extent-tree.c

   1 #include <stdio.h>
   2 #include <stdlib.h>
   3 #include "kerncompat.h"
   4 #include "radix-tree.h"
   5 #include "ctree.h"
   6 #include "disk-io.h"
   7 #include "print-tree.h"
   8
   9 static int find_free_extent(struct ctree_root *orig_root, u64 num_blocks,
  10                             u64 search_start, u64 search_end, struct key *ins);
  11 static int finish_current_insert(struct ctree_root *extent_root);
  12 static int run_pending(struct ctree_root *extent_root);
  13
  14 /*
  15  * pending extents are blocks that we're trying to allocate in the extent
  16  * map while trying to grow the map because of other allocations.  To avoid
  17  * recursing, they are tagged in the radix tree and cleaned up after
  18  * other allocations are done.  The pending tag is also used in the same
  19  * manner for deletes.
  20  */
  21 #define CTREE_EXTENT_PENDING_DEL 0
  22
  23 static int inc_block_ref(struct ctree_root *root, u64 blocknr)
  24 {
  25         struct ctree_path path;
  26         int ret;
  27         struct key key;
  28         struct leaf *l;
  29         struct extent_item *item;
  30         struct key ins;
  31
  32         find_free_extent(root->extent_root, 0, 0, (u64)-1, &ins);
  33         init_path(&path);
  34         key.objectid = blocknr;
  35         key.flags = 0;
  36         key.offset = 1;
  37         ret = search_slot(root->extent_root, &key, &path, 0, 1);
  38         if (ret != 0)
  39                 BUG();
  40         BUG_ON(ret != 0);
  41         l = &path.nodes[0]->leaf;
  42         item = (struct extent_item *)(l->data +
  43                                       l->items[path.slots[0]].offset);
  44         item->refs++;
  45
  46         BUG_ON(list_empty(&path.nodes[0]->dirty));
  47         release_path(root->extent_root, &path);
  48         finish_current_insert(root->extent_root);
  49         run_pending(root->extent_root);
  50         return 0;
  51 }
  52
  53 static int lookup_block_ref(struct ctree_root *root, u64 blocknr, u32 *refs)
  54 {
  55         struct ctree_path path;
  56         int ret;
  57         struct key key;
  58         struct leaf *l;
  59         struct extent_item *item;
  60         init_path(&path);
  61         key.objectid = blocknr;
  62         key.flags = 0;
  63         key.offset = 1;
  64         ret = search_slot(root->extent_root, &key, &path, 0, 0);
  65         if (ret != 0)
  66                 BUG();
  67         l = &path.nodes[0]->leaf;
  68         item = (struct extent_item *)(l->data +
  69                                       l->items[path.slots[0]].offset);
  70         *refs = item->refs;
  71         release_path(root->extent_root, &path);
  72         return 0;
  73 }
  74
  75 int btrfs_inc_ref(struct ctree_root *root, struct tree_buffer *buf)
  76 {
  77         u64 blocknr;
  78         int i;
  79
  80         if (root == root->extent_root)
  81                 return 0;
  82         if (is_leaf(buf->node.header.flags))
  83                 return 0;
  84
  85         for (i = 0; i < buf->node.header.nritems; i++) {
  86                 blocknr = buf->node.blockptrs[i];
  87                 inc_block_ref(root, blocknr);
  88         }
  89         return 0;
  90 }
  91
  92 int btrfs_finish_extent_commit(struct ctree_root *root)
  93 {
  94         struct ctree_root *extent_root = root->extent_root;
  95         unsigned long gang[8];
  96         int ret;
  97         int i;
  98
  99         while(1) {
 100                 ret = radix_tree_gang_lookup(&extent_root->pinned_radix,
 101                                                  (void **)gang, 0,
 102                                                  ARRAY_SIZE(gang));
 103                 if (!ret)
 104                         break;
 105                 for (i = 0; i < ret; i++) {
 106                         radix_tree_delete(&extent_root->pinned_radix, gang[i]);
 107                 }
 108         }
 109         extent_root->last_insert.objectid = 0;
 110         extent_root->last_insert.offset = 0;
 111         return 0;
 112 }
 113
 114 static int finish_current_insert(struct ctree_root *extent_root)
 115 {
 116         struct key ins;
 117         struct extent_item extent_item;
 118         int i;
 119         int ret;
 120
 121         extent_item.refs = 1;
 122         extent_item.owner = extent_root->node->node.header.parentid;
 123         ins.offset = 1;
 124         ins.flags = 0;
 125
 126         for (i = 0; i < extent_root->current_insert.flags; i++) {
 127                 ins.objectid = extent_root->current_insert.objectid + i;
 128                 ret = insert_item(extent_root, &ins, &extent_item,
 129                                   sizeof(extent_item));
 130                 BUG_ON(ret);
 131         }
 132         extent_root->current_insert.offset = 0;
 133         return 0;
 134 }
 135
 136 /*
 137  * remove an extent from the root, returns 0 on success
 138  */
 139 int __free_extent(struct ctree_root *root, u64 blocknr, u64 num_blocks)
 140 {
 141         struct ctree_path path;
 142         struct key key;
 143         struct ctree_root *extent_root = root->extent_root;
 144         int ret;
 145         struct item *item;
 146         struct extent_item *ei;
 147         struct key ins;
 148
 149         key.objectid = blocknr;
 150         key.flags = 0;
 151         key.offset = num_blocks;
 152
 153         find_free_extent(root, 0, 0, (u64)-1, &ins);
 154         init_path(&path);
 155         ret = search_slot(extent_root, &key, &path, -1, 1);
 156         if (ret) {
 157                 printf("failed to find %Lu\n", key.objectid);
 158                 print_tree(extent_root, extent_root->node);
 159                 printf("failed to find %Lu\n", key.objectid);
 160                 BUG();
 161         }
 162         item = path.nodes[0]->leaf.items + path.slots[0];
 163         ei = (struct extent_item *)(path.nodes[0]->leaf.data + item->offset);
 164         BUG_ON(ei->refs == 0);
 165         ei->refs--;
 166         if (ei->refs == 0) {
 167                 if (root == extent_root) {
 168                         int err;
 169                         radix_tree_preload(GFP_KERNEL);
 170                         err = radix_tree_insert(&extent_root->pinned_radix,
 171                                           blocknr, (void *)blocknr);
 172                         BUG_ON(err);
 173                         radix_tree_preload_end();
 174                 }
 175                 ret = del_item(extent_root, &path);
 176                 if (root != extent_root &&
 177                     extent_root->last_insert.objectid < blocknr)
 178                         extent_root->last_insert.objectid = blocknr;
 179                 if (ret)
 180                         BUG();
 181         }
 182         release_path(extent_root, &path);
 183         finish_current_insert(extent_root);
 184         return ret;
 185 }
 186
 187 /*
 188  * find all the blocks marked as pending in the radix tree and remove
 189  * them from the extent map
 190  */
 191 static int del_pending_extents(struct ctree_root *extent_root)
 192 {
 193         int ret;
 194         struct tree_buffer *gang[4];
 195         int i;
 196
 197         while(1) {
 198                 ret = radix_tree_gang_lookup_tag(&extent_root->cache_radix,
 199                                                  (void **)gang, 0,
 200                                                  ARRAY_SIZE(gang),
 201                                                  CTREE_EXTENT_PENDING_DEL);
 202                 if (!ret)
 203                         break;
 204                 for (i = 0; i < ret; i++) {
 205                         ret = __free_extent(extent_root, gang[i]->blocknr, 1);
 206                         radix_tree_tag_clear(&extent_root->cache_radix,
 207                                                 gang[i]->blocknr,
 208                                                 CTREE_EXTENT_PENDING_DEL);
 209                         tree_block_release(extent_root, gang[i]);
 210                 }
 211         }
 212         return 0;
 213 }
 214
 215 static int run_pending(struct ctree_root *extent_root)
 216 {
 217         while(radix_tree_tagged(&extent_root->cache_radix,
 218                                 CTREE_EXTENT_PENDING_DEL))
 219                 del_pending_extents(extent_root);
 220         return 0;
 221 }
 222
 223
 224 /*
 225  * remove an extent from the root, returns 0 on success
 226  */
 227 int free_extent(struct ctree_root *root, u64 blocknr, u64 num_blocks)
 228 {
 229         struct key key;
 230         struct ctree_root *extent_root = root->extent_root;
 231         struct tree_buffer *t;
 232         int pending_ret;
 233         int ret;
 234
 235         if (root == extent_root) {
 236                 t = find_tree_block(root, blocknr);
 237                 radix_tree_tag_set(&root->cache_radix, blocknr,
 238                                    CTREE_EXTENT_PENDING_DEL);
 239                 return 0;
 240         }
 241         key.objectid = blocknr;
 242         key.flags = 0;
 243         key.offset = num_blocks;
 244         ret = __free_extent(root, blocknr, num_blocks);
 245         pending_ret = run_pending(root->extent_root);
 246         return ret ? ret : pending_ret;
 247 }
 248
 249 /*
 250  * walks the btree of allocated extents and find a hole of a given size.
 251  * The key ins is changed to record the hole:
 252  * ins->objectid == block start
 253  * ins->flags = 0
 254  * ins->offset == number of blocks
 255  * Any available blocks before search_start are skipped.
 256  */
 257 static int find_free_extent(struct ctree_root *orig_root, u64 num_blocks,
 258                             u64 search_start, u64 search_end, struct key *ins)
 259 {
 260         struct ctree_path path;
 261         struct key *key;
 262         int ret;
 263         u64 hole_size = 0;
 264         int slot = 0;
 265         u64 last_block;
 266         u64 test_block;
 267         int start_found;
 268         struct leaf *l;
 269         struct ctree_root * root = orig_root->extent_root;
 270         int total_needed = num_blocks;
 271
 272         total_needed += (node_level(root->node->node.header.flags) + 1) * 3;
 273         if (root->last_insert.objectid > search_start)
 274                 search_start = root->last_insert.objectid;
 275 check_failed:
 276         init_path(&path);
 277         ins->objectid = search_start;
 278         ins->offset = 0;
 279         ins->flags = 0;
 280         start_found = 0;
 281         ret = search_slot(root, ins, &path, 0, 0);
 282         if (ret < 0)
 283                 goto error;
 284
 285         if (path.slots[0] > 0)
 286                 path.slots[0]--;
 287
 288         while (1) {
 289                 l = &path.nodes[0]->leaf;
 290                 slot = path.slots[0];
 291                 if (slot >= l->header.nritems) {
 292                         ret = next_leaf(root, &path);
 293                         if (ret == 0)
 294                                 continue;
 295                         if (ret < 0)
 296                                 goto error;
 297                         if (!start_found) {
 298                                 ins->objectid = search_start;
 299                                 ins->offset = (u64)-1;
 300                                 start_found = 1;
 301                                 goto check_pending;
 302                         }
 303                         ins->objectid = last_block > search_start ?
 304                                         last_block : search_start;
 305                         ins->offset = (u64)-1;
 306                         goto check_pending;
 307                 }
 308                 key = &l->items[slot].key;
 309                 if (key->objectid >= search_start) {
 310                         if (start_found) {
 311                                 if (last_block < search_start)
 312                                         last_block = search_start;
 313                                 hole_size = key->objectid - last_block;
 314                                 if (hole_size > total_needed) {
 315                                         ins->objectid = last_block;
 316                                         ins->offset = hole_size;
 317                                         goto check_pending;
 318                                 }
 319                         }
 320                 }
 321                 start_found = 1;
 322                 last_block = key->objectid + key->offset;
 323                 path.slots[0]++;
 324         }
 325         // FIXME -ENOSPC
 326 check_pending:
 327         /* we have to make sure we didn't find an extent that has already
 328          * been allocated by the map tree or the original allocation
 329          */
 330         release_path(root, &path);
 331         BUG_ON(ins->objectid < search_start);
 332         for (test_block = ins->objectid;
 333              test_block < ins->objectid + total_needed; test_block++) {
 334                 if (radix_tree_lookup(&root->pinned_radix, test_block)) {
 335                         search_start = test_block + 1;
 336                         goto check_failed;
 337                 }
 338         }
 339         BUG_ON(root->current_insert.offset);
 340         root->current_insert.offset = total_needed - num_blocks;
 341         root->current_insert.objectid = ins->objectid + num_blocks;
 342         root->current_insert.flags = 0;
 343         root->last_insert.objectid = ins->objectid;
 344         ins->offset = num_blocks;
 345         return 0;
 346 error:
 347         release_path(root, &path);
 348         return ret;
 349 }
 350
 351 /*
 352  * finds a free extent and does all the dirty work required for allocation
 353  * returns the key for the extent through ins, and a tree buffer for
 354  * the first block of the extent through buf.
 355  *
 356  * returns 0 if everything worked, non-zero otherwise.
 357  */
 358 int alloc_extent(struct ctree_root *root, u64 num_blocks, u64 search_start,
 359                          u64 search_end, u64 owner, struct key *ins)
 360 {
 361         int ret;
 362         int pending_ret;
 363         struct ctree_root *extent_root = root->extent_root;
 364         struct extent_item extent_item;
 365
 366         extent_item.refs = 1;
 367         extent_item.owner = owner;
 368
 369         if (root == extent_root) {
 370                 BUG_ON(extent_root->current_insert.offset == 0);
 371                 BUG_ON(num_blocks != 1);
 372                 BUG_ON(extent_root->current_insert.flags ==
 373                        extent_root->current_insert.offset);
 374                 ins->offset = 1;
 375                 ins->objectid = extent_root->current_insert.objectid +
 376                                 extent_root->current_insert.flags++;
 377                 return 0;
 378         }
 379         ret = find_free_extent(root, num_blocks, search_start,
 380                                search_end, ins);
 381         if (ret)
 382                 return ret;
 383
 384         ret = insert_item(extent_root, ins, &extent_item,
 385                           sizeof(extent_item));
 386
 387         finish_current_insert(extent_root);
 388         pending_ret = run_pending(extent_root);
 389         if (ret)
 390                 return ret;
 391         if (pending_ret)
 392                 return pending_ret;
 393         return 0;
 394 }
 395
 396 /*
 397  * helper function to allocate a block for a given tree
 398  * returns the tree buffer or NULL.
 399  */
 400 struct tree_buffer *alloc_free_block(struct ctree_root *root)
 401 {
 402         struct key ins;
 403         int ret;
 404         struct tree_buffer *buf;
 405
 406         ret = alloc_extent(root, 1, 0, (unsigned long)-1,
 407                            root->node->node.header.parentid,
 408                            &ins);
 409         if (ret) {
 410                 BUG();
 411                 return NULL;
 412         }
 413         buf = find_tree_block(root, ins.objectid);
 414         dirty_tree_block(root, buf);
 415         return buf;
 416 }
 417
 418 int walk_down_tree(struct ctree_root *root, struct ctree_path *path, int *level)
 419 {
 420         struct tree_buffer *next;
 421         struct tree_buffer *cur;
 422         u64 blocknr;
 423         int ret;
 424         u32 refs;
 425
 426         ret = lookup_block_ref(root, path->nodes[*level]->blocknr, &refs);
 427         BUG_ON(ret);
 428         if (refs > 1)
 429                 goto out;
 430         while(*level > 0) {
 431                 cur = path->nodes[*level];
 432                 if (path->slots[*level] >= cur->node.header.nritems)
 433                         break;
 434                 blocknr = cur->node.blockptrs[path->slots[*level]];
 435                 ret = lookup_block_ref(root, blocknr, &refs);
 436                 if (refs != 1 || *level == 1) {
 437                         path->slots[*level]++;
 438                         ret = free_extent(root, blocknr, 1);
 439                         BUG_ON(ret);
 440                         continue;
 441                 }
 442                 BUG_ON(ret);
 443                 next = read_tree_block(root, blocknr);
 444                 if (path->nodes[*level-1]) {
 445                         tree_block_release(root, path->nodes[*level-1]);
 446                 }
 447                 path->nodes[*level-1] = next;
 448                 *level = node_level(next->node.header.flags);
 449                 path->slots[*level] = 0;
 450         }
 451 out:
 452         ret = free_extent(root, path->nodes[*level]->blocknr, 1);
 453         path->nodes[*level] = NULL;
 454         *level += 1;
 455         BUG_ON(ret);
 456         return 0;
 457 }
 458
 459 int walk_up_tree(struct ctree_root *root, struct ctree_path *path, int *level)
 460 {
 461         int i;
 462         int slot;
 463         int ret;
 464         for(i = *level; i < MAX_LEVEL - 1 && path->nodes[i]; i++) {
 465                 slot = path->slots[i];
 466                 if (slot < path->nodes[i]->node.header.nritems - 1) {
 467                         path->slots[i]++;
 468                         *level = i;
 469                         return 0;
 470                 } else {
 471                         ret = free_extent(root,
 472                                           path->nodes[*level]->blocknr, 1);
 473                         *level = i + 1;
 474                         BUG_ON(ret);
 475                 }
 476         }
 477         return 1;
 478 }
 479
 480 int btrfs_drop_snapshot(struct ctree_root *root, struct tree_buffer *snap)
 481 {
 482         int ret;
 483         int level;
 484         struct ctree_path path;
 485         int i;
 486         int orig_level;
 487
 488         init_path(&path);
 489
 490         level = node_level(snap->node.header.flags);
 491         orig_level = level;
 492         path.nodes[level] = snap;
 493         path.slots[level] = 0;
 494         while(1) {
 495                 ret = walk_down_tree(root, &path, &level);
 496                 if (ret > 0)
 497                         break;
 498                 ret = walk_up_tree(root, &path, &level);
 499                 if (ret > 0)
 500                         break;
 501         }
 502         for (i = 0; i < orig_level; i++) {
 503                 if (path.nodes[i])
 504                         tree_block_release(root, path.nodes[i]);
 505         }
 506
 507         return 0;
 508 }
 509
 510
 511 #if 0
 512 int btrfs_drop_snapshot(struct ctree_root *root, struct tree_buffer *snap)
 513 {
 514         int ret;
 515         int level;
 516         int refs;
 517         u64 blocknr = snap->blocknr;
 518
 519         level = node_level(snap->node.header.flags);
 520         ret = lookup_block_ref(root, snap->blocknr, &refs);
 521         BUG_ON(ret);
 522         if (refs == 1 && level != 0) {
 523                 struct node *n = &snap->node;
 524                 struct tree_buffer *b;
 525                 int i;
 526                 for (i = 0; i < n->header.nritems; i++) {
 527                         b = read_tree_block(root, n->blockptrs[i]);
 528                         /* FIXME, don't recurse here */
 529                         ret = btrfs_drop_snapshot(root, b);
 530                         BUG_ON(ret);
 531                         tree_block_release(root, b);
 532                 }
 533         }
 534         ret = free_extent(root, blocknr, 1);
 535         BUG_ON(ret);
 536         return 0;
 537 }
 538 #endif