fs/btrfs/free-space-cache.c

   1 /*
   2  * Copyright (C) 2008 Red Hat.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/pagemap.h>
  20 #include <linux/sched.h>
  21 #include <linux/slab.h>
  22 #include <linux/math64.h>
  23 #include "ctree.h"
  24 #include "free-space-cache.h"
  25 #include "transaction.h"
  26 #include "disk-io.h"
  27
  28 #define BITS_PER_BITMAP         (PAGE_CACHE_SIZE * 8)
  29 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
  30
  31 static void recalculate_thresholds(struct btrfs_block_group_cache
  32                                    *block_group);
  33 static int link_free_space(struct btrfs_block_group_cache *block_group,
  34                            struct btrfs_free_space *info);
  35
  36 struct inode *lookup_free_space_inode(struct btrfs_root *root,
  37                                       struct btrfs_block_group_cache
  38                                       *block_group, struct btrfs_path *path)
  39 {
  40         struct btrfs_key key;
  41         struct btrfs_key location;
  42         struct btrfs_disk_key disk_key;
  43         struct btrfs_free_space_header *header;
  44         struct extent_buffer *leaf;
  45         struct inode *inode = NULL;
  46         int ret;
  47
  48         spin_lock(&block_group->lock);
  49         if (block_group->inode)
  50                 inode = igrab(block_group->inode);
  51         spin_unlock(&block_group->lock);
  52         if (inode)
  53                 return inode;
  54
  55         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  56         key.offset = block_group->key.objectid;
  57         key.type = 0;
  58
  59         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  60         if (ret < 0)
  61                 return ERR_PTR(ret);
  62         if (ret > 0) {
  63                 btrfs_release_path(root, path);
  64                 return ERR_PTR(-ENOENT);
  65         }
  66
  67         leaf = path->nodes[0];
  68         header = btrfs_item_ptr(leaf, path->slots[0],
  69                                 struct btrfs_free_space_header);
  70         btrfs_free_space_key(leaf, header, &disk_key);
  71         btrfs_disk_key_to_cpu(&location, &disk_key);
  72         btrfs_release_path(root, path);
  73
  74         inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
  75         if (!inode)
  76                 return ERR_PTR(-ENOENT);
  77         if (IS_ERR(inode))
  78                 return inode;
  79         if (is_bad_inode(inode)) {
  80                 iput(inode);
  81                 return ERR_PTR(-ENOENT);
  82         }
  83
  84         spin_lock(&block_group->lock);
  85         if (!root->fs_info->closing) {
  86                 block_group->inode = igrab(inode);
  87                 block_group->iref = 1;
  88         }
  89         spin_unlock(&block_group->lock);
  90
  91         return inode;
  92 }
  93
  94 int create_free_space_inode(struct btrfs_root *root,
  95                             struct btrfs_trans_handle *trans,
  96                             struct btrfs_block_group_cache *block_group,
  97                             struct btrfs_path *path)
  98 {
  99         struct btrfs_key key;
 100         struct btrfs_disk_key disk_key;
 101         struct btrfs_free_space_header *header;
 102         struct btrfs_inode_item *inode_item;
 103         struct extent_buffer *leaf;
 104         u64 objectid;
 105         int ret;
 106
 107         ret = btrfs_find_free_objectid(trans, root, 0, &objectid);
 108         if (ret < 0)
 109                 return ret;
 110
 111         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
 112         if (ret)
 113                 return ret;
 114
 115         leaf = path->nodes[0];
 116         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 117                                     struct btrfs_inode_item);
 118         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 119         memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
 120                              sizeof(*inode_item));
 121         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 122         btrfs_set_inode_size(leaf, inode_item, 0);
 123         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 124         btrfs_set_inode_uid(leaf, inode_item, 0);
 125         btrfs_set_inode_gid(leaf, inode_item, 0);
 126         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 127         btrfs_set_inode_flags(leaf, inode_item, BTRFS_INODE_NOCOMPRESS |
 128                               BTRFS_INODE_PREALLOC | BTRFS_INODE_NODATASUM);
 129         btrfs_set_inode_nlink(leaf, inode_item, 1);
 130         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 131         btrfs_set_inode_block_group(leaf, inode_item,
 132                                     block_group->key.objectid);
 133         btrfs_mark_buffer_dirty(leaf);
 134         btrfs_release_path(root, path);
 135
 136         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 137         key.offset = block_group->key.objectid;
 138         key.type = 0;
 139
 140         ret = btrfs_insert_empty_item(trans, root, path, &key,
 141                                       sizeof(struct btrfs_free_space_header));
 142         if (ret < 0) {
 143                 btrfs_release_path(root, path);
 144                 return ret;
 145         }
 146         leaf = path->nodes[0];
 147         header = btrfs_item_ptr(leaf, path->slots[0],
 148                                 struct btrfs_free_space_header);
 149         memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
 150         btrfs_set_free_space_key(leaf, header, &disk_key);
 151         btrfs_mark_buffer_dirty(leaf);
 152         btrfs_release_path(root, path);
 153
 154         return 0;
 155 }
 156
 157 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
 158                                     struct btrfs_trans_handle *trans,
 159                                     struct btrfs_path *path,
 160                                     struct inode *inode)
 161 {
 162         loff_t oldsize;
 163         int ret = 0;
 164
 165         trans->block_rsv = root->orphan_block_rsv;
 166         ret = btrfs_block_rsv_check(trans, root,
 167                                     root->orphan_block_rsv,
 168                                     0, 5);
 169         if (ret)
 170                 return ret;
 171
 172         oldsize = i_size_read(inode);
 173         btrfs_i_size_write(inode, 0);
 174         truncate_pagecache(inode, oldsize, 0);
 175
 176         /*
 177          * We don't need an orphan item because truncating the free space cache
 178          * will never be split across transactions.
 179          */
 180         ret = btrfs_truncate_inode_items(trans, root, inode,
 181                                          0, BTRFS_EXTENT_DATA_KEY);
 182         if (ret) {
 183                 WARN_ON(1);
 184                 return ret;
 185         }
 186
 187         return btrfs_update_inode(trans, root, inode);
 188 }
 189
 190 static int readahead_cache(struct inode *inode)
 191 {
 192         struct file_ra_state *ra;
 193         unsigned long last_index;
 194
 195         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 196         if (!ra)
 197                 return -ENOMEM;
 198
 199         file_ra_state_init(ra, inode->i_mapping);
 200         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 201
 202         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 203
 204         kfree(ra);
 205
 206         return 0;
 207 }
 208
 209 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 210                           struct btrfs_block_group_cache *block_group)
 211 {
 212         struct btrfs_root *root = fs_info->tree_root;
 213         struct inode *inode;
 214         struct btrfs_free_space_header *header;
 215         struct extent_buffer *leaf;
 216         struct page *page;
 217         struct btrfs_path *path;
 218         u32 *checksums = NULL, *crc;
 219         char *disk_crcs = NULL;
 220         struct btrfs_key key;
 221         struct list_head bitmaps;
 222         u64 num_entries;
 223         u64 num_bitmaps;
 224         u64 generation;
 225         u32 cur_crc = ~(u32)0;
 226         pgoff_t index = 0;
 227         unsigned long first_page_offset;
 228         int num_checksums;
 229         int ret = 0;
 230
 231         /*
 232          * If we're unmounting then just return, since this does a search on the
 233          * normal root and not the commit root and we could deadlock.
 234          */
 235         smp_mb();
 236         if (fs_info->closing)
 237                 return 0;
 238
 239         /*
 240          * If this block group has been marked to be cleared for one reason or
 241          * another then we can't trust the on disk cache, so just return.
 242          */
 243         spin_lock(&block_group->lock);
 244         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 245                 spin_unlock(&block_group->lock);
 246                 return 0;
 247         }
 248         spin_unlock(&block_group->lock);
 249
 250         INIT_LIST_HEAD(&bitmaps);
 251
 252         path = btrfs_alloc_path();
 253         if (!path)
 254                 return 0;
 255
 256         inode = lookup_free_space_inode(root, block_group, path);
 257         if (IS_ERR(inode)) {
 258                 btrfs_free_path(path);
 259                 return 0;
 260         }
 261
 262         /* Nothing in the space cache, goodbye */
 263         if (!i_size_read(inode)) {
 264                 btrfs_free_path(path);
 265                 goto out;
 266         }
 267
 268         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 269         key.offset = block_group->key.objectid;
 270         key.type = 0;
 271
 272         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 273         if (ret) {
 274                 btrfs_free_path(path);
 275                 goto out;
 276         }
 277
 278         leaf = path->nodes[0];
 279         header = btrfs_item_ptr(leaf, path->slots[0],
 280                                 struct btrfs_free_space_header);
 281         num_entries = btrfs_free_space_entries(leaf, header);
 282         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 283         generation = btrfs_free_space_generation(leaf, header);
 284         btrfs_free_path(path);
 285
 286         if (BTRFS_I(inode)->generation != generation) {
 287                 printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
 288                        " not match free space cache generation (%llu) for "
 289                        "block group %llu\n",
 290                        (unsigned long long)BTRFS_I(inode)->generation,
 291                        (unsigned long long)generation,
 292                        (unsigned long long)block_group->key.objectid);
 293                 goto free_cache;
 294         }
 295
 296         if (!num_entries)
 297                 goto out;
 298
 299         /* Setup everything for doing checksumming */
 300         num_checksums = i_size_read(inode) / PAGE_CACHE_SIZE;
 301         checksums = crc = kzalloc(sizeof(u32) * num_checksums, GFP_NOFS);
 302         if (!checksums)
 303                 goto out;
 304         first_page_offset = (sizeof(u32) * num_checksums) + sizeof(u64);
 305         disk_crcs = kzalloc(first_page_offset, GFP_NOFS);
 306         if (!disk_crcs)
 307                 goto out;
 308
 309         ret = readahead_cache(inode);
 310         if (ret) {
 311                 ret = 0;
 312                 goto out;
 313         }
 314
 315         while (1) {
 316                 struct btrfs_free_space_entry *entry;
 317                 struct btrfs_free_space *e;
 318                 void *addr;
 319                 unsigned long offset = 0;
 320                 unsigned long start_offset = 0;
 321                 int need_loop = 0;
 322
 323                 if (!num_entries && !num_bitmaps)
 324                         break;
 325
 326                 if (index == 0) {
 327                         start_offset = first_page_offset;
 328                         offset = start_offset;
 329                 }
 330
 331                 page = grab_cache_page(inode->i_mapping, index);
 332                 if (!page) {
 333                         ret = 0;
 334                         goto free_cache;
 335                 }
 336
 337                 if (!PageUptodate(page)) {
 338                         btrfs_readpage(NULL, page);
 339                         lock_page(page);
 340                         if (!PageUptodate(page)) {
 341                                 unlock_page(page);
 342                                 page_cache_release(page);
 343                                 printk(KERN_ERR "btrfs: error reading free "
 344                                        "space cache: %llu\n",
 345                                        (unsigned long long)
 346                                        block_group->key.objectid);
 347                                 goto free_cache;
 348                         }
 349                 }
 350                 addr = kmap(page);
 351
 352                 if (index == 0) {
 353                         u64 *gen;
 354
 355                         memcpy(disk_crcs, addr, first_page_offset);
 356                         gen = addr + (sizeof(u32) * num_checksums);
 357                         if (*gen != BTRFS_I(inode)->generation) {
 358                                 printk(KERN_ERR "btrfs: space cache generation"
 359                                        " (%llu) does not match inode (%llu) "
 360                                        "for block group %llu\n",
 361                                        (unsigned long long)*gen,
 362                                        (unsigned long long)
 363                                        BTRFS_I(inode)->generation,
 364                                        (unsigned long long)
 365                                        block_group->key.objectid);
 366                                 kunmap(page);
 367                                 unlock_page(page);
 368                                 page_cache_release(page);
 369                                 goto free_cache;
 370                         }
 371                         crc = (u32 *)disk_crcs;
 372                 }
 373                 entry = addr + start_offset;
 374
 375                 /* First lets check our crc before we do anything fun */
 376                 cur_crc = ~(u32)0;
 377                 cur_crc = btrfs_csum_data(root, addr + start_offset, cur_crc,
 378                                           PAGE_CACHE_SIZE - start_offset);
 379                 btrfs_csum_final(cur_crc, (char *)&cur_crc);
 380                 if (cur_crc != *crc) {
 381                         printk(KERN_ERR "btrfs: crc mismatch for page %lu in "
 382                                "block group %llu\n", index,
 383                                (unsigned long long)block_group->key.objectid);
 384                         kunmap(page);
 385                         unlock_page(page);
 386                         page_cache_release(page);
 387                         goto free_cache;
 388                 }
 389                 crc++;
 390
 391                 while (1) {
 392                         if (!num_entries)
 393                                 break;
 394
 395                         need_loop = 1;
 396                         e = kzalloc(sizeof(struct btrfs_free_space), GFP_NOFS);
 397                         if (!e) {
 398                                 kunmap(page);
 399                                 unlock_page(page);
 400                                 page_cache_release(page);
 401                                 goto free_cache;
 402                         }
 403
 404                         e->offset = le64_to_cpu(entry->offset);
 405                         e->bytes = le64_to_cpu(entry->bytes);
 406                         if (!e->bytes) {
 407                                 kunmap(page);
 408                                 kfree(e);
 409                                 unlock_page(page);
 410                                 page_cache_release(page);
 411                                 goto free_cache;
 412                         }
 413
 414                         if (entry->type == BTRFS_FREE_SPACE_EXTENT) {
 415                                 spin_lock(&block_group->tree_lock);
 416                                 ret = link_free_space(block_group, e);
 417                                 spin_unlock(&block_group->tree_lock);
 418                                 BUG_ON(ret);
 419                         } else {
 420                                 e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
 421                                 if (!e->bitmap) {
 422                                         kunmap(page);
 423                                         kfree(e);
 424                                         unlock_page(page);
 425                                         page_cache_release(page);
 426                                         goto free_cache;
 427                                 }
 428                                 spin_lock(&block_group->tree_lock);
 429                                 ret = link_free_space(block_group, e);
 430                                 block_group->total_bitmaps++;
 431                                 recalculate_thresholds(block_group);
 432                                 spin_unlock(&block_group->tree_lock);
 433                                 list_add_tail(&e->list, &bitmaps);
 434                         }
 435
 436                         num_entries--;
 437                         offset += sizeof(struct btrfs_free_space_entry);
 438                         if (offset + sizeof(struct btrfs_free_space_entry) >=
 439                             PAGE_CACHE_SIZE)
 440                                 break;
 441                         entry++;
 442                 }
 443
 444                 /*
 445                  * We read an entry out of this page, we need to move on to the
 446                  * next page.
 447                  */
 448                 if (need_loop) {
 449                         kunmap(page);
 450                         goto next;
 451                 }
 452
 453                 /*
 454                  * We add the bitmaps at the end of the entries in order that
 455                  * the bitmap entries are added to the cache.
 456                  */
 457                 e = list_entry(bitmaps.next, struct btrfs_free_space, list);
 458                 list_del_init(&e->list);
 459                 memcpy(e->bitmap, addr, PAGE_CACHE_SIZE);
 460                 kunmap(page);
 461                 num_bitmaps--;
 462 next:
 463                 unlock_page(page);
 464                 page_cache_release(page);
 465                 index++;
 466         }
 467
 468         ret = 1;
 469 out:
 470         kfree(checksums);
 471         kfree(disk_crcs);
 472         iput(inode);
 473         return ret;
 474
 475 free_cache:
 476         /* This cache is bogus, make sure it gets cleared */
 477         spin_lock(&block_group->lock);
 478         block_group->disk_cache_state = BTRFS_DC_CLEAR;
 479         spin_unlock(&block_group->lock);
 480         btrfs_remove_free_space_cache(block_group);
 481         goto out;
 482 }
 483
 484 int btrfs_write_out_cache(struct btrfs_root *root,
 485                           struct btrfs_trans_handle *trans,
 486                           struct btrfs_block_group_cache *block_group,
 487                           struct btrfs_path *path)
 488 {
 489         struct btrfs_free_space_header *header;
 490         struct extent_buffer *leaf;
 491         struct inode *inode;
 492         struct rb_node *node;
 493         struct list_head *pos, *n;
 494         struct page *page;
 495         struct extent_state *cached_state = NULL;
 496         struct list_head bitmap_list;
 497         struct btrfs_key key;
 498         u64 bytes = 0;
 499         u32 *crc, *checksums;
 500         pgoff_t index = 0, last_index = 0;
 501         unsigned long first_page_offset;
 502         int num_checksums;
 503         int entries = 0;
 504         int bitmaps = 0;
 505         int ret = 0;
 506
 507         root = root->fs_info->tree_root;
 508
 509         INIT_LIST_HEAD(&bitmap_list);
 510
 511         spin_lock(&block_group->lock);
 512         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
 513                 spin_unlock(&block_group->lock);
 514                 return 0;
 515         }
 516         spin_unlock(&block_group->lock);
 517
 518         inode = lookup_free_space_inode(root, block_group, path);
 519         if (IS_ERR(inode))
 520                 return 0;
 521
 522         if (!i_size_read(inode)) {
 523                 iput(inode);
 524                 return 0;
 525         }
 526
 527         node = rb_first(&block_group->free_space_offset);
 528         if (!node) {
 529                 iput(inode);
 530                 return 0;
 531         }
 532
 533         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 534         filemap_write_and_wait(inode->i_mapping);
 535         btrfs_wait_ordered_range(inode, inode->i_size &
 536                                  ~(root->sectorsize - 1), (u64)-1);
 537
 538         /* We need a checksum per page. */
 539         num_checksums = i_size_read(inode) / PAGE_CACHE_SIZE;
 540         crc = checksums  = kzalloc(sizeof(u32) * num_checksums, GFP_NOFS);
 541         if (!crc) {
 542                 iput(inode);
 543                 return 0;
 544         }
 545
 546         /* Since the first page has all of our checksums and our generation we
 547          * need to calculate the offset into the page that we can start writing
 548          * our entries.
 549          */
 550         first_page_offset = (sizeof(u32) * num_checksums) + sizeof(u64);
 551
 552         /*
 553          * Lock all pages first so we can lock the extent safely.
 554          *
 555          * NOTE: Because we hold the ref the entire time we're going to write to
 556          * the page find_get_page should never fail, so we don't do a check
 557          * after find_get_page at this point.  Just putting this here so people
 558          * know and don't freak out.
 559          */
 560         while (index <= last_index) {
 561                 page = grab_cache_page(inode->i_mapping, index);
 562                 if (!page) {
 563                         pgoff_t i = 0;
 564
 565                         while (i < index) {
 566                                 page = find_get_page(inode->i_mapping, i);
 567                                 unlock_page(page);
 568                                 page_cache_release(page);
 569                                 page_cache_release(page);
 570                                 i++;
 571                         }
 572                         goto out_free;
 573                 }
 574                 index++;
 575         }
 576
 577         index = 0;
 578         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
 579                          0, &cached_state, GFP_NOFS);
 580
 581         /* Write out the extent entries */
 582         do {
 583                 struct btrfs_free_space_entry *entry;
 584                 void *addr;
 585                 unsigned long offset = 0;
 586                 unsigned long start_offset = 0;
 587
 588                 if (index == 0) {
 589                         start_offset = first_page_offset;
 590                         offset = start_offset;
 591                 }
 592
 593                 page = find_get_page(inode->i_mapping, index);
 594
 595                 addr = kmap(page);
 596                 entry = addr + start_offset;
 597
 598                 memset(addr, 0, PAGE_CACHE_SIZE);
 599                 while (1) {
 600                         struct btrfs_free_space *e;
 601
 602                         e = rb_entry(node, struct btrfs_free_space, offset_index);
 603                         entries++;
 604
 605                         entry->offset = cpu_to_le64(e->offset);
 606                         entry->bytes = cpu_to_le64(e->bytes);
 607                         if (e->bitmap) {
 608                                 entry->type = BTRFS_FREE_SPACE_BITMAP;
 609                                 list_add_tail(&e->list, &bitmap_list);
 610                                 bitmaps++;
 611                         } else {
 612                                 entry->type = BTRFS_FREE_SPACE_EXTENT;
 613                         }
 614                         node = rb_next(node);
 615                         if (!node)
 616                                 break;
 617                         offset += sizeof(struct btrfs_free_space_entry);
 618                         if (offset + sizeof(struct btrfs_free_space_entry) >=
 619                             PAGE_CACHE_SIZE)
 620                                 break;
 621                         entry++;
 622                 }
 623                 *crc = ~(u32)0;
 624                 *crc = btrfs_csum_data(root, addr + start_offset, *crc,
 625                                        PAGE_CACHE_SIZE - start_offset);
 626                 kunmap(page);
 627
 628                 btrfs_csum_final(*crc, (char *)crc);
 629                 crc++;
 630
 631                 bytes += PAGE_CACHE_SIZE;
 632
 633                 ClearPageChecked(page);
 634                 set_page_extent_mapped(page);
 635                 SetPageUptodate(page);
 636                 set_page_dirty(page);
 637
 638                 /*
 639                  * We need to release our reference we got for grab_cache_page,
 640                  * except for the first page which will hold our checksums, we
 641                  * do that below.
 642                  */
 643                 if (index != 0) {
 644                         unlock_page(page);
 645                         page_cache_release(page);
 646                 }
 647
 648                 page_cache_release(page);
 649
 650                 index++;
 651         } while (node);
 652
 653         /* Write out the bitmaps */
 654         list_for_each_safe(pos, n, &bitmap_list) {
 655                 void *addr;
 656                 struct btrfs_free_space *entry =
 657                         list_entry(pos, struct btrfs_free_space, list);
 658
 659                 page = find_get_page(inode->i_mapping, index);
 660
 661                 addr = kmap(page);
 662                 memcpy(addr, entry->bitmap, PAGE_CACHE_SIZE);
 663                 *crc = ~(u32)0;
 664                 *crc = btrfs_csum_data(root, addr, *crc, PAGE_CACHE_SIZE);
 665                 kunmap(page);
 666                 btrfs_csum_final(*crc, (char *)crc);
 667                 crc++;
 668                 bytes += PAGE_CACHE_SIZE;
 669
 670                 ClearPageChecked(page);
 671                 set_page_extent_mapped(page);
 672                 SetPageUptodate(page);
 673                 set_page_dirty(page);
 674                 unlock_page(page);
 675                 page_cache_release(page);
 676                 page_cache_release(page);
 677                 list_del_init(&entry->list);
 678                 index++;
 679         }
 680
 681         /* Zero out the rest of the pages just to make sure */
 682         while (index <= last_index) {
 683                 void *addr;
 684
 685                 page = find_get_page(inode->i_mapping, index);
 686
 687                 addr = kmap(page);
 688                 memset(addr, 0, PAGE_CACHE_SIZE);
 689                 kunmap(page);
 690                 ClearPageChecked(page);
 691                 set_page_extent_mapped(page);
 692                 SetPageUptodate(page);
 693                 set_page_dirty(page);
 694                 unlock_page(page);
 695                 page_cache_release(page);
 696                 page_cache_release(page);
 697                 bytes += PAGE_CACHE_SIZE;
 698                 index++;
 699         }
 700
 701         btrfs_set_extent_delalloc(inode, 0, bytes - 1, &cached_state);
 702
 703         /* Write the checksums and trans id to the first page */
 704         {
 705                 void *addr;
 706                 u64 *gen;
 707
 708                 page = find_get_page(inode->i_mapping, 0);
 709
 710                 addr = kmap(page);
 711                 memcpy(addr, checksums, sizeof(u32) * num_checksums);
 712                 gen = addr + (sizeof(u32) * num_checksums);
 713                 *gen = trans->transid;
 714                 kunmap(page);
 715                 ClearPageChecked(page);
 716                 set_page_extent_mapped(page);
 717                 SetPageUptodate(page);
 718                 set_page_dirty(page);
 719                 unlock_page(page);
 720                 page_cache_release(page);
 721                 page_cache_release(page);
 722         }
 723         BTRFS_I(inode)->generation = trans->transid;
 724
 725         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
 726                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
 727
 728         filemap_write_and_wait(inode->i_mapping);
 729
 730         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 731         key.offset = block_group->key.objectid;
 732         key.type = 0;
 733
 734         ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
 735         if (ret < 0) {
 736                 ret = 0;
 737                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
 738                                  EXTENT_DIRTY | EXTENT_DELALLOC |
 739                                  EXTENT_DO_ACCOUNTING, 0, 0, NULL, GFP_NOFS);
 740                 goto out_free;
 741         }
 742         leaf = path->nodes[0];
 743         if (ret > 0) {
 744                 struct btrfs_key found_key;
 745                 BUG_ON(!path->slots[0]);
 746                 path->slots[0]--;
 747                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
 748                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
 749                     found_key.offset != block_group->key.objectid) {
 750                         ret = 0;
 751                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
 752                                          EXTENT_DIRTY | EXTENT_DELALLOC |
 753                                          EXTENT_DO_ACCOUNTING, 0, 0, NULL,
 754                                          GFP_NOFS);
 755                         btrfs_release_path(root, path);
 756                         goto out_free;
 757                 }
 758         }
 759         header = btrfs_item_ptr(leaf, path->slots[0],
 760                                 struct btrfs_free_space_header);
 761         btrfs_set_free_space_entries(leaf, header, entries);
 762         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
 763         btrfs_set_free_space_generation(leaf, header, trans->transid);
 764         btrfs_mark_buffer_dirty(leaf);
 765         btrfs_release_path(root, path);
 766
 767         ret = 1;
 768
 769 out_free:
 770         if (ret == 0) {
 771                 invalidate_inode_pages2_range(inode->i_mapping, 0, index);
 772                 spin_lock(&block_group->lock);
 773                 block_group->disk_cache_state = BTRFS_DC_ERROR;
 774                 spin_unlock(&block_group->lock);
 775                 BTRFS_I(inode)->generation = 0;
 776         }
 777         kfree(checksums);
 778         btrfs_update_inode(trans, root, inode);
 779         iput(inode);
 780         return ret;
 781 }
 782
 783 static inline unsigned long offset_to_bit(u64 bitmap_start, u64 sectorsize,
 784                                           u64 offset)
 785 {
 786         BUG_ON(offset < bitmap_start);
 787         offset -= bitmap_start;
 788         return (unsigned long)(div64_u64(offset, sectorsize));
 789 }
 790
 791 static inline unsigned long bytes_to_bits(u64 bytes, u64 sectorsize)
 792 {
 793         return (unsigned long)(div64_u64(bytes, sectorsize));
 794 }
 795
 796 static inline u64 offset_to_bitmap(struct btrfs_block_group_cache *block_group,
 797                                    u64 offset)
 798 {
 799         u64 bitmap_start;
 800         u64 bytes_per_bitmap;
 801
 802         bytes_per_bitmap = BITS_PER_BITMAP * block_group->sectorsize;
 803         bitmap_start = offset - block_group->key.objectid;
 804         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
 805         bitmap_start *= bytes_per_bitmap;
 806         bitmap_start += block_group->key.objectid;
 807
 808         return bitmap_start;
 809 }
 810
 811 static int tree_insert_offset(struct rb_root *root, u64 offset,
 812                               struct rb_node *node, int bitmap)
 813 {
 814         struct rb_node **p = &root->rb_node;
 815         struct rb_node *parent = NULL;
 816         struct btrfs_free_space *info;
 817
 818         while (*p) {
 819                 parent = *p;
 820                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
 821
 822                 if (offset < info->offset) {
 823                         p = &(*p)->rb_left;
 824                 } else if (offset > info->offset) {
 825                         p = &(*p)->rb_right;
 826                 } else {
 827                         /*
 828                          * we could have a bitmap entry and an extent entry
 829                          * share the same offset.  If this is the case, we want
 830                          * the extent entry to always be found first if we do a
 831                          * linear search through the tree, since we want to have
 832                          * the quickest allocation time, and allocating from an
 833                          * extent is faster than allocating from a bitmap.  So
 834                          * if we're inserting a bitmap and we find an entry at
 835                          * this offset, we want to go right, or after this entry
 836                          * logically.  If we are inserting an extent and we've
 837                          * found a bitmap, we want to go left, or before
 838                          * logically.
 839                          */
 840                         if (bitmap) {
 841                                 WARN_ON(info->bitmap);
 842                                 p = &(*p)->rb_right;
 843                         } else {
 844                                 WARN_ON(!info->bitmap);
 845                                 p = &(*p)->rb_left;
 846                         }
 847                 }
 848         }
 849
 850         rb_link_node(node, parent, p);
 851         rb_insert_color(node, root);
 852
 853         return 0;
 854 }
 855
 856 /*
 857  * searches the tree for the given offset.
 858  *
 859  * fuzzy - If this is set, then we are trying to make an allocation, and we just
 860  * want a section that has at least bytes size and comes at or after the given
 861  * offset.
 862  */
 863 static struct btrfs_free_space *
 864 tree_search_offset(struct btrfs_block_group_cache *block_group,
 865                    u64 offset, int bitmap_only, int fuzzy)
 866 {
 867         struct rb_node *n = block_group->free_space_offset.rb_node;
 868         struct btrfs_free_space *entry, *prev = NULL;
 869
 870         /* find entry that is closest to the 'offset' */
 871         while (1) {
 872                 if (!n) {
 873                         entry = NULL;
 874                         break;
 875                 }
 876
 877                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
 878                 prev = entry;
 879
 880                 if (offset < entry->offset)
 881                         n = n->rb_left;
 882                 else if (offset > entry->offset)
 883                         n = n->rb_right;
 884                 else
 885                         break;
 886         }
 887
 888         if (bitmap_only) {
 889                 if (!entry)
 890                         return NULL;
 891                 if (entry->bitmap)
 892                         return entry;
 893
 894                 /*
 895                  * bitmap entry and extent entry may share same offset,
 896                  * in that case, bitmap entry comes after extent entry.
 897                  */
 898                 n = rb_next(n);
 899                 if (!n)
 900                         return NULL;
 901                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
 902                 if (entry->offset != offset)
 903                         return NULL;
 904
 905                 WARN_ON(!entry->bitmap);
 906                 return entry;
 907         } else if (entry) {
 908                 if (entry->bitmap) {
 909                         /*
 910                          * if previous extent entry covers the offset,
 911                          * we should return it instead of the bitmap entry
 912                          */
 913                         n = &entry->offset_index;
 914                         while (1) {
 915                                 n = rb_prev(n);
 916                                 if (!n)
 917                                         break;
 918                                 prev = rb_entry(n, struct btrfs_free_space,
 919                                                 offset_index);
 920                                 if (!prev->bitmap) {
 921                                         if (prev->offset + prev->bytes > offset)
 922                                                 entry = prev;
 923                                         break;
 924                                 }
 925                         }
 926                 }
 927                 return entry;
 928         }
 929
 930         if (!prev)
 931                 return NULL;
 932
 933         /* find last entry before the 'offset' */
 934         entry = prev;
 935         if (entry->offset > offset) {
 936                 n = rb_prev(&entry->offset_index);
 937                 if (n) {
 938                         entry = rb_entry(n, struct btrfs_free_space,
 939                                         offset_index);
 940                         BUG_ON(entry->offset > offset);
 941                 } else {
 942                         if (fuzzy)
 943                                 return entry;
 944                         else
 945                                 return NULL;
 946                 }
 947         }
 948
 949         if (entry->bitmap) {
 950                 n = &entry->offset_index;
 951                 while (1) {
 952                         n = rb_prev(n);
 953                         if (!n)
 954                                 break;
 955                         prev = rb_entry(n, struct btrfs_free_space,
 956                                         offset_index);
 957                         if (!prev->bitmap) {
 958                                 if (prev->offset + prev->bytes > offset)
 959                                         return prev;
 960                                 break;
 961                         }
 962                 }
 963                 if (entry->offset + BITS_PER_BITMAP *
 964                     block_group->sectorsize > offset)
 965                         return entry;
 966         } else if (entry->offset + entry->bytes > offset)
 967                 return entry;
 968
 969         if (!fuzzy)
 970                 return NULL;
 971
 972         while (1) {
 973                 if (entry->bitmap) {
 974                         if (entry->offset + BITS_PER_BITMAP *
 975                             block_group->sectorsize > offset)
 976                                 break;
 977                 } else {
 978                         if (entry->offset + entry->bytes > offset)
 979                                 break;
 980                 }
 981
 982                 n = rb_next(&entry->offset_index);
 983                 if (!n)
 984                         return NULL;
 985                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
 986         }
 987         return entry;
 988 }
 989
 990 static inline void
 991 __unlink_free_space(struct btrfs_block_group_cache *block_group,
 992                     struct btrfs_free_space *info)
 993 {
 994         rb_erase(&info->offset_index, &block_group->free_space_offset);
 995         block_group->free_extents--;
 996 }
 997
 998 static void unlink_free_space(struct btrfs_block_group_cache *block_group,
 999                               struct btrfs_free_space *info)
1000 {
1001         __unlink_free_space(block_group, info);
1002         block_group->free_space -= info->bytes;
1003 }
1004
1005 static int link_free_space(struct btrfs_block_group_cache *block_group,
1006                            struct btrfs_free_space *info)
1007 {
1008         int ret = 0;
1009
1010         BUG_ON(!info->bitmap && !info->bytes);
1011         ret = tree_insert_offset(&block_group->free_space_offset, info->offset,
1012                                  &info->offset_index, (info->bitmap != NULL));
1013         if (ret)
1014                 return ret;
1015
1016         block_group->free_space += info->bytes;
1017         block_group->free_extents++;
1018         return ret;
1019 }
1020
1021 static void recalculate_thresholds(struct btrfs_block_group_cache *block_group)
1022 {
1023         u64 max_bytes;
1024         u64 bitmap_bytes;
1025         u64 extent_bytes;
1026         u64 size = block_group->key.offset;
1027
1028         /*
1029          * The goal is to keep the total amount of memory used per 1gb of space
1030          * at or below 32k, so we need to adjust how much memory we allow to be
1031          * used by extent based free space tracking
1032          */
1033         if (size < 1024 * 1024 * 1024)
1034                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1035         else
1036                 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1037                         div64_u64(size, 1024 * 1024 * 1024);
1038
1039         /*
1040          * we want to account for 1 more bitmap than what we have so we can make
1041          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1042          * we add more bitmaps.
1043          */
1044         bitmap_bytes = (block_group->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1045
1046         if (bitmap_bytes >= max_bytes) {
1047                 block_group->extents_thresh = 0;
1048                 return;
1049         }
1050
1051         /*
1052          * we want the extent entry threshold to always be at most 1/2 the maxw
1053          * bytes we can have, or whatever is less than that.
1054          */
1055         extent_bytes = max_bytes - bitmap_bytes;
1056         extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1057
1058         block_group->extents_thresh =
1059                 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1060 }
1061
1062 static void bitmap_clear_bits(struct btrfs_block_group_cache *block_group,
1063                               struct btrfs_free_space *info, u64 offset,
1064                               u64 bytes)
1065 {
1066         unsigned long start, end;
1067         unsigned long i;
1068
1069         start = offset_to_bit(info->offset, block_group->sectorsize, offset);
1070         end = start + bytes_to_bits(bytes, block_group->sectorsize);
1071         BUG_ON(end > BITS_PER_BITMAP);
1072
1073         for (i = start; i < end; i++)
1074                 clear_bit(i, info->bitmap);
1075
1076         info->bytes -= bytes;
1077         block_group->free_space -= bytes;
1078 }
1079
1080 static void bitmap_set_bits(struct btrfs_block_group_cache *block_group,
1081                             struct btrfs_free_space *info, u64 offset,
1082                             u64 bytes)
1083 {
1084         unsigned long start, end;
1085         unsigned long i;
1086
1087         start = offset_to_bit(info->offset, block_group->sectorsize, offset);
1088         end = start + bytes_to_bits(bytes, block_group->sectorsize);
1089         BUG_ON(end > BITS_PER_BITMAP);
1090
1091         for (i = start; i < end; i++)
1092                 set_bit(i, info->bitmap);
1093
1094         info->bytes += bytes;
1095         block_group->free_space += bytes;
1096 }
1097
1098 static int search_bitmap(struct btrfs_block_group_cache *block_group,
1099                          struct btrfs_free_space *bitmap_info, u64 *offset,
1100                          u64 *bytes)
1101 {
1102         unsigned long found_bits = 0;
1103         unsigned long bits, i;
1104         unsigned long next_zero;
1105
1106         i = offset_to_bit(bitmap_info->offset, block_group->sectorsize,
1107                           max_t(u64, *offset, bitmap_info->offset));
1108         bits = bytes_to_bits(*bytes, block_group->sectorsize);
1109
1110         for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
1111              i < BITS_PER_BITMAP;
1112              i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
1113                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1114                                                BITS_PER_BITMAP, i);
1115                 if ((next_zero - i) >= bits) {
1116                         found_bits = next_zero - i;
1117                         break;
1118                 }
1119                 i = next_zero;
1120         }
1121
1122         if (found_bits) {
1123                 *offset = (u64)(i * block_group->sectorsize) +
1124                         bitmap_info->offset;
1125                 *bytes = (u64)(found_bits) * block_group->sectorsize;
1126                 return 0;
1127         }
1128
1129         return -1;
1130 }
1131
1132 static struct btrfs_free_space *find_free_space(struct btrfs_block_group_cache
1133                                                 *block_group, u64 *offset,
1134                                                 u64 *bytes, int debug)
1135 {
1136         struct btrfs_free_space *entry;
1137         struct rb_node *node;
1138         int ret;
1139
1140         if (!block_group->free_space_offset.rb_node)
1141                 return NULL;
1142
1143         entry = tree_search_offset(block_group,
1144                                    offset_to_bitmap(block_group, *offset),
1145                                    0, 1);
1146         if (!entry)
1147                 return NULL;
1148
1149         for (node = &entry->offset_index; node; node = rb_next(node)) {
1150                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1151                 if (entry->bytes < *bytes)
1152                         continue;
1153
1154                 if (entry->bitmap) {
1155                         ret = search_bitmap(block_group, entry, offset, bytes);
1156                         if (!ret)
1157                                 return entry;
1158                         continue;
1159                 }
1160
1161                 *offset = entry->offset;
1162                 *bytes = entry->bytes;
1163                 return entry;
1164         }
1165
1166         return NULL;
1167 }
1168
1169 static void add_new_bitmap(struct btrfs_block_group_cache *block_group,
1170                            struct btrfs_free_space *info, u64 offset)
1171 {
1172         u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
1173         int max_bitmaps = (int)div64_u64(block_group->key.offset +
1174                                          bytes_per_bg - 1, bytes_per_bg);
1175         BUG_ON(block_group->total_bitmaps >= max_bitmaps);
1176
1177         info->offset = offset_to_bitmap(block_group, offset);
1178         info->bytes = 0;
1179         link_free_space(block_group, info);
1180         block_group->total_bitmaps++;
1181
1182         recalculate_thresholds(block_group);
1183 }
1184
1185 static void free_bitmap(struct btrfs_block_group_cache *block_group,
1186                         struct btrfs_free_space *bitmap_info)
1187 {
1188         unlink_free_space(block_group, bitmap_info);
1189         kfree(bitmap_info->bitmap);
1190         kfree(bitmap_info);
1191         block_group->total_bitmaps--;
1192         recalculate_thresholds(block_group);
1193 }
1194
1195 static noinline int remove_from_bitmap(struct btrfs_block_group_cache *block_group,
1196                               struct btrfs_free_space *bitmap_info,
1197                               u64 *offset, u64 *bytes)
1198 {
1199         u64 end;
1200         u64 search_start, search_bytes;
1201         int ret;
1202
1203 again:
1204         end = bitmap_info->offset +
1205                 (u64)(BITS_PER_BITMAP * block_group->sectorsize) - 1;
1206
1207         /*
1208          * XXX - this can go away after a few releases.
1209          *
1210          * since the only user of btrfs_remove_free_space is the tree logging
1211          * stuff, and the only way to test that is under crash conditions, we
1212          * want to have this debug stuff here just in case somethings not
1213          * working.  Search the bitmap for the space we are trying to use to
1214          * make sure its actually there.  If its not there then we need to stop
1215          * because something has gone wrong.
1216          */
1217         search_start = *offset;
1218         search_bytes = *bytes;
1219         search_bytes = min(search_bytes, end - search_start + 1);
1220         ret = search_bitmap(block_group, bitmap_info, &search_start,
1221                             &search_bytes);
1222         BUG_ON(ret < 0 || search_start != *offset);
1223
1224         if (*offset > bitmap_info->offset && *offset + *bytes > end) {
1225                 bitmap_clear_bits(block_group, bitmap_info, *offset,
1226                                   end - *offset + 1);
1227                 *bytes -= end - *offset + 1;
1228                 *offset = end + 1;
1229         } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
1230                 bitmap_clear_bits(block_group, bitmap_info, *offset, *bytes);
1231                 *bytes = 0;
1232         }
1233
1234         if (*bytes) {
1235                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1236                 if (!bitmap_info->bytes)
1237                         free_bitmap(block_group, bitmap_info);
1238
1239                 /*
1240                  * no entry after this bitmap, but we still have bytes to
1241                  * remove, so something has gone wrong.
1242                  */
1243                 if (!next)
1244                         return -EINVAL;
1245
1246                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1247                                        offset_index);
1248
1249                 /*
1250                  * if the next entry isn't a bitmap we need to return to let the
1251                  * extent stuff do its work.
1252                  */
1253                 if (!bitmap_info->bitmap)
1254                         return -EAGAIN;
1255
1256                 /*
1257                  * Ok the next item is a bitmap, but it may not actually hold
1258                  * the information for the rest of this free space stuff, so
1259                  * look for it, and if we don't find it return so we can try
1260                  * everything over again.
1261                  */
1262                 search_start = *offset;
1263                 search_bytes = *bytes;
1264                 ret = search_bitmap(block_group, bitmap_info, &search_start,
1265                                     &search_bytes);
1266                 if (ret < 0 || search_start != *offset)
1267                         return -EAGAIN;
1268
1269                 goto again;
1270         } else if (!bitmap_info->bytes)
1271                 free_bitmap(block_group, bitmap_info);
1272
1273         return 0;
1274 }
1275
1276 static int insert_into_bitmap(struct btrfs_block_group_cache *block_group,
1277                               struct btrfs_free_space *info)
1278 {
1279         struct btrfs_free_space *bitmap_info;
1280         int added = 0;
1281         u64 bytes, offset, end;
1282         int ret;
1283
1284         /*
1285          * If we are below the extents threshold then we can add this as an
1286          * extent, and don't have to deal with the bitmap
1287          */
1288         if (block_group->free_extents < block_group->extents_thresh &&
1289             info->bytes > block_group->sectorsize * 4)
1290                 return 0;
1291
1292         /*
1293          * some block groups are so tiny they can't be enveloped by a bitmap, so
1294          * don't even bother to create a bitmap for this
1295          */
1296         if (BITS_PER_BITMAP * block_group->sectorsize >
1297             block_group->key.offset)
1298                 return 0;
1299
1300         bytes = info->bytes;
1301         offset = info->offset;
1302
1303 again:
1304         bitmap_info = tree_search_offset(block_group,
1305                                          offset_to_bitmap(block_group, offset),
1306                                          1, 0);
1307         if (!bitmap_info) {
1308                 BUG_ON(added);
1309                 goto new_bitmap;
1310         }
1311
1312         end = bitmap_info->offset +
1313                 (u64)(BITS_PER_BITMAP * block_group->sectorsize);
1314
1315         if (offset >= bitmap_info->offset && offset + bytes > end) {
1316                 bitmap_set_bits(block_group, bitmap_info, offset,
1317                                 end - offset);
1318                 bytes -= end - offset;
1319                 offset = end;
1320                 added = 0;
1321         } else if (offset >= bitmap_info->offset && offset + bytes <= end) {
1322                 bitmap_set_bits(block_group, bitmap_info, offset, bytes);
1323                 bytes = 0;
1324         } else {
1325                 BUG();
1326         }
1327
1328         if (!bytes) {
1329                 ret = 1;
1330                 goto out;
1331         } else
1332                 goto again;
1333
1334 new_bitmap:
1335         if (info && info->bitmap) {
1336                 add_new_bitmap(block_group, info, offset);
1337                 added = 1;
1338                 info = NULL;
1339                 goto again;
1340         } else {
1341                 spin_unlock(&block_group->tree_lock);
1342
1343                 /* no pre-allocated info, allocate a new one */
1344                 if (!info) {
1345                         info = kzalloc(sizeof(struct btrfs_free_space),
1346                                        GFP_NOFS);
1347                         if (!info) {
1348                                 spin_lock(&block_group->tree_lock);
1349                                 ret = -ENOMEM;
1350                                 goto out;
1351                         }
1352                 }
1353
1354                 /* allocate the bitmap */
1355                 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1356                 spin_lock(&block_group->tree_lock);
1357                 if (!info->bitmap) {
1358                         ret = -ENOMEM;
1359                         goto out;
1360                 }
1361                 goto again;
1362         }
1363
1364 out:
1365         if (info) {
1366                 if (info->bitmap)
1367                         kfree(info->bitmap);
1368                 kfree(info);
1369         }
1370
1371         return ret;
1372 }
1373
1374 bool try_merge_free_space(struct btrfs_block_group_cache *block_group,
1375                           struct btrfs_free_space *info, bool update_stat)
1376 {
1377         struct btrfs_free_space *left_info;
1378         struct btrfs_free_space *right_info;
1379         bool merged = false;
1380         u64 offset = info->offset;
1381         u64 bytes = info->bytes;
1382
1383         /*
1384          * first we want to see if there is free space adjacent to the range we
1385          * are adding, if there is remove that struct and add a new one to
1386          * cover the entire range
1387          */
1388         right_info = tree_search_offset(block_group, offset + bytes, 0, 0);
1389         if (right_info && rb_prev(&right_info->offset_index))
1390                 left_info = rb_entry(rb_prev(&right_info->offset_index),
1391                                      struct btrfs_free_space, offset_index);
1392         else
1393                 left_info = tree_search_offset(block_group, offset - 1, 0, 0);
1394
1395         if (right_info && !right_info->bitmap) {
1396                 if (update_stat)
1397                         unlink_free_space(block_group, right_info);
1398                 else
1399                         __unlink_free_space(block_group, right_info);
1400                 info->bytes += right_info->bytes;
1401                 kfree(right_info);
1402                 merged = true;
1403         }
1404
1405         if (left_info && !left_info->bitmap &&
1406             left_info->offset + left_info->bytes == offset) {
1407                 if (update_stat)
1408                         unlink_free_space(block_group, left_info);
1409                 else
1410                         __unlink_free_space(block_group, left_info);
1411                 info->offset = left_info->offset;
1412                 info->bytes += left_info->bytes;
1413                 kfree(left_info);
1414                 merged = true;
1415         }
1416
1417         return merged;
1418 }
1419
1420 int btrfs_add_free_space(struct btrfs_block_group_cache *block_group,
1421                          u64 offset, u64 bytes)
1422 {
1423         struct btrfs_free_space *info;
1424         int ret = 0;
1425
1426         info = kzalloc(sizeof(struct btrfs_free_space), GFP_NOFS);
1427         if (!info)
1428                 return -ENOMEM;
1429
1430         info->offset = offset;
1431         info->bytes = bytes;
1432
1433         spin_lock(&block_group->tree_lock);
1434
1435         if (try_merge_free_space(block_group, info, true))
1436                 goto link;
1437
1438         /*
1439          * There was no extent directly to the left or right of this new
1440          * extent then we know we're going to have to allocate a new extent, so
1441          * before we do that see if we need to drop this into a bitmap
1442          */
1443         ret = insert_into_bitmap(block_group, info);
1444         if (ret < 0) {
1445                 goto out;
1446         } else if (ret) {
1447                 ret = 0;
1448                 goto out;
1449         }
1450 link:
1451         ret = link_free_space(block_group, info);
1452         if (ret)
1453                 kfree(info);
1454 out:
1455         spin_unlock(&block_group->tree_lock);
1456
1457         if (ret) {
1458                 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1459                 BUG_ON(ret == -EEXIST);
1460         }
1461
1462         return ret;
1463 }
1464
1465 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1466                             u64 offset, u64 bytes)
1467 {
1468         struct btrfs_free_space *info;
1469         struct btrfs_free_space *next_info = NULL;
1470         int ret = 0;
1471
1472         spin_lock(&block_group->tree_lock);
1473
1474 again:
1475         info = tree_search_offset(block_group, offset, 0, 0);
1476         if (!info) {
1477                 /*
1478                  * oops didn't find an extent that matched the space we wanted
1479                  * to remove, look for a bitmap instead
1480                  */
1481                 info = tree_search_offset(block_group,
1482                                           offset_to_bitmap(block_group, offset),
1483                                           1, 0);
1484                 if (!info) {
1485                         WARN_ON(1);
1486                         goto out_lock;
1487                 }
1488         }
1489
1490         if (info->bytes < bytes && rb_next(&info->offset_index)) {
1491                 u64 end;
1492                 next_info = rb_entry(rb_next(&info->offset_index),
1493                                              struct btrfs_free_space,
1494                                              offset_index);
1495
1496                 if (next_info->bitmap)
1497                         end = next_info->offset + BITS_PER_BITMAP *
1498                                 block_group->sectorsize - 1;
1499                 else
1500                         end = next_info->offset + next_info->bytes;
1501
1502                 if (next_info->bytes < bytes ||
1503                     next_info->offset > offset || offset > end) {
1504                         printk(KERN_CRIT "Found free space at %llu, size %llu,"
1505                               " trying to use %llu\n",
1506                               (unsigned long long)info->offset,
1507                               (unsigned long long)info->bytes,
1508                               (unsigned long long)bytes);
1509                         WARN_ON(1);
1510                         ret = -EINVAL;
1511                         goto out_lock;
1512                 }
1513
1514                 info = next_info;
1515         }
1516
1517         if (info->bytes == bytes) {
1518                 unlink_free_space(block_group, info);
1519                 if (info->bitmap) {
1520                         kfree(info->bitmap);
1521                         block_group->total_bitmaps--;
1522                 }
1523                 kfree(info);
1524                 goto out_lock;
1525         }
1526
1527         if (!info->bitmap && info->offset == offset) {
1528                 unlink_free_space(block_group, info);
1529                 info->offset += bytes;
1530                 info->bytes -= bytes;
1531                 link_free_space(block_group, info);
1532                 goto out_lock;
1533         }
1534
1535         if (!info->bitmap && info->offset <= offset &&
1536             info->offset + info->bytes >= offset + bytes) {
1537                 u64 old_start = info->offset;
1538                 /*
1539                  * we're freeing space in the middle of the info,
1540                  * this can happen during tree log replay
1541                  *
1542                  * first unlink the old info and then
1543                  * insert it again after the hole we're creating
1544                  */
1545                 unlink_free_space(block_group, info);
1546                 if (offset + bytes < info->offset + info->bytes) {
1547                         u64 old_end = info->offset + info->bytes;
1548
1549                         info->offset = offset + bytes;
1550                         info->bytes = old_end - info->offset;
1551                         ret = link_free_space(block_group, info);
1552                         WARN_ON(ret);
1553                         if (ret)
1554                                 goto out_lock;
1555                 } else {
1556                         /* the hole we're creating ends at the end
1557                          * of the info struct, just free the info
1558                          */
1559                         kfree(info);
1560                 }
1561                 spin_unlock(&block_group->tree_lock);
1562
1563                 /* step two, insert a new info struct to cover
1564                  * anything before the hole
1565                  */
1566                 ret = btrfs_add_free_space(block_group, old_start,
1567                                            offset - old_start);
1568                 WARN_ON(ret);
1569                 goto out;
1570         }
1571
1572         ret = remove_from_bitmap(block_group, info, &offset, &bytes);
1573         if (ret == -EAGAIN)
1574                 goto again;
1575         BUG_ON(ret);
1576 out_lock:
1577         spin_unlock(&block_group->tree_lock);
1578 out:
1579         return ret;
1580 }
1581
1582 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1583                            u64 bytes)
1584 {
1585         struct btrfs_free_space *info;
1586         struct rb_node *n;
1587         int count = 0;
1588
1589         for (n = rb_first(&block_group->free_space_offset); n; n = rb_next(n)) {
1590                 info = rb_entry(n, struct btrfs_free_space, offset_index);
1591                 if (info->bytes >= bytes)
1592                         count++;
1593                 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1594                        (unsigned long long)info->offset,
1595                        (unsigned long long)info->bytes,
1596                        (info->bitmap) ? "yes" : "no");
1597         }
1598         printk(KERN_INFO "block group has cluster?: %s\n",
1599                list_empty(&block_group->cluster_list) ? "no" : "yes");
1600         printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1601                "\n", count);
1602 }
1603
1604 u64 btrfs_block_group_free_space(struct btrfs_block_group_cache *block_group)
1605 {
1606         struct btrfs_free_space *info;
1607         struct rb_node *n;
1608         u64 ret = 0;
1609
1610         for (n = rb_first(&block_group->free_space_offset); n;
1611              n = rb_next(n)) {
1612                 info = rb_entry(n, struct btrfs_free_space, offset_index);
1613                 ret += info->bytes;
1614         }
1615
1616         return ret;
1617 }
1618
1619 /*
1620  * for a given cluster, put all of its extents back into the free
1621  * space cache.  If the block group passed doesn't match the block group
1622  * pointed to by the cluster, someone else raced in and freed the
1623  * cluster already.  In that case, we just return without changing anything
1624  */
1625 static int
1626 __btrfs_return_cluster_to_free_space(
1627                              struct btrfs_block_group_cache *block_group,
1628                              struct btrfs_free_cluster *cluster)
1629 {
1630         struct btrfs_free_space *entry;
1631         struct rb_node *node;
1632         bool bitmap;
1633
1634         spin_lock(&cluster->lock);
1635         if (cluster->block_group != block_group)
1636                 goto out;
1637
1638         bitmap = cluster->points_to_bitmap;
1639         cluster->block_group = NULL;
1640         cluster->window_start = 0;
1641         list_del_init(&cluster->block_group_list);
1642         cluster->points_to_bitmap = false;
1643
1644         if (bitmap)
1645                 goto out;
1646
1647         node = rb_first(&cluster->root);
1648         while (node) {
1649                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1650                 node = rb_next(&entry->offset_index);
1651                 rb_erase(&entry->offset_index, &cluster->root);
1652                 BUG_ON(entry->bitmap);
1653                 try_merge_free_space(block_group, entry, false);
1654                 tree_insert_offset(&block_group->free_space_offset,
1655                                    entry->offset, &entry->offset_index, 0);
1656         }
1657         cluster->root = RB_ROOT;
1658
1659 out:
1660         spin_unlock(&cluster->lock);
1661         btrfs_put_block_group(block_group);
1662         return 0;
1663 }
1664
1665 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
1666 {
1667         struct btrfs_free_space *info;
1668         struct rb_node *node;
1669         struct btrfs_free_cluster *cluster;
1670         struct list_head *head;
1671
1672         spin_lock(&block_group->tree_lock);
1673         while ((head = block_group->cluster_list.next) !=
1674                &block_group->cluster_list) {
1675                 cluster = list_entry(head, struct btrfs_free_cluster,
1676                                      block_group_list);
1677
1678                 WARN_ON(cluster->block_group != block_group);
1679                 __btrfs_return_cluster_to_free_space(block_group, cluster);
1680                 if (need_resched()) {
1681                         spin_unlock(&block_group->tree_lock);
1682                         cond_resched();
1683                         spin_lock(&block_group->tree_lock);
1684                 }
1685         }
1686
1687         while ((node = rb_last(&block_group->free_space_offset)) != NULL) {
1688                 info = rb_entry(node, struct btrfs_free_space, offset_index);
1689                 unlink_free_space(block_group, info);
1690                 if (info->bitmap)
1691                         kfree(info->bitmap);
1692                 kfree(info);
1693                 if (need_resched()) {
1694                         spin_unlock(&block_group->tree_lock);
1695                         cond_resched();
1696                         spin_lock(&block_group->tree_lock);
1697                 }
1698         }
1699
1700         spin_unlock(&block_group->tree_lock);
1701 }
1702
1703 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
1704                                u64 offset, u64 bytes, u64 empty_size)
1705 {
1706         struct btrfs_free_space *entry = NULL;
1707         u64 bytes_search = bytes + empty_size;
1708         u64 ret = 0;
1709
1710         spin_lock(&block_group->tree_lock);
1711         entry = find_free_space(block_group, &offset, &bytes_search, 0);
1712         if (!entry)
1713                 goto out;
1714
1715         ret = offset;
1716         if (entry->bitmap) {
1717                 bitmap_clear_bits(block_group, entry, offset, bytes);
1718                 if (!entry->bytes)
1719                         free_bitmap(block_group, entry);
1720         } else {
1721                 unlink_free_space(block_group, entry);
1722                 entry->offset += bytes;
1723                 entry->bytes -= bytes;
1724                 if (!entry->bytes)
1725                         kfree(entry);
1726                 else
1727                         link_free_space(block_group, entry);
1728         }
1729
1730 out:
1731         spin_unlock(&block_group->tree_lock);
1732
1733         return ret;
1734 }
1735
1736 /*
1737  * given a cluster, put all of its extents back into the free space
1738  * cache.  If a block group is passed, this function will only free
1739  * a cluster that belongs to the passed block group.
1740  *
1741  * Otherwise, it'll get a reference on the block group pointed to by the
1742  * cluster and remove the cluster from it.
1743  */
1744 int btrfs_return_cluster_to_free_space(
1745                                struct btrfs_block_group_cache *block_group,
1746                                struct btrfs_free_cluster *cluster)
1747 {
1748         int ret;
1749
1750         /* first, get a safe pointer to the block group */
1751         spin_lock(&cluster->lock);
1752         if (!block_group) {
1753                 block_group = cluster->block_group;
1754                 if (!block_group) {
1755                         spin_unlock(&cluster->lock);
1756                         return 0;
1757                 }
1758         } else if (cluster->block_group != block_group) {
1759                 /* someone else has already freed it don't redo their work */
1760                 spin_unlock(&cluster->lock);
1761                 return 0;
1762         }
1763         atomic_inc(&block_group->count);
1764         spin_unlock(&cluster->lock);
1765
1766         /* now return any extents the cluster had on it */
1767         spin_lock(&block_group->tree_lock);
1768         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
1769         spin_unlock(&block_group->tree_lock);
1770
1771         /* finally drop our ref */
1772         btrfs_put_block_group(block_group);
1773         return ret;
1774 }
1775
1776 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
1777                                    struct btrfs_free_cluster *cluster,
1778                                    u64 bytes, u64 min_start)
1779 {
1780         struct btrfs_free_space *entry;
1781         int err;
1782         u64 search_start = cluster->window_start;
1783         u64 search_bytes = bytes;
1784         u64 ret = 0;
1785
1786         spin_lock(&block_group->tree_lock);
1787         spin_lock(&cluster->lock);
1788
1789         if (!cluster->points_to_bitmap)
1790                 goto out;
1791
1792         if (cluster->block_group != block_group)
1793                 goto out;
1794
1795         /*
1796          * search_start is the beginning of the bitmap, but at some point it may
1797          * be a good idea to point to the actual start of the free area in the
1798          * bitmap, so do the offset_to_bitmap trick anyway, and set bitmap_only
1799          * to 1 to make sure we get the bitmap entry
1800          */
1801         entry = tree_search_offset(block_group,
1802                                    offset_to_bitmap(block_group, search_start),
1803                                    1, 0);
1804         if (!entry || !entry->bitmap)
1805                 goto out;
1806
1807         search_start = min_start;
1808         search_bytes = bytes;
1809
1810         err = search_bitmap(block_group, entry, &search_start,
1811                             &search_bytes);
1812         if (err)
1813                 goto out;
1814
1815         ret = search_start;
1816         bitmap_clear_bits(block_group, entry, ret, bytes);
1817         if (entry->bytes == 0)
1818                 free_bitmap(block_group, entry);
1819 out:
1820         spin_unlock(&cluster->lock);
1821         spin_unlock(&block_group->tree_lock);
1822
1823         return ret;
1824 }
1825
1826 /*
1827  * given a cluster, try to allocate 'bytes' from it, returns 0
1828  * if it couldn't find anything suitably large, or a logical disk offset
1829  * if things worked out
1830  */
1831 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
1832                              struct btrfs_free_cluster *cluster, u64 bytes,
1833                              u64 min_start)
1834 {
1835         struct btrfs_free_space *entry = NULL;
1836         struct rb_node *node;
1837         u64 ret = 0;
1838
1839         if (cluster->points_to_bitmap)
1840                 return btrfs_alloc_from_bitmap(block_group, cluster, bytes,
1841                                                min_start);
1842
1843         spin_lock(&cluster->lock);
1844         if (bytes > cluster->max_size)
1845                 goto out;
1846
1847         if (cluster->block_group != block_group)
1848                 goto out;
1849
1850         node = rb_first(&cluster->root);
1851         if (!node)
1852                 goto out;
1853
1854         entry = rb_entry(node, struct btrfs_free_space, offset_index);
1855
1856         while(1) {
1857                 if (entry->bytes < bytes || entry->offset < min_start) {
1858                         struct rb_node *node;
1859
1860                         node = rb_next(&entry->offset_index);
1861                         if (!node)
1862                                 break;
1863                         entry = rb_entry(node, struct btrfs_free_space,
1864                                          offset_index);
1865                         continue;
1866                 }
1867                 ret = entry->offset;
1868
1869                 entry->offset += bytes;
1870                 entry->bytes -= bytes;
1871
1872                 if (entry->bytes == 0)
1873                         rb_erase(&entry->offset_index, &cluster->root);
1874                 break;
1875         }
1876 out:
1877         spin_unlock(&cluster->lock);
1878
1879         if (!ret)
1880                 return 0;
1881
1882         spin_lock(&block_group->tree_lock);
1883
1884         block_group->free_space -= bytes;
1885         if (entry->bytes == 0) {
1886                 block_group->free_extents--;
1887                 kfree(entry);
1888         }
1889
1890         spin_unlock(&block_group->tree_lock);
1891
1892         return ret;
1893 }
1894
1895 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
1896                                 struct btrfs_free_space *entry,
1897                                 struct btrfs_free_cluster *cluster,
1898                                 u64 offset, u64 bytes, u64 min_bytes)
1899 {
1900         unsigned long next_zero;
1901         unsigned long i;
1902         unsigned long search_bits;
1903         unsigned long total_bits;
1904         unsigned long found_bits;
1905         unsigned long start = 0;
1906         unsigned long total_found = 0;
1907         bool found = false;
1908
1909         i = offset_to_bit(entry->offset, block_group->sectorsize,
1910                           max_t(u64, offset, entry->offset));
1911         search_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
1912         total_bits = bytes_to_bits(bytes, block_group->sectorsize);
1913
1914 again:
1915         found_bits = 0;
1916         for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
1917              i < BITS_PER_BITMAP;
1918              i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
1919                 next_zero = find_next_zero_bit(entry->bitmap,
1920                                                BITS_PER_BITMAP, i);
1921                 if (next_zero - i >= search_bits) {
1922                         found_bits = next_zero - i;
1923                         break;
1924                 }
1925                 i = next_zero;
1926         }
1927
1928         if (!found_bits)
1929                 return -1;
1930
1931         if (!found) {
1932                 start = i;
1933                 found = true;
1934         }
1935
1936         total_found += found_bits;
1937
1938         if (cluster->max_size < found_bits * block_group->sectorsize)
1939                 cluster->max_size = found_bits * block_group->sectorsize;
1940
1941         if (total_found < total_bits) {
1942                 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
1943                 if (i - start > total_bits * 2) {
1944                         total_found = 0;
1945                         cluster->max_size = 0;
1946                         found = false;
1947                 }
1948                 goto again;
1949         }
1950
1951         cluster->window_start = start * block_group->sectorsize +
1952                 entry->offset;
1953         cluster->points_to_bitmap = true;
1954
1955         return 0;
1956 }
1957
1958 /*
1959  * here we try to find a cluster of blocks in a block group.  The goal
1960  * is to find at least bytes free and up to empty_size + bytes free.
1961  * We might not find them all in one contiguous area.
1962  *
1963  * returns zero and sets up cluster if things worked out, otherwise
1964  * it returns -enospc
1965  */
1966 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
1967                              struct btrfs_root *root,
1968                              struct btrfs_block_group_cache *block_group,
1969                              struct btrfs_free_cluster *cluster,
1970                              u64 offset, u64 bytes, u64 empty_size)
1971 {
1972         struct btrfs_free_space *entry = NULL;
1973         struct rb_node *node;
1974         struct btrfs_free_space *next;
1975         struct btrfs_free_space *last = NULL;
1976         u64 min_bytes;
1977         u64 window_start;
1978         u64 window_free;
1979         u64 max_extent = 0;
1980         bool found_bitmap = false;
1981         int ret;
1982
1983         /* for metadata, allow allocates with more holes */
1984         if (btrfs_test_opt(root, SSD_SPREAD)) {
1985                 min_bytes = bytes + empty_size;
1986         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
1987                 /*
1988                  * we want to do larger allocations when we are
1989                  * flushing out the delayed refs, it helps prevent
1990                  * making more work as we go along.
1991                  */
1992                 if (trans->transaction->delayed_refs.flushing)
1993                         min_bytes = max(bytes, (bytes + empty_size) >> 1);
1994                 else
1995                         min_bytes = max(bytes, (bytes + empty_size) >> 4);
1996         } else
1997                 min_bytes = max(bytes, (bytes + empty_size) >> 2);
1998
1999         spin_lock(&block_group->tree_lock);
2000         spin_lock(&cluster->lock);
2001
2002         /* someone already found a cluster, hooray */
2003         if (cluster->block_group) {
2004                 ret = 0;
2005                 goto out;
2006         }
2007 again:
2008         entry = tree_search_offset(block_group, offset, found_bitmap, 1);
2009         if (!entry) {
2010                 ret = -ENOSPC;
2011                 goto out;
2012         }
2013
2014         /*
2015          * If found_bitmap is true, we exhausted our search for extent entries,
2016          * and we just want to search all of the bitmaps that we can find, and
2017          * ignore any extent entries we find.
2018          */
2019         while (entry->bitmap || found_bitmap ||
2020                (!entry->bitmap && entry->bytes < min_bytes)) {
2021                 struct rb_node *node = rb_next(&entry->offset_index);
2022
2023                 if (entry->bitmap && entry->bytes > bytes + empty_size) {
2024                         ret = btrfs_bitmap_cluster(block_group, entry, cluster,
2025                                                    offset, bytes + empty_size,
2026                                                    min_bytes);
2027                         if (!ret)
2028                                 goto got_it;
2029                 }
2030
2031                 if (!node) {
2032                         ret = -ENOSPC;
2033                         goto out;
2034                 }
2035                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2036         }
2037
2038         /*
2039          * We already searched all the extent entries from the passed in offset
2040          * to the end and didn't find enough space for the cluster, and we also
2041          * didn't find any bitmaps that met our criteria, just go ahead and exit
2042          */
2043         if (found_bitmap) {
2044                 ret = -ENOSPC;
2045                 goto out;
2046         }
2047
2048         cluster->points_to_bitmap = false;
2049         window_start = entry->offset;
2050         window_free = entry->bytes;
2051         last = entry;
2052         max_extent = entry->bytes;
2053
2054         while (1) {
2055                 /* out window is just right, lets fill it */
2056                 if (window_free >= bytes + empty_size)
2057                         break;
2058
2059                 node = rb_next(&last->offset_index);
2060                 if (!node) {
2061                         if (found_bitmap)
2062                                 goto again;
2063                         ret = -ENOSPC;
2064                         goto out;
2065                 }
2066                 next = rb_entry(node, struct btrfs_free_space, offset_index);
2067
2068                 /*
2069                  * we found a bitmap, so if this search doesn't result in a
2070                  * cluster, we know to go and search again for the bitmaps and
2071                  * start looking for space there
2072                  */
2073                 if (next->bitmap) {
2074                         if (!found_bitmap)
2075                                 offset = next->offset;
2076                         found_bitmap = true;
2077                         last = next;
2078                         continue;
2079                 }
2080
2081                 /*
2082                  * we haven't filled the empty size and the window is
2083                  * very large.  reset and try again
2084                  */
2085                 if (next->offset - (last->offset + last->bytes) > 128 * 1024 ||
2086                     next->offset - window_start > (bytes + empty_size) * 2) {
2087                         entry = next;
2088                         window_start = entry->offset;
2089                         window_free = entry->bytes;
2090                         last = entry;
2091                         max_extent = entry->bytes;
2092                 } else {
2093                         last = next;
2094                         window_free += next->bytes;
2095                         if (entry->bytes > max_extent)
2096                                 max_extent = entry->bytes;
2097                 }
2098         }
2099
2100         cluster->window_start = entry->offset;
2101
2102         /*
2103          * now we've found our entries, pull them out of the free space
2104          * cache and put them into the cluster rbtree
2105          *
2106          * The cluster includes an rbtree, but only uses the offset index
2107          * of each free space cache entry.
2108          */
2109         while (1) {
2110                 node = rb_next(&entry->offset_index);
2111                 if (entry->bitmap && node) {
2112                         entry = rb_entry(node, struct btrfs_free_space,
2113                                          offset_index);
2114                         continue;
2115                 } else if (entry->bitmap && !node) {
2116                         break;
2117                 }
2118
2119                 rb_erase(&entry->offset_index, &block_group->free_space_offset);
2120                 ret = tree_insert_offset(&cluster->root, entry->offset,
2121                                          &entry->offset_index, 0);
2122                 BUG_ON(ret);
2123
2124                 if (!node || entry == last)
2125                         break;
2126
2127                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2128         }
2129
2130         cluster->max_size = max_extent;
2131 got_it:
2132         ret = 0;
2133         atomic_inc(&block_group->count);
2134         list_add_tail(&cluster->block_group_list, &block_group->cluster_list);
2135         cluster->block_group = block_group;
2136 out:
2137         spin_unlock(&cluster->lock);
2138         spin_unlock(&block_group->tree_lock);
2139
2140         return ret;
2141 }
2142
2143 /*
2144  * simple code to zero out a cluster
2145  */
2146 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2147 {
2148         spin_lock_init(&cluster->lock);
2149         spin_lock_init(&cluster->refill_lock);
2150         cluster->root = RB_ROOT;
2151         cluster->max_size = 0;
2152         cluster->points_to_bitmap = false;
2153         INIT_LIST_HEAD(&cluster->block_group_list);
2154         cluster->block_group = NULL;
2155 }
2156