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         if (err && err != -EEXIST) {
 172                 BUG();
 173                 return err;
 174         }
 175         radix_tree_tag_set(&root->fs_info->pinned_radix, blocknr,
 176                            tag);
 177         return 0;
 178 }
 179
 180 /*
 181  * remove an extent from the root, returns 0 on success
 182  */
 183 static int __free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 184                          *root, u64 blocknr, u64 num_blocks)
 185 {
 186         struct btrfs_path path;
 187         struct btrfs_key key;
 188         struct btrfs_fs_info *info = root->fs_info;
 189         struct btrfs_root *extent_root = info->extent_root;
 190         int ret;
 191         struct btrfs_extent_item *ei;
 192         struct btrfs_key ins;
 193         u32 refs;
 194
 195         key.objectid = blocknr;
 196         key.flags = 0;
 197         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 198         key.offset = num_blocks;
 199
 200         find_free_extent(trans, root, 0, 0, (u64)-1, &ins);
 201         btrfs_init_path(&path);
 202         ret = btrfs_search_slot(trans, extent_root, &key, &path, -1, 1);
 203         if (ret) {
 204                 printk("failed to find %Lu\n", key.objectid);
 205                 btrfs_print_tree(extent_root, extent_root->node);
 206                 printk("failed to find %Lu\n", key.objectid);
 207                 BUG();
 208         }
 209         ei = btrfs_item_ptr(btrfs_buffer_leaf(path.nodes[0]), path.slots[0],
 210                             struct btrfs_extent_item);
 211         BUG_ON(ei->refs == 0);
 212         refs = btrfs_extent_refs(ei) - 1;
 213         btrfs_set_extent_refs(ei, refs);
 214         if (refs == 0) {
 215                 u64 super_blocks_used;
 216                 super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 217                 btrfs_set_super_blocks_used(info->disk_super,
 218                                             super_blocks_used - num_blocks);
 219                 ret = btrfs_del_item(trans, extent_root, &path);
 220                 if (extent_root->fs_info->last_insert.objectid >
 221                     blocknr)
 222                         extent_root->fs_info->last_insert.objectid = blocknr;
 223                 if (ret)
 224                         BUG();
 225         }
 226         mark_buffer_dirty(path.nodes[0]);
 227         btrfs_release_path(extent_root, &path);
 228         finish_current_insert(trans, extent_root);
 229         return ret;
 230 }
 231
 232 /*
 233  * find all the blocks marked as pending in the radix tree and remove
 234  * them from the extent map
 235  */
 236 static int del_pending_extents(struct btrfs_trans_handle *trans, struct
 237                                btrfs_root *extent_root)
 238 {
 239         int ret;
 240         int wret;
 241         int err = 0;
 242         struct buffer_head *gang[4];
 243         int i;
 244         struct radix_tree_root *radix = &extent_root->fs_info->pinned_radix;
 245
 246         while(1) {
 247                 ret = radix_tree_gang_lookup_tag(
 248                                         &extent_root->fs_info->pinned_radix,
 249                                         (void **)gang, 0,
 250                                         ARRAY_SIZE(gang),
 251                                         CTREE_EXTENT_PENDING_DEL);
 252                 if (!ret)
 253                         break;
 254                 for (i = 0; i < ret; i++) {
 255                         radix_tree_tag_set(radix, gang[i]->b_blocknr,
 256                                            CTREE_EXTENT_PINNED);
 257                         radix_tree_tag_clear(radix, gang[i]->b_blocknr,
 258                                              CTREE_EXTENT_PENDING_DEL);
 259                         wret = __free_extent(trans, extent_root,
 260                                              gang[i]->b_blocknr, 1);
 261                         if (wret)
 262                                 err = wret;
 263                 }
 264         }
 265         return err;
 266 }
 267
 268 /*
 269  * remove an extent from the root, returns 0 on success
 270  */
 271 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 272                       *root, u64 blocknr, u64 num_blocks, int pin)
 273 {
 274         struct btrfs_root *extent_root = root->fs_info->extent_root;
 275         struct buffer_head *t;
 276         int pending_ret;
 277         int ret;
 278
 279         if (root == extent_root) {
 280                 t = find_tree_block(root, blocknr);
 281                 pin_down_block(root, blocknr, CTREE_EXTENT_PENDING_DEL);
 282                 return 0;
 283         }
 284         if (pin) {
 285                 ret = pin_down_block(root, blocknr, CTREE_EXTENT_PINNED);
 286                 BUG_ON(ret);
 287         }
 288         ret = __free_extent(trans, root, blocknr, num_blocks);
 289         pending_ret = del_pending_extents(trans, root->fs_info->extent_root);
 290         return ret ? ret : pending_ret;
 291 }
 292
 293 /*
 294  * walks the btree of allocated extents and find a hole of a given size.
 295  * The key ins is changed to record the hole:
 296  * ins->objectid == block start
 297  * ins->flags = BTRFS_EXTENT_ITEM_KEY
 298  * ins->offset == number of blocks
 299  * Any available blocks before search_start are skipped.
 300  */
 301 static int find_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 302                             *orig_root, u64 num_blocks, u64 search_start, u64
 303                             search_end, struct btrfs_key *ins)
 304 {
 305         struct btrfs_path path;
 306         struct btrfs_key key;
 307         int ret;
 308         u64 hole_size = 0;
 309         int slot = 0;
 310         u64 last_block = 0;
 311         u64 test_block;
 312         int start_found;
 313         struct btrfs_leaf *l;
 314         struct btrfs_root * root = orig_root->fs_info->extent_root;
 315         int total_needed = num_blocks;
 316         int level;
 317
 318         level = btrfs_header_level(btrfs_buffer_header(root->node));
 319         total_needed += (level + 1) * 3;
 320         if (root->fs_info->last_insert.objectid > search_start)
 321                 search_start = root->fs_info->last_insert.objectid;
 322
 323         ins->flags = 0;
 324         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
 325
 326 check_failed:
 327         btrfs_init_path(&path);
 328         ins->objectid = search_start;
 329         ins->offset = 0;
 330         start_found = 0;
 331         ret = btrfs_search_slot(trans, root, ins, &path, 0, 0);
 332         if (ret < 0)
 333                 goto error;
 334
 335         if (path.slots[0] > 0)
 336                 path.slots[0]--;
 337
 338         while (1) {
 339                 l = btrfs_buffer_leaf(path.nodes[0]);
 340                 slot = path.slots[0];
 341                 if (slot >= btrfs_header_nritems(&l->header)) {
 342                         ret = btrfs_next_leaf(root, &path);
 343                         if (ret == 0)
 344                                 continue;
 345                         if (ret < 0)
 346                                 goto error;
 347                         if (!start_found) {
 348                                 ins->objectid = search_start;
 349                                 ins->offset = (u64)-1;
 350                                 start_found = 1;
 351                                 goto check_pending;
 352                         }
 353                         ins->objectid = last_block > search_start ?
 354                                         last_block : search_start;
 355                         ins->offset = (u64)-1;
 356                         goto check_pending;
 357                 }
 358                 btrfs_disk_key_to_cpu(&key, &l->items[slot].key);
 359                 if (key.objectid >= search_start) {
 360                         if (start_found) {
 361                                 if (last_block < search_start)
 362                                         last_block = search_start;
 363                                 hole_size = key.objectid - last_block;
 364                                 if (hole_size > total_needed) {
 365                                         ins->objectid = last_block;
 366                                         ins->offset = hole_size;
 367                                         goto check_pending;
 368                                 }
 369                         }
 370                 }
 371                 start_found = 1;
 372                 last_block = key.objectid + key.offset;
 373                 path.slots[0]++;
 374         }
 375         // FIXME -ENOSPC
 376 check_pending:
 377         /* we have to make sure we didn't find an extent that has already
 378          * been allocated by the map tree or the original allocation
 379          */
 380         btrfs_release_path(root, &path);
 381         BUG_ON(ins->objectid < search_start);
 382         for (test_block = ins->objectid;
 383              test_block < ins->objectid + total_needed; test_block++) {
 384                 if (radix_tree_lookup(&root->fs_info->pinned_radix,
 385                                       test_block)) {
 386                         search_start = test_block + 1;
 387                         goto check_failed;
 388                 }
 389         }
 390         BUG_ON(root->fs_info->current_insert.offset);
 391         root->fs_info->current_insert.offset = total_needed - num_blocks;
 392         root->fs_info->current_insert.objectid = ins->objectid + num_blocks;
 393         root->fs_info->current_insert.flags = 0;
 394         root->fs_info->last_insert.objectid = ins->objectid;
 395         ins->offset = num_blocks;
 396         return 0;
 397 error:
 398         btrfs_release_path(root, &path);
 399         return ret;
 400 }
 401
 402 /*
 403  * finds a free extent and does all the dirty work required for allocation
 404  * returns the key for the extent through ins, and a tree buffer for
 405  * the first block of the extent through buf.
 406  *
 407  * returns 0 if everything worked, non-zero otherwise.
 408  */
 409 static int alloc_extent(struct btrfs_trans_handle *trans, struct btrfs_root
 410                         *root, u64 num_blocks, u64 search_start, u64
 411                         search_end, u64 owner, struct btrfs_key *ins)
 412 {
 413         int ret;
 414         int pending_ret;
 415         u64 super_blocks_used;
 416         struct btrfs_fs_info *info = root->fs_info;
 417         struct btrfs_root *extent_root = info->extent_root;
 418         struct btrfs_extent_item extent_item;
 419
 420         btrfs_set_extent_refs(&extent_item, 1);
 421         btrfs_set_extent_owner(&extent_item, owner);
 422
 423         if (root == extent_root) {
 424                 BUG_ON(extent_root->fs_info->current_insert.offset == 0);
 425                 BUG_ON(num_blocks != 1);
 426                 BUG_ON(extent_root->fs_info->current_insert.flags ==
 427                        extent_root->fs_info->current_insert.offset);
 428                 ins->offset = 1;
 429                 ins->objectid = extent_root->fs_info->current_insert.objectid +
 430                                 extent_root->fs_info->current_insert.flags++;
 431                 return 0;
 432         }
 433         ret = find_free_extent(trans, root, num_blocks, search_start,
 434                                search_end, ins);
 435         if (ret)
 436                 return ret;
 437
 438         super_blocks_used = btrfs_super_blocks_used(info->disk_super);
 439         btrfs_set_super_blocks_used(info->disk_super, super_blocks_used +
 440                                     num_blocks);
 441         ret = btrfs_insert_item(trans, extent_root, ins, &extent_item,
 442                                 sizeof(extent_item));
 443
 444         finish_current_insert(trans, extent_root);
 445         pending_ret = del_pending_extents(trans, extent_root);
 446         if (ret)
 447                 return ret;
 448         if (pending_ret)
 449                 return pending_ret;
 450         return 0;
 451 }
 452
 453 /*
 454  * helper function to allocate a block for a given tree
 455  * returns the tree buffer or NULL.
 456  */
 457 struct buffer_head *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
 458                                             struct btrfs_root *root)
 459 {
 460         struct btrfs_key ins;
 461         int ret;
 462         struct buffer_head *buf;
 463
 464         ret = alloc_extent(trans, root, 1, 0, (unsigned long)-1,
 465                 btrfs_header_parentid(btrfs_buffer_header(root->node)), &ins);
 466         if (ret) {
 467                 BUG();
 468                 return NULL;
 469         }
 470         buf = find_tree_block(root, ins.objectid);
 471         set_buffer_uptodate(buf);
 472         return buf;
 473 }
 474
 475 /*
 476  * helper function for drop_snapshot, this walks down the tree dropping ref
 477  * counts as it goes.
 478  */
 479 static int walk_down_tree(struct btrfs_trans_handle *trans, struct btrfs_root
 480                           *root, struct btrfs_path *path, int *level)
 481 {
 482         struct buffer_head *next;
 483         struct buffer_head *cur;
 484         u64 blocknr;
 485         int ret;
 486         u32 refs;
 487
 488         ret = lookup_block_ref(trans, root, path->nodes[*level]->b_blocknr,
 489                                &refs);
 490         BUG_ON(ret);
 491         if (refs > 1)
 492                 goto out;
 493         /*
 494          * walk down to the last node level and free all the leaves
 495          */
 496         while(*level > 0) {
 497                 cur = path->nodes[*level];
 498                 if (path->slots[*level] >=
 499                     btrfs_header_nritems(btrfs_buffer_header(cur)))
 500                         break;
 501                 blocknr = btrfs_node_blockptr(btrfs_buffer_node(cur),
 502                                               path->slots[*level]);
 503                 ret = lookup_block_ref(trans, root, blocknr, &refs);
 504                 if (refs != 1 || *level == 1) {
 505                         path->slots[*level]++;
 506                         ret = btrfs_free_extent(trans, root, blocknr, 1, 1);
 507                         BUG_ON(ret);
 508                         continue;
 509                 }
 510                 BUG_ON(ret);
 511                 next = read_tree_block(root, blocknr);
 512                 if (path->nodes[*level-1])
 513                         btrfs_block_release(root, path->nodes[*level-1]);
 514                 path->nodes[*level-1] = next;
 515                 *level = btrfs_header_level(btrfs_buffer_header(next));
 516                 path->slots[*level] = 0;
 517         }
 518 out:
 519         ret = btrfs_free_extent(trans, root, path->nodes[*level]->b_blocknr,
 520                                 1, 1);
 521         btrfs_block_release(root, path->nodes[*level]);
 522         path->nodes[*level] = NULL;
 523         *level += 1;
 524         BUG_ON(ret);
 525         return 0;
 526 }
 527
 528 /*
 529  * helper for dropping snapshots.  This walks back up the tree in the path
 530  * to find the first node higher up where we haven't yet gone through
 531  * all the slots
 532  */
 533 static int walk_up_tree(struct btrfs_trans_handle *trans, struct btrfs_root
 534                         *root, struct btrfs_path *path, int *level)
 535 {
 536         int i;
 537         int slot;
 538         int ret;
 539         for(i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
 540                 slot = path->slots[i];
 541                 if (slot < btrfs_header_nritems(
 542                     btrfs_buffer_header(path->nodes[i])) - 1) {
 543                         path->slots[i]++;
 544                         *level = i;
 545                         return 0;
 546                 } else {
 547                         ret = btrfs_free_extent(trans, root,
 548                                                 path->nodes[*level]->b_blocknr,
 549                                                 1, 1);
 550                         btrfs_block_release(root, path->nodes[*level]);
 551                         path->nodes[*level] = NULL;
 552                         *level = i + 1;
 553                         BUG_ON(ret);
 554                 }
 555         }
 556         return 1;
 557 }
 558
 559 /*
 560  * drop the reference count on the tree rooted at 'snap'.  This traverses
 561  * the tree freeing any blocks that have a ref count of zero after being
 562  * decremented.
 563  */
 564 int btrfs_drop_snapshot(struct btrfs_trans_handle *trans, struct btrfs_root
 565                         *root, struct buffer_head *snap)
 566 {
 567         int ret = 0;
 568         int wret;
 569         int level;
 570         struct btrfs_path path;
 571         int i;
 572         int orig_level;
 573
 574         btrfs_init_path(&path);
 575
 576         level = btrfs_header_level(btrfs_buffer_header(snap));
 577         orig_level = level;
 578         path.nodes[level] = snap;
 579         path.slots[level] = 0;
 580         while(1) {
 581                 wret = walk_down_tree(trans, root, &path, &level);
 582                 if (wret > 0)
 583                         break;
 584                 if (wret < 0)
 585                         ret = wret;
 586
 587                 wret = walk_up_tree(trans, root, &path, &level);
 588                 if (wret > 0)
 589                         break;
 590                 if (wret < 0)
 591                         ret = wret;
 592         }
 593         for (i = 0; i <= orig_level; i++) {
 594                 if (path.nodes[i]) {
 595                         btrfs_block_release(root, path.nodes[i]);
 596                 }
 597         }
 598         return ret;
 599 }