_bit(In_sync, &rs->dev[i].rdev.flags))
1087                                 raid_param_cnt += 2; /* for rebuilds */
1088                         if (rs->dev[i].data_dev &&
1089                             test_bit(WriteMostly, &rs->dev[i].rdev.flags))
1090                                 raid_param_cnt += 2;
1091                 }
1092
1093                 raid_param_cnt += (hweight32(rs->print_flags & ~DMPF_REBUILD) * 2);
1094                 if (rs->print_flags & (DMPF_SYNC | DMPF_NOSYNC))
1095                         raid_param_cnt--;
1096
1097                 DMEMIT("%s %u %u", rs->raid_type->name,
1098                        raid_param_cnt, rs->md.chunk_sectors);
1099
1100                 if ((rs->print_flags & DMPF_SYNC) &&
1101                     (rs->md.recovery_cp == MaxSector))
1102                         DMEMIT(" sync");
1103                 if (rs->print_flags & DMPF_NOSYNC)
1104                         DMEMIT(" nosync");
1105
1106                 for (i = 0; i < rs->md.raid_disks; i++)
1107                         if ((rs->print_flags & DMPF_REBUILD) &&
1108                             rs->dev[i].data_dev &&
1109                             !test_bit(In_sync, &rs->dev[i].rdev.flags))
1110                                 DMEMIT(" rebuild %u", i);
1111
1112                 if (rs->print_flags & DMPF_DAEMON_SLEEP)
1113                         DMEMIT(" daemon_sleep %lu",
1114                                rs->md.bitmap_info.daemon_sleep);
1115
1116                 if (rs->print_flags & DMPF_MIN_RECOVERY_RATE)
1117                         DMEMIT(" min_recovery_rate %d", rs->md.sync_speed_min);
1118
1119                 if (rs->print_flags & DMPF_MAX_RECOVERY_RATE)
1120                         DMEMIT(" max_recovery_rate %d", rs->md.sync_speed_max);
1121
1122                 for (i = 0; i < rs->md.raid_disks; i++)
1123                         if (rs->dev[i].data_dev &&
1124                             test_bit(WriteMostly, &rs->dev[i].rdev.flags))
1125                                 DMEMIT(" write_mostly %u", i);
1126
1127                 if (rs->print_flags & DMPF_MAX_WRITE_BEHIND)
1128                         DMEMIT(" max_write_behind %lu",
1129                                rs->md.bitmap_info.max_write_behind);
1130
1131                 if (rs->print_flags & DMPF_STRIPE_CACHE) {
1132                         struct r5conf *conf = rs->md.private;
1133
1134                         /* convert from kiB to sectors */
1135                         DMEMIT(" stripe_cache %d",
1136                                conf ? conf->max_nr_stripes * 2 : 0);
1137                 }
1138
1139                 if (rs->print_flags & DMPF_REGION_SIZE)
1140                         DMEMIT(" region_size %lu",
1141                                rs->md.bitmap_info.chunksize >> 9);
1142
1143                 DMEMIT(" %d", rs->md.raid_disks);
1144                 for (i = 0; i < rs->md.raid_disks; i++) {
1145                         if (rs->dev[i].meta_dev)
1146                                 DMEMIT(" %s", rs->dev[i].meta_dev->name);
1147                         else
1148                                 DMEMIT(" -");
1149
1150                         if (rs->dev[i].data_dev)
1151                                 DMEMIT(" %s", rs->dev[i].data_dev->name);
1152                         else
1153                                 DMEMIT(" -");
1154                 }
1155         }
1156
1157         return 0;
1158 }
1159
1160 static int raid_iterate_devices(struct dm_target *ti, iterate_devices_callout_fn fn, void *data)
1161 {
1162         struct raid_set *rs = ti->private;
1163         unsigned i;
1164         int ret = 0;
1165
1166         for (i = 0; !ret && i < rs->md.raid_disks; i++)
1167                 if (rs->dev[i].data_dev)
1168                         ret = fn(ti,
1169                                  rs->dev[i].data_dev,
1170                                  0, /* No offset on data devs */
1171                                  rs->md.dev_sectors,
1172                                  data);
1173
1174         return ret;
1175 }
1176
1177 static void raid_io_hints(struct dm_target *ti, struct queue_limits *limits)
1178 {
1179         struct raid_set *rs = ti->private;
1180         unsigned chunk_size = rs->md.chunk_sectors << 9;
1181         struct r5conf *conf = rs->md.private;
1182
1183         blk_limits_io_min(limits, chunk_size);
1184         blk_limits_io_opt(limits, chunk_size * (conf->raid_disks - conf->max_degraded));
1185 }
1186
1187 static void raid_presuspend(struct dm_target *ti)
1188 {
1189         struct raid_set *rs = ti->private;
1190
1191         md_stop_writes(&rs->md);
1192 }
1193
1194 static void raid_postsuspend(struct dm_target *ti)
1195 {
1196         struct raid_set *rs = ti->private;
1197
1198         mddev_suspend(&rs->md);
1199 }
1200
1201 static void raid_resume(struct dm_target *ti)
1202 {
1203         struct raid_set *rs = ti->private;
1204
1205         if (!rs->bitmap_loaded) {
1206                 bitmap_load(&rs->md);
1207                 rs->bitmap_loaded = 1;
1208         } else
1209                 md_wakeup_thread(rs->md.thread);
1210
1211         mddev_resume(&rs->md);
1212 }
1213
1214 static struct target_type raid_target = {
1215         .name = "raid",
1216         .version = {1, 1, 0},
1217         .module = THIS_MODULE,
1218         .ctr = raid_ctr,
1219         .dtr = raid_dtr,
1220         .map = raid_map,
1221         .status = raid_status,
1222         .iterate_devices = raid_iterate_devices,
1223         .io_hints = raid_io_hints,
1224         .presuspend = raid_presuspend,
1225         .postsuspend = raid_postsuspend,
1226         .resume = raid_resume,
1227 };
1228
1229 static int __init dm_raid_init(void)
1230 {
1231         return dm_register_target(&raid_target);
1232 }
1233
1234 static void __exit dm_raid_exit(void)
1235 {
1236         dm_unregister_target(&raid_target);
1237 }
1238
1239 module_init(dm_raid_init);
1240 module_exit(dm_raid_exit);
1241
1242 MODULE_DESCRIPTION(DM_NAME " raid4/5/6 target");
1243 MODULE_ALIAS("dm-raid4");
1244 MODULE_ALIAS("dm-raid5");
1245 MODULE_ALIAS("dm-raid6");
1246 MODULE_AUTHOR("Neil Brown <dm-devel@redhat.com>");
1247 MODULE_LICENSE("GPL");