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 #include "extent_io.h"
  28
  29 #define BITS_PER_BITMAP         (PAGE_CACHE_SIZE * 8)
  30 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
  31
  32 static void recalculate_thresholds(struct btrfs_block_group_cache
  33                                    *block_group);
  34 static int link_free_space(struct btrfs_block_group_cache *block_group,
  35                            struct btrfs_free_space *info);
  36
  37 struct inode *lookup_free_space_inode(struct btrfs_root *root,
  38                                       struct btrfs_block_group_cache
  39                                       *block_group, struct btrfs_path *path)
  40 {
  41         struct btrfs_key key;
  42         struct btrfs_key location;
  43         struct btrfs_disk_key disk_key;
  44         struct btrfs_free_space_header *header;
  45         struct extent_buffer *leaf;
  46         struct inode *inode = NULL;
  47         int ret;
  48
  49         spin_lock(&block_group->lock);
  50         if (block_group->inode)
  51                 inode = igrab(block_group->inode);
  52         spin_unlock(&block_group->lock);
  53         if (inode)
  54                 return inode;
  55
  56         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  57         key.offset = block_group->key.objectid;
  58         key.type = 0;
  59
  60         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  61         if (ret < 0)
  62                 return ERR_PTR(ret);
  63         if (ret > 0) {
  64                 btrfs_release_path(root, path);
  65                 return ERR_PTR(-ENOENT);
  66         }
  67
  68         leaf = path->nodes[0];
  69         header = btrfs_item_ptr(leaf, path->slots[0],
  70                                 struct btrfs_free_space_header);
  71         btrfs_free_space_key(leaf, header, &disk_key);
  72         btrfs_disk_key_to_cpu(&location, &disk_key);
  73         btrfs_release_path(root, path);
  74
  75         inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
  76         if (!inode)
  77                 return ERR_PTR(-ENOENT);
  78         if (IS_ERR(inode))
  79                 return inode;
  80         if (is_bad_inode(inode)) {
  81                 iput(inode);
  82                 return ERR_PTR(-ENOENT);
  83         }
  84
  85         inode->i_mapping->flags &= ~__GFP_FS;
  86
  87         spin_lock(&block_group->lock);
  88         if (!root->fs_info->closing) {
  89                 block_group->inode = igrab(inode);
  90                 block_group->iref = 1;
  91         }
  92         spin_unlock(&block_group->lock);
  93
  94         return inode;
  95 }
  96
  97 int create_free_space_inode(struct btrfs_root *root,
  98                             struct btrfs_trans_handle *trans,
  99                             struct btrfs_block_group_cache *block_group,
 100                             struct btrfs_path *path)
 101 {
 102         struct btrfs_key key;
 103         struct btrfs_disk_key disk_key;
 104         struct btrfs_free_space_header *header;
 105         struct btrfs_inode_item *inode_item;
 106         struct extent_buffer *leaf;
 107         u64 objectid;
 108         int ret;
 109
 110         ret = btrfs_find_free_objectid(trans, root, 0, &objectid);
 111         if (ret < 0)
 112                 return ret;
 113
 114         ret = btrfs_insert_empty_inode(trans, root, path, objectid);
 115         if (ret)
 116                 return ret;
 117
 118         leaf = path->nodes[0];
 119         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 120                                     struct btrfs_inode_item);
 121         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 122         memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
 123                              sizeof(*inode_item));
 124         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 125         btrfs_set_inode_size(leaf, inode_item, 0);
 126         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 127         btrfs_set_inode_uid(leaf, inode_item, 0);
 128         btrfs_set_inode_gid(leaf, inode_item, 0);
 129         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 130         btrfs_set_inode_flags(leaf, inode_item, BTRFS_INODE_NOCOMPRESS |
 131                               BTRFS_INODE_PREALLOC | BTRFS_INODE_NODATASUM);
 132         btrfs_set_inode_nlink(leaf, inode_item, 1);
 133         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 134         btrfs_set_inode_block_group(leaf, inode_item,
 135                                     block_group->key.objectid);
 136         btrfs_mark_buffer_dirty(leaf);
 137         btrfs_release_path(root, path);
 138
 139         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 140         key.offset = block_group->key.objectid;
 141         key.type = 0;
 142
 143         ret = btrfs_insert_empty_item(trans, root, path, &key,
 144                                       sizeof(struct btrfs_free_space_header));
 145         if (ret < 0) {
 146                 btrfs_release_path(root, path);
 147                 return ret;
 148         }
 149         leaf = path->nodes[0];
 150         header = btrfs_item_ptr(leaf, path->slots[0],
 151                                 struct btrfs_free_space_header);
 152         memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
 153         btrfs_set_free_space_key(leaf, header, &disk_key);
 154         btrfs_mark_buffer_dirty(leaf);
 155         btrfs_release_path(root, path);
 156
 157         return 0;
 158 }
 159
 160 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
 161                                     struct btrfs_trans_handle *trans,
 162                                     struct btrfs_path *path,
 163                                     struct inode *inode)
 164 {
 165         loff_t oldsize;
 166         int ret = 0;
 167
 168         trans->block_rsv = root->orphan_block_rsv;
 169         ret = btrfs_block_rsv_check(trans, root,
 170                                     root->orphan_block_rsv,
 171                                     0, 5);
 172         if (ret)
 173                 return ret;
 174
 175         oldsize = i_size_read(inode);
 176         btrfs_i_size_write(inode, 0);
 177         truncate_pagecache(inode, oldsize, 0);
 178
 179         /*
 180          * We don't need an orphan item because truncating the free space cache
 181          * will never be split across transactions.
 182          */
 183         ret = btrfs_truncate_inode_items(trans, root, inode,
 184                                          0, BTRFS_EXTENT_DATA_KEY);
 185         if (ret) {
 186                 WARN_ON(1);
 187                 return ret;
 188         }
 189
 190         return btrfs_update_inode(trans, root, inode);
 191 }
 192
 193 static int readahead_cache(struct inode *inode)
 194 {
 195         struct file_ra_state *ra;
 196         unsigned long last_index;
 197
 198         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 199         if (!ra)
 200                 return -ENOMEM;
 201
 202         file_ra_state_init(ra, inode->i_mapping);
 203         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 204
 205         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 206
 207         kfree(ra);
 208
 209         return 0;
 210 }
 211
 212 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 213                           struct btrfs_block_group_cache *block_group)
 214 {
 215         struct btrfs_root *root = fs_info->tree_root;
 216         struct inode *inode;
 217         struct btrfs_free_space_header *header;
 218         struct extent_buffer *leaf;
 219         struct page *page;
 220         struct btrfs_path *path;
 221         u32 *checksums = NULL, *crc;
 222         char *disk_crcs = NULL;
 223         struct btrfs_key key;
 224         struct list_head bitmaps;
 225         u64 num_entries;
 226         u64 num_bitmaps;
 227         u64 generation;
 228         u64 used = btrfs_block_group_used(&block_group->item);
 229         u32 cur_crc = ~(u32)0;
 230         pgoff_t index = 0;
 231         unsigned long first_page_offset;
 232         int num_checksums;
 233         int ret = 0;
 234
 235         /*
 236          * If we're unmounting then just return, since this does a search on the
 237          * normal root and not the commit root and we could deadlock.
 238          */
 239         smp_mb();
 240         if (fs_info->closing)
 241                 return 0;
 242
 243         /*
 244          * If this block group has been marked to be cleared for one reason or
 245          * another then we can't trust the on disk cache, so just return.
 246          */
 247         spin_lock(&block_group->lock);
 248         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 249                 spin_unlock(&block_group->lock);
 250                 return 0;
 251         }
 252         spin_unlock(&block_group->lock);
 253
 254         INIT_LIST_HEAD(&bitmaps);
 255
 256         path = btrfs_alloc_path();
 257         if (!path)
 258                 return 0;
 259
 260         inode = lookup_free_space_inode(root, block_group, path);
 261         if (IS_ERR(inode)) {
 262                 btrfs_free_path(path);
 263                 return 0;
 264         }
 265
 266         /* Nothing in the space cache, goodbye */
 267         if (!i_size_read(inode)) {
 268                 btrfs_free_path(path);
 269                 goto out;
 270         }
 271
 272         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 273         key.offset = block_group->key.objectid;
 274         key.type = 0;
 275
 276         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 277         if (ret) {
 278                 btrfs_free_path(path);
 279                 goto out;
 280         }
 281
 282         leaf = path->nodes[0];
 283         header = btrfs_item_ptr(leaf, path->slots[0],
 284                                 struct btrfs_free_space_header);
 285         num_entries = btrfs_free_space_entries(leaf, header);
 286         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 287         generation = btrfs_free_space_generation(leaf, header);
 288         btrfs_free_path(path);
 289
 290         if (BTRFS_I(inode)->generation != generation) {
 291                 printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
 292                        " not match free space cache generation (%llu) for "
 293                        "block group %llu\n",
 294                        (unsigned long long)BTRFS_I(inode)->generation,
 295                        (unsigned long long)generation,
 296                        (unsigned long long)block_group->key.objectid);
 297                 goto free_cache;
 298         }
 299
 300         if (!num_entries)
 301                 goto out;
 302
 303         /* Setup everything for doing checksumming */
 304         num_checksums = i_size_read(inode) / PAGE_CACHE_SIZE;
 305         checksums = crc = kzalloc(sizeof(u32) * num_checksums, GFP_NOFS);
 306         if (!checksums)
 307                 goto out;
 308         first_page_offset = (sizeof(u32) * num_checksums) + sizeof(u64);
 309         disk_crcs = kzalloc(first_page_offset, GFP_NOFS);
 310         if (!disk_crcs)
 311                 goto out;
 312
 313         ret = readahead_cache(inode);
 314         if (ret) {
 315                 ret = 0;
 316                 goto out;
 317         }
 318
 319         while (1) {
 320                 struct btrfs_free_space_entry *entry;
 321                 struct btrfs_free_space *e;
 322                 void *addr;
 323                 unsigned long offset = 0;
 324                 unsigned long start_offset = 0;
 325                 int need_loop = 0;
 326
 327                 if (!num_entries && !num_bitmaps)
 328                         break;
 329
 330                 if (index == 0) {
 331                         start_offset = first_page_offset;
 332                         offset = start_offset;
 333                 }
 334
 335                 page = grab_cache_page(inode->i_mapping, index);
 336                 if (!page) {
 337                         ret = 0;
 338                         goto free_cache;
 339                 }
 340
 341                 if (!PageUptodate(page)) {
 342                         btrfs_readpage(NULL, page);
 343                         lock_page(page);
 344                         if (!PageUptodate(page)) {
 345                                 unlock_page(page);
 346                                 page_cache_release(page);
 347                                 printk(KERN_ERR "btrfs: error reading free "
 348                                        "space cache: %llu\n",
 349                                        (unsigned long long)
 350                                        block_group->key.objectid);
 351                                 goto free_cache;
 352                         }
 353                 }
 354                 addr = kmap(page);
 355
 356                 if (index == 0) {
 357                         u64 *gen;
 358
 359                         memcpy(disk_crcs, addr, first_page_offset);
 360                         gen = addr + (sizeof(u32) * num_checksums);
 361                         if (*gen != BTRFS_I(inode)->generation) {
 362                                 printk(KERN_ERR "btrfs: space cache generation"
 363                                        " (%llu) does not match inode (%llu) "
 364                                        "for block group %llu\n",
 365                                        (unsigned long long)*gen,
 366                                        (unsigned long long)
 367                                        BTRFS_I(inode)->generation,
 368                                        (unsigned long long)
 369                                        block_group->key.objectid);
 370                                 kunmap(page);
 371                                 unlock_page(page);
 372                                 page_cache_release(page);
 373                                 goto free_cache;
 374                         }
 375                         crc = (u32 *)disk_crcs;
 376                 }
 377                 entry = addr + start_offset;
 378
 379                 /* First lets check our crc before we do anything fun */
 380                 cur_crc = ~(u32)0;
 381                 cur_crc = btrfs_csum_data(root, addr + start_offset, cur_crc,
 382                                           PAGE_CACHE_SIZE - start_offset);
 383                 btrfs_csum_final(cur_crc, (char *)&cur_crc);
 384                 if (cur_crc != *crc) {
 385                         printk(KERN_ERR "btrfs: crc mismatch for page %lu in "
 386                                "block group %llu\n", index,
 387                                (unsigned long long)block_group->key.objectid);
 388                         kunmap(page);
 389                         unlock_page(page);
 390                         page_cache_release(page);
 391                         goto free_cache;
 392                 }
 393                 crc++;
 394
 395                 while (1) {
 396                         if (!num_entries)
 397                                 break;
 398
 399                         need_loop = 1;
 400                         e = kmem_cache_zalloc(btrfs_free_space_cachep,
 401                                               GFP_NOFS);
 402                         if (!e) {
 403                                 kunmap(page);
 404                                 unlock_page(page);
 405                                 page_cache_release(page);
 406                                 goto free_cache;
 407                         }
 408
 409                         e->offset = le64_to_cpu(entry->offset);
 410                         e->bytes = le64_to_cpu(entry->bytes);
 411                         if (!e->bytes) {
 412                                 kunmap(page);
 413                                 kmem_cache_free(btrfs_free_space_cachep, e);
 414                                 unlock_page(page);
 415                                 page_cache_release(page);
 416                                 goto free_cache;
 417                         }
 418
 419                         if (entry->type == BTRFS_FREE_SPACE_EXTENT) {
 420                                 spin_lock(&block_group->tree_lock);
 421                                 ret = link_free_space(block_group, e);
 422                                 spin_unlock(&block_group->tree_lock);
 423                                 BUG_ON(ret);
 424                         } else {
 425                                 e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
 426                                 if (!e->bitmap) {
 427                                         kunmap(page);
 428                                         kmem_cache_free(
 429                                                 btrfs_free_space_cachep, e);
 430                                         unlock_page(page);
 431                                         page_cache_release(page);
 432                                         goto free_cache;
 433                                 }
 434                                 spin_lock(&block_group->tree_lock);
 435                                 ret = link_free_space(block_group, e);
 436                                 block_group->total_bitmaps++;
 437                                 recalculate_thresholds(block_group);
 438                                 spin_unlock(&block_group->tree_lock);
 439                                 list_add_tail(&e->list, &bitmaps);
 440                         }
 441
 442                         num_entries--;
 443                         offset += sizeof(struct btrfs_free_space_entry);
 444                         if (offset + sizeof(struct btrfs_free_space_entry) >=
 445                             PAGE_CACHE_SIZE)
 446                                 break;
 447                         entry++;
 448                 }
 449
 450                 /*
 451                  * We read an entry out of this page, we need to move on to the
 452                  * next page.
 453                  */
 454                 if (need_loop) {
 455                         kunmap(page);
 456                         goto next;
 457                 }
 458
 459                 /*
 460                  * We add the bitmaps at the end of the entries in order that
 461                  * the bitmap entries are added to the cache.
 462                  */
 463                 e = list_entry(bitmaps.next, struct btrfs_free_space, list);
 464                 list_del_init(&e->list);
 465                 memcpy(e->bitmap, addr, PAGE_CACHE_SIZE);
 466                 kunmap(page);
 467                 num_bitmaps--;
 468 next:
 469                 unlock_page(page);
 470                 page_cache_release(page);
 471                 index++;
 472         }
 473
 474         spin_lock(&block_group->tree_lock);
 475         if (block_group->free_space != (block_group->key.offset - used -
 476                                         block_group->bytes_super)) {
 477                 spin_unlock(&block_group->tree_lock);
 478                 printk(KERN_ERR "block group %llu has an wrong amount of free "
 479                        "space\n", block_group->key.objectid);
 480                 ret = 0;
 481                 goto free_cache;
 482         }
 483         spin_unlock(&block_group->tree_lock);
 484
 485         ret = 1;
 486 out:
 487         kfree(checksums);
 488         kfree(disk_crcs);
 489         iput(inode);
 490         return ret;
 491
 492 free_cache:
 493         /* This cache is bogus, make sure it gets cleared */
 494         spin_lock(&block_group->lock);
 495         block_group->disk_cache_state = BTRFS_DC_CLEAR;
 496         spin_unlock(&block_group->lock);
 497         btrfs_remove_free_space_cache(block_group);
 498         goto out;
 499 }
 500
 501 int btrfs_write_out_cache(struct btrfs_root *root,
 502                           struct btrfs_trans_handle *trans,
 503                           struct btrfs_block_group_cache *block_group,
 504                           struct btrfs_path *path)
 505 {
 506         struct btrfs_free_space_header *header;
 507         struct extent_buffer *leaf;
 508         struct inode *inode;
 509         struct rb_node *node;
 510         struct list_head *pos, *n;
 511         struct page **pages;
 512         struct page *page;
 513         struct extent_state *cached_state = NULL;
 514         struct btrfs_free_cluster *cluster = NULL;
 515         struct extent_io_tree *unpin = NULL;
 516         struct list_head bitmap_list;
 517         struct btrfs_key key;
 518         u64 start, end, len;
 519         u64 bytes = 0;
 520         u32 *crc, *checksums;
 521         unsigned long first_page_offset;
 522         int index = 0, num_pages = 0;
 523         int entries = 0;
 524         int bitmaps = 0;
 525         int ret = 0;
 526         bool next_page = false;
 527         bool out_of_space = false;
 528
 529         root = root->fs_info->tree_root;
 530
 531         INIT_LIST_HEAD(&bitmap_list);
 532
 533         spin_lock(&block_group->lock);
 534         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
 535                 spin_unlock(&block_group->lock);
 536                 return 0;
 537         }
 538         spin_unlock(&block_group->lock);
 539
 540         inode = lookup_free_space_inode(root, block_group, path);
 541         if (IS_ERR(inode))
 542                 return 0;
 543
 544         if (!i_size_read(inode)) {
 545                 iput(inode);
 546                 return 0;
 547         }
 548
 549         node = rb_first(&block_group->free_space_offset);
 550         if (!node) {
 551                 iput(inode);
 552                 return 0;
 553         }
 554
 555         num_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
 556                 PAGE_CACHE_SHIFT;
 557         filemap_write_and_wait(inode->i_mapping);
 558         btrfs_wait_ordered_range(inode, inode->i_size &
 559                                  ~(root->sectorsize - 1), (u64)-1);
 560
 561         /* We need a checksum per page. */
 562         crc = checksums = kzalloc(sizeof(u32) * num_pages, GFP_NOFS);
 563         if (!crc) {
 564                 iput(inode);
 565                 return 0;
 566         }
 567
 568         pages = kzalloc(sizeof(struct page *) * num_pages, GFP_NOFS);
 569         if (!pages) {
 570                 kfree(crc);
 571                 iput(inode);
 572                 return 0;
 573         }
 574
 575         /* Since the first page has all of our checksums and our generation we
 576          * need to calculate the offset into the page that we can start writing
 577          * our entries.
 578          */
 579         first_page_offset = (sizeof(u32) * num_pages) + sizeof(u64);
 580
 581         /* Get the cluster for this block_group if it exists */
 582         if (!list_empty(&block_group->cluster_list))
 583                 cluster = list_entry(block_group->cluster_list.next,
 584                                      struct btrfs_free_cluster,
 585                                      block_group_list);
 586
 587         /*
 588          * We shouldn't have switched the pinned extents yet so this is the
 589          * right one
 590          */
 591         unpin = root->fs_info->pinned_extents;
 592
 593         /*
 594          * Lock all pages first so we can lock the extent safely.
 595          *
 596          * NOTE: Because we hold the ref the entire time we're going to write to
 597          * the page find_get_page should never fail, so we don't do a check
 598          * after find_get_page at this point.  Just putting this here so people
 599          * know and don't freak out.
 600          */
 601         while (index < num_pages) {
 602                 page = grab_cache_page(inode->i_mapping, index);
 603                 if (!page) {
 604                         int i;
 605
 606                         for (i = 0; i < num_pages; i++) {
 607                                 unlock_page(pages[i]);
 608                                 page_cache_release(pages[i]);
 609                         }
 610                         goto out_free;
 611                 }
 612                 pages[index] = page;
 613                 index++;
 614         }
 615
 616         index = 0;
 617         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
 618                          0, &cached_state, GFP_NOFS);
 619
 620         /*
 621          * When searching for pinned extents, we need to start at our start
 622          * offset.
 623          */
 624         start = block_group->key.objectid;
 625
 626         /* Write out the extent entries */
 627         do {
 628                 struct btrfs_free_space_entry *entry;
 629                 void *addr;
 630                 unsigned long offset = 0;
 631                 unsigned long start_offset = 0;
 632
 633                 next_page = false;
 634
 635                 if (index == 0) {
 636                         start_offset = first_page_offset;
 637                         offset = start_offset;
 638                 }
 639
 640                 if (index >= num_pages) {
 641                         out_of_space = true;
 642                         break;
 643                 }
 644
 645                 page = pages[index];
 646
 647                 addr = kmap(page);
 648                 entry = addr + start_offset;
 649
 650                 memset(addr, 0, PAGE_CACHE_SIZE);
 651                 while (node && !next_page) {
 652                         struct btrfs_free_space *e;
 653
 654                         e = rb_entry(node, struct btrfs_free_space, offset_index);
 655                         entries++;
 656
 657                         entry->offset = cpu_to_le64(e->offset);
 658                         entry->bytes = cpu_to_le64(e->bytes);
 659                         if (e->bitmap) {
 660                                 entry->type = BTRFS_FREE_SPACE_BITMAP;
 661                                 list_add_tail(&e->list, &bitmap_list);
 662                                 bitmaps++;
 663                         } else {
 664                                 entry->type = BTRFS_FREE_SPACE_EXTENT;
 665                         }
 666                         node = rb_next(node);
 667                         if (!node && cluster) {
 668                                 node = rb_first(&cluster->root);
 669                                 cluster = NULL;
 670                         }
 671                         offset += sizeof(struct btrfs_free_space_entry);
 672                         if (offset + sizeof(struct btrfs_free_space_entry) >=
 673                             PAGE_CACHE_SIZE)
 674                                 next_page = true;
 675                         entry++;
 676                 }
 677
 678                 /*
 679                  * We want to add any pinned extents to our free space cache
 680                  * so we don't leak the space
 681                  */
 682                 while (!next_page && (start < block_group->key.objectid +
 683                                       block_group->key.offset)) {
 684                         ret = find_first_extent_bit(unpin, start, &start, &end,
 685                                                     EXTENT_DIRTY);
 686                         if (ret) {
 687                                 ret = 0;
 688                                 break;
 689                         }
 690
 691                         /* This pinned extent is out of our range */
 692                         if (start >= block_group->key.objectid +
 693                             block_group->key.offset)
 694                                 break;
 695
 696                         len = block_group->key.objectid +
 697                                 block_group->key.offset - start;
 698                         len = min(len, end + 1 - start);
 699
 700                         entries++;
 701                         entry->offset = cpu_to_le64(start);
 702                         entry->bytes = cpu_to_le64(len);
 703                         entry->type = BTRFS_FREE_SPACE_EXTENT;
 704
 705                         start = end + 1;
 706                         offset += sizeof(struct btrfs_free_space_entry);
 707                         if (offset + sizeof(struct btrfs_free_space_entry) >=
 708                             PAGE_CACHE_SIZE)
 709                                 next_page = true;
 710                         entry++;
 711                 }
 712                 *crc = ~(u32)0;
 713                 *crc = btrfs_csum_data(root, addr + start_offset, *crc,
 714                                        PAGE_CACHE_SIZE - start_offset);
 715                 kunmap(page);
 716
 717                 btrfs_csum_final(*crc, (char *)crc);
 718                 crc++;
 719
 720                 bytes += PAGE_CACHE_SIZE;
 721
 722                 index++;
 723         } while (node || next_page);
 724
 725         /* Write out the bitmaps */
 726         list_for_each_safe(pos, n, &bitmap_list) {
 727                 void *addr;
 728                 struct btrfs_free_space *entry =
 729                         list_entry(pos, struct btrfs_free_space, list);
 730
 731                 if (index >= num_pages) {
 732                         out_of_space = true;
 733                         break;
 734                 }
 735                 page = pages[index];
 736
 737                 addr = kmap(page);
 738                 memcpy(addr, entry->bitmap, PAGE_CACHE_SIZE);
 739                 *crc = ~(u32)0;
 740                 *crc = btrfs_csum_data(root, addr, *crc, PAGE_CACHE_SIZE);
 741                 kunmap(page);
 742                 btrfs_csum_final(*crc, (char *)crc);
 743                 crc++;
 744                 bytes += PAGE_CACHE_SIZE;
 745
 746                 list_del_init(&entry->list);
 747                 index++;
 748         }
 749
 750         if (out_of_space) {
 751                 btrfs_drop_pages(pages, num_pages);
 752                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
 753                                      i_size_read(inode) - 1, &cached_state,
 754                                      GFP_NOFS);
 755                 ret = 0;
 756                 goto out_free;
 757         }
 758
 759         /* Zero out the rest of the pages just to make sure */
 760         while (index < num_pages) {
 761                 void *addr;
 762
 763                 page = pages[index];
 764                 addr = kmap(page);
 765                 memset(addr, 0, PAGE_CACHE_SIZE);
 766                 kunmap(page);
 767                 bytes += PAGE_CACHE_SIZE;
 768                 index++;
 769         }
 770
 771         /* Write the checksums and trans id to the first page */
 772         {
 773                 void *addr;
 774                 u64 *gen;
 775
 776                 page = pages[0];
 777
 778                 addr = kmap(page);
 779                 memcpy(addr, checksums, sizeof(u32) * num_pages);
 780                 gen = addr + (sizeof(u32) * num_pages);
 781                 *gen = trans->transid;
 782                 kunmap(page);
 783         }
 784
 785         ret = btrfs_dirty_pages(root, inode, pages, num_pages, 0,
 786                                             bytes, &cached_state);
 787         btrfs_drop_pages(pages, num_pages);
 788         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
 789                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
 790
 791         if (ret) {
 792                 ret = 0;
 793                 goto out_free;
 794         }
 795
 796         BTRFS_I(inode)->generation = trans->transid;
 797
 798         filemap_write_and_wait(inode->i_mapping);
 799
 800         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 801         key.offset = block_group->key.objectid;
 802         key.type = 0;
 803
 804         ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
 805         if (ret < 0) {
 806                 ret = 0;
 807                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
 808                                  EXTENT_DIRTY | EXTENT_DELALLOC |
 809                                  EXTENT_DO_ACCOUNTING, 0, 0, NULL, GFP_NOFS);
 810                 goto out_free;
 811         }
 812         leaf = path->nodes[0];
 813         if (ret > 0) {
 814                 struct btrfs_key found_key;
 815                 BUG_ON(!path->slots[0]);
 816                 path->slots[0]--;
 817                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
 818                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
 819                     found_key.offset != block_group->key.objectid) {
 820                         ret = 0;
 821                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, bytes - 1,
 822                                          EXTENT_DIRTY | EXTENT_DELALLOC |
 823                                          EXTENT_DO_ACCOUNTING, 0, 0, NULL,
 824                                          GFP_NOFS);
 825                         btrfs_release_path(root, path);
 826                         goto out_free;
 827                 }
 828         }
 829         header = btrfs_item_ptr(leaf, path->slots[0],
 830                                 struct btrfs_free_space_header);
 831         btrfs_set_free_space_entries(leaf, header, entries);
 832         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
 833         btrfs_set_free_space_generation(leaf, header, trans->transid);
 834         btrfs_mark_buffer_dirty(leaf);
 835         btrfs_release_path(root, path);
 836
 837         ret = 1;
 838
 839 out_free:
 840         if (ret == 0) {
 841                 invalidate_inode_pages2_range(inode->i_mapping, 0, index);
 842                 spin_lock(&block_group->lock);
 843                 block_group->disk_cache_state = BTRFS_DC_ERROR;
 844                 spin_unlock(&block_group->lock);
 845                 BTRFS_I(inode)->generation = 0;
 846         }
 847         kfree(checksums);
 848         kfree(pages);
 849         btrfs_update_inode(trans, root, inode);
 850         iput(inode);
 851         return ret;
 852 }
 853
 854 static inline unsigned long offset_to_bit(u64 bitmap_start, u64 sectorsize,
 855                                           u64 offset)
 856 {
 857         BUG_ON(offset < bitmap_start);
 858         offset -= bitmap_start;
 859         return (unsigned long)(div64_u64(offset, sectorsize));
 860 }
 861
 862 static inline unsigned long bytes_to_bits(u64 bytes, u64 sectorsize)
 863 {
 864         return (unsigned long)(div64_u64(bytes, sectorsize));
 865 }
 866
 867 static inline u64 offset_to_bitmap(struct btrfs_block_group_cache *block_group,
 868                                    u64 offset)
 869 {
 870         u64 bitmap_start;
 871         u64 bytes_per_bitmap;
 872
 873         bytes_per_bitmap = BITS_PER_BITMAP * block_group->sectorsize;
 874         bitmap_start = offset - block_group->key.objectid;
 875         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
 876         bitmap_start *= bytes_per_bitmap;
 877         bitmap_start += block_group->key.objectid;
 878
 879         return bitmap_start;
 880 }
 881
 882 static int tree_insert_offset(struct rb_root *root, u64 offset,
 883                               struct rb_node *node, int bitmap)
 884 {
 885         struct rb_node **p = &root->rb_node;
 886         struct rb_node *parent = NULL;
 887         struct btrfs_free_space *info;
 888
 889         while (*p) {
 890                 parent = *p;
 891                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
 892
 893                 if (offset < info->offset) {
 894                         p = &(*p)->rb_left;
 895                 } else if (offset > info->offset) {
 896                         p = &(*p)->rb_right;
 897                 } else {
 898                         /*
 899                          * we could have a bitmap entry and an extent entry
 900                          * share the same offset.  If this is the case, we want
 901                          * the extent entry to always be found first if we do a
 902                          * linear search through the tree, since we want to have
 903                          * the quickest allocation time, and allocating from an
 904                          * extent is faster than allocating from a bitmap.  So
 905                          * if we're inserting a bitmap and we find an entry at
 906                          * this offset, we want to go right, or after this entry
 907                          * logically.  If we are inserting an extent and we've
 908                          * found a bitmap, we want to go left, or before
 909                          * logically.
 910                          */
 911                         if (bitmap) {
 912                                 WARN_ON(info->bitmap);
 913                                 p = &(*p)->rb_right;
 914                         } else {
 915                                 WARN_ON(!info->bitmap);
 916                                 p = &(*p)->rb_left;
 917                         }
 918                 }
 919         }
 920
 921         rb_link_node(node, parent, p);
 922         rb_insert_color(node, root);
 923
 924         return 0;
 925 }
 926
 927 /*
 928  * searches the tree for the given offset.
 929  *
 930  * fuzzy - If this is set, then we are trying to make an allocation, and we just
 931  * want a section that has at least bytes size and comes at or after the given
 932  * offset.
 933  */
 934 static struct btrfs_free_space *
 935 tree_search_offset(struct btrfs_block_group_cache *block_group,
 936                    u64 offset, int bitmap_only, int fuzzy)
 937 {
 938         struct rb_node *n = block_group->free_space_offset.rb_node;
 939         struct btrfs_free_space *entry, *prev = NULL;
 940
 941         /* find entry that is closest to the 'offset' */
 942         while (1) {
 943                 if (!n) {
 944                         entry = NULL;
 945                         break;
 946                 }
 947
 948                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
 949                 prev = entry;
 950
 951                 if (offset < entry->offset)
 952                         n = n->rb_left;
 953                 else if (offset > entry->offset)
 954                         n = n->rb_right;
 955                 else
 956                         break;
 957         }
 958
 959         if (bitmap_only) {
 960                 if (!entry)
 961                         return NULL;
 962                 if (entry->bitmap)
 963                         return entry;
 964
 965                 /*
 966                  * bitmap entry and extent entry may share same offset,
 967                  * in that case, bitmap entry comes after extent entry.
 968                  */
 969                 n = rb_next(n);
 970                 if (!n)
 971                         return NULL;
 972                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
 973                 if (entry->offset != offset)
 974                         return NULL;
 975
 976                 WARN_ON(!entry->bitmap);
 977                 return entry;
 978         } else if (entry) {
 979                 if (entry->bitmap) {
 980                         /*
 981                          * if previous extent entry covers the offset,
 982                          * we should return it instead of the bitmap entry
 983                          */
 984                         n = &entry->offset_index;
 985                         while (1) {
 986                                 n = rb_prev(n);
 987                                 if (!n)
 988                                         break;
 989                                 prev = rb_entry(n, struct btrfs_free_space,
 990                                                 offset_index);
 991                                 if (!prev->bitmap) {
 992                                         if (prev->offset + prev->bytes > offset)
 993                                                 entry = prev;
 994                                         break;
 995                                 }
 996                         }
 997                 }
 998                 return entry;
 999         }
1000
1001         if (!prev)
1002                 return NULL;
1003
1004         /* find last entry before the 'offset' */
1005         entry = prev;
1006         if (entry->offset > offset) {
1007                 n = rb_prev(&entry->offset_index);
1008                 if (n) {
1009                         entry = rb_entry(n, struct btrfs_free_space,
1010                                         offset_index);
1011                         BUG_ON(entry->offset > offset);
1012                 } else {
1013                         if (fuzzy)
1014                                 return entry;
1015                         else
1016                                 return NULL;
1017                 }
1018         }
1019
1020         if (entry->bitmap) {
1021                 n = &entry->offset_index;
1022                 while (1) {
1023                         n = rb_prev(n);
1024                         if (!n)
1025                                 break;
1026                         prev = rb_entry(n, struct btrfs_free_space,
1027                                         offset_index);
1028                         if (!prev->bitmap) {
1029                                 if (prev->offset + prev->bytes > offset)
1030                                         return prev;
1031                                 break;
1032                         }
1033                 }
1034                 if (entry->offset + BITS_PER_BITMAP *
1035                     block_group->sectorsize > offset)
1036                         return entry;
1037         } else if (entry->offset + entry->bytes > offset)
1038                 return entry;
1039
1040         if (!fuzzy)
1041                 return NULL;
1042
1043         while (1) {
1044                 if (entry->bitmap) {
1045                         if (entry->offset + BITS_PER_BITMAP *
1046                             block_group->sectorsize > offset)
1047                                 break;
1048                 } else {
1049                         if (entry->offset + entry->bytes > offset)
1050                                 break;
1051                 }
1052
1053                 n = rb_next(&entry->offset_index);
1054                 if (!n)
1055                         return NULL;
1056                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1057         }
1058         return entry;
1059 }
1060
1061 static inline void
1062 __unlink_free_space(struct btrfs_block_group_cache *block_group,
1063                     struct btrfs_free_space *info)
1064 {
1065         rb_erase(&info->offset_index, &block_group->free_space_offset);
1066         block_group->free_extents--;
1067 }
1068
1069 static void unlink_free_space(struct btrfs_block_group_cache *block_group,
1070                               struct btrfs_free_space *info)
1071 {
1072         __unlink_free_space(block_group, info);
1073         block_group->free_space -= info->bytes;
1074 }
1075
1076 static int link_free_space(struct btrfs_block_group_cache *block_group,
1077                            struct btrfs_free_space *info)
1078 {
1079         int ret = 0;
1080
1081         BUG_ON(!info->bitmap && !info->bytes);
1082         ret = tree_insert_offset(&block_group->free_space_offset, info->offset,
1083                                  &info->offset_index, (info->bitmap != NULL));
1084         if (ret)
1085                 return ret;
1086
1087         block_group->free_space += info->bytes;
1088         block_group->free_extents++;
1089         return ret;
1090 }
1091
1092 static void recalculate_thresholds(struct btrfs_block_group_cache *block_group)
1093 {
1094         u64 max_bytes;
1095         u64 bitmap_bytes;
1096         u64 extent_bytes;
1097         u64 size = block_group->key.offset;
1098
1099         /*
1100          * The goal is to keep the total amount of memory used per 1gb of space
1101          * at or below 32k, so we need to adjust how much memory we allow to be
1102          * used by extent based free space tracking
1103          */
1104         if (size < 1024 * 1024 * 1024)
1105                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1106         else
1107                 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1108                         div64_u64(size, 1024 * 1024 * 1024);
1109
1110         /*
1111          * we want to account for 1 more bitmap than what we have so we can make
1112          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1113          * we add more bitmaps.
1114          */
1115         bitmap_bytes = (block_group->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1116
1117         if (bitmap_bytes >= max_bytes) {
1118                 block_group->extents_thresh = 0;
1119                 return;
1120         }
1121
1122         /*
1123          * we want the extent entry threshold to always be at most 1/2 the maxw
1124          * bytes we can have, or whatever is less than that.
1125          */
1126         extent_bytes = max_bytes - bitmap_bytes;
1127         extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1128
1129         block_group->extents_thresh =
1130                 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1131 }
1132
1133 static void bitmap_clear_bits(struct btrfs_block_group_cache *block_group,
1134                               struct btrfs_free_space *info, u64 offset,
1135                               u64 bytes)
1136 {
1137         unsigned long start, end;
1138         unsigned long i;
1139
1140         start = offset_to_bit(info->offset, block_group->sectorsize, offset);
1141         end = start + bytes_to_bits(bytes, block_group->sectorsize);
1142         BUG_ON(end > BITS_PER_BITMAP);
1143
1144         for (i = start; i < end; i++)
1145                 clear_bit(i, info->bitmap);
1146
1147         info->bytes -= bytes;
1148         block_group->free_space -= bytes;
1149 }
1150
1151 static void bitmap_set_bits(struct btrfs_block_group_cache *block_group,
1152                             struct btrfs_free_space *info, u64 offset,
1153                             u64 bytes)
1154 {
1155         unsigned long start, end;
1156         unsigned long i;
1157
1158         start = offset_to_bit(info->offset, block_group->sectorsize, offset);
1159         end = start + bytes_to_bits(bytes, block_group->sectorsize);
1160         BUG_ON(end > BITS_PER_BITMAP);
1161
1162         for (i = start; i < end; i++)
1163                 set_bit(i, info->bitmap);
1164
1165         info->bytes += bytes;
1166         block_group->free_space += bytes;
1167 }
1168
1169 static int search_bitmap(struct btrfs_block_group_cache *block_group,
1170                          struct btrfs_free_space *bitmap_info, u64 *offset,
1171                          u64 *bytes)
1172 {
1173         unsigned long found_bits = 0;
1174         unsigned long bits, i;
1175         unsigned long next_zero;
1176
1177         i = offset_to_bit(bitmap_info->offset, block_group->sectorsize,
1178                           max_t(u64, *offset, bitmap_info->offset));
1179         bits = bytes_to_bits(*bytes, block_group->sectorsize);
1180
1181         for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
1182              i < BITS_PER_BITMAP;
1183              i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
1184                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1185                                                BITS_PER_BITMAP, i);
1186                 if ((next_zero - i) >= bits) {
1187                         found_bits = next_zero - i;
1188                         break;
1189                 }
1190                 i = next_zero;
1191         }
1192
1193         if (found_bits) {
1194                 *offset = (u64)(i * block_group->sectorsize) +
1195                         bitmap_info->offset;
1196                 *bytes = (u64)(found_bits) * block_group->sectorsize;
1197                 return 0;
1198         }
1199
1200         return -1;
1201 }
1202
1203 static struct btrfs_free_space *find_free_space(struct btrfs_block_group_cache
1204                                                 *block_group, u64 *offset,
1205                                                 u64 *bytes, int debug)
1206 {
1207         struct btrfs_free_space *entry;
1208         struct rb_node *node;
1209         int ret;
1210
1211         if (!block_group->free_space_offset.rb_node)
1212                 return NULL;
1213
1214         entry = tree_search_offset(block_group,
1215                                    offset_to_bitmap(block_group, *offset),
1216                                    0, 1);
1217         if (!entry)
1218                 return NULL;
1219
1220         for (node = &entry->offset_index; node; node = rb_next(node)) {
1221                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1222                 if (entry->bytes < *bytes)
1223                         continue;
1224
1225                 if (entry->bitmap) {
1226                         ret = search_bitmap(block_group, entry, offset, bytes);
1227                         if (!ret)
1228                                 return entry;
1229                         continue;
1230                 }
1231
1232                 *offset = entry->offset;
1233                 *bytes = entry->bytes;
1234                 return entry;
1235         }
1236
1237         return NULL;
1238 }
1239
1240 static void add_new_bitmap(struct btrfs_block_group_cache *block_group,
1241                            struct btrfs_free_space *info, u64 offset)
1242 {
1243         u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
1244         int max_bitmaps = (int)div64_u64(block_group->key.offset +
1245                                          bytes_per_bg - 1, bytes_per_bg);
1246         BUG_ON(block_group->total_bitmaps >= max_bitmaps);
1247
1248         info->offset = offset_to_bitmap(block_group, offset);
1249         info->bytes = 0;
1250         link_free_space(block_group, info);
1251         block_group->total_bitmaps++;
1252
1253         recalculate_thresholds(block_group);
1254 }
1255
1256 static void free_bitmap(struct btrfs_block_group_cache *block_group,
1257                         struct btrfs_free_space *bitmap_info)
1258 {
1259         unlink_free_space(block_group, bitmap_info);
1260         kfree(bitmap_info->bitmap);
1261         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1262         block_group->total_bitmaps--;
1263         recalculate_thresholds(block_group);
1264 }
1265
1266 static noinline int remove_from_bitmap(struct btrfs_block_group_cache *block_group,
1267                               struct btrfs_free_space *bitmap_info,
1268                               u64 *offset, u64 *bytes)
1269 {
1270         u64 end;
1271         u64 search_start, search_bytes;
1272         int ret;
1273
1274 again:
1275         end = bitmap_info->offset +
1276                 (u64)(BITS_PER_BITMAP * block_group->sectorsize) - 1;
1277
1278         /*
1279          * XXX - this can go away after a few releases.
1280          *
1281          * since the only user of btrfs_remove_free_space is the tree logging
1282          * stuff, and the only way to test that is under crash conditions, we
1283          * want to have this debug stuff here just in case somethings not
1284          * working.  Search the bitmap for the space we are trying to use to
1285          * make sure its actually there.  If its not there then we need to stop
1286          * because something has gone wrong.
1287          */
1288         search_start = *offset;
1289         search_bytes = *bytes;
1290         search_bytes = min(search_bytes, end - search_start + 1);
1291         ret = search_bitmap(block_group, bitmap_info, &search_start,
1292                             &search_bytes);
1293         BUG_ON(ret < 0 || search_start != *offset);
1294
1295         if (*offset > bitmap_info->offset && *offset + *bytes > end) {
1296                 bitmap_clear_bits(block_group, bitmap_info, *offset,
1297                                   end - *offset + 1);
1298                 *bytes -= end - *offset + 1;
1299                 *offset = end + 1;
1300         } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
1301                 bitmap_clear_bits(block_group, bitmap_info, *offset, *bytes);
1302                 *bytes = 0;
1303         }
1304
1305         if (*bytes) {
1306                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1307                 if (!bitmap_info->bytes)
1308                         free_bitmap(block_group, bitmap_info);
1309
1310                 /*
1311                  * no entry after this bitmap, but we still have bytes to
1312                  * remove, so something has gone wrong.
1313                  */
1314                 if (!next)
1315                         return -EINVAL;
1316
1317                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1318                                        offset_index);
1319
1320                 /*
1321                  * if the next entry isn't a bitmap we need to return to let the
1322                  * extent stuff do its work.
1323                  */
1324                 if (!bitmap_info->bitmap)
1325                         return -EAGAIN;
1326
1327                 /*
1328                  * Ok the next item is a bitmap, but it may not actually hold
1329                  * the information for the rest of this free space stuff, so
1330                  * look for it, and if we don't find it return so we can try
1331                  * everything over again.
1332                  */
1333                 search_start = *offset;
1334                 search_bytes = *bytes;
1335                 ret = search_bitmap(block_group, bitmap_info, &search_start,
1336                                     &search_bytes);
1337                 if (ret < 0 || search_start != *offset)
1338                         return -EAGAIN;
1339
1340                 goto again;
1341         } else if (!bitmap_info->bytes)
1342                 free_bitmap(block_group, bitmap_info);
1343
1344         return 0;
1345 }
1346
1347 static int insert_into_bitmap(struct btrfs_block_group_cache *block_group,
1348                               struct btrfs_free_space *info)
1349 {
1350         struct btrfs_free_space *bitmap_info;
1351         int added = 0;
1352         u64 bytes, offset, end;
1353         int ret;
1354
1355         /*
1356          * If we are below the extents threshold then we can add this as an
1357          * extent, and don't have to deal with the bitmap
1358          */
1359         if (block_group->free_extents < block_group->extents_thresh) {
1360                 /*
1361                  * If this block group has some small extents we don't want to
1362                  * use up all of our free slots in the cache with them, we want
1363                  * to reserve them to larger extents, however if we have plent
1364                  * of cache left then go ahead an dadd them, no sense in adding
1365                  * the overhead of a bitmap if we don't have to.
1366                  */
1367                 if (info->bytes <= block_group->sectorsize * 4) {
1368                         if (block_group->free_extents * 2 <=
1369                             block_group->extents_thresh)
1370                                 return 0;
1371                 } else {
1372                         return 0;
1373                 }
1374         }
1375
1376         /*
1377          * some block groups are so tiny they can't be enveloped by a bitmap, so
1378          * don't even bother to create a bitmap for this
1379          */
1380         if (BITS_PER_BITMAP * block_group->sectorsize >
1381             block_group->key.offset)
1382                 return 0;
1383
1384         bytes = info->bytes;
1385         offset = info->offset;
1386
1387 again:
1388         bitmap_info = tree_search_offset(block_group,
1389                                          offset_to_bitmap(block_group, offset),
1390                                          1, 0);
1391         if (!bitmap_info) {
1392                 BUG_ON(added);
1393                 goto new_bitmap;
1394         }
1395
1396         end = bitmap_info->offset +
1397                 (u64)(BITS_PER_BITMAP * block_group->sectorsize);
1398
1399         if (offset >= bitmap_info->offset && offset + bytes > end) {
1400                 bitmap_set_bits(block_group, bitmap_info, offset,
1401                                 end - offset);
1402                 bytes -= end - offset;
1403                 offset = end;
1404                 added = 0;
1405         } else if (offset >= bitmap_info->offset && offset + bytes <= end) {
1406                 bitmap_set_bits(block_group, bitmap_info, offset, bytes);
1407                 bytes = 0;
1408         } else {
1409                 BUG();
1410         }
1411
1412         if (!bytes) {
1413                 ret = 1;
1414                 goto out;
1415         } else
1416                 goto again;
1417
1418 new_bitmap:
1419         if (info && info->bitmap) {
1420                 add_new_bitmap(block_group, info, offset);
1421                 added = 1;
1422                 info = NULL;
1423                 goto again;
1424         } else {
1425                 spin_unlock(&block_group->tree_lock);
1426
1427                 /* no pre-allocated info, allocate a new one */
1428                 if (!info) {
1429                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
1430                                                  GFP_NOFS);
1431                         if (!info) {
1432                                 spin_lock(&block_group->tree_lock);
1433                                 ret = -ENOMEM;
1434                                 goto out;
1435                         }
1436                 }
1437
1438                 /* allocate the bitmap */
1439                 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1440                 spin_lock(&block_group->tree_lock);
1441                 if (!info->bitmap) {
1442                         ret = -ENOMEM;
1443                         goto out;
1444                 }
1445                 goto again;
1446         }
1447
1448 out:
1449         if (info) {
1450                 if (info->bitmap)
1451                         kfree(info->bitmap);
1452                 kmem_cache_free(btrfs_free_space_cachep, info);
1453         }
1454
1455         return ret;
1456 }
1457
1458 bool try_merge_free_space(struct btrfs_block_group_cache *block_group,
1459                           struct btrfs_free_space *info, bool update_stat)
1460 {
1461         struct btrfs_free_space *left_info;
1462         struct btrfs_free_space *right_info;
1463         bool merged = false;
1464         u64 offset = info->offset;
1465         u64 bytes = info->bytes;
1466
1467         /*
1468          * first we want to see if there is free space adjacent to the range we
1469          * are adding, if there is remove that struct and add a new one to
1470          * cover the entire range
1471          */
1472         right_info = tree_search_offset(block_group, offset + bytes, 0, 0);
1473         if (right_info && rb_prev(&right_info->offset_index))
1474                 left_info = rb_entry(rb_prev(&right_info->offset_index),
1475                                      struct btrfs_free_space, offset_index);
1476         else
1477                 left_info = tree_search_offset(block_group, offset - 1, 0, 0);
1478
1479         if (right_info && !right_info->bitmap) {
1480                 if (update_stat)
1481                         unlink_free_space(block_group, right_info);
1482                 else
1483                         __unlink_free_space(block_group, right_info);
1484                 info->bytes += right_info->bytes;
1485                 kmem_cache_free(btrfs_free_space_cachep, right_info);
1486                 merged = true;
1487         }
1488
1489         if (left_info && !left_info->bitmap &&
1490             left_info->offset + left_info->bytes == offset) {
1491                 if (update_stat)
1492                         unlink_free_space(block_group, left_info);
1493                 else
1494                         __unlink_free_space(block_group, left_info);
1495                 info->offset = left_info->offset;
1496                 info->bytes += left_info->bytes;
1497                 kmem_cache_free(btrfs_free_space_cachep, left_info);
1498                 merged = true;
1499         }
1500
1501         return merged;
1502 }
1503
1504 int btrfs_add_free_space(struct btrfs_block_group_cache *block_group,
1505                          u64 offset, u64 bytes)
1506 {
1507         struct btrfs_free_space *info;
1508         int ret = 0;
1509
1510         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
1511         if (!info)
1512                 return -ENOMEM;
1513
1514         info->offset = offset;
1515         info->bytes = bytes;
1516
1517         spin_lock(&block_group->tree_lock);
1518
1519         if (try_merge_free_space(block_group, info, true))
1520                 goto link;
1521
1522         /*
1523          * There was no extent directly to the left or right of this new
1524          * extent then we know we're going to have to allocate a new extent, so
1525          * before we do that see if we need to drop this into a bitmap
1526          */
1527         ret = insert_into_bitmap(block_group, info);
1528         if (ret < 0) {
1529                 goto out;
1530         } else if (ret) {
1531                 ret = 0;
1532                 goto out;
1533         }
1534 link:
1535         ret = link_free_space(block_group, info);
1536         if (ret)
1537                 kmem_cache_free(btrfs_free_space_cachep, info);
1538 out:
1539         spin_unlock(&block_group->tree_lock);
1540
1541         if (ret) {
1542                 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1543                 BUG_ON(ret == -EEXIST);
1544         }
1545
1546         return ret;
1547 }
1548
1549 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1550                             u64 offset, u64 bytes)
1551 {
1552         struct btrfs_free_space *info;
1553         struct btrfs_free_space *next_info = NULL;
1554         int ret = 0;
1555
1556         spin_lock(&block_group->tree_lock);
1557
1558 again:
1559         info = tree_search_offset(block_group, offset, 0, 0);
1560         if (!info) {
1561                 /*
1562                  * oops didn't find an extent that matched the space we wanted
1563                  * to remove, look for a bitmap instead
1564                  */
1565                 info = tree_search_offset(block_group,
1566                                           offset_to_bitmap(block_group, offset),
1567                                           1, 0);
1568                 if (!info) {
1569                         WARN_ON(1);
1570                         goto out_lock;
1571                 }
1572         }
1573
1574         if (info->bytes < bytes && rb_next(&info->offset_index)) {
1575                 u64 end;
1576                 next_info = rb_entry(rb_next(&info->offset_index),
1577                                              struct btrfs_free_space,
1578                                              offset_index);
1579
1580                 if (next_info->bitmap)
1581                         end = next_info->offset + BITS_PER_BITMAP *
1582                                 block_group->sectorsize - 1;
1583                 else
1584                         end = next_info->offset + next_info->bytes;
1585
1586                 if (next_info->bytes < bytes ||
1587                     next_info->offset > offset || offset > end) {
1588                         printk(KERN_CRIT "Found free space at %llu, size %llu,"
1589                               " trying to use %llu\n",
1590                               (unsigned long long)info->offset,
1591                               (unsigned long long)info->bytes,
1592                               (unsigned long long)bytes);
1593                         WARN_ON(1);
1594                         ret = -EINVAL;
1595                         goto out_lock;
1596                 }
1597
1598                 info = next_info;
1599         }
1600
1601         if (info->bytes == bytes) {
1602                 unlink_free_space(block_group, info);
1603                 if (info->bitmap) {
1604                         kfree(info->bitmap);
1605                         block_group->total_bitmaps--;
1606                 }
1607                 kmem_cache_free(btrfs_free_space_cachep, info);
1608                 goto out_lock;
1609         }
1610
1611         if (!info->bitmap && info->offset == offset) {
1612                 unlink_free_space(block_group, info);
1613                 info->offset += bytes;
1614                 info->bytes -= bytes;
1615                 link_free_space(block_group, info);
1616                 goto out_lock;
1617         }
1618
1619         if (!info->bitmap && info->offset <= offset &&
1620             info->offset + info->bytes >= offset + bytes) {
1621                 u64 old_start = info->offset;
1622                 /*
1623                  * we're freeing space in the middle of the info,
1624                  * this can happen during tree log replay
1625                  *
1626                  * first unlink the old info and then
1627                  * insert it again after the hole we're creating
1628                  */
1629                 unlink_free_space(block_group, info);
1630                 if (offset + bytes < info->offset + info->bytes) {
1631                         u64 old_end = info->offset + info->bytes;
1632
1633                         info->offset = offset + bytes;
1634                         info->bytes = old_end - info->offset;
1635                         ret = link_free_space(block_group, info);
1636                         WARN_ON(ret);
1637                         if (ret)
1638                                 goto out_lock;
1639                 } else {
1640                         /* the hole we're creating ends at the end
1641                          * of the info struct, just free the info
1642                          */
1643                         kmem_cache_free(btrfs_free_space_cachep, info);
1644                 }
1645                 spin_unlock(&block_group->tree_lock);
1646
1647                 /* step two, insert a new info struct to cover
1648                  * anything before the hole
1649                  */
1650                 ret = btrfs_add_free_space(block_group, old_start,
1651                                            offset - old_start);
1652                 WARN_ON(ret);
1653                 goto out;
1654         }
1655
1656         ret = remove_from_bitmap(block_group, info, &offset, &bytes);
1657         if (ret == -EAGAIN)
1658                 goto again;
1659         BUG_ON(ret);
1660 out_lock:
1661         spin_unlock(&block_group->tree_lock);
1662 out:
1663         return ret;
1664 }
1665
1666 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1667                            u64 bytes)
1668 {
1669         struct btrfs_free_space *info;
1670         struct rb_node *n;
1671         int count = 0;
1672
1673         for (n = rb_first(&block_group->free_space_offset); n; n = rb_next(n)) {
1674                 info = rb_entry(n, struct btrfs_free_space, offset_index);
1675                 if (info->bytes >= bytes)
1676                         count++;
1677                 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1678                        (unsigned long long)info->offset,
1679                        (unsigned long long)info->bytes,
1680                        (info->bitmap) ? "yes" : "no");
1681         }
1682         printk(KERN_INFO "block group has cluster?: %s\n",
1683                list_empty(&block_group->cluster_list) ? "no" : "yes");
1684         printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1685                "\n", count);
1686 }
1687
1688 u64 btrfs_block_group_free_space(struct btrfs_block_group_cache *block_group)
1689 {
1690         struct btrfs_free_space *info;
1691         struct rb_node *n;
1692         u64 ret = 0;
1693
1694         for (n = rb_first(&block_group->free_space_offset); n;
1695              n = rb_next(n)) {
1696                 info = rb_entry(n, struct btrfs_free_space, offset_index);
1697                 ret += info->bytes;
1698         }
1699
1700         return ret;
1701 }
1702
1703 /*
1704  * for a given cluster, put all of its extents back into the free
1705  * space cache.  If the block group passed doesn't match the block group
1706  * pointed to by the cluster, someone else raced in and freed the
1707  * cluster already.  In that case, we just return without changing anything
1708  */
1709 static int
1710 __btrfs_return_cluster_to_free_space(
1711                              struct btrfs_block_group_cache *block_group,
1712                              struct btrfs_free_cluster *cluster)
1713 {
1714         struct btrfs_free_space *entry;
1715         struct rb_node *node;
1716
1717         spin_lock(&cluster->lock);
1718         if (cluster->block_group != block_group)
1719                 goto out;
1720
1721         cluster->block_group = NULL;
1722         cluster->window_start = 0;
1723         list_del_init(&cluster->block_group_list);
1724
1725         node = rb_first(&cluster->root);
1726         while (node) {
1727                 bool bitmap;
1728
1729                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1730                 node = rb_next(&entry->offset_index);
1731                 rb_erase(&entry->offset_index, &cluster->root);
1732
1733                 bitmap = (entry->bitmap != NULL);
1734                 if (!bitmap)
1735                         try_merge_free_space(block_group, entry, false);
1736                 tree_insert_offset(&block_group->free_space_offset,
1737                                    entry->offset, &entry->offset_index, bitmap);
1738         }
1739         cluster->root = RB_ROOT;
1740
1741 out:
1742         spin_unlock(&cluster->lock);
1743         btrfs_put_block_group(block_group);
1744         return 0;
1745 }
1746
1747 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
1748 {
1749         struct btrfs_free_space *info;
1750         struct rb_node *node;
1751         struct btrfs_free_cluster *cluster;
1752         struct list_head *head;
1753
1754         spin_lock(&block_group->tree_lock);
1755         while ((head = block_group->cluster_list.next) !=
1756                &block_group->cluster_list) {
1757                 cluster = list_entry(head, struct btrfs_free_cluster,
1758                                      block_group_list);
1759
1760                 WARN_ON(cluster->block_group != block_group);
1761                 __btrfs_return_cluster_to_free_space(block_group, cluster);
1762                 if (need_resched()) {
1763                         spin_unlock(&block_group->tree_lock);
1764                         cond_resched();
1765                         spin_lock(&block_group->tree_lock);
1766                 }
1767         }
1768
1769         while ((node = rb_last(&block_group->free_space_offset)) != NULL) {
1770                 info = rb_entry(node, struct btrfs_free_space, offset_index);
1771                 if (!info->bitmap) {
1772                         unlink_free_space(block_group, info);
1773                         kmem_cache_free(btrfs_free_space_cachep, info);
1774                 } else {
1775                         free_bitmap(block_group, info);
1776                 }
1777
1778                 if (need_resched()) {
1779                         spin_unlock(&block_group->tree_lock);
1780                         cond_resched();
1781                         spin_lock(&block_group->tree_lock);
1782                 }
1783         }
1784
1785         spin_unlock(&block_group->tree_lock);
1786 }
1787
1788 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
1789                                u64 offset, u64 bytes, u64 empty_size)
1790 {
1791         struct btrfs_free_space *entry = NULL;
1792         u64 bytes_search = bytes + empty_size;
1793         u64 ret = 0;
1794
1795         spin_lock(&block_group->tree_lock);
1796         entry = find_free_space(block_group, &offset, &bytes_search, 0);
1797         if (!entry)
1798                 goto out;
1799
1800         ret = offset;
1801         if (entry->bitmap) {
1802                 bitmap_clear_bits(block_group, entry, offset, bytes);
1803                 if (!entry->bytes)
1804                         free_bitmap(block_group, entry);
1805         } else {
1806                 unlink_free_space(block_group, entry);
1807                 entry->offset += bytes;
1808                 entry->bytes -= bytes;
1809                 if (!entry->bytes)
1810                         kmem_cache_free(btrfs_free_space_cachep, entry);
1811                 else
1812                         link_free_space(block_group, entry);
1813         }
1814
1815 out:
1816         spin_unlock(&block_group->tree_lock);
1817
1818         return ret;
1819 }
1820
1821 /*
1822  * given a cluster, put all of its extents back into the free space
1823  * cache.  If a block group is passed, this function will only free
1824  * a cluster that belongs to the passed block group.
1825  *
1826  * Otherwise, it'll get a reference on the block group pointed to by the
1827  * cluster and remove the cluster from it.
1828  */
1829 int btrfs_return_cluster_to_free_space(
1830                                struct btrfs_block_group_cache *block_group,
1831                                struct btrfs_free_cluster *cluster)
1832 {
1833         int ret;
1834
1835         /* first, get a safe pointer to the block group */
1836         spin_lock(&cluster->lock);
1837         if (!block_group) {
1838                 block_group = cluster->block_group;
1839                 if (!block_group) {
1840                         spin_unlock(&cluster->lock);
1841                         return 0;
1842                 }
1843         } else if (cluster->block_group != block_group) {
1844                 /* someone else has already freed it don't redo their work */
1845                 spin_unlock(&cluster->lock);
1846                 return 0;
1847         }
1848         atomic_inc(&block_group->count);
1849         spin_unlock(&cluster->lock);
1850
1851         /* now return any extents the cluster had on it */
1852         spin_lock(&block_group->tree_lock);
1853         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
1854         spin_unlock(&block_group->tree_lock);
1855
1856         /* finally drop our ref */
1857         btrfs_put_block_group(block_group);
1858         return ret;
1859 }
1860
1861 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
1862                                    struct btrfs_free_cluster *cluster,
1863                                    struct btrfs_free_space *entry,
1864                                    u64 bytes, u64 min_start)
1865 {
1866         int err;
1867         u64 search_start = cluster->window_start;
1868         u64 search_bytes = bytes;
1869         u64 ret = 0;
1870
1871         search_start = min_start;
1872         search_bytes = bytes;
1873
1874         err = search_bitmap(block_group, entry, &search_start,
1875                             &search_bytes);
1876         if (err)
1877                 return 0;
1878
1879         ret = search_start;
1880         bitmap_clear_bits(block_group, entry, ret, bytes);
1881
1882         return ret;
1883 }
1884
1885 /*
1886  * given a cluster, try to allocate 'bytes' from it, returns 0
1887  * if it couldn't find anything suitably large, or a logical disk offset
1888  * if things worked out
1889  */
1890 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
1891                              struct btrfs_free_cluster *cluster, u64 bytes,
1892                              u64 min_start)
1893 {
1894         struct btrfs_free_space *entry = NULL;
1895         struct rb_node *node;
1896         u64 ret = 0;
1897
1898         spin_lock(&cluster->lock);
1899         if (bytes > cluster->max_size)
1900                 goto out;
1901
1902         if (cluster->block_group != block_group)
1903                 goto out;
1904
1905         node = rb_first(&cluster->root);
1906         if (!node)
1907                 goto out;
1908
1909         entry = rb_entry(node, struct btrfs_free_space, offset_index);
1910         while(1) {
1911                 if (entry->bytes < bytes ||
1912                     (!entry->bitmap && entry->offset < min_start)) {
1913                         struct rb_node *node;
1914
1915                         node = rb_next(&entry->offset_index);
1916                         if (!node)
1917                                 break;
1918                         entry = rb_entry(node, struct btrfs_free_space,
1919                                          offset_index);
1920                         continue;
1921                 }
1922
1923                 if (entry->bitmap) {
1924                         ret = btrfs_alloc_from_bitmap(block_group,
1925                                                       cluster, entry, bytes,
1926                                                       min_start);
1927                         if (ret == 0) {
1928                                 struct rb_node *node;
1929                                 node = rb_next(&entry->offset_index);
1930                                 if (!node)
1931                                         break;
1932                                 entry = rb_entry(node, struct btrfs_free_space,
1933                                                  offset_index);
1934                                 continue;
1935                         }
1936                 } else {
1937
1938                         ret = entry->offset;
1939
1940                         entry->offset += bytes;
1941                         entry->bytes -= bytes;
1942                 }
1943
1944                 if (entry->bytes == 0)
1945                         rb_erase(&entry->offset_index, &cluster->root);
1946                 break;
1947         }
1948 out:
1949         spin_unlock(&cluster->lock);
1950
1951         if (!ret)
1952                 return 0;
1953
1954         spin_lock(&block_group->tree_lock);
1955
1956         block_group->free_space -= bytes;
1957         if (entry->bytes == 0) {
1958                 block_group->free_extents--;
1959                 if (entry->bitmap) {
1960                         kfree(entry->bitmap);
1961                         block_group->total_bitmaps--;
1962                         recalculate_thresholds(block_group);
1963                 }
1964                 kmem_cache_free(btrfs_free_space_cachep, entry);
1965         }
1966
1967         spin_unlock(&block_group->tree_lock);
1968
1969         return ret;
1970 }
1971
1972 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
1973                                 struct btrfs_free_space *entry,
1974                                 struct btrfs_free_cluster *cluster,
1975                                 u64 offset, u64 bytes, u64 min_bytes)
1976 {
1977         unsigned long next_zero;
1978         unsigned long i;
1979         unsigned long search_bits;
1980         unsigned long total_bits;
1981         unsigned long found_bits;
1982         unsigned long start = 0;
1983         unsigned long total_found = 0;
1984         int ret;
1985         bool found = false;
1986
1987         i = offset_to_bit(entry->offset, block_group->sectorsize,
1988                           max_t(u64, offset, entry->offset));
1989         search_bits = bytes_to_bits(bytes, block_group->sectorsize);
1990         total_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
1991
1992 again:
1993         found_bits = 0;
1994         for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
1995              i < BITS_PER_BITMAP;
1996              i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
1997                 next_zero = find_next_zero_bit(entry->bitmap,
1998                                                BITS_PER_BITMAP, i);
1999                 if (next_zero - i >= search_bits) {
2000                         found_bits = next_zero - i;
2001                         break;
2002                 }
2003                 i = next_zero;
2004         }
2005
2006         if (!found_bits)
2007                 return -ENOSPC;
2008
2009         if (!found) {
2010                 start = i;
2011                 found = true;
2012         }
2013
2014         total_found += found_bits;
2015
2016         if (cluster->max_size < found_bits * block_group->sectorsize)
2017                 cluster->max_size = found_bits * block_group->sectorsize;
2018
2019         if (total_found < total_bits) {
2020                 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
2021                 if (i - start > total_bits * 2) {
2022                         total_found = 0;
2023                         cluster->max_size = 0;
2024                         found = false;
2025                 }
2026                 goto again;
2027         }
2028
2029         cluster->window_start = start * block_group->sectorsize +
2030                 entry->offset;
2031         rb_erase(&entry->offset_index, &block_group->free_space_offset);
2032         ret = tree_insert_offset(&cluster->root, entry->offset,
2033                                  &entry->offset_index, 1);
2034         BUG_ON(ret);
2035
2036         return 0;
2037 }
2038
2039 /*
2040  * This searches the block group for just extents to fill the cluster with.
2041  */
2042 static int setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2043                                    struct btrfs_free_cluster *cluster,
2044                                    u64 offset, u64 bytes, u64 min_bytes)
2045 {
2046         struct btrfs_free_space *first = NULL;
2047         struct btrfs_free_space *entry = NULL;
2048         struct btrfs_free_space *prev = NULL;
2049         struct btrfs_free_space *last;
2050         struct rb_node *node;
2051         u64 window_start;
2052         u64 window_free;
2053         u64 max_extent;
2054         u64 max_gap = 128 * 1024;
2055
2056         entry = tree_search_offset(block_group, offset, 0, 1);
2057         if (!entry)
2058                 return -ENOSPC;
2059
2060         /*
2061          * We don't want bitmaps, so just move along until we find a normal
2062          * extent entry.
2063          */
2064         while (entry->bitmap) {
2065                 node = rb_next(&entry->offset_index);
2066                 if (!node)
2067                         return -ENOSPC;
2068                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2069         }
2070
2071         window_start = entry->offset;
2072         window_free = entry->bytes;
2073         max_extent = entry->bytes;
2074         first = entry;
2075         last = entry;
2076         prev = entry;
2077
2078         while (window_free <= min_bytes) {
2079                 node = rb_next(&entry->offset_index);
2080                 if (!node)
2081                         return -ENOSPC;
2082                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2083
2084                 if (entry->bitmap)
2085                         continue;
2086                 /*
2087                  * we haven't filled the empty size and the window is
2088                  * very large.  reset and try again
2089                  */
2090                 if (entry->offset - (prev->offset + prev->bytes) > max_gap ||
2091                     entry->offset - window_start > (min_bytes * 2)) {
2092                         first = entry;
2093                         window_start = entry->offset;
2094                         window_free = entry->bytes;
2095                         last = entry;
2096                         max_extent = entry->bytes;
2097                 } else {
2098                         last = entry;
2099                         window_free += entry->bytes;
2100                         if (entry->bytes > max_extent)
2101                                 max_extent = entry->bytes;
2102                 }
2103                 prev = entry;
2104         }
2105
2106         cluster->window_start = first->offset;
2107
2108         node = &first->offset_index;
2109
2110         /*
2111          * now we've found our entries, pull them out of the free space
2112          * cache and put them into the cluster rbtree
2113          */
2114         do {
2115                 int ret;
2116
2117                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2118                 node = rb_next(&entry->offset_index);
2119                 if (entry->bitmap)
2120                         continue;
2121
2122                 rb_erase(&entry->offset_index, &block_group->free_space_offset);
2123                 ret = tree_insert_offset(&cluster->root, entry->offset,
2124                                          &entry->offset_index, 0);
2125                 BUG_ON(ret);
2126         } while (node && entry != last);
2127
2128         cluster->max_size = max_extent;
2129
2130         return 0;
2131 }
2132
2133 /*
2134  * This specifically looks for bitmaps that may work in the cluster, we assume
2135  * that we have already failed to find extents that will work.
2136  */
2137 static int setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2138                                 struct btrfs_free_cluster *cluster,
2139                                 u64 offset, u64 bytes, u64 min_bytes)
2140 {
2141         struct btrfs_free_space *entry;
2142         struct rb_node *node;
2143         int ret = -ENOSPC;
2144
2145         if (block_group->total_bitmaps == 0)
2146                 return -ENOSPC;
2147
2148         entry = tree_search_offset(block_group,
2149                                    offset_to_bitmap(block_group, offset),
2150                                    0, 1);
2151         if (!entry)
2152                 return -ENOSPC;
2153
2154         node = &entry->offset_index;
2155         do {
2156                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2157                 node = rb_next(&entry->offset_index);
2158                 if (!entry->bitmap)
2159                         continue;
2160                 if (entry->bytes < min_bytes)
2161                         continue;
2162                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2163                                            bytes, min_bytes);
2164         } while (ret && node);
2165
2166         return ret;
2167 }
2168
2169 /*
2170  * here we try to find a cluster of blocks in a block group.  The goal
2171  * is to find at least bytes free and up to empty_size + bytes free.
2172  * We might not find them all in one contiguous area.
2173  *
2174  * returns zero and sets up cluster if things worked out, otherwise
2175  * it returns -enospc
2176  */
2177 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
2178                              struct btrfs_root *root,
2179                              struct btrfs_block_group_cache *block_group,
2180                              struct btrfs_free_cluster *cluster,
2181                              u64 offset, u64 bytes, u64 empty_size)
2182 {
2183         u64 min_bytes;
2184         int ret;
2185
2186         /* for metadata, allow allocates with more holes */
2187         if (btrfs_test_opt(root, SSD_SPREAD)) {
2188                 min_bytes = bytes + empty_size;
2189         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2190                 /*
2191                  * we want to do larger allocations when we are
2192                  * flushing out the delayed refs, it helps prevent
2193                  * making more work as we go along.
2194                  */
2195                 if (trans->transaction->delayed_refs.flushing)
2196                         min_bytes = max(bytes, (bytes + empty_size) >> 1);
2197                 else
2198                         min_bytes = max(bytes, (bytes + empty_size) >> 4);
2199         } else
2200                 min_bytes = max(bytes, (bytes + empty_size) >> 2);
2201
2202         spin_lock(&block_group->tree_lock);
2203
2204         /*
2205          * If we know we don't have enough space to make a cluster don't even
2206          * bother doing all the work to try and find one.
2207          */
2208         if (block_group->free_space < min_bytes) {
2209                 spin_unlock(&block_group->tree_lock);
2210                 return -ENOSPC;
2211         }
2212
2213         spin_lock(&cluster->lock);
2214
2215         /* someone already found a cluster, hooray */
2216         if (cluster->block_group) {
2217                 ret = 0;
2218                 goto out;
2219         }
2220
2221         ret = setup_cluster_no_bitmap(block_group, cluster, offset, bytes,
2222                                       min_bytes);
2223         if (ret)
2224                 ret = setup_cluster_bitmap(block_group, cluster, offset,
2225                                            bytes, min_bytes);
2226
2227         if (!ret) {
2228                 atomic_inc(&block_group->count);
2229                 list_add_tail(&cluster->block_group_list,
2230                               &block_group->cluster_list);
2231                 cluster->block_group = block_group;
2232         }
2233 out:
2234         spin_unlock(&cluster->lock);
2235         spin_unlock(&block_group->tree_lock);
2236
2237         return ret;
2238 }
2239
2240 /*
2241  * simple code to zero out a cluster
2242  */
2243 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2244 {
2245         spin_lock_init(&cluster->lock);
2246         spin_lock_init(&cluster->refill_lock);
2247         cluster->root = RB_ROOT;
2248         cluster->max_size = 0;
2249         INIT_LIST_HEAD(&cluster->block_group_list);
2250         cluster->block_group = NULL;
2251 }
2252
2253 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
2254                            u64 *trimmed, u64 start, u64 end, u64 minlen)
2255 {
2256         struct btrfs_free_space *entry = NULL;
2257         struct btrfs_fs_info *fs_info = block_group->fs_info;
2258         u64 bytes = 0;
2259         u64 actually_trimmed;
2260         int ret = 0;
2261
2262         *trimmed = 0;
2263
2264         while (start < end) {
2265                 spin_lock(&block_group->tree_lock);
2266
2267                 if (block_group->free_space < minlen) {
2268                         spin_unlock(&block_group->tree_lock);
2269                         break;
2270                 }
2271
2272                 entry = tree_search_offset(block_group, start, 0, 1);
2273                 if (!entry)
2274                         entry = tree_search_offset(block_group,
2275                                                    offset_to_bitmap(block_group,
2276                                                                     start),
2277                                                    1, 1);
2278
2279                 if (!entry || entry->offset >= end) {
2280                         spin_unlock(&block_group->tree_lock);
2281                         break;
2282                 }
2283
2284                 if (entry->bitmap) {
2285                         ret = search_bitmap(block_group, entry, &start, &bytes);
2286                         if (!ret) {
2287                                 if (start >= end) {
2288                                         spin_unlock(&block_group->tree_lock);
2289                                         break;
2290                                 }
2291                                 bytes = min(bytes, end - start);
2292                                 bitmap_clear_bits(block_group, entry,
2293                                                   start, bytes);
2294                                 if (entry->bytes == 0)
2295                                         free_bitmap(block_group, entry);
2296                         } else {
2297                                 start = entry->offset + BITS_PER_BITMAP *
2298                                         block_group->sectorsize;
2299                                 spin_unlock(&block_group->tree_lock);
2300                                 ret = 0;
2301                                 continue;
2302                         }
2303                 } else {
2304                         start = entry->offset;
2305                         bytes = min(entry->bytes, end - start);
2306                         unlink_free_space(block_group, entry);
2307                         kmem_cache_free(btrfs_free_space_cachep, entry);
2308                 }
2309
2310                 spin_unlock(&block_group->tree_lock);
2311
2312                 if (bytes >= minlen) {
2313                         int update_ret;
2314                         update_ret = btrfs_update_reserved_bytes(block_group,
2315                                                                  bytes, 1, 1);
2316
2317                         ret = btrfs_error_discard_extent(fs_info->extent_root,
2318                                                          start,
2319                                                          bytes,
2320                                                          &actually_trimmed);
2321
2322                         btrfs_add_free_space(block_group,
2323                                              start, bytes);
2324                         if (!update_ret)
2325                                 btrfs_update_reserved_bytes(block_group,
2326                                                             bytes, 0, 1);
2327
2328                         if (ret)
2329                                 break;
2330                         *trimmed += actually_trimmed;
2331                 }
2332                 start += bytes;
2333                 bytes = 0;
2334
2335                 if (fatal_signal_pending(current)) {
2336                         ret = -ERESTARTSYS;
2337                         break;
2338                 }
2339
2340                 cond_resched();
2341         }
2342
2343         return ret;
2344 }