fs/btrfs/extent-tree.c

   1 #include <linux/module.h>
   2 #include <linux/radix-tree.h>
   3 #include "ctree.h"
   4 #include "disk-io.h"
   5 #include "print-tree.h"
   6 #include "transaction.h"
   7
   8 static int find_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
   9                             *orig_root, u64 num_blocks, u64 search_start, u64
  10                             search_end, struct btrfs_key *ins);
  11 static int finish_current_insert(struct btrfs_trans_handle *trans, struct
  12                                  btrfs_root *extent_root);
  13 static int del_pending_extents(struct btrfs_trans_handle *trans, struct
  14                                btrfs_root *extent_root);
  15 /*
  16  * pending extents are blocks that we're trying to allocate in the extent
  17  * map while trying to grow the map because of other allocations.  To avoid
  18  * recursing, they are tagged in the radix tree and cleaned up after
  19  * other allocations are done.  The pending tag is also used in the same
  20  * manner for deletes.
  21  */
  22 #define CTREE_EXTENT_PENDING_DEL 0
  23 #define CTREE_EXTENT_PINNED 1
  24
  25 static int inc_block_ref(struct btrfs_trans_handle *trans, struct btrfs_root
  26                          *root, u64 blocknr)
  27 {
  28         struct btrfs_path path;
  29         int ret;
  30         struct btrfs_key key;
  31         struct btrfs_leaf *l;
  32         struct btrfs_extent_item *item;
  33         struct btrfs_key ins;
  34         u32 refs;
  35
  36         find_free_extent(trans, root->fs_info->extent_root, 0, 0, (u64)-1,
  37                          &ins);
  38         btrfs_init_path(&path);
  39         key.objectid = blocknr;
  40         key.flags = 0;
  41         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
  42         key.offset = 1;
  43         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key, &path,
  44                                 0, 1);
  45         if (ret != 0)
  46                 BUG();
  47         BUG_ON(ret != 0);
  48         l = btrfs_buffer_leaf(path.nodes[0]);
  49         item = btrfs_item_ptr(l, path.slots[0], struct btrfs_extent_item);
  50         refs = btrfs_extent_refs(item);
  51         btrfs_set_extent_refs(item, refs + 1);
  52         mark_buffer_dirty(path.nodes[0]);
  53
  54         btrfs_release_path(root->fs_info->extent_root, &path);
  55         finish_current_insert(trans, root->fs_info->extent_root);
  56         del_pending_extents(trans, root->fs_info->extent_root);
  57         return 0;
  58 }
  59
  60 static int lookup_block_ref(struct btrfs_trans_handle *trans, struct btrfs_root
  61                             *root, u64 blocknr, u32 *refs)
  62 {
  63         struct btrfs_path path;
  64         int ret;
  65         struct btrfs_key key;
  66         struct btrfs_leaf *l;
  67         struct btrfs_extent_item *item;
  68         btrfs_init_path(&path);
  69         key.objectid = blocknr;
  70         key.offset = 1;
  71         key.flags = 0;
  72         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
  73         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key, &path,
  74                                 0, 0);
  75         if (ret != 0)
  76                 BUG();
  77         l = btrfs_buffer_leaf(path.nodes[0]);
  78         item = btrfs_item_ptr(l, path.slots[0], struct btrfs_extent_item);
  79         *refs = btrfs_extent_refs(item);
  80         btrfs_release_path(root->fs_info->extent_root, &path);
  81         return 0;
  82 }
  83
  84 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  85                   struct buffer_head *buf)
  86 {
  87         u64 blocknr;
  88         struct btrfs_node *buf_node;
  89         int i;
  90
  91         if (!root->ref_cows)
  92                 return 0;
  93         buf_node = btrfs_buffer_node(buf);
  94         if (btrfs_is_leaf(buf_node))
  95                 return 0;
  96
  97         for (i = 0; i < btrfs_header_nritems(&buf_node->header); i++) {
  98                 blocknr = btrfs_node_blockptr(buf_node, i);
  99                 inc_block_ref(trans, root, blocknr);
 100         }
 101         return 0;
 102 }
 103
 104 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans, struct
 105                                btrfs_root *root)
 106 {
 107         struct buffer_head *gang[8];
 108         u64 first = 0;
 109         int ret;
 110         int i;
 111
 112         while(1) {
 113                 ret = radix_tree_gang_lookup_tag(&root->fs_info->pinned_radix,
 114                                              (void **)gang, 0,
 115                                              ARRAY_SIZE(gang),
 116                                              CTREE_EXTENT_PINNED);
 117                 if (!ret)
 118                         break;
 119                 if (!first)
 120                         first = gang[0]->b_blocknr;
 121                 for (i = 0; i < ret; i++) {
 122                         radix_tree_delete(&root->fs_info->pinned_radix,
 123                                           gang[i]->b_blocknr);
 124                         brelse(gang[i]);
 125                 }
 126         }
 127         if (root->fs_info->last_insert.objectid > first)
 128                 root->fs_info->last_insert.objectid = first;
 129         root->fs_info->last_insert.offset = 0;
 130         return 0;
 131 }
 132
 133 static int finish_current_insert(struct btrfs_trans_handle *trans, struct
 134                                  btrfs_root *extent_root)
 135 {
 136         struct btrfs_key ins;
 137         struct btrfs_extent_item extent_item;
 138         int i;
 139         int ret;
 140         u64 super_blocks_used;
 141         struct btrfs_fs_info *info = extent_root->fs_info;
 142
 143         btrfs_set_extent_refs(&extent_item, 1);
 144         btrfs_set_extent_owner(&extent_item,
 145                 btrfs_header_parentid(btrfs_buffer_header(extent_root->node)));
 146         ins.offset = 1;
 147         ins.flags = 0;
 148         btrfs_set_key_type(&ins, BTRFS_EXTENT_ITEM_KEY);
 149
 150         for (i = 0; i < extent_root->fs_info->current_insert.flags; i++) {
 151                 ins.objectid = extent_root->fs_info->current_insert.objectid +
 152                                 i;
 153                 super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 154                 btrfs_set_super_blocks_used(info->disk_super,
 155                                             super_blocks_used + 1);
 156                 ret = btrfs_insert_item(trans, extent_root, &ins, &extent_item,
 157                                         sizeof(extent_item));
 158                 BUG_ON(ret);
 159         }
 160         extent_root->fs_info->current_insert.offset = 0;
 161         return 0;
 162 }
 163
 164 static int pin_down_block(struct btrfs_root *root, u64 blocknr, int tag)
 165 {
 166         int err;
 167         struct buffer_head *bh = sb_getblk(root->fs_info->sb, blocknr);
 168         BUG_ON(!bh);
 169         err = radix_tree_insert(&root->fs_info->pinned_radix,
 170                                 blocknr, bh);
 171         BUG_ON(err);
 172         if (err)
 173                 return err;
 174         radix_tree_tag_set(&root->fs_info->pinned_radix, blocknr,
 175                            tag);
 176         return 0;
 177 }
 178
 179 /*
 180  * remove an extent from the root, returns 0 on success
 181  */
 182 static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 183                          *root, u64 blocknr, u64 num_blocks)
 184 {
 185         struct btrfs_path path;
 186         struct btrfs_key key;
 187         struct btrfs_fs_info *info = root->fs_info;
 188         struct btrfs_root *extent_root = info->extent_root;
 189         int ret;
 190         struct btrfs_extent_item *ei;
 191         struct btrfs_key ins;
 192         u32 refs;
 193
 194         key.objectid = blocknr;
 195         key.flags = 0;
 196         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 197         key.offset = num_blocks;
 198
 199         find_free_extent(trans, root, 0, 0, (u64)-1, &ins);
 200         btrfs_init_path(&path);
 201         ret = btrfs_search_slot(trans, extent_root, &key, &path, -1, 1);
 202         if (ret) {
 203                 printk("failed to find %Lu\n", key.objectid);
 204                 btrfs_print_tree(extent_root, extent_root->node);
 205                 printk("failed to find %Lu\n", key.objectid);
 206                 BUG();
 207         }
 208         ei = btrfs_item_ptr(btrfs_buffer_leaf(path.nodes[0]), path.slots[0],
 209                             struct btrfs_extent_item);
 210         BUG_ON(ei->refs == 0);
 211         refs = btrfs_extent_refs(ei) - 1;
 212         btrfs_set_extent_refs(ei, refs);
 213         if (refs == 0) {
 214                 u64 super_blocks_used;
 215                 super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 216                 btrfs_set_super_blocks_used(info->disk_super,
 217                                             super_blocks_used - num_blocks);
 218                 ret = btrfs_del_item(trans, extent_root, &path);
 219                 if (extent_root->fs_info->last_insert.objectid >
 220                     blocknr)
 221                         extent_root->fs_info->last_insert.objectid = blocknr;
 222                 if (ret)
 223                         BUG();
 224         }
 225         mark_buffer_dirty(path.nodes[0]);
 226         btrfs_release_path(extent_root, &path);
 227         finish_current_insert(trans, extent_root);
 228         return ret;
 229 }
 230
 231 /*
 232  * find all the blocks marked as pending in the radix tree and remove
 233  * them from the extent map
 234  */
 235 static int del_pending_extents(struct btrfs_trans_handle *trans, struct
 236                                btrfs_root *extent_root)
 237 {
 238         int ret;
 239         int wret;
 240         int err = 0;
 241         struct buffer_head *gang[4];
 242         int i;
 243         struct radix_tree_root *radix = &extent_root->fs_info->pinned_radix;
 244
 245         while(1) {
 246                 ret = radix_tree_gang_lookup_tag(
 247                                         &extent_root->fs_info->pinned_radix,
 248                                         (void **)gang, 0,
 249                                         ARRAY_SIZE(gang),
 250                                         CTREE_EXTENT_PENDING_DEL);
 251                 if (!ret)
 252                         break;
 253                 for (i = 0; i < ret; i++) {
 254                         radix_tree_tag_set(radix, gang[i]->b_blocknr,
 255                                            CTREE_EXTENT_PINNED);
 256                         radix_tree_tag_clear(radix, gang[i]->b_blocknr,
 257                                              CTREE_EXTENT_PENDING_DEL);
 258                         wret = __free_extent(trans, extent_root,
 259                                              gang[i]->b_blocknr, 1);
 260                         if (wret)
 261                                 err = wret;
 262                 }
 263         }
 264         return err;
 265 }
 266
 267 /*
 268  * remove an extent from the root, returns 0 on success
 269  */
 270 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 271                       *root, u64 blocknr, u64 num_blocks, int pin)
 272 {
 273         struct btrfs_root *extent_root = root->fs_info->extent_root;
 274         struct buffer_head *t;
 275         int pending_ret;
 276         int ret;
 277
 278         if (root == extent_root) {
 279                 t = find_tree_block(root, blocknr);
 280                 pin_down_block(root, blocknr, CTREE_EXTENT_PENDING_DEL);
 281                 return 0;
 282         }
 283         if (pin) {
 284                 ret = pin_down_block(root, blocknr, CTREE_EXTENT_PINNED);
 285                 BUG_ON(ret);
 286         }
 287         ret = __free_extent(trans, root, blocknr, num_blocks);
 288         pending_ret = del_pending_extents(trans, root->fs_info->extent_root);
 289         return ret ? ret : pending_ret;
 290 }
 291
 292 /*
 293  * walks the btree of allocated extents and find a hole of a given size.
 294  * The key ins is changed to record the hole:
 295  * ins->objectid == block start
 296  * ins->flags = BTRFS_EXTENT_ITEM_KEY
 297  * ins->offset == number of blocks
 298  * Any available blocks before search_start are skipped.
 299  */
 300 static int find_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 301                             *orig_root, u64 num_blocks, u64 search_start, u64
 302                             search_end, struct btrfs_key *ins)
 303 {
 304         struct btrfs_path path;
 305         struct btrfs_key key;
 306         int ret;
 307         u64 hole_size = 0;
 308         int slot = 0;
 309         u64 last_block = 0;
 310         u64 test_block;
 311         int start_found;
 312         struct btrfs_leaf *l;
 313         struct btrfs_root * root = orig_root->fs_info->extent_root;
 314         int total_needed = num_blocks;
 315         int level;
 316
 317         level = btrfs_header_level(btrfs_buffer_header(root->node));
 318         total_needed += (level + 1) * 3;
 319         if (root->fs_info->last_insert.objectid > search_start)
 320                 search_start = root->fs_info->last_insert.objectid;
 321
 322         ins->flags = 0;
 323         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
 324
 325 check_failed:
 326         btrfs_init_path(&path);
 327         ins->objectid = search_start;
 328         ins->offset = 0;
 329         start_found = 0;
 330         ret = btrfs_search_slot(trans, root, ins, &path, 0, 0);
 331         if (ret < 0)
 332                 goto error;
 333
 334         if (path.slots[0] > 0)
 335                 path.slots[0]--;
 336
 337         while (1) {
 338                 l = btrfs_buffer_leaf(path.nodes[0]);
 339                 slot = path.slots[0];
 340                 if (slot >= btrfs_header_nritems(&l->header)) {
 341                         ret = btrfs_next_leaf(root, &path);
 342                         if (ret == 0)
 343                                 continue;
 344                         if (ret < 0)
 345                                 goto error;
 346                         if (!start_found) {
 347                                 ins->objectid = search_start;
 348                                 ins->offset = (u64)-1;
 349                                 start_found = 1;
 350                                 goto check_pending;
 351                         }
 352                         ins->objectid = last_block > search_start ?
 353                                         last_block : search_start;
 354                         ins->offset = (u64)-1;
 355                         goto check_pending;
 356                 }
 357                 btrfs_disk_key_to_cpu(&key, &l->items[slot].key);
 358                 if (key.objectid >= search_start) {
 359                         if (start_found) {
 360                                 if (last_block < search_start)
 361                                         last_block = search_start;
 362                                 hole_size = key.objectid - last_block;
 363                                 if (hole_size > total_needed) {
 364                                         ins->objectid = last_block;
 365                                         ins->offset = hole_size;
 366                                         goto check_pending;
 367                                 }
 368                         }
 369                 }
 370                 start_found = 1;
 371                 last_block = key.objectid + key.offset;
 372                 path.slots[0]++;
 373         }
 374         // FIXME -ENOSPC
 375 check_pending:
 376         /* we have to make sure we didn't find an extent that has already
 377          * been allocated by the map tree or the original allocation
 378          */
 379         btrfs_release_path(root, &path);
 380         BUG_ON(ins->objectid < search_start);
 381         for (test_block = ins->objectid;
 382              test_block < ins->objectid + total_needed; test_block++) {
 383                 if (radix_tree_lookup(&root->fs_info->pinned_radix,
 384                                       test_block)) {
 385                         search_start = test_block + 1;
 386                         goto check_failed;
 387                 }
 388         }
 389         BUG_ON(root->fs_info->current_insert.offset);
 390         root->fs_info->current_insert.offset = total_needed - num_blocks;
 391         root->fs_info->current_insert.objectid = ins->objectid + num_blocks;
 392         root->fs_info->current_insert.flags = 0;
 393         root->fs_info->last_insert.objectid = ins->objectid;
 394         ins->offset = num_blocks;
 395         return 0;
 396 error:
 397         btrfs_release_path(root, &path);
 398         return ret;
 399 }
 400
 401 /*
 402  * finds a free extent and does all the dirty work required for allocation
 403  * returns the key for the extent through ins, and a tree buffer for
 404  * the first block of the extent through buf.
 405  *
 406  * returns 0 if everything worked, non-zero otherwise.
 407  */
 408 static int alloc_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 409                         *root, u64 num_blocks, u64 search_start, u64
 410                         search_end, u64 owner, struct btrfs_key *ins)
 411 {
 412         int ret;
 413         int pending_ret;
 414         u64 super_blocks_used;
 415         struct btrfs_fs_info *info = root->fs_info;
 416         struct btrfs_root *extent_root = info->extent_root;
 417         struct btrfs_extent_item extent_item;
 418
 419         btrfs_set_extent_refs(&extent_item, 1);
 420         btrfs_set_extent_owner(&extent_item, owner);
 421
 422         if (root == extent_root) {
 423                 BUG_ON(extent_root->fs_info->current_insert.offset == 0);
 424                 BUG_ON(num_blocks != 1);
 425                 BUG_ON(extent_root->fs_info->current_insert.flags ==
 426                        extent_root->fs_info->current_insert.offset);
 427                 ins->offset = 1;
 428                 ins->objectid = extent_root->fs_info->current_insert.objectid +
 429                                 extent_root->fs_info->current_insert.flags++;
 430                 return 0;
 431         }
 432         ret = find_free_extent(trans, root, num_blocks, search_start,
 433                                search_end, ins);
 434         if (ret)
 435                 return ret;
 436
 437         super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 438         btrfs_set_super_blocks_used(info->disk_super, super_blocks_used +
 439                                     num_blocks);
 440         ret = btrfs_insert_item(trans, extent_root, ins, &extent_item,
 441                                 sizeof(extent_item));
 442
 443         finish_current_insert(trans, extent_root);
 444         pending_ret = del_pending_extents(trans, extent_root);
 445         if (ret)
 446                 return ret;
 447         if (pending_ret)
 448                 return pending_ret;
 449         return 0;
 450 }
 451
 452 /*
 453  * helper function to allocate a block for a given tree
 454  * returns the tree buffer or NULL.
 455  */
 456 struct buffer_head *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
 457                                             struct btrfs_root *root)
 458 {
 459         struct btrfs_key ins;
 460         int ret;
 461         struct buffer_head *buf;
 462
 463         ret = alloc_extent(trans, root, 1, 0, (unsigned long)-1,
 464                 btrfs_header_parentid(btrfs_buffer_header(root->node)), &ins);
 465         if (ret) {
 466                 BUG();
 467                 return NULL;
 468         }
 469         buf = find_tree_block(root, ins.objectid);
 470         set_buffer_uptodate(buf);
 471         return buf;
 472 }
 473
 474 /*
 475  * helper function for drop_snapshot, this walks down the tree dropping ref
 476  * counts as it goes.
 477  */
 478 static int walk_down_tree(struct btrfs_trans_handle *trans, struct btrfs_root
 479                           *root, struct btrfs_path *path, int *level)
 480 {
 481         struct buffer_head *next;
 482         struct buffer_head *cur;
 483         u64 blocknr;
 484         int ret;
 485         u32 refs;
 486
 487         ret = lookup_block_ref(trans, root, path->nodes[*level]->b_blocknr,
 488                                &refs);
 489         BUG_ON(ret);
 490         if (refs > 1)
 491                 goto out;
 492         /*
 493          * walk down to the last node level and free all the leaves
 494          */
 495         while(*level > 0) {
 496                 cur = path->nodes[*level];
 497                 if (path->slots[*level] >=
 498                     btrfs_header_nritems(btrfs_buffer_header(cur)))
 499                         break;
 500                 blocknr = btrfs_node_blockptr(btrfs_buffer_node(cur),
 501                                               path->slots[*level]);
 502                 ret = lookup_block_ref(trans, root, blocknr, &refs);
 503                 if (refs != 1 || *level == 1) {
 504                         path->slots[*level]++;
 505                         ret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 506                         BUG_ON(ret);
 507                         continue;
 508                 }
 509                 BUG_ON(ret);
 510                 next = read_tree_block(root, blocknr);
 511                 if (path->nodes[*level-1])
 512                         btrfs_block_release(root, path->nodes[*level-1]);
 513                 path->nodes[*level-1] = next;
 514                 *level = btrfs_header_level(btrfs_buffer_header(next));
 515                 path->slots[*level] = 0;
 516         }
 517 out:
 518         ret = btrfs_free_extent(trans, root, path->nodes[*level]->b_blocknr,
 519                                 1, 1);
 520         btrfs_block_release(root, path->nodes[*level]);
 521         path->nodes[*level] = NULL;
 522         *level += 1;
 523         BUG_ON(ret);
 524         return 0;
 525 }
 526
 527 /*
 528  * helper for dropping snapshots.  This walks back up the tree in the path
 529  * to find the first node higher up where we haven't yet gone through
 530  * all the slots
 531  */
 532 static int walk_up_tree(struct btrfs_trans_handle *trans, struct btrfs_root
 533                         *root, struct btrfs_path *path, int *level)
 534 {
 535         int i;
 536         int slot;
 537         int ret;
 538         for(i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
 539                 slot = path->slots[i];
 540                 if (slot < btrfs_header_nritems(
 541                     btrfs_buffer_header(path->nodes[i])) - 1) {
 542                         path->slots[i]++;
 543                         *level = i;
 544                         return 0;
 545                 } else {
 546                         ret = btrfs_free_extent(trans, root,
 547                                                 path->nodes[*level]->b_blocknr,
 548                                                 1, 1);
 549                         btrfs_block_release(root, path->nodes[*level]);
 550                         path->nodes[*level] = NULL;
 551                         *level = i + 1;
 552                         BUG_ON(ret);
 553                 }
 554         }
 555         return 1;
 556 }
 557
 558 /*
 559  * drop the reference count on the tree rooted at 'snap'.  This traverses
 560  * the tree freeing any blocks that have a ref count of zero after being
 561  * decremented.
 562  */
 563 int btrfs_drop_snapshot(struct btrfs_trans_handle *trans, struct btrfs_root
 564                         *root, struct buffer_head *snap)
 565 {
 566         int ret = 0;
 567         int wret;
 568         int level;
 569         struct btrfs_path path;
 570         int i;
 571         int orig_level;
 572
 573         btrfs_init_path(&path);
 574
 575         level = btrfs_header_level(btrfs_buffer_header(snap));
 576         orig_level = level;
 577         path.nodes[level] = snap;
 578         path.slots[level] = 0;
 579         while(1) {
 580                 wret = walk_down_tree(trans, root, &path, &level);
 581                 if (wret > 0)
 582                         break;
 583                 if (wret < 0)
 584                         ret = wret;
 585
 586                 wret = walk_up_tree(trans, root, &path, &level);
 587                 if (wret > 0)
 588                         break;
 589                 if (wret < 0)
 590                         ret = wret;
 591         }
 592         for (i = 0; i <= orig_level; i++) {
 593                 if (path.nodes[i]) {
 594                         btrfs_block_release(root, path.nodes[i]);
 595                 }
 596         }
 597         return ret;
 598 }