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         u64 needed_bytes;
 202         loff_t oldsize;
 203         int ret = 0;
 204
 205         rsv = trans->block_rsv;
 206         trans->block_rsv = &root->fs_info->global_block_rsv;
 207
 208         /* 1 for slack space, 1 for updating the inode */
 209         needed_bytes = btrfs_calc_trunc_metadata_size(root, 1) +
 210                 btrfs_calc_trans_metadata_size(root, 1);
 211
 212         spin_lock(&trans->block_rsv->lock);
 213         if (trans->block_rsv->reserved < needed_bytes) {
 214                 spin_unlock(&trans->block_rsv->lock);
 215                 trans->block_rsv = rsv;
 216                 return -ENOSPC;
 217         }
 218         spin_unlock(&trans->block_rsv->lock);
 219
 220         oldsize = i_size_read(inode);
 221         btrfs_i_size_write(inode, 0);
 222         truncate_pagecache(inode, oldsize, 0);
 223
 224         /*
 225          * We don't need an orphan item because truncating the free space cache
 226          * will never be split across transactions.
 227          */
 228         ret = btrfs_truncate_inode_items(trans, root, inode,
 229                                          0, BTRFS_EXTENT_DATA_KEY);
 230
 231         if (ret) {
 232                 trans->block_rsv = rsv;
 233                 WARN_ON(1);
 234                 return ret;
 235         }
 236
 237         ret = btrfs_update_inode(trans, root, inode);
 238         trans->block_rsv = rsv;
 239
 240         return ret;
 241 }
 242
 243 static int readahead_cache(struct inode *inode)
 244 {
 245         struct file_ra_state *ra;
 246         unsigned long last_index;
 247
 248         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 249         if (!ra)
 250                 return -ENOMEM;
 251
 252         file_ra_state_init(ra, inode->i_mapping);
 253         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 254
 255         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 256
 257         kfree(ra);
 258
 259         return 0;
 260 }
 261
 262 struct io_ctl {
 263         void *cur, *orig;
 264         struct page *page;
 265         struct page **pages;
 266         struct btrfs_root *root;
 267         unsigned long size;
 268         int index;
 269         int num_pages;
 270         unsigned check_crcs:1;
 271 };
 272
 273 static int io_ctl_init(struct io_ctl *io_ctl, struct inode *inode,
 274                        struct btrfs_root *root)
 275 {
 276         memset(io_ctl, 0, sizeof(struct io_ctl));
 277         io_ctl->num_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
 278                 PAGE_CACHE_SHIFT;
 279         io_ctl->pages = kzalloc(sizeof(struct page *) * io_ctl->num_pages,
 280                                 GFP_NOFS);
 281         if (!io_ctl->pages)
 282                 return -ENOMEM;
 283         io_ctl->root = root;
 284         if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
 285                 io_ctl->check_crcs = 1;
 286         return 0;
 287 }
 288
 289 static void io_ctl_free(struct io_ctl *io_ctl)
 290 {
 291         kfree(io_ctl->pages);
 292 }
 293
 294 static void io_ctl_unmap_page(struct io_ctl *io_ctl)
 295 {
 296         if (io_ctl->cur) {
 297                 kunmap(io_ctl->page);
 298                 io_ctl->cur = NULL;
 299                 io_ctl->orig = NULL;
 300         }
 301 }
 302
 303 static void io_ctl_map_page(struct io_ctl *io_ctl, int clear)
 304 {
 305         WARN_ON(io_ctl->cur);
 306         BUG_ON(io_ctl->index >= io_ctl->num_pages);
 307         io_ctl->page = io_ctl->pages[io_ctl->index++];
 308         io_ctl->cur = kmap(io_ctl->page);
 309         io_ctl->orig = io_ctl->cur;
 310         io_ctl->size = PAGE_CACHE_SIZE;
 311         if (clear)
 312                 memset(io_ctl->cur, 0, PAGE_CACHE_SIZE);
 313 }
 314
 315 static void io_ctl_drop_pages(struct io_ctl *io_ctl)
 316 {
 317         int i;
 318
 319         io_ctl_unmap_page(io_ctl);
 320
 321         for (i = 0; i < io_ctl->num_pages; i++) {
 322                 ClearPageChecked(io_ctl->pages[i]);
 323                 unlock_page(io_ctl->pages[i]);
 324                 page_cache_release(io_ctl->pages[i]);
 325         }
 326 }
 327
 328 static int io_ctl_prepare_pages(struct io_ctl *io_ctl, struct inode *inode,
 329                                 int uptodate)
 330 {
 331         struct page *page;
 332         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 333         int i;
 334
 335         for (i = 0; i < io_ctl->num_pages; i++) {
 336                 page = find_or_create_page(inode->i_mapping, i, mask);
 337                 if (!page) {
 338                         io_ctl_drop_pages(io_ctl);
 339                         return -ENOMEM;
 340                 }
 341                 io_ctl->pages[i] = page;
 342                 if (uptodate && !PageUptodate(page)) {
 343                         btrfs_readpage(NULL, page);
 344                         lock_page(page);
 345                         if (!PageUptodate(page)) {
 346                                 printk(KERN_ERR "btrfs: error reading free "
 347                                        "space cache\n");
 348                                 io_ctl_drop_pages(io_ctl);
 349                                 return -EIO;
 350                         }
 351                 }
 352         }
 353
 354         for (i = 0; i < io_ctl->num_pages; i++) {
 355                 clear_page_dirty_for_io(io_ctl->pages[i]);
 356                 set_page_extent_mapped(io_ctl->pages[i]);
 357         }
 358
 359         return 0;
 360 }
 361
 362 static void io_ctl_set_generation(struct io_ctl *io_ctl, u64 generation)
 363 {
 364         u64 *val;
 365
 366         io_ctl_map_page(io_ctl, 1);
 367
 368         /*
 369          * Skip the csum areas.  If we don't check crcs then we just have a
 370          * 64bit chunk at the front of the first page.
 371          */
 372         if (io_ctl->check_crcs) {
 373                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 374                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 375         } else {
 376                 io_ctl->cur += sizeof(u64);
 377                 io_ctl->size -= sizeof(u64) * 2;
 378         }
 379
 380         val = io_ctl->cur;
 381         *val = cpu_to_le64(generation);
 382         io_ctl->cur += sizeof(u64);
 383 }
 384
 385 static int io_ctl_check_generation(struct io_ctl *io_ctl, u64 generation)
 386 {
 387         u64 *gen;
 388
 389         /*
 390          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 391          * chunk at the front of the first page.
 392          */
 393         if (io_ctl->check_crcs) {
 394                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 395                 io_ctl->size -= sizeof(u64) +
 396                         (sizeof(u32) * io_ctl->num_pages);
 397         } else {
 398                 io_ctl->cur += sizeof(u64);
 399                 io_ctl->size -= sizeof(u64) * 2;
 400         }
 401
 402         gen = io_ctl->cur;
 403         if (le64_to_cpu(*gen) != generation) {
 404                 printk_ratelimited(KERN_ERR "btrfs: space cache generation "
 405                                    "(%Lu) does not match inode (%Lu)\n", *gen,
 406                                    generation);
 407                 io_ctl_unmap_page(io_ctl);
 408                 return -EIO;
 409         }
 410         io_ctl->cur += sizeof(u64);
 411         return 0;
 412 }
 413
 414 static void io_ctl_set_crc(struct io_ctl *io_ctl, int index)
 415 {
 416         u32 *tmp;
 417         u32 crc = ~(u32)0;
 418         unsigned offset = 0;
 419
 420         if (!io_ctl->check_crcs) {
 421                 io_ctl_unmap_page(io_ctl);
 422                 return;
 423         }
 424
 425         if (index == 0)
 426                 offset = sizeof(u32) * io_ctl->num_pages;;
 427
 428         crc = btrfs_csum_data(io_ctl->root, io_ctl->orig + offset, crc,
 429                               PAGE_CACHE_SIZE - offset);
 430         btrfs_csum_final(crc, (char *)&crc);
 431         io_ctl_unmap_page(io_ctl);
 432         tmp = kmap(io_ctl->pages[0]);
 433         tmp += index;
 434         *tmp = crc;
 435         kunmap(io_ctl->pages[0]);
 436 }
 437
 438 static int io_ctl_check_crc(struct io_ctl *io_ctl, int index)
 439 {
 440         u32 *tmp, val;
 441         u32 crc = ~(u32)0;
 442         unsigned offset = 0;
 443
 444         if (!io_ctl->check_crcs) {
 445                 io_ctl_map_page(io_ctl, 0);
 446                 return 0;
 447         }
 448
 449         if (index == 0)
 450                 offset = sizeof(u32) * io_ctl->num_pages;
 451
 452         tmp = kmap(io_ctl->pages[0]);
 453         tmp += index;
 454         val = *tmp;
 455         kunmap(io_ctl->pages[0]);
 456
 457         io_ctl_map_page(io_ctl, 0);
 458         crc = btrfs_csum_data(io_ctl->root, io_ctl->orig + offset, crc,
 459                               PAGE_CACHE_SIZE - offset);
 460         btrfs_csum_final(crc, (char *)&crc);
 461         if (val != crc) {
 462                 printk_ratelimited(KERN_ERR "btrfs: csum mismatch on free "
 463                                    "space cache\n");
 464                 io_ctl_unmap_page(io_ctl);
 465                 return -EIO;
 466         }
 467
 468         return 0;
 469 }
 470
 471 static int io_ctl_add_entry(struct io_ctl *io_ctl, u64 offset, u64 bytes,
 472                             void *bitmap)
 473 {
 474         struct btrfs_free_space_entry *entry;
 475
 476         if (!io_ctl->cur)
 477                 return -ENOSPC;
 478
 479         entry = io_ctl->cur;
 480         entry->offset = cpu_to_le64(offset);
 481         entry->bytes = cpu_to_le64(bytes);
 482         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 483                 BTRFS_FREE_SPACE_EXTENT;
 484         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 485         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 486
 487         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 488                 return 0;
 489
 490         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 491
 492         /* No more pages to map */
 493         if (io_ctl->index >= io_ctl->num_pages)
 494                 return 0;
 495
 496         /* map the next page */
 497         io_ctl_map_page(io_ctl, 1);
 498         return 0;
 499 }
 500
 501 static int io_ctl_add_bitmap(struct io_ctl *io_ctl, void *bitmap)
 502 {
 503         if (!io_ctl->cur)
 504                 return -ENOSPC;
 505
 506         /*
 507          * If we aren't at the start of the current page, unmap this one and
 508          * map the next one if there is any left.
 509          */
 510         if (io_ctl->cur != io_ctl->orig) {
 511                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 512                 if (io_ctl->index >= io_ctl->num_pages)
 513                         return -ENOSPC;
 514                 io_ctl_map_page(io_ctl, 0);
 515         }
 516
 517         memcpy(io_ctl->cur, bitmap, PAGE_CACHE_SIZE);
 518         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 519         if (io_ctl->index < io_ctl->num_pages)
 520                 io_ctl_map_page(io_ctl, 0);
 521         return 0;
 522 }
 523
 524 static void io_ctl_zero_remaining_pages(struct io_ctl *io_ctl)
 525 {
 526         /*
 527          * If we're not on the boundary we know we've modified the page and we
 528          * need to crc the page.
 529          */
 530         if (io_ctl->cur != io_ctl->orig)
 531                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 532         else
 533                 io_ctl_unmap_page(io_ctl);
 534
 535         while (io_ctl->index < io_ctl->num_pages) {
 536                 io_ctl_map_page(io_ctl, 1);
 537                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 538         }
 539 }
 540
 541 static int io_ctl_read_entry(struct io_ctl *io_ctl,
 542                             struct btrfs_free_space *entry, u8 *type)
 543 {
 544         struct btrfs_free_space_entry *e;
 545         int ret;
 546
 547         if (!io_ctl->cur) {
 548                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 549                 if (ret)
 550                         return ret;
 551         }
 552
 553         e = io_ctl->cur;
 554         entry->offset = le64_to_cpu(e->offset);
 555         entry->bytes = le64_to_cpu(e->bytes);
 556         *type = e->type;
 557         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 558         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 559
 560         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 561                 return 0;
 562
 563         io_ctl_unmap_page(io_ctl);
 564
 565         return 0;
 566 }
 567
 568 static int io_ctl_read_bitmap(struct io_ctl *io_ctl,
 569                               struct btrfs_free_space *entry)
 570 {
 571         int ret;
 572
 573         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 574         if (ret)
 575                 return ret;
 576
 577         memcpy(entry->bitmap, io_ctl->cur, PAGE_CACHE_SIZE);
 578         io_ctl_unmap_page(io_ctl);
 579
 580         return 0;
 581 }
 582
 583 int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 584                             struct btrfs_free_space_ctl *ctl,
 585                             struct btrfs_path *path, u64 offset)
 586 {
 587         struct btrfs_free_space_header *header;
 588         struct extent_buffer *leaf;
 589         struct io_ctl io_ctl;
 590         struct btrfs_key key;
 591         struct btrfs_free_space *e, *n;
 592         struct list_head bitmaps;
 593         u64 num_entries;
 594         u64 num_bitmaps;
 595         u64 generation;
 596         u8 type;
 597         int ret = 0;
 598
 599         INIT_LIST_HEAD(&bitmaps);
 600
 601         /* Nothing in the space cache, goodbye */
 602         if (!i_size_read(inode))
 603                 return 0;
 604
 605         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 606         key.offset = offset;
 607         key.type = 0;
 608
 609         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 610         if (ret < 0)
 611                 return 0;
 612         else if (ret > 0) {
 613                 btrfs_release_path(path);
 614                 return 0;
 615         }
 616
 617         ret = -1;
 618
 619         leaf = path->nodes[0];
 620         header = btrfs_item_ptr(leaf, path->slots[0],
 621                                 struct btrfs_free_space_header);
 622         num_entries = btrfs_free_space_entries(leaf, header);
 623         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 624         generation = btrfs_free_space_generation(leaf, header);
 625         btrfs_release_path(path);
 626
 627         if (BTRFS_I(inode)->generation != generation) {
 628                 printk(KERN_ERR "btrfs: free space inode generation (%llu) did"
 629                        " not match free space cache generation (%llu)\n",
 630                        (unsigned long long)BTRFS_I(inode)->generation,
 631                        (unsigned long long)generation);
 632                 return 0;
 633         }
 634
 635         if (!num_entries)
 636                 return 0;
 637
 638         io_ctl_init(&io_ctl, inode, root);
 639         ret = readahead_cache(inode);
 640         if (ret)
 641                 goto out;
 642
 643         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 644         if (ret)
 645                 goto out;
 646
 647         ret = io_ctl_check_crc(&io_ctl, 0);
 648         if (ret)
 649                 goto free_cache;
 650
 651         ret = io_ctl_check_generation(&io_ctl, generation);
 652         if (ret)
 653                 goto free_cache;
 654
 655         while (num_entries) {
 656                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 657                                       GFP_NOFS);
 658                 if (!e)
 659                         goto free_cache;
 660
 661                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 662                 if (ret) {
 663                         kmem_cache_free(btrfs_free_space_cachep, e);
 664                         goto free_cache;
 665                 }
 666
 667                 if (!e->bytes) {
 668                         kmem_cache_free(btrfs_free_space_cachep, e);
 669                         goto free_cache;
 670                 }
 671
 672                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 673                         spin_lock(&ctl->tree_lock);
 674                         ret = link_free_space(ctl, e);
 675                         spin_unlock(&ctl->tree_lock);
 676                         if (ret) {
 677                                 printk(KERN_ERR "Duplicate entries in "
 678                                        "free space cache, dumping\n");
 679                                 kmem_cache_free(btrfs_free_space_cachep, e);
 680                                 goto free_cache;
 681                         }
 682                 } else {
 683                         BUG_ON(!num_bitmaps);
 684                         num_bitmaps--;
 685                         e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
 686                         if (!e->bitmap) {
 687                                 kmem_cache_free(
 688                                         btrfs_free_space_cachep, e);
 689                                 goto free_cache;
 690                         }
 691                         spin_lock(&ctl->tree_lock);
 692                         ret = link_free_space(ctl, e);
 693                         ctl->total_bitmaps++;
 694                         ctl->op->recalc_thresholds(ctl);
 695                         spin_unlock(&ctl->tree_lock);
 696                         if (ret) {
 697                                 printk(KERN_ERR "Duplicate entries in "
 698                                        "free space cache, dumping\n");
 699                                 kmem_cache_free(btrfs_free_space_cachep, e);
 700                                 goto free_cache;
 701                         }
 702                         list_add_tail(&e->list, &bitmaps);
 703                 }
 704
 705                 num_entries--;
 706         }
 707
 708         io_ctl_unmap_page(&io_ctl);
 709
 710         /*
 711          * We add the bitmaps at the end of the entries in order that
 712          * the bitmap entries are added to the cache.
 713          */
 714         list_for_each_entry_safe(e, n, &bitmaps, list) {
 715                 list_del_init(&e->list);
 716                 ret = io_ctl_read_bitmap(&io_ctl, e);
 717                 if (ret)
 718                         goto free_cache;
 719         }
 720
 721         io_ctl_drop_pages(&io_ctl);
 722         ret = 1;
 723 out:
 724         io_ctl_free(&io_ctl);
 725         return ret;
 726 free_cache:
 727         io_ctl_drop_pages(&io_ctl);
 728         __btrfs_remove_free_space_cache(ctl);
 729         goto out;
 730 }
 731
 732 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 733                           struct btrfs_block_group_cache *block_group)
 734 {
 735         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 736         struct btrfs_root *root = fs_info->tree_root;
 737         struct inode *inode;
 738         struct btrfs_path *path;
 739         int ret = 0;
 740         bool matched;
 741         u64 used = btrfs_block_group_used(&block_group->item);
 742
 743         /*
 744          * If we're unmounting then just return, since this does a search on the
 745          * normal root and not the commit root and we could deadlock.
 746          */
 747         if (btrfs_fs_closing(fs_info))
 748                 return 0;
 749
 750         /*
 751          * If this block group has been marked to be cleared for one reason or
 752          * another then we can't trust the on disk cache, so just return.
 753          */
 754         spin_lock(&block_group->lock);
 755         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 756                 spin_unlock(&block_group->lock);
 757                 return 0;
 758         }
 759         spin_unlock(&block_group->lock);
 760
 761         path = btrfs_alloc_path();
 762         if (!path)
 763                 return 0;
 764
 765         inode = lookup_free_space_inode(root, block_group, path);
 766         if (IS_ERR(inode)) {
 767                 btrfs_free_path(path);
 768                 return 0;
 769         }
 770
 771         /* We may have converted the inode and made the cache invalid. */
 772         spin_lock(&block_group->lock);
 773         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 774                 spin_unlock(&block_group->lock);
 775                 goto out;
 776         }
 777         spin_unlock(&block_group->lock);
 778
 779         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 780                                       path, block_group->key.objectid);
 781         btrfs_free_path(path);
 782         if (ret <= 0)
 783                 goto out;
 784
 785         spin_lock(&ctl->tree_lock);
 786         matched = (ctl->free_space == (block_group->key.offset - used -
 787                                        block_group->bytes_super));
 788         spin_unlock(&ctl->tree_lock);
 789
 790         if (!matched) {
 791                 __btrfs_remove_free_space_cache(ctl);
 792                 printk(KERN_ERR "block group %llu has an wrong amount of free "
 793                        "space\n", block_group->key.objectid);
 794                 ret = -1;
 795         }
 796 out:
 797         if (ret < 0) {
 798                 /* This cache is bogus, make sure it gets cleared */
 799                 spin_lock(&block_group->lock);
 800                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 801                 spin_unlock(&block_group->lock);
 802                 ret = 0;
 803
 804                 printk(KERN_ERR "btrfs: failed to load free space cache "
 805                        "for block group %llu\n", block_group->key.objectid);
 806         }
 807
 808         iput(inode);
 809         return ret;
 810 }
 811
 812 /**
 813  * __btrfs_write_out_cache - write out cached info to an inode
 814  * @root - the root the inode belongs to
 815  * @ctl - the free space cache we are going to write out
 816  * @block_group - the block_group for this cache if it belongs to a block_group
 817  * @trans - the trans handle
 818  * @path - the path to use
 819  * @offset - the offset for the key we'll insert
 820  *
 821  * This function writes out a free space cache struct to disk for quick recovery
 822  * on mount.  This will return 0 if it was successfull in writing the cache out,
 823  * and -1 if it was not.
 824  */
 825 int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
 826                             struct btrfs_free_space_ctl *ctl,
 827                             struct btrfs_block_group_cache *block_group,
 828                             struct btrfs_trans_handle *trans,
 829                             struct btrfs_path *path, u64 offset)
 830 {
 831         struct btrfs_free_space_header *header;
 832         struct extent_buffer *leaf;
 833         struct rb_node *node;
 834         struct list_head *pos, *n;
 835         struct extent_state *cached_state = NULL;
 836         struct btrfs_free_cluster *cluster = NULL;
 837         struct extent_io_tree *unpin = NULL;
 838         struct io_ctl io_ctl;
 839         struct list_head bitmap_list;
 840         struct btrfs_key key;
 841         u64 start, end, len;
 842         int entries = 0;
 843         int bitmaps = 0;
 844         int ret;
 845         int err = -1;
 846
 847         INIT_LIST_HEAD(&bitmap_list);
 848
 849         if (!i_size_read(inode))
 850                 return -1;
 851
 852         io_ctl_init(&io_ctl, inode, root);
 853
 854         /* Get the cluster for this block_group if it exists */
 855         if (block_group && !list_empty(&block_group->cluster_list))
 856                 cluster = list_entry(block_group->cluster_list.next,
 857                                      struct btrfs_free_cluster,
 858                                      block_group_list);
 859
 860         /*
 861          * We shouldn't have switched the pinned extents yet so this is the
 862          * right one
 863          */
 864         unpin = root->fs_info->pinned_extents;
 865
 866         /* Lock all pages first so we can lock the extent safely. */
 867         io_ctl_prepare_pages(&io_ctl, inode, 0);
 868
 869         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
 870                          0, &cached_state, GFP_NOFS);
 871
 872         /*
 873          * When searching for pinned extents, we need to start at our start
 874          * offset.
 875          */
 876         if (block_group)
 877                 start = block_group->key.objectid;
 878
 879         node = rb_first(&ctl->free_space_offset);
 880         if (!node && cluster) {
 881                 node = rb_first(&cluster->root);
 882                 cluster = NULL;
 883         }
 884
 885         /* Make sure we can fit our crcs into the first page */
 886         if (io_ctl.check_crcs &&
 887             (io_ctl.num_pages * sizeof(u32)) >= PAGE_CACHE_SIZE) {
 888                 WARN_ON(1);
 889                 goto out_nospc;
 890         }
 891
 892         io_ctl_set_generation(&io_ctl, trans->transid);
 893
 894         /* Write out the extent entries */
 895         while (node) {
 896                 struct btrfs_free_space *e;
 897
 898                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 899                 entries++;
 900
 901                 ret = io_ctl_add_entry(&io_ctl, e->offset, e->bytes,
 902                                        e->bitmap);
 903                 if (ret)
 904                         goto out_nospc;
 905
 906                 if (e->bitmap) {
 907                         list_add_tail(&e->list, &bitmap_list);
 908                         bitmaps++;
 909                 }
 910                 node = rb_next(node);
 911                 if (!node && cluster) {
 912                         node = rb_first(&cluster->root);
 913                         cluster = NULL;
 914                 }
 915         }
 916
 917         /*
 918          * We want to add any pinned extents to our free space cache
 919          * so we don't leak the space
 920          */
 921         while (block_group && (start < block_group->key.objectid +
 922                                block_group->key.offset)) {
 923                 ret = find_first_extent_bit(unpin, start, &start, &end,
 924                                             EXTENT_DIRTY);
 925                 if (ret) {
 926                         ret = 0;
 927                         break;
 928                 }
 929
 930                 /* This pinned extent is out of our range */
 931                 if (start >= block_group->key.objectid +
 932                     block_group->key.offset)
 933                         break;
 934
 935                 len = block_group->key.objectid +
 936                         block_group->key.offset - start;
 937                 len = min(len, end + 1 - start);
 938
 939                 entries++;
 940                 ret = io_ctl_add_entry(&io_ctl, start, len, NULL);
 941                 if (ret)
 942                         goto out_nospc;
 943
 944                 start = end + 1;
 945         }
 946
 947         /* Write out the bitmaps */
 948         list_for_each_safe(pos, n, &bitmap_list) {
 949                 struct btrfs_free_space *entry =
 950                         list_entry(pos, struct btrfs_free_space, list);
 951
 952                 ret = io_ctl_add_bitmap(&io_ctl, entry->bitmap);
 953                 if (ret)
 954                         goto out_nospc;
 955                 list_del_init(&entry->list);
 956         }
 957
 958         /* Zero out the rest of the pages just to make sure */
 959         io_ctl_zero_remaining_pages(&io_ctl);
 960
 961         ret = btrfs_dirty_pages(root, inode, io_ctl.pages, io_ctl.num_pages,
 962                                 0, i_size_read(inode), &cached_state);
 963         io_ctl_drop_pages(&io_ctl);
 964         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
 965                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
 966
 967         if (ret)
 968                 goto out;
 969
 970
 971         ret = filemap_write_and_wait(inode->i_mapping);
 972         if (ret)
 973                 goto out;
 974
 975         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 976         key.offset = offset;
 977         key.type = 0;
 978
 979         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 980         if (ret < 0) {
 981                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
 982                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
 983                                  GFP_NOFS);
 984                 goto out;
 985         }
 986         leaf = path->nodes[0];
 987         if (ret > 0) {
 988                 struct btrfs_key found_key;
 989                 BUG_ON(!path->slots[0]);
 990                 path->slots[0]--;
 991                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
 992                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
 993                     found_key.offset != offset) {
 994                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
 995                                          inode->i_size - 1,
 996                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
 997                                          NULL, GFP_NOFS);
 998                         btrfs_release_path(path);
 999                         goto out;
1000                 }
1001         }
1002
1003         BTRFS_I(inode)->generation = trans->transid;
1004         header = btrfs_item_ptr(leaf, path->slots[0],
1005                                 struct btrfs_free_space_header);
1006         btrfs_set_free_space_entries(leaf, header, entries);
1007         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1008         btrfs_set_free_space_generation(leaf, header, trans->transid);
1009         btrfs_mark_buffer_dirty(leaf);
1010         btrfs_release_path(path);
1011
1012         err = 0;
1013 out:
1014         io_ctl_free(&io_ctl);
1015         if (err) {
1016                 invalidate_inode_pages2(inode->i_mapping);
1017                 BTRFS_I(inode)->generation = 0;
1018         }
1019         btrfs_update_inode(trans, root, inode);
1020         return err;
1021
1022 out_nospc:
1023         list_for_each_safe(pos, n, &bitmap_list) {
1024                 struct btrfs_free_space *entry =
1025                         list_entry(pos, struct btrfs_free_space, list);
1026                 list_del_init(&entry->list);
1027         }
1028         io_ctl_drop_pages(&io_ctl);
1029         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1030                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1031         goto out;
1032 }
1033
1034 int btrfs_write_out_cache(struct btrfs_root *root,
1035                           struct btrfs_trans_handle *trans,
1036                           struct btrfs_block_group_cache *block_group,
1037                           struct btrfs_path *path)
1038 {
1039         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1040         struct inode *inode;
1041         int ret = 0;
1042
1043         root = root->fs_info->tree_root;
1044
1045         spin_lock(&block_group->lock);
1046         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1047                 spin_unlock(&block_group->lock);
1048                 return 0;
1049         }
1050         spin_unlock(&block_group->lock);
1051
1052         inode = lookup_free_space_inode(root, block_group, path);
1053         if (IS_ERR(inode))
1054                 return 0;
1055
1056         ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
1057                                       path, block_group->key.objectid);
1058         if (ret) {
1059                 spin_lock(&block_group->lock);
1060                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1061                 spin_unlock(&block_group->lock);
1062                 ret = 0;
1063 #ifdef DEBUG
1064                 printk(KERN_ERR "btrfs: failed to write free space cace "
1065                        "for block group %llu\n", block_group->key.objectid);
1066 #endif
1067         }
1068
1069         iput(inode);
1070         return ret;
1071 }
1072
1073 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1074                                           u64 offset)
1075 {
1076         BUG_ON(offset < bitmap_start);
1077         offset -= bitmap_start;
1078         return (unsigned long)(div_u64(offset, unit));
1079 }
1080
1081 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1082 {
1083         return (unsigned long)(div_u64(bytes, unit));
1084 }
1085
1086 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1087                                    u64 offset)
1088 {
1089         u64 bitmap_start;
1090         u64 bytes_per_bitmap;
1091
1092         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1093         bitmap_start = offset - ctl->start;
1094         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1095         bitmap_start *= bytes_per_bitmap;
1096         bitmap_start += ctl->start;
1097
1098         return bitmap_start;
1099 }
1100
1101 static int tree_insert_offset(struct rb_root *root, u64 offset,
1102                               struct rb_node *node, int bitmap)
1103 {
1104         struct rb_node **p = &root->rb_node;
1105         struct rb_node *parent = NULL;
1106         struct btrfs_free_space *info;
1107
1108         while (*p) {
1109                 parent = *p;
1110                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1111
1112                 if (offset < info->offset) {
1113                         p = &(*p)->rb_left;
1114                 } else if (offset > info->offset) {
1115                         p = &(*p)->rb_right;
1116                 } else {
1117                         /*
1118                          * we could have a bitmap entry and an extent entry
1119                          * share the same offset.  If this is the case, we want
1120                          * the extent entry to always be found first if we do a
1121                          * linear search through the tree, since we want to have
1122                          * the quickest allocation time, and allocating from an
1123                          * extent is faster than allocating from a bitmap.  So
1124                          * if we're inserting a bitmap and we find an entry at
1125                          * this offset, we want to go right, or after this entry
1126                          * logically.  If we are inserting an extent and we've
1127                          * found a bitmap, we want to go left, or before
1128                          * logically.
1129                          */
1130                         if (bitmap) {
1131                                 if (info->bitmap) {
1132                                         WARN_ON_ONCE(1);
1133                                         return -EEXIST;
1134                                 }
1135                                 p = &(*p)->rb_right;
1136                         } else {
1137                                 if (!info->bitmap) {
1138                                         WARN_ON_ONCE(1);
1139                                         return -EEXIST;
1140                                 }
1141                                 p = &(*p)->rb_left;
1142                         }
1143                 }
1144         }
1145
1146         rb_link_node(node, parent, p);
1147         rb_insert_color(node, root);
1148
1149         return 0;
1150 }
1151
1152 /*
1153  * searches the tree for the given offset.
1154  *
1155  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1156  * want a section that has at least bytes size and comes at or after the given
1157  * offset.
1158  */
1159 static struct btrfs_free_space *
1160 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1161                    u64 offset, int bitmap_only, int fuzzy)
1162 {
1163         struct rb_node *n = ctl->free_space_offset.rb_node;
1164         struct btrfs_free_space *entry, *prev = NULL;
1165
1166         /* find entry that is closest to the 'offset' */
1167         while (1) {
1168                 if (!n) {
1169                         entry = NULL;
1170                         break;
1171                 }
1172
1173                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1174                 prev = entry;
1175
1176                 if (offset < entry->offset)
1177                         n = n->rb_left;
1178                 else if (offset > entry->offset)
1179                         n = n->rb_right;
1180                 else
1181                         break;
1182         }
1183
1184         if (bitmap_only) {
1185                 if (!entry)
1186                         return NULL;
1187                 if (entry->bitmap)
1188                         return entry;
1189
1190                 /*
1191                  * bitmap entry and extent entry may share same offset,
1192                  * in that case, bitmap entry comes after extent entry.
1193                  */
1194                 n = rb_next(n);
1195                 if (!n)
1196                         return NULL;
1197                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1198                 if (entry->offset != offset)
1199                         return NULL;
1200
1201                 WARN_ON(!entry->bitmap);
1202                 return entry;
1203         } else if (entry) {
1204                 if (entry->bitmap) {
1205                         /*
1206                          * if previous extent entry covers the offset,
1207                          * we should return it instead of the bitmap entry
1208                          */
1209                         n = &entry->offset_index;
1210                         while (1) {
1211                                 n = rb_prev(n);
1212                                 if (!n)
1213                                         break;
1214                                 prev = rb_entry(n, struct btrfs_free_space,
1215                                                 offset_index);
1216                                 if (!prev->bitmap) {
1217                                         if (prev->offset + prev->bytes > offset)
1218                                                 entry = prev;
1219                                         break;
1220                                 }
1221                         }
1222                 }
1223                 return entry;
1224         }
1225
1226         if (!prev)
1227                 return NULL;
1228
1229         /* find last entry before the 'offset' */
1230         entry = prev;
1231         if (entry->offset > offset) {
1232                 n = rb_prev(&entry->offset_index);
1233                 if (n) {
1234                         entry = rb_entry(n, struct btrfs_free_space,
1235                                         offset_index);
1236                         BUG_ON(entry->offset > offset);
1237                 } else {
1238                         if (fuzzy)
1239                                 return entry;
1240                         else
1241                                 return NULL;
1242                 }
1243         }
1244
1245         if (entry->bitmap) {
1246                 n = &entry->offset_index;
1247                 while (1) {
1248                         n = rb_prev(n);
1249                         if (!n)
1250                                 break;
1251                         prev = rb_entry(n, struct btrfs_free_space,
1252                                         offset_index);
1253                         if (!prev->bitmap) {
1254                                 if (prev->offset + prev->bytes > offset)
1255                                         return prev;
1256                                 break;
1257                         }
1258                 }
1259                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1260                         return entry;
1261         } else if (entry->offset + entry->bytes > offset)
1262                 return entry;
1263
1264         if (!fuzzy)
1265                 return NULL;
1266
1267         while (1) {
1268                 if (entry->bitmap) {
1269                         if (entry->offset + BITS_PER_BITMAP *
1270                             ctl->unit > offset)
1271                                 break;
1272                 } else {
1273                         if (entry->offset + entry->bytes > offset)
1274                                 break;
1275                 }
1276
1277                 n = rb_next(&entry->offset_index);
1278                 if (!n)
1279                         return NULL;
1280                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1281         }
1282         return entry;
1283 }
1284
1285 static inline void
1286 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1287                     struct btrfs_free_space *info)
1288 {
1289         rb_erase(&info->offset_index, &ctl->free_space_offset);
1290         ctl->free_extents--;
1291 }
1292
1293 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1294                               struct btrfs_free_space *info)
1295 {
1296         __unlink_free_space(ctl, info);
1297         ctl->free_space -= info->bytes;
1298 }
1299
1300 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1301                            struct btrfs_free_space *info)
1302 {
1303         int ret = 0;
1304
1305         BUG_ON(!info->bitmap && !info->bytes);
1306         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1307                                  &info->offset_index, (info->bitmap != NULL));
1308         if (ret)
1309                 return ret;
1310
1311         ctl->free_space += info->bytes;
1312         ctl->free_extents++;
1313         return ret;
1314 }
1315
1316 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1317 {
1318         struct btrfs_block_group_cache *block_group = ctl->private;
1319         u64 max_bytes;
1320         u64 bitmap_bytes;
1321         u64 extent_bytes;
1322         u64 size = block_group->key.offset;
1323         u64 bytes_per_bg = BITS_PER_BITMAP * block_group->sectorsize;
1324         int max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1325
1326         BUG_ON(ctl->total_bitmaps > max_bitmaps);
1327
1328         /*
1329          * The goal is to keep the total amount of memory used per 1gb of space
1330          * at or below 32k, so we need to adjust how much memory we allow to be
1331          * used by extent based free space tracking
1332          */
1333         if (size < 1024 * 1024 * 1024)
1334                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1335         else
1336                 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1337                         div64_u64(size, 1024 * 1024 * 1024);
1338
1339         /*
1340          * we want to account for 1 more bitmap than what we have so we can make
1341          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1342          * we add more bitmaps.
1343          */
1344         bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1345
1346         if (bitmap_bytes >= max_bytes) {
1347                 ctl->extents_thresh = 0;
1348                 return;
1349         }
1350
1351         /*
1352          * we want the extent entry threshold to always be at most 1/2 the maxw
1353          * bytes we can have, or whatever is less than that.
1354          */
1355         extent_bytes = max_bytes - bitmap_bytes;
1356         extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1357
1358         ctl->extents_thresh =
1359                 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1360 }
1361
1362 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1363                                        struct btrfs_free_space *info,
1364                                        u64 offset, u64 bytes)
1365 {
1366         unsigned long start, count;
1367
1368         start = offset_to_bit(info->offset, ctl->unit, offset);
1369         count = bytes_to_bits(bytes, ctl->unit);
1370         BUG_ON(start + count > BITS_PER_BITMAP);
1371
1372         bitmap_clear(info->bitmap, start, count);
1373
1374         info->bytes -= bytes;
1375 }
1376
1377 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1378                               struct btrfs_free_space *info, u64 offset,
1379                               u64 bytes)
1380 {
1381         __bitmap_clear_bits(ctl, info, offset, bytes);
1382         ctl->free_space -= bytes;
1383 }
1384
1385 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1386                             struct btrfs_free_space *info, u64 offset,
1387                             u64 bytes)
1388 {
1389         unsigned long start, count;
1390
1391         start = offset_to_bit(info->offset, ctl->unit, offset);
1392         count = bytes_to_bits(bytes, ctl->unit);
1393         BUG_ON(start + count > BITS_PER_BITMAP);
1394
1395         bitmap_set(info->bitmap, start, count);
1396
1397         info->bytes += bytes;
1398         ctl->free_space += bytes;
1399 }
1400
1401 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1402                          struct btrfs_free_space *bitmap_info, u64 *offset,
1403                          u64 *bytes)
1404 {
1405         unsigned long found_bits = 0;
1406         unsigned long bits, i;
1407         unsigned long next_zero;
1408
1409         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1410                           max_t(u64, *offset, bitmap_info->offset));
1411         bits = bytes_to_bits(*bytes, ctl->unit);
1412
1413         for (i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i);
1414              i < BITS_PER_BITMAP;
1415              i = find_next_bit(bitmap_info->bitmap, BITS_PER_BITMAP, i + 1)) {
1416                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1417                                                BITS_PER_BITMAP, i);
1418                 if ((next_zero - i) >= bits) {
1419                         found_bits = next_zero - i;
1420                         break;
1421                 }
1422                 i = next_zero;
1423         }
1424
1425         if (found_bits) {
1426                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1427                 *bytes = (u64)(found_bits) * ctl->unit;
1428                 return 0;
1429         }
1430
1431         return -1;
1432 }
1433
1434 static struct btrfs_free_space *
1435 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes)
1436 {
1437         struct btrfs_free_space *entry;
1438         struct rb_node *node;
1439         int ret;
1440
1441         if (!ctl->free_space_offset.rb_node)
1442                 return NULL;
1443
1444         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1445         if (!entry)
1446                 return NULL;
1447
1448         for (node = &entry->offset_index; node; node = rb_next(node)) {
1449                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1450                 if (entry->bytes < *bytes)
1451                         continue;
1452
1453                 if (entry->bitmap) {
1454                         ret = search_bitmap(ctl, entry, offset, bytes);
1455                         if (!ret)
1456                                 return entry;
1457                         continue;
1458                 }
1459
1460                 *offset = entry->offset;
1461                 *bytes = entry->bytes;
1462                 return entry;
1463         }
1464
1465         return NULL;
1466 }
1467
1468 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1469                            struct btrfs_free_space *info, u64 offset)
1470 {
1471         info->offset = offset_to_bitmap(ctl, offset);
1472         info->bytes = 0;
1473         link_free_space(ctl, info);
1474         ctl->total_bitmaps++;
1475
1476         ctl->op->recalc_thresholds(ctl);
1477 }
1478
1479 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1480                         struct btrfs_free_space *bitmap_info)
1481 {
1482         unlink_free_space(ctl, bitmap_info);
1483         kfree(bitmap_info->bitmap);
1484         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1485         ctl->total_bitmaps--;
1486         ctl->op->recalc_thresholds(ctl);
1487 }
1488
1489 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1490                               struct btrfs_free_space *bitmap_info,
1491                               u64 *offset, u64 *bytes)
1492 {
1493         u64 end;
1494         u64 search_start, search_bytes;
1495         int ret;
1496
1497 again:
1498         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1499
1500         /*
1501          * XXX - this can go away after a few releases.
1502          *
1503          * since the only user of btrfs_remove_free_space is the tree logging
1504          * stuff, and the only way to test that is under crash conditions, we
1505          * want to have this debug stuff here just in case somethings not
1506          * working.  Search the bitmap for the space we are trying to use to
1507          * make sure its actually there.  If its not there then we need to stop
1508          * because something has gone wrong.
1509          */
1510         search_start = *offset;
1511         search_bytes = *bytes;
1512         search_bytes = min(search_bytes, end - search_start + 1);
1513         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1514         BUG_ON(ret < 0 || search_start != *offset);
1515
1516         if (*offset > bitmap_info->offset && *offset + *bytes > end) {
1517                 bitmap_clear_bits(ctl, bitmap_info, *offset, end - *offset + 1);
1518                 *bytes -= end - *offset + 1;
1519                 *offset = end + 1;
1520         } else if (*offset >= bitmap_info->offset && *offset + *bytes <= end) {
1521                 bitmap_clear_bits(ctl, bitmap_info, *offset, *bytes);
1522                 *bytes = 0;
1523         }
1524
1525         if (*bytes) {
1526                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1527                 if (!bitmap_info->bytes)
1528                         free_bitmap(ctl, bitmap_info);
1529
1530                 /*
1531                  * no entry after this bitmap, but we still have bytes to
1532                  * remove, so something has gone wrong.
1533                  */
1534                 if (!next)
1535                         return -EINVAL;
1536
1537                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1538                                        offset_index);
1539
1540                 /*
1541                  * if the next entry isn't a bitmap we need to return to let the
1542                  * extent stuff do its work.
1543                  */
1544                 if (!bitmap_info->bitmap)
1545                         return -EAGAIN;
1546
1547                 /*
1548                  * Ok the next item is a bitmap, but it may not actually hold
1549                  * the information for the rest of this free space stuff, so
1550                  * look for it, and if we don't find it return so we can try
1551                  * everything over again.
1552                  */
1553                 search_start = *offset;
1554                 search_bytes = *bytes;
1555                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1556                                     &search_bytes);
1557                 if (ret < 0 || search_start != *offset)
1558                         return -EAGAIN;
1559
1560                 goto again;
1561         } else if (!bitmap_info->bytes)
1562                 free_bitmap(ctl, bitmap_info);
1563
1564         return 0;
1565 }
1566
1567 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1568                                struct btrfs_free_space *info, u64 offset,
1569                                u64 bytes)
1570 {
1571         u64 bytes_to_set = 0;
1572         u64 end;
1573
1574         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1575
1576         bytes_to_set = min(end - offset, bytes);
1577
1578         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1579
1580         return bytes_to_set;
1581
1582 }
1583
1584 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1585                       struct btrfs_free_space *info)
1586 {
1587         struct btrfs_block_group_cache *block_group = ctl->private;
1588
1589         /*
1590          * If we are below the extents threshold then we can add this as an
1591          * extent, and don't have to deal with the bitmap
1592          */
1593         if (ctl->free_extents < ctl->extents_thresh) {
1594                 /*
1595                  * If this block group has some small extents we don't want to
1596                  * use up all of our free slots in the cache with them, we want
1597                  * to reserve them to larger extents, however if we have plent
1598                  * of cache left then go ahead an dadd them, no sense in adding
1599                  * the overhead of a bitmap if we don't have to.
1600                  */
1601                 if (info->bytes <= block_group->sectorsize * 4) {
1602                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
1603                                 return false;
1604                 } else {
1605                         return false;
1606                 }
1607         }
1608
1609         /*
1610          * some block groups are so tiny they can't be enveloped by a bitmap, so
1611          * don't even bother to create a bitmap for this
1612          */
1613         if (BITS_PER_BITMAP * block_group->sectorsize >
1614             block_group->key.offset)
1615                 return false;
1616
1617         return true;
1618 }
1619
1620 static struct btrfs_free_space_op free_space_op = {
1621         .recalc_thresholds      = recalculate_thresholds,
1622         .use_bitmap             = use_bitmap,
1623 };
1624
1625 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1626                               struct btrfs_free_space *info)
1627 {
1628         struct btrfs_free_space *bitmap_info;
1629         struct btrfs_block_group_cache *block_group = NULL;
1630         int added = 0;
1631         u64 bytes, offset, bytes_added;
1632         int ret;
1633
1634         bytes = info->bytes;
1635         offset = info->offset;
1636
1637         if (!ctl->op->use_bitmap(ctl, info))
1638                 return 0;
1639
1640         if (ctl->op == &free_space_op)
1641                 block_group = ctl->private;
1642 again:
1643         /*
1644          * Since we link bitmaps right into the cluster we need to see if we
1645          * have a cluster here, and if so and it has our bitmap we need to add
1646          * the free space to that bitmap.
1647          */
1648         if (block_group && !list_empty(&block_group->cluster_list)) {
1649                 struct btrfs_free_cluster *cluster;
1650                 struct rb_node *node;
1651                 struct btrfs_free_space *entry;
1652
1653                 cluster = list_entry(block_group->cluster_list.next,
1654                                      struct btrfs_free_cluster,
1655                                      block_group_list);
1656                 spin_lock(&cluster->lock);
1657                 node = rb_first(&cluster->root);
1658                 if (!node) {
1659                         spin_unlock(&cluster->lock);
1660                         goto no_cluster_bitmap;
1661                 }
1662
1663                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1664                 if (!entry->bitmap) {
1665                         spin_unlock(&cluster->lock);
1666                         goto no_cluster_bitmap;
1667                 }
1668
1669                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
1670                         bytes_added = add_bytes_to_bitmap(ctl, entry,
1671                                                           offset, bytes);
1672                         bytes -= bytes_added;
1673                         offset += bytes_added;
1674                 }
1675                 spin_unlock(&cluster->lock);
1676                 if (!bytes) {
1677                         ret = 1;
1678                         goto out;
1679                 }
1680         }
1681
1682 no_cluster_bitmap:
1683         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1684                                          1, 0);
1685         if (!bitmap_info) {
1686                 BUG_ON(added);
1687                 goto new_bitmap;
1688         }
1689
1690         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
1691         bytes -= bytes_added;
1692         offset += bytes_added;
1693         added = 0;
1694
1695         if (!bytes) {
1696                 ret = 1;
1697                 goto out;
1698         } else
1699                 goto again;
1700
1701 new_bitmap:
1702         if (info && info->bitmap) {
1703                 add_new_bitmap(ctl, info, offset);
1704                 added = 1;
1705                 info = NULL;
1706                 goto again;
1707         } else {
1708                 spin_unlock(&ctl->tree_lock);
1709
1710                 /* no pre-allocated info, allocate a new one */
1711                 if (!info) {
1712                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
1713                                                  GFP_NOFS);
1714                         if (!info) {
1715                                 spin_lock(&ctl->tree_lock);
1716                                 ret = -ENOMEM;
1717                                 goto out;
1718                         }
1719                 }
1720
1721                 /* allocate the bitmap */
1722                 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1723                 spin_lock(&ctl->tree_lock);
1724                 if (!info->bitmap) {
1725                         ret = -ENOMEM;
1726                         goto out;
1727                 }
1728                 goto again;
1729         }
1730
1731 out:
1732         if (info) {
1733                 if (info->bitmap)
1734                         kfree(info->bitmap);
1735                 kmem_cache_free(btrfs_free_space_cachep, info);
1736         }
1737
1738         return ret;
1739 }
1740
1741 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
1742                           struct btrfs_free_space *info, bool update_stat)
1743 {
1744         struct btrfs_free_space *left_info;
1745         struct btrfs_free_space *right_info;
1746         bool merged = false;
1747         u64 offset = info->offset;
1748         u64 bytes = info->bytes;
1749
1750         /*
1751          * first we want to see if there is free space adjacent to the range we
1752          * are adding, if there is remove that struct and add a new one to
1753          * cover the entire range
1754          */
1755         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
1756         if (right_info && rb_prev(&right_info->offset_index))
1757                 left_info = rb_entry(rb_prev(&right_info->offset_index),
1758                                      struct btrfs_free_space, offset_index);
1759         else
1760                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
1761
1762         if (right_info && !right_info->bitmap) {
1763                 if (update_stat)
1764                         unlink_free_space(ctl, right_info);
1765                 else
1766                         __unlink_free_space(ctl, right_info);
1767                 info->bytes += right_info->bytes;
1768                 kmem_cache_free(btrfs_free_space_cachep, right_info);
1769                 merged = true;
1770         }
1771
1772         if (left_info && !left_info->bitmap &&
1773             left_info->offset + left_info->bytes == offset) {
1774                 if (update_stat)
1775                         unlink_free_space(ctl, left_info);
1776                 else
1777                         __unlink_free_space(ctl, left_info);
1778                 info->offset = left_info->offset;
1779                 info->bytes += left_info->bytes;
1780                 kmem_cache_free(btrfs_free_space_cachep, left_info);
1781                 merged = true;
1782         }
1783
1784         return merged;
1785 }
1786
1787 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
1788                            u64 offset, u64 bytes)
1789 {
1790         struct btrfs_free_space *info;
1791         int ret = 0;
1792
1793         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
1794         if (!info)
1795                 return -ENOMEM;
1796
1797         info->offset = offset;
1798         info->bytes = bytes;
1799
1800         spin_lock(&ctl->tree_lock);
1801
1802         if (try_merge_free_space(ctl, info, true))
1803                 goto link;
1804
1805         /*
1806          * There was no extent directly to the left or right of this new
1807          * extent then we know we're going to have to allocate a new extent, so
1808          * before we do that see if we need to drop this into a bitmap
1809          */
1810         ret = insert_into_bitmap(ctl, info);
1811         if (ret < 0) {
1812                 goto out;
1813         } else if (ret) {
1814                 ret = 0;
1815                 goto out;
1816         }
1817 link:
1818         ret = link_free_space(ctl, info);
1819         if (ret)
1820                 kmem_cache_free(btrfs_free_space_cachep, info);
1821 out:
1822         spin_unlock(&ctl->tree_lock);
1823
1824         if (ret) {
1825                 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1826                 BUG_ON(ret == -EEXIST);
1827         }
1828
1829         return ret;
1830 }
1831
1832 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1833                             u64 offset, u64 bytes)
1834 {
1835         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1836         struct btrfs_free_space *info;
1837         struct btrfs_free_space *next_info = NULL;
1838         int ret = 0;
1839
1840         spin_lock(&ctl->tree_lock);
1841
1842 again:
1843         info = tree_search_offset(ctl, offset, 0, 0);
1844         if (!info) {
1845                 /*
1846                  * oops didn't find an extent that matched the space we wanted
1847                  * to remove, look for a bitmap instead
1848                  */
1849                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1850                                           1, 0);
1851                 if (!info) {
1852                         /* the tree logging code might be calling us before we
1853                          * have fully loaded the free space rbtree for this
1854                          * block group.  So it is possible the entry won't
1855                          * be in the rbtree yet at all.  The caching code
1856                          * will make sure not to put it in the rbtree if
1857                          * the logging code has pinned it.
1858                          */
1859                         goto out_lock;
1860                 }
1861         }
1862
1863         if (info->bytes < bytes && rb_next(&info->offset_index)) {
1864                 u64 end;
1865                 next_info = rb_entry(rb_next(&info->offset_index),
1866                                              struct btrfs_free_space,
1867                                              offset_index);
1868
1869                 if (next_info->bitmap)
1870                         end = next_info->offset +
1871                               BITS_PER_BITMAP * ctl->unit - 1;
1872                 else
1873                         end = next_info->offset + next_info->bytes;
1874
1875                 if (next_info->bytes < bytes ||
1876                     next_info->offset > offset || offset > end) {
1877                         printk(KERN_CRIT "Found free space at %llu, size %llu,"
1878                               " trying to use %llu\n",
1879                               (unsigned long long)info->offset,
1880                               (unsigned long long)info->bytes,
1881                               (unsigned long long)bytes);
1882                         WARN_ON(1);
1883                         ret = -EINVAL;
1884                         goto out_lock;
1885                 }
1886
1887                 info = next_info;
1888         }
1889
1890         if (info->bytes == bytes) {
1891                 unlink_free_space(ctl, info);
1892                 if (info->bitmap) {
1893                         kfree(info->bitmap);
1894                         ctl->total_bitmaps--;
1895                 }
1896                 kmem_cache_free(btrfs_free_space_cachep, info);
1897                 ret = 0;
1898                 goto out_lock;
1899         }
1900
1901         if (!info->bitmap && info->offset == offset) {
1902                 unlink_free_space(ctl, info);
1903                 info->offset += bytes;
1904                 info->bytes -= bytes;
1905                 ret = link_free_space(ctl, info);
1906                 WARN_ON(ret);
1907                 goto out_lock;
1908         }
1909
1910         if (!info->bitmap && info->offset <= offset &&
1911             info->offset + info->bytes >= offset + bytes) {
1912                 u64 old_start = info->offset;
1913                 /*
1914                  * we're freeing space in the middle of the info,
1915                  * this can happen during tree log replay
1916                  *
1917                  * first unlink the old info and then
1918                  * insert it again after the hole we're creating
1919                  */
1920                 unlink_free_space(ctl, info);
1921                 if (offset + bytes < info->offset + info->bytes) {
1922                         u64 old_end = info->offset + info->bytes;
1923
1924                         info->offset = offset + bytes;
1925                         info->bytes = old_end - info->offset;
1926                         ret = link_free_space(ctl, info);
1927                         WARN_ON(ret);
1928                         if (ret)
1929                                 goto out_lock;
1930                 } else {
1931                         /* the hole we're creating ends at the end
1932                          * of the info struct, just free the info
1933                          */
1934                         kmem_cache_free(btrfs_free_space_cachep, info);
1935                 }
1936                 spin_unlock(&ctl->tree_lock);
1937
1938                 /* step two, insert a new info struct to cover
1939                  * anything before the hole
1940                  */
1941                 ret = btrfs_add_free_space(block_group, old_start,
1942                                            offset - old_start);
1943                 WARN_ON(ret);
1944                 goto out;
1945         }
1946
1947         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
1948         if (ret == -EAGAIN)
1949                 goto again;
1950         BUG_ON(ret);
1951 out_lock:
1952         spin_unlock(&ctl->tree_lock);
1953 out:
1954         return ret;
1955 }
1956
1957 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1958                            u64 bytes)
1959 {
1960         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1961         struct btrfs_free_space *info;
1962         struct rb_node *n;
1963         int count = 0;
1964
1965         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
1966                 info = rb_entry(n, struct btrfs_free_space, offset_index);
1967                 if (info->bytes >= bytes)
1968                         count++;
1969                 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1970                        (unsigned long long)info->offset,
1971                        (unsigned long long)info->bytes,
1972                        (info->bitmap) ? "yes" : "no");
1973         }
1974         printk(KERN_INFO "block group has cluster?: %s\n",
1975                list_empty(&block_group->cluster_list) ? "no" : "yes");
1976         printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1977                "\n", count);
1978 }
1979
1980 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
1981 {
1982         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1983
1984         spin_lock_init(&ctl->tree_lock);
1985         ctl->unit = block_group->sectorsize;
1986         ctl->start = block_group->key.objectid;
1987         ctl->private = block_group;
1988         ctl->op = &free_space_op;
1989
1990         /*
1991          * we only want to have 32k of ram per block group for keeping
1992          * track of free space, and if we pass 1/2 of that we want to
1993          * start converting things over to using bitmaps
1994          */
1995         ctl->extents_thresh = ((1024 * 32) / 2) /
1996                                 sizeof(struct btrfs_free_space);
1997 }
1998
1999 /*
2000  * for a given cluster, put all of its extents back into the free
2001  * space cache.  If the block group passed doesn't match the block group
2002  * pointed to by the cluster, someone else raced in and freed the
2003  * cluster already.  In that case, we just return without changing anything
2004  */
2005 static int
2006 __btrfs_return_cluster_to_free_space(
2007                              struct btrfs_block_group_cache *block_group,
2008                              struct btrfs_free_cluster *cluster)
2009 {
2010         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2011         struct btrfs_free_space *entry;
2012         struct rb_node *node;
2013
2014         spin_lock(&cluster->lock);
2015         if (cluster->block_group != block_group)
2016                 goto out;
2017
2018         cluster->block_group = NULL;
2019         cluster->window_start = 0;
2020         list_del_init(&cluster->block_group_list);
2021
2022         node = rb_first(&cluster->root);
2023         while (node) {
2024                 bool bitmap;
2025
2026                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2027                 node = rb_next(&entry->offset_index);
2028                 rb_erase(&entry->offset_index, &cluster->root);
2029
2030                 bitmap = (entry->bitmap != NULL);
2031                 if (!bitmap)
2032                         try_merge_free_space(ctl, entry, false);
2033                 tree_insert_offset(&ctl->free_space_offset,
2034                                    entry->offset, &entry->offset_index, bitmap);
2035         }
2036         cluster->root = RB_ROOT;
2037
2038 out:
2039         spin_unlock(&cluster->lock);
2040         btrfs_put_block_group(block_group);
2041         return 0;
2042 }
2043
2044 void __btrfs_remove_free_space_cache_locked(struct btrfs_free_space_ctl *ctl)
2045 {
2046         struct btrfs_free_space *info;
2047         struct rb_node *node;
2048
2049         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2050                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2051                 if (!info->bitmap) {
2052                         unlink_free_space(ctl, info);
2053                         kmem_cache_free(btrfs_free_space_cachep, info);
2054                 } else {
2055                         free_bitmap(ctl, info);
2056                 }
2057                 if (need_resched()) {
2058                         spin_unlock(&ctl->tree_lock);
2059                         cond_resched();
2060                         spin_lock(&ctl->tree_lock);
2061                 }
2062         }
2063 }
2064
2065 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2066 {
2067         spin_lock(&ctl->tree_lock);
2068         __btrfs_remove_free_space_cache_locked(ctl);
2069         spin_unlock(&ctl->tree_lock);
2070 }
2071
2072 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2073 {
2074         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2075         struct btrfs_free_cluster *cluster;
2076         struct list_head *head;
2077
2078         spin_lock(&ctl->tree_lock);
2079         while ((head = block_group->cluster_list.next) !=
2080                &block_group->cluster_list) {
2081                 cluster = list_entry(head, struct btrfs_free_cluster,
2082                                      block_group_list);
2083
2084                 WARN_ON(cluster->block_group != block_group);
2085                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2086                 if (need_resched()) {
2087                         spin_unlock(&ctl->tree_lock);
2088                         cond_resched();
2089                         spin_lock(&ctl->tree_lock);
2090                 }
2091         }
2092         __btrfs_remove_free_space_cache_locked(ctl);
2093         spin_unlock(&ctl->tree_lock);
2094
2095 }
2096
2097 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2098                                u64 offset, u64 bytes, u64 empty_size)
2099 {
2100         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2101         struct btrfs_free_space *entry = NULL;
2102         u64 bytes_search = bytes + empty_size;
2103         u64 ret = 0;
2104
2105         spin_lock(&ctl->tree_lock);
2106         entry = find_free_space(ctl, &offset, &bytes_search);
2107         if (!entry)
2108                 goto out;
2109
2110         ret = offset;
2111         if (entry->bitmap) {
2112                 bitmap_clear_bits(ctl, entry, offset, bytes);
2113                 if (!entry->bytes)
2114                         free_bitmap(ctl, entry);
2115         } else {
2116                 unlink_free_space(ctl, entry);
2117                 entry->offset += bytes;
2118                 entry->bytes -= bytes;
2119                 if (!entry->bytes)
2120                         kmem_cache_free(btrfs_free_space_cachep, entry);
2121                 else
2122                         link_free_space(ctl, entry);
2123         }
2124
2125 out:
2126         spin_unlock(&ctl->tree_lock);
2127
2128         return ret;
2129 }
2130
2131 /*
2132  * given a cluster, put all of its extents back into the free space
2133  * cache.  If a block group is passed, this function will only free
2134  * a cluster that belongs to the passed block group.
2135  *
2136  * Otherwise, it'll get a reference on the block group pointed to by the
2137  * cluster and remove the cluster from it.
2138  */
2139 int btrfs_return_cluster_to_free_space(
2140                                struct btrfs_block_group_cache *block_group,
2141                                struct btrfs_free_cluster *cluster)
2142 {
2143         struct btrfs_free_space_ctl *ctl;
2144         int ret;
2145
2146         /* first, get a safe pointer to the block group */
2147         spin_lock(&cluster->lock);
2148         if (!block_group) {
2149                 block_group = cluster->block_group;
2150                 if (!block_group) {
2151                         spin_unlock(&cluster->lock);
2152                         return 0;
2153                 }
2154         } else if (cluster->block_group != block_group) {
2155                 /* someone else has already freed it don't redo their work */
2156                 spin_unlock(&cluster->lock);
2157                 return 0;
2158         }
2159         atomic_inc(&block_group->count);
2160         spin_unlock(&cluster->lock);
2161
2162         ctl = block_group->free_space_ctl;
2163
2164         /* now return any extents the cluster had on it */
2165         spin_lock(&ctl->tree_lock);
2166         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2167         spin_unlock(&ctl->tree_lock);
2168
2169         /* finally drop our ref */
2170         btrfs_put_block_group(block_group);
2171         return ret;
2172 }
2173
2174 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2175                                    struct btrfs_free_cluster *cluster,
2176                                    struct btrfs_free_space *entry,
2177                                    u64 bytes, u64 min_start)
2178 {
2179         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2180         int err;
2181         u64 search_start = cluster->window_start;
2182         u64 search_bytes = bytes;
2183         u64 ret = 0;
2184
2185         search_start = min_start;
2186         search_bytes = bytes;
2187
2188         err = search_bitmap(ctl, entry, &search_start, &search_bytes);
2189         if (err)
2190                 return 0;
2191
2192         ret = search_start;
2193         __bitmap_clear_bits(ctl, entry, ret, bytes);
2194
2195         return ret;
2196 }
2197
2198 /*
2199  * given a cluster, try to allocate 'bytes' from it, returns 0
2200  * if it couldn't find anything suitably large, or a logical disk offset
2201  * if things worked out
2202  */
2203 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2204                              struct btrfs_free_cluster *cluster, u64 bytes,
2205                              u64 min_start)
2206 {
2207         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2208         struct btrfs_free_space *entry = NULL;
2209         struct rb_node *node;
2210         u64 ret = 0;
2211
2212         spin_lock(&cluster->lock);
2213         if (bytes > cluster->max_size)
2214                 goto out;
2215
2216         if (cluster->block_group != block_group)
2217                 goto out;
2218
2219         node = rb_first(&cluster->root);
2220         if (!node)
2221                 goto out;
2222
2223         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2224         while(1) {
2225                 if (entry->bytes < bytes ||
2226                     (!entry->bitmap && entry->offset < min_start)) {
2227                         node = rb_next(&entry->offset_index);
2228                         if (!node)
2229                                 break;
2230                         entry = rb_entry(node, struct btrfs_free_space,
2231                                          offset_index);
2232                         continue;
2233                 }
2234
2235                 if (entry->bitmap) {
2236                         ret = btrfs_alloc_from_bitmap(block_group,
2237                                                       cluster, entry, bytes,
2238                                                       min_start);
2239                         if (ret == 0) {
2240                                 node = rb_next(&entry->offset_index);
2241                                 if (!node)
2242                                         break;
2243                                 entry = rb_entry(node, struct btrfs_free_space,
2244                                                  offset_index);
2245                                 continue;
2246                         }
2247                 } else {
2248                         ret = entry->offset;
2249
2250                         entry->offset += bytes;
2251                         entry->bytes -= bytes;
2252                 }
2253
2254                 if (entry->bytes == 0)
2255                         rb_erase(&entry->offset_index, &cluster->root);
2256                 break;
2257         }
2258 out:
2259         spin_unlock(&cluster->lock);
2260
2261         if (!ret)
2262                 return 0;
2263
2264         spin_lock(&ctl->tree_lock);
2265
2266         ctl->free_space -= bytes;
2267         if (entry->bytes == 0) {
2268                 ctl->free_extents--;
2269                 if (entry->bitmap) {
2270                         kfree(entry->bitmap);
2271                         ctl->total_bitmaps--;
2272                         ctl->op->recalc_thresholds(ctl);
2273                 }
2274                 kmem_cache_free(btrfs_free_space_cachep, entry);
2275         }
2276
2277         spin_unlock(&ctl->tree_lock);
2278
2279         return ret;
2280 }
2281
2282 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2283                                 struct btrfs_free_space *entry,
2284                                 struct btrfs_free_cluster *cluster,
2285                                 u64 offset, u64 bytes, u64 min_bytes)
2286 {
2287         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2288         unsigned long next_zero;
2289         unsigned long i;
2290         unsigned long search_bits;
2291         unsigned long total_bits;
2292         unsigned long found_bits;
2293         unsigned long start = 0;
2294         unsigned long total_found = 0;
2295         int ret;
2296         bool found = false;
2297
2298         i = offset_to_bit(entry->offset, block_group->sectorsize,
2299                           max_t(u64, offset, entry->offset));
2300         search_bits = bytes_to_bits(bytes, block_group->sectorsize);
2301         total_bits = bytes_to_bits(min_bytes, block_group->sectorsize);
2302
2303 again:
2304         found_bits = 0;
2305         for (i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i);
2306              i < BITS_PER_BITMAP;
2307              i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, i + 1)) {
2308                 next_zero = find_next_zero_bit(entry->bitmap,
2309                                                BITS_PER_BITMAP, i);
2310                 if (next_zero - i >= search_bits) {
2311                         found_bits = next_zero - i;
2312                         break;
2313                 }
2314                 i = next_zero;
2315         }
2316
2317         if (!found_bits)
2318                 return -ENOSPC;
2319
2320         if (!found) {
2321                 start = i;
2322                 found = true;
2323         }
2324
2325         total_found += found_bits;
2326
2327         if (cluster->max_size < found_bits * block_group->sectorsize)
2328                 cluster->max_size = found_bits * block_group->sectorsize;
2329
2330         if (total_found < total_bits) {
2331                 i = find_next_bit(entry->bitmap, BITS_PER_BITMAP, next_zero);
2332                 if (i - start > total_bits * 2) {
2333                         total_found = 0;
2334                         cluster->max_size = 0;
2335                         found = false;
2336                 }
2337                 goto again;
2338         }
2339
2340         cluster->window_start = start * block_group->sectorsize +
2341                 entry->offset;
2342         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2343         ret = tree_insert_offset(&cluster->root, entry->offset,
2344                                  &entry->offset_index, 1);
2345         BUG_ON(ret);
2346
2347         return 0;
2348 }
2349
2350 /*
2351  * This searches the block group for just extents to fill the cluster with.
2352  */
2353 static noinline int
2354 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2355                         struct btrfs_free_cluster *cluster,
2356                         struct list_head *bitmaps, u64 offset, u64 bytes,
2357                         u64 min_bytes)
2358 {
2359         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2360         struct btrfs_free_space *first = NULL;
2361         struct btrfs_free_space *entry = NULL;
2362         struct btrfs_free_space *prev = NULL;
2363         struct btrfs_free_space *last;
2364         struct rb_node *node;
2365         u64 window_start;
2366         u64 window_free;
2367         u64 max_extent;
2368         u64 max_gap = 128 * 1024;
2369
2370         entry = tree_search_offset(ctl, offset, 0, 1);
2371         if (!entry)
2372                 return -ENOSPC;
2373
2374         /*
2375          * We don't want bitmaps, so just move along until we find a normal
2376          * extent entry.
2377          */
2378         while (entry->bitmap) {
2379                 if (list_empty(&entry->list))
2380                         list_add_tail(&entry->list, bitmaps);
2381                 node = rb_next(&entry->offset_index);
2382                 if (!node)
2383                         return -ENOSPC;
2384                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2385         }
2386
2387         window_start = entry->offset;
2388         window_free = entry->bytes;
2389         max_extent = entry->bytes;
2390         first = entry;
2391         last = entry;
2392         prev = entry;
2393
2394         while (window_free <= min_bytes) {
2395                 node = rb_next(&entry->offset_index);
2396                 if (!node)
2397                         return -ENOSPC;
2398                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2399
2400                 if (entry->bitmap) {
2401                         if (list_empty(&entry->list))
2402                                 list_add_tail(&entry->list, bitmaps);
2403                         continue;
2404                 }
2405
2406                 /*
2407                  * we haven't filled the empty size and the window is
2408                  * very large.  reset and try again
2409                  */
2410                 if (entry->offset - (prev->offset + prev->bytes) > max_gap ||
2411                     entry->offset - window_start > (min_bytes * 2)) {
2412                         first = entry;
2413                         window_start = entry->offset;
2414                         window_free = entry->bytes;
2415                         last = entry;
2416                         max_extent = entry->bytes;
2417                 } else {
2418                         last = entry;
2419                         window_free += entry->bytes;
2420                         if (entry->bytes > max_extent)
2421                                 max_extent = entry->bytes;
2422                 }
2423                 prev = entry;
2424         }
2425
2426         cluster->window_start = first->offset;
2427
2428         node = &first->offset_index;
2429
2430         /*
2431          * now we've found our entries, pull them out of the free space
2432          * cache and put them into the cluster rbtree
2433          */
2434         do {
2435                 int ret;
2436
2437                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2438                 node = rb_next(&entry->offset_index);
2439                 if (entry->bitmap)
2440                         continue;
2441
2442                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2443                 ret = tree_insert_offset(&cluster->root, entry->offset,
2444                                          &entry->offset_index, 0);
2445                 BUG_ON(ret);
2446         } while (node && entry != last);
2447
2448         cluster->max_size = max_extent;
2449
2450         return 0;
2451 }
2452
2453 /*
2454  * This specifically looks for bitmaps that may work in the cluster, we assume
2455  * that we have already failed to find extents that will work.
2456  */
2457 static noinline int
2458 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2459                      struct btrfs_free_cluster *cluster,
2460                      struct list_head *bitmaps, u64 offset, u64 bytes,
2461                      u64 min_bytes)
2462 {
2463         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2464         struct btrfs_free_space *entry;
2465         int ret = -ENOSPC;
2466         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2467
2468         if (ctl->total_bitmaps == 0)
2469                 return -ENOSPC;
2470
2471         /*
2472          * The bitmap that covers offset won't be in the list unless offset
2473          * is just its start offset.
2474          */
2475         entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2476         if (entry->offset != bitmap_offset) {
2477                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2478                 if (entry && list_empty(&entry->list))
2479                         list_add(&entry->list, bitmaps);
2480         }
2481
2482         list_for_each_entry(entry, bitmaps, list) {
2483                 if (entry->bytes < min_bytes)
2484                         continue;
2485                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2486                                            bytes, min_bytes);
2487                 if (!ret)
2488                         return 0;
2489         }
2490
2491         /*
2492          * The bitmaps list has all the bitmaps that record free space
2493          * starting after offset, so no more search is required.
2494          */
2495         return -ENOSPC;
2496 }
2497
2498 /*
2499  * here we try to find a cluster of blocks in a block group.  The goal
2500  * is to find at least bytes free and up to empty_size + bytes free.
2501  * We might not find them all in one contiguous area.
2502  *
2503  * returns zero and sets up cluster if things worked out, otherwise
2504  * it returns -enospc
2505  */
2506 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
2507                              struct btrfs_root *root,
2508                              struct btrfs_block_group_cache *block_group,
2509                              struct btrfs_free_cluster *cluster,
2510                              u64 offset, u64 bytes, u64 empty_size)
2511 {
2512         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2513         struct btrfs_free_space *entry, *tmp;
2514         LIST_HEAD(bitmaps);
2515         u64 min_bytes;
2516         int ret;
2517
2518         /* for metadata, allow allocates with more holes */
2519         if (btrfs_test_opt(root, SSD_SPREAD)) {
2520                 min_bytes = bytes + empty_size;
2521         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2522                 /*
2523                  * we want to do larger allocations when we are
2524                  * flushing out the delayed refs, it helps prevent
2525                  * making more work as we go along.
2526                  */
2527                 if (trans->transaction->delayed_refs.flushing)
2528                         min_bytes = max(bytes, (bytes + empty_size) >> 1);
2529                 else
2530                         min_bytes = max(bytes, (bytes + empty_size) >> 4);
2531         } else
2532                 min_bytes = max(bytes, (bytes + empty_size) >> 2);
2533
2534         spin_lock(&ctl->tree_lock);
2535
2536         /*
2537          * If we know we don't have enough space to make a cluster don't even
2538          * bother doing all the work to try and find one.
2539          */
2540         if (ctl->free_space < min_bytes) {
2541                 spin_unlock(&ctl->tree_lock);
2542                 return -ENOSPC;
2543         }
2544
2545         spin_lock(&cluster->lock);
2546
2547         /* someone already found a cluster, hooray */
2548         if (cluster->block_group) {
2549                 ret = 0;
2550                 goto out;
2551         }
2552
2553         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
2554                                       bytes, min_bytes);
2555         if (ret)
2556                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
2557                                            offset, bytes, min_bytes);
2558
2559         /* Clear our temporary list */
2560         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
2561                 list_del_init(&entry->list);
2562
2563         if (!ret) {
2564                 atomic_inc(&block_group->count);
2565                 list_add_tail(&cluster->block_group_list,
2566                               &block_group->cluster_list);
2567                 cluster->block_group = block_group;
2568         }
2569 out:
2570         spin_unlock(&cluster->lock);
2571         spin_unlock(&ctl->tree_lock);
2572
2573         return ret;
2574 }
2575
2576 /*
2577  * simple code to zero out a cluster
2578  */
2579 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2580 {
2581         spin_lock_init(&cluster->lock);
2582         spin_lock_init(&cluster->refill_lock);
2583         cluster->root = RB_ROOT;
2584         cluster->max_size = 0;
2585         INIT_LIST_HEAD(&cluster->block_group_list);
2586         cluster->block_group = NULL;
2587 }
2588
2589 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
2590                            u64 *trimmed, u64 start, u64 end, u64 minlen)
2591 {
2592         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2593         struct btrfs_free_space *entry = NULL;
2594         struct btrfs_fs_info *fs_info = block_group->fs_info;
2595         u64 bytes = 0;
2596         u64 actually_trimmed;
2597         int ret = 0;
2598
2599         *trimmed = 0;
2600
2601         while (start < end) {
2602                 spin_lock(&ctl->tree_lock);
2603
2604                 if (ctl->free_space < minlen) {
2605                         spin_unlock(&ctl->tree_lock);
2606                         break;
2607                 }
2608
2609                 entry = tree_search_offset(ctl, start, 0, 1);
2610                 if (!entry)
2611                         entry = tree_search_offset(ctl,
2612                                                    offset_to_bitmap(ctl, start),
2613                                                    1, 1);
2614
2615                 if (!entry || entry->offset >= end) {
2616                         spin_unlock(&ctl->tree_lock);
2617                         break;
2618                 }
2619
2620                 if (entry->bitmap) {
2621                         ret = search_bitmap(ctl, entry, &start, &bytes);
2622                         if (!ret) {
2623                                 if (start >= end) {
2624                                         spin_unlock(&ctl->tree_lock);
2625                                         break;
2626                                 }
2627                                 bytes = min(bytes, end - start);
2628                                 bitmap_clear_bits(ctl, entry, start, bytes);
2629                                 if (entry->bytes == 0)
2630                                         free_bitmap(ctl, entry);
2631                         } else {
2632                                 start = entry->offset + BITS_PER_BITMAP *
2633                                         block_group->sectorsize;
2634                                 spin_unlock(&ctl->tree_lock);
2635                                 ret = 0;
2636                                 continue;
2637                         }
2638                 } else {
2639                         start = entry->offset;
2640                         bytes = min(entry->bytes, end - start);
2641                         unlink_free_space(ctl, entry);
2642                         kmem_cache_free(btrfs_free_space_cachep, entry);
2643                 }
2644
2645                 spin_unlock(&ctl->tree_lock);
2646
2647                 if (bytes >= minlen) {
2648                         struct btrfs_space_info *space_info;
2649                         int update = 0;
2650
2651                         space_info = block_group->space_info;
2652                         spin_lock(&space_info->lock);
2653                         spin_lock(&block_group->lock);
2654                         if (!block_group->ro) {
2655                                 block_group->reserved += bytes;
2656                                 space_info->bytes_reserved += bytes;
2657                                 update = 1;
2658                         }
2659                         spin_unlock(&block_group->lock);
2660                         spin_unlock(&space_info->lock);
2661
2662                         ret = btrfs_error_discard_extent(fs_info->extent_root,
2663                                                          start,
2664                                                          bytes,
2665                                                          &actually_trimmed);
2666
2667                         btrfs_add_free_space(block_group, start, bytes);
2668                         if (update) {
2669                                 spin_lock(&space_info->lock);
2670                                 spin_lock(&block_group->lock);
2671                                 if (block_group->ro)
2672                                         space_info->bytes_readonly += bytes;
2673                                 block_group->reserved -= bytes;
2674                                 space_info->bytes_reserved -= bytes;
2675                                 spin_unlock(&space_info->lock);
2676                                 spin_unlock(&block_group->lock);
2677                         }
2678
2679                         if (ret)
2680                                 break;
2681                         *trimmed += actually_trimmed;
2682                 }
2683                 start += bytes;
2684                 bytes = 0;
2685
2686                 if (fatal_signal_pending(current)) {
2687                         ret = -ERESTARTSYS;
2688                         break;
2689                 }
2690
2691                 cond_resched();
2692         }
2693
2694         return ret;
2695 }
2696
2697 /*
2698  * Find the left-most item in the cache tree, and then return the
2699  * smallest inode number in the item.
2700  *
2701  * Note: the returned inode number may not be the smallest one in
2702  * the tree, if the left-most item is a bitmap.
2703  */
2704 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
2705 {
2706         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
2707         struct btrfs_free_space *entry = NULL;
2708         u64 ino = 0;
2709
2710         spin_lock(&ctl->tree_lock);
2711
2712         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
2713                 goto out;
2714
2715         entry = rb_entry(rb_first(&ctl->free_space_offset),
2716                          struct btrfs_free_space, offset_index);
2717
2718         if (!entry->bitmap) {
2719                 ino = entry->offset;
2720
2721                 unlink_free_space(ctl, entry);
2722                 entry->offset++;
2723                 entry->bytes--;
2724                 if (!entry->bytes)
2725                         kmem_cache_free(btrfs_free_space_cachep, entry);
2726                 else
2727                         link_free_space(ctl, entry);
2728         } else {
2729                 u64 offset = 0;
2730                 u64 count = 1;
2731                 int ret;
2732
2733                 ret = search_bitmap(ctl, entry, &offset, &count);
2734                 BUG_ON(ret);
2735
2736                 ino = offset;
2737                 bitmap_clear_bits(ctl, entry, offset, 1);
2738                 if (entry->bytes == 0)
2739                         free_bitmap(ctl, entry);
2740         }
2741 out:
2742         spin_unlock(&ctl->tree_lock);
2743
2744         return ino;
2745 }
2746
2747 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
2748                                     struct btrfs_path *path)
2749 {
2750         struct inode *inode = NULL;
2751
2752         spin_lock(&root->cache_lock);
2753         if (root->cache_inode)
2754                 inode = igrab(root->cache_inode);
2755         spin_unlock(&root->cache_lock);
2756         if (inode)
2757                 return inode;
2758
2759         inode = __lookup_free_space_inode(root, path, 0);
2760         if (IS_ERR(inode))
2761                 return inode;
2762
2763         spin_lock(&root->cache_lock);
2764         if (!btrfs_fs_closing(root->fs_info))
2765                 root->cache_inode = igrab(inode);
2766         spin_unlock(&root->cache_lock);
2767
2768         return inode;
2769 }
2770
2771 int create_free_ino_inode(struct btrfs_root *root,
2772                           struct btrfs_trans_handle *trans,
2773                           struct btrfs_path *path)
2774 {
2775         return __create_free_space_inode(root, trans, path,
2776                                          BTRFS_FREE_INO_OBJECTID, 0);
2777 }
2778
2779 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
2780 {
2781         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2782         struct btrfs_path *path;
2783         struct inode *inode;
2784         int ret = 0;
2785         u64 root_gen = btrfs_root_generation(&root->root_item);
2786
2787         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2788                 return 0;
2789
2790         /*
2791          * If we're unmounting then just return, since this does a search on the
2792          * normal root and not the commit root and we could deadlock.
2793          */
2794         if (btrfs_fs_closing(fs_info))
2795                 return 0;
2796
2797         path = btrfs_alloc_path();
2798         if (!path)
2799                 return 0;
2800
2801         inode = lookup_free_ino_inode(root, path);
2802         if (IS_ERR(inode))
2803                 goto out;
2804
2805         if (root_gen != BTRFS_I(inode)->generation)
2806                 goto out_put;
2807
2808         ret = __load_free_space_cache(root, inode, ctl, path, 0);
2809
2810         if (ret < 0)
2811                 printk(KERN_ERR "btrfs: failed to load free ino cache for "
2812                        "root %llu\n", root->root_key.objectid);
2813 out_put:
2814         iput(inode);
2815 out:
2816         btrfs_free_path(path);
2817         return ret;
2818 }
2819
2820 int btrfs_write_out_ino_cache(struct btrfs_root *root,
2821                               struct btrfs_trans_handle *trans,
2822                               struct btrfs_path *path)
2823 {
2824         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2825         struct inode *inode;
2826         int ret;
2827
2828         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2829                 return 0;
2830
2831         inode = lookup_free_ino_inode(root, path);
2832         if (IS_ERR(inode))
2833                 return 0;
2834
2835         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
2836         if (ret) {
2837                 btrfs_delalloc_release_metadata(inode, inode->i_size);
2838 #ifdef DEBUG
2839                 printk(KERN_ERR "btrfs: failed to write free ino cache "
2840                        "for root %llu\n", root->root_key.objectid);
2841 #endif
2842         }
2843
2844         iput(inode);
2845         return ret;
2846 }