drivers/md/persistent-data/dm-space-map-metadata.c

   1 /*
   2  * Copyright (C) 2011 Red Hat, Inc.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include "dm-space-map.h"
   8 #include "dm-space-map-common.h"
   9 #include "dm-space-map-metadata.h"
  10
  11 #include <linux/list.h>
  12 #include <linux/slab.h>
  13 #include <linux/device-mapper.h>
  14
  15 #define DM_MSG_PREFIX "space map metadata"
  16
  17 /*----------------------------------------------------------------*/
  18
  19 /*
  20  * Space map interface.
  21  *
  22  * The low level disk format is written using the standard btree and
  23  * transaction manager.  This means that performing disk operations may
  24  * cause us to recurse into the space map in order to allocate new blocks.
  25  * For this reason we have a pool of pre-allocated blocks large enough to
  26  * service any metadata_ll_disk operation.
  27  */
  28
  29 /*
  30  * FIXME: we should calculate this based on the size of the device.
  31  * Only the metadata space map needs this functionality.
  32  */
  33 #define MAX_RECURSIVE_ALLOCATIONS 1024
  34
  35 enum block_op_type {
  36         BOP_INC,
  37         BOP_DEC
  38 };
  39
  40 struct block_op {
  41         enum block_op_type type;
  42         dm_block_t block;
  43 };
  44
  45 struct sm_metadata {
  46         struct dm_space_map sm;
  47
  48         struct ll_disk ll;
  49         struct ll_disk old_ll;
  50
  51         dm_block_t begin;
  52
  53         unsigned recursion_count;
  54         unsigned allocated_this_transaction;
  55         unsigned nr_uncommitted;
  56         struct block_op uncommitted[MAX_RECURSIVE_ALLOCATIONS];
  57 };
  58
  59 static int add_bop(struct sm_metadata *smm, enum block_op_type type, dm_block_t b)
  60 {
  61         struct block_op *op;
  62
  63         if (smm->nr_uncommitted == MAX_RECURSIVE_ALLOCATIONS) {
  64                 DMERR("too many recursive allocations");
  65                 return -ENOMEM;
  66         }
  67
  68         op = smm->uncommitted + smm->nr_uncommitted++;
  69         op->type = type;
  70         op->block = b;
  71
  72         return 0;
  73 }
  74
  75 static int commit_bop(struct sm_metadata *smm, struct block_op *op)
  76 {
  77         int r = 0;
  78         enum allocation_event ev;
  79
  80         switch (op->type) {
  81         case BOP_INC:
  82                 r = sm_ll_inc(&smm->ll, op->block, &ev);
  83                 break;
  84
  85         case BOP_DEC:
  86                 r = sm_ll_dec(&smm->ll, op->block, &ev);
  87                 break;
  88         }
  89
  90         return r;
  91 }
  92
  93 static void in(struct sm_metadata *smm)
  94 {
  95         smm->recursion_count++;
  96 }
  97
  98 static int out(struct sm_metadata *smm)
  99 {
 100         int r = 0;
 101
 102         /*
 103          * If we're not recursing then very bad things are happening.
 104          */
 105         if (!smm->recursion_count) {
 106                 DMERR("lost track of recursion depth");
 107                 return -ENOMEM;
 108         }
 109
 110         if (smm->recursion_count == 1 && smm->nr_uncommitted) {
 111                 while (smm->nr_uncommitted && !r) {
 112                         smm->nr_uncommitted--;
 113                         r = commit_bop(smm, smm->uncommitted +
 114                                        smm->nr_uncommitted);
 115                         if (r)
 116                                 break;
 117                 }
 118         }
 119
 120         smm->recursion_count--;
 121
 122         return r;
 123 }
 124
 125 /*
 126  * When using the out() function above, we often want to combine an error
 127  * code for the operation run in the recursive context with that from
 128  * out().
 129  */
 130 static int combine_errors(int r1, int r2)
 131 {
 132         return r1 ? r1 : r2;
 133 }
 134
 135 static int recursing(struct sm_metadata *smm)
 136 {
 137         return smm->recursion_count;
 138 }
 139
 140 static void sm_metadata_destroy(struct dm_space_map *sm)
 141 {
 142         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 143
 144         kfree(smm);
 145 }
 146
 147 static int sm_metadata_get_nr_blocks(struct dm_space_map *sm, dm_block_t *count)
 148 {
 149         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 150
 151         *count = smm->ll.nr_blocks;
 152
 153         return 0;
 154 }
 155
 156 static int sm_metadata_get_nr_free(struct dm_space_map *sm, dm_block_t *count)
 157 {
 158         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 159
 160         *count = smm->old_ll.nr_blocks - smm->old_ll.nr_allocated -
 161                  smm->allocated_this_transaction;
 162
 163         return 0;
 164 }
 165
 166 static int sm_metadata_get_count(struct dm_space_map *sm, dm_block_t b,
 167                                  uint32_t *result)
 168 {
 169         int r, i;
 170         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 171         unsigned adjustment = 0;
 172
 173         /*
 174          * We may have some uncommitted adjustments to add.  This list
 175          * should always be really short.
 176          */
 177         for (i = 0; i < smm->nr_uncommitted; i++) {
 178                 struct block_op *op = smm->uncommitted + i;
 179
 180                 if (op->block != b)
 181                         continue;
 182
 183                 switch (op->type) {
 184                 case BOP_INC:
 185                         adjustment++;
 186                         break;
 187
 188                 case BOP_DEC:
 189                         adjustment--;
 190                         break;
 191                 }
 192         }
 193
 194         r = sm_ll_lookup(&smm->ll, b, result);
 195         if (r)
 196                 return r;
 197
 198         *result += adjustment;
 199
 200         return 0;
 201 }
 202
 203 static int sm_metadata_count_is_more_than_one(struct dm_space_map *sm,
 204                                               dm_block_t b, int *result)
 205 {
 206         int r, i, adjustment = 0;
 207         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 208         uint32_t rc;
 209
 210         /*
 211          * We may have some uncommitted adjustments to add.  This list
 212          * should always be really short.
 213          */
 214         for (i = 0; i < smm->nr_uncommitted; i++) {
 215                 struct block_op *op = smm->uncommitted + i;
 216
 217                 if (op->block != b)
 218                         continue;
 219
 220                 switch (op->type) {
 221                 case BOP_INC:
 222                         adjustment++;
 223                         break;
 224
 225                 case BOP_DEC:
 226                         adjustment--;
 227                         break;
 228                 }
 229         }
 230
 231         if (adjustment > 1) {
 232                 *result = 1;
 233                 return 0;
 234         }
 235
 236         r = sm_ll_lookup_bitmap(&smm->ll, b, &rc);
 237         if (r)
 238                 return r;
 239
 240         if (rc == 3)
 241                 /*
 242                  * We err on the side of caution, and always return true.
 243                  */
 244                 *result = 1;
 245         else
 246                 *result = rc + adjustment > 1;
 247
 248         return 0;
 249 }
 250
 251 static int sm_metadata_set_count(struct dm_space_map *sm, dm_block_t b,
 252                                  uint32_t count)
 253 {
 254         int r, r2;
 255         enum allocation_event ev;
 256         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 257
 258         if (smm->recursion_count) {
 259                 DMERR("cannot recurse set_count()");
 260                 return -EINVAL;
 261         }
 262
 263         in(smm);
 264         r = sm_ll_insert(&smm->ll, b, count, &ev);
 265         r2 = out(smm);
 266
 267         return combine_errors(r, r2);
 268 }
 269
 270 static int sm_metadata_inc_block(struct dm_space_map *sm, dm_block_t b)
 271 {
 272         int r, r2 = 0;
 273         enum allocation_event ev;
 274         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 275
 276         if (recursing(smm))
 277                 r = add_bop(smm, BOP_INC, b);
 278         else {
 279                 in(smm);
 280                 r = sm_ll_inc(&smm->ll, b, &ev);
 281                 r2 = out(smm);
 282         }
 283
 284         return combine_errors(r, r2);
 285 }
 286
 287 static int sm_metadata_dec_block(struct dm_space_map *sm, dm_block_t b)
 288 {
 289         int r, r2 = 0;
 290         enum allocation_event ev;
 291         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 292
 293         if (recursing(smm))
 294                 r = add_bop(smm, BOP_DEC, b);
 295         else {
 296                 in(smm);
 297                 r = sm_ll_dec(&smm->ll, b, &ev);
 298                 r2 = out(smm);
 299         }
 300
 301         return combine_errors(r, r2);
 302 }
 303
 304 static int sm_metadata_new_block_(struct dm_space_map *sm, dm_block_t *b)
 305 {
 306         int r, r2 = 0;
 307         enum allocation_event ev;
 308         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 309
 310         r = sm_ll_find_free_block(&smm->old_ll, smm->begin, smm->old_ll.nr_blocks, b);
 311         if (r)
 312                 return r;
 313
 314         smm->begin = *b + 1;
 315
 316         if (recursing(smm))
 317                 r = add_bop(smm, BOP_INC, *b);
 318         else {
 319                 in(smm);
 320                 r = sm_ll_inc(&smm->ll, *b, &ev);
 321                 r2 = out(smm);
 322         }
 323
 324         if (!r)
 325                 smm->allocated_this_transaction++;
 326
 327         return combine_errors(r, r2);
 328 }
 329
 330 static int sm_metadata_new_block(struct dm_space_map *sm, dm_block_t *b)
 331 {
 332         int r = sm_metadata_new_block_(sm, b);
 333         if (r)
 334                 DMERR("unable to allocate new metadata block");
 335         return r;
 336 }
 337
 338 static int sm_metadata_commit(struct dm_space_map *sm)
 339 {
 340         int r;
 341         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 342
 343         r = sm_ll_commit(&smm->ll);
 344         if (r)
 345                 return r;
 346
 347         memcpy(&smm->old_ll, &smm->ll, sizeof(smm->old_ll));
 348         smm->begin = 0;
 349         smm->allocated_this_transaction = 0;
 350
 351         return 0;
 352 }
 353
 354 static int sm_metadata_root_size(struct dm_space_map *sm, size_t *result)
 355 {
 356         *result = sizeof(struct disk_sm_root);
 357
 358         return 0;
 359 }
 360
 361 static int sm_metadata_copy_root(struct dm_space_map *sm, void *where_le, size_t max)
 362 {
 363         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 364         struct disk_sm_root root_le;
 365
 366         root_le.nr_blocks = cpu_to_le64(smm->ll.nr_blocks);
 367         root_le.nr_allocated = cpu_to_le64(smm->ll.nr_allocated);
 368         root_le.bitmap_root = cpu_to_le64(smm->ll.bitmap_root);
 369         root_le.ref_count_root = cpu_to_le64(smm->ll.ref_count_root);
 370
 371         if (max < sizeof(root_le))
 372                 return -ENOSPC;
 373
 374         memcpy(where_le, &root_le, sizeof(root_le));
 375
 376         return 0;
 377 }
 378
 379 static int sm_metadata_extend(struct dm_space_map *sm, dm_block_t extra_blocks);
 380
 381 static struct dm_space_map ops = {
 382         .destroy = sm_metadata_destroy,
 383         .extend = sm_metadata_extend,
 384         .get_nr_blocks = sm_metadata_get_nr_blocks,
 385         .get_nr_free = sm_metadata_get_nr_free,
 386         .get_count = sm_metadata_get_count,
 387         .count_is_more_than_one = sm_metadata_count_is_more_than_one,
 388         .set_count = sm_metadata_set_count,
 389         .inc_block = sm_metadata_inc_block,
 390         .dec_block = sm_metadata_dec_block,
 391         .new_block = sm_metadata_new_block,
 392         .commit = sm_metadata_commit,
 393         .root_size = sm_metadata_root_size,
 394         .copy_root = sm_metadata_copy_root,
 395         .register_threshold_callback = NULL
 396 };
 397
 398 /*----------------------------------------------------------------*/
 399
 400 /*
 401  * When a new space map is created that manages its own space.  We use
 402  * this tiny bootstrap allocator.
 403  */
 404 static void sm_bootstrap_destroy(struct dm_space_map *sm)
 405 {
 406 }
 407
 408 static int sm_bootstrap_extend(struct dm_space_map *sm, dm_block_t extra_blocks)
 409 {
 410         DMERR("bootstrap doesn't support extend");
 411
 412         return -EINVAL;
 413 }
 414
 415 static int sm_bootstrap_get_nr_blocks(struct dm_space_map *sm, dm_block_t *count)
 416 {
 417         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 418
 419         return smm->ll.nr_blocks;
 420 }
 421
 422 static int sm_bootstrap_get_nr_free(struct dm_space_map *sm, dm_block_t *count)
 423 {
 424         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 425
 426         *count = smm->ll.nr_blocks - smm->begin;
 427
 428         return 0;
 429 }
 430
 431 static int sm_bootstrap_get_count(struct dm_space_map *sm, dm_block_t b,
 432                                   uint32_t *result)
 433 {
 434         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 435
 436         return b < smm->begin ? 1 : 0;
 437 }
 438
 439 static int sm_bootstrap_count_is_more_than_one(struct dm_space_map *sm,
 440                                                dm_block_t b, int *result)
 441 {
 442         *result = 0;
 443
 444         return 0;
 445 }
 446
 447 static int sm_bootstrap_set_count(struct dm_space_map *sm, dm_block_t b,
 448                                   uint32_t count)
 449 {
 450         DMERR("bootstrap doesn't support set_count");
 451
 452         return -EINVAL;
 453 }
 454
 455 static int sm_bootstrap_new_block(struct dm_space_map *sm, dm_block_t *b)
 456 {
 457         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 458
 459         /*
 460          * We know the entire device is unused.
 461          */
 462         if (smm->begin == smm->ll.nr_blocks)
 463                 return -ENOSPC;
 464
 465         *b = smm->begin++;
 466
 467         return 0;
 468 }
 469
 470 static int sm_bootstrap_inc_block(struct dm_space_map *sm, dm_block_t b)
 471 {
 472         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 473
 474         return add_bop(smm, BOP_INC, b);
 475 }
 476
 477 static int sm_bootstrap_dec_block(struct dm_space_map *sm, dm_block_t b)
 478 {
 479         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 480
 481         return add_bop(smm, BOP_DEC, b);
 482 }
 483
 484 static int sm_bootstrap_commit(struct dm_space_map *sm)
 485 {
 486         return 0;
 487 }
 488
 489 static int sm_bootstrap_root_size(struct dm_space_map *sm, size_t *result)
 490 {
 491         DMERR("bootstrap doesn't support root_size");
 492
 493         return -EINVAL;
 494 }
 495
 496 static int sm_bootstrap_copy_root(struct dm_space_map *sm, void *where,
 497                                   size_t max)
 498 {
 499         DMERR("bootstrap doesn't support copy_root");
 500
 501         return -EINVAL;
 502 }
 503
 504 static struct dm_space_map bootstrap_ops = {
 505         .destroy = sm_bootstrap_destroy,
 506         .extend = sm_bootstrap_extend,
 507         .get_nr_blocks = sm_bootstrap_get_nr_blocks,
 508         .get_nr_free = sm_bootstrap_get_nr_free,
 509         .get_count = sm_bootstrap_get_count,
 510         .count_is_more_than_one = sm_bootstrap_count_is_more_than_one,
 511         .set_count = sm_bootstrap_set_count,
 512         .inc_block = sm_bootstrap_inc_block,
 513         .dec_block = sm_bootstrap_dec_block,
 514         .new_block = sm_bootstrap_new_block,
 515         .commit = sm_bootstrap_commit,
 516         .root_size = sm_bootstrap_root_size,
 517         .copy_root = sm_bootstrap_copy_root,
 518         .register_threshold_callback = NULL
 519 };
 520
 521 /*----------------------------------------------------------------*/
 522
 523 static int sm_metadata_extend(struct dm_space_map *sm, dm_block_t extra_blocks)
 524 {
 525         int r, i;
 526         enum allocation_event ev;
 527         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 528         dm_block_t old_len = smm->ll.nr_blocks;
 529
 530         /*
 531          * Flick into a mode where all blocks get allocated in the new area.
 532          */
 533         smm->begin = old_len;
 534         memcpy(&smm->sm, &bootstrap_ops, sizeof(smm->sm));
 535
 536         /*
 537          * Extend.
 538          */
 539         r = sm_ll_extend(&smm->ll, extra_blocks);
 540
 541         /*
 542          * Switch back to normal behaviour.
 543          */
 544         memcpy(&smm->sm, &ops, sizeof(smm->sm));
 545         for (i = old_len; !r && i < smm->begin; i++)
 546                 r = sm_ll_inc(&smm->ll, i, &ev);
 547
 548         return r;
 549 }
 550
 551 /*----------------------------------------------------------------*/
 552
 553 struct dm_space_map *dm_sm_metadata_init(void)
 554 {
 555         struct sm_metadata *smm;
 556
 557         smm = kmalloc(sizeof(*smm), GFP_KERNEL);
 558         if (!smm)
 559                 return ERR_PTR(-ENOMEM);
 560
 561         memcpy(&smm->sm, &ops, sizeof(smm->sm));
 562
 563         return &smm->sm;
 564 }
 565
 566 int dm_sm_metadata_create(struct dm_space_map *sm,
 567                           struct dm_transaction_manager *tm,
 568                           dm_block_t nr_blocks,
 569                           dm_block_t superblock)
 570 {
 571         int r;
 572         dm_block_t i;
 573         enum allocation_event ev;
 574         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 575
 576         smm->begin = superblock + 1;
 577         smm->recursion_count = 0;
 578         smm->allocated_this_transaction = 0;
 579         smm->nr_uncommitted = 0;
 580
 581         memcpy(&smm->sm, &bootstrap_ops, sizeof(smm->sm));
 582
 583         r = sm_ll_new_metadata(&smm->ll, tm);
 584         if (r)
 585                 return r;
 586
 587         r = sm_ll_extend(&smm->ll, nr_blocks);
 588         if (r)
 589                 return r;
 590
 591         memcpy(&smm->sm, &ops, sizeof(smm->sm));
 592
 593         /*
 594          * Now we need to update the newly created data structures with the
 595          * allocated blocks that they were built from.
 596          */
 597         for (i = superblock; !r && i < smm->begin; i++)
 598                 r = sm_ll_inc(&smm->ll, i, &ev);
 599
 600         if (r)
 601                 return r;
 602
 603         return sm_metadata_commit(sm);
 604 }
 605
 606 int dm_sm_metadata_open(struct dm_space_map *sm,
 607                         struct dm_transaction_manager *tm,
 608                         void *root_le, size_t len)
 609 {
 610         int r;
 611         struct sm_metadata *smm = container_of(sm, struct sm_metadata, sm);
 612
 613         r = sm_ll_open_metadata(&smm->ll, tm, root_le, len);
 614         if (r)
 615                 return r;
 616
 617         smm->begin = 0;
 618         smm->recursion_count = 0;
 619         smm->allocated_this_transaction = 0;
 620         smm->nr_uncommitted = 0;
 621
 622         memcpy(&smm->old_ll, &smm->ll, sizeof(smm->old_ll));
 623         return 0;
 624 }