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 <linux/ratelimit.h>
  24 #include "ctree.h"
  25 #include "free-space-cache.h"
  26 #include "transaction.h"
  27 #include "disk-io.h"
  28 #include "extent_io.h"
  29 #include "inode-map.h"
  30
  31 #define BITS_PER_BITMAP         (PAGE_CACHE_SIZE * 8)
  32 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
  33
  34 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  35                            struct btrfs_free_space *info);
  36
  37 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  38                                                struct btrfs_path *path,
  39                                                u64 offset)
  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         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  50         key.offset = offset;
  51         key.type = 0;
  52
  53         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  54         if (ret < 0)
  55                 return ERR_PTR(ret);
  56         if (ret > 0) {
  57                 btrfs_release_path(path);
  58                 return ERR_PTR(-ENOENT);
  59         }
  60
  61         leaf = path->nodes[0];
  62         header = btrfs_item_ptr(leaf, path->slots[0],
  63                                 struct btrfs_free_space_header);
  64         btrfs_free_space_key(leaf, header, &disk_key);
  65         btrfs_disk_key_to_cpu(&location, &disk_key);
  66         btrfs_release_path(path);
  67
  68         inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
  69         if (!inode)
  70                 return ERR_PTR(-ENOENT);
  71         if (IS_ERR(inode))
  72                 return inode;
  73         if (is_bad_inode(inode)) {
  74                 iput(inode);
  75                 return ERR_PTR(-ENOENT);
  76         }
  77
  78         inode->i_mapping->flags &= ~__GFP_FS;
  79
  80         return inode;
  81 }
  82
  83 struct inode *lookup_free_space_inode(struct btrfs_root *root,
  84                                       struct btrfs_block_group_cache
  85                                       *block_group, struct btrfs_path *path)
  86 {
  87         struct inode *inode = NULL;
  88         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
  89
  90         spin_lock(&block_group->lock);
  91         if (block_group->inode)
  92                 inode = igrab(block_group->inode);
  93         spin_unlock(&block_group->lock);
  94         if (inode)
  95                 return inode;
  96
  97         inode = __lookup_free_space_inode(root, path,
  98                                           block_group->key.objectid);
  99         if (IS_ERR(inode))
 100                 return inode;
 101
 102         spin_lock(&block_group->lock);
 103         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 104                 printk(KERN_INFO "Old style space inode found, converting.\n");
 105                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 106                         BTRFS_INODE_NODATACOW;
 107                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 108         }
 109
 110         if (!block_group->iref) {
 111                 block_group->inode = igrab(inode);
 112                 block_group->iref = 1;
 113         }
 114         spin_unlock(&block_group->lock);
 115
 116         return inode;
 117 }
 118
 119 int __create_free_space_inode(struct btrfs_root *root,
 120                               struct btrfs_trans_handle *trans,
 121                               struct btrfs_path *path, u64 ino, u64 offset)
 122 {
 123         struct btrfs_key key;
 124         struct btrfs_disk_key disk_key;
 125         struct btrfs_free_space_header *header;
 126         struct btrfs_inode_item *inode_item;
 127         struct extent_buffer *leaf;
 128         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 129         int ret;
 130
 131         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 132         if (ret)
 133                 return ret;
 134
 135         /* We inline crc's for the free disk space cache */
 136         if (ino != BTRFS_FREE_INO_OBJECTID)
 137                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 138
 139         leaf = path->nodes[0];
 140         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 141                                     struct btrfs_inode_item);
 142         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 143         memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
 144                              sizeof(*inode_item));
 145         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 146         btrfs_set_inode_size(leaf, inode_item, 0);
 147         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 148         btrfs_set_inode_uid(leaf, inode_item, 0);
 149         btrfs_set_inode_gid(leaf, inode_item, 0);
 150         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 151         btrfs_set_inode_flags(leaf, inode_item, flags);
 152         btrfs_set_inode_nlink(leaf, inode_item, 1);
 153         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 154         btrfs_set_inode_block_group(leaf, inode_item, offset);
 155         btrfs_mark_buffer_dirty(leaf);
 156         btrfs_release_path(path);
 157
 158         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 159         key.offset = offset;
 160         key.type = 0;
 161
 162         ret = btrfs_insert_empty_item(trans, root, path, &key,
 163                                       sizeof(struct btrfs_free_space_header));
 164         if (ret < 0) {
 165                 btrfs_release_path(path);
 166                 return ret;
 167         }
 168         leaf = path->nodes[0];
 169         header = btrfs_item_ptr(leaf, path->slots[0],
 170                                 struct btrfs_free_space_header);
 171         memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
 172         btrfs_set_free_space_key(leaf, header, &disk_key);
 173         btrfs_mark_buffer_dirty(leaf);
 174         btrfs_release_path(path);
 175
 176         return 0;
 177 }
 178
 179 int create_free_space_inode(struct btrfs_root *root,
 180                             struct btrfs_trans_handle *trans,
 181                             struct btrfs_block_group_cache *block_group,
 182                             struct btrfs_path *path)
 183 {
 184         int ret;
 185         u64 ino;
 186
 187         ret = btrfs_find_free_objectid(root, &ino);
 188         if (ret < 0)
 189                 return ret;
 190
 191         return __create_free_space_inode(root, trans, path, ino,
 192                                          block_group->key.objectid);
 193 }
 194
 195 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
 196                                     struct btrfs_trans_handle *trans,
 197                                     struct btrfs_path *path,
 198                                     struct inode *inode)
 199 {
 200         struct btrfs_block_rsv *rsv;
 201         loff_t oldsize;
 202         int ret = 0;
 203
 204         rsv = trans->block_rsv;
 205         trans->block_rsv = root->orphan_block_rsv;
 206         ret = btrfs_block_rsv_check(root, root->orphan_block_rsv, 5);
 207         if (ret)
 208                 return ret;
 209
 210         oldsize = i_size_read(inode);
 211         btrfs_i_size_write(inode, 0);
 212         truncate_pagecache(inode, oldsize, 0);
 213
 214         /*
 215          * We don't need an orphan item because truncating the free space cache
 216          * will never be split across transactions.
 217          */
 218         ret = btrfs_truncate_inode_items(trans, root, inode,
 219                                          0, BTRFS_EXTENT_DATA_KEY);
 220
 221         trans->block_rsv = rsv;
 222         if (ret) {
 223                 WARN_ON(1);
 224                 return ret;
 225         }
 226
 227         ret = btrfs_update_inode(trans, root, inode);
 228         return ret;
 229 }
 230
 231 static int readahead_cache(struct inode *inode)
 232 {
 233         struct file_ra_state *ra;
 234         unsigned long last_index;
 235
 236         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 237         if (!ra)
 238                 return -ENOMEM;
 239
 240         file_ra_state_init(ra, inode->i_mapping);
 241         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 242
 243         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 244
 245         kfree(ra);
 246
 247         return 0;
 248 }
 249
 250 struct io_ctl {
 251         void *cur, *orig;
 252         struct page *page;
 253         struct page **pages;
 254         struct btrfs_root *root;
 255         unsigned long size;
 256         int index;
 257         int num_pages;
 258         unsigned check_crcs:1;
 259 };
 260
 261 static int io_ctl_init(struct io_ctl *io_ctl, struct inode *inode,
 262                        struct btrfs_root *root)
 263 {
 264         memset(io_ctl, 0, sizeof(struct io_ctl));
 265         io_ctl->num_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
 266                 PAGE_CACHE_SHIFT;
 267         io_ctl->pages = kzalloc(sizeof(struct page *) * io_ctl->num_pages,
 268                                 GFP_NOFS);
 269         if (!io_ctl->pages)
 270                 return -ENOMEM;
 271         io_ctl->root = root;
 272         if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
 273                 io_ctl->check_crcs = 1;
 274         return 0;
 275 }
 276
 277 static void io_ctl_free(struct io_ctl *io_ctl)
 278 {
 279         kfree(io_ctl->pages);
 280 }
 281
 282 static void io_ctl_unmap_page(struct io_ctl *io_ctl)
 283 {
 284         if (io_ctl->cur) {
 285                 kunmap(io_ctl->page);
 286                 io_ctl->cur = NULL;
 287                 io_ctl->orig = NULL;
 288         }
 289 }
 290
 291 static void io_ctl_map_page(struct io_ctl *io_ctl, int clear)
 292 {
 293         WARN_ON(io_ctl->cur);
 294         BUG_ON(io_ctl->index >= io_ctl->num_pages);
 295         io_ctl->page = io_ctl->pages[io_ctl->index++];
 296         io_ctl->cur = kmap(io_ctl->page);
 297         io_ctl->orig = io_ctl->cur;
 298         io_ctl->size = PAGE_CACHE_SIZE;
 299         if (clear)
 300                 memset(io_ctl->cur, 0, PAGE_CACHE_SIZE);
 301 }
 302
 303 static void io_ctl_drop_pages(struct io_ctl *io_ctl)
 304 {
 305         int i;
 306
 307         io_ctl_unmap_page(io_ctl);
 308
 309         for (i = 0; i < io_ctl->num_pages; i++) {
 310                 ClearPageChecked(io_ctl->pages[i]);
 311                 unlock_page(io_ctl->pages[i]);
 312                 page_cache_release(io_ctl->pages[i]);
 313         }
 314 }
 315
 316 static int io_ctl_prepare_pages(struct io_ctl *io_ctl, struct inode *inode,
 317                                 int uptodate)
 318 {
 319         struct page *page;
 320         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 321         int i;
 322
 323         for (i = 0; i < io_ctl->num_pages; i++) {
 324                 page = find_or_create_page(inode->i_mapping, i, mask);
 325                 if (!page) {
 326                         io_ctl_drop_pages(io_ctl);
 327                         return -ENOMEM;
 328                 }
 329                 io_ctl->pages[i] = page;
 330                 if (uptodate && !PageUptodate(page)) {
 331                         btrfs_readpage(NULL, page);
 332                         lock_page(page);
 333                         if (!PageUptodate(page)) {
 334                                 printk(KERN_ERR "btrfs: error reading free "
 335                                        "space cache\n");
 336                                 io_ctl_drop_pages(io_ctl);
 337                                 return -EIO;
 338                         }
 339                 }
 340         }
 341
 342         return 0;
 343 }
 344
 345 static void io_ctl_set_generation(struct io_ctl *io_ctl, u64 generation)
 346 {
 347         u64 *val;
 348
 349         io_ctl_map_page(io_ctl, 1);
 350
 351         /*
 352          * Skip the csum areas.  If we don't check crcs then we just have a
 353          * 64bit chunk at the front of the first page.
 354          */
 355         if (io_ctl->check_crcs) {
 356                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 357                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 358         } else {
 359                 io_ctl->cur += sizeof(u64);
 360                 io_ctl->size -= sizeof(u64) * 2;
 361         }
 362
 363         val = io_ctl->cur;
 364         *val = cpu_to_le64(generation);
 365         io_ctl->cur += sizeof(u64);
 366 }
 367
 368 static int io_ctl_check_generation(struct io_ctl *io_ctl, u64 generation)
 369 {
 370         u64 *gen;
 371
 372         /*
 373          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 374          * chunk at the front of the first page.
 375          */
 376         if (io_ctl->check_crcs) {
 377                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 378                 io_ctl->size -= sizeof(u64) +
 379                         (sizeof(u32) * io_ctl->num_pages);
 380         } else {
 381                 io_ctl->cur += sizeof(u64);
 382                 io_ctl->size -= sizeof(u64) * 2;
 383         }
 384
 385         gen = io_ctl->cur;
 386         if (le64_to_cpu(*gen) != generation) {
 387                 printk_ratelimited(KERN_ERR "btrfs: space cache generation "
 388                                    "(%Lu) does not match inode (%Lu)\n", *gen,
 389                                    generation);
 390                 io_ctl_unmap_page(io_ctl);
 391                 return -EIO;
 392         }
 393         io_ctl->cur += sizeof(u64);
 394         return 0;
 395 }
 396
 397 static void io_ctl_set_crc(struct io_ctl *io_ctl, int index)
 398 {
 399         u32 *tmp;
 400         u32 crc = ~(u32)0;
 401         unsigned offset = 0;
 402
 403         if (!io_ctl->check_crcs) {
 404                 io_ctl_unmap_page(io_ctl);
 405                 return;
 406         }
 407
 408         if (index == 0)
 409                 offset = sizeof(u32) * io_ctl->num_pages;;
 410
 411         crc = btrfs_csum_data(io_ctl->root, io_ctl->orig + offset, crc,
 412                               PAGE_CACHE_SIZE - offset);
 413         btrfs_csum_final(crc, (char *)&crc);
 414         io_ctl_unmap_page(io_ctl);
 415         tmp = kmap(io_ctl->pages[0]);
 416         tmp += index;
 417         *tmp = crc;
 418         kunmap(io_ctl->pages[0]);
 419 }
 420
 421 static int io_ctl_check_crc(struct io_ctl *io_ctl, int index)
 422 {
 423         u32 *tmp, val;
 424         u32 crc = ~(u32)0;
 425         unsigned offset = 0;
 426
 427         if (!io_ctl->check_crcs) {
 428                 io_ctl_map_page(io_ctl, 0);
 429                 return 0;
 430         }
 431
 432         if (index == 0)
 433                 offset = sizeof(u32) * io_ctl->num_pages;
 434
 435         tmp = kmap(io_ctl->pages[0]);
 436         tmp += index;
 437         val = *tmp;
 438         kunmap(io_ctl->pages[0]);
 439
 440         io_ctl_map_page(io_ctl, 0);
 441         crc = btrfs_csum_data(io_ctl->root, io_ctl->orig + offset, crc,
 442                               PAGE_CACHE_SIZE - offset);
 443         btrfs_csum_final(crc, (char *)&crc);
 444         if (val != crc) {
 445                 printk_ratelimited(KERN_ERR "btrfs: csum mismatch on free "
 446                                    "space cache\n");
 447                 io_ctl_unmap_page(io_ctl);
 448                 return -EIO;
 449         }
 450
 451         return 0;
 452 }
 453
 454 static int io_ctl_add_entry(struct io_ctl *io_ctl, u64 offset, u64 bytes,
 455                             void *bitmap)
 456 {
 457         struct btrfs_free_space_entry *entry;
 458
 459         if (!io_ctl->cur)
 460                 return -ENOSPC;
 461
 462         entry = io_ctl->cur;
 463         entry->offset = cpu_to_le64(offset);
 464         entry->bytes = cpu_to_le64(bytes);
 465         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 466                 BTRFS_FREE_SPACE_EXTENT;
 467         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 468         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 469
 470         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 471                 return 0;
 472
 473         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 474
 475         /* No more pages to map */
 476         if (io_ctl->index >= io_ctl->num_pages)
 477                 return 0;
 478
 479         /* map the next page */
 480         io_ctl_map_page(io_ctl, 1);
 481         return 0;
 482 }
 483
 484 static int io_ctl_add_bitmap(struct io_ctl *io_ctl, void *bitmap)
 485 {
 486         if (!io_ctl->cur)
 487                 return -ENOSPC;
 488
 489         /*
 490          * If we aren't at the start of the current page, unmap this one and
 491          * map the next one if there is any left.
 492          */
 493         if (io_ctl->cur != io_ctl->orig) {
 494                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 495                 if (io_ctl->index >= io_ctl->num_pages)
 496                         return -ENOSPC;
 497                 io_ctl_map_page(io_ctl, 0);
 498         }
 499
 500         memcpy(io_ctl->cur, bitmap, PAGE_CACHE_SIZE);
 501         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 502         if (io_ctl->index < io_ctl->num_pages)
 503                 io_ctl_map_page(io_ctl, 0);
 504         return 0;
 505 }
 506
 507 static void io_ctl_zero_remaining_pages(struct io_ctl *io_ctl)
 508 {
 509         /*
 510          * If we're not on the boundary we know we've modified the page and we
 511          * need to crc the page.
 512          */
 513         if (io_ctl->cur != io_ctl->orig)
 514                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 515         else
 516                 io_ctl_unmap_page(io_ctl);
 517
 518         while (io_ctl->index < io_ctl->num_pages) {
 519                 io_ctl_map_page(io_ctl, 1);
 520                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 521         }
 522 }
 523
 524 static int io_ctl_read_entry(struct io_ctl *io_ctl,
 525                             struct btrfs_free_space *entry, u8 *type)
 526 {
 527         struct btrfs_free_space_entry *e;
 528
 529         e = io_ctl->cur;
 530         entry->offset = le64_to_cpu(e->offset);
 531         entry->bytes = le64_to_cpu(e->bytes);
 532         *type = e->type;
 533         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 534         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 535
 536         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 537                 return 0;
 538
 539         io_ctl_unmap_page(io_ctl);
 540
 541         if (io_ctl->index >= io_ctl->num_pages)
 542                 return 0;
 543
 544         return io_ctl_check_crc(io_ctl, io_ctl->index);
 545 }
 546
 547 static int io_ctl_read_bitmap(struct io_ctl *io_ctl,
 548                               struct btrfs_free_space *entry)
 549 {
 550         int ret;
 551
 552         if (io_ctl->cur && io_ctl->cur != io_ctl->orig)
 553                 io_ctl_unmap_page(io_ctl);
 554
 555         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 556         if (ret)
 557                 return ret;
 558
 559         memcpy(entry->bitmap, io_ctl->cur, PAGE_CACHE_SIZE);
 560         io_ctl_unmap_page(io_ctl);
 561
 562         return 0;
 563 }
 564
 565 int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 566                             struct btrfs_free_space_ctl *ctl,
 567                             struct btrfs_path *path, u64 offset)
 568 {
 569         struct btrfs_free_space_header *header;
 570         struct extent_buffer *leaf;
 571         struct io_ctl io_ctl;
 572         struct btrfs_key key;
 573         struct btrfs_free_space *e, *n;
 574         struct list_head bitmaps;
 575         u64 num_entries;
 576         u64 num_bitmaps;
 577         u64 generation;
 578         u8 type;
 579         int ret = 0;
 580
 581         INIT_LIST_HEAD(&bitmaps);
 582
 583         /* Nothing in the space cache, goodbye */
 584         if (!i_size_read(inode))
 585                 return 0;
 586
 587         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 588         key.offset = offset;
 589         key.type = 0;
 590
 591         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 592         if (ret < 0)
 593                 return 0;
 594         else if (ret > 0) {
 595                 btrfs_release_path(path);
 596                 return 0;
 597         }
 598
 599         ret = -1;
 600
 601         leaf = path->nodes[0];
 602         header = btrfs_item_ptr(leaf, path->slots[0],
 603                                 struct btrfs_free_space_header);
 604         num_entries = btrfs_free_space_entries(leaf, header);
 605         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 606         generation = btrfs_free_space_generation(leaf, header);
 607         btrfs_release_path(path);
 608
 609         if (BTRFS_I(inode)->generation != generation) {
 610                 printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
 611                        " not match free space cache generation (%llu)\n",
 612                        (unsigned long long)BTRFS_I(inode)->generation,
 613                        (unsigned long long)generation);
 614                 return 0;
 615         }
 616
 617         if (!num_entries)
 618                 return 0;
 619
 620         io_ctl_init(&io_ctl, inode, root);
 621         ret = readahead_cache(inode);
 622         if (ret)
 623                 goto out;
 624
 625         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 626         if (ret)
 627                 goto out;
 628
 629         ret = io_ctl_check_crc(&io_ctl, 0);
 630         if (ret)
 631                 goto free_cache;
 632
 633         ret = io_ctl_check_generation(&io_ctl, generation);
 634         if (ret)
 635                 goto free_cache;
 636
 637         while (num_entries) {
 638                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 639                                       GFP_NOFS);
 640                 if (!e)
 641                         goto free_cache;
 642
 643                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 644                 if (ret) {
 645                         kmem_cache_free(btrfs_free_space_cachep, e);
 646                         goto free_cache;
 647                 }
 648
 649                 if (!e->bytes) {
 650                         kmem_cache_free(btrfs_free_space_cachep, e);
 651                         goto free_cache;
 652                 }
 653
 654                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 655                         spin_lock(&ctl->tree_lock);
 656                         ret = link_free_space(ctl, e);
 657                         spin_unlock(&ctl->tree_lock);
 658                         if (ret) {
 659                                 printk(KERN_ERR "Duplicate entries in "
 660                                        "free space cache, dumping\n");
 661                                 kmem_cache_free(btrfs_free_space_cachep, e);
 662                                 goto free_cache;
 663                         }
 664                 } else {
 665                         BUG_ON(!num_bitmaps);
 666                         num_bitmaps--;
 667                         e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
 668                         if (!e->bitmap) {
 669                                 kmem_cache_free(
 670                                         btrfs_free_space_cachep, e);
 671                                 goto free_cache;
 672                         }
 673                         spin_lock(&ctl->tree_lock);
 674                         ret = link_free_space(ctl, e);
 675                         ctl->total_bitmaps++;
 676                         ctl->op->recalc_thresholds(ctl);
 677                         spin_unlock(&ctl->tree_lock);
 678                         if (ret) {
 679                                 printk(KERN_ERR "Duplicate entries in "
 680                                        "free space cache, dumping\n");
 681                                 kmem_cache_free(btrfs_free_space_cachep, e);
 682                                 goto free_cache;
 683                         }
 684                         list_add_tail(&e->list, &bitmaps);
 685                 }
 686
 687                 num_entries--;
 688         }
 689
 690         /*
 691          * We add the bitmaps at the end of the entries in order that
 692          * the bitmap entries are added to the cache.
 693          */
 694         list_for_each_entry_safe(e, n, &bitmaps, list) {
 695                 list_del_init(&e->list);
 696                 ret = io_ctl_read_bitmap(&io_ctl, e);
 697                 if (ret)
 698                         goto free_cache;
 699         }
 700
 701         io_ctl_drop_pages(&io_ctl);
 702         ret = 1;
 703 out:
 704         io_ctl_free(&io_ctl);
 705         return ret;
 706 free_cache:
 707         io_ctl_drop_pages(&io_ctl);
 708         __btrfs_remove_free_space_cache(ctl);
 709         goto out;
 710 }
 711
 712 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 713                           struct btrfs_block_group_cache *block_group)
 714 {
 715         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 716         struct btrfs_root *root = fs_info->tree_root;
 717         struct inode *inode;
 718         struct btrfs_path *path;
 719         int ret = 0;
 720         bool matched;
 721         u64 used = btrfs_block_group_used(&block_group->item);
 722
 723         /*
 724          * If we're unmounting then just return, since this does a search on the
 725          * normal root and not the commit root and we could deadlock.
 726          */
 727         if (btrfs_fs_closing(fs_info))
 728                 return 0;
 729
 730         /*
 731          * If this block group has been marked to be cleared for one reason or
 732          * another then we can't trust the on disk cache, so just return.
 733          */
 734         spin_lock(&block_group->lock);
 735         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 736                 spin_unlock(&block_group->lock);
 737                 return 0;
 738         }
 739         spin_unlock(&block_group->lock);
 740
 741         path = btrfs_alloc_path();
 742         if (!path)
 743                 return 0;
 744
 745         inode = lookup_free_space_inode(root, block_group, path);
 746         if (IS_ERR(inode)) {
 747                 btrfs_free_path(path);
 748                 return 0;
 749         }
 750
 751         /* We may have converted the inode and made the cache invalid. */
 752         spin_lock(&block_group->lock);
 753         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 754                 spin_unlock(&block_group->lock);
 755                 goto out;
 756         }
 757         spin_unlock(&block_group->lock);
 758
 759         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 760                                       path, block_group->key.objectid);
 761         btrfs_free_path(path);
 762         if (ret <= 0)
 763                 goto out;
 764
 765         spin_lock(&ctl->tree_lock);
 766         matched = (ctl->free_space == (block_group->key.offset - used -
 767                                        block_group->bytes_super));
 768         spin_unlock(&ctl->tree_lock);
 769
 770         if (!matched) {
 771                 __btrfs_remove_free_space_cache(ctl);
 772                 printk(KERN_ERR "block group %llu has an wrong amount of free "
 773                        "space\n", block_group->key.objectid);
 774                 ret = -1;
 775         }
 776 out:
 777         if (ret < 0) {
 778                 /* This cache is bogus, make sure it gets cleared */
 779                 spin_lock(&block_group->lock);
 780                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 781                 spin_unlock(&block_group->lock);
 782                 ret = 0;
 783
 784                 printk(KERN_ERR "btrfs: failed to load free space cache "
 785                        "for block group %llu\n", block_group->key.objectid);
 786         }
 787
 788         iput(inode);
 789         return ret;
 790 }
 791
 792 /**
 793  * __btrfs_write_out_cache - write out cached info to an inode
 794  * @root - the root the inode belongs to
 795  * @ctl - the free space cache we are going to write out
 796  * @block_group - the block_group for this cache if it belongs to a block_group
 797  * @trans - the trans handle
 798  * @path - the path to use
 799  * @offset - the offset for the key we'll insert
 800  *
 801  * This function writes out a free space cache struct to disk for quick recovery
 802  * on mount.  This will return 0 if it was successfull in writing the cache out,
 803  * and -1 if it was not.
 804  */
 805 int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
 806                             struct btrfs_free_space_ctl *ctl,
 807                             struct btrfs_block_group_cache *block_group,
 808                             struct btrfs_trans_handle *trans,
 809                             struct btrfs_path *path, u64 offset)
 810 {
 811         struct btrfs_free_space_header *header;
 812         struct extent_buffer *leaf;
 813         struct rb_node *node;
 814         struct list_head *pos, *n;
 815         struct extent_state *cached_state = NULL;
 816         struct btrfs_free_cluster *cluster = NULL;
 817         struct extent_io_tree *unpin = NULL;
 818         struct io_ctl io_ctl;
 819         struct list_head bitmap_list;
 820         struct btrfs_key key;
 821         u64 start, end, len;
 822         int entries = 0;
 823         int bitmaps = 0;
 824         int ret;
 825         int err = -1;
 826
 827         INIT_LIST_HEAD(&bitmap_list);
 828
 829         if (!i_size_read(inode))
 830                 return -1;
 831
 832         filemap_write_and_wait(inode->i_mapping);
 833         btrfs_wait_ordered_range(inode, inode->i_size &
 834                                  ~(root->sectorsize - 1), (u64)-1);
 835
 836         io_ctl_init(&io_ctl, inode, root);
 837
 838         /* Get the cluster for this block_group if it exists */
 839         if (block_group && !list_empty(&block_group->cluster_list))
 840                 cluster = list_entry(block_group->cluster_list.next,
 841                                      struct btrfs_free_cluster,
 842                                      block_group_list);
 843
 844         /*
 845          * We shouldn't have switched the pinned extents yet so this is the
 846          * right one
 847          */
 848         unpin = root->fs_info->pinned_extents;
 849
 850         /* Lock all pages first so we can lock the extent safely. */
 851         io_ctl_prepare_pages(&io_ctl, inode, 0);
 852
 853         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
 854                          0, &cached_state, GFP_NOFS);
 855
 856         /*
 857          * When searching for pinned extents, we need to start at our start
 858          * offset.
 859          */
 860         if (block_group)
 861                 start = block_group->key.objectid;
 862
 863         node = rb_first(&ctl->free_space_offset);
 864         if (!node && cluster) {
 865                 node = rb_first(&cluster->root);
 866                 cluster = NULL;
 867         }
 868
 869         /* Make sure we can fit our crcs into the first page */
 870         if (io_ctl.check_crcs &&
 871             (io_ctl.num_pages * sizeof(u32)) >= PAGE_CACHE_SIZE) {
 872                 WARN_ON(1);
 873                 goto out_nospc;
 874         }
 875
 876         io_ctl_set_generation(&io_ctl, trans->transid);
 877
 878         /* Write out the extent entries */
 879         while (node) {
 880                 struct btrfs_free_space *e;
 881
 882                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 883                 entries++;
 884
 885                 ret = io_ctl_add_entry(&io_ctl, e->offset, e->bytes,
 886                                        e->bitmap);
 887                 if (ret)
 888                         goto out_nospc;
 889
 890                 if (e->bitmap) {
 891                         list_add_tail(&e->list, &bitmap_list);
 892                         bitmaps++;
 893                 }
 894                 node = rb_next(node);
 895                 if (!node && cluster) {
 896                         node = rb_first(&cluster->root);
 897                         cluster = NULL;
 898                 }
 899         }
 900
 901         /*
 902          * We want to add any pinned extents to our free space cache
 903          * so we don't leak the space
 904          */
 905         while (block_group && (start < block_group->key.objectid +
 906                                block_group->key.offset)) {
 907                 ret = find_first_extent_bit(unpin, start, &start, &end,
 908                                             EXTENT_DIRTY);
 909                 if (ret) {
 910                         ret = 0;
 911                         break;
 912                 }
 913
 914                 /* This pinned extent is out of our range */
 915                 if (start >= block_group->key.objectid +
 916                     block_group->key.offset)
 917                         break;
 918
 919                 len = block_group->key.objectid +
 920                         block_group->key.offset - start;
 921                 len = min(len, end + 1 - start);
 922
 923                 entries++;
 924                 ret = io_ctl_add_entry(&io_ctl, start, len, NULL);
 925                 if (ret)
 926                         goto out_nospc;
 927
 928                 start = end + 1;
 929         }
 930
 931         /* Write out the bitmaps */
 932         list_for_each_safe(pos, n, &bitmap_list) {
 933                 struct btrfs_free_space *entry =
 934                         list_entry(pos, struct btrfs_free_space, list);
 935
 936                 ret = io_ctl_add_bitmap(&io_ctl, entry->bitmap);
 937                 if (ret)
 938                         goto out_nospc;
 939                 list_del_init(&entry->list);
 940         }
 941
 942         /* Zero out the rest of the pages just to make sure */
 943         io_ctl_zero_remaining_pages(&io_ctl);
 944
 945         ret = btrfs_dirty_pages(root, inode, io_ctl.pages, io_ctl.num_pages,
 946                                 0, i_size_read(inode), &cached_state);
 947         io_ctl_drop_pages(&io_ctl);
 948         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
 949                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
 950
 951         if (ret)
 952                 goto out;
 953
 954
 955         ret = filemap_write_and_wait(inode->i_mapping);
 956         if (ret)
 957                 goto out;
 958
 959         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 960         key.offset = offset;
 961         key.type = 0;
 962
 963         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 964         if (ret < 0) {
 965                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
 966                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
 967                                  GFP_NOFS);
 968                 goto out;
 969         }
 970         leaf = path->nodes[0];
 971         if (ret > 0) {
 972                 struct btrfs_key found_key;
 973                 BUG_ON(!path->slots[0]);
 974                 path->slots[0]--;
 975                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
 976                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
 977                     found_key.offset != offset) {
 978                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
 979                                          inode->i_size - 1,
 980                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
 981                                          NULL, GFP_NOFS);
 982                         btrfs_release_path(path);
 983                         goto out;
 984                 }
 985         }
 986
 987         BTRFS_I(inode)->generation = trans->transid;
 988         header = btrfs_item_ptr(leaf, path->slots[0],
 989                                 struct btrfs_free_space_header);
 990         btrfs_set_free_space_entries(leaf, header, entries);
 991         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
 992         btrfs_set_free_space_generation(leaf, header, trans->transid);
 993         btrfs_mark_buffer_dirty(leaf);
 994         btrfs_release_path(path);
 995
 996         err = 0;
 997 out:
 998         io_ctl_free(&io_ctl);
 999         if (err) {
1000                 invalidate_inode_pages2(inode->i_mapping);
1001                 BTRFS_I(inode)->generation = 0;
1002         }
1003         btrfs_update_inode(trans, root, inode);
1004         return err;
1005
1006 out_nospc:
1007         list_for_each_safe(pos, n, &bitmap_list) {
1008                 struct btrfs_free_space *entry =
1009                         list_entry(pos, struct btrfs_free_space, list);
1010                 list_del_init(&entry->list);
1011         }
1012         io_ctl_drop_pages(&io_ctl);
1013         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1014                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1015         goto out;
1016 }
1017
1018 int btrfs_write_out_cache(struct btrfs_root *root,
1019                           struct btrfs_trans_handle *trans,
1020                           struct btrfs_block_group_cache *block_group,
1021                           struct btrfs_path *path)
1022 {
1023         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1024         struct inode *inode;
1025         int ret = 0;
1026
1027         root = root->fs_info->tree_root;
1028
1029         spin_lock(&block_group->lock);
1030         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1031                 spin_unlock(&block_group->lock);
1032                 return 0;
1033         }
1034         spin_unlock(&block_group->lock);
1035
1036         inode = lookup_free_space_inode(root, block_group, path);
1037         if (IS_ERR(inode))
1038                 return 0;
1039
1040         ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
1041                                       path, block_group->key.objectid);
1042         if (ret) {
1043                 spin_lock(&block_group->lock);
1044                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1045                 spin_unlock(&block_group->lock);
1046                 ret = 0;
1047 #ifdef DEBUG
1048                 printk(KERN_ERR "btrfs: failed to write free space cace "
1049                        "for block group %llu\n", block_group->key.objectid);
1050 #endif
1051         }
1052
1053         iput(inode);
1054         return ret;
1055 }
1056
1057 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1058                                           u64 offset)
1059 {
1060         BUG_ON(offset < bitmap_start);
1061         offset -= bitmap_start;
1062         return (unsigned long)(div_u64(offset, unit));
1063 }
1064
1065 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1066 {
1067         return (unsigned long)(div_u64(bytes, unit));
1068 }
1069
1070 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1071                                    u64 offset)
1072 {
1073         u64 bitmap_start;
1074         u64 bytes_per_bitmap;
1075
1076         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1077         bitmap_start = offset - ctl->start;
1078         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1079         bitmap_start *= bytes_per_bitmap;
1080         bitmap_start += ctl->start;
1081
1082         return bitmap_start;
1083 }
1084
1085 static int tree_insert_offset(struct rb_root *root, u64 offset,
1086                               struct rb_node *node, int bitmap)
1087 {
1088         struct rb_node **p = &root->rb_node;
1089         struct rb_node *parent = NULL;
1090         struct btrfs_free_space *info;
1091
1092         while (*p) {
1093                 parent = *p;
1094                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1095
1096                 if (offset < info->offset) {
1097                         p = &(*p)->rb_left;
1098                 } else if (offset > info->offset) {
1099                         p = &(*p)->rb_right;
1100                 } else {
1101                         /*
1102                          * we could have a bitmap entry and an extent entry
1103                          * share the same offset.  If this is the case, we want
1104                          * the extent entry to always be found first if we do a
1105                          * linear search through the tree, since we want to have
1106                          * the quickest allocation time, and allocating from an
1107                          * extent is faster than allocating from a bitmap.  So
1108                          * if we're inserting a bitmap and we find an entry at
1109                          * this offset, we want to go right, or after this entry
1110                          * logically.  If we are inserting an extent and we've
1111                          * found a bitmap, we want to go left, or before
1112                          * logically.
1113                          */
1114                         if (bitmap) {
1115                                 if (info->bitmap) {
1116                                         WARN_ON_ONCE(1);
1117                                         return -EEXIST;
1118                                 }
1119                                 p = &(*p)->rb_right;
1120                         } else {
1121                                 if (!info->bitmap) {
1122                                         WARN_ON_ONCE(1);
1123                                         return -EEXIST;
1124                                 }
1125                                 p = &(*p)->rb_left;
1126                         }
1127                 }
1128         }
1129
1130         rb_link_node(node, parent, p);
1131         rb_insert_color(node, root);
1132
1133         return 0;
1134 }
1135
1136 /*
1137  * searches the tree for the given offset.
1138  *
1139  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1140  * want a section that has at least bytes size and comes at or after the given
1141  * offset.
1142  */
1143 static struct btrfs_free_space *
1144 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1145                    u64 offset, int bitmap_only, int fuzzy)
1146 {
1147         struct rb_node *n = ctl->free_space_offset.rb_node;
1148         struct btrfs_free_space *entry, *prev = NULL;
1149
1150         /* find entry that is closest to the 'offset' */
1151         while (1) {
1152                 if (!n) {
1153                         entry = NULL;
1154                         break;
1155                 }
1156
1157                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1158                 prev = entry;
1159
1160                 if (offset < entry->offset)
1161                         n = n->rb_left;
1162                 else if (offset > entry->offset)
1163                         n = n->rb_right;
1164                 else
1165                         break;
1166         }
1167
1168         if (bitmap_only) {
1169                 if (!entry)
1170                         return NULL;
1171                 if (entry->bitmap)
1172                         return entry;
1173
1174                 /*
1175                  * bitmap entry and extent entry may share same offset,
1176                  * in that case, bitmap entry comes after extent entry.
1177                  */
1178                 n = rb_next(n);
1179                 if (!n)
1180                         return NULL;
1181                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1182                 if (entry->offset != offset)
1183                         return NULL;
1184
1185                 WARN_ON(!entry->bitmap);
1186                 return entry;
1187         } else if (entry) {
1188                 if (entry->bitmap) {
1189                         /*
1190                          * if previous extent entry covers the offset,
1191                          * we should return it instead of the bitmap entry
1192                          */
1193                         n = &entry->offset_index;
1194                         while (1) {
1195                                 n = rb_prev(n);
1196                                 if (!n)
1197                                         break;
1198                                 prev = rb_entry(n, struct btrfs_free_space,
1199                                                 offset_index);
1200                                 if (!prev->bitmap) {
1201                                         if (prev->offset + prev->bytes > offset)
1202                                                 entry = prev;
1203                                         break;
1204                                 }
1205                         }
1206                 }
1207                 return entry;
1208         }
1209
1210         if (!prev)
1211                 return NULL;
1212
1213         /* find last entry before the 'offset' */
1214         entry = prev;
1215         if (entry->offset > offset) {
1216                 n = rb_prev(&entry->offset_index);
1217                 if (n) {
1218                         entry = rb_entry(n, struct btrfs_free_space,
1219                                         offset_index);
1220                         BUG_ON(entry->offset > offset);
1221                 } else {
1222                         if (fuzzy)
1223                                 return entry;
1224                         else
1225                                 return NULL;
1226                 }
1227         }
1228
1229         if (entry->bitmap) {
1230                 n = &entry->offset_index;
1231                 while (1) {
1232                         n = rb_prev(n);
1233                         if (!n)
1234                                 break;
1235                         prev = rb_entry(n, struct btrfs_free_space,
1236                                         offset_index);
1237                         if (!prev->bitmap) {
1238                                 if (prev->offset + prev->bytes > offset)
1239                                         return prev;
1240                                 break;
1241                         }
1242                 }
1243                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1244                         return entry;
1245         } else if (entry->offset + entry->bytes > offset)
1246                 return entry;
1247
1248         if (!fuzzy)
1249                 return NULL;
1250
1251         while (1) {
1252                 if (entry->bitmap) {
1253                         if (entry->offset + BITS_PER_BITMAP *
1254                             ctl->unit > offset)
1255                                 break;
1256                 } else {
1257                         if (entry->offset + entry->bytes > offset)
1258                                 break;
1259                 }
1260
1261                 n = rb_next(&entry->offset_index);
1262                 if (!n)
1263                         return NULL;
1264                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1265         }
1266         return entry;
1267 }
1268
1269 static inline void
1270 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1271                     struct btrfs_free_space *info)
1272 {
1273         rb_erase(&info->offset_index, &ctl->free_space_offset);
1274         ctl->free_extents--;
1275 }
1276
1277 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1278                               struct btrfs_free_space *info)
1279 {
1280         __unlink_free_space(ctl, info);
1281         ctl->free_space -= info->bytes;
1282 }
1283
1284 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1285                            struct btrfs_free_space *info)
1286 {
1287         int ret = 0;
1288
1289         BUG_ON(!info->bitmap && !info->bytes);
1290         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1291                                  &info->offset_index, (info->bitmap != NULL));
1292         if (ret)
1293                 return ret;
1294
1295         ctl->free_space += info->bytes;
1296         ctl->free_extents++;
1297         return ret;
1298 }
1299
1300 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1301 {
1302         struct btrfs_block_group_cache *block_group = ctl->private;
1303         u64 max_bytes;
1304         u64 bitmap_bytes;
1305         u64 extent_bytes;
1306         u64 size = block_group->key.offset;
1307         u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
1308         int max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1309
1310         BUG_ON(ctl->total_bitmaps > max_bitmaps);
1311
1312         /*
1313          * The goal is to keep the total amount of memory used per 1gb of space
1314          * at or below 32k, so we need to adjust how much memory we allow to be
1315          * used by extent based free space tracking
1316          */
1317         if (size < 1024 * 1024 * 1024)
1318                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1319         else
1320                 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1321                         div64_u64(size, 1024 * 1024 * 1024);
1322
1323         /*
1324          * we want to account for 1 more bitmap than what we have so we can make
1325          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1326          * we add more bitmaps.
1327          */
1328         bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1329
1330         if (bitmap_bytes >= max_bytes) {
1331                 ctl->extents_thresh = 0;
1332                 return;
1333         }
1334
1335         /*
1336          * we want the extent entry threshold to always be at most 1/2 the maxw
1337          * bytes we can have, or whatever is less than that.
1338          */
1339         extent_bytes = max_bytes - bitmap_bytes;
1340         extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1341
1342         ctl->extents_thresh =
1343                 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1344 }
1345
1346 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1347                                        struct btrfs_free_space *info,
1348                                        u64 offset, u64 bytes)
1349 {
1350         unsigned long start, count;
1351
1352         start = offset_to_bit(info->offset, ctl->unit, offset);
1353         count = bytes_to_bits(bytes, ctl->unit);
1354         BUG_ON(start + count > BITS_PER_BITMAP);
1355
1356         bitmap_clear(info->bitmap, start, count);
1357
1358         info->bytes -= bytes;
1359 }
1360
1361 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1362                               struct btrfs_free_space *info, u64 offset,
1363                               u64 bytes)
1364 {
1365         __bitmap_clear_bits(ctl, info, offset, bytes);
1366         ctl->free_space -= bytes;
1367 }
1368
1369 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1370                             struct btrfs_free_space *info, u64 offset,
1371                             u64 bytes)
1372 {
1373         unsigned long start, count;
1374
1375         start = offset_to_bit(info->offset, ctl->unit, offset);
1376         count = bytes_to_bits(bytes, ctl->unit);
1377         BUG_ON(start + count > BITS_PER_BITMAP);
1378
1379         bitmap_set(info->bitmap, start, count);
1380
1381         info->bytes += bytes;
1382         ctl->free_space += bytes;
1383 }
1384
1385 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1386                          struct btrfs_free_space *bitmap_info, u64 *offset,
1387                          u64 *bytes)
1388 {
1389         unsigned long found_bits = 0;
1390         unsigned long bits, i;
1391         unsigned long next_zero;
1392
1393         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1394                           max_t(u64, *offset, bitmap_info->offset));
1395         bits = bytes_to_bits(*bytes, ctl->unit);
1396
1397         for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
1398              i < BITS_PER_BITMAP;
1399              i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
1400                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1401                                                BITS_PER_BITMAP, i);
1402                 if ((next_zero - i) >= bits) {
1403                         found_bits = next_zero - i;
1404                         break;
1405                 }
1406                 i = next_zero;
1407         }
1408
1409         if (found_bits) {
1410                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1411                 *bytes = (u64)(found_bits) * ctl->unit;
1412                 return 0;
1413         }
1414
1415         return -1;
1416 }
1417
1418 static struct btrfs_free_space *
1419 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes)
1420 {
1421         struct btrfs_free_space *entry;
1422         struct rb_node *node;
1423         int ret;
1424
1425         if (!ctl->free_space_offset.rb_node)
1426                 return NULL;
1427
1428         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1429         if (!entry)
1430                 return NULL;
1431
1432         for (node = &entry->offset_index; node; node = rb_next(node)) {
1433                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1434                 if (entry->bytes < *bytes)
1435                         continue;
1436
1437                 if (entry->bitmap) {
1438                         ret = search_bitmap(ctl, entry, offset, bytes);
1439                         if (!ret)
1440                                 return entry;
1441                         continue;
1442                 }
1443
1444                 *offset = entry->offset;
1445                 *bytes = entry->bytes;
1446                 return entry;
1447         }
1448
1449         return NULL;
1450 }
1451
1452 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1453                            struct btrfs_free_space *info, u64 offset)
1454 {
1455         info->offset = offset_to_bitmap(ctl, offset);
1456         info->bytes = 0;
1457         link_free_space(ctl, info);
1458         ctl->total_bitmaps++;
1459
1460         ctl->op->recalc_thresholds(ctl);
1461 }
1462
1463 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1464                         struct btrfs_free_space *bitmap_info)
1465 {
1466         unlink_free_space(ctl, bitmap_info);
1467         kfree(bitmap_info->bitmap);
1468         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1469         ctl->total_bitmaps--;
1470         ctl->op->recalc_thresholds(ctl);
1471 }
1472
1473 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1474                               struct btrfs_free_space *bitmap_info,
1475                               u64 *offset, u64 *bytes)
1476 {
1477         u64 end;
1478         u64 search_start, search_bytes;
1479         int ret;
1480
1481 again:
1482         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1483
1484         /*
1485          * XXX - this can go away after a few releases.
1486          *
1487          * since the only user of btrfs_remove_free_space is the tree logging
1488          * stuff, and the only way to test that is under crash conditions, we
1489          * want to have this debug stuff here just in case somethings not
1490          * working.  Search the bitmap for the space we are trying to use to
1491          * make sure its actually there.  If its not there then we need to stop
1492          * because something has gone wrong.
1493          */
1494         search_start = *offset;
1495         search_bytes = *bytes;
1496         search_bytes = min(search_bytes, end - search_start + 1);
1497         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1498         BUG_ON(ret < 0 || search_start != *offset);
1499
1500         if (*offset > bitmap_info->offset && *offset + *bytes > end) {
1501                 bitmap_clear_bits(ctl, bitmap_info, *offset, end - *offset + 1);
1502                 *bytes -= end - *offset + 1;
1503                 *offset = end + 1;
1504         } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
1505                 bitmap_clear_bits(ctl, bitmap_info, *offset, *bytes);
1506                 *bytes = 0;
1507         }
1508
1509         if (*bytes) {
1510                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1511                 if (!bitmap_info->bytes)
1512                         free_bitmap(ctl, bitmap_info);
1513
1514                 /*
1515                  * no entry after this bitmap, but we still have bytes to
1516                  * remove, so something has gone wrong.
1517                  */
1518                 if (!next)
1519                         return -EINVAL;
1520
1521                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1522                                        offset_index);
1523
1524                 /*
1525                  * if the next entry isn't a bitmap we need to return to let the
1526                  * extent stuff do its work.
1527                  */
1528                 if (!bitmap_info->bitmap)
1529                         return -EAGAIN;
1530
1531                 /*
1532                  * Ok the next item is a bitmap, but it may not actually hold
1533                  * the information for the rest of this free space stuff, so
1534                  * look for it, and if we don't find it return so we can try
1535                  * everything over again.
1536                  */
1537                 search_start = *offset;
1538                 search_bytes = *bytes;
1539                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1540                                     &search_bytes);
1541                 if (ret < 0 || search_start != *offset)
1542                         return -EAGAIN;
1543
1544                 goto again;
1545         } else if (!bitmap_info->bytes)
1546                 free_bitmap(ctl, bitmap_info);
1547
1548         return 0;
1549 }
1550
1551 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1552                                struct btrfs_free_space *info, u64 offset,
1553                                u64 bytes)
1554 {
1555         u64 bytes_to_set = 0;
1556         u64 end;
1557
1558         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1559
1560         bytes_to_set = min(end - offset, bytes);
1561
1562         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1563
1564         return bytes_to_set;
1565
1566 }
1567
1568 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1569                       struct btrfs_free_space *info)
1570 {
1571         struct btrfs_block_group_cache *block_group = ctl->private;
1572
1573         /*
1574          * If we are below the extents threshold then we can add this as an
1575          * extent, and don't have to deal with the bitmap
1576          */
1577         if (ctl->free_extents < ctl->extents_thresh) {
1578                 /*
1579                  * If this block group has some small extents we don't want to
1580                  * use up all of our free slots in the cache with them, we want
1581                  * to reserve them to larger extents, however if we have plent
1582                  * of cache left then go ahead an dadd them, no sense in adding
1583                  * the overhead of a bitmap if we don't have to.
1584                  */
1585                 if (info->bytes <= block_group->sectorsize * 4) {
1586                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
1587                                 return false;
1588                 } else {
1589                         return false;
1590                 }
1591         }
1592
1593         /*
1594          * some block groups are so tiny they can't be enveloped by a bitmap, so
1595          * don't even bother to create a bitmap for this
1596          */
1597         if (BITS_PER_BITMAP * block_group->sectorsize >
1598             block_group->key.offset)
1599                 return false;
1600
1601         return true;
1602 }
1603
1604 static struct btrfs_free_space_op free_space_op = {
1605         .recalc_thresholds      = recalculate_thresholds,
1606         .use_bitmap             = use_bitmap,
1607 };
1608
1609 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1610                               struct btrfs_free_space *info)
1611 {
1612         struct btrfs_free_space *bitmap_info;
1613         struct btrfs_block_group_cache *block_group = NULL;
1614         int added = 0;
1615         u64 bytes, offset, bytes_added;
1616         int ret;
1617
1618         bytes = info->bytes;
1619         offset = info->offset;
1620
1621         if (!ctl->op->use_bitmap(ctl, info))
1622                 return 0;
1623
1624         if (ctl->op == &free_space_op)
1625                 block_group = ctl->private;
1626 again:
1627         /*
1628          * Since we link bitmaps right into the cluster we need to see if we
1629          * have a cluster here, and if so and it has our bitmap we need to add
1630          * the free space to that bitmap.
1631          */
1632         if (block_group && !list_empty(&block_group->cluster_list)) {
1633                 struct btrfs_free_cluster *cluster;
1634                 struct rb_node *node;
1635                 struct btrfs_free_space *entry;
1636
1637                 cluster = list_entry(block_group->cluster_list.next,
1638                                      struct btrfs_free_cluster,
1639                                      block_group_list);
1640                 spin_lock(&cluster->lock);
1641                 node = rb_first(&cluster->root);
1642                 if (!node) {
1643                         spin_unlock(&cluster->lock);
1644                         goto no_cluster_bitmap;
1645                 }
1646
1647                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1648                 if (!entry->bitmap) {
1649                         spin_unlock(&cluster->lock);
1650                         goto no_cluster_bitmap;
1651                 }
1652
1653                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
1654                         bytes_added = add_bytes_to_bitmap(ctl, entry,
1655                                                           offset, bytes);
1656                         bytes -= bytes_added;
1657                         offset += bytes_added;
1658                 }
1659                 spin_unlock(&cluster->lock);
1660                 if (!bytes) {
1661                         ret = 1;
1662                         goto out;
1663                 }
1664         }
1665
1666 no_cluster_bitmap:
1667         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1668                                          1, 0);
1669         if (!bitmap_info) {
1670                 BUG_ON(added);
1671                 goto new_bitmap;
1672         }
1673
1674         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
1675         bytes -= bytes_added;
1676         offset += bytes_added;
1677         added = 0;
1678
1679         if (!bytes) {
1680                 ret = 1;
1681                 goto out;
1682         } else
1683                 goto again;
1684
1685 new_bitmap:
1686         if (info && info->bitmap) {
1687                 add_new_bitmap(ctl, info, offset);
1688                 added = 1;
1689                 info = NULL;
1690                 goto again;
1691         } else {
1692                 spin_unlock(&ctl->tree_lock);
1693
1694                 /* no pre-allocated info, allocate a new one */
1695                 if (!info) {
1696                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
1697                                                  GFP_NOFS);
1698                         if (!info) {
1699                                 spin_lock(&ctl->tree_lock);
1700                                 ret = -ENOMEM;
1701                                 goto out;
1702                         }
1703                 }
1704
1705                 /* allocate the bitmap */
1706                 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1707                 spin_lock(&ctl->tree_lock);
1708                 if (!info->bitmap) {
1709                         ret = -ENOMEM;
1710                         goto out;
1711                 }
1712                 goto again;
1713         }
1714
1715 out:
1716         if (info) {
1717                 if (info->bitmap)
1718                         kfree(info->bitmap);
1719                 kmem_cache_free(btrfs_free_space_cachep, info);
1720         }
1721
1722         return ret;
1723 }
1724
1725 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
1726                           struct btrfs_free_space *info, bool update_stat)
1727 {
1728         struct btrfs_free_space *left_info;
1729         struct btrfs_free_space *right_info;
1730         bool merged = false;
1731         u64 offset = info->offset;
1732         u64 bytes = info->bytes;
1733
1734         /*
1735          * first we want to see if there is free space adjacent to the range we
1736          * are adding, if there is remove that struct and add a new one to
1737          * cover the entire range
1738          */
1739         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
1740         if (right_info && rb_prev(&right_info->offset_index))
1741                 left_info = rb_entry(rb_prev(&right_info->offset_index),
1742                                      struct btrfs_free_space, offset_index);
1743         else
1744                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
1745
1746         if (right_info && !right_info->bitmap) {
1747                 if (update_stat)
1748                         unlink_free_space(ctl, right_info);
1749                 else
1750                         __unlink_free_space(ctl, right_info);
1751                 info->bytes += right_info->bytes;
1752                 kmem_cache_free(btrfs_free_space_cachep, right_info);
1753                 merged = true;
1754         }
1755
1756         if (left_info && !left_info->bitmap &&
1757             left_info->offset + left_info->bytes == offset) {
1758                 if (update_stat)
1759                         unlink_free_space(ctl, left_info);
1760                 else
1761                         __unlink_free_space(ctl, left_info);
1762                 info->offset = left_info->offset;
1763                 info->bytes += left_info->bytes;
1764                 kmem_cache_free(btrfs_free_space_cachep, left_info);
1765                 merged = true;
1766         }
1767
1768         return merged;
1769 }
1770
1771 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
1772                            u64 offset, u64 bytes)
1773 {
1774         struct btrfs_free_space *info;
1775         int ret = 0;
1776
1777         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
1778         if (!info)
1779                 return -ENOMEM;
1780
1781         info->offset = offset;
1782         info->bytes = bytes;
1783
1784         spin_lock(&ctl->tree_lock);
1785
1786         if (try_merge_free_space(ctl, info, true))
1787                 goto link;
1788
1789         /*
1790          * There was no extent directly to the left or right of this new
1791          * extent then we know we're going to have to allocate a new extent, so
1792          * before we do that see if we need to drop this into a bitmap
1793          */
1794         ret = insert_into_bitmap(ctl, info);
1795         if (ret < 0) {
1796                 goto out;
1797         } else if (ret) {
1798                 ret = 0;
1799                 goto out;
1800         }
1801 link:
1802         ret = link_free_space(ctl, info);
1803         if (ret)
1804                 kmem_cache_free(btrfs_free_space_cachep, info);
1805 out:
1806         spin_unlock(&ctl->tree_lock);
1807
1808         if (ret) {
1809                 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1810                 BUG_ON(ret == -EEXIST);
1811         }
1812
1813         return ret;
1814 }
1815
1816 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1817                             u64 offset, u64 bytes)
1818 {
1819         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1820         struct btrfs_free_space *info;
1821         struct btrfs_free_space *next_info = NULL;
1822         int ret = 0;
1823
1824         spin_lock(&ctl->tree_lock);
1825
1826 again:
1827         info = tree_search_offset(ctl, offset, 0, 0);
1828         if (!info) {
1829                 /*
1830                  * oops didn't find an extent that matched the space we wanted
1831                  * to remove, look for a bitmap instead
1832                  */
1833                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1834                                           1, 0);
1835                 if (!info) {
1836                         WARN_ON(1);
1837                         goto out_lock;
1838                 }
1839         }
1840
1841         if (info->bytes < bytes && rb_next(&info->offset_index)) {
1842                 u64 end;
1843                 next_info = rb_entry(rb_next(&info->offset_index),
1844                                              struct btrfs_free_space,
1845                                              offset_index);
1846
1847                 if (next_info->bitmap)
1848                         end = next_info->offset +
1849                               BITS_PER_BITMAP * ctl->unit - 1;
1850                 else
1851                         end = next_info->offset + next_info->bytes;
1852
1853                 if (next_info->bytes < bytes ||
1854                     next_info->offset > offset || offset > end) {
1855                         printk(KERN_CRIT "Found free space at %llu, size %llu,"
1856                               " trying to use %llu\n",
1857                               (unsigned long long)info->offset,
1858                               (unsigned long long)info->bytes,
1859                               (unsigned long long)bytes);
1860                         WARN_ON(1);
1861                         ret = -EINVAL;
1862                         goto out_lock;
1863                 }
1864
1865                 info = next_info;
1866         }
1867
1868         if (info->bytes == bytes) {
1869                 unlink_free_space(ctl, info);
1870                 if (info->bitmap) {
1871                         kfree(info->bitmap);
1872                         ctl->total_bitmaps--;
1873                 }
1874                 kmem_cache_free(btrfs_free_space_cachep, info);
1875                 goto out_lock;
1876         }
1877
1878         if (!info->bitmap && info->offset == offset) {
1879                 unlink_free_space(ctl, info);
1880                 info->offset += bytes;
1881                 info->bytes -= bytes;
1882                 link_free_space(ctl, info);
1883                 goto out_lock;
1884         }
1885
1886         if (!info->bitmap && info->offset <= offset &&
1887             info->offset + info->bytes >= offset + bytes) {
1888                 u64 old_start = info->offset;
1889                 /*
1890                  * we're freeing space in the middle of the info,
1891                  * this can happen during tree log replay
1892                  *
1893                  * first unlink the old info and then
1894                  * insert it again after the hole we're creating
1895                  */
1896                 unlink_free_space(ctl, info);
1897                 if (offset + bytes < info->offset + info->bytes) {
1898                         u64 old_end = info->offset + info->bytes;
1899
1900                         info->offset = offset + bytes;
1901                         info->bytes = old_end - info->offset;
1902                         ret = link_free_space(ctl, info);
1903                         WARN_ON(ret);
1904                         if (ret)
1905                                 goto out_lock;
1906                 } else {
1907                         /* the hole we're creating ends at the end
1908                          * of the info struct, just free the info
1909                          */
1910                         kmem_cache_free(btrfs_free_space_cachep, info);
1911                 }
1912                 spin_unlock(&ctl->tree_lock);
1913
1914                 /* step two, insert a new info struct to cover
1915                  * anything before the hole
1916                  */
1917                 ret = btrfs_add_free_space(block_group, old_start,
1918                                            offset - old_start);
1919                 WARN_ON(ret);
1920                 goto out;
1921         }
1922
1923         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
1924         if (ret == -EAGAIN)
1925                 goto again;
1926         BUG_ON(ret);
1927 out_lock:
1928         spin_unlock(&ctl->tree_lock);
1929 out:
1930         return ret;
1931 }
1932
1933 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1934                            u64 bytes)
1935 {
1936         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1937         struct btrfs_free_space *info;
1938         struct rb_node *n;
1939         int count = 0;
1940
1941         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
1942                 info = rb_entry(n, struct btrfs_free_space, offset_index);
1943                 if (info->bytes >= bytes)
1944                         count++;
1945                 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1946                        (unsigned long long)info->offset,
1947                        (unsigned long long)info->bytes,
1948                        (info->bitmap) ? "yes" : "no");
1949         }
1950         printk(KERN_INFO "block group has cluster?: %s\n",
1951                list_empty(&block_group->cluster_list) ? "no" : "yes");
1952         printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1953                "\n", count);
1954 }
1955
1956 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
1957 {
1958         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1959
1960         spin_lock_init(&ctl->tree_lock);
1961         ctl->unit = block_group->sectorsize;
1962         ctl->start = block_group->key.objectid;
1963         ctl->private = block_group;
1964         ctl->op = &free_space_op;
1965
1966         /*
1967          * we only want to have 32k of ram per block group for keeping
1968          * track of free space, and if we pass 1/2 of that we want to
1969          * start converting things over to using bitmaps
1970          */
1971         ctl->extents_thresh = ((1024 * 32) / 2) /
1972                                 sizeof(struct btrfs_free_space);
1973 }
1974
1975 /*
1976  * for a given cluster, put all of its extents back into the free
1977  * space cache.  If the block group passed doesn't match the block group
1978  * pointed to by the cluster, someone else raced in and freed the
1979  * cluster already.  In that case, we just return without changing anything
1980  */
1981 static int
1982 __btrfs_return_cluster_to_free_space(
1983                              struct btrfs_block_group_cache *block_group,
1984                              struct btrfs_free_cluster *cluster)
1985 {
1986         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1987         struct btrfs_free_space *entry;
1988         struct rb_node *node;
1989
1990         spin_lock(&cluster->lock);
1991         if (cluster->block_group != block_group)
1992                 goto out;
1993
1994         cluster->block_group = NULL;
1995         cluster->window_start = 0;
1996         list_del_init(&cluster->block_group_list);
1997
1998         node = rb_first(&cluster->root);
1999         while (node) {
2000                 bool bitmap;
2001
2002                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2003                 node = rb_next(&entry->offset_index);
2004                 rb_erase(&entry->offset_index, &cluster->root);
2005
2006                 bitmap = (entry->bitmap != NULL);
2007                 if (!bitmap)
2008                         try_merge_free_space(ctl, entry, false);
2009                 tree_insert_offset(&ctl->free_space_offset,
2010                                    entry->offset, &entry->offset_index, bitmap);
2011         }
2012         cluster->root = RB_ROOT;
2013
2014 out:
2015         spin_unlock(&cluster->lock);
2016         btrfs_put_block_group(block_group);
2017         return 0;
2018 }
2019
2020 void __btrfs_remove_free_space_cache_locked(struct btrfs_free_space_ctl *ctl)
2021 {
2022         struct btrfs_free_space *info;
2023         struct rb_node *node;
2024
2025         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2026                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2027                 if (!info->bitmap) {
2028                         unlink_free_space(ctl, info);
2029                         kmem_cache_free(btrfs_free_space_cachep, info);
2030                 } else {
2031                         free_bitmap(ctl, info);
2032                 }
2033                 if (need_resched()) {
2034                         spin_unlock(&ctl->tree_lock);
2035                         cond_resched();
2036                         spin_lock(&ctl->tree_lock);
2037                 }
2038         }
2039 }
2040
2041 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2042 {
2043         spin_lock(&ctl->tree_lock);
2044         __btrfs_remove_free_space_cache_locked(ctl);
2045         spin_unlock(&ctl->tree_lock);
2046 }
2047
2048 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2049 {
2050         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2051         struct btrfs_free_cluster *cluster;
2052         struct list_head *head;
2053
2054         spin_lock(&ctl->tree_lock);
2055         while ((head = block_group->cluster_list.next) !=
2056                &block_group->cluster_list) {
2057                 cluster = list_entry(head, struct btrfs_free_cluster,
2058                                      block_group_list);
2059
2060                 WARN_ON(cluster->block_group != block_group);
2061                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2062                 if (need_resched()) {
2063                         spin_unlock(&ctl->tree_lock);
2064                         cond_resched();
2065                         spin_lock(&ctl->tree_lock);
2066                 }
2067         }
2068         __btrfs_remove_free_space_cache_locked(ctl);
2069         spin_unlock(&ctl->tree_lock);
2070
2071 }
2072
2073 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2074                                u64 offset, u64 bytes, u64 empty_size)
2075 {
2076         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2077         struct btrfs_free_space *entry = NULL;
2078         u64 bytes_search = bytes + empty_size;
2079         u64 ret = 0;
2080
2081         spin_lock(&ctl->tree_lock);
2082         entry = find_free_space(ctl, &offset, &bytes_search);
2083         if (!entry)
2084                 goto out;
2085
2086         ret = offset;
2087         if (entry->bitmap) {
2088                 bitmap_clear_bits(ctl, entry, offset, bytes);
2089                 if (!entry->bytes)
2090                         free_bitmap(ctl, entry);
2091         } else {
2092                 unlink_free_space(ctl, entry);
2093                 entry->offset += bytes;
2094                 entry->bytes -= bytes;
2095                 if (!entry->bytes)
2096                         kmem_cache_free(btrfs_free_space_cachep, entry);
2097                 else
2098                         link_free_space(ctl, entry);
2099         }
2100
2101 out:
2102         spin_unlock(&ctl->tree_lock);
2103
2104         return ret;
2105 }
2106
2107 /*
2108  * given a cluster, put all of its extents back into the free space
2109  * cache.  If a block group is passed, this function will only free
2110  * a cluster that belongs to the passed block group.
2111  *
2112  * Otherwise, it'll get a reference on the block group pointed to by the
2113  * cluster and remove the cluster from it.
2114  */
2115 int btrfs_return_cluster_to_free_space(
2116                                struct btrfs_block_group_cache *block_group,
2117                                struct btrfs_free_cluster *cluster)
2118 {
2119         struct btrfs_free_space_ctl *ctl;
2120         int ret;
2121
2122         /* first, get a safe pointer to the block group */
2123         spin_lock(&cluster->lock);
2124         if (!block_group) {
2125                 block_group = cluster->block_group;
2126                 if (!block_group) {
2127                         spin_unlock(&cluster->lock);
2128                         return 0;
2129                 }
2130         } else if (cluster->block_group != block_group) {
2131                 /* someone else has already freed it don't redo their work */
2132                 spin_unlock(&cluster->lock);
2133                 return 0;
2134         }
2135         atomic_inc(&block_group->count);
2136         spin_unlock(&cluster->lock);
2137
2138         ctl = block_group->free_space_ctl;
2139
2140         /* now return any extents the cluster had on it */
2141         spin_lock(&ctl->tree_lock);
2142         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2143         spin_unlock(&ctl->tree_lock);
2144
2145         /* finally drop our ref */
2146         btrfs_put_block_group(block_group);
2147         return ret;
2148 }
2149
2150 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2151                                    struct btrfs_free_cluster *cluster,
2152                                    struct btrfs_free_space *entry,
2153                                    u64 bytes, u64 min_start)
2154 {
2155         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2156         int err;
2157         u64 search_start = cluster->window_start;
2158         u64 search_bytes = bytes;
2159         u64 ret = 0;
2160
2161         search_start = min_start;
2162         search_bytes = bytes;
2163
2164         err = search_bitmap(ctl, entry, &search_start, &search_bytes);
2165         if (err)
2166                 return 0;
2167
2168         ret = search_start;
2169         __bitmap_clear_bits(ctl, entry, ret, bytes);
2170
2171         return ret;
2172 }
2173
2174 /*
2175  * given a cluster, try to allocate 'bytes' from it, returns 0
2176  * if it couldn't find anything suitably large, or a logical disk offset
2177  * if things worked out
2178  */
2179 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2180                              struct btrfs_free_cluster *cluster, u64 bytes,
2181                              u64 min_start)
2182 {
2183         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2184         struct btrfs_free_space *entry = NULL;
2185         struct rb_node *node;
2186         u64 ret = 0;
2187
2188         spin_lock(&cluster->lock);
2189         if (bytes > cluster->max_size)
2190                 goto out;
2191
2192         if (cluster->block_group != block_group)
2193                 goto out;
2194
2195         node = rb_first(&cluster->root);
2196         if (!node)
2197                 goto out;
2198
2199         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2200         while(1) {
2201                 if (entry->bytes < bytes ||
2202                     (!entry->bitmap && entry->offset < min_start)) {
2203                         node = rb_next(&entry->offset_index);
2204                         if (!node)
2205                                 break;
2206                         entry = rb_entry(node, struct btrfs_free_space,
2207                                          offset_index);
2208                         continue;
2209                 }
2210
2211                 if (entry->bitmap) {
2212                         ret = btrfs_alloc_from_bitmap(block_group,
2213                                                       cluster, entry, bytes,
2214                                                       min_start);
2215                         if (ret == 0) {
2216                                 node = rb_next(&entry->offset_index);
2217                                 if (!node)
2218                                         break;
2219                                 entry = rb_entry(node, struct btrfs_free_space,
2220                                                  offset_index);
2221                                 continue;
2222                         }
2223                 } else {
2224                         ret = entry->offset;
2225
2226                         entry->offset += bytes;
2227                         entry->bytes -= bytes;
2228                 }
2229
2230                 if (entry->bytes == 0)
2231                         rb_erase(&entry->offset_index, &cluster->root);
2232                 break;
2233         }
2234 out:
2235         spin_unlock(&cluster->lock);
2236
2237         if (!ret)
2238                 return 0;
2239
2240         spin_lock(&ctl->tree_lock);
2241
2242         ctl->free_space -= bytes;
2243         if (entry->bytes == 0) {
2244                 ctl->free_extents--;
2245                 if (entry->bitmap) {
2246                         kfree(entry->bitmap);
2247                         ctl->total_bitmaps--;
2248                         ctl->op->recalc_thresholds(ctl);
2249                 }
2250                 kmem_cache_free(btrfs_free_space_cachep, entry);
2251         }
2252
2253         spin_unlock(&ctl->tree_lock);
2254
2255         return ret;
2256 }
2257
2258 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2259                                 struct btrfs_free_space *entry,
2260                                 struct btrfs_free_cluster *cluster,
2261                                 u64 offset, u64 bytes, u64 min_bytes)
2262 {
2263         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2264         unsigned long next_zero;
2265         unsigned long i;
2266         unsigned long search_bits;
2267         unsigned long total_bits;
2268         unsigned long found_bits;
2269         unsigned long start = 0;
2270         unsigned long total_found = 0;
2271         int ret;
2272         bool found = false;
2273
2274         i = offset_to_bit(entry->offset, block_group->sectorsize,
2275                           max_t(u64, offset, entry->offset));
2276         search_bits = bytes_to_bits(bytes, block_group->sectorsize);
2277         total_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
2278
2279 again:
2280         found_bits = 0;
2281         for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
2282              i < BITS_PER_BITMAP;
2283              i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
2284                 next_zero = find_next_zero_bit(entry->bitmap,
2285                                                BITS_PER_BITMAP, i);
2286                 if (next_zero - i >= search_bits) {
2287                         found_bits = next_zero - i;
2288                         break;
2289                 }
2290                 i = next_zero;
2291         }
2292
2293         if (!found_bits)
2294                 return -ENOSPC;
2295
2296         if (!found) {
2297                 start = i;
2298                 found = true;
2299         }
2300
2301         total_found += found_bits;
2302
2303         if (cluster->max_size < found_bits * block_group->sectorsize)
2304                 cluster->max_size = found_bits * block_group->sectorsize;
2305
2306         if (total_found < total_bits) {
2307                 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
2308                 if (i - start > total_bits * 2) {
2309                         total_found = 0;
2310                         cluster->max_size = 0;
2311                         found = false;
2312                 }
2313                 goto again;
2314         }
2315
2316         cluster->window_start = start * block_group->sectorsize +
2317                 entry->offset;
2318         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2319         ret = tree_insert_offset(&cluster->root, entry->offset,
2320                                  &entry->offset_index, 1);
2321         BUG_ON(ret);
2322
2323         return 0;
2324 }
2325
2326 /*
2327  * This searches the block group for just extents to fill the cluster with.
2328  */
2329 static noinline int
2330 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2331                         struct btrfs_free_cluster *cluster,
2332                         struct list_head *bitmaps, u64 offset, u64 bytes,
2333                         u64 min_bytes)
2334 {
2335         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2336         struct btrfs_free_space *first = NULL;
2337         struct btrfs_free_space *entry = NULL;
2338         struct btrfs_free_space *prev = NULL;
2339         struct btrfs_free_space *last;
2340         struct rb_node *node;
2341         u64 window_start;
2342         u64 window_free;
2343         u64 max_extent;
2344         u64 max_gap = 128 * 1024;
2345
2346         entry = tree_search_offset(ctl, offset, 0, 1);
2347         if (!entry)
2348                 return -ENOSPC;
2349
2350         /*
2351          * We don't want bitmaps, so just move along until we find a normal
2352          * extent entry.
2353          */
2354         while (entry->bitmap) {
2355                 if (list_empty(&entry->list))
2356                         list_add_tail(&entry->list, bitmaps);
2357                 node = rb_next(&entry->offset_index);
2358                 if (!node)
2359                         return -ENOSPC;
2360                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2361         }
2362
2363         window_start = entry->offset;
2364         window_free = entry->bytes;
2365         max_extent = entry->bytes;
2366         first = entry;
2367         last = entry;
2368         prev = entry;
2369
2370         while (window_free <= min_bytes) {
2371                 node = rb_next(&entry->offset_index);
2372                 if (!node)
2373                         return -ENOSPC;
2374                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2375
2376                 if (entry->bitmap) {
2377                         if (list_empty(&entry->list))
2378                                 list_add_tail(&entry->list, bitmaps);
2379                         continue;
2380                 }
2381
2382                 /*
2383                  * we haven't filled the empty size and the window is
2384                  * very large.  reset and try again
2385                  */
2386                 if (entry->offset - (prev->offset + prev->bytes) > max_gap ||
2387                     entry->offset - window_start > (min_bytes * 2)) {
2388                         first = entry;
2389                         window_start = entry->offset;
2390                         window_free = entry->bytes;
2391                         last = entry;
2392                         max_extent = entry->bytes;
2393                 } else {
2394                         last = entry;
2395                         window_free += entry->bytes;
2396                         if (entry->bytes > max_extent)
2397                                 max_extent = entry->bytes;
2398                 }
2399                 prev = entry;
2400         }
2401
2402         cluster->window_start = first->offset;
2403
2404         node = &first->offset_index;
2405
2406         /*
2407          * now we've found our entries, pull them out of the free space
2408          * cache and put them into the cluster rbtree
2409          */
2410         do {
2411                 int ret;
2412
2413                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2414                 node = rb_next(&entry->offset_index);
2415                 if (entry->bitmap)
2416                         continue;
2417
2418                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2419                 ret = tree_insert_offset(&cluster->root, entry->offset,
2420                                          &entry->offset_index, 0);
2421                 BUG_ON(ret);
2422         } while (node && entry != last);
2423
2424         cluster->max_size = max_extent;
2425
2426         return 0;
2427 }
2428
2429 /*
2430  * This specifically looks for bitmaps that may work in the cluster, we assume
2431  * that we have already failed to find extents that will work.
2432  */
2433 static noinline int
2434 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2435                      struct btrfs_free_cluster *cluster,
2436                      struct list_head *bitmaps, u64 offset, u64 bytes,
2437                      u64 min_bytes)
2438 {
2439         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2440         struct btrfs_free_space *entry;
2441         struct rb_node *node;
2442         int ret = -ENOSPC;
2443
2444         if (ctl->total_bitmaps == 0)
2445                 return -ENOSPC;
2446
2447         /*
2448          * First check our cached list of bitmaps and see if there is an entry
2449          * here that will work.
2450          */
2451         list_for_each_entry(entry, bitmaps, list) {
2452                 if (entry->bytes < min_bytes)
2453                         continue;
2454                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2455                                            bytes, min_bytes);
2456                 if (!ret)
2457                         return 0;
2458         }
2459
2460         /*
2461          * If we do have entries on our list and we are here then we didn't find
2462          * anything, so go ahead and get the next entry after the last entry in
2463          * this list and start the search from there.
2464          */
2465         if (!list_empty(bitmaps)) {
2466                 entry = list_entry(bitmaps->prev, struct btrfs_free_space,
2467                                    list);
2468                 node = rb_next(&entry->offset_index);
2469                 if (!node)
2470                         return -ENOSPC;
2471                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2472                 goto search;
2473         }
2474
2475         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, offset), 0, 1);
2476         if (!entry)
2477                 return -ENOSPC;
2478
2479 search:
2480         node = &entry->offset_index;
2481         do {
2482                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2483                 node = rb_next(&entry->offset_index);
2484                 if (!entry->bitmap)
2485                         continue;
2486                 if (entry->bytes < min_bytes)
2487                         continue;
2488                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2489                                            bytes, min_bytes);
2490         } while (ret && node);
2491
2492         return ret;
2493 }
2494
2495 /*
2496  * here we try to find a cluster of blocks in a block group.  The goal
2497  * is to find at least bytes free and up to empty_size + bytes free.
2498  * We might not find them all in one contiguous area.
2499  *
2500  * returns zero and sets up cluster if things worked out, otherwise
2501  * it returns -enospc
2502  */
2503 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
2504                              struct btrfs_root *root,
2505                              struct btrfs_block_group_cache *block_group,
2506                              struct btrfs_free_cluster *cluster,
2507                              u64 offset, u64 bytes, u64 empty_size)
2508 {
2509         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2510         struct list_head bitmaps;
2511         struct btrfs_free_space *entry, *tmp;
2512         u64 min_bytes;
2513         int ret;
2514
2515         /* for metadata, allow allocates with more holes */
2516         if (btrfs_test_opt(root, SSD_SPREAD)) {
2517                 min_bytes = bytes + empty_size;
2518         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2519                 /*
2520                  * we want to do larger allocations when we are
2521                  * flushing out the delayed refs, it helps prevent
2522                  * making more work as we go along.
2523                  */
2524                 if (trans->transaction->delayed_refs.flushing)
2525                         min_bytes = max(bytes, (bytes + empty_size) >> 1);
2526                 else
2527                         min_bytes = max(bytes, (bytes + empty_size) >> 4);
2528         } else
2529                 min_bytes = max(bytes, (bytes + empty_size) >> 2);
2530
2531         spin_lock(&ctl->tree_lock);
2532
2533         /*
2534          * If we know we don't have enough space to make a cluster don't even
2535          * bother doing all the work to try and find one.
2536          */
2537         if (ctl->free_space < min_bytes) {
2538                 spin_unlock(&ctl->tree_lock);
2539                 return -ENOSPC;
2540         }
2541
2542         spin_lock(&cluster->lock);
2543
2544         /* someone already found a cluster, hooray */
2545         if (cluster->block_group) {
2546                 ret = 0;
2547                 goto out;
2548         }
2549
2550         INIT_LIST_HEAD(&bitmaps);
2551         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
2552                                       bytes, min_bytes);
2553         if (ret)
2554                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
2555                                            offset, bytes, min_bytes);
2556
2557         /* Clear our temporary list */
2558         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
2559                 list_del_init(&entry->list);
2560
2561         if (!ret) {
2562                 atomic_inc(&block_group->count);
2563                 list_add_tail(&cluster->block_group_list,
2564                               &block_group->cluster_list);
2565                 cluster->block_group = block_group;
2566         }
2567 out:
2568         spin_unlock(&cluster->lock);
2569         spin_unlock(&ctl->tree_lock);
2570
2571         return ret;
2572 }
2573
2574 /*
2575  * simple code to zero out a cluster
2576  */
2577 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2578 {
2579         spin_lock_init(&cluster->lock);
2580         spin_lock_init(&cluster->refill_lock);
2581         cluster->root = RB_ROOT;
2582         cluster->max_size = 0;
2583         INIT_LIST_HEAD(&cluster->block_group_list);
2584         cluster->block_group = NULL;
2585 }
2586
2587 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
2588                            u64 *trimmed, u64 start, u64 end, u64 minlen)
2589 {
2590         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2591         struct btrfs_free_space *entry = NULL;
2592         struct btrfs_fs_info *fs_info = block_group->fs_info;
2593         u64 bytes = 0;
2594         u64 actually_trimmed;
2595         int ret = 0;
2596
2597         *trimmed = 0;
2598
2599         while (start < end) {
2600                 spin_lock(&ctl->tree_lock);
2601
2602                 if (ctl->free_space < minlen) {
2603                         spin_unlock(&ctl->tree_lock);
2604                         break;
2605                 }
2606
2607                 entry = tree_search_offset(ctl, start, 0, 1);
2608                 if (!entry)
2609                         entry = tree_search_offset(ctl,
2610                                                    offset_to_bitmap(ctl, start),
2611                                                    1, 1);
2612
2613                 if (!entry || entry->offset >= end) {
2614                         spin_unlock(&ctl->tree_lock);
2615                         break;
2616                 }
2617
2618                 if (entry->bitmap) {
2619                         ret = search_bitmap(ctl, entry, &start, &bytes);
2620                         if (!ret) {
2621                                 if (start >= end) {
2622                                         spin_unlock(&ctl->tree_lock);
2623                                         break;
2624                                 }
2625                                 bytes = min(bytes, end - start);
2626                                 bitmap_clear_bits(ctl, entry, start, bytes);
2627                                 if (entry->bytes == 0)
2628                                         free_bitmap(ctl, entry);
2629                         } else {
2630                                 start = entry->offset + BITS_PER_BITMAP *
2631                                         block_group->sectorsize;
2632                                 spin_unlock(&ctl->tree_lock);
2633                                 ret = 0;
2634                                 continue;
2635                         }
2636                 } else {
2637                         start = entry->offset;
2638                         bytes = min(entry->bytes, end - start);
2639                         unlink_free_space(ctl, entry);
2640                         kmem_cache_free(btrfs_free_space_cachep, entry);
2641                 }
2642
2643                 spin_unlock(&ctl->tree_lock);
2644
2645                 if (bytes >= minlen) {
2646                         struct btrfs_space_info *space_info;
2647                         int update = 0;
2648
2649                         space_info = block_group->space_info;
2650                         spin_lock(&space_info->lock);
2651                         spin_lock(&block_group->lock);
2652                         if (!block_group->ro) {
2653                                 block_group->reserved += bytes;
2654                                 space_info->bytes_reserved += bytes;
2655                                 update = 1;
2656                         }
2657                         spin_unlock(&block_group->lock);
2658                         spin_unlock(&space_info->lock);
2659
2660                         ret = btrfs_error_discard_extent(fs_info->extent_root,
2661                                                          start,
2662                                                          bytes,
2663                                                          &actually_trimmed);
2664
2665                         btrfs_add_free_space(block_group, start, bytes);
2666                         if (update) {
2667                                 spin_lock(&space_info->lock);
2668                                 spin_lock(&block_group->lock);
2669                                 if (block_group->ro)
2670                                         space_info->bytes_readonly += bytes;
2671                                 block_group->reserved -= bytes;
2672                                 space_info->bytes_reserved -= bytes;
2673                                 spin_unlock(&space_info->lock);
2674                                 spin_unlock(&block_group->lock);
2675                         }
2676
2677                         if (ret)
2678                                 break;
2679                         *trimmed += actually_trimmed;
2680                 }
2681                 start += bytes;
2682                 bytes = 0;
2683
2684                 if (fatal_signal_pending(current)) {
2685                         ret = -ERESTARTSYS;
2686                         break;
2687                 }
2688
2689                 cond_resched();
2690         }
2691
2692         return ret;
2693 }
2694
2695 /*
2696  * Find the left-most item in the cache tree, and then return the
2697  * smallest inode number in the item.
2698  *
2699  * Note: the returned inode number may not be the smallest one in
2700  * the tree, if the left-most item is a bitmap.
2701  */
2702 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
2703 {
2704         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
2705         struct btrfs_free_space *entry = NULL;
2706         u64 ino = 0;
2707
2708         spin_lock(&ctl->tree_lock);
2709
2710         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
2711                 goto out;
2712
2713         entry = rb_entry(rb_first(&ctl->free_space_offset),
2714                          struct btrfs_free_space, offset_index);
2715
2716         if (!entry->bitmap) {
2717                 ino = entry->offset;
2718
2719                 unlink_free_space(ctl, entry);
2720                 entry->offset++;
2721                 entry->bytes--;
2722                 if (!entry->bytes)
2723                         kmem_cache_free(btrfs_free_space_cachep, entry);
2724                 else
2725                         link_free_space(ctl, entry);
2726         } else {
2727                 u64 offset = 0;
2728                 u64 count = 1;
2729                 int ret;
2730
2731                 ret = search_bitmap(ctl, entry, &offset, &count);
2732                 BUG_ON(ret);
2733
2734                 ino = offset;
2735                 bitmap_clear_bits(ctl, entry, offset, 1);
2736                 if (entry->bytes == 0)
2737                         free_bitmap(ctl, entry);
2738         }
2739 out:
2740         spin_unlock(&ctl->tree_lock);
2741
2742         return ino;
2743 }
2744
2745 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
2746                                     struct btrfs_path *path)
2747 {
2748         struct inode *inode = NULL;
2749
2750         spin_lock(&root->cache_lock);
2751         if (root->cache_inode)
2752                 inode = igrab(root->cache_inode);
2753         spin_unlock(&root->cache_lock);
2754         if (inode)
2755                 return inode;
2756
2757         inode = __lookup_free_space_inode(root, path, 0);
2758         if (IS_ERR(inode))
2759                 return inode;
2760
2761         spin_lock(&root->cache_lock);
2762         if (!btrfs_fs_closing(root->fs_info))
2763                 root->cache_inode = igrab(inode);
2764         spin_unlock(&root->cache_lock);
2765
2766         return inode;
2767 }
2768
2769 int create_free_ino_inode(struct btrfs_root *root,
2770                           struct btrfs_trans_handle *trans,
2771                           struct btrfs_path *path)
2772 {
2773         return __create_free_space_inode(root, trans, path,
2774                                          BTRFS_FREE_INO_OBJECTID, 0);
2775 }
2776
2777 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
2778 {
2779         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2780         struct btrfs_path *path;
2781         struct inode *inode;
2782         int ret = 0;
2783         u64 root_gen = btrfs_root_generation(&root->root_item);
2784
2785         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2786                 return 0;
2787
2788         /*
2789          * If we're unmounting then just return, since this does a search on the
2790          * normal root and not the commit root and we could deadlock.
2791          */
2792         if (btrfs_fs_closing(fs_info))
2793                 return 0;
2794
2795         path = btrfs_alloc_path();
2796         if (!path)
2797                 return 0;
2798
2799         inode = lookup_free_ino_inode(root, path);
2800         if (IS_ERR(inode))
2801                 goto out;
2802
2803         if (root_gen != BTRFS_I(inode)->generation)
2804                 goto out_put;
2805
2806         ret = __load_free_space_cache(root, inode, ctl, path, 0);
2807
2808         if (ret < 0)
2809                 printk(KERN_ERR "btrfs: failed to load free ino cache for "
2810                        "root %llu\n", root->root_key.objectid);
2811 out_put:
2812         iput(inode);
2813 out:
2814         btrfs_free_path(path);
2815         return ret;
2816 }
2817
2818 int btrfs_write_out_ino_cache(struct btrfs_root *root,
2819                               struct btrfs_trans_handle *trans,
2820                               struct btrfs_path *path)
2821 {
2822         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2823         struct inode *inode;
2824         int ret;
2825
2826         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2827                 return 0;
2828
2829         inode = lookup_free_ino_inode(root, path);
2830         if (IS_ERR(inode))
2831                 return 0;
2832
2833         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
2834         if (ret) {
2835                 btrfs_delalloc_release_metadata(inode, inode->i_size);
2836 #ifdef DEBUG
2837                 printk(KERN_ERR "btrfs: failed to write free ino cache "
2838                        "for root %llu\n", root->root_key.objectid);
2839 #endif
2840         }
2841
2842         iput(inode);
2843         return ret;
2844 }