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         uint64_t print_flags;
  60
  61         struct mddev md;
  62         struct raid_type *raid_type;
  63         struct dm_target_callbacks callbacks;
  64
  65         struct raid_dev dev[0];
  66 };
  67
  68 /* Supported raid types and properties. */
  69 static struct raid_type {
  70         const char *name;               /* RAID algorithm. */
  71         const char *descr;              /* Descriptor text for logging. */
  72         const unsigned parity_devs;     /* # of parity devices. */
  73         const unsigned minimal_devs;    /* minimal # of devices in set. */
  74         const unsigned level;           /* RAID level. */
  75         const unsigned algorithm;       /* RAID algorithm. */
  76 } raid_types[] = {
  77         {"raid1",    "RAID1 (mirroring)",               0, 2, 1, 0 /* NONE */},
  78         {"raid4",    "RAID4 (dedicated parity disk)",   1, 2, 5, ALGORITHM_PARITY_0},
  79         {"raid5_la", "RAID5 (left asymmetric)",         1, 2, 5, ALGORITHM_LEFT_ASYMMETRIC},
  80         {"raid5_ra", "RAID5 (right asymmetric)",        1, 2, 5, ALGORITHM_RIGHT_ASYMMETRIC},
  81         {"raid5_ls", "RAID5 (left symmetric)",          1, 2, 5, ALGORITHM_LEFT_SYMMETRIC},
  82         {"raid5_rs", "RAID5 (right symmetric)",         1, 2, 5, ALGORITHM_RIGHT_SYMMETRIC},
  83         {"raid6_zr", "RAID6 (zero restart)",            2, 4, 6, ALGORITHM_ROTATING_ZERO_RESTART},
  84         {"raid6_nr", "RAID6 (N restart)",               2, 4, 6, ALGORITHM_ROTATING_N_RESTART},
  85         {"raid6_nc", "RAID6 (N continue)",              2, 4, 6, ALGORITHM_ROTATING_N_CONTINUE}
  86 };
  87
  88 static struct raid_type *get_raid_type(char *name)
  89 {
  90         int i;
  91
  92         for (i = 0; i < ARRAY_SIZE(raid_types); i++)
  93                 if (!strcmp(raid_types[i].name, name))
  94                         return &raid_types[i];
  95
  96         return NULL;
  97 }
  98
  99 static struct raid_set *context_alloc(struct dm_target *ti, struct raid_type *raid_type, unsigned raid_devs)
 100 {
 101         unsigned i;
 102         struct raid_set *rs;
 103         sector_t sectors_per_dev;
 104
 105         if (raid_devs <= raid_type->parity_devs) {
 106                 ti->error = "Insufficient number of devices";
 107                 return ERR_PTR(-EINVAL);
 108         }
 109
 110         sectors_per_dev = ti->len;
 111         if ((raid_type->level > 1) &&
 112             sector_div(sectors_per_dev, (raid_devs - raid_type->parity_devs))) {
 113                 ti->error = "Target length not divisible by number of data devices";
 114                 return ERR_PTR(-EINVAL);
 115         }
 116
 117         rs = kzalloc(sizeof(*rs) + raid_devs * sizeof(rs->dev[0]), GFP_KERNEL);
 118         if (!rs) {
 119                 ti->error = "Cannot allocate raid context";
 120                 return ERR_PTR(-ENOMEM);
 121         }
 122
 123         mddev_init(&rs->md);
 124
 125         rs->ti = ti;
 126         rs->raid_type = raid_type;
 127         rs->md.raid_disks = raid_devs;
 128         rs->md.level = raid_type->level;
 129         rs->md.new_level = rs->md.level;
 130         rs->md.dev_sectors = sectors_per_dev;
 131         rs->md.layout = raid_type->algorithm;
 132         rs->md.new_layout = rs->md.layout;
 133         rs->md.delta_disks = 0;
 134         rs->md.recovery_cp = 0;
 135
 136         for (i = 0; i < raid_devs; i++)
 137                 md_rdev_init(&rs->dev[i].rdev);
 138
 139         /*
 140          * Remaining items to be initialized by further RAID params:
 141          *  rs->md.persistent
 142          *  rs->md.external
 143          *  rs->md.chunk_sectors
 144          *  rs->md.new_chunk_sectors
 145          */
 146
 147         return rs;
 148 }
 149
 150 static void context_free(struct raid_set *rs)
 151 {
 152         int i;
 153
 154         for (i = 0; i < rs->md.raid_disks; i++) {
 155                 if (rs->dev[i].meta_dev)
 156                         dm_put_device(rs->ti, rs->dev[i].meta_dev);
 157                 if (rs->dev[i].rdev.sb_page)
 158                         put_page(rs->dev[i].rdev.sb_page);
 159                 rs->dev[i].rdev.sb_page = NULL;
 160                 rs->dev[i].rdev.sb_loaded = 0;
 161                 if (rs->dev[i].data_dev)
 162                         dm_put_device(rs->ti, rs->dev[i].data_dev);
 163         }
 164
 165         kfree(rs);
 166 }
 167
 168 /*
 169  * For every device we have two words
 170  *  <meta_dev>: meta device name or '-' if missing
 171  *  <data_dev>: data device name or '-' if missing
 172  *
 173  * The following are permitted:
 174  *    - -
 175  *    - <data_dev>
 176  *    <meta_dev> <data_dev>
 177  *
 178  * The following is not allowed:
 179  *    <meta_dev> -
 180  *
 181  * This code parses those words.  If there is a failure,
 182  * the caller must use context_free to unwind the operations.
 183  */
 184 static int dev_parms(struct raid_set *rs, char **argv)
 185 {
 186         int i;
 187         int rebuild = 0;
 188         int metadata_available = 0;
 189         int ret = 0;
 190
 191         for (i = 0; i < rs->md.raid_disks; i++, argv += 2) {
 192                 rs->dev[i].rdev.raid_disk = i;
 193
 194                 rs->dev[i].meta_dev = NULL;
 195                 rs->dev[i].data_dev = NULL;
 196
 197                 /*
 198                  * There are no offsets, since there is a separate device
 199                  * for data and metadata.
 200                  */
 201                 rs->dev[i].rdev.data_offset = 0;
 202                 rs->dev[i].rdev.mddev = &rs->md;
 203
 204                 if (strcmp(argv[0], "-")) {
 205                         ret = dm_get_device(rs->ti, argv[0],
 206                                             dm_table_get_mode(rs->ti->table),
 207                                             &rs->dev[i].meta_dev);
 208                         rs->ti->error = "RAID metadata device lookup failure";
 209                         if (ret)
 210                                 return ret;
 211
 212                         rs->dev[i].rdev.sb_page = alloc_page(GFP_KERNEL);
 213                         if (!rs->dev[i].rdev.sb_page)
 214                                 return -ENOMEM;
 215                 }
 216
 217                 if (!strcmp(argv[1], "-")) {
 218                         if (!test_bit(In_sync, &rs->dev[i].rdev.flags) &&
 219                             (!rs->dev[i].rdev.recovery_offset)) {
 220                                 rs->ti->error = "Drive designated for rebuild not specified";
 221                                 return -EINVAL;
 222                         }
 223
 224                         rs->ti->error = "No data device supplied with metadata device";
 225                         if (rs->dev[i].meta_dev)
 226                                 return -EINVAL;
 227
 228                         continue;
 229                 }
 230
 231                 ret = dm_get_device(rs->ti, argv[1],
 232                                     dm_table_get_mode(rs->ti->table),
 233                                     &rs->dev[i].data_dev);
 234                 if (ret) {
 235                         rs->ti->error = "RAID device lookup failure";
 236                         return ret;
 237                 }
 238
 239                 if (rs->dev[i].meta_dev) {
 240                         metadata_available = 1;
 241                         rs->dev[i].rdev.meta_bdev = rs->dev[i].meta_dev->bdev;
 242                 }
 243                 rs->dev[i].rdev.bdev = rs->dev[i].data_dev->bdev;
 244                 list_add(&rs->dev[i].rdev.same_set, &rs->md.disks);
 245                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
 246                         rebuild++;
 247         }
 248
 249         if (metadata_available) {
 250                 rs->md.external = 0;
 251                 rs->md.persistent = 1;
 252                 rs->md.major_version = 2;
 253         } else if (rebuild && !rs->md.recovery_cp) {
 254                 /*
 255                  * Without metadata, we will not be able to tell if the array
 256                  * is in-sync or not - we must assume it is not.  Therefore,
 257                  * it is impossible to rebuild a drive.
 258                  *
 259                  * Even if there is metadata, the on-disk information may
 260                  * indicate that the array is not in-sync and it will then
 261                  * fail at that time.
 262                  *
 263                  * User could specify 'nosync' option if desperate.
 264                  */
 265                 DMERR("Unable to rebuild drive while array is not in-sync");
 266                 rs->ti->error = "RAID device lookup failure";
 267                 return -EINVAL;
 268         }
 269
 270         return 0;
 271 }
 272
 273 /*
 274  * validate_region_size
 275  * @rs
 276  * @region_size:  region size in sectors.  If 0, pick a size (4MiB default).
 277  *
 278  * Set rs->md.bitmap_info.chunksize (which really refers to 'region size').
 279  * Ensure that (ti->len/region_size < 2^21) - required by MD bitmap.
 280  *
 281  * Returns: 0 on success, -EINVAL on failure.
 282  */
 283 static int validate_region_size(struct raid_set *rs, unsigned long region_size)
 284 {
 285         unsigned long min_region_size = rs->ti->len / (1 << 21);
 286
 287         if (!region_size) {
 288                 /*
 289                  * Choose a reasonable default.  All figures in sectors.
 290                  */
 291                 if (min_region_size > (1 << 13)) {
 292                         DMINFO("Choosing default region size of %lu sectors",
 293                                region_size);
 294                         region_size = min_region_size;
 295                 } else {
 296                         DMINFO("Choosing default region size of 4MiB");
 297                         region_size = 1 << 13; /* sectors */
 298                 }
 299         } else {
 300                 /*
 301                  * Validate user-supplied value.
 302                  */
 303                 if (region_size > rs->ti->len) {
 304                         rs->ti->error = "Supplied region size is too large";
 305                         return -EINVAL;
 306                 }
 307
 308                 if (region_size < min_region_size) {
 309                         DMERR("Supplied region_size (%lu sectors) below minimum (%lu)",
 310                               region_size, min_region_size);
 311                         rs->ti->error = "Supplied region size is too small";
 312                         return -EINVAL;
 313                 }
 314
 315                 if (!is_power_of_2(region_size)) {
 316                         rs->ti->error = "Region size is not a power of 2";
 317                         return -EINVAL;
 318                 }
 319
 320                 if (region_size < rs->md.chunk_sectors) {
 321                         rs->ti->error = "Region size is smaller than the chunk size";
 322                         return -EINVAL;
 323                 }
 324         }
 325
 326         /*
 327          * Convert sectors to bytes.
 328          */
 329         rs->md.bitmap_info.chunksize = (region_size << 9);
 330
 331         return 0;
 332 }
 333
 334 /*
 335  * Possible arguments are...
 336  *      <chunk_size> [optional_args]
 337  *
 338  * Argument definitions
 339  *    <chunk_size>                      The number of sectors per disk that
 340  *                                      will form the "stripe"
 341  *    [[no]sync]                        Force or prevent recovery of the
 342  *                                      entire array
 343  *    [rebuild <idx>]                   Rebuild the drive indicated by the index
 344  *    [daemon_sleep <ms>]               Time between bitmap daemon work to
 345  *                                      clear bits
 346  *    [min_recovery_rate <kB/sec/disk>] Throttle RAID initialization
 347  *    [max_recovery_rate <kB/sec/disk>] Throttle RAID initialization
 348  *    [write_mostly <idx>]              Indicate a write mostly drive via index
 349  *    [max_write_behind <sectors>]      See '-write-behind=' (man mdadm)
 350  *    [stripe_cache <sectors>]          Stripe cache size for higher RAIDs
 351  *    [region_size <sectors>]           Defines granularity of bitmap
 352  */
 353 static int parse_raid_params(struct raid_set *rs, char **argv,
 354                              unsigned num_raid_params)
 355 {
 356         unsigned i, rebuild_cnt = 0;
 357         unsigned long value, region_size = 0;
 358         char *key;
 359
 360         /*
 361          * First, parse the in-order required arguments
 362          * "chunk_size" is the only argument of this type.
 363          */
 364         if ((strict_strtoul(argv[0], 10, &value) < 0)) {
 365                 rs->ti->error = "Bad chunk size";
 366                 return -EINVAL;
 367         } else if (rs->raid_type->level == 1) {
 368                 if (value)
 369                         DMERR("Ignoring chunk size parameter for RAID 1");
 370                 value = 0;
 371         } else if (!is_power_of_2(value)) {
 372                 rs->ti->error = "Chunk size must be a power of 2";
 373                 return -EINVAL;
 374         } else if (value < 8) {
 375                 rs->ti->error = "Chunk size value is too small";
 376                 return -EINVAL;
 377         }
 378
 379         rs->md.new_chunk_sectors = rs->md.chunk_sectors = value;
 380         argv++;
 381         num_raid_params--;
 382
 383         /*
 384          * We set each individual device as In_sync with a completed
 385          * 'recovery_offset'.  If there has been a device failure or
 386          * replacement then one of the following cases applies:
 387          *
 388          *   1) User specifies 'rebuild'.
 389          *      - Device is reset when param is read.
 390          *   2) A new device is supplied.
 391          *      - No matching superblock found, resets device.
 392          *   3) Device failure was transient and returns on reload.
 393          *      - Failure noticed, resets device for bitmap replay.
 394          *   4) Device hadn't completed recovery after previous failure.
 395          *      - Superblock is read and overrides recovery_offset.
 396          *
 397          * What is found in the superblocks of the devices is always
 398          * authoritative, unless 'rebuild' or '[no]sync' was specified.
 399          */
 400         for (i = 0; i < rs->md.raid_disks; i++) {
 401                 set_bit(In_sync, &rs->dev[i].rdev.flags);
 402                 rs->dev[i].rdev.recovery_offset = MaxSector;
 403         }
 404
 405         /*
 406          * Second, parse the unordered optional arguments
 407          */
 408         for (i = 0; i < num_raid_params; i++) {
 409                 if (!strcasecmp(argv[i], "nosync")) {
 410                         rs->md.recovery_cp = MaxSector;
 411                         rs->print_flags |= DMPF_NOSYNC;
 412                         continue;
 413                 }
 414                 if (!strcasecmp(argv[i], "sync")) {
 415                         rs->md.recovery_cp = 0;
 416                         rs->print_flags |= DMPF_SYNC;
 417                         continue;
 418                 }
 419
 420                 /* The rest of the optional arguments come in key/value pairs */
 421                 if ((i + 1) >= num_raid_params) {
 422                         rs->ti->error = "Wrong number of raid parameters given";
 423                         return -EINVAL;
 424                 }
 425
 426                 key = argv[i++];
 427                 if (strict_strtoul(argv[i], 10, &value) < 0) {
 428                         rs->ti->error = "Bad numerical argument given in raid params";
 429                         return -EINVAL;
 430                 }
 431
 432                 if (!strcasecmp(key, "rebuild")) {
 433                         rebuild_cnt++;
 434                         if (((rs->raid_type->level != 1) &&
 435                              (rebuild_cnt > rs->raid_type->parity_devs)) ||
 436                             ((rs->raid_type->level == 1) &&
 437                              (rebuild_cnt > (rs->md.raid_disks - 1)))) {
 438                                 rs->ti->error = "Too many rebuild devices specified for given RAID type";
 439                                 return -EINVAL;
 440                         }
 441                         if (value > rs->md.raid_disks) {
 442                                 rs->ti->error = "Invalid rebuild index given";
 443                                 return -EINVAL;
 444                         }
 445                         clear_bit(In_sync, &rs->dev[value].rdev.flags);
 446                         rs->dev[value].rdev.recovery_offset = 0;
 447                         rs->print_flags |= DMPF_REBUILD;
 448                 } else if (!strcasecmp(key, "write_mostly")) {
 449                         if (rs->raid_type->level != 1) {
 450                                 rs->ti->error = "write_mostly option is only valid for RAID1";
 451                                 return -EINVAL;
 452                         }
 453                         if (value >= rs->md.raid_disks) {
 454                                 rs->ti->error = "Invalid write_mostly drive index given";
 455                                 return -EINVAL;
 456                         }
 457                         set_bit(WriteMostly, &rs->dev[value].rdev.flags);
 458                 } else if (!strcasecmp(key, "max_write_behind")) {
 459                         if (rs->raid_type->level != 1) {
 460                                 rs->ti->error = "max_write_behind option is only valid for RAID1";
 461                                 return -EINVAL;
 462                         }
 463                         rs->print_flags |= DMPF_MAX_WRITE_BEHIND;
 464
 465                         /*
 466                          * In device-mapper, we specify things in sectors, but
 467                          * MD records this value in kB
 468                          */
 469                         value /= 2;
 470                         if (value > COUNTER_MAX) {
 471                                 rs->ti->error = "Max write-behind limit out of range";
 472                                 return -EINVAL;
 473                         }
 474                         rs->md.bitmap_info.max_write_behind = value;
 475                 } else if (!strcasecmp(key, "daemon_sleep")) {
 476                         rs->print_flags |= DMPF_DAEMON_SLEEP;
 477                         if (!value || (value > MAX_SCHEDULE_TIMEOUT)) {
 478                                 rs->ti->error = "daemon sleep period out of range";
 479                                 return -EINVAL;
 480                         }
 481                         rs->md.bitmap_info.daemon_sleep = value;
 482                 } else if (!strcasecmp(key, "stripe_cache")) {
 483                         rs->print_flags |= DMPF_STRIPE_CACHE;
 484
 485                         /*
 486                          * In device-mapper, we specify things in sectors, but
 487                          * MD records this value in kB
 488                          */
 489                         value /= 2;
 490
 491                         if (rs->raid_type->level < 5) {
 492                                 rs->ti->error = "Inappropriate argument: stripe_cache";
 493                                 return -EINVAL;
 494                         }
 495                         if (raid5_set_cache_size(&rs->md, (int)value)) {
 496                                 rs->ti->error = "Bad stripe_cache size";
 497                                 return -EINVAL;
 498                         }
 499                 } else if (!strcasecmp(key, "min_recovery_rate")) {
 500                         rs->print_flags |= DMPF_MIN_RECOVERY_RATE;
 501                         if (value > INT_MAX) {
 502                                 rs->ti->error = "min_recovery_rate out of range";
 503                                 return -EINVAL;
 504                         }
 505                         rs->md.sync_speed_min = (int)value;
 506                 } else if (!strcasecmp(key, "max_recovery_rate")) {
 507                         rs->print_flags |= DMPF_MAX_RECOVERY_RATE;
 508                         if (value > INT_MAX) {
 509                                 rs->ti->error = "max_recovery_rate out of range";
 510                                 return -EINVAL;
 511                         }
 512                         rs->md.sync_speed_max = (int)value;
 513                 } else if (!strcasecmp(key, "region_size")) {
 514                         rs->print_flags |= DMPF_REGION_SIZE;
 515                         region_size = value;
 516                 } else {
 517                         DMERR("Unable to parse RAID parameter: %s", key);
 518                         rs->ti->error = "Unable to parse RAID parameters";
 519                         return -EINVAL;
 520                 }
 521         }
 522
 523         if (validate_region_size(rs, region_size))
 524                 return -EINVAL;
 525
 526         if (rs->md.chunk_sectors)
 527                 rs->ti->split_io = rs->md.chunk_sectors;
 528         else
 529                 rs->ti->split_io = region_size;
 530
 531         if (rs->md.chunk_sectors)
 532                 rs->ti->split_io = rs->md.chunk_sectors;
 533         else
 534                 rs->ti->split_io = region_size;
 535
 536         /* Assume there are no metadata devices until the drives are parsed */
 537         rs->md.persistent = 0;
 538         rs->md.external = 1;
 539
 540         return 0;
 541 }
 542
 543 static void do_table_event(struct work_struct *ws)
 544 {
 545         struct raid_set *rs = container_of(ws, struct raid_set, md.event_work);
 546
 547         dm_table_event(rs->ti->table);
 548 }
 549
 550 static int raid_is_congested(struct dm_target_callbacks *cb, int bits)
 551 {
 552         struct raid_set *rs = container_of(cb, struct raid_set, callbacks);
 553
 554         if (rs->raid_type->level == 1)
 555                 return md_raid1_congested(&rs->md, bits);
 556
 557         return md_raid5_congested(&rs->md, bits);
 558 }
 559
 560 /*
 561  * This structure is never routinely used by userspace, unlike md superblocks.
 562  * Devices with this superblock should only ever be accessed via device-mapper.
 563  */
 564 #define DM_RAID_MAGIC 0x64526D44
 565 struct dm_raid_superblock {
 566         __le32 magic;           /* "DmRd" */
 567         __le32 features;        /* Used to indicate possible future changes */
 568
 569         __le32 num_devices;     /* Number of devices in this array. (Max 64) */
 570         __le32 array_position;  /* The position of this drive in the array */
 571
 572         __le64 events;          /* Incremented by md when superblock updated */
 573         __le64 failed_devices;  /* Bit field of devices to indicate failures */
 574
 575         /*
 576          * This offset tracks the progress of the repair or replacement of
 577          * an individual drive.
 578          */
 579         __le64 disk_recovery_offset;
 580
 581         /*
 582          * This offset tracks the progress of the initial array
 583          * synchronisation/parity calculation.
 584          */
 585         __le64 array_resync_offset;
 586
 587         /*
 588          * RAID characteristics
 589          */
 590         __le32 level;
 591         __le32 layout;
 592         __le32 stripe_sectors;
 593
 594         /* Remainder of a logical block is zero-filled when writing (see super_sync()). */
 595 } __packed;
 596
 597 static int read_disk_sb(struct md_rdev *rdev, int size)
 598 {
 599         BUG_ON(!rdev->sb_page);
 600
 601         if (rdev->sb_loaded)
 602                 return 0;
 603
 604         if (!sync_page_io(rdev, 0, size, rdev->sb_page, READ, 1)) {
 605                 DMERR("Failed to read device superblock");
 606                 return -EINVAL;
 607         }
 608
 609         rdev->sb_loaded = 1;
 610
 611         return 0;
 612 }
 613
 614 static void super_sync(struct mddev *mddev, struct md_rdev *rdev)
 615 {
 616         struct md_rdev *r, *t;
 617         uint64_t failed_devices;
 618         struct dm_raid_superblock *sb;
 619
 620         sb = page_address(rdev->sb_page);
 621         failed_devices = le64_to_cpu(sb->failed_devices);
 622
 623         rdev_for_each(r, t, mddev)
 624                 if ((r->raid_disk >= 0) && test_bit(Faulty, &r->flags))
 625                         failed_devices |= (1ULL << r->raid_disk);
 626
 627         memset(sb + 1, 0, rdev->sb_size - sizeof(*sb));
 628
 629         sb->magic = cpu_to_le32(DM_RAID_MAGIC);
 630         sb->features = cpu_to_le32(0);  /* No features yet */
 631
 632         sb->num_devices = cpu_to_le32(mddev->raid_disks);
 633         sb->array_position = cpu_to_le32(rdev->raid_disk);
 634
 635         sb->events = cpu_to_le64(mddev->events);
 636         sb->failed_devices = cpu_to_le64(failed_devices);
 637
 638         sb->disk_recovery_offset = cpu_to_le64(rdev->recovery_offset);
 639         sb->array_resync_offset = cpu_to_le64(mddev->recovery_cp);
 640
 641         sb->level = cpu_to_le32(mddev->level);
 642         sb->layout = cpu_to_le32(mddev->layout);
 643         sb->stripe_sectors = cpu_to_le32(mddev->chunk_sectors);
 644 }
 645
 646 /*
 647  * super_load
 648  *
 649  * This function creates a superblock if one is not found on the device
 650  * and will decide which superblock to use if there's a choice.
 651  *
 652  * Return: 1 if use rdev, 0 if use refdev, -Exxx otherwise
 653  */
 654 static int super_load(struct md_rdev *rdev, struct md_rdev *refdev)
 655 {
 656         int ret;
 657         struct dm_raid_superblock *sb;
 658         struct dm_raid_superblock *refsb;
 659         uint64_t events_sb, events_refsb;
 660
 661         rdev->sb_start = 0;
 662         rdev->sb_size = bdev_logical_block_size(rdev->meta_bdev);
 663         if (rdev->sb_size < sizeof(*sb) || rdev->sb_size > PAGE_SIZE) {
 664                 DMERR("superblock size of a logical block is no longer valid");
 665                 return -EINVAL;
 666         }
 667
 668         ret = read_disk_sb(rdev, rdev->sb_size);
 669         if (ret)
 670                 return ret;
 671
 672         sb = page_address(rdev->sb_page);
 673
 674         /*
 675          * Two cases that we want to write new superblocks and rebuild:
 676          * 1) New device (no matching magic number)
 677          * 2) Device specified for rebuild (!In_sync w/ offset == 0)
 678          */
 679         if ((sb->magic != cpu_to_le32(DM_RAID_MAGIC)) ||
 680             (!test_bit(In_sync, &rdev->flags) && !rdev->recovery_offset)) {
 681                 super_sync(rdev->mddev, rdev);
 682
 683                 set_bit(FirstUse, &rdev->flags);
 684
 685                 /* Force writing of superblocks to disk */
 686                 set_bit(MD_CHANGE_DEVS, &rdev->mddev->flags);
 687
 688                 /* Any superblock is better than none, choose that if given */
 689                 return refdev ? 0 : 1;
 690         }
 691
 692         if (!refdev)
 693                 return 1;
 694
 695         events_sb = le64_to_cpu(sb->events);
 696
 697         refsb = page_address(refdev->sb_page);
 698         events_refsb = le64_to_cpu(refsb->events);
 699
 700         return (events_sb > events_refsb) ? 1 : 0;
 701 }
 702
 703 static int super_init_validation(struct mddev *mddev, struct md_rdev *rdev)
 704 {
 705         int role;
 706         struct raid_set *rs = container_of(mddev, struct raid_set, md);
 707         uint64_t events_sb;
 708         uint64_t failed_devices;
 709         struct dm_raid_superblock *sb;
 710         uint32_t new_devs = 0;
 711         uint32_t rebuilds = 0;
 712         struct md_rdev *r, *t;
 713         struct dm_raid_superblock *sb2;
 714
 715         sb = page_address(rdev->sb_page);
 716         events_sb = le64_to_cpu(sb->events);
 717         failed_devices = le64_to_cpu(sb->failed_devices);
 718
 719         /*
 720          * Initialise to 1 if this is a new superblock.
 721          */
 722         mddev->events = events_sb ? : 1;
 723
 724         /*
 725          * Reshaping is not currently allowed
 726          */
 727         if ((le32_to_cpu(sb->level) != mddev->level) ||
 728             (le32_to_cpu(sb->layout) != mddev->layout) ||
 729             (le32_to_cpu(sb->stripe_sectors) != mddev->chunk_sectors)) {
 730                 DMERR("Reshaping arrays not yet supported.");
 731                 return -EINVAL;
 732         }
 733
 734         /* We can only change the number of devices in RAID1 right now */
 735         if ((rs->raid_type->level != 1) &&
 736             (le32_to_cpu(sb->num_devices) != mddev->raid_disks)) {
 737                 DMERR("Reshaping arrays not yet supported.");
 738                 return -EINVAL;
 739         }
 740
 741         if (!(rs->print_flags & (DMPF_SYNC | DMPF_NOSYNC)))
 742                 mddev->recovery_cp = le64_to_cpu(sb->array_resync_offset);
 743
 744         /*
 745          * During load, we set FirstUse if a new superblock was written.
 746          * There are two reasons we might not have a superblock:
 747          * 1) The array is brand new - in which case, all of the
 748          *    devices must have their In_sync bit set.  Also,
 749          *    recovery_cp must be 0, unless forced.
 750          * 2) This is a new device being added to an old array
 751          *    and the new device needs to be rebuilt - in which
 752          *    case the In_sync bit will /not/ be set and
 753          *    recovery_cp must be MaxSector.
 754          */
 755         rdev_for_each(r, t, mddev) {
 756                 if (!test_bit(In_sync, &r->flags)) {
 757                         DMINFO("Device %d specified for rebuild: "
 758                                "Clearing superblock", r->raid_disk);
 759                         rebuilds++;
 760                 } else if (test_bit(FirstUse, &r->flags))
 761                         new_devs++;
 762         }
 763
 764         if (!rebuilds) {
 765                 if (new_devs == mddev->raid_disks) {
 766                         DMINFO("Superblocks created for new array");
 767                         set_bit(MD_ARRAY_FIRST_USE, &mddev->flags);
 768                 } else if (new_devs) {
 769                         DMERR("New device injected "
 770                               "into existing array without 'rebuild' "
 771                               "parameter specified");
 772                         return -EINVAL;
 773                 }
 774         } else if (new_devs) {
 775                 DMERR("'rebuild' devices cannot be "
 776                       "injected into an array with other first-time devices");
 777                 return -EINVAL;
 778         } else if (mddev->recovery_cp != MaxSector) {
 779                 DMERR("'rebuild' specified while array is not in-sync");
 780                 return -EINVAL;
 781         }
 782
 783         /*
 784          * Now we set the Faulty bit for those devices that are
 785          * recorded in the superblock as failed.
 786          */
 787         rdev_for_each(r, t, mddev) {
 788                 if (!r->sb_page)
 789                         continue;
 790                 sb2 = page_address(r->sb_page);
 791                 sb2->failed_devices = 0;
 792
 793                 /*
 794                  * Check for any device re-ordering.
 795                  */
 796                 if (!test_bit(FirstUse, &r->flags) && (r->raid_disk >= 0)) {
 797                         role = le32_to_cpu(sb2->array_position);
 798                         if (role != r->raid_disk) {
 799                                 if (rs->raid_type->level != 1) {
 800                                         rs->ti->error = "Cannot change device "
 801                                                 "positions in RAID array";
 802                                         return -EINVAL;
 803                                 }
 804                                 DMINFO("RAID1 device #%d now at position #%d",
 805                                        role, r->raid_disk);
 806                         }
 807
 808                         /*
 809                          * Partial recovery is performed on
 810                          * returning failed devices.
 811                          */
 812                         if (failed_devices & (1 << role))
 813                                 set_bit(Faulty, &r->flags);
 814                 }
 815         }
 816
 817         return 0;
 818 }
 819
 820 static int super_validate(struct mddev *mddev, struct md_rdev *rdev)
 821 {
 822         struct dm_raid_superblock *sb = page_address(rdev->sb_page);
 823
 824         /*
 825          * If mddev->events is not set, we know we have not yet initialized
 826          * the array.
 827          */
 828         if (!mddev->events && super_init_validation(mddev, rdev))
 829                 return -EINVAL;
 830
 831         mddev->bitmap_info.offset = 4096 >> 9; /* Enable bitmap creation */
 832         rdev->mddev->bitmap_info.default_offset = 4096 >> 9;
 833         if (!test_bit(FirstUse, &rdev->flags)) {
 834                 rdev->recovery_offset = le64_to_cpu(sb->disk_recovery_offset);
 835                 if (rdev->recovery_offset != MaxSector)
 836                         clear_bit(In_sync, &rdev->flags);
 837         }
 838
 839         /*
 840          * If a device comes back, set it as not In_sync and no longer faulty.
 841          */
 842         if (test_bit(Faulty, &rdev->flags)) {
 843                 clear_bit(Faulty, &rdev->flags);
 844                 clear_bit(In_sync, &rdev->flags);
 845                 rdev->saved_raid_disk = rdev->raid_disk;
 846                 rdev->recovery_offset = 0;
 847         }
 848
 849         clear_bit(FirstUse, &rdev->flags);
 850
 851         return 0;
 852 }
 853
 854 /*
 855  * Analyse superblocks and select the freshest.
 856  */
 857 static int analyse_superblocks(struct dm_target *ti, struct raid_set *rs)
 858 {
 859         int ret;
 860         struct md_rdev *rdev, *freshest, *tmp;
 861         struct mddev *mddev = &rs->md;
 862
 863         freshest = NULL;
 864         rdev_for_each(rdev, tmp, mddev) {
 865                 if (!rdev->meta_bdev)
 866                         continue;
 867
 868                 ret = super_load(rdev, freshest);
 869
 870                 switch (ret) {
 871                 case 1:
 872                         freshest = rdev;
 873                         break;
 874                 case 0:
 875                         break;
 876                 default:
 877                         ti->error = "Failed to load superblock";
 878                         return ret;
 879                 }
 880         }
 881
 882         if (!freshest)
 883                 return 0;
 884
 885         /*
 886          * Validation of the freshest device provides the source of
 887          * validation for the remaining devices.
 888          */
 889         ti->error = "Unable to assemble array: Invalid superblocks";
 890         if (super_validate(mddev, freshest))
 891                 return -EINVAL;
 892
 893         rdev_for_each(rdev, tmp, mddev)
 894                 if ((rdev != freshest) && super_validate(mddev, rdev))
 895                         return -EINVAL;
 896
 897         return 0;
 898 }
 899
 900 /*
 901  * Construct a RAID4/5/6 mapping:
 902  * Args:
 903  *      <raid_type> <#raid_params> <raid_params>                \
 904  *      <#raid_devs> { <meta_dev1> <dev1> .. <meta_devN> <devN> }
 905  *
 906  * <raid_params> varies by <raid_type>.  See 'parse_raid_params' for
 907  * details on possible <raid_params>.
 908  */
 909 static int raid_ctr(struct dm_target *ti, unsigned argc, char **argv)
 910 {
 911         int ret;
 912         struct raid_type *rt;
 913         unsigned long num_raid_params, num_raid_devs;
 914         struct raid_set *rs = NULL;
 915
 916         /* Must have at least <raid_type> <#raid_params> */
 917         if (argc < 2) {
 918                 ti->error = "Too few arguments";
 919                 return -EINVAL;
 920         }
 921
 922         /* raid type */
 923         rt = get_raid_type(argv[0]);
 924         if (!rt) {
 925                 ti->error = "Unrecognised raid_type";
 926                 return -EINVAL;
 927         }
 928         argc--;
 929         argv++;
 930
 931         /* number of RAID parameters */
 932         if (strict_strtoul(argv[0], 10, &num_raid_params) < 0) {
 933                 ti->error = "Cannot understand number of RAID parameters";
 934                 return -EINVAL;
 935         }
 936         argc--;
 937         argv++;
 938
 939         /* Skip over RAID params for now and find out # of devices */
 940         if (num_raid_params + 1 > argc) {
 941                 ti->error = "Arguments do not agree with counts given";
 942                 return -EINVAL;
 943         }
 944
 945         if ((strict_strtoul(argv[num_raid_params], 10, &num_raid_devs) < 0) ||
 946             (num_raid_devs >= INT_MAX)) {
 947                 ti->error = "Cannot understand number of raid devices";
 948                 return -EINVAL;
 949         }
 950
 951         rs = context_alloc(ti, rt, (unsigned)num_raid_devs);
 952         if (IS_ERR(rs))
 953                 return PTR_ERR(rs);
 954
 955         ret = parse_raid_params(rs, argv, (unsigned)num_raid_params);
 956         if (ret)
 957                 goto bad;
 958
 959         ret = -EINVAL;
 960
 961         argc -= num_raid_params + 1; /* +1: we already have num_raid_devs */
 962         argv += num_raid_params + 1;
 963
 964         if (argc != (num_raid_devs * 2)) {
 965                 ti->error = "Supplied RAID devices does not match the count given";
 966                 goto bad;
 967         }
 968
 969         ret = dev_parms(rs, argv);
 970         if (ret)
 971                 goto bad;
 972
 973         rs->md.sync_super = super_sync;
 974         ret = analyse_superblocks(ti, rs);
 975         if (ret)
 976                 goto bad;
 977
 978         INIT_WORK(&rs->md.event_work, do_table_event);
 979         ti->private = rs;
 980         ti->num_flush_requests = 1;
 981
 982         mutex_lock(&rs->md.reconfig_mutex);
 983         ret = md_run(&rs->md);
 984         rs->md.in_sync = 0; /* Assume already marked dirty */
 985         mutex_unlock(&rs->md.reconfig_mutex);
 986
 987         if (ret) {
 988                 ti->error = "Fail to run raid array";
 989                 goto bad;
 990         }
 991
 992         rs->callbacks.congested_fn = raid_is_congested;
 993         dm_table_add_target_callbacks(ti->table, &rs->callbacks);
 994
 995         mddev_suspend(&rs->md);
 996         return 0;
 997
 998 bad:
 999         context_free(rs);
1000
1001         return ret;
1002 }
1003
1004 static void raid_dtr(struct dm_target *ti)
1005 {
1006         struct raid_set *rs = ti->private;
1007
1008         list_del_init(&rs->callbacks.list);
1009         md_stop(&rs->md);
1010         context_free(rs);
1011 }
1012
1013 static int raid_map(struct dm_target *ti, struct bio *bio, union map_info *map_context)
1014 {
1015         struct raid_set *rs = ti->private;
1016         struct mddev *mddev = &rs->md;
1017
1018         mddev->pers->make_request(mddev, bio);
1019
1020         return DM_MAPIO_SUBMITTED;
1021 }
1022
1023 static void raid_status(struct dm_target *ti, status_type_t type,
1024                         char *result, unsigned maxlen)
1025 {
1026         struct raid_set *rs = ti->private;
1027         unsigned raid_param_cnt = 1; /* at least 1 for chunksize */
1028         unsigned sz = 0;
1029         int i, array_in_sync = 0;
1030         sector_t sync;
1031
1032         switch (type) {
1033         case STATUSTYPE_INFO:
1034                 DMEMIT("%s %d ", rs->raid_type->name, rs->md.raid_disks);
1035
1036                 if (test_bit(MD_RECOVERY_RUNNING, &rs->md.recovery))
1037                         sync = rs->md.curr_resync_completed;
1038                 else
1039                         sync = rs->md.recovery_cp;
1040
1041                 if (sync >= rs->md.resync_max_sectors) {
1042                         array_in_sync = 1;
1043                         sync = rs->md.resync_max_sectors;
1044                 } else {
1045                         /*
1046                          * The array may be doing an initial sync, or it may
1047                          * be rebuilding individual components.  If all the
1048                          * devices are In_sync, then it is the array that is
1049                          * being initialized.
1050                          */
1051                         for (i = 0; i < rs->md.raid_disks; i++)
1052                                 if (!test_bit(In_sync, &rs->dev[i].rdev.flags))
1053                                         array_in_sync = 1;
1054                 }
1055                 /*
1056                  * Status characters:
1057                  *  'D' = Dead/Failed device
1058                  *  'a' = Alive but not in-sync
1059                  *  'A' = Alive and in-sync
1060                  */
1061                 for (i = 0; i < rs->md.raid_disks; i++) {
1062                         if (test_bit(Faulty, &rs->dev[i].rdev.flags))
1063                                 DMEMIT("D");
1064                         else if (!array_in_sync ||
1065                                  !test_bit(In_sync, &rs->dev[i].rdev.flags))
1066                                 DMEMIT("a");
1067                         else
1068                                 DMEMIT("A");
1069                 }
1070
1071                 /*
1072                  * In-sync ratio:
1073                  *  The in-sync ratio shows the progress of:
1074                  *   - Initializing the array
1075                  *   - Rebuilding a subset of devices of the array
1076                  *  The user can distinguish between the two by referring
1077                  *  to the status characters.
1078                  */
1079                 DMEMIT(" %llu/%llu",
1080                        (unsigned long long) sync,
1081                        (unsigned long long) rs->md.resync_max_sectors);
1082
1083                 break;
1084         case STATUSTYPE_TABLE:
1085                 /* The string you would use to construct this array */
1086                 for (i = 0; i < rs->md.raid_disks; i++) {
1087                         if ((rs->print_flags & DMPF_REBUILD) &&
1088                             rs->dev[i].data_dev &&
1089                             !test_bit(In_sync, &rs->dev[i].rdev.flags))
1090                                 raid_param_cnt += 2; /* for rebuilds */
1091                         if (rs->dev[i].data_dev &&
1092                             test_bit(WriteMostly, &rs->dev[i].rdev.flags))
1093                                 raid_param_cnt += 2;
1094                 }
1095
1096                 raid_param_cnt += (hweight64(rs->print_flags & ~DMPF_REBUILD) * 2);
1097                 if (rs->print_flags & (DMPF_SYNC | DMPF_NOSYNC))
1098                         raid_param_cnt--;
1099
1100                 DMEMIT("%s %u %u", rs->raid_type->name,
1101                        raid_param_cnt, rs->md.chunk_sectors);
1102
1103                 if ((rs->print_flags & DMPF_SYNC) &&
1104                     (rs->md.recovery_cp == MaxSector))
1105                         DMEMIT(" sync");
1106                 if (rs->print_flags & DMPF_NOSYNC)
1107                         DMEMIT(" nosync");
1108
1109                 for (i = 0; i < rs->md.raid_disks; i++)
1110                         if ((rs->print_flags & DMPF_REBUILD) &&
1111                             rs->dev[i].data_dev &&
1112                             !test_bit(In_sync, &rs->dev[i].rdev.flags))
1113                                 DMEMIT(" rebuild %u", i);
1114
1115                 if (rs->print_flags & DMPF_DAEMON_SLEEP)
1116                         DMEMIT(" daemon_sleep %lu",
1117                                rs->md.bitmap_info.daemon_sleep);
1118
1119                 if (rs->print_flags & DMPF_MIN_RECOVERY_RATE)
1120                         DMEMIT(" min_recovery_rate %d", rs->md.sync_speed_min);
1121
1122                 if (rs->print_flags & DMPF_MAX_RECOVERY_RATE)
1123                         DMEMIT(" max_recovery_rate %d", rs->md.sync_speed_max);
1124
1125                 for (i = 0; i < rs->md.raid_disks; i++)
1126                         if (rs->dev[i].data_dev &&
1127                             test_bit(WriteMostly, &rs->dev[i].rdev.flags))
1128                                 DMEMIT(" write_mostly %u", i);
1129
1130                 if (rs->print_flags & DMPF_MAX_WRITE_BEHIND)
1131                         DMEMIT(" max_write_behind %lu",
1132                                rs->md.bitmap_info.max_write_behind);
1133
1134                 if (rs->print_flags & DMPF_STRIPE_CACHE) {
1135                         struct r5conf *conf = rs->md.private;
1136
1137                         /* convert from kiB to sectors */
1138                         DMEMIT(" stripe_cache %d",
1139                                conf ? conf->max_nr_stripes * 2 : 0);
1140                 }
1141
1142                 if (rs->print_flags & DMPF_REGION_SIZE)
1143                         DMEMIT(" region_size %lu",
1144                                rs->md.bitmap_info.chunksize >> 9);
1145
1146                 DMEMIT(" %d", rs->md.raid_disks);
1147                 for (i = 0; i < rs->md.raid_disks; i++) {
1148                         if (rs->dev[i].meta_dev)
1149                                 DMEMIT(" %s", rs->dev[i].meta_dev->name);
1150                         else
1151                                 DMEMIT(" -");
1152
1153                         if (rs->dev[i].data_dev)
1154                                 DMEMIT(" %s", rs->dev[i].data_dev->name);
1155                         else
1156                                 DMEMIT(" -");
1157                 }
1158         }
1159 }
1160
1161 static int raid_iterate_devices(struct dm_target *ti, iterate_devices_callout_fn fn, void *data)
1162 {
1163         struct raid_set *rs = ti->private;
1164         unsigned i;
1165         int ret = 0;
1166
1167         for (i = 0; !ret && i < rs->md.raid_disks; i++)
1168                 if (rs->dev[i].data_dev)
1169                         ret = fn(ti,
1170                                  rs->dev[i].data_dev,
1171                                  0, /* No offset on data devs */
1172                                  rs->md.dev_sectors,
1173                                  data);
1174
1175         return ret;
1176 }
1177
1178 static void raid_io_hints(struct dm_target *ti, struct queue_limits *limits)
1179 {
1180         struct raid_set *rs = ti->private;
1181         unsigned chunk_size = rs->md.chunk_sectors << 9;
1182         struct r5conf *conf = rs->md.private;
1183
1184         blk_limits_io_min(limits, chunk_size);
1185         blk_limits_io_opt(limits, chunk_size * (conf->raid_disks - conf->max_degraded));
1186 }
1187
1188 static void raid_presuspend(struct dm_target *ti)
1189 {
1190         struct raid_set *rs = ti->private;
1191
1192         md_stop_writes(&rs->md);
1193 }
1194
1195 static void raid_postsuspend(struct dm_target *ti)
1196 {
1197         struct raid_set *rs = ti->private;
1198
1199         mddev_suspend(&rs->md);
1200 }
1201
1202 static void raid_resume(struct dm_target *ti)
1203 {
1204         struct raid_set *rs = ti->private;
1205
1206         bitmap_load(&rs->md);
1207         mddev_resume(&rs->md);
1208 }
1209
1210 static struct target_type raid_target = {
1211         .name = "raid",
1212         .version = {1, 1, 1},
1213         .module = THIS_MODULE,
1214         .ctr = raid_ctr,
1215         .dtr = raid_dtr,
1216         .map = raid_map,
1217         .status = raid_status,
1218         .iterate_devices = raid_iterate_devices,
1219         .io_hints = raid_io_hints,
1220         .presuspend = raid_presuspend,
1221         .postsuspend = raid_postsuspend,
1222         .resume = raid_resume,
1223 };
1224
1225 static int __init dm_raid_init(void)
1226 {
1227         return dm_register_target(&raid_target);
1228 }
1229
1230 static void __exit dm_raid_exit(void)
1231 {
1232         dm_unregister_target(&raid_target);
1233 }
1234
1235 module_init(dm_raid_init);
1236 module_exit(dm_raid_exit);
1237
1238 MODULE_DESCRIPTION(DM_NAME " raid4/5/6 target");
1239 MODULE_ALIAS("dm-raid4");
1240 MODULE_ALIAS("dm-raid5");
1241 MODULE_ALIAS("dm-raid6");
1242 MODULE_AUTHOR("Neil Brown <dm-devel@redhat.com>");
1243 MODULE_LICENSE("GPL");