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 #include "volumes.h"
  31
  32 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  33 #define MAX_CACHE_BYTES_PER_GIG SZ_32K
  34
  35 struct btrfs_trim_range {
  36         u64 start;
  37         u64 bytes;
  38         struct list_head list;
  39 };
  40
  41 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  42                            struct btrfs_free_space *info);
  43 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  44                               struct btrfs_free_space *info);
  45
  46 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  47                                                struct btrfs_path *path,
  48                                                u64 offset)
  49 {
  50         struct btrfs_key key;
  51         struct btrfs_key location;
  52         struct btrfs_disk_key disk_key;
  53         struct btrfs_free_space_header *header;
  54         struct extent_buffer *leaf;
  55         struct inode *inode = NULL;
  56         int ret;
  57
  58         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  59         key.offset = offset;
  60         key.type = 0;
  61
  62         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  63         if (ret < 0)
  64                 return ERR_PTR(ret);
  65         if (ret > 0) {
  66                 btrfs_release_path(path);
  67                 return ERR_PTR(-ENOENT);
  68         }
  69
  70         leaf = path->nodes[0];
  71         header = btrfs_item_ptr(leaf, path->slots[0],
  72                                 struct btrfs_free_space_header);
  73         btrfs_free_space_key(leaf, header, &disk_key);
  74         btrfs_disk_key_to_cpu(&location, &disk_key);
  75         btrfs_release_path(path);
  76
  77         inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
  78         if (!inode)
  79                 return ERR_PTR(-ENOENT);
  80         if (IS_ERR(inode))
  81                 return inode;
  82         if (is_bad_inode(inode)) {
  83                 iput(inode);
  84                 return ERR_PTR(-ENOENT);
  85         }
  86
  87         mapping_set_gfp_mask(inode->i_mapping,
  88                         mapping_gfp_constraint(inode->i_mapping,
  89                         ~(__GFP_FS | __GFP_HIGHMEM)));
  90
  91         return inode;
  92 }
  93
  94 struct inode *lookup_free_space_inode(struct btrfs_root *root,
  95                                       struct btrfs_block_group_cache
  96                                       *block_group, struct btrfs_path *path)
  97 {
  98         struct inode *inode = NULL;
  99         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 100
 101         spin_lock(&block_group->lock);
 102         if (block_group->inode)
 103                 inode = igrab(block_group->inode);
 104         spin_unlock(&block_group->lock);
 105         if (inode)
 106                 return inode;
 107
 108         inode = __lookup_free_space_inode(root, path,
 109                                           block_group->key.objectid);
 110         if (IS_ERR(inode))
 111                 return inode;
 112
 113         spin_lock(&block_group->lock);
 114         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 115                 btrfs_info(root->fs_info,
 116                         "Old style space inode found, converting.");
 117                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 118                         BTRFS_INODE_NODATACOW;
 119                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 120         }
 121
 122         if (!block_group->iref) {
 123                 block_group->inode = igrab(inode);
 124                 block_group->iref = 1;
 125         }
 126         spin_unlock(&block_group->lock);
 127
 128         return inode;
 129 }
 130
 131 static int __create_free_space_inode(struct btrfs_root *root,
 132                                      struct btrfs_trans_handle *trans,
 133                                      struct btrfs_path *path,
 134                                      u64 ino, u64 offset)
 135 {
 136         struct btrfs_key key;
 137         struct btrfs_disk_key disk_key;
 138         struct btrfs_free_space_header *header;
 139         struct btrfs_inode_item *inode_item;
 140         struct extent_buffer *leaf;
 141         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 142         int ret;
 143
 144         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 145         if (ret)
 146                 return ret;
 147
 148         /* We inline crc's for the free disk space cache */
 149         if (ino != BTRFS_FREE_INO_OBJECTID)
 150                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 151
 152         leaf = path->nodes[0];
 153         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 154                                     struct btrfs_inode_item);
 155         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 156         memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
 157                              sizeof(*inode_item));
 158         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 159         btrfs_set_inode_size(leaf, inode_item, 0);
 160         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 161         btrfs_set_inode_uid(leaf, inode_item, 0);
 162         btrfs_set_inode_gid(leaf, inode_item, 0);
 163         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 164         btrfs_set_inode_flags(leaf, inode_item, flags);
 165         btrfs_set_inode_nlink(leaf, inode_item, 1);
 166         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 167         btrfs_set_inode_block_group(leaf, inode_item, offset);
 168         btrfs_mark_buffer_dirty(leaf);
 169         btrfs_release_path(path);
 170
 171         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 172         key.offset = offset;
 173         key.type = 0;
 174         ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                       sizeof(struct btrfs_free_space_header));
 176         if (ret < 0) {
 177                 btrfs_release_path(path);
 178                 return ret;
 179         }
 180
 181         leaf = path->nodes[0];
 182         header = btrfs_item_ptr(leaf, path->slots[0],
 183                                 struct btrfs_free_space_header);
 184         memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
 185         btrfs_set_free_space_key(leaf, header, &disk_key);
 186         btrfs_mark_buffer_dirty(leaf);
 187         btrfs_release_path(path);
 188
 189         return 0;
 190 }
 191
 192 int create_free_space_inode(struct btrfs_root *root,
 193                             struct btrfs_trans_handle *trans,
 194                             struct btrfs_block_group_cache *block_group,
 195                             struct btrfs_path *path)
 196 {
 197         int ret;
 198         u64 ino;
 199
 200         ret = btrfs_find_free_objectid(root, &ino);
 201         if (ret < 0)
 202                 return ret;
 203
 204         return __create_free_space_inode(root, trans, path, ino,
 205                                          block_group->key.objectid);
 206 }
 207
 208 int btrfs_check_trunc_cache_free_space(struct btrfs_root *root,
 209                                        struct btrfs_block_rsv *rsv)
 210 {
 211         u64 needed_bytes;
 212         int ret;
 213
 214         /* 1 for slack space, 1 for updating the inode */
 215         needed_bytes = btrfs_calc_trunc_metadata_size(root, 1) +
 216                 btrfs_calc_trans_metadata_size(root, 1);
 217
 218         spin_lock(&rsv->lock);
 219         if (rsv->reserved < needed_bytes)
 220                 ret = -ENOSPC;
 221         else
 222                 ret = 0;
 223         spin_unlock(&rsv->lock);
 224         return ret;
 225 }
 226
 227 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
 228                                     struct btrfs_trans_handle *trans,
 229                                     struct btrfs_block_group_cache *block_group,
 230                                     struct inode *inode)
 231 {
 232         int ret = 0;
 233         struct btrfs_path *path = btrfs_alloc_path();
 234         bool locked = false;
 235
 236         if (!path) {
 237                 ret = -ENOMEM;
 238                 goto fail;
 239         }
 240
 241         if (block_group) {
 242                 locked = true;
 243                 mutex_lock(&trans->transaction->cache_write_mutex);
 244                 if (!list_empty(&block_group->io_list)) {
 245                         list_del_init(&block_group->io_list);
 246
 247                         btrfs_wait_cache_io(root, trans, block_group,
 248                                             &block_group->io_ctl, path,
 249                                             block_group->key.objectid);
 250                         btrfs_put_block_group(block_group);
 251                 }
 252
 253                 /*
 254                  * now that we've truncated the cache away, its no longer
 255                  * setup or written
 256                  */
 257                 spin_lock(&block_group->lock);
 258                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 259                 spin_unlock(&block_group->lock);
 260         }
 261         btrfs_free_path(path);
 262
 263         btrfs_i_size_write(inode, 0);
 264         truncate_pagecache(inode, 0);
 265
 266         /*
 267          * We don't need an orphan item because truncating the free space cache
 268          * will never be split across transactions.
 269          * We don't need to check for -EAGAIN because we're a free space
 270          * cache inode
 271          */
 272         ret = btrfs_truncate_inode_items(trans, root, inode,
 273                                          0, BTRFS_EXTENT_DATA_KEY);
 274         if (ret)
 275                 goto fail;
 276
 277         ret = btrfs_update_inode(trans, root, inode);
 278
 279 fail:
 280         if (locked)
 281                 mutex_unlock(&trans->transaction->cache_write_mutex);
 282         if (ret)
 283                 btrfs_abort_transaction(trans, ret);
 284
 285         return ret;
 286 }
 287
 288 static int readahead_cache(struct inode *inode)
 289 {
 290         struct file_ra_state *ra;
 291         unsigned long last_index;
 292
 293         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 294         if (!ra)
 295                 return -ENOMEM;
 296
 297         file_ra_state_init(ra, inode->i_mapping);
 298         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 299
 300         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 301
 302         kfree(ra);
 303
 304         return 0;
 305 }
 306
 307 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 308                        struct btrfs_root *root, int write)
 309 {
 310         int num_pages;
 311         int check_crcs = 0;
 312
 313         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 314
 315         if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
 316                 check_crcs = 1;
 317
 318         /* Make sure we can fit our crcs into the first page */
 319         if (write && check_crcs &&
 320             (num_pages * sizeof(u32)) >= PAGE_SIZE)
 321                 return -ENOSPC;
 322
 323         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 324
 325         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 326         if (!io_ctl->pages)
 327                 return -ENOMEM;
 328
 329         io_ctl->num_pages = num_pages;
 330         io_ctl->root = root;
 331         io_ctl->check_crcs = check_crcs;
 332         io_ctl->inode = inode;
 333
 334         return 0;
 335 }
 336
 337 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 338 {
 339         kfree(io_ctl->pages);
 340         io_ctl->pages = NULL;
 341 }
 342
 343 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 344 {
 345         if (io_ctl->cur) {
 346                 io_ctl->cur = NULL;
 347                 io_ctl->orig = NULL;
 348         }
 349 }
 350
 351 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 352 {
 353         ASSERT(io_ctl->index < io_ctl->num_pages);
 354         io_ctl->page = io_ctl->pages[io_ctl->index++];
 355         io_ctl->cur = page_address(io_ctl->page);
 356         io_ctl->orig = io_ctl->cur;
 357         io_ctl->size = PAGE_SIZE;
 358         if (clear)
 359                 memset(io_ctl->cur, 0, PAGE_SIZE);
 360 }
 361
 362 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 363 {
 364         int i;
 365
 366         io_ctl_unmap_page(io_ctl);
 367
 368         for (i = 0; i < io_ctl->num_pages; i++) {
 369                 if (io_ctl->pages[i]) {
 370                         ClearPageChecked(io_ctl->pages[i]);
 371                         unlock_page(io_ctl->pages[i]);
 372                         put_page(io_ctl->pages[i]);
 373                 }
 374         }
 375 }
 376
 377 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 378                                 int uptodate)
 379 {
 380         struct page *page;
 381         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 382         int i;
 383
 384         for (i = 0; i < io_ctl->num_pages; i++) {
 385                 page = find_or_create_page(inode->i_mapping, i, mask);
 386                 if (!page) {
 387                         io_ctl_drop_pages(io_ctl);
 388                         return -ENOMEM;
 389                 }
 390                 io_ctl->pages[i] = page;
 391                 if (uptodate && !PageUptodate(page)) {
 392                         btrfs_readpage(NULL, page);
 393                         lock_page(page);
 394                         if (!PageUptodate(page)) {
 395                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 396                                            "error reading free space cache");
 397                                 io_ctl_drop_pages(io_ctl);
 398                                 return -EIO;
 399                         }
 400                 }
 401         }
 402
 403         for (i = 0; i < io_ctl->num_pages; i++) {
 404                 clear_page_dirty_for_io(io_ctl->pages[i]);
 405                 set_page_extent_mapped(io_ctl->pages[i]);
 406         }
 407
 408         return 0;
 409 }
 410
 411 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 412 {
 413         __le64 *val;
 414
 415         io_ctl_map_page(io_ctl, 1);
 416
 417         /*
 418          * Skip the csum areas.  If we don't check crcs then we just have a
 419          * 64bit chunk at the front of the first page.
 420          */
 421         if (io_ctl->check_crcs) {
 422                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 423                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 424         } else {
 425                 io_ctl->cur += sizeof(u64);
 426                 io_ctl->size -= sizeof(u64) * 2;
 427         }
 428
 429         val = io_ctl->cur;
 430         *val = cpu_to_le64(generation);
 431         io_ctl->cur += sizeof(u64);
 432 }
 433
 434 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 435 {
 436         __le64 *gen;
 437
 438         /*
 439          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 440          * chunk at the front of the first page.
 441          */
 442         if (io_ctl->check_crcs) {
 443                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 444                 io_ctl->size -= sizeof(u64) +
 445                         (sizeof(u32) * io_ctl->num_pages);
 446         } else {
 447                 io_ctl->cur += sizeof(u64);
 448                 io_ctl->size -= sizeof(u64) * 2;
 449         }
 450
 451         gen = io_ctl->cur;
 452         if (le64_to_cpu(*gen) != generation) {
 453                 btrfs_err_rl(io_ctl->root->fs_info,
 454                         "space cache generation (%llu) does not match inode (%llu)",
 455                                 *gen, generation);
 456                 io_ctl_unmap_page(io_ctl);
 457                 return -EIO;
 458         }
 459         io_ctl->cur += sizeof(u64);
 460         return 0;
 461 }
 462
 463 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 464 {
 465         u32 *tmp;
 466         u32 crc = ~(u32)0;
 467         unsigned offset = 0;
 468
 469         if (!io_ctl->check_crcs) {
 470                 io_ctl_unmap_page(io_ctl);
 471                 return;
 472         }
 473
 474         if (index == 0)
 475                 offset = sizeof(u32) * io_ctl->num_pages;
 476
 477         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 478                               PAGE_SIZE - offset);
 479         btrfs_csum_final(crc, (char *)&crc);
 480         io_ctl_unmap_page(io_ctl);
 481         tmp = page_address(io_ctl->pages[0]);
 482         tmp += index;
 483         *tmp = crc;
 484 }
 485
 486 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 487 {
 488         u32 *tmp, val;
 489         u32 crc = ~(u32)0;
 490         unsigned offset = 0;
 491
 492         if (!io_ctl->check_crcs) {
 493                 io_ctl_map_page(io_ctl, 0);
 494                 return 0;
 495         }
 496
 497         if (index == 0)
 498                 offset = sizeof(u32) * io_ctl->num_pages;
 499
 500         tmp = page_address(io_ctl->pages[0]);
 501         tmp += index;
 502         val = *tmp;
 503
 504         io_ctl_map_page(io_ctl, 0);
 505         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 506                               PAGE_SIZE - offset);
 507         btrfs_csum_final(crc, (char *)&crc);
 508         if (val != crc) {
 509                 btrfs_err_rl(io_ctl->root->fs_info,
 510                         "csum mismatch on free space cache");
 511                 io_ctl_unmap_page(io_ctl);
 512                 return -EIO;
 513         }
 514
 515         return 0;
 516 }
 517
 518 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 519                             void *bitmap)
 520 {
 521         struct btrfs_free_space_entry *entry;
 522
 523         if (!io_ctl->cur)
 524                 return -ENOSPC;
 525
 526         entry = io_ctl->cur;
 527         entry->offset = cpu_to_le64(offset);
 528         entry->bytes = cpu_to_le64(bytes);
 529         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 530                 BTRFS_FREE_SPACE_EXTENT;
 531         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 532         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 533
 534         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 535                 return 0;
 536
 537         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 538
 539         /* No more pages to map */
 540         if (io_ctl->index >= io_ctl->num_pages)
 541                 return 0;
 542
 543         /* map the next page */
 544         io_ctl_map_page(io_ctl, 1);
 545         return 0;
 546 }
 547
 548 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 549 {
 550         if (!io_ctl->cur)
 551                 return -ENOSPC;
 552
 553         /*
 554          * If we aren't at the start of the current page, unmap this one and
 555          * map the next one if there is any left.
 556          */
 557         if (io_ctl->cur != io_ctl->orig) {
 558                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 559                 if (io_ctl->index >= io_ctl->num_pages)
 560                         return -ENOSPC;
 561                 io_ctl_map_page(io_ctl, 0);
 562         }
 563
 564         memcpy(io_ctl->cur, bitmap, PAGE_SIZE);
 565         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 566         if (io_ctl->index < io_ctl->num_pages)
 567                 io_ctl_map_page(io_ctl, 0);
 568         return 0;
 569 }
 570
 571 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 572 {
 573         /*
 574          * If we're not on the boundary we know we've modified the page and we
 575          * need to crc the page.
 576          */
 577         if (io_ctl->cur != io_ctl->orig)
 578                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 579         else
 580                 io_ctl_unmap_page(io_ctl);
 581
 582         while (io_ctl->index < io_ctl->num_pages) {
 583                 io_ctl_map_page(io_ctl, 1);
 584                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 585         }
 586 }
 587
 588 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 589                             struct btrfs_free_space *entry, u8 *type)
 590 {
 591         struct btrfs_free_space_entry *e;
 592         int ret;
 593
 594         if (!io_ctl->cur) {
 595                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 596                 if (ret)
 597                         return ret;
 598         }
 599
 600         e = io_ctl->cur;
 601         entry->offset = le64_to_cpu(e->offset);
 602         entry->bytes = le64_to_cpu(e->bytes);
 603         *type = e->type;
 604         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 605         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 606
 607         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 608                 return 0;
 609
 610         io_ctl_unmap_page(io_ctl);
 611
 612         return 0;
 613 }
 614
 615 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 616                               struct btrfs_free_space *entry)
 617 {
 618         int ret;
 619
 620         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 621         if (ret)
 622                 return ret;
 623
 624         memcpy(entry->bitmap, io_ctl->cur, PAGE_SIZE);
 625         io_ctl_unmap_page(io_ctl);
 626
 627         return 0;
 628 }
 629
 630 /*
 631  * Since we attach pinned extents after the fact we can have contiguous sections
 632  * of free space that are split up in entries.  This poses a problem with the
 633  * tree logging stuff since it could have allocated across what appears to be 2
 634  * entries since we would have merged the entries when adding the pinned extents
 635  * back to the free space cache.  So run through the space cache that we just
 636  * loaded and merge contiguous entries.  This will make the log replay stuff not
 637  * blow up and it will make for nicer allocator behavior.
 638  */
 639 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 640 {
 641         struct btrfs_free_space *e, *prev = NULL;
 642         struct rb_node *n;
 643
 644 again:
 645         spin_lock(&ctl->tree_lock);
 646         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 647                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 648                 if (!prev)
 649                         goto next;
 650                 if (e->bitmap || prev->bitmap)
 651                         goto next;
 652                 if (prev->offset + prev->bytes == e->offset) {
 653                         unlink_free_space(ctl, prev);
 654                         unlink_free_space(ctl, e);
 655                         prev->bytes += e->bytes;
 656                         kmem_cache_free(btrfs_free_space_cachep, e);
 657                         link_free_space(ctl, prev);
 658                         prev = NULL;
 659                         spin_unlock(&ctl->tree_lock);
 660                         goto again;
 661                 }
 662 next:
 663                 prev = e;
 664         }
 665         spin_unlock(&ctl->tree_lock);
 666 }
 667
 668 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 669                                    struct btrfs_free_space_ctl *ctl,
 670                                    struct btrfs_path *path, u64 offset)
 671 {
 672         struct btrfs_free_space_header *header;
 673         struct extent_buffer *leaf;
 674         struct btrfs_io_ctl io_ctl;
 675         struct btrfs_key key;
 676         struct btrfs_free_space *e, *n;
 677         LIST_HEAD(bitmaps);
 678         u64 num_entries;
 679         u64 num_bitmaps;
 680         u64 generation;
 681         u8 type;
 682         int ret = 0;
 683
 684         /* Nothing in the space cache, goodbye */
 685         if (!i_size_read(inode))
 686                 return 0;
 687
 688         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 689         key.offset = offset;
 690         key.type = 0;
 691
 692         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 693         if (ret < 0)
 694                 return 0;
 695         else if (ret > 0) {
 696                 btrfs_release_path(path);
 697                 return 0;
 698         }
 699
 700         ret = -1;
 701
 702         leaf = path->nodes[0];
 703         header = btrfs_item_ptr(leaf, path->slots[0],
 704                                 struct btrfs_free_space_header);
 705         num_entries = btrfs_free_space_entries(leaf, header);
 706         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 707         generation = btrfs_free_space_generation(leaf, header);
 708         btrfs_release_path(path);
 709
 710         if (!BTRFS_I(inode)->generation) {
 711                 btrfs_info(root->fs_info,
 712                            "The free space cache file (%llu) is invalid. skip it\n",
 713                            offset);
 714                 return 0;
 715         }
 716
 717         if (BTRFS_I(inode)->generation != generation) {
 718                 btrfs_err(root->fs_info,
 719                         "free space inode generation (%llu) "
 720                         "did not match free space cache generation (%llu)",
 721                         BTRFS_I(inode)->generation, generation);
 722                 return 0;
 723         }
 724
 725         if (!num_entries)
 726                 return 0;
 727
 728         ret = io_ctl_init(&io_ctl, inode, root, 0);
 729         if (ret)
 730                 return ret;
 731
 732         ret = readahead_cache(inode);
 733         if (ret)
 734                 goto out;
 735
 736         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 737         if (ret)
 738                 goto out;
 739
 740         ret = io_ctl_check_crc(&io_ctl, 0);
 741         if (ret)
 742                 goto free_cache;
 743
 744         ret = io_ctl_check_generation(&io_ctl, generation);
 745         if (ret)
 746                 goto free_cache;
 747
 748         while (num_entries) {
 749                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 750                                       GFP_NOFS);
 751                 if (!e)
 752                         goto free_cache;
 753
 754                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 755                 if (ret) {
 756                         kmem_cache_free(btrfs_free_space_cachep, e);
 757                         goto free_cache;
 758                 }
 759
 760                 if (!e->bytes) {
 761                         kmem_cache_free(btrfs_free_space_cachep, e);
 762                         goto free_cache;
 763                 }
 764
 765                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 766                         spin_lock(&ctl->tree_lock);
 767                         ret = link_free_space(ctl, e);
 768                         spin_unlock(&ctl->tree_lock);
 769                         if (ret) {
 770                                 btrfs_err(root->fs_info,
 771                                         "Duplicate entries in free space cache, dumping");
 772                                 kmem_cache_free(btrfs_free_space_cachep, e);
 773                                 goto free_cache;
 774                         }
 775                 } else {
 776                         ASSERT(num_bitmaps);
 777                         num_bitmaps--;
 778                         e->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
 779                         if (!e->bitmap) {
 780                                 kmem_cache_free(
 781                                         btrfs_free_space_cachep, e);
 782                                 goto free_cache;
 783                         }
 784                         spin_lock(&ctl->tree_lock);
 785                         ret = link_free_space(ctl, e);
 786                         ctl->total_bitmaps++;
 787                         ctl->op->recalc_thresholds(ctl);
 788                         spin_unlock(&ctl->tree_lock);
 789                         if (ret) {
 790                                 btrfs_err(root->fs_info,
 791                                         "Duplicate entries in free space cache, dumping");
 792                                 kmem_cache_free(btrfs_free_space_cachep, e);
 793                                 goto free_cache;
 794                         }
 795                         list_add_tail(&e->list, &bitmaps);
 796                 }
 797
 798                 num_entries--;
 799         }
 800
 801         io_ctl_unmap_page(&io_ctl);
 802
 803         /*
 804          * We add the bitmaps at the end of the entries in order that
 805          * the bitmap entries are added to the cache.
 806          */
 807         list_for_each_entry_safe(e, n, &bitmaps, list) {
 808                 list_del_init(&e->list);
 809                 ret = io_ctl_read_bitmap(&io_ctl, e);
 810                 if (ret)
 811                         goto free_cache;
 812         }
 813
 814         io_ctl_drop_pages(&io_ctl);
 815         merge_space_tree(ctl);
 816         ret = 1;
 817 out:
 818         io_ctl_free(&io_ctl);
 819         return ret;
 820 free_cache:
 821         io_ctl_drop_pages(&io_ctl);
 822         __btrfs_remove_free_space_cache(ctl);
 823         goto out;
 824 }
 825
 826 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 827                           struct btrfs_block_group_cache *block_group)
 828 {
 829         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 830         struct btrfs_root *root = fs_info->tree_root;
 831         struct inode *inode;
 832         struct btrfs_path *path;
 833         int ret = 0;
 834         bool matched;
 835         u64 used = btrfs_block_group_used(&block_group->item);
 836
 837         /*
 838          * If this block group has been marked to be cleared for one reason or
 839          * another then we can't trust the on disk cache, so just return.
 840          */
 841         spin_lock(&block_group->lock);
 842         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 843                 spin_unlock(&block_group->lock);
 844                 return 0;
 845         }
 846         spin_unlock(&block_group->lock);
 847
 848         path = btrfs_alloc_path();
 849         if (!path)
 850                 return 0;
 851         path->search_commit_root = 1;
 852         path->skip_locking = 1;
 853
 854         inode = lookup_free_space_inode(root, block_group, path);
 855         if (IS_ERR(inode)) {
 856                 btrfs_free_path(path);
 857                 return 0;
 858         }
 859
 860         /* We may have converted the inode and made the cache invalid. */
 861         spin_lock(&block_group->lock);
 862         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 863                 spin_unlock(&block_group->lock);
 864                 btrfs_free_path(path);
 865                 goto out;
 866         }
 867         spin_unlock(&block_group->lock);
 868
 869         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 870                                       path, block_group->key.objectid);
 871         btrfs_free_path(path);
 872         if (ret <= 0)
 873                 goto out;
 874
 875         spin_lock(&ctl->tree_lock);
 876         matched = (ctl->free_space == (block_group->key.offset - used -
 877                                        block_group->bytes_super));
 878         spin_unlock(&ctl->tree_lock);
 879
 880         if (!matched) {
 881                 __btrfs_remove_free_space_cache(ctl);
 882                 btrfs_warn(fs_info, "block group %llu has wrong amount of free space",
 883                         block_group->key.objectid);
 884                 ret = -1;
 885         }
 886 out:
 887         if (ret < 0) {
 888                 /* This cache is bogus, make sure it gets cleared */
 889                 spin_lock(&block_group->lock);
 890                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 891                 spin_unlock(&block_group->lock);
 892                 ret = 0;
 893
 894                 btrfs_warn(fs_info, "failed to load free space cache for block group %llu, rebuilding it now",
 895                         block_group->key.objectid);
 896         }
 897
 898         iput(inode);
 899         return ret;
 900 }
 901
 902 static noinline_for_stack
 903 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
 904                               struct btrfs_free_space_ctl *ctl,
 905                               struct btrfs_block_group_cache *block_group,
 906                               int *entries, int *bitmaps,
 907                               struct list_head *bitmap_list)
 908 {
 909         int ret;
 910         struct btrfs_free_cluster *cluster = NULL;
 911         struct btrfs_free_cluster *cluster_locked = NULL;
 912         struct rb_node *node = rb_first(&ctl->free_space_offset);
 913         struct btrfs_trim_range *trim_entry;
 914
 915         /* Get the cluster for this block_group if it exists */
 916         if (block_group && !list_empty(&block_group->cluster_list)) {
 917                 cluster = list_entry(block_group->cluster_list.next,
 918                                      struct btrfs_free_cluster,
 919                                      block_group_list);
 920         }
 921
 922         if (!node && cluster) {
 923                 cluster_locked = cluster;
 924                 spin_lock(&cluster_locked->lock);
 925                 node = rb_first(&cluster->root);
 926                 cluster = NULL;
 927         }
 928
 929         /* Write out the extent entries */
 930         while (node) {
 931                 struct btrfs_free_space *e;
 932
 933                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 934                 *entries += 1;
 935
 936                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
 937                                        e->bitmap);
 938                 if (ret)
 939                         goto fail;
 940
 941                 if (e->bitmap) {
 942                         list_add_tail(&e->list, bitmap_list);
 943                         *bitmaps += 1;
 944                 }
 945                 node = rb_next(node);
 946                 if (!node && cluster) {
 947                         node = rb_first(&cluster->root);
 948                         cluster_locked = cluster;
 949                         spin_lock(&cluster_locked->lock);
 950                         cluster = NULL;
 951                 }
 952         }
 953         if (cluster_locked) {
 954                 spin_unlock(&cluster_locked->lock);
 955                 cluster_locked = NULL;
 956         }
 957
 958         /*
 959          * Make sure we don't miss any range that was removed from our rbtree
 960          * because trimming is running. Otherwise after a umount+mount (or crash
 961          * after committing the transaction) we would leak free space and get
 962          * an inconsistent free space cache report from fsck.
 963          */
 964         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
 965                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
 966                                        trim_entry->bytes, NULL);
 967                 if (ret)
 968                         goto fail;
 969                 *entries += 1;
 970         }
 971
 972         return 0;
 973 fail:
 974         if (cluster_locked)
 975                 spin_unlock(&cluster_locked->lock);
 976         return -ENOSPC;
 977 }
 978
 979 static noinline_for_stack int
 980 update_cache_item(struct btrfs_trans_handle *trans,
 981                   struct btrfs_root *root,
 982                   struct inode *inode,
 983                   struct btrfs_path *path, u64 offset,
 984                   int entries, int bitmaps)
 985 {
 986         struct btrfs_key key;
 987         struct btrfs_free_space_header *header;
 988         struct extent_buffer *leaf;
 989         int ret;
 990
 991         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 992         key.offset = offset;
 993         key.type = 0;
 994
 995         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 996         if (ret < 0) {
 997                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
 998                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
 999                                  GFP_NOFS);
1000                 goto fail;
1001         }
1002         leaf = path->nodes[0];
1003         if (ret > 0) {
1004                 struct btrfs_key found_key;
1005                 ASSERT(path->slots[0]);
1006                 path->slots[0]--;
1007                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1008                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1009                     found_key.offset != offset) {
1010                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1011                                          inode->i_size - 1,
1012                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1013                                          NULL, GFP_NOFS);
1014                         btrfs_release_path(path);
1015                         goto fail;
1016                 }
1017         }
1018
1019         BTRFS_I(inode)->generation = trans->transid;
1020         header = btrfs_item_ptr(leaf, path->slots[0],
1021                                 struct btrfs_free_space_header);
1022         btrfs_set_free_space_entries(leaf, header, entries);
1023         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1024         btrfs_set_free_space_generation(leaf, header, trans->transid);
1025         btrfs_mark_buffer_dirty(leaf);
1026         btrfs_release_path(path);
1027
1028         return 0;
1029
1030 fail:
1031         return -1;
1032 }
1033
1034 static noinline_for_stack int
1035 write_pinned_extent_entries(struct btrfs_root *root,
1036                             struct btrfs_block_group_cache *block_group,
1037                             struct btrfs_io_ctl *io_ctl,
1038                             int *entries)
1039 {
1040         u64 start, extent_start, extent_end, len;
1041         struct extent_io_tree *unpin = NULL;
1042         int ret;
1043
1044         if (!block_group)
1045                 return 0;
1046
1047         /*
1048          * We want to add any pinned extents to our free space cache
1049          * so we don't leak the space
1050          *
1051          * We shouldn't have switched the pinned extents yet so this is the
1052          * right one
1053          */
1054         unpin = root->fs_info->pinned_extents;
1055
1056         start = block_group->key.objectid;
1057
1058         while (start < block_group->key.objectid + block_group->key.offset) {
1059                 ret = find_first_extent_bit(unpin, start,
1060                                             &extent_start, &extent_end,
1061                                             EXTENT_DIRTY, NULL);
1062                 if (ret)
1063                         return 0;
1064
1065                 /* This pinned extent is out of our range */
1066                 if (extent_start >= block_group->key.objectid +
1067                     block_group->key.offset)
1068                         return 0;
1069
1070                 extent_start = max(extent_start, start);
1071                 extent_end = min(block_group->key.objectid +
1072                                  block_group->key.offset, extent_end + 1);
1073                 len = extent_end - extent_start;
1074
1075                 *entries += 1;
1076                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1077                 if (ret)
1078                         return -ENOSPC;
1079
1080                 start = extent_end;
1081         }
1082
1083         return 0;
1084 }
1085
1086 static noinline_for_stack int
1087 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1088 {
1089         struct btrfs_free_space *entry, *next;
1090         int ret;
1091
1092         /* Write out the bitmaps */
1093         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1094                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1095                 if (ret)
1096                         return -ENOSPC;
1097                 list_del_init(&entry->list);
1098         }
1099
1100         return 0;
1101 }
1102
1103 static int flush_dirty_cache(struct inode *inode)
1104 {
1105         int ret;
1106
1107         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1108         if (ret)
1109                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1110                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1111                                  GFP_NOFS);
1112
1113         return ret;
1114 }
1115
1116 static void noinline_for_stack
1117 cleanup_bitmap_list(struct list_head *bitmap_list)
1118 {
1119         struct btrfs_free_space *entry, *next;
1120
1121         list_for_each_entry_safe(entry, next, bitmap_list, list)
1122                 list_del_init(&entry->list);
1123 }
1124
1125 static void noinline_for_stack
1126 cleanup_write_cache_enospc(struct inode *inode,
1127                            struct btrfs_io_ctl *io_ctl,
1128                            struct extent_state **cached_state,
1129                            struct list_head *bitmap_list)
1130 {
1131         io_ctl_drop_pages(io_ctl);
1132         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1133                              i_size_read(inode) - 1, cached_state,
1134                              GFP_NOFS);
1135 }
1136
1137 int btrfs_wait_cache_io(struct btrfs_root *root,
1138                         struct btrfs_trans_handle *trans,
1139                         struct btrfs_block_group_cache *block_group,
1140                         struct btrfs_io_ctl *io_ctl,
1141                         struct btrfs_path *path, u64 offset)
1142 {
1143         int ret;
1144         struct inode *inode = io_ctl->inode;
1145
1146         if (!inode)
1147                 return 0;
1148
1149         if (block_group)
1150                 root = root->fs_info->tree_root;
1151
1152         /* Flush the dirty pages in the cache file. */
1153         ret = flush_dirty_cache(inode);
1154         if (ret)
1155                 goto out;
1156
1157         /* Update the cache item to tell everyone this cache file is valid. */
1158         ret = update_cache_item(trans, root, inode, path, offset,
1159                                 io_ctl->entries, io_ctl->bitmaps);
1160 out:
1161         io_ctl_free(io_ctl);
1162         if (ret) {
1163                 invalidate_inode_pages2(inode->i_mapping);
1164                 BTRFS_I(inode)->generation = 0;
1165                 if (block_group) {
1166 #ifdef DEBUG
1167                         btrfs_err(root->fs_info,
1168                                 "failed to write free space cache for block group %llu",
1169                                 block_group->key.objectid);
1170 #endif
1171                 }
1172         }
1173         btrfs_update_inode(trans, root, inode);
1174
1175         if (block_group) {
1176                 /* the dirty list is protected by the dirty_bgs_lock */
1177                 spin_lock(&trans->transaction->dirty_bgs_lock);
1178
1179                 /* the disk_cache_state is protected by the block group lock */
1180                 spin_lock(&block_group->lock);
1181
1182                 /*
1183                  * only mark this as written if we didn't get put back on
1184                  * the dirty list while waiting for IO.   Otherwise our
1185                  * cache state won't be right, and we won't get written again
1186                  */
1187                 if (!ret && list_empty(&block_group->dirty_list))
1188                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1189                 else if (ret)
1190                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1191
1192                 spin_unlock(&block_group->lock);
1193                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1194                 io_ctl->inode = NULL;
1195                 iput(inode);
1196         }
1197
1198         return ret;
1199
1200 }
1201
1202 /**
1203  * __btrfs_write_out_cache - write out cached info to an inode
1204  * @root - the root the inode belongs to
1205  * @ctl - the free space cache we are going to write out
1206  * @block_group - the block_group for this cache if it belongs to a block_group
1207  * @trans - the trans handle
1208  * @path - the path to use
1209  * @offset - the offset for the key we'll insert
1210  *
1211  * This function writes out a free space cache struct to disk for quick recovery
1212  * on mount.  This will return 0 if it was successful in writing the cache out,
1213  * or an errno if it was not.
1214  */
1215 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1216                                    struct btrfs_free_space_ctl *ctl,
1217                                    struct btrfs_block_group_cache *block_group,
1218                                    struct btrfs_io_ctl *io_ctl,
1219                                    struct btrfs_trans_handle *trans,
1220                                    struct btrfs_path *path, u64 offset)
1221 {
1222         struct extent_state *cached_state = NULL;
1223         LIST_HEAD(bitmap_list);
1224         int entries = 0;
1225         int bitmaps = 0;
1226         int ret;
1227         int must_iput = 0;
1228
1229         if (!i_size_read(inode))
1230                 return -EIO;
1231
1232         WARN_ON(io_ctl->pages);
1233         ret = io_ctl_init(io_ctl, inode, root, 1);
1234         if (ret)
1235                 return ret;
1236
1237         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1238                 down_write(&block_group->data_rwsem);
1239                 spin_lock(&block_group->lock);
1240                 if (block_group->delalloc_bytes) {
1241                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1242                         spin_unlock(&block_group->lock);
1243                         up_write(&block_group->data_rwsem);
1244                         BTRFS_I(inode)->generation = 0;
1245                         ret = 0;
1246                         must_iput = 1;
1247                         goto out;
1248                 }
1249                 spin_unlock(&block_group->lock);
1250         }
1251
1252         /* Lock all pages first so we can lock the extent safely. */
1253         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1254         if (ret)
1255                 goto out;
1256
1257         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1258                          &cached_state);
1259
1260         io_ctl_set_generation(io_ctl, trans->transid);
1261
1262         mutex_lock(&ctl->cache_writeout_mutex);
1263         /* Write out the extent entries in the free space cache */
1264         spin_lock(&ctl->tree_lock);
1265         ret = write_cache_extent_entries(io_ctl, ctl,
1266                                          block_group, &entries, &bitmaps,
1267                                          &bitmap_list);
1268         if (ret)
1269                 goto out_nospc_locked;
1270
1271         /*
1272          * Some spaces that are freed in the current transaction are pinned,
1273          * they will be added into free space cache after the transaction is
1274          * committed, we shouldn't lose them.
1275          *
1276          * If this changes while we are working we'll get added back to
1277          * the dirty list and redo it.  No locking needed
1278          */
1279         ret = write_pinned_extent_entries(root, block_group, io_ctl, &entries);
1280         if (ret)
1281                 goto out_nospc_locked;
1282
1283         /*
1284          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1285          * locked while doing it because a concurrent trim can be manipulating
1286          * or freeing the bitmap.
1287          */
1288         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1289         spin_unlock(&ctl->tree_lock);
1290         mutex_unlock(&ctl->cache_writeout_mutex);
1291         if (ret)
1292                 goto out_nospc;
1293
1294         /* Zero out the rest of the pages just to make sure */
1295         io_ctl_zero_remaining_pages(io_ctl);
1296
1297         /* Everything is written out, now we dirty the pages in the file. */
1298         ret = btrfs_dirty_pages(root, inode, io_ctl->pages, io_ctl->num_pages,
1299                                 0, i_size_read(inode), &cached_state);
1300         if (ret)
1301                 goto out_nospc;
1302
1303         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1304                 up_write(&block_group->data_rwsem);
1305         /*
1306          * Release the pages and unlock the extent, we will flush
1307          * them out later
1308          */
1309         io_ctl_drop_pages(io_ctl);
1310
1311         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1312                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1313
1314         /*
1315          * at this point the pages are under IO and we're happy,
1316          * The caller is responsible for waiting on them and updating the
1317          * the cache and the inode
1318          */
1319         io_ctl->entries = entries;
1320         io_ctl->bitmaps = bitmaps;
1321
1322         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1323         if (ret)
1324                 goto out;
1325
1326         return 0;
1327
1328 out:
1329         io_ctl->inode = NULL;
1330         io_ctl_free(io_ctl);
1331         if (ret) {
1332                 invalidate_inode_pages2(inode->i_mapping);
1333                 BTRFS_I(inode)->generation = 0;
1334         }
1335         btrfs_update_inode(trans, root, inode);
1336         if (must_iput)
1337                 iput(inode);
1338         return ret;
1339
1340 out_nospc_locked:
1341         cleanup_bitmap_list(&bitmap_list);
1342         spin_unlock(&ctl->tree_lock);
1343         mutex_unlock(&ctl->cache_writeout_mutex);
1344
1345 out_nospc:
1346         cleanup_write_cache_enospc(inode, io_ctl, &cached_state, &bitmap_list);
1347
1348         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1349                 up_write(&block_group->data_rwsem);
1350
1351         goto out;
1352 }
1353
1354 int btrfs_write_out_cache(struct btrfs_root *root,
1355                           struct btrfs_trans_handle *trans,
1356                           struct btrfs_block_group_cache *block_group,
1357                           struct btrfs_path *path)
1358 {
1359         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1360         struct inode *inode;
1361         int ret = 0;
1362
1363         root = root->fs_info->tree_root;
1364
1365         spin_lock(&block_group->lock);
1366         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1367                 spin_unlock(&block_group->lock);
1368                 return 0;
1369         }
1370         spin_unlock(&block_group->lock);
1371
1372         inode = lookup_free_space_inode(root, block_group, path);
1373         if (IS_ERR(inode))
1374                 return 0;
1375
1376         ret = __btrfs_write_out_cache(root, inode, ctl, block_group,
1377                                       &block_group->io_ctl, trans,
1378                                       path, block_group->key.objectid);
1379         if (ret) {
1380 #ifdef DEBUG
1381                 btrfs_err(root->fs_info,
1382                         "failed to write free space cache for block group %llu",
1383                         block_group->key.objectid);
1384 #endif
1385                 spin_lock(&block_group->lock);
1386                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1387                 spin_unlock(&block_group->lock);
1388
1389                 block_group->io_ctl.inode = NULL;
1390                 iput(inode);
1391         }
1392
1393         /*
1394          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1395          * to wait for IO and put the inode
1396          */
1397
1398         return ret;
1399 }
1400
1401 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1402                                           u64 offset)
1403 {
1404         ASSERT(offset >= bitmap_start);
1405         offset -= bitmap_start;
1406         return (unsigned long)(div_u64(offset, unit));
1407 }
1408
1409 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1410 {
1411         return (unsigned long)(div_u64(bytes, unit));
1412 }
1413
1414 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1415                                    u64 offset)
1416 {
1417         u64 bitmap_start;
1418         u64 bytes_per_bitmap;
1419
1420         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1421         bitmap_start = offset - ctl->start;
1422         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1423         bitmap_start *= bytes_per_bitmap;
1424         bitmap_start += ctl->start;
1425
1426         return bitmap_start;
1427 }
1428
1429 static int tree_insert_offset(struct rb_root *root, u64 offset,
1430                               struct rb_node *node, int bitmap)
1431 {
1432         struct rb_node **p = &root->rb_node;
1433         struct rb_node *parent = NULL;
1434         struct btrfs_free_space *info;
1435
1436         while (*p) {
1437                 parent = *p;
1438                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1439
1440                 if (offset < info->offset) {
1441                         p = &(*p)->rb_left;
1442                 } else if (offset > info->offset) {
1443                         p = &(*p)->rb_right;
1444                 } else {
1445                         /*
1446                          * we could have a bitmap entry and an extent entry
1447                          * share the same offset.  If this is the case, we want
1448                          * the extent entry to always be found first if we do a
1449                          * linear search through the tree, since we want to have
1450                          * the quickest allocation time, and allocating from an
1451                          * extent is faster than allocating from a bitmap.  So
1452                          * if we're inserting a bitmap and we find an entry at
1453                          * this offset, we want to go right, or after this entry
1454                          * logically.  If we are inserting an extent and we've
1455                          * found a bitmap, we want to go left, or before
1456                          * logically.
1457                          */
1458                         if (bitmap) {
1459                                 if (info->bitmap) {
1460                                         WARN_ON_ONCE(1);
1461                                         return -EEXIST;
1462                                 }
1463                                 p = &(*p)->rb_right;
1464                         } else {
1465                                 if (!info->bitmap) {
1466                                         WARN_ON_ONCE(1);
1467                                         return -EEXIST;
1468                                 }
1469                                 p = &(*p)->rb_left;
1470                         }
1471                 }
1472         }
1473
1474         rb_link_node(node, parent, p);
1475         rb_insert_color(node, root);
1476
1477         return 0;
1478 }
1479
1480 /*
1481  * searches the tree for the given offset.
1482  *
1483  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1484  * want a section that has at least bytes size and comes at or after the given
1485  * offset.
1486  */
1487 static struct btrfs_free_space *
1488 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1489                    u64 offset, int bitmap_only, int fuzzy)
1490 {
1491         struct rb_node *n = ctl->free_space_offset.rb_node;
1492         struct btrfs_free_space *entry, *prev = NULL;
1493
1494         /* find entry that is closest to the 'offset' */
1495         while (1) {
1496                 if (!n) {
1497                         entry = NULL;
1498                         break;
1499                 }
1500
1501                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1502                 prev = entry;
1503
1504                 if (offset < entry->offset)
1505                         n = n->rb_left;
1506                 else if (offset > entry->offset)
1507                         n = n->rb_right;
1508                 else
1509                         break;
1510         }
1511
1512         if (bitmap_only) {
1513                 if (!entry)
1514                         return NULL;
1515                 if (entry->bitmap)
1516                         return entry;
1517
1518                 /*
1519                  * bitmap entry and extent entry may share same offset,
1520                  * in that case, bitmap entry comes after extent entry.
1521                  */
1522                 n = rb_next(n);
1523                 if (!n)
1524                         return NULL;
1525                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1526                 if (entry->offset != offset)
1527                         return NULL;
1528
1529                 WARN_ON(!entry->bitmap);
1530                 return entry;
1531         } else if (entry) {
1532                 if (entry->bitmap) {
1533                         /*
1534                          * if previous extent entry covers the offset,
1535                          * we should return it instead of the bitmap entry
1536                          */
1537                         n = rb_prev(&entry->offset_index);
1538                         if (n) {
1539                                 prev = rb_entry(n, struct btrfs_free_space,
1540                                                 offset_index);
1541                                 if (!prev->bitmap &&
1542                                     prev->offset + prev->bytes > offset)
1543                                         entry = prev;
1544                         }
1545                 }
1546                 return entry;
1547         }
1548
1549         if (!prev)
1550                 return NULL;
1551
1552         /* find last entry before the 'offset' */
1553         entry = prev;
1554         if (entry->offset > offset) {
1555                 n = rb_prev(&entry->offset_index);
1556                 if (n) {
1557                         entry = rb_entry(n, struct btrfs_free_space,
1558                                         offset_index);
1559                         ASSERT(entry->offset <= offset);
1560                 } else {
1561                         if (fuzzy)
1562                                 return entry;
1563                         else
1564                                 return NULL;
1565                 }
1566         }
1567
1568         if (entry->bitmap) {
1569                 n = rb_prev(&entry->offset_index);
1570                 if (n) {
1571                         prev = rb_entry(n, struct btrfs_free_space,
1572                                         offset_index);
1573                         if (!prev->bitmap &&
1574                             prev->offset + prev->bytes > offset)
1575                                 return prev;
1576                 }
1577                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1578                         return entry;
1579         } else if (entry->offset + entry->bytes > offset)
1580                 return entry;
1581
1582         if (!fuzzy)
1583                 return NULL;
1584
1585         while (1) {
1586                 if (entry->bitmap) {
1587                         if (entry->offset + BITS_PER_BITMAP *
1588                             ctl->unit > offset)
1589                                 break;
1590                 } else {
1591                         if (entry->offset + entry->bytes > offset)
1592                                 break;
1593                 }
1594
1595                 n = rb_next(&entry->offset_index);
1596                 if (!n)
1597                         return NULL;
1598                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1599         }
1600         return entry;
1601 }
1602
1603 static inline void
1604 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1605                     struct btrfs_free_space *info)
1606 {
1607         rb_erase(&info->offset_index, &ctl->free_space_offset);
1608         ctl->free_extents--;
1609 }
1610
1611 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1612                               struct btrfs_free_space *info)
1613 {
1614         __unlink_free_space(ctl, info);
1615         ctl->free_space -= info->bytes;
1616 }
1617
1618 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1619                            struct btrfs_free_space *info)
1620 {
1621         int ret = 0;
1622
1623         ASSERT(info->bytes || info->bitmap);
1624         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1625                                  &info->offset_index, (info->bitmap != NULL));
1626         if (ret)
1627                 return ret;
1628
1629         ctl->free_space += info->bytes;
1630         ctl->free_extents++;
1631         return ret;
1632 }
1633
1634 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1635 {
1636         struct btrfs_block_group_cache *block_group = ctl->private;
1637         u64 max_bytes;
1638         u64 bitmap_bytes;
1639         u64 extent_bytes;
1640         u64 size = block_group->key.offset;
1641         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1642         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1643
1644         max_bitmaps = max_t(u64, max_bitmaps, 1);
1645
1646         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1647
1648         /*
1649          * The goal is to keep the total amount of memory used per 1gb of space
1650          * at or below 32k, so we need to adjust how much memory we allow to be
1651          * used by extent based free space tracking
1652          */
1653         if (size < SZ_1G)
1654                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1655         else
1656                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1657
1658         /*
1659          * we want to account for 1 more bitmap than what we have so we can make
1660          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1661          * we add more bitmaps.
1662          */
1663         bitmap_bytes = (ctl->total_bitmaps + 1) * ctl->unit;
1664
1665         if (bitmap_bytes >= max_bytes) {
1666                 ctl->extents_thresh = 0;
1667                 return;
1668         }
1669
1670         /*
1671          * we want the extent entry threshold to always be at most 1/2 the max
1672          * bytes we can have, or whatever is less than that.
1673          */
1674         extent_bytes = max_bytes - bitmap_bytes;
1675         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1676
1677         ctl->extents_thresh =
1678                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1679 }
1680
1681 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1682                                        struct btrfs_free_space *info,
1683                                        u64 offset, u64 bytes)
1684 {
1685         unsigned long start, count;
1686
1687         start = offset_to_bit(info->offset, ctl->unit, offset);
1688         count = bytes_to_bits(bytes, ctl->unit);
1689         ASSERT(start + count <= BITS_PER_BITMAP);
1690
1691         bitmap_clear(info->bitmap, start, count);
1692
1693         info->bytes -= bytes;
1694 }
1695
1696 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1697                               struct btrfs_free_space *info, u64 offset,
1698                               u64 bytes)
1699 {
1700         __bitmap_clear_bits(ctl, info, offset, bytes);
1701         ctl->free_space -= bytes;
1702 }
1703
1704 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1705                             struct btrfs_free_space *info, u64 offset,
1706                             u64 bytes)
1707 {
1708         unsigned long start, count;
1709
1710         start = offset_to_bit(info->offset, ctl->unit, offset);
1711         count = bytes_to_bits(bytes, ctl->unit);
1712         ASSERT(start + count <= BITS_PER_BITMAP);
1713
1714         bitmap_set(info->bitmap, start, count);
1715
1716         info->bytes += bytes;
1717         ctl->free_space += bytes;
1718 }
1719
1720 /*
1721  * If we can not find suitable extent, we will use bytes to record
1722  * the size of the max extent.
1723  */
1724 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1725                          struct btrfs_free_space *bitmap_info, u64 *offset,
1726                          u64 *bytes, bool for_alloc)
1727 {
1728         unsigned long found_bits = 0;
1729         unsigned long max_bits = 0;
1730         unsigned long bits, i;
1731         unsigned long next_zero;
1732         unsigned long extent_bits;
1733
1734         /*
1735          * Skip searching the bitmap if we don't have a contiguous section that
1736          * is large enough for this allocation.
1737          */
1738         if (for_alloc &&
1739             bitmap_info->max_extent_size &&
1740             bitmap_info->max_extent_size < *bytes) {
1741                 *bytes = bitmap_info->max_extent_size;
1742                 return -1;
1743         }
1744
1745         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1746                           max_t(u64, *offset, bitmap_info->offset));
1747         bits = bytes_to_bits(*bytes, ctl->unit);
1748
1749         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1750                 if (for_alloc && bits == 1) {
1751                         found_bits = 1;
1752                         break;
1753                 }
1754                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1755                                                BITS_PER_BITMAP, i);
1756                 extent_bits = next_zero - i;
1757                 if (extent_bits >= bits) {
1758                         found_bits = extent_bits;
1759                         break;
1760                 } else if (extent_bits > max_bits) {
1761                         max_bits = extent_bits;
1762                 }
1763                 i = next_zero;
1764         }
1765
1766         if (found_bits) {
1767                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1768                 *bytes = (u64)(found_bits) * ctl->unit;
1769                 return 0;
1770         }
1771
1772         *bytes = (u64)(max_bits) * ctl->unit;
1773         bitmap_info->max_extent_size = *bytes;
1774         return -1;
1775 }
1776
1777 /* Cache the size of the max extent in bytes */
1778 static struct btrfs_free_space *
1779 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1780                 unsigned long align, u64 *max_extent_size)
1781 {
1782         struct btrfs_free_space *entry;
1783         struct rb_node *node;
1784         u64 tmp;
1785         u64 align_off;
1786         int ret;
1787
1788         if (!ctl->free_space_offset.rb_node)
1789                 goto out;
1790
1791         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1792         if (!entry)
1793                 goto out;
1794
1795         for (node = &entry->offset_index; node; node = rb_next(node)) {
1796                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1797                 if (entry->bytes < *bytes) {
1798                         if (entry->bytes > *max_extent_size)
1799                                 *max_extent_size = entry->bytes;
1800                         continue;
1801                 }
1802
1803                 /* make sure the space returned is big enough
1804                  * to match our requested alignment
1805                  */
1806                 if (*bytes >= align) {
1807                         tmp = entry->offset - ctl->start + align - 1;
1808                         tmp = div64_u64(tmp, align);
1809                         tmp = tmp * align + ctl->start;
1810                         align_off = tmp - entry->offset;
1811                 } else {
1812                         align_off = 0;
1813                         tmp = entry->offset;
1814                 }
1815
1816                 if (entry->bytes < *bytes + align_off) {
1817                         if (entry->bytes > *max_extent_size)
1818                                 *max_extent_size = entry->bytes;
1819                         continue;
1820                 }
1821
1822                 if (entry->bitmap) {
1823                         u64 size = *bytes;
1824
1825                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1826                         if (!ret) {
1827                                 *offset = tmp;
1828                                 *bytes = size;
1829                                 return entry;
1830                         } else if (size > *max_extent_size) {
1831                                 *max_extent_size = size;
1832                         }
1833                         continue;
1834                 }
1835
1836                 *offset = tmp;
1837                 *bytes = entry->bytes - align_off;
1838                 return entry;
1839         }
1840 out:
1841         return NULL;
1842 }
1843
1844 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1845                            struct btrfs_free_space *info, u64 offset)
1846 {
1847         info->offset = offset_to_bitmap(ctl, offset);
1848         info->bytes = 0;
1849         INIT_LIST_HEAD(&info->list);
1850         link_free_space(ctl, info);
1851         ctl->total_bitmaps++;
1852
1853         ctl->op->recalc_thresholds(ctl);
1854 }
1855
1856 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1857                         struct btrfs_free_space *bitmap_info)
1858 {
1859         unlink_free_space(ctl, bitmap_info);
1860         kfree(bitmap_info->bitmap);
1861         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1862         ctl->total_bitmaps--;
1863         ctl->op->recalc_thresholds(ctl);
1864 }
1865
1866 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1867                               struct btrfs_free_space *bitmap_info,
1868                               u64 *offset, u64 *bytes)
1869 {
1870         u64 end;
1871         u64 search_start, search_bytes;
1872         int ret;
1873
1874 again:
1875         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1876
1877         /*
1878          * We need to search for bits in this bitmap.  We could only cover some
1879          * of the extent in this bitmap thanks to how we add space, so we need
1880          * to search for as much as it as we can and clear that amount, and then
1881          * go searching for the next bit.
1882          */
1883         search_start = *offset;
1884         search_bytes = ctl->unit;
1885         search_bytes = min(search_bytes, end - search_start + 1);
1886         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1887                             false);
1888         if (ret < 0 || search_start != *offset)
1889                 return -EINVAL;
1890
1891         /* We may have found more bits than what we need */
1892         search_bytes = min(search_bytes, *bytes);
1893
1894         /* Cannot clear past the end of the bitmap */
1895         search_bytes = min(search_bytes, end - search_start + 1);
1896
1897         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1898         *offset += search_bytes;
1899         *bytes -= search_bytes;
1900
1901         if (*bytes) {
1902                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1903                 if (!bitmap_info->bytes)
1904                         free_bitmap(ctl, bitmap_info);
1905
1906                 /*
1907                  * no entry after this bitmap, but we still have bytes to
1908                  * remove, so something has gone wrong.
1909                  */
1910                 if (!next)
1911                         return -EINVAL;
1912
1913                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1914                                        offset_index);
1915
1916                 /*
1917                  * if the next entry isn't a bitmap we need to return to let the
1918                  * extent stuff do its work.
1919                  */
1920                 if (!bitmap_info->bitmap)
1921                         return -EAGAIN;
1922
1923                 /*
1924                  * Ok the next item is a bitmap, but it may not actually hold
1925                  * the information for the rest of this free space stuff, so
1926                  * look for it, and if we don't find it return so we can try
1927                  * everything over again.
1928                  */
1929                 search_start = *offset;
1930                 search_bytes = ctl->unit;
1931                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1932                                     &search_bytes, false);
1933                 if (ret < 0 || search_start != *offset)
1934                         return -EAGAIN;
1935
1936                 goto again;
1937         } else if (!bitmap_info->bytes)
1938                 free_bitmap(ctl, bitmap_info);
1939
1940         return 0;
1941 }
1942
1943 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1944                                struct btrfs_free_space *info, u64 offset,
1945                                u64 bytes)
1946 {
1947         u64 bytes_to_set = 0;
1948         u64 end;
1949
1950         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1951
1952         bytes_to_set = min(end - offset, bytes);
1953
1954         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1955
1956         /*
1957          * We set some bytes, we have no idea what the max extent size is
1958          * anymore.
1959          */
1960         info->max_extent_size = 0;
1961
1962         return bytes_to_set;
1963
1964 }
1965
1966 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1967                       struct btrfs_free_space *info)
1968 {
1969         struct btrfs_block_group_cache *block_group = ctl->private;
1970         bool forced = false;
1971
1972 #ifdef CONFIG_BTRFS_DEBUG
1973         if (btrfs_should_fragment_free_space(block_group->fs_info->extent_root,
1974                                              block_group))
1975                 forced = true;
1976 #endif
1977
1978         /*
1979          * If we are below the extents threshold then we can add this as an
1980          * extent, and don't have to deal with the bitmap
1981          */
1982         if (!forced && ctl->free_extents < ctl->extents_thresh) {
1983                 /*
1984                  * If this block group has some small extents we don't want to
1985                  * use up all of our free slots in the cache with them, we want
1986                  * to reserve them to larger extents, however if we have plenty
1987                  * of cache left then go ahead an dadd them, no sense in adding
1988                  * the overhead of a bitmap if we don't have to.
1989                  */
1990                 if (info->bytes <= block_group->sectorsize * 4) {
1991                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
1992                                 return false;
1993                 } else {
1994                         return false;
1995                 }
1996         }
1997
1998         /*
1999          * The original block groups from mkfs can be really small, like 8
2000          * megabytes, so don't bother with a bitmap for those entries.  However
2001          * some block groups can be smaller than what a bitmap would cover but
2002          * are still large enough that they could overflow the 32k memory limit,
2003          * so allow those block groups to still be allowed to have a bitmap
2004          * entry.
2005          */
2006         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
2007                 return false;
2008
2009         return true;
2010 }
2011
2012 static const struct btrfs_free_space_op free_space_op = {
2013         .recalc_thresholds      = recalculate_thresholds,
2014         .use_bitmap             = use_bitmap,
2015 };
2016
2017 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2018                               struct btrfs_free_space *info)
2019 {
2020         struct btrfs_free_space *bitmap_info;
2021         struct btrfs_block_group_cache *block_group = NULL;
2022         int added = 0;
2023         u64 bytes, offset, bytes_added;
2024         int ret;
2025
2026         bytes = info->bytes;
2027         offset = info->offset;
2028
2029         if (!ctl->op->use_bitmap(ctl, info))
2030                 return 0;
2031
2032         if (ctl->op == &free_space_op)
2033                 block_group = ctl->private;
2034 again:
2035         /*
2036          * Since we link bitmaps right into the cluster we need to see if we
2037          * have a cluster here, and if so and it has our bitmap we need to add
2038          * the free space to that bitmap.
2039          */
2040         if (block_group && !list_empty(&block_group->cluster_list)) {
2041                 struct btrfs_free_cluster *cluster;
2042                 struct rb_node *node;
2043                 struct btrfs_free_space *entry;
2044
2045                 cluster = list_entry(block_group->cluster_list.next,
2046                                      struct btrfs_free_cluster,
2047                                      block_group_list);
2048                 spin_lock(&cluster->lock);
2049                 node = rb_first(&cluster->root);
2050                 if (!node) {
2051                         spin_unlock(&cluster->lock);
2052                         goto no_cluster_bitmap;
2053                 }
2054
2055                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2056                 if (!entry->bitmap) {
2057                         spin_unlock(&cluster->lock);
2058                         goto no_cluster_bitmap;
2059                 }
2060
2061                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2062                         bytes_added = add_bytes_to_bitmap(ctl, entry,
2063                                                           offset, bytes);
2064                         bytes -= bytes_added;
2065                         offset += bytes_added;
2066                 }
2067                 spin_unlock(&cluster->lock);
2068                 if (!bytes) {
2069                         ret = 1;
2070                         goto out;
2071                 }
2072         }
2073
2074 no_cluster_bitmap:
2075         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2076                                          1, 0);
2077         if (!bitmap_info) {
2078                 ASSERT(added == 0);
2079                 goto new_bitmap;
2080         }
2081
2082         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2083         bytes -= bytes_added;
2084         offset += bytes_added;
2085         added = 0;
2086
2087         if (!bytes) {
2088                 ret = 1;
2089                 goto out;
2090         } else
2091                 goto again;
2092
2093 new_bitmap:
2094         if (info && info->bitmap) {
2095                 add_new_bitmap(ctl, info, offset);
2096                 added = 1;
2097                 info = NULL;
2098                 goto again;
2099         } else {
2100                 spin_unlock(&ctl->tree_lock);
2101
2102                 /* no pre-allocated info, allocate a new one */
2103                 if (!info) {
2104                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2105                                                  GFP_NOFS);
2106                         if (!info) {
2107                                 spin_lock(&ctl->tree_lock);
2108                                 ret = -ENOMEM;
2109                                 goto out;
2110                         }
2111                 }
2112
2113                 /* allocate the bitmap */
2114                 info->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
2115                 spin_lock(&ctl->tree_lock);
2116                 if (!info->bitmap) {
2117                         ret = -ENOMEM;
2118                         goto out;
2119                 }
2120                 goto again;
2121         }
2122
2123 out:
2124         if (info) {
2125                 if (info->bitmap)
2126                         kfree(info->bitmap);
2127                 kmem_cache_free(btrfs_free_space_cachep, info);
2128         }
2129
2130         return ret;
2131 }
2132
2133 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2134                           struct btrfs_free_space *info, bool update_stat)
2135 {
2136         struct btrfs_free_space *left_info;
2137         struct btrfs_free_space *right_info;
2138         bool merged = false;
2139         u64 offset = info->offset;
2140         u64 bytes = info->bytes;
2141
2142         /*
2143          * first we want to see if there is free space adjacent to the range we
2144          * are adding, if there is remove that struct and add a new one to
2145          * cover the entire range
2146          */
2147         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2148         if (right_info && rb_prev(&right_info->offset_index))
2149                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2150                                      struct btrfs_free_space, offset_index);
2151         else
2152                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2153
2154         if (right_info && !right_info->bitmap) {
2155                 if (update_stat)
2156                         unlink_free_space(ctl, right_info);
2157                 else
2158                         __unlink_free_space(ctl, right_info);
2159                 info->bytes += right_info->bytes;
2160                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2161                 merged = true;
2162         }
2163
2164         if (left_info && !left_info->bitmap &&
2165             left_info->offset + left_info->bytes == offset) {
2166                 if (update_stat)
2167                         unlink_free_space(ctl, left_info);
2168                 else
2169                         __unlink_free_space(ctl, left_info);
2170                 info->offset = left_info->offset;
2171                 info->bytes += left_info->bytes;
2172                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2173                 merged = true;
2174         }
2175
2176         return merged;
2177 }
2178
2179 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2180                                      struct btrfs_free_space *info,
2181                                      bool update_stat)
2182 {
2183         struct btrfs_free_space *bitmap;
2184         unsigned long i;
2185         unsigned long j;
2186         const u64 end = info->offset + info->bytes;
2187         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2188         u64 bytes;
2189
2190         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2191         if (!bitmap)
2192                 return false;
2193
2194         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2195         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2196         if (j == i)
2197                 return false;
2198         bytes = (j - i) * ctl->unit;
2199         info->bytes += bytes;
2200
2201         if (update_stat)
2202                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2203         else
2204                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2205
2206         if (!bitmap->bytes)
2207                 free_bitmap(ctl, bitmap);
2208
2209         return true;
2210 }
2211
2212 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2213                                        struct btrfs_free_space *info,
2214                                        bool update_stat)
2215 {
2216         struct btrfs_free_space *bitmap;
2217         u64 bitmap_offset;
2218         unsigned long i;
2219         unsigned long j;
2220         unsigned long prev_j;
2221         u64 bytes;
2222
2223         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2224         /* If we're on a boundary, try the previous logical bitmap. */
2225         if (bitmap_offset == info->offset) {
2226                 if (info->offset == 0)
2227                         return false;
2228                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2229         }
2230
2231         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2232         if (!bitmap)
2233                 return false;
2234
2235         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2236         j = 0;
2237         prev_j = (unsigned long)-1;
2238         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2239                 if (j > i)
2240                         break;
2241                 prev_j = j;
2242         }
2243         if (prev_j == i)
2244                 return false;
2245
2246         if (prev_j == (unsigned long)-1)
2247                 bytes = (i + 1) * ctl->unit;
2248         else
2249                 bytes = (i - prev_j) * ctl->unit;
2250
2251         info->offset -= bytes;
2252         info->bytes += bytes;
2253
2254         if (update_stat)
2255                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2256         else
2257                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2258
2259         if (!bitmap->bytes)
2260                 free_bitmap(ctl, bitmap);
2261
2262         return true;
2263 }
2264
2265 /*
2266  * We prefer always to allocate from extent entries, both for clustered and
2267  * non-clustered allocation requests. So when attempting to add a new extent
2268  * entry, try to see if there's adjacent free space in bitmap entries, and if
2269  * there is, migrate that space from the bitmaps to the extent.
2270  * Like this we get better chances of satisfying space allocation requests
2271  * because we attempt to satisfy them based on a single cache entry, and never
2272  * on 2 or more entries - even if the entries represent a contiguous free space
2273  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2274  * ends).
2275  */
2276 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2277                               struct btrfs_free_space *info,
2278                               bool update_stat)
2279 {
2280         /*
2281          * Only work with disconnected entries, as we can change their offset,
2282          * and must be extent entries.
2283          */
2284         ASSERT(!info->bitmap);
2285         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2286
2287         if (ctl->total_bitmaps > 0) {
2288                 bool stole_end;
2289                 bool stole_front = false;
2290
2291                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2292                 if (ctl->total_bitmaps > 0)
2293                         stole_front = steal_from_bitmap_to_front(ctl, info,
2294                                                                  update_stat);
2295
2296                 if (stole_end || stole_front)
2297                         try_merge_free_space(ctl, info, update_stat);
2298         }
2299 }
2300
2301 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
2302                            u64 offset, u64 bytes)
2303 {
2304         struct btrfs_free_space *info;
2305         int ret = 0;
2306
2307         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2308         if (!info)
2309                 return -ENOMEM;
2310
2311         info->offset = offset;
2312         info->bytes = bytes;
2313         RB_CLEAR_NODE(&info->offset_index);
2314
2315         spin_lock(&ctl->tree_lock);
2316
2317         if (try_merge_free_space(ctl, info, true))
2318                 goto link;
2319
2320         /*
2321          * There was no extent directly to the left or right of this new
2322          * extent then we know we're going to have to allocate a new extent, so
2323          * before we do that see if we need to drop this into a bitmap
2324          */
2325         ret = insert_into_bitmap(ctl, info);
2326         if (ret < 0) {
2327                 goto out;
2328         } else if (ret) {
2329                 ret = 0;
2330                 goto out;
2331         }
2332 link:
2333         /*
2334          * Only steal free space from adjacent bitmaps if we're sure we're not
2335          * going to add the new free space to existing bitmap entries - because
2336          * that would mean unnecessary work that would be reverted. Therefore
2337          * attempt to steal space from bitmaps if we're adding an extent entry.
2338          */
2339         steal_from_bitmap(ctl, info, true);
2340
2341         ret = link_free_space(ctl, info);
2342         if (ret)
2343                 kmem_cache_free(btrfs_free_space_cachep, info);
2344 out:
2345         spin_unlock(&ctl->tree_lock);
2346
2347         if (ret) {
2348                 printk(KERN_CRIT "BTRFS: unable to add free space :%d\n", ret);
2349                 ASSERT(ret != -EEXIST);
2350         }
2351
2352         return ret;
2353 }
2354
2355 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2356                             u64 offset, u64 bytes)
2357 {
2358         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2359         struct btrfs_free_space *info;
2360         int ret;
2361         bool re_search = false;
2362
2363         spin_lock(&ctl->tree_lock);
2364
2365 again:
2366         ret = 0;
2367         if (!bytes)
2368                 goto out_lock;
2369
2370         info = tree_search_offset(ctl, offset, 0, 0);
2371         if (!info) {
2372                 /*
2373                  * oops didn't find an extent that matched the space we wanted
2374                  * to remove, look for a bitmap instead
2375                  */
2376                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2377                                           1, 0);
2378                 if (!info) {
2379                         /*
2380                          * If we found a partial bit of our free space in a
2381                          * bitmap but then couldn't find the other part this may
2382                          * be a problem, so WARN about it.
2383                          */
2384                         WARN_ON(re_search);
2385                         goto out_lock;
2386                 }
2387         }
2388
2389         re_search = false;
2390         if (!info->bitmap) {
2391                 unlink_free_space(ctl, info);
2392                 if (offset == info->offset) {
2393                         u64 to_free = min(bytes, info->bytes);
2394
2395                         info->bytes -= to_free;
2396                         info->offset += to_free;
2397                         if (info->bytes) {
2398                                 ret = link_free_space(ctl, info);
2399                                 WARN_ON(ret);
2400                         } else {
2401                                 kmem_cache_free(btrfs_free_space_cachep, info);
2402                         }
2403
2404                         offset += to_free;
2405                         bytes -= to_free;
2406                         goto again;
2407                 } else {
2408                         u64 old_end = info->bytes + info->offset;
2409
2410                         info->bytes = offset - info->offset;
2411                         ret = link_free_space(ctl, info);
2412                         WARN_ON(ret);
2413                         if (ret)
2414                                 goto out_lock;
2415
2416                         /* Not enough bytes in this entry to satisfy us */
2417                         if (old_end < offset + bytes) {
2418                                 bytes -= old_end - offset;
2419                                 offset = old_end;
2420                                 goto again;
2421                         } else if (old_end == offset + bytes) {
2422                                 /* all done */
2423                                 goto out_lock;
2424                         }
2425                         spin_unlock(&ctl->tree_lock);
2426
2427                         ret = btrfs_add_free_space(block_group, offset + bytes,
2428                                                    old_end - (offset + bytes));
2429                         WARN_ON(ret);
2430                         goto out;
2431                 }
2432         }
2433
2434         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2435         if (ret == -EAGAIN) {
2436                 re_search = true;
2437                 goto again;
2438         }
2439 out_lock:
2440         spin_unlock(&ctl->tree_lock);
2441 out:
2442         return ret;
2443 }
2444
2445 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2446                            u64 bytes)
2447 {
2448         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2449         struct btrfs_free_space *info;
2450         struct rb_node *n;
2451         int count = 0;
2452
2453         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2454                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2455                 if (info->bytes >= bytes && !block_group->ro)
2456                         count++;
2457                 btrfs_crit(block_group->fs_info,
2458                            "entry offset %llu, bytes %llu, bitmap %s",
2459                            info->offset, info->bytes,
2460                        (info->bitmap) ? "yes" : "no");
2461         }
2462         btrfs_info(block_group->fs_info, "block group has cluster?: %s",
2463                list_empty(&block_group->cluster_list) ? "no" : "yes");
2464         btrfs_info(block_group->fs_info,
2465                    "%d blocks of free space at or bigger than bytes is", count);
2466 }
2467
2468 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2469 {
2470         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2471
2472         spin_lock_init(&ctl->tree_lock);
2473         ctl->unit = block_group->sectorsize;
2474         ctl->start = block_group->key.objectid;
2475         ctl->private = block_group;
2476         ctl->op = &free_space_op;
2477         INIT_LIST_HEAD(&ctl->trimming_ranges);
2478         mutex_init(&ctl->cache_writeout_mutex);
2479
2480         /*
2481          * we only want to have 32k of ram per block group for keeping
2482          * track of free space, and if we pass 1/2 of that we want to
2483          * start converting things over to using bitmaps
2484          */
2485         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2486 }
2487
2488 /*
2489  * for a given cluster, put all of its extents back into the free
2490  * space cache.  If the block group passed doesn't match the block group
2491  * pointed to by the cluster, someone else raced in and freed the
2492  * cluster already.  In that case, we just return without changing anything
2493  */
2494 static int
2495 __btrfs_return_cluster_to_free_space(
2496                              struct btrfs_block_group_cache *block_group,
2497                              struct btrfs_free_cluster *cluster)
2498 {
2499         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2500         struct btrfs_free_space *entry;
2501         struct rb_node *node;
2502
2503         spin_lock(&cluster->lock);
2504         if (cluster->block_group != block_group)
2505                 goto out;
2506
2507         cluster->block_group = NULL;
2508         cluster->window_start = 0;
2509         list_del_init(&cluster->block_group_list);
2510
2511         node = rb_first(&cluster->root);
2512         while (node) {
2513                 bool bitmap;
2514
2515                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2516                 node = rb_next(&entry->offset_index);
2517                 rb_erase(&entry->offset_index, &cluster->root);
2518                 RB_CLEAR_NODE(&entry->offset_index);
2519
2520                 bitmap = (entry->bitmap != NULL);
2521                 if (!bitmap) {
2522                         try_merge_free_space(ctl, entry, false);
2523                         steal_from_bitmap(ctl, entry, false);
2524                 }
2525                 tree_insert_offset(&ctl->free_space_offset,
2526                                    entry->offset, &entry->offset_index, bitmap);
2527         }
2528         cluster->root = RB_ROOT;
2529
2530 out:
2531         spin_unlock(&cluster->lock);
2532         btrfs_put_block_group(block_group);
2533         return 0;
2534 }
2535
2536 static void __btrfs_remove_free_space_cache_locked(
2537                                 struct btrfs_free_space_ctl *ctl)
2538 {
2539         struct btrfs_free_space *info;
2540         struct rb_node *node;
2541
2542         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2543                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2544                 if (!info->bitmap) {
2545                         unlink_free_space(ctl, info);
2546                         kmem_cache_free(btrfs_free_space_cachep, info);
2547                 } else {
2548                         free_bitmap(ctl, info);
2549                 }
2550
2551                 cond_resched_lock(&ctl->tree_lock);
2552         }
2553 }
2554
2555 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2556 {
2557         spin_lock(&ctl->tree_lock);
2558         __btrfs_remove_free_space_cache_locked(ctl);
2559         spin_unlock(&ctl->tree_lock);
2560 }
2561
2562 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2563 {
2564         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2565         struct btrfs_free_cluster *cluster;
2566         struct list_head *head;
2567
2568         spin_lock(&ctl->tree_lock);
2569         while ((head = block_group->cluster_list.next) !=
2570                &block_group->cluster_list) {
2571                 cluster = list_entry(head, struct btrfs_free_cluster,
2572                                      block_group_list);
2573
2574                 WARN_ON(cluster->block_group != block_group);
2575                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2576
2577                 cond_resched_lock(&ctl->tree_lock);
2578         }
2579         __btrfs_remove_free_space_cache_locked(ctl);
2580         spin_unlock(&ctl->tree_lock);
2581
2582 }
2583
2584 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2585                                u64 offset, u64 bytes, u64 empty_size,
2586                                u64 *max_extent_size)
2587 {
2588         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2589         struct btrfs_free_space *entry = NULL;
2590         u64 bytes_search = bytes + empty_size;
2591         u64 ret = 0;
2592         u64 align_gap = 0;
2593         u64 align_gap_len = 0;
2594
2595         spin_lock(&ctl->tree_lock);
2596         entry = find_free_space(ctl, &offset, &bytes_search,
2597                                 block_group->full_stripe_len, max_extent_size);
2598         if (!entry)
2599                 goto out;
2600
2601         ret = offset;
2602         if (entry->bitmap) {
2603                 bitmap_clear_bits(ctl, entry, offset, bytes);
2604                 if (!entry->bytes)
2605                         free_bitmap(ctl, entry);
2606         } else {
2607                 unlink_free_space(ctl, entry);
2608                 align_gap_len = offset - entry->offset;
2609                 align_gap = entry->offset;
2610
2611                 entry->offset = offset + bytes;
2612                 WARN_ON(entry->bytes < bytes + align_gap_len);
2613
2614                 entry->bytes -= bytes + align_gap_len;
2615                 if (!entry->bytes)
2616                         kmem_cache_free(btrfs_free_space_cachep, entry);
2617                 else
2618                         link_free_space(ctl, entry);
2619         }
2620 out:
2621         spin_unlock(&ctl->tree_lock);
2622
2623         if (align_gap_len)
2624                 __btrfs_add_free_space(ctl, align_gap, align_gap_len);
2625         return ret;
2626 }
2627
2628 /*
2629  * given a cluster, put all of its extents back into the free space
2630  * cache.  If a block group is passed, this function will only free
2631  * a cluster that belongs to the passed block group.
2632  *
2633  * Otherwise, it'll get a reference on the block group pointed to by the
2634  * cluster and remove the cluster from it.
2635  */
2636 int btrfs_return_cluster_to_free_space(
2637                                struct btrfs_block_group_cache *block_group,
2638                                struct btrfs_free_cluster *cluster)
2639 {
2640         struct btrfs_free_space_ctl *ctl;
2641         int ret;
2642
2643         /* first, get a safe pointer to the block group */
2644         spin_lock(&cluster->lock);
2645         if (!block_group) {
2646                 block_group = cluster->block_group;
2647                 if (!block_group) {
2648                         spin_unlock(&cluster->lock);
2649                         return 0;
2650                 }
2651         } else if (cluster->block_group != block_group) {
2652                 /* someone else has already freed it don't redo their work */
2653                 spin_unlock(&cluster->lock);
2654                 return 0;
2655         }
2656         atomic_inc(&block_group->count);
2657         spin_unlock(&cluster->lock);
2658
2659         ctl = block_group->free_space_ctl;
2660
2661         /* now return any extents the cluster had on it */
2662         spin_lock(&ctl->tree_lock);
2663         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2664         spin_unlock(&ctl->tree_lock);
2665
2666         /* finally drop our ref */
2667         btrfs_put_block_group(block_group);
2668         return ret;
2669 }
2670
2671 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2672                                    struct btrfs_free_cluster *cluster,
2673                                    struct btrfs_free_space *entry,
2674                                    u64 bytes, u64 min_start,
2675                                    u64 *max_extent_size)
2676 {
2677         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2678         int err;
2679         u64 search_start = cluster->window_start;
2680         u64 search_bytes = bytes;
2681         u64 ret = 0;
2682
2683         search_start = min_start;
2684         search_bytes = bytes;
2685
2686         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2687         if (err) {
2688                 if (search_bytes > *max_extent_size)
2689                         *max_extent_size = search_bytes;
2690                 return 0;
2691         }
2692
2693         ret = search_start;
2694         __bitmap_clear_bits(ctl, entry, ret, bytes);
2695
2696         return ret;
2697 }
2698
2699 /*
2700  * given a cluster, try to allocate 'bytes' from it, returns 0
2701  * if it couldn't find anything suitably large, or a logical disk offset
2702  * if things worked out
2703  */
2704 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2705                              struct btrfs_free_cluster *cluster, u64 bytes,
2706                              u64 min_start, u64 *max_extent_size)
2707 {
2708         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2709         struct btrfs_free_space *entry = NULL;
2710         struct rb_node *node;
2711         u64 ret = 0;
2712
2713         spin_lock(&cluster->lock);
2714         if (bytes > cluster->max_size)
2715                 goto out;
2716
2717         if (cluster->block_group != block_group)
2718                 goto out;
2719
2720         node = rb_first(&cluster->root);
2721         if (!node)
2722                 goto out;
2723
2724         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2725         while (1) {
2726                 if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2727                         *max_extent_size = entry->bytes;
2728
2729                 if (entry->bytes < bytes ||
2730                     (!entry->bitmap && entry->offset < min_start)) {
2731                         node = rb_next(&entry->offset_index);
2732                         if (!node)
2733                                 break;
2734                         entry = rb_entry(node, struct btrfs_free_space,
2735                                          offset_index);
2736                         continue;
2737                 }
2738
2739                 if (entry->bitmap) {
2740                         ret = btrfs_alloc_from_bitmap(block_group,
2741                                                       cluster, entry, bytes,
2742                                                       cluster->window_start,
2743                                                       max_extent_size);
2744                         if (ret == 0) {
2745                                 node = rb_next(&entry->offset_index);
2746                                 if (!node)
2747                                         break;
2748                                 entry = rb_entry(node, struct btrfs_free_space,
2749                                                  offset_index);
2750                                 continue;
2751                         }
2752                         cluster->window_start += bytes;
2753                 } else {
2754                         ret = entry->offset;
2755
2756                         entry->offset += bytes;
2757                         entry->bytes -= bytes;
2758                 }
2759
2760                 if (entry->bytes == 0)
2761                         rb_erase(&entry->offset_index, &cluster->root);
2762                 break;
2763         }
2764 out:
2765         spin_unlock(&cluster->lock);
2766
2767         if (!ret)
2768                 return 0;
2769
2770         spin_lock(&ctl->tree_lock);
2771
2772         ctl->free_space -= bytes;
2773         if (entry->bytes == 0) {
2774                 ctl->free_extents--;
2775                 if (entry->bitmap) {
2776                         kfree(entry->bitmap);
2777                         ctl->total_bitmaps--;
2778                         ctl->op->recalc_thresholds(ctl);
2779                 }
2780                 kmem_cache_free(btrfs_free_space_cachep, entry);
2781         }
2782
2783         spin_unlock(&ctl->tree_lock);
2784
2785         return ret;
2786 }
2787
2788 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2789                                 struct btrfs_free_space *entry,
2790                                 struct btrfs_free_cluster *cluster,
2791                                 u64 offset, u64 bytes,
2792                                 u64 cont1_bytes, u64 min_bytes)
2793 {
2794         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2795         unsigned long next_zero;
2796         unsigned long i;
2797         unsigned long want_bits;
2798         unsigned long min_bits;
2799         unsigned long found_bits;
2800         unsigned long max_bits = 0;
2801         unsigned long start = 0;
2802         unsigned long total_found = 0;
2803         int ret;
2804
2805         i = offset_to_bit(entry->offset, ctl->unit,
2806                           max_t(u64, offset, entry->offset));
2807         want_bits = bytes_to_bits(bytes, ctl->unit);
2808         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2809
2810         /*
2811          * Don't bother looking for a cluster in this bitmap if it's heavily
2812          * fragmented.
2813          */
2814         if (entry->max_extent_size &&
2815             entry->max_extent_size < cont1_bytes)
2816                 return -ENOSPC;
2817 again:
2818         found_bits = 0;
2819         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2820                 next_zero = find_next_zero_bit(entry->bitmap,
2821                                                BITS_PER_BITMAP, i);
2822                 if (next_zero - i >= min_bits) {
2823                         found_bits = next_zero - i;
2824                         if (found_bits > max_bits)
2825                                 max_bits = found_bits;
2826                         break;
2827                 }
2828                 if (next_zero - i > max_bits)
2829                         max_bits = next_zero - i;
2830                 i = next_zero;
2831         }
2832
2833         if (!found_bits) {
2834                 entry->max_extent_size = (u64)max_bits * ctl->unit;
2835                 return -ENOSPC;
2836         }
2837
2838         if (!total_found) {
2839                 start = i;
2840                 cluster->max_size = 0;
2841         }
2842
2843         total_found += found_bits;
2844
2845         if (cluster->max_size < found_bits * ctl->unit)
2846                 cluster->max_size = found_bits * ctl->unit;
2847
2848         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2849                 i = next_zero + 1;
2850                 goto again;
2851         }
2852
2853         cluster->window_start = start * ctl->unit + entry->offset;
2854         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2855         ret = tree_insert_offset(&cluster->root, entry->offset,
2856                                  &entry->offset_index, 1);
2857         ASSERT(!ret); /* -EEXIST; Logic error */
2858
2859         trace_btrfs_setup_cluster(block_group, cluster,
2860                                   total_found * ctl->unit, 1);
2861         return 0;
2862 }
2863
2864 /*
2865  * This searches the block group for just extents to fill the cluster with.
2866  * Try to find a cluster with at least bytes total bytes, at least one
2867  * extent of cont1_bytes, and other clusters of at least min_bytes.
2868  */
2869 static noinline int
2870 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2871                         struct btrfs_free_cluster *cluster,
2872                         struct list_head *bitmaps, u64 offset, u64 bytes,
2873                         u64 cont1_bytes, u64 min_bytes)
2874 {
2875         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2876         struct btrfs_free_space *first = NULL;
2877         struct btrfs_free_space *entry = NULL;
2878         struct btrfs_free_space *last;
2879         struct rb_node *node;
2880         u64 window_free;
2881         u64 max_extent;
2882         u64 total_size = 0;
2883
2884         entry = tree_search_offset(ctl, offset, 0, 1);
2885         if (!entry)
2886                 return -ENOSPC;
2887
2888         /*
2889          * We don't want bitmaps, so just move along until we find a normal
2890          * extent entry.
2891          */
2892         while (entry->bitmap || entry->bytes < min_bytes) {
2893                 if (entry->bitmap && list_empty(&entry->list))
2894                         list_add_tail(&entry->list, bitmaps);
2895                 node = rb_next(&entry->offset_index);
2896                 if (!node)
2897                         return -ENOSPC;
2898                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2899         }
2900
2901         window_free = entry->bytes;
2902         max_extent = entry->bytes;
2903         first = entry;
2904         last = entry;
2905
2906         for (node = rb_next(&entry->offset_index); node;
2907              node = rb_next(&entry->offset_index)) {
2908                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2909
2910                 if (entry->bitmap) {
2911                         if (list_empty(&entry->list))
2912                                 list_add_tail(&entry->list, bitmaps);
2913                         continue;
2914                 }
2915
2916                 if (entry->bytes < min_bytes)
2917                         continue;
2918
2919                 last = entry;
2920                 window_free += entry->bytes;
2921                 if (entry->bytes > max_extent)
2922                         max_extent = entry->bytes;
2923         }
2924
2925         if (window_free < bytes || max_extent < cont1_bytes)
2926                 return -ENOSPC;
2927
2928         cluster->window_start = first->offset;
2929
2930         node = &first->offset_index;
2931
2932         /*
2933          * now we've found our entries, pull them out of the free space
2934          * cache and put them into the cluster rbtree
2935          */
2936         do {
2937                 int ret;
2938
2939                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2940                 node = rb_next(&entry->offset_index);
2941                 if (entry->bitmap || entry->bytes < min_bytes)
2942                         continue;
2943
2944                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2945                 ret = tree_insert_offset(&cluster->root, entry->offset,
2946                                          &entry->offset_index, 0);
2947                 total_size += entry->bytes;
2948                 ASSERT(!ret); /* -EEXIST; Logic error */
2949         } while (node && entry != last);
2950
2951         cluster->max_size = max_extent;
2952         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2953         return 0;
2954 }
2955
2956 /*
2957  * This specifically looks for bitmaps that may work in the cluster, we assume
2958  * that we have already failed to find extents that will work.
2959  */
2960 static noinline int
2961 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2962                      struct btrfs_free_cluster *cluster,
2963                      struct list_head *bitmaps, u64 offset, u64 bytes,
2964                      u64 cont1_bytes, u64 min_bytes)
2965 {
2966         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2967         struct btrfs_free_space *entry = NULL;
2968         int ret = -ENOSPC;
2969         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2970
2971         if (ctl->total_bitmaps == 0)
2972                 return -ENOSPC;
2973
2974         /*
2975          * The bitmap that covers offset won't be in the list unless offset
2976          * is just its start offset.
2977          */
2978         if (!list_empty(bitmaps))
2979                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2980
2981         if (!entry || entry->offset != bitmap_offset) {
2982                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2983                 if (entry && list_empty(&entry->list))
2984                         list_add(&entry->list, bitmaps);
2985         }
2986
2987         list_for_each_entry(entry, bitmaps, list) {
2988                 if (entry->bytes < bytes)
2989                         continue;
2990                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2991                                            bytes, cont1_bytes, min_bytes);
2992                 if (!ret)
2993                         return 0;
2994         }
2995
2996         /*
2997          * The bitmaps list has all the bitmaps that record free space
2998          * starting after offset, so no more search is required.
2999          */
3000         return -ENOSPC;
3001 }
3002
3003 /*
3004  * here we try to find a cluster of blocks in a block group.  The goal
3005  * is to find at least bytes+empty_size.
3006  * We might not find them all in one contiguous area.
3007  *
3008  * returns zero and sets up cluster if things worked out, otherwise
3009  * it returns -enospc
3010  */
3011 int btrfs_find_space_cluster(struct btrfs_root *root,
3012                              struct btrfs_block_group_cache *block_group,
3013                              struct btrfs_free_cluster *cluster,
3014                              u64 offset, u64 bytes, u64 empty_size)
3015 {
3016         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3017         struct btrfs_free_space *entry, *tmp;
3018         LIST_HEAD(bitmaps);
3019         u64 min_bytes;
3020         u64 cont1_bytes;
3021         int ret;
3022
3023         /*
3024          * Choose the minimum extent size we'll require for this
3025          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3026          * For metadata, allow allocates with smaller extents.  For
3027          * data, keep it dense.
3028          */
3029         if (btrfs_test_opt(root->fs_info, SSD_SPREAD)) {
3030                 cont1_bytes = min_bytes = bytes + empty_size;
3031         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3032                 cont1_bytes = bytes;
3033                 min_bytes = block_group->sectorsize;
3034         } else {
3035                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3036                 min_bytes = block_group->sectorsize;
3037         }
3038
3039         spin_lock(&ctl->tree_lock);
3040
3041         /*
3042          * If we know we don't have enough space to make a cluster don't even
3043          * bother doing all the work to try and find one.
3044          */
3045         if (ctl->free_space < bytes) {
3046                 spin_unlock(&ctl->tree_lock);
3047                 return -ENOSPC;
3048         }
3049
3050         spin_lock(&cluster->lock);
3051
3052         /* someone already found a cluster, hooray */
3053         if (cluster->block_group) {
3054                 ret = 0;
3055                 goto out;
3056         }
3057
3058         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3059                                  min_bytes);
3060
3061         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3062                                       bytes + empty_size,
3063                                       cont1_bytes, min_bytes);
3064         if (ret)
3065                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3066                                            offset, bytes + empty_size,
3067                                            cont1_bytes, min_bytes);
3068
3069         /* Clear our temporary list */
3070         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3071                 list_del_init(&entry->list);
3072
3073         if (!ret) {
3074                 atomic_inc(&block_group->count);
3075                 list_add_tail(&cluster->block_group_list,
3076                               &block_group->cluster_list);
3077                 cluster->block_group = block_group;
3078         } else {
3079                 trace_btrfs_failed_cluster_setup(block_group);
3080         }
3081 out:
3082         spin_unlock(&cluster->lock);
3083         spin_unlock(&ctl->tree_lock);
3084
3085         return ret;
3086 }
3087
3088 /*
3089  * simple code to zero out a cluster
3090  */
3091 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3092 {
3093         spin_lock_init(&cluster->lock);
3094         spin_lock_init(&cluster->refill_lock);
3095         cluster->root = RB_ROOT;
3096         cluster->max_size = 0;
3097         cluster->fragmented = false;
3098         INIT_LIST_HEAD(&cluster->block_group_list);
3099         cluster->block_group = NULL;
3100 }
3101
3102 static int do_trimming(struct btrfs_block_group_cache *block_group,
3103                        u64 *total_trimmed, u64 start, u64 bytes,
3104                        u64 reserved_start, u64 reserved_bytes,
3105                        struct btrfs_trim_range *trim_entry)
3106 {
3107         struct btrfs_space_info *space_info = block_group->space_info;
3108         struct btrfs_fs_info *fs_info = block_group->fs_info;
3109         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3110         int ret;
3111         int update = 0;
3112         u64 trimmed = 0;
3113
3114         spin_lock(&space_info->lock);
3115         spin_lock(&block_group->lock);
3116         if (!block_group->ro) {
3117                 block_group->reserved += reserved_bytes;
3118                 space_info->bytes_reserved += reserved_bytes;
3119                 update = 1;
3120         }
3121         spin_unlock(&block_group->lock);
3122         spin_unlock(&space_info->lock);
3123
3124         ret = btrfs_discard_extent(fs_info->extent_root,
3125                                    start, bytes, &trimmed);
3126         if (!ret)
3127                 *total_trimmed += trimmed;
3128
3129         mutex_lock(&ctl->cache_writeout_mutex);
3130         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3131         list_del(&trim_entry->list);
3132         mutex_unlock(&ctl->cache_writeout_mutex);
3133
3134         if (update) {
3135                 spin_lock(&space_info->lock);
3136                 spin_lock(&block_group->lock);
3137                 if (block_group->ro)
3138                         space_info->bytes_readonly += reserved_bytes;
3139                 block_group->reserved -= reserved_bytes;
3140                 space_info->bytes_reserved -= reserved_bytes;
3141                 spin_unlock(&space_info->lock);
3142                 spin_unlock(&block_group->lock);
3143         }
3144
3145         return ret;
3146 }
3147
3148 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3149                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3150 {
3151         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3152         struct btrfs_free_space *entry;
3153         struct rb_node *node;
3154         int ret = 0;
3155         u64 extent_start;
3156         u64 extent_bytes;
3157         u64 bytes;
3158
3159         while (start < end) {
3160                 struct btrfs_trim_range trim_entry;
3161
3162                 mutex_lock(&ctl->cache_writeout_mutex);
3163                 spin_lock(&ctl->tree_lock);
3164
3165                 if (ctl->free_space < minlen) {
3166                         spin_unlock(&ctl->tree_lock);
3167                         mutex_unlock(&ctl->cache_writeout_mutex);
3168                         break;
3169                 }
3170
3171                 entry = tree_search_offset(ctl, start, 0, 1);
3172                 if (!entry) {
3173                         spin_unlock(&ctl->tree_lock);
3174                         mutex_unlock(&ctl->cache_writeout_mutex);
3175                         break;
3176                 }
3177
3178                 /* skip bitmaps */
3179                 while (entry->bitmap) {
3180                         node = rb_next(&entry->offset_index);
3181                         if (!node) {
3182                                 spin_unlock(&ctl->tree_lock);
3183                                 mutex_unlock(&ctl->cache_writeout_mutex);
3184                                 goto out;
3185                         }
3186                         entry = rb_entry(node, struct btrfs_free_space,
3187                                          offset_index);
3188                 }
3189
3190                 if (entry->offset >= end) {
3191                         spin_unlock(&ctl->tree_lock);
3192                         mutex_unlock(&ctl->cache_writeout_mutex);
3193                         break;
3194                 }
3195
3196                 extent_start = entry->offset;
3197                 extent_bytes = entry->bytes;
3198                 start = max(start, extent_start);
3199                 bytes = min(extent_start + extent_bytes, end) - start;
3200                 if (bytes < minlen) {
3201                         spin_unlock(&ctl->tree_lock);
3202                         mutex_unlock(&ctl->cache_writeout_mutex);
3203                         goto next;
3204                 }
3205
3206                 unlink_free_space(ctl, entry);
3207                 kmem_cache_free(btrfs_free_space_cachep, entry);
3208
3209                 spin_unlock(&ctl->tree_lock);
3210                 trim_entry.start = extent_start;
3211                 trim_entry.bytes = extent_bytes;
3212                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3213                 mutex_unlock(&ctl->cache_writeout_mutex);
3214
3215                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3216                                   extent_start, extent_bytes, &trim_entry);
3217                 if (ret)
3218                         break;
3219 next:
3220                 start += bytes;
3221
3222                 if (fatal_signal_pending(current)) {
3223                         ret = -ERESTARTSYS;
3224                         break;
3225                 }
3226
3227                 cond_resched();
3228         }
3229 out:
3230         return ret;
3231 }
3232
3233 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3234                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3235 {
3236         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3237         struct btrfs_free_space *entry;
3238         int ret = 0;
3239         int ret2;
3240         u64 bytes;
3241         u64 offset = offset_to_bitmap(ctl, start);
3242
3243         while (offset < end) {
3244                 bool next_bitmap = false;
3245                 struct btrfs_trim_range trim_entry;
3246
3247                 mutex_lock(&ctl->cache_writeout_mutex);
3248                 spin_lock(&ctl->tree_lock);
3249
3250                 if (ctl->free_space < minlen) {
3251                         spin_unlock(&ctl->tree_lock);
3252                         mutex_unlock(&ctl->cache_writeout_mutex);
3253                         break;
3254                 }
3255
3256                 entry = tree_search_offset(ctl, offset, 1, 0);
3257                 if (!entry) {
3258                         spin_unlock(&ctl->tree_lock);
3259                         mutex_unlock(&ctl->cache_writeout_mutex);
3260                         next_bitmap = true;
3261                         goto next;
3262                 }
3263
3264                 bytes = minlen;
3265                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3266                 if (ret2 || start >= end) {
3267                         spin_unlock(&ctl->tree_lock);
3268                         mutex_unlock(&ctl->cache_writeout_mutex);
3269                         next_bitmap = true;
3270                         goto next;
3271                 }
3272
3273                 bytes = min(bytes, end - start);
3274                 if (bytes < minlen) {
3275                         spin_unlock(&ctl->tree_lock);
3276                         mutex_unlock(&ctl->cache_writeout_mutex);
3277                         goto next;
3278                 }
3279
3280                 bitmap_clear_bits(ctl, entry, start, bytes);
3281                 if (entry->bytes == 0)
3282                         free_bitmap(ctl, entry);
3283
3284                 spin_unlock(&ctl->tree_lock);
3285                 trim_entry.start = start;
3286                 trim_entry.bytes = bytes;
3287                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3288                 mutex_unlock(&ctl->cache_writeout_mutex);
3289
3290                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3291                                   start, bytes, &trim_entry);
3292                 if (ret)
3293                         break;
3294 next:
3295                 if (next_bitmap) {
3296                         offset += BITS_PER_BITMAP * ctl->unit;
3297                 } else {
3298                         start += bytes;
3299                         if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3300                                 offset += BITS_PER_BITMAP * ctl->unit;
3301                 }
3302
3303                 if (fatal_signal_pending(current)) {
3304                         ret = -ERESTARTSYS;
3305                         break;
3306                 }
3307
3308                 cond_resched();
3309         }
3310
3311         return ret;
3312 }
3313
3314 void btrfs_get_block_group_trimming(struct btrfs_block_group_cache *cache)
3315 {
3316         atomic_inc(&cache->trimming);
3317 }
3318
3319 void btrfs_put_block_group_trimming(struct btrfs_block_group_cache *block_group)
3320 {
3321         struct extent_map_tree *em_tree;
3322         struct extent_map *em;
3323         bool cleanup;
3324
3325         spin_lock(&block_group->lock);
3326         cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3327                    block_group->removed);
3328         spin_unlock(&block_group->lock);
3329
3330         if (cleanup) {
3331                 lock_chunks(block_group->fs_info->chunk_root);
3332                 em_tree = &block_group->fs_info->mapping_tree.map_tree;
3333                 write_lock(&em_tree->lock);
3334                 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3335                                            1);
3336                 BUG_ON(!em); /* logic error, can't happen */
3337                 /*
3338                  * remove_extent_mapping() will delete us from the pinned_chunks
3339                  * list, which is protected by the chunk mutex.
3340                  */
3341                 remove_extent_mapping(em_tree, em);
3342                 write_unlock(&em_tree->lock);
3343                 unlock_chunks(block_group->fs_info->chunk_root);
3344
3345                 /* once for us and once for the tree */
3346                 free_extent_map(em);
3347                 free_extent_map(em);
3348
3349                 /*
3350                  * We've left one free space entry and other tasks trimming
3351                  * this block group have left 1 entry each one. Free them.
3352                  */
3353                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3354         }
3355 }
3356
3357 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3358                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3359 {
3360         int ret;
3361
3362         *trimmed = 0;
3363
3364         spin_lock(&block_group->lock);
3365         if (block_group->removed) {
3366                 spin_unlock(&block_group->lock);
3367                 return 0;
3368         }
3369         btrfs_get_block_group_trimming(block_group);
3370         spin_unlock(&block_group->lock);
3371
3372         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3373         if (ret)
3374                 goto out;
3375
3376         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3377 out:
3378         btrfs_put_block_group_trimming(block_group);
3379         return ret;
3380 }
3381
3382 /*
3383  * Find the left-most item in the cache tree, and then return the
3384  * smallest inode number in the item.
3385  *
3386  * Note: the returned inode number may not be the smallest one in
3387  * the tree, if the left-most item is a bitmap.
3388  */
3389 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3390 {
3391         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3392         struct btrfs_free_space *entry = NULL;
3393         u64 ino = 0;
3394
3395         spin_lock(&ctl->tree_lock);
3396
3397         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3398                 goto out;
3399
3400         entry = rb_entry(rb_first(&ctl->free_space_offset),
3401                          struct btrfs_free_space, offset_index);
3402
3403         if (!entry->bitmap) {
3404                 ino = entry->offset;
3405
3406                 unlink_free_space(ctl, entry);
3407                 entry->offset++;
3408                 entry->bytes--;
3409                 if (!entry->bytes)
3410                         kmem_cache_free(btrfs_free_space_cachep, entry);
3411                 else
3412                         link_free_space(ctl, entry);
3413         } else {
3414                 u64 offset = 0;
3415                 u64 count = 1;
3416                 int ret;
3417
3418                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3419                 /* Logic error; Should be empty if it can't find anything */
3420                 ASSERT(!ret);
3421
3422                 ino = offset;
3423                 bitmap_clear_bits(ctl, entry, offset, 1);
3424                 if (entry->bytes == 0)
3425                         free_bitmap(ctl, entry);
3426         }
3427 out:
3428         spin_unlock(&ctl->tree_lock);
3429
3430         return ino;
3431 }
3432
3433 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3434                                     struct btrfs_path *path)
3435 {
3436         struct inode *inode = NULL;
3437
3438         spin_lock(&root->ino_cache_lock);
3439         if (root->ino_cache_inode)
3440                 inode = igrab(root->ino_cache_inode);
3441         spin_unlock(&root->ino_cache_lock);
3442         if (inode)
3443                 return inode;
3444
3445         inode = __lookup_free_space_inode(root, path, 0);
3446         if (IS_ERR(inode))
3447                 return inode;
3448
3449         spin_lock(&root->ino_cache_lock);
3450         if (!btrfs_fs_closing(root->fs_info))
3451                 root->ino_cache_inode = igrab(inode);
3452         spin_unlock(&root->ino_cache_lock);
3453
3454         return inode;
3455 }
3456
3457 int create_free_ino_inode(struct btrfs_root *root,
3458                           struct btrfs_trans_handle *trans,
3459                           struct btrfs_path *path)
3460 {
3461         return __create_free_space_inode(root, trans, path,
3462                                          BTRFS_FREE_INO_OBJECTID, 0);
3463 }
3464
3465 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3466 {
3467         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3468         struct btrfs_path *path;
3469         struct inode *inode;
3470         int ret = 0;
3471         u64 root_gen = btrfs_root_generation(&root->root_item);
3472
3473         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
3474                 return 0;
3475
3476         /*
3477          * If we're unmounting then just return, since this does a search on the
3478          * normal root and not the commit root and we could deadlock.
3479          */
3480         if (btrfs_fs_closing(fs_info))
3481                 return 0;
3482
3483         path = btrfs_alloc_path();
3484         if (!path)
3485                 return 0;
3486
3487         inode = lookup_free_ino_inode(root, path);
3488         if (IS_ERR(inode))
3489                 goto out;
3490
3491         if (root_gen != BTRFS_I(inode)->generation)
3492                 goto out_put;
3493
3494         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3495
3496         if (ret < 0)
3497                 btrfs_err(fs_info,
3498                         "failed to load free ino cache for root %llu",
3499                         root->root_key.objectid);
3500 out_put:
3501         iput(inode);
3502 out:
3503         btrfs_free_path(path);
3504         return ret;
3505 }
3506
3507 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3508                               struct btrfs_trans_handle *trans,
3509                               struct btrfs_path *path,
3510                               struct inode *inode)
3511 {
3512         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3513         int ret;
3514         struct btrfs_io_ctl io_ctl;
3515         bool release_metadata = true;
3516
3517         if (!btrfs_test_opt(root->fs_info, INODE_MAP_CACHE))
3518                 return 0;
3519
3520         memset(&io_ctl, 0, sizeof(io_ctl));
3521         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl,
3522                                       trans, path, 0);
3523         if (!ret) {
3524                 /*
3525                  * At this point writepages() didn't error out, so our metadata
3526                  * reservation is released when the writeback finishes, at
3527                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3528                  * with or without an error.
3529                  */
3530                 release_metadata = false;
3531                 ret = btrfs_wait_cache_io(root, trans, NULL, &io_ctl, path, 0);
3532         }
3533
3534         if (ret) {
3535                 if (release_metadata)
3536                         btrfs_delalloc_release_metadata(inode, inode->i_size);
3537 #ifdef DEBUG
3538                 btrfs_err(root->fs_info,
3539                         "failed to write free ino cache for root %llu",
3540                         root->root_key.objectid);
3541 #endif
3542         }
3543
3544         return ret;
3545 }
3546
3547 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3548 /*
3549  * Use this if you need to make a bitmap or extent entry specifically, it
3550  * doesn't do any of the merging that add_free_space does, this acts a lot like
3551  * how the free space cache loading stuff works, so you can get really weird
3552  * configurations.
3553  */
3554 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3555                               u64 offset, u64 bytes, bool bitmap)
3556 {
3557         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3558         struct btrfs_free_space *info = NULL, *bitmap_info;
3559         void *map = NULL;
3560         u64 bytes_added;
3561         int ret;
3562
3563 again:
3564         if (!info) {
3565                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3566                 if (!info)
3567                         return -ENOMEM;
3568         }
3569
3570         if (!bitmap) {
3571                 spin_lock(&ctl->tree_lock);
3572                 info->offset = offset;
3573                 info->bytes = bytes;
3574                 info->max_extent_size = 0;
3575                 ret = link_free_space(ctl, info);
3576                 spin_unlock(&ctl->tree_lock);
3577                 if (ret)
3578                         kmem_cache_free(btrfs_free_space_cachep, info);
3579                 return ret;
3580         }
3581
3582         if (!map) {
3583                 map = kzalloc(PAGE_SIZE, GFP_NOFS);
3584                 if (!map) {
3585                         kmem_cache_free(btrfs_free_space_cachep, info);
3586                         return -ENOMEM;
3587                 }
3588         }
3589
3590         spin_lock(&ctl->tree_lock);
3591         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3592                                          1, 0);
3593         if (!bitmap_info) {
3594                 info->bitmap = map;
3595                 map = NULL;
3596                 add_new_bitmap(ctl, info, offset);
3597                 bitmap_info = info;
3598                 info = NULL;
3599         }
3600
3601         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3602
3603         bytes -= bytes_added;
3604         offset += bytes_added;
3605         spin_unlock(&ctl->tree_lock);
3606
3607         if (bytes)
3608                 goto again;
3609
3610         if (info)
3611                 kmem_cache_free(btrfs_free_space_cachep, info);
3612         if (map)
3613                 kfree(map);
3614         return 0;
3615 }
3616
3617 /*
3618  * Checks to see if the given range is in the free space cache.  This is really
3619  * just used to check the absence of space, so if there is free space in the
3620  * range at all we will return 1.
3621  */
3622 int test_check_exists(struct btrfs_block_group_cache *cache,
3623                       u64 offset, u64 bytes)
3624 {
3625         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3626         struct btrfs_free_space *info;
3627         int ret = 0;
3628
3629         spin_lock(&ctl->tree_lock);
3630         info = tree_search_offset(ctl, offset, 0, 0);
3631         if (!info) {
3632                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3633                                           1, 0);
3634                 if (!info)
3635                         goto out;
3636         }
3637
3638 have_info:
3639         if (info->bitmap) {
3640                 u64 bit_off, bit_bytes;
3641                 struct rb_node *n;
3642                 struct btrfs_free_space *tmp;
3643
3644                 bit_off = offset;
3645                 bit_bytes = ctl->unit;
3646                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3647                 if (!ret) {
3648                         if (bit_off == offset) {
3649                                 ret = 1;
3650                                 goto out;
3651                         } else if (bit_off > offset &&
3652                                    offset + bytes > bit_off) {
3653                                 ret = 1;
3654                                 goto out;
3655                         }
3656                 }
3657
3658                 n = rb_prev(&info->offset_index);
3659                 while (n) {
3660                         tmp = rb_entry(n, struct btrfs_free_space,
3661                                        offset_index);
3662                         if (tmp->offset + tmp->bytes < offset)
3663                                 break;
3664                         if (offset + bytes < tmp->offset) {
3665                                 n = rb_prev(&tmp->offset_index);
3666                                 continue;
3667                         }
3668                         info = tmp;
3669                         goto have_info;
3670                 }
3671
3672                 n = rb_next(&info->offset_index);
3673                 while (n) {
3674                         tmp = rb_entry(n, struct btrfs_free_space,
3675                                        offset_index);
3676                         if (offset + bytes < tmp->offset)
3677                                 break;
3678                         if (tmp->offset + tmp->bytes < offset) {
3679                                 n = rb_next(&tmp->offset_index);
3680                                 continue;
3681                         }
3682                         info = tmp;
3683                         goto have_info;
3684                 }
3685
3686                 ret = 0;
3687                 goto out;
3688         }
3689
3690         if (info->offset == offset) {
3691                 ret = 1;
3692                 goto out;
3693         }
3694
3695         if (offset > info->offset && offset < info->offset + info->bytes)
3696                 ret = 1;
3697 out:
3698         spin_unlock(&ctl->tree_lock);
3699         return ret;
3700 }
3701 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */