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_CACHE_SIZE * 8)
  33 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
  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, root, 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_CACHE_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_CACHE_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_CACHE_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_CACHE_SIZE;
 358         if (clear)
 359                 memset(io_ctl->cur, 0, PAGE_CACHE_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                         page_cache_release(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_CACHE_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_CACHE_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_CACHE_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_CACHE_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_CACHE_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 list_head *pos, *n;
1090         int ret;
1091
1092         /* Write out the bitmaps */
1093         list_for_each_safe(pos, n, bitmap_list) {
1094                 struct btrfs_free_space *entry =
1095                         list_entry(pos, struct btrfs_free_space, list);
1096
1097                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1098                 if (ret)
1099                         return -ENOSPC;
1100                 list_del_init(&entry->list);
1101         }
1102
1103         return 0;
1104 }
1105
1106 static int flush_dirty_cache(struct inode *inode)
1107 {
1108         int ret;
1109
1110         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1111         if (ret)
1112                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1113                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1114                                  GFP_NOFS);
1115
1116         return ret;
1117 }
1118
1119 static void noinline_for_stack
1120 cleanup_bitmap_list(struct list_head *bitmap_list)
1121 {
1122         struct list_head *pos, *n;
1123
1124         list_for_each_safe(pos, n, bitmap_list) {
1125                 struct btrfs_free_space *entry =
1126                         list_entry(pos, struct btrfs_free_space, list);
1127                 list_del_init(&entry->list);
1128         }
1129 }
1130
1131 static void noinline_for_stack
1132 cleanup_write_cache_enospc(struct inode *inode,
1133                            struct btrfs_io_ctl *io_ctl,
1134                            struct extent_state **cached_state,
1135                            struct list_head *bitmap_list)
1136 {
1137         io_ctl_drop_pages(io_ctl);
1138         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1139                              i_size_read(inode) - 1, cached_state,
1140                              GFP_NOFS);
1141 }
1142
1143 int btrfs_wait_cache_io(struct btrfs_root *root,
1144                         struct btrfs_trans_handle *trans,
1145                         struct btrfs_block_group_cache *block_group,
1146                         struct btrfs_io_ctl *io_ctl,
1147                         struct btrfs_path *path, u64 offset)
1148 {
1149         int ret;
1150         struct inode *inode = io_ctl->inode;
1151
1152         if (!inode)
1153                 return 0;
1154
1155         if (block_group)
1156                 root = root->fs_info->tree_root;
1157
1158         /* Flush the dirty pages in the cache file. */
1159         ret = flush_dirty_cache(inode);
1160         if (ret)
1161                 goto out;
1162
1163         /* Update the cache item to tell everyone this cache file is valid. */
1164         ret = update_cache_item(trans, root, inode, path, offset,
1165                                 io_ctl->entries, io_ctl->bitmaps);
1166 out:
1167         io_ctl_free(io_ctl);
1168         if (ret) {
1169                 invalidate_inode_pages2(inode->i_mapping);
1170                 BTRFS_I(inode)->generation = 0;
1171                 if (block_group) {
1172 #ifdef DEBUG
1173                         btrfs_err(root->fs_info,
1174                                 "failed to write free space cache for block group %llu",
1175                                 block_group->key.objectid);
1176 #endif
1177                 }
1178         }
1179         btrfs_update_inode(trans, root, inode);
1180
1181         if (block_group) {
1182                 /* the dirty list is protected by the dirty_bgs_lock */
1183                 spin_lock(&trans->transaction->dirty_bgs_lock);
1184
1185                 /* the disk_cache_state is protected by the block group lock */
1186                 spin_lock(&block_group->lock);
1187
1188                 /*
1189                  * only mark this as written if we didn't get put back on
1190                  * the dirty list while waiting for IO.   Otherwise our
1191                  * cache state won't be right, and we won't get written again
1192                  */
1193                 if (!ret && list_empty(&block_group->dirty_list))
1194                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1195                 else if (ret)
1196                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1197
1198                 spin_unlock(&block_group->lock);
1199                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1200                 io_ctl->inode = NULL;
1201                 iput(inode);
1202         }
1203
1204         return ret;
1205
1206 }
1207
1208 /**
1209  * __btrfs_write_out_cache - write out cached info to an inode
1210  * @root - the root the inode belongs to
1211  * @ctl - the free space cache we are going to write out
1212  * @block_group - the block_group for this cache if it belongs to a block_group
1213  * @trans - the trans handle
1214  * @path - the path to use
1215  * @offset - the offset for the key we'll insert
1216  *
1217  * This function writes out a free space cache struct to disk for quick recovery
1218  * on mount.  This will return 0 if it was successful in writing the cache out,
1219  * or an errno if it was not.
1220  */
1221 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1222                                    struct btrfs_free_space_ctl *ctl,
1223                                    struct btrfs_block_group_cache *block_group,
1224                                    struct btrfs_io_ctl *io_ctl,
1225                                    struct btrfs_trans_handle *trans,
1226                                    struct btrfs_path *path, u64 offset)
1227 {
1228         struct extent_state *cached_state = NULL;
1229         LIST_HEAD(bitmap_list);
1230         int entries = 0;
1231         int bitmaps = 0;
1232         int ret;
1233         int must_iput = 0;
1234
1235         if (!i_size_read(inode))
1236                 return -EIO;
1237
1238         WARN_ON(io_ctl->pages);
1239         ret = io_ctl_init(io_ctl, inode, root, 1);
1240         if (ret)
1241                 return ret;
1242
1243         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1244                 down_write(&block_group->data_rwsem);
1245                 spin_lock(&block_group->lock);
1246                 if (block_group->delalloc_bytes) {
1247                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1248                         spin_unlock(&block_group->lock);
1249                         up_write(&block_group->data_rwsem);
1250                         BTRFS_I(inode)->generation = 0;
1251                         ret = 0;
1252                         must_iput = 1;
1253                         goto out;
1254                 }
1255                 spin_unlock(&block_group->lock);
1256         }
1257
1258         /* Lock all pages first so we can lock the extent safely. */
1259         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1260         if (ret)
1261                 goto out_unlock;
1262
1263         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1264                          0, &cached_state);
1265
1266         io_ctl_set_generation(io_ctl, trans->transid);
1267
1268         mutex_lock(&ctl->cache_writeout_mutex);
1269         /* Write out the extent entries in the free space cache */
1270         spin_lock(&ctl->tree_lock);
1271         ret = write_cache_extent_entries(io_ctl, ctl,
1272                                          block_group, &entries, &bitmaps,
1273                                          &bitmap_list);
1274         if (ret)
1275                 goto out_nospc_locked;
1276
1277         /*
1278          * Some spaces that are freed in the current transaction are pinned,
1279          * they will be added into free space cache after the transaction is
1280          * committed, we shouldn't lose them.
1281          *
1282          * If this changes while we are working we'll get added back to
1283          * the dirty list and redo it.  No locking needed
1284          */
1285         ret = write_pinned_extent_entries(root, block_group, io_ctl, &entries);
1286         if (ret)
1287                 goto out_nospc_locked;
1288
1289         /*
1290          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1291          * locked while doing it because a concurrent trim can be manipulating
1292          * or freeing the bitmap.
1293          */
1294         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1295         spin_unlock(&ctl->tree_lock);
1296         mutex_unlock(&ctl->cache_writeout_mutex);
1297         if (ret)
1298                 goto out_nospc;
1299
1300         /* Zero out the rest of the pages just to make sure */
1301         io_ctl_zero_remaining_pages(io_ctl);
1302
1303         /* Everything is written out, now we dirty the pages in the file. */
1304         ret = btrfs_dirty_pages(root, inode, io_ctl->pages, io_ctl->num_pages,
1305                                 0, i_size_read(inode), &cached_state);
1306         if (ret)
1307                 goto out_nospc;
1308
1309         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1310                 up_write(&block_group->data_rwsem);
1311         /*
1312          * Release the pages and unlock the extent, we will flush
1313          * them out later
1314          */
1315         io_ctl_drop_pages(io_ctl);
1316
1317         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1318                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1319
1320         /*
1321          * at this point the pages are under IO and we're happy,
1322          * The caller is responsible for waiting on them and updating the
1323          * the cache and the inode
1324          */
1325         io_ctl->entries = entries;
1326         io_ctl->bitmaps = bitmaps;
1327
1328         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1329         if (ret)
1330                 goto out;
1331
1332         return 0;
1333
1334 out:
1335         io_ctl->inode = NULL;
1336         io_ctl_free(io_ctl);
1337         if (ret) {
1338                 invalidate_inode_pages2(inode->i_mapping);
1339                 BTRFS_I(inode)->generation = 0;
1340         }
1341         btrfs_update_inode(trans, root, inode);
1342         if (must_iput)
1343                 iput(inode);
1344         return ret;
1345
1346 out_nospc_locked:
1347         cleanup_bitmap_list(&bitmap_list);
1348         spin_unlock(&ctl->tree_lock);
1349         mutex_unlock(&ctl->cache_writeout_mutex);
1350
1351 out_nospc:
1352         cleanup_write_cache_enospc(inode, io_ctl, &cached_state, &bitmap_list);
1353
1354 out_unlock:
1355         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1356                 up_write(&block_group->data_rwsem);
1357
1358         goto out;
1359 }
1360
1361 int btrfs_write_out_cache(struct btrfs_root *root,
1362                           struct btrfs_trans_handle *trans,
1363                           struct btrfs_block_group_cache *block_group,
1364                           struct btrfs_path *path)
1365 {
1366         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1367         struct inode *inode;
1368         int ret = 0;
1369
1370         root = root->fs_info->tree_root;
1371
1372         spin_lock(&block_group->lock);
1373         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1374                 spin_unlock(&block_group->lock);
1375                 return 0;
1376         }
1377         spin_unlock(&block_group->lock);
1378
1379         inode = lookup_free_space_inode(root, block_group, path);
1380         if (IS_ERR(inode))
1381                 return 0;
1382
1383         ret = __btrfs_write_out_cache(root, inode, ctl, block_group,
1384                                       &block_group->io_ctl, trans,
1385                                       path, block_group->key.objectid);
1386         if (ret) {
1387 #ifdef DEBUG
1388                 btrfs_err(root->fs_info,
1389                         "failed to write free space cache for block group %llu",
1390                         block_group->key.objectid);
1391 #endif
1392                 spin_lock(&block_group->lock);
1393                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1394                 spin_unlock(&block_group->lock);
1395
1396                 block_group->io_ctl.inode = NULL;
1397                 iput(inode);
1398         }
1399
1400         /*
1401          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1402          * to wait for IO and put the inode
1403          */
1404
1405         return ret;
1406 }
1407
1408 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1409                                           u64 offset)
1410 {
1411         ASSERT(offset >= bitmap_start);
1412         offset -= bitmap_start;
1413         return (unsigned long)(div_u64(offset, unit));
1414 }
1415
1416 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1417 {
1418         return (unsigned long)(div_u64(bytes, unit));
1419 }
1420
1421 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1422                                    u64 offset)
1423 {
1424         u64 bitmap_start;
1425         u32 bytes_per_bitmap;
1426
1427         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1428         bitmap_start = offset - ctl->start;
1429         bitmap_start = div_u64(bitmap_start, bytes_per_bitmap);
1430         bitmap_start *= bytes_per_bitmap;
1431         bitmap_start += ctl->start;
1432
1433         return bitmap_start;
1434 }
1435
1436 static int tree_insert_offset(struct rb_root *root, u64 offset,
1437                               struct rb_node *node, int bitmap)
1438 {
1439         struct rb_node **p = &root->rb_node;
1440         struct rb_node *parent = NULL;
1441         struct btrfs_free_space *info;
1442
1443         while (*p) {
1444                 parent = *p;
1445                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1446
1447                 if (offset < info->offset) {
1448                         p = &(*p)->rb_left;
1449                 } else if (offset > info->offset) {
1450                         p = &(*p)->rb_right;
1451                 } else {
1452                         /*
1453                          * we could have a bitmap entry and an extent entry
1454                          * share the same offset.  If this is the case, we want
1455                          * the extent entry to always be found first if we do a
1456                          * linear search through the tree, since we want to have
1457                          * the quickest allocation time, and allocating from an
1458                          * extent is faster than allocating from a bitmap.  So
1459                          * if we're inserting a bitmap and we find an entry at
1460                          * this offset, we want to go right, or after this entry
1461                          * logically.  If we are inserting an extent and we've
1462                          * found a bitmap, we want to go left, or before
1463                          * logically.
1464                          */
1465                         if (bitmap) {
1466                                 if (info->bitmap) {
1467                                         WARN_ON_ONCE(1);
1468                                         return -EEXIST;
1469                                 }
1470                                 p = &(*p)->rb_right;
1471                         } else {
1472                                 if (!info->bitmap) {
1473                                         WARN_ON_ONCE(1);
1474                                         return -EEXIST;
1475                                 }
1476                                 p = &(*p)->rb_left;
1477                         }
1478                 }
1479         }
1480
1481         rb_link_node(node, parent, p);
1482         rb_insert_color(node, root);
1483
1484         return 0;
1485 }
1486
1487 /*
1488  * searches the tree for the given offset.
1489  *
1490  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1491  * want a section that has at least bytes size and comes at or after the given
1492  * offset.
1493  */
1494 static struct btrfs_free_space *
1495 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1496                    u64 offset, int bitmap_only, int fuzzy)
1497 {
1498         struct rb_node *n = ctl->free_space_offset.rb_node;
1499         struct btrfs_free_space *entry, *prev = NULL;
1500
1501         /* find entry that is closest to the 'offset' */
1502         while (1) {
1503                 if (!n) {
1504                         entry = NULL;
1505                         break;
1506                 }
1507
1508                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1509                 prev = entry;
1510
1511                 if (offset < entry->offset)
1512                         n = n->rb_left;
1513                 else if (offset > entry->offset)
1514                         n = n->rb_right;
1515                 else
1516                         break;
1517         }
1518
1519         if (bitmap_only) {
1520                 if (!entry)
1521                         return NULL;
1522                 if (entry->bitmap)
1523                         return entry;
1524
1525                 /*
1526                  * bitmap entry and extent entry may share same offset,
1527                  * in that case, bitmap entry comes after extent entry.
1528                  */
1529                 n = rb_next(n);
1530                 if (!n)
1531                         return NULL;
1532                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1533                 if (entry->offset != offset)
1534                         return NULL;
1535
1536                 WARN_ON(!entry->bitmap);
1537                 return entry;
1538         } else if (entry) {
1539                 if (entry->bitmap) {
1540                         /*
1541                          * if previous extent entry covers the offset,
1542                          * we should return it instead of the bitmap entry
1543                          */
1544                         n = rb_prev(&entry->offset_index);
1545                         if (n) {
1546                                 prev = rb_entry(n, struct btrfs_free_space,
1547                                                 offset_index);
1548                                 if (!prev->bitmap &&
1549                                     prev->offset + prev->bytes > offset)
1550                                         entry = prev;
1551                         }
1552                 }
1553                 return entry;
1554         }
1555
1556         if (!prev)
1557                 return NULL;
1558
1559         /* find last entry before the 'offset' */
1560         entry = prev;
1561         if (entry->offset > offset) {
1562                 n = rb_prev(&entry->offset_index);
1563                 if (n) {
1564                         entry = rb_entry(n, struct btrfs_free_space,
1565                                         offset_index);
1566                         ASSERT(entry->offset <= offset);
1567                 } else {
1568                         if (fuzzy)
1569                                 return entry;
1570                         else
1571                                 return NULL;
1572                 }
1573         }
1574
1575         if (entry->bitmap) {
1576                 n = rb_prev(&entry->offset_index);
1577                 if (n) {
1578                         prev = rb_entry(n, struct btrfs_free_space,
1579                                         offset_index);
1580                         if (!prev->bitmap &&
1581                             prev->offset + prev->bytes > offset)
1582                                 return prev;
1583                 }
1584                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1585                         return entry;
1586         } else if (entry->offset + entry->bytes > offset)
1587                 return entry;
1588
1589         if (!fuzzy)
1590                 return NULL;
1591
1592         while (1) {
1593                 if (entry->bitmap) {
1594                         if (entry->offset + BITS_PER_BITMAP *
1595                             ctl->unit > offset)
1596                                 break;
1597                 } else {
1598                         if (entry->offset + entry->bytes > offset)
1599                                 break;
1600                 }
1601
1602                 n = rb_next(&entry->offset_index);
1603                 if (!n)
1604                         return NULL;
1605                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1606         }
1607         return entry;
1608 }
1609
1610 static inline void
1611 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1612                     struct btrfs_free_space *info)
1613 {
1614         rb_erase(&info->offset_index, &ctl->free_space_offset);
1615         ctl->free_extents--;
1616 }
1617
1618 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1619                               struct btrfs_free_space *info)
1620 {
1621         __unlink_free_space(ctl, info);
1622         ctl->free_space -= info->bytes;
1623 }
1624
1625 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1626                            struct btrfs_free_space *info)
1627 {
1628         int ret = 0;
1629
1630         ASSERT(info->bytes || info->bitmap);
1631         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1632                                  &info->offset_index, (info->bitmap != NULL));
1633         if (ret)
1634                 return ret;
1635
1636         ctl->free_space += info->bytes;
1637         ctl->free_extents++;
1638         return ret;
1639 }
1640
1641 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1642 {
1643         struct btrfs_block_group_cache *block_group = ctl->private;
1644         u64 max_bytes;
1645         u64 bitmap_bytes;
1646         u64 extent_bytes;
1647         u64 size = block_group->key.offset;
1648         u32 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1649         u32 max_bitmaps = div_u64(size + bytes_per_bg - 1, bytes_per_bg);
1650
1651         max_bitmaps = max_t(u32, max_bitmaps, 1);
1652
1653         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1654
1655         /*
1656          * The goal is to keep the total amount of memory used per 1gb of space
1657          * at or below 32k, so we need to adjust how much memory we allow to be
1658          * used by extent based free space tracking
1659          */
1660         if (size < 1024 * 1024 * 1024)
1661                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1662         else
1663                 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1664                         div_u64(size, 1024 * 1024 * 1024);
1665
1666         /*
1667          * we want to account for 1 more bitmap than what we have so we can make
1668          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1669          * we add more bitmaps.
1670          */
1671         bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1672
1673         if (bitmap_bytes >= max_bytes) {
1674                 ctl->extents_thresh = 0;
1675                 return;
1676         }
1677
1678         /*
1679          * we want the extent entry threshold to always be at most 1/2 the max
1680          * bytes we can have, or whatever is less than that.
1681          */
1682         extent_bytes = max_bytes - bitmap_bytes;
1683         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1684
1685         ctl->extents_thresh =
1686                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1687 }
1688
1689 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1690                                        struct btrfs_free_space *info,
1691                                        u64 offset, u64 bytes)
1692 {
1693         unsigned long start, count;
1694
1695         start = offset_to_bit(info->offset, ctl->unit, offset);
1696         count = bytes_to_bits(bytes, ctl->unit);
1697         ASSERT(start + count <= BITS_PER_BITMAP);
1698
1699         bitmap_clear(info->bitmap, start, count);
1700
1701         info->bytes -= bytes;
1702         if (info->max_extent_size > ctl->unit)
1703                 info->max_extent_size = 0;
1704 }
1705
1706 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1707                               struct btrfs_free_space *info, u64 offset,
1708                               u64 bytes)
1709 {
1710         __bitmap_clear_bits(ctl, info, offset, bytes);
1711         ctl->free_space -= bytes;
1712 }
1713
1714 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1715                             struct btrfs_free_space *info, u64 offset,
1716                             u64 bytes)
1717 {
1718         unsigned long start, count;
1719
1720         start = offset_to_bit(info->offset, ctl->unit, offset);
1721         count = bytes_to_bits(bytes, ctl->unit);
1722         ASSERT(start + count <= BITS_PER_BITMAP);
1723
1724         bitmap_set(info->bitmap, start, count);
1725
1726         info->bytes += bytes;
1727         ctl->free_space += bytes;
1728 }
1729
1730 /*
1731  * If we can not find suitable extent, we will use bytes to record
1732  * the size of the max extent.
1733  */
1734 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1735                          struct btrfs_free_space *bitmap_info, u64 *offset,
1736                          u64 *bytes, bool for_alloc)
1737 {
1738         unsigned long found_bits = 0;
1739         unsigned long max_bits = 0;
1740         unsigned long bits, i;
1741         unsigned long next_zero;
1742         unsigned long extent_bits;
1743
1744         /*
1745          * Skip searching the bitmap if we don't have a contiguous section that
1746          * is large enough for this allocation.
1747          */
1748         if (for_alloc &&
1749             bitmap_info->max_extent_size &&
1750             bitmap_info->max_extent_size < *bytes) {
1751                 *bytes = bitmap_info->max_extent_size;
1752                 return -1;
1753         }
1754
1755         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1756                           max_t(u64, *offset, bitmap_info->offset));
1757         bits = bytes_to_bits(*bytes, ctl->unit);
1758
1759         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1760                 if (for_alloc && bits == 1) {
1761                         found_bits = 1;
1762                         break;
1763                 }
1764                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1765                                                BITS_PER_BITMAP, i);
1766                 extent_bits = next_zero - i;
1767                 if (extent_bits >= bits) {
1768                         found_bits = extent_bits;
1769                         break;
1770                 } else if (extent_bits > max_bits) {
1771                         max_bits = extent_bits;
1772                 }
1773                 i = next_zero;
1774         }
1775
1776         if (found_bits) {
1777                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1778                 *bytes = (u64)(found_bits) * ctl->unit;
1779                 return 0;
1780         }
1781
1782         *bytes = (u64)(max_bits) * ctl->unit;
1783         bitmap_info->max_extent_size = *bytes;
1784         return -1;
1785 }
1786
1787 static inline u64 get_max_extent_size(struct btrfs_free_space *entry)
1788 {
1789         if (entry->bitmap)
1790                 return entry->max_extent_size;
1791         return entry->bytes;
1792 }
1793
1794 /* Cache the size of the max extent in bytes */
1795 static struct btrfs_free_space *
1796 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1797                 unsigned long align, u64 *max_extent_size)
1798 {
1799         struct btrfs_free_space *entry;
1800         struct rb_node *node;
1801         u64 tmp;
1802         u64 align_off;
1803         int ret;
1804
1805         if (!ctl->free_space_offset.rb_node)
1806                 goto out;
1807
1808         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1809         if (!entry)
1810                 goto out;
1811
1812         for (node = &entry->offset_index; node; node = rb_next(node)) {
1813                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1814                 if (entry->bytes < *bytes) {
1815                         *max_extent_size = max(get_max_extent_size(entry),
1816                                                *max_extent_size);
1817                         continue;
1818                 }
1819
1820                 /* make sure the space returned is big enough
1821                  * to match our requested alignment
1822                  */
1823                 if (*bytes >= align) {
1824                         tmp = entry->offset - ctl->start + align - 1;
1825                         tmp = div64_u64(tmp, align);
1826                         tmp = tmp * align + ctl->start;
1827                         align_off = tmp - entry->offset;
1828                 } else {
1829                         align_off = 0;
1830                         tmp = entry->offset;
1831                 }
1832
1833                 if (entry->bytes < *bytes + align_off) {
1834                         *max_extent_size = max(get_max_extent_size(entry),
1835                                                *max_extent_size);
1836                         continue;
1837                 }
1838
1839                 if (entry->bitmap) {
1840                         u64 size = *bytes;
1841
1842                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1843                         if (!ret) {
1844                                 *offset = tmp;
1845                                 *bytes = size;
1846                                 return entry;
1847                         } else {
1848                                 *max_extent_size =
1849                                         max(get_max_extent_size(entry),
1850                                             *max_extent_size);
1851                         }
1852                         continue;
1853                 }
1854
1855                 *offset = tmp;
1856                 *bytes = entry->bytes - align_off;
1857                 return entry;
1858         }
1859 out:
1860         return NULL;
1861 }
1862
1863 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1864                            struct btrfs_free_space *info, u64 offset)
1865 {
1866         info->offset = offset_to_bitmap(ctl, offset);
1867         info->bytes = 0;
1868         INIT_LIST_HEAD(&info->list);
1869         link_free_space(ctl, info);
1870         ctl->total_bitmaps++;
1871
1872         ctl->op->recalc_thresholds(ctl);
1873 }
1874
1875 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1876                         struct btrfs_free_space *bitmap_info)
1877 {
1878         unlink_free_space(ctl, bitmap_info);
1879         kfree(bitmap_info->bitmap);
1880         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1881         ctl->total_bitmaps--;
1882         ctl->op->recalc_thresholds(ctl);
1883 }
1884
1885 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1886                               struct btrfs_free_space *bitmap_info,
1887                               u64 *offset, u64 *bytes)
1888 {
1889         u64 end;
1890         u64 search_start, search_bytes;
1891         int ret;
1892
1893 again:
1894         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1895
1896         /*
1897          * We need to search for bits in this bitmap.  We could only cover some
1898          * of the extent in this bitmap thanks to how we add space, so we need
1899          * to search for as much as it as we can and clear that amount, and then
1900          * go searching for the next bit.
1901          */
1902         search_start = *offset;
1903         search_bytes = ctl->unit;
1904         search_bytes = min(search_bytes, end - search_start + 1);
1905         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1906                             false);
1907         if (ret < 0 || search_start != *offset)
1908                 return -EINVAL;
1909
1910         /* We may have found more bits than what we need */
1911         search_bytes = min(search_bytes, *bytes);
1912
1913         /* Cannot clear past the end of the bitmap */
1914         search_bytes = min(search_bytes, end - search_start + 1);
1915
1916         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1917         *offset += search_bytes;
1918         *bytes -= search_bytes;
1919
1920         if (*bytes) {
1921                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1922                 if (!bitmap_info->bytes)
1923                         free_bitmap(ctl, bitmap_info);
1924
1925                 /*
1926                  * no entry after this bitmap, but we still have bytes to
1927                  * remove, so something has gone wrong.
1928                  */
1929                 if (!next)
1930                         return -EINVAL;
1931
1932                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1933                                        offset_index);
1934
1935                 /*
1936                  * if the next entry isn't a bitmap we need to return to let the
1937                  * extent stuff do its work.
1938                  */
1939                 if (!bitmap_info->bitmap)
1940                         return -EAGAIN;
1941
1942                 /*
1943                  * Ok the next item is a bitmap, but it may not actually hold
1944                  * the information for the rest of this free space stuff, so
1945                  * look for it, and if we don't find it return so we can try
1946                  * everything over again.
1947                  */
1948                 search_start = *offset;
1949                 search_bytes = ctl->unit;
1950                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1951                                     &search_bytes, false);
1952                 if (ret < 0 || search_start != *offset)
1953                         return -EAGAIN;
1954
1955                 goto again;
1956         } else if (!bitmap_info->bytes)
1957                 free_bitmap(ctl, bitmap_info);
1958
1959         return 0;
1960 }
1961
1962 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1963                                struct btrfs_free_space *info, u64 offset,
1964                                u64 bytes)
1965 {
1966         u64 bytes_to_set = 0;
1967         u64 end;
1968
1969         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1970
1971         bytes_to_set = min(end - offset, bytes);
1972
1973         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1974
1975         /*
1976          * We set some bytes, we have no idea what the max extent size is
1977          * anymore.
1978          */
1979         info->max_extent_size = 0;
1980
1981         return bytes_to_set;
1982
1983 }
1984
1985 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1986                       struct btrfs_free_space *info)
1987 {
1988         struct btrfs_block_group_cache *block_group = ctl->private;
1989         bool forced = false;
1990
1991 #ifdef CONFIG_BTRFS_DEBUG
1992         if (btrfs_should_fragment_free_space(block_group->fs_info->extent_root,
1993                                              block_group))
1994                 forced = true;
1995 #endif
1996
1997         /*
1998          * If we are below the extents threshold then we can add this as an
1999          * extent, and don't have to deal with the bitmap
2000          */
2001         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2002                 /*
2003                  * If this block group has some small extents we don't want to
2004                  * use up all of our free slots in the cache with them, we want
2005                  * to reserve them to larger extents, however if we have plent
2006                  * of cache left then go ahead an dadd them, no sense in adding
2007                  * the overhead of a bitmap if we don't have to.
2008                  */
2009                 if (info->bytes <= block_group->sectorsize * 4) {
2010                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
2011                                 return false;
2012                 } else {
2013                         return false;
2014                 }
2015         }
2016
2017         /*
2018          * The original block groups from mkfs can be really small, like 8
2019          * megabytes, so don't bother with a bitmap for those entries.  However
2020          * some block groups can be smaller than what a bitmap would cover but
2021          * are still large enough that they could overflow the 32k memory limit,
2022          * so allow those block groups to still be allowed to have a bitmap
2023          * entry.
2024          */
2025         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
2026                 return false;
2027
2028         return true;
2029 }
2030
2031 static struct btrfs_free_space_op free_space_op = {
2032         .recalc_thresholds      = recalculate_thresholds,
2033         .use_bitmap             = use_bitmap,
2034 };
2035
2036 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2037                               struct btrfs_free_space *info)
2038 {
2039         struct btrfs_free_space *bitmap_info;
2040         struct btrfs_block_group_cache *block_group = NULL;
2041         int added = 0;
2042         u64 bytes, offset, bytes_added;
2043         int ret;
2044
2045         bytes = info->bytes;
2046         offset = info->offset;
2047
2048         if (!ctl->op->use_bitmap(ctl, info))
2049                 return 0;
2050
2051         if (ctl->op == &free_space_op)
2052                 block_group = ctl->private;
2053 again:
2054         /*
2055          * Since we link bitmaps right into the cluster we need to see if we
2056          * have a cluster here, and if so and it has our bitmap we need to add
2057          * the free space to that bitmap.
2058          */
2059         if (block_group && !list_empty(&block_group->cluster_list)) {
2060                 struct btrfs_free_cluster *cluster;
2061                 struct rb_node *node;
2062                 struct btrfs_free_space *entry;
2063
2064                 cluster = list_entry(block_group->cluster_list.next,
2065                                      struct btrfs_free_cluster,
2066                                      block_group_list);
2067                 spin_lock(&cluster->lock);
2068                 node = rb_first(&cluster->root);
2069                 if (!node) {
2070                         spin_unlock(&cluster->lock);
2071                         goto no_cluster_bitmap;
2072                 }
2073
2074                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2075                 if (!entry->bitmap) {
2076                         spin_unlock(&cluster->lock);
2077                         goto no_cluster_bitmap;
2078                 }
2079
2080                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2081                         bytes_added = add_bytes_to_bitmap(ctl, entry,
2082                                                           offset, bytes);
2083                         bytes -= bytes_added;
2084                         offset += bytes_added;
2085                 }
2086                 spin_unlock(&cluster->lock);
2087                 if (!bytes) {
2088                         ret = 1;
2089                         goto out;
2090                 }
2091         }
2092
2093 no_cluster_bitmap:
2094         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2095                                          1, 0);
2096         if (!bitmap_info) {
2097                 ASSERT(added == 0);
2098                 goto new_bitmap;
2099         }
2100
2101         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2102         bytes -= bytes_added;
2103         offset += bytes_added;
2104         added = 0;
2105
2106         if (!bytes) {
2107                 ret = 1;
2108                 goto out;
2109         } else
2110                 goto again;
2111
2112 new_bitmap:
2113         if (info && info->bitmap) {
2114                 add_new_bitmap(ctl, info, offset);
2115                 added = 1;
2116                 info = NULL;
2117                 goto again;
2118         } else {
2119                 spin_unlock(&ctl->tree_lock);
2120
2121                 /* no pre-allocated info, allocate a new one */
2122                 if (!info) {
2123                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2124                                                  GFP_NOFS);
2125                         if (!info) {
2126                                 spin_lock(&ctl->tree_lock);
2127                                 ret = -ENOMEM;
2128                                 goto out;
2129                         }
2130                 }
2131
2132                 /* allocate the bitmap */
2133                 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
2134                 spin_lock(&ctl->tree_lock);
2135                 if (!info->bitmap) {
2136                         ret = -ENOMEM;
2137                         goto out;
2138                 }
2139                 goto again;
2140         }
2141
2142 out:
2143         if (info) {
2144                 if (info->bitmap)
2145                         kfree(info->bitmap);
2146                 kmem_cache_free(btrfs_free_space_cachep, info);
2147         }
2148
2149         return ret;
2150 }
2151
2152 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2153                           struct btrfs_free_space *info, bool update_stat)
2154 {
2155         struct btrfs_free_space *left_info;
2156         struct btrfs_free_space *right_info;
2157         bool merged = false;
2158         u64 offset = info->offset;
2159         u64 bytes = info->bytes;
2160
2161         /*
2162          * first we want to see if there is free space adjacent to the range we
2163          * are adding, if there is remove that struct and add a new one to
2164          * cover the entire range
2165          */
2166         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2167         if (right_info && rb_prev(&right_info->offset_index))
2168                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2169                                      struct btrfs_free_space, offset_index);
2170         else
2171                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2172
2173         if (right_info && !right_info->bitmap) {
2174                 if (update_stat)
2175                         unlink_free_space(ctl, right_info);
2176                 else
2177                         __unlink_free_space(ctl, right_info);
2178                 info->bytes += right_info->bytes;
2179                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2180                 merged = true;
2181         }
2182
2183         if (left_info && !left_info->bitmap &&
2184             left_info->offset + left_info->bytes == offset) {
2185                 if (update_stat)
2186                         unlink_free_space(ctl, left_info);
2187                 else
2188                         __unlink_free_space(ctl, left_info);
2189                 info->offset = left_info->offset;
2190                 info->bytes += left_info->bytes;
2191                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2192                 merged = true;
2193         }
2194
2195         return merged;
2196 }
2197
2198 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2199                                      struct btrfs_free_space *info,
2200                                      bool update_stat)
2201 {
2202         struct btrfs_free_space *bitmap;
2203         unsigned long i;
2204         unsigned long j;
2205         const u64 end = info->offset + info->bytes;
2206         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2207         u64 bytes;
2208
2209         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2210         if (!bitmap)
2211                 return false;
2212
2213         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2214         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2215         if (j == i)
2216                 return false;
2217         bytes = (j - i) * ctl->unit;
2218         info->bytes += bytes;
2219
2220         if (update_stat)
2221                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2222         else
2223                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2224
2225         if (!bitmap->bytes)
2226                 free_bitmap(ctl, bitmap);
2227
2228         return true;
2229 }
2230
2231 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2232                                        struct btrfs_free_space *info,
2233                                        bool update_stat)
2234 {
2235         struct btrfs_free_space *bitmap;
2236         u64 bitmap_offset;
2237         unsigned long i;
2238         unsigned long j;
2239         unsigned long prev_j;
2240         u64 bytes;
2241
2242         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2243         /* If we're on a boundary, try the previous logical bitmap. */
2244         if (bitmap_offset == info->offset) {
2245                 if (info->offset == 0)
2246                         return false;
2247                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2248         }
2249
2250         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2251         if (!bitmap)
2252                 return false;
2253
2254         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2255         j = 0;
2256         prev_j = (unsigned long)-1;
2257         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2258                 if (j > i)
2259                         break;
2260                 prev_j = j;
2261         }
2262         if (prev_j == i)
2263                 return false;
2264
2265         if (prev_j == (unsigned long)-1)
2266                 bytes = (i + 1) * ctl->unit;
2267         else
2268                 bytes = (i - prev_j) * ctl->unit;
2269
2270         info->offset -= bytes;
2271         info->bytes += bytes;
2272
2273         if (update_stat)
2274                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2275         else
2276                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2277
2278         if (!bitmap->bytes)
2279                 free_bitmap(ctl, bitmap);
2280
2281         return true;
2282 }
2283
2284 /*
2285  * We prefer always to allocate from extent entries, both for clustered and
2286  * non-clustered allocation requests. So when attempting to add a new extent
2287  * entry, try to see if there's adjacent free space in bitmap entries, and if
2288  * there is, migrate that space from the bitmaps to the extent.
2289  * Like this we get better chances of satisfying space allocation requests
2290  * because we attempt to satisfy them based on a single cache entry, and never
2291  * on 2 or more entries - even if the entries represent a contiguous free space
2292  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2293  * ends).
2294  */
2295 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2296                               struct btrfs_free_space *info,
2297                               bool update_stat)
2298 {
2299         /*
2300          * Only work with disconnected entries, as we can change their offset,
2301          * and must be extent entries.
2302          */
2303         ASSERT(!info->bitmap);
2304         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2305
2306         if (ctl->total_bitmaps > 0) {
2307                 bool stole_end;
2308                 bool stole_front = false;
2309
2310                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2311                 if (ctl->total_bitmaps > 0)
2312                         stole_front = steal_from_bitmap_to_front(ctl, info,
2313                                                                  update_stat);
2314
2315                 if (stole_end || stole_front)
2316                         try_merge_free_space(ctl, info, update_stat);
2317         }
2318 }
2319
2320 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
2321                            u64 offset, u64 bytes)
2322 {
2323         struct btrfs_free_space *info;
2324         int ret = 0;
2325
2326         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2327         if (!info)
2328                 return -ENOMEM;
2329
2330         info->offset = offset;
2331         info->bytes = bytes;
2332         RB_CLEAR_NODE(&info->offset_index);
2333
2334         spin_lock(&ctl->tree_lock);
2335
2336         if (try_merge_free_space(ctl, info, true))
2337                 goto link;
2338
2339         /*
2340          * There was no extent directly to the left or right of this new
2341          * extent then we know we're going to have to allocate a new extent, so
2342          * before we do that see if we need to drop this into a bitmap
2343          */
2344         ret = insert_into_bitmap(ctl, info);
2345         if (ret < 0) {
2346                 goto out;
2347         } else if (ret) {
2348                 ret = 0;
2349                 goto out;
2350         }
2351 link:
2352         /*
2353          * Only steal free space from adjacent bitmaps if we're sure we're not
2354          * going to add the new free space to existing bitmap entries - because
2355          * that would mean unnecessary work that would be reverted. Therefore
2356          * attempt to steal space from bitmaps if we're adding an extent entry.
2357          */
2358         steal_from_bitmap(ctl, info, true);
2359
2360         ret = link_free_space(ctl, info);
2361         if (ret)
2362                 kmem_cache_free(btrfs_free_space_cachep, info);
2363 out:
2364         spin_unlock(&ctl->tree_lock);
2365
2366         if (ret) {
2367                 printk(KERN_CRIT "BTRFS: unable to add free space :%d\n", ret);
2368                 ASSERT(ret != -EEXIST);
2369         }
2370
2371         return ret;
2372 }
2373
2374 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2375                             u64 offset, u64 bytes)
2376 {
2377         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2378         struct btrfs_free_space *info;
2379         int ret;
2380         bool re_search = false;
2381
2382         spin_lock(&ctl->tree_lock);
2383
2384 again:
2385         ret = 0;
2386         if (!bytes)
2387                 goto out_lock;
2388
2389         info = tree_search_offset(ctl, offset, 0, 0);
2390         if (!info) {
2391                 /*
2392                  * oops didn't find an extent that matched the space we wanted
2393                  * to remove, look for a bitmap instead
2394                  */
2395                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2396                                           1, 0);
2397                 if (!info) {
2398                         /*
2399                          * If we found a partial bit of our free space in a
2400                          * bitmap but then couldn't find the other part this may
2401                          * be a problem, so WARN about it.
2402                          */
2403                         WARN_ON(re_search);
2404                         goto out_lock;
2405                 }
2406         }
2407
2408         re_search = false;
2409         if (!info->bitmap) {
2410                 unlink_free_space(ctl, info);
2411                 if (offset == info->offset) {
2412                         u64 to_free = min(bytes, info->bytes);
2413
2414                         info->bytes -= to_free;
2415                         info->offset += to_free;
2416                         if (info->bytes) {
2417                                 ret = link_free_space(ctl, info);
2418                                 WARN_ON(ret);
2419                         } else {
2420                                 kmem_cache_free(btrfs_free_space_cachep, info);
2421                         }
2422
2423                         offset += to_free;
2424                         bytes -= to_free;
2425                         goto again;
2426                 } else {
2427                         u64 old_end = info->bytes + info->offset;
2428
2429                         info->bytes = offset - info->offset;
2430                         ret = link_free_space(ctl, info);
2431                         WARN_ON(ret);
2432                         if (ret)
2433                                 goto out_lock;
2434
2435                         /* Not enough bytes in this entry to satisfy us */
2436                         if (old_end < offset + bytes) {
2437                                 bytes -= old_end - offset;
2438                                 offset = old_end;
2439                                 goto again;
2440                         } else if (old_end == offset + bytes) {
2441                                 /* all done */
2442                                 goto out_lock;
2443                         }
2444                         spin_unlock(&ctl->tree_lock);
2445
2446                         ret = btrfs_add_free_space(block_group, offset + bytes,
2447                                                    old_end - (offset + bytes));
2448                         WARN_ON(ret);
2449                         goto out;
2450                 }
2451         }
2452
2453         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2454         if (ret == -EAGAIN) {
2455                 re_search = true;
2456                 goto again;
2457         }
2458 out_lock:
2459         spin_unlock(&ctl->tree_lock);
2460 out:
2461         return ret;
2462 }
2463
2464 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2465                            u64 bytes)
2466 {
2467         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2468         struct btrfs_free_space *info;
2469         struct rb_node *n;
2470         int count = 0;
2471
2472         spin_lock(&ctl->tree_lock);
2473         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2474                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2475                 if (info->bytes >= bytes && !block_group->ro)
2476                         count++;
2477                 btrfs_crit(block_group->fs_info,
2478                            "entry offset %llu, bytes %llu, bitmap %s",
2479                            info->offset, info->bytes,
2480                        (info->bitmap) ? "yes" : "no");
2481         }
2482         spin_unlock(&ctl->tree_lock);
2483         btrfs_info(block_group->fs_info, "block group has cluster?: %s",
2484                list_empty(&block_group->cluster_list) ? "no" : "yes");
2485         btrfs_info(block_group->fs_info,
2486                    "%d blocks of free space at or bigger than bytes is", count);
2487 }
2488
2489 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2490 {
2491         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2492
2493         spin_lock_init(&ctl->tree_lock);
2494         ctl->unit = block_group->sectorsize;
2495         ctl->start = block_group->key.objectid;
2496         ctl->private = block_group;
2497         ctl->op = &free_space_op;
2498         INIT_LIST_HEAD(&ctl->trimming_ranges);
2499         mutex_init(&ctl->cache_writeout_mutex);
2500
2501         /*
2502          * we only want to have 32k of ram per block group for keeping
2503          * track of free space, and if we pass 1/2 of that we want to
2504          * start converting things over to using bitmaps
2505          */
2506         ctl->extents_thresh = ((1024 * 32) / 2) /
2507                                 sizeof(struct btrfs_free_space);
2508 }
2509
2510 /*
2511  * for a given cluster, put all of its extents back into the free
2512  * space cache.  If the block group passed doesn't match the block group
2513  * pointed to by the cluster, someone else raced in and freed the
2514  * cluster already.  In that case, we just return without changing anything
2515  */
2516 static int
2517 __btrfs_return_cluster_to_free_space(
2518                              struct btrfs_block_group_cache *block_group,
2519                              struct btrfs_free_cluster *cluster)
2520 {
2521         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2522         struct btrfs_free_space *entry;
2523         struct rb_node *node;
2524
2525         spin_lock(&cluster->lock);
2526         if (cluster->block_group != block_group)
2527                 goto out;
2528
2529         cluster->block_group = NULL;
2530         cluster->window_start = 0;
2531         list_del_init(&cluster->block_group_list);
2532
2533         node = rb_first(&cluster->root);
2534         while (node) {
2535                 bool bitmap;
2536
2537                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2538                 node = rb_next(&entry->offset_index);
2539                 rb_erase(&entry->offset_index, &cluster->root);
2540                 RB_CLEAR_NODE(&entry->offset_index);
2541
2542                 bitmap = (entry->bitmap != NULL);
2543                 if (!bitmap) {
2544                         try_merge_free_space(ctl, entry, false);
2545                         steal_from_bitmap(ctl, entry, false);
2546                 }
2547                 tree_insert_offset(&ctl->free_space_offset,
2548                                    entry->offset, &entry->offset_index, bitmap);
2549         }
2550         cluster->root = RB_ROOT;
2551
2552 out:
2553         spin_unlock(&cluster->lock);
2554         btrfs_put_block_group(block_group);
2555         return 0;
2556 }
2557
2558 static void __btrfs_remove_free_space_cache_locked(
2559                                 struct btrfs_free_space_ctl *ctl)
2560 {
2561         struct btrfs_free_space *info;
2562         struct rb_node *node;
2563
2564         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2565                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2566                 if (!info->bitmap) {
2567                         unlink_free_space(ctl, info);
2568                         kmem_cache_free(btrfs_free_space_cachep, info);
2569                 } else {
2570                         free_bitmap(ctl, info);
2571                 }
2572
2573                 cond_resched_lock(&ctl->tree_lock);
2574         }
2575 }
2576
2577 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2578 {
2579         spin_lock(&ctl->tree_lock);
2580         __btrfs_remove_free_space_cache_locked(ctl);
2581         spin_unlock(&ctl->tree_lock);
2582 }
2583
2584 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2585 {
2586         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2587         struct btrfs_free_cluster *cluster;
2588         struct list_head *head;
2589
2590         spin_lock(&ctl->tree_lock);
2591         while ((head = block_group->cluster_list.next) !=
2592                &block_group->cluster_list) {
2593                 cluster = list_entry(head, struct btrfs_free_cluster,
2594                                      block_group_list);
2595
2596                 WARN_ON(cluster->block_group != block_group);
2597                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2598
2599                 cond_resched_lock(&ctl->tree_lock);
2600         }
2601         __btrfs_remove_free_space_cache_locked(ctl);
2602         spin_unlock(&ctl->tree_lock);
2603
2604 }
2605
2606 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2607                                u64 offset, u64 bytes, u64 empty_size,
2608                                u64 *max_extent_size)
2609 {
2610         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2611         struct btrfs_free_space *entry = NULL;
2612         u64 bytes_search = bytes + empty_size;
2613         u64 ret = 0;
2614         u64 align_gap = 0;
2615         u64 align_gap_len = 0;
2616
2617         spin_lock(&ctl->tree_lock);
2618         entry = find_free_space(ctl, &offset, &bytes_search,
2619                                 block_group->full_stripe_len, max_extent_size);
2620         if (!entry)
2621                 goto out;
2622
2623         ret = offset;
2624         if (entry->bitmap) {
2625                 bitmap_clear_bits(ctl, entry, offset, bytes);
2626                 if (!entry->bytes)
2627                         free_bitmap(ctl, entry);
2628         } else {
2629                 unlink_free_space(ctl, entry);
2630                 align_gap_len = offset - entry->offset;
2631                 align_gap = entry->offset;
2632
2633                 entry->offset = offset + bytes;
2634                 WARN_ON(entry->bytes < bytes + align_gap_len);
2635
2636                 entry->bytes -= bytes + align_gap_len;
2637                 if (!entry->bytes)
2638                         kmem_cache_free(btrfs_free_space_cachep, entry);
2639                 else
2640                         link_free_space(ctl, entry);
2641         }
2642 out:
2643         spin_unlock(&ctl->tree_lock);
2644
2645         if (align_gap_len)
2646                 __btrfs_add_free_space(ctl, align_gap, align_gap_len);
2647         return ret;
2648 }
2649
2650 /*
2651  * given a cluster, put all of its extents back into the free space
2652  * cache.  If a block group is passed, this function will only free
2653  * a cluster that belongs to the passed block group.
2654  *
2655  * Otherwise, it'll get a reference on the block group pointed to by the
2656  * cluster and remove the cluster from it.
2657  */
2658 int btrfs_return_cluster_to_free_space(
2659                                struct btrfs_block_group_cache *block_group,
2660                                struct btrfs_free_cluster *cluster)
2661 {
2662         struct btrfs_free_space_ctl *ctl;
2663         int ret;
2664
2665         /* first, get a safe pointer to the block group */
2666         spin_lock(&cluster->lock);
2667         if (!block_group) {
2668                 block_group = cluster->block_group;
2669                 if (!block_group) {
2670                         spin_unlock(&cluster->lock);
2671                         return 0;
2672                 }
2673         } else if (cluster->block_group != block_group) {
2674                 /* someone else has already freed it don't redo their work */
2675                 spin_unlock(&cluster->lock);
2676                 return 0;
2677         }
2678         atomic_inc(&block_group->count);
2679         spin_unlock(&cluster->lock);
2680
2681         ctl = block_group->free_space_ctl;
2682
2683         /* now return any extents the cluster had on it */
2684         spin_lock(&ctl->tree_lock);
2685         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2686         spin_unlock(&ctl->tree_lock);
2687
2688         /* finally drop our ref */
2689         btrfs_put_block_group(block_group);
2690         return ret;
2691 }
2692
2693 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2694                                    struct btrfs_free_cluster *cluster,
2695                                    struct btrfs_free_space *entry,
2696                                    u64 bytes, u64 min_start,
2697                                    u64 *max_extent_size)
2698 {
2699         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2700         int err;
2701         u64 search_start = cluster->window_start;
2702         u64 search_bytes = bytes;
2703         u64 ret = 0;
2704
2705         search_start = min_start;
2706         search_bytes = bytes;
2707
2708         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2709         if (err) {
2710                 *max_extent_size = max(get_max_extent_size(entry),
2711                                        *max_extent_size);
2712                 return 0;
2713         }
2714
2715         ret = search_start;
2716         __bitmap_clear_bits(ctl, entry, ret, bytes);
2717
2718         return ret;
2719 }
2720
2721 /*
2722  * given a cluster, try to allocate 'bytes' from it, returns 0
2723  * if it couldn't find anything suitably large, or a logical disk offset
2724  * if things worked out
2725  */
2726 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2727                              struct btrfs_free_cluster *cluster, u64 bytes,
2728                              u64 min_start, u64 *max_extent_size)
2729 {
2730         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2731         struct btrfs_free_space *entry = NULL;
2732         struct rb_node *node;
2733         u64 ret = 0;
2734
2735         spin_lock(&cluster->lock);
2736         if (bytes > cluster->max_size)
2737                 goto out;
2738
2739         if (cluster->block_group != block_group)
2740                 goto out;
2741
2742         node = rb_first(&cluster->root);
2743         if (!node)
2744                 goto out;
2745
2746         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2747         while (1) {
2748                 if (entry->bytes < bytes)
2749                         *max_extent_size = max(get_max_extent_size(entry),
2750                                                *max_extent_size);
2751
2752                 if (entry->bytes < bytes ||
2753                     (!entry->bitmap && entry->offset < min_start)) {
2754                         node = rb_next(&entry->offset_index);
2755                         if (!node)
2756                                 break;
2757                         entry = rb_entry(node, struct btrfs_free_space,
2758                                          offset_index);
2759                         continue;
2760                 }
2761
2762                 if (entry->bitmap) {
2763                         ret = btrfs_alloc_from_bitmap(block_group,
2764                                                       cluster, entry, bytes,
2765                                                       cluster->window_start,
2766                                                       max_extent_size);
2767                         if (ret == 0) {
2768                                 node = rb_next(&entry->offset_index);
2769                                 if (!node)
2770                                         break;
2771                                 entry = rb_entry(node, struct btrfs_free_space,
2772                                                  offset_index);
2773                                 continue;
2774                         }
2775                         cluster->window_start += bytes;
2776                 } else {
2777                         ret = entry->offset;
2778
2779                         entry->offset += bytes;
2780                         entry->bytes -= bytes;
2781                 }
2782
2783                 if (entry->bytes == 0)
2784                         rb_erase(&entry->offset_index, &cluster->root);
2785                 break;
2786         }
2787 out:
2788         spin_unlock(&cluster->lock);
2789
2790         if (!ret)
2791                 return 0;
2792
2793         spin_lock(&ctl->tree_lock);
2794
2795         ctl->free_space -= bytes;
2796         if (entry->bytes == 0) {
2797                 ctl->free_extents--;
2798                 if (entry->bitmap) {
2799                         kfree(entry->bitmap);
2800                         ctl->total_bitmaps--;
2801                         ctl->op->recalc_thresholds(ctl);
2802                 }
2803                 kmem_cache_free(btrfs_free_space_cachep, entry);
2804         }
2805
2806         spin_unlock(&ctl->tree_lock);
2807
2808         return ret;
2809 }
2810
2811 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2812                                 struct btrfs_free_space *entry,
2813                                 struct btrfs_free_cluster *cluster,
2814                                 u64 offset, u64 bytes,
2815                                 u64 cont1_bytes, u64 min_bytes)
2816 {
2817         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2818         unsigned long next_zero;
2819         unsigned long i;
2820         unsigned long want_bits;
2821         unsigned long min_bits;
2822         unsigned long found_bits;
2823         unsigned long max_bits = 0;
2824         unsigned long start = 0;
2825         unsigned long total_found = 0;
2826         int ret;
2827
2828         i = offset_to_bit(entry->offset, ctl->unit,
2829                           max_t(u64, offset, entry->offset));
2830         want_bits = bytes_to_bits(bytes, ctl->unit);
2831         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2832
2833         /*
2834          * Don't bother looking for a cluster in this bitmap if it's heavily
2835          * fragmented.
2836          */
2837         if (entry->max_extent_size &&
2838             entry->max_extent_size < cont1_bytes)
2839                 return -ENOSPC;
2840 again:
2841         found_bits = 0;
2842         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2843                 next_zero = find_next_zero_bit(entry->bitmap,
2844                                                BITS_PER_BITMAP, i);
2845                 if (next_zero - i >= min_bits) {
2846                         found_bits = next_zero - i;
2847                         if (found_bits > max_bits)
2848                                 max_bits = found_bits;
2849                         break;
2850                 }
2851                 if (next_zero - i > max_bits)
2852                         max_bits = next_zero - i;
2853                 i = next_zero;
2854         }
2855
2856         if (!found_bits) {
2857                 entry->max_extent_size = (u64)max_bits * ctl->unit;
2858                 return -ENOSPC;
2859         }
2860
2861         if (!total_found) {
2862                 start = i;
2863                 cluster->max_size = 0;
2864         }
2865
2866         total_found += found_bits;
2867
2868         if (cluster->max_size < found_bits * ctl->unit)
2869                 cluster->max_size = found_bits * ctl->unit;
2870
2871         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2872                 i = next_zero + 1;
2873                 goto again;
2874         }
2875
2876         cluster->window_start = start * ctl->unit + entry->offset;
2877         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2878         ret = tree_insert_offset(&cluster->root, entry->offset,
2879                                  &entry->offset_index, 1);
2880         ASSERT(!ret); /* -EEXIST; Logic error */
2881
2882         trace_btrfs_setup_cluster(block_group, cluster,
2883                                   total_found * ctl->unit, 1);
2884         return 0;
2885 }
2886
2887 /*
2888  * This searches the block group for just extents to fill the cluster with.
2889  * Try to find a cluster with at least bytes total bytes, at least one
2890  * extent of cont1_bytes, and other clusters of at least min_bytes.
2891  */
2892 static noinline int
2893 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2894                         struct btrfs_free_cluster *cluster,
2895                         struct list_head *bitmaps, u64 offset, u64 bytes,
2896                         u64 cont1_bytes, u64 min_bytes)
2897 {
2898         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2899         struct btrfs_free_space *first = NULL;
2900         struct btrfs_free_space *entry = NULL;
2901         struct btrfs_free_space *last;
2902         struct rb_node *node;
2903         u64 window_free;
2904         u64 max_extent;
2905         u64 total_size = 0;
2906
2907         entry = tree_search_offset(ctl, offset, 0, 1);
2908         if (!entry)
2909                 return -ENOSPC;
2910
2911         /*
2912          * We don't want bitmaps, so just move along until we find a normal
2913          * extent entry.
2914          */
2915         while (entry->bitmap || entry->bytes < min_bytes) {
2916                 if (entry->bitmap && list_empty(&entry->list))
2917                         list_add_tail(&entry->list, bitmaps);
2918                 node = rb_next(&entry->offset_index);
2919                 if (!node)
2920                         return -ENOSPC;
2921                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2922         }
2923
2924         window_free = entry->bytes;
2925         max_extent = entry->bytes;
2926         first = entry;
2927         last = entry;
2928
2929         for (node = rb_next(&entry->offset_index); node;
2930              node = rb_next(&entry->offset_index)) {
2931                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2932
2933                 if (entry->bitmap) {
2934                         if (list_empty(&entry->list))
2935                                 list_add_tail(&entry->list, bitmaps);
2936                         continue;
2937                 }
2938
2939                 if (entry->bytes < min_bytes)
2940                         continue;
2941
2942                 last = entry;
2943                 window_free += entry->bytes;
2944                 if (entry->bytes > max_extent)
2945                         max_extent = entry->bytes;
2946         }
2947
2948         if (window_free < bytes || max_extent < cont1_bytes)
2949                 return -ENOSPC;
2950
2951         cluster->window_start = first->offset;
2952
2953         node = &first->offset_index;
2954
2955         /*
2956          * now we've found our entries, pull them out of the free space
2957          * cache and put them into the cluster rbtree
2958          */
2959         do {
2960                 int ret;
2961
2962                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2963                 node = rb_next(&entry->offset_index);
2964                 if (entry->bitmap || entry->bytes < min_bytes)
2965                         continue;
2966
2967                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2968                 ret = tree_insert_offset(&cluster->root, entry->offset,
2969                                          &entry->offset_index, 0);
2970                 total_size += entry->bytes;
2971                 ASSERT(!ret); /* -EEXIST; Logic error */
2972         } while (node && entry != last);
2973
2974         cluster->max_size = max_extent;
2975         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2976         return 0;
2977 }
2978
2979 /*
2980  * This specifically looks for bitmaps that may work in the cluster, we assume
2981  * that we have already failed to find extents that will work.
2982  */
2983 static noinline int
2984 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2985                      struct btrfs_free_cluster *cluster,
2986                      struct list_head *bitmaps, u64 offset, u64 bytes,
2987                      u64 cont1_bytes, u64 min_bytes)
2988 {
2989         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2990         struct btrfs_free_space *entry = NULL;
2991         int ret = -ENOSPC;
2992         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2993
2994         if (ctl->total_bitmaps == 0)
2995                 return -ENOSPC;
2996
2997         /*
2998          * The bitmap that covers offset won't be in the list unless offset
2999          * is just its start offset.
3000          */
3001         if (!list_empty(bitmaps))
3002                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3003
3004         if (!entry || entry->offset != bitmap_offset) {
3005                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3006                 if (entry && list_empty(&entry->list))
3007                         list_add(&entry->list, bitmaps);
3008         }
3009
3010         list_for_each_entry(entry, bitmaps, list) {
3011                 if (entry->bytes < bytes)
3012                         continue;
3013                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3014                                            bytes, cont1_bytes, min_bytes);
3015                 if (!ret)
3016                         return 0;
3017         }
3018
3019         /*
3020          * The bitmaps list has all the bitmaps that record free space
3021          * starting after offset, so no more search is required.
3022          */
3023         return -ENOSPC;
3024 }
3025
3026 /*
3027  * here we try to find a cluster of blocks in a block group.  The goal
3028  * is to find at least bytes+empty_size.
3029  * We might not find them all in one contiguous area.
3030  *
3031  * returns zero and sets up cluster if things worked out, otherwise
3032  * it returns -enospc
3033  */
3034 int btrfs_find_space_cluster(struct btrfs_root *root,
3035                              struct btrfs_block_group_cache *block_group,
3036                              struct btrfs_free_cluster *cluster,
3037                              u64 offset, u64 bytes, u64 empty_size)
3038 {
3039         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3040         struct btrfs_free_space *entry, *tmp;
3041         LIST_HEAD(bitmaps);
3042         u64 min_bytes;
3043         u64 cont1_bytes;
3044         int ret;
3045
3046         /*
3047          * Choose the minimum extent size we'll require for this
3048          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3049          * For metadata, allow allocates with smaller extents.  For
3050          * data, keep it dense.
3051          */
3052         if (btrfs_test_opt(root, SSD_SPREAD)) {
3053                 cont1_bytes = min_bytes = bytes + empty_size;
3054         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3055                 cont1_bytes = bytes;
3056                 min_bytes = block_group->sectorsize;
3057         } else {
3058                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3059                 min_bytes = block_group->sectorsize;
3060         }
3061
3062         spin_lock(&ctl->tree_lock);
3063
3064         /*
3065          * If we know we don't have enough space to make a cluster don't even
3066          * bother doing all the work to try and find one.
3067          */
3068         if (ctl->free_space < bytes) {
3069                 spin_unlock(&ctl->tree_lock);
3070                 return -ENOSPC;
3071         }
3072
3073         spin_lock(&cluster->lock);
3074
3075         /* someone already found a cluster, hooray */
3076         if (cluster->block_group) {
3077                 ret = 0;
3078                 goto out;
3079         }
3080
3081         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3082                                  min_bytes);
3083
3084         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3085                                       bytes + empty_size,
3086                                       cont1_bytes, min_bytes);
3087         if (ret)
3088                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3089                                            offset, bytes + empty_size,
3090                                            cont1_bytes, min_bytes);
3091
3092         /* Clear our temporary list */
3093         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3094                 list_del_init(&entry->list);
3095
3096         if (!ret) {
3097                 atomic_inc(&block_group->count);
3098                 list_add_tail(&cluster->block_group_list,
3099                               &block_group->cluster_list);
3100                 cluster->block_group = block_group;
3101         } else {
3102                 trace_btrfs_failed_cluster_setup(block_group);
3103         }
3104 out:
3105         spin_unlock(&cluster->lock);
3106         spin_unlock(&ctl->tree_lock);
3107
3108         return ret;
3109 }
3110
3111 /*
3112  * simple code to zero out a cluster
3113  */
3114 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3115 {
3116         spin_lock_init(&cluster->lock);
3117         spin_lock_init(&cluster->refill_lock);
3118         cluster->root = RB_ROOT;
3119         cluster->max_size = 0;
3120         cluster->fragmented = false;
3121         INIT_LIST_HEAD(&cluster->block_group_list);
3122         cluster->block_group = NULL;
3123 }
3124
3125 static int do_trimming(struct btrfs_block_group_cache *block_group,
3126                        u64 *total_trimmed, u64 start, u64 bytes,
3127                        u64 reserved_start, u64 reserved_bytes,
3128                        struct btrfs_trim_range *trim_entry)
3129 {
3130         struct btrfs_space_info *space_info = block_group->space_info;
3131         struct btrfs_fs_info *fs_info = block_group->fs_info;
3132         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3133         int ret;
3134         int update = 0;
3135         u64 trimmed = 0;
3136
3137         spin_lock(&space_info->lock);
3138         spin_lock(&block_group->lock);
3139         if (!block_group->ro) {
3140                 block_group->reserved += reserved_bytes;
3141                 space_info->bytes_reserved += reserved_bytes;
3142                 update = 1;
3143         }
3144         spin_unlock(&block_group->lock);
3145         spin_unlock(&space_info->lock);
3146
3147         ret = btrfs_discard_extent(fs_info->extent_root,
3148                                    start, bytes, &trimmed);
3149         if (!ret)
3150                 *total_trimmed += trimmed;
3151
3152         mutex_lock(&ctl->cache_writeout_mutex);
3153         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3154         list_del(&trim_entry->list);
3155         mutex_unlock(&ctl->cache_writeout_mutex);
3156
3157         if (update) {
3158                 spin_lock(&space_info->lock);
3159                 spin_lock(&block_group->lock);
3160                 if (block_group->ro)
3161                         space_info->bytes_readonly += reserved_bytes;
3162                 block_group->reserved -= reserved_bytes;
3163                 space_info->bytes_reserved -= reserved_bytes;
3164                 spin_unlock(&space_info->lock);
3165                 spin_unlock(&block_group->lock);
3166         }
3167
3168         return ret;
3169 }
3170
3171 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3172                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3173 {
3174         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3175         struct btrfs_free_space *entry;
3176         struct rb_node *node;
3177         int ret = 0;
3178         u64 extent_start;
3179         u64 extent_bytes;
3180         u64 bytes;
3181
3182         while (start < end) {
3183                 struct btrfs_trim_range trim_entry;
3184
3185                 mutex_lock(&ctl->cache_writeout_mutex);
3186                 spin_lock(&ctl->tree_lock);
3187
3188                 if (ctl->free_space < minlen) {
3189                         spin_unlock(&ctl->tree_lock);
3190                         mutex_unlock(&ctl->cache_writeout_mutex);
3191                         break;
3192                 }
3193
3194                 entry = tree_search_offset(ctl, start, 0, 1);
3195                 if (!entry) {
3196                         spin_unlock(&ctl->tree_lock);
3197                         mutex_unlock(&ctl->cache_writeout_mutex);
3198                         break;
3199                 }
3200
3201                 /* skip bitmaps */
3202                 while (entry->bitmap) {
3203                         node = rb_next(&entry->offset_index);
3204                         if (!node) {
3205                                 spin_unlock(&ctl->tree_lock);
3206                                 mutex_unlock(&ctl->cache_writeout_mutex);
3207                                 goto out;
3208                         }
3209                         entry = rb_entry(node, struct btrfs_free_space,
3210                                          offset_index);
3211                 }
3212
3213                 if (entry->offset >= end) {
3214                         spin_unlock(&ctl->tree_lock);
3215                         mutex_unlock(&ctl->cache_writeout_mutex);
3216                         break;
3217                 }
3218
3219                 extent_start = entry->offset;
3220                 extent_bytes = entry->bytes;
3221                 start = max(start, extent_start);
3222                 bytes = min(extent_start + extent_bytes, end) - start;
3223                 if (bytes < minlen) {
3224                         spin_unlock(&ctl->tree_lock);
3225                         mutex_unlock(&ctl->cache_writeout_mutex);
3226                         goto next;
3227                 }
3228
3229                 unlink_free_space(ctl, entry);
3230                 kmem_cache_free(btrfs_free_space_cachep, entry);
3231
3232                 spin_unlock(&ctl->tree_lock);
3233                 trim_entry.start = extent_start;
3234                 trim_entry.bytes = extent_bytes;
3235                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3236                 mutex_unlock(&ctl->cache_writeout_mutex);
3237
3238                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3239                                   extent_start, extent_bytes, &trim_entry);
3240                 if (ret)
3241                         break;
3242 next:
3243                 start += bytes;
3244
3245                 if (fatal_signal_pending(current)) {
3246                         ret = -ERESTARTSYS;
3247                         break;
3248                 }
3249
3250                 cond_resched();
3251         }
3252 out:
3253         return ret;
3254 }
3255
3256 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3257                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3258 {
3259         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3260         struct btrfs_free_space *entry;
3261         int ret = 0;
3262         int ret2;
3263         u64 bytes;
3264         u64 offset = offset_to_bitmap(ctl, start);
3265
3266         while (offset < end) {
3267                 bool next_bitmap = false;
3268                 struct btrfs_trim_range trim_entry;
3269
3270                 mutex_lock(&ctl->cache_writeout_mutex);
3271                 spin_lock(&ctl->tree_lock);
3272
3273                 if (ctl->free_space < minlen) {
3274                         spin_unlock(&ctl->tree_lock);
3275                         mutex_unlock(&ctl->cache_writeout_mutex);
3276                         break;
3277                 }
3278
3279                 entry = tree_search_offset(ctl, offset, 1, 0);
3280                 if (!entry) {
3281                         spin_unlock(&ctl->tree_lock);
3282                         mutex_unlock(&ctl->cache_writeout_mutex);
3283                         next_bitmap = true;
3284                         goto next;
3285                 }
3286
3287                 bytes = minlen;
3288                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3289                 if (ret2 || start >= end) {
3290                         spin_unlock(&ctl->tree_lock);
3291                         mutex_unlock(&ctl->cache_writeout_mutex);
3292                         next_bitmap = true;
3293                         goto next;
3294                 }
3295
3296                 bytes = min(bytes, end - start);
3297                 if (bytes < minlen) {
3298                         spin_unlock(&ctl->tree_lock);
3299                         mutex_unlock(&ctl->cache_writeout_mutex);
3300                         goto next;
3301                 }
3302
3303                 bitmap_clear_bits(ctl, entry, start, bytes);
3304                 if (entry->bytes == 0)
3305                         free_bitmap(ctl, entry);
3306
3307                 spin_unlock(&ctl->tree_lock);
3308                 trim_entry.start = start;
3309                 trim_entry.bytes = bytes;
3310                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3311                 mutex_unlock(&ctl->cache_writeout_mutex);
3312
3313                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3314                                   start, bytes, &trim_entry);
3315                 if (ret)
3316                         break;
3317 next:
3318                 if (next_bitmap) {
3319                         offset += BITS_PER_BITMAP * ctl->unit;
3320                 } else {
3321                         start += bytes;
3322                         if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3323                                 offset += BITS_PER_BITMAP * ctl->unit;
3324                 }
3325
3326                 if (fatal_signal_pending(current)) {
3327                         ret = -ERESTARTSYS;
3328                         break;
3329                 }
3330
3331                 cond_resched();
3332         }
3333
3334         return ret;
3335 }
3336
3337 void btrfs_get_block_group_trimming(struct btrfs_block_group_cache *cache)
3338 {
3339         atomic_inc(&cache->trimming);
3340 }
3341
3342 void btrfs_put_block_group_trimming(struct btrfs_block_group_cache *block_group)
3343 {
3344         struct extent_map_tree *em_tree;
3345         struct extent_map *em;
3346         bool cleanup;
3347
3348         spin_lock(&block_group->lock);
3349         cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3350                    block_group->removed);
3351         spin_unlock(&block_group->lock);
3352
3353         if (cleanup) {
3354                 lock_chunks(block_group->fs_info->chunk_root);
3355                 em_tree = &block_group->fs_info->mapping_tree.map_tree;
3356                 write_lock(&em_tree->lock);
3357                 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3358                                            1);
3359                 BUG_ON(!em); /* logic error, can't happen */
3360                 /*
3361                  * remove_extent_mapping() will delete us from the pinned_chunks
3362                  * list, which is protected by the chunk mutex.
3363                  */
3364                 remove_extent_mapping(em_tree, em);
3365                 write_unlock(&em_tree->lock);
3366                 unlock_chunks(block_group->fs_info->chunk_root);
3367
3368                 /* once for us and once for the tree */
3369                 free_extent_map(em);
3370                 free_extent_map(em);
3371
3372                 /*
3373                  * We've left one free space entry and other tasks trimming
3374                  * this block group have left 1 entry each one. Free them.
3375                  */
3376                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3377         }
3378 }
3379
3380 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3381                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3382 {
3383         int ret;
3384
3385         *trimmed = 0;
3386
3387         spin_lock(&block_group->lock);
3388         if (block_group->removed) {
3389                 spin_unlock(&block_group->lock);
3390                 return 0;
3391         }
3392         btrfs_get_block_group_trimming(block_group);
3393         spin_unlock(&block_group->lock);
3394
3395         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3396         if (ret)
3397                 goto out;
3398
3399         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3400 out:
3401         btrfs_put_block_group_trimming(block_group);
3402         return ret;
3403 }
3404
3405 /*
3406  * Find the left-most item in the cache tree, and then return the
3407  * smallest inode number in the item.
3408  *
3409  * Note: the returned inode number may not be the smallest one in
3410  * the tree, if the left-most item is a bitmap.
3411  */
3412 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3413 {
3414         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3415         struct btrfs_free_space *entry = NULL;
3416         u64 ino = 0;
3417
3418         spin_lock(&ctl->tree_lock);
3419
3420         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3421                 goto out;
3422
3423         entry = rb_entry(rb_first(&ctl->free_space_offset),
3424                          struct btrfs_free_space, offset_index);
3425
3426         if (!entry->bitmap) {
3427                 ino = entry->offset;
3428
3429                 unlink_free_space(ctl, entry);
3430                 entry->offset++;
3431                 entry->bytes--;
3432                 if (!entry->bytes)
3433                         kmem_cache_free(btrfs_free_space_cachep, entry);
3434                 else
3435                         link_free_space(ctl, entry);
3436         } else {
3437                 u64 offset = 0;
3438                 u64 count = 1;
3439                 int ret;
3440
3441                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3442                 /* Logic error; Should be empty if it can't find anything */
3443                 ASSERT(!ret);
3444
3445                 ino = offset;
3446                 bitmap_clear_bits(ctl, entry, offset, 1);
3447                 if (entry->bytes == 0)
3448                         free_bitmap(ctl, entry);
3449         }
3450 out:
3451         spin_unlock(&ctl->tree_lock);
3452
3453         return ino;
3454 }
3455
3456 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3457                                     struct btrfs_path *path)
3458 {
3459         struct inode *inode = NULL;
3460
3461         spin_lock(&root->ino_cache_lock);
3462         if (root->ino_cache_inode)
3463                 inode = igrab(root->ino_cache_inode);
3464         spin_unlock(&root->ino_cache_lock);
3465         if (inode)
3466                 return inode;
3467
3468         inode = __lookup_free_space_inode(root, path, 0);
3469         if (IS_ERR(inode))
3470                 return inode;
3471
3472         spin_lock(&root->ino_cache_lock);
3473         if (!btrfs_fs_closing(root->fs_info))
3474                 root->ino_cache_inode = igrab(inode);
3475         spin_unlock(&root->ino_cache_lock);
3476
3477         return inode;
3478 }
3479
3480 int create_free_ino_inode(struct btrfs_root *root,
3481                           struct btrfs_trans_handle *trans,
3482                           struct btrfs_path *path)
3483 {
3484         return __create_free_space_inode(root, trans, path,
3485                                          BTRFS_FREE_INO_OBJECTID, 0);
3486 }
3487
3488 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3489 {
3490         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3491         struct btrfs_path *path;
3492         struct inode *inode;
3493         int ret = 0;
3494         u64 root_gen = btrfs_root_generation(&root->root_item);
3495
3496         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3497                 return 0;
3498
3499         /*
3500          * If we're unmounting then just return, since this does a search on the
3501          * normal root and not the commit root and we could deadlock.
3502          */
3503         if (btrfs_fs_closing(fs_info))
3504                 return 0;
3505
3506         path = btrfs_alloc_path();
3507         if (!path)
3508                 return 0;
3509
3510         inode = lookup_free_ino_inode(root, path);
3511         if (IS_ERR(inode))
3512                 goto out;
3513
3514         if (root_gen != BTRFS_I(inode)->generation)
3515                 goto out_put;
3516
3517         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3518
3519         if (ret < 0)
3520                 btrfs_err(fs_info,
3521                         "failed to load free ino cache for root %llu",
3522                         root->root_key.objectid);
3523 out_put:
3524         iput(inode);
3525 out:
3526         btrfs_free_path(path);
3527         return ret;
3528 }
3529
3530 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3531                               struct btrfs_trans_handle *trans,
3532                               struct btrfs_path *path,
3533                               struct inode *inode)
3534 {
3535         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3536         int ret;
3537         struct btrfs_io_ctl io_ctl;
3538         bool release_metadata = true;
3539
3540         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3541                 return 0;
3542
3543         memset(&io_ctl, 0, sizeof(io_ctl));
3544         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl,
3545                                       trans, path, 0);
3546         if (!ret) {
3547                 /*
3548                  * At this point writepages() didn't error out, so our metadata
3549                  * reservation is released when the writeback finishes, at
3550                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3551                  * with or without an error.
3552                  */
3553                 release_metadata = false;
3554                 ret = btrfs_wait_cache_io(root, trans, NULL, &io_ctl, path, 0);
3555         }
3556
3557         if (ret) {
3558                 if (release_metadata)
3559                         btrfs_delalloc_release_metadata(inode, inode->i_size);
3560 #ifdef DEBUG
3561                 btrfs_err(root->fs_info,
3562                         "failed to write free ino cache for root %llu",
3563                         root->root_key.objectid);
3564 #endif
3565         }
3566
3567         return ret;
3568 }
3569
3570 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3571 /*
3572  * Use this if you need to make a bitmap or extent entry specifically, it
3573  * doesn't do any of the merging that add_free_space does, this acts a lot like
3574  * how the free space cache loading stuff works, so you can get really weird
3575  * configurations.
3576  */
3577 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3578                               u64 offset, u64 bytes, bool bitmap)
3579 {
3580         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3581         struct btrfs_free_space *info = NULL, *bitmap_info;
3582         void *map = NULL;
3583         u64 bytes_added;
3584         int ret;
3585
3586 again:
3587         if (!info) {
3588                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3589                 if (!info)
3590                         return -ENOMEM;
3591         }
3592
3593         if (!bitmap) {
3594                 spin_lock(&ctl->tree_lock);
3595                 info->offset = offset;
3596                 info->bytes = bytes;
3597                 info->max_extent_size = 0;
3598                 ret = link_free_space(ctl, info);
3599                 spin_unlock(&ctl->tree_lock);
3600                 if (ret)
3601                         kmem_cache_free(btrfs_free_space_cachep, info);
3602                 return ret;
3603         }
3604
3605         if (!map) {
3606                 map = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
3607                 if (!map) {
3608                         kmem_cache_free(btrfs_free_space_cachep, info);
3609                         return -ENOMEM;
3610                 }
3611         }
3612
3613         spin_lock(&ctl->tree_lock);
3614         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3615                                          1, 0);
3616         if (!bitmap_info) {
3617                 info->bitmap = map;
3618                 map = NULL;
3619                 add_new_bitmap(ctl, info, offset);
3620                 bitmap_info = info;
3621                 info = NULL;
3622         }
3623
3624         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3625
3626         bytes -= bytes_added;
3627         offset += bytes_added;
3628         spin_unlock(&ctl->tree_lock);
3629
3630         if (bytes)
3631                 goto again;
3632
3633         if (info)
3634                 kmem_cache_free(btrfs_free_space_cachep, info);
3635         if (map)
3636                 kfree(map);
3637         return 0;
3638 }
3639
3640 /*
3641  * Checks to see if the given range is in the free space cache.  This is really
3642  * just used to check the absence of space, so if there is free space in the
3643  * range at all we will return 1.
3644  */
3645 int test_check_exists(struct btrfs_block_group_cache *cache,
3646                       u64 offset, u64 bytes)
3647 {
3648         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3649         struct btrfs_free_space *info;
3650         int ret = 0;
3651
3652         spin_lock(&ctl->tree_lock);
3653         info = tree_search_offset(ctl, offset, 0, 0);
3654         if (!info) {
3655                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3656                                           1, 0);
3657                 if (!info)
3658                         goto out;
3659         }
3660
3661 have_info:
3662         if (info->bitmap) {
3663                 u64 bit_off, bit_bytes;
3664                 struct rb_node *n;
3665                 struct btrfs_free_space *tmp;
3666
3667                 bit_off = offset;
3668                 bit_bytes = ctl->unit;
3669                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3670                 if (!ret) {
3671                         if (bit_off == offset) {
3672                                 ret = 1;
3673                                 goto out;
3674                         } else if (bit_off > offset &&
3675                                    offset + bytes > bit_off) {
3676                                 ret = 1;
3677                                 goto out;
3678                         }
3679                 }
3680
3681                 n = rb_prev(&info->offset_index);
3682                 while (n) {
3683                         tmp = rb_entry(n, struct btrfs_free_space,
3684                                        offset_index);
3685                         if (tmp->offset + tmp->bytes < offset)
3686                                 break;
3687                         if (offset + bytes < tmp->offset) {
3688                                 n = rb_prev(&info->offset_index);
3689                                 continue;
3690                         }
3691                         info = tmp;
3692                         goto have_info;
3693                 }
3694
3695                 n = rb_next(&info->offset_index);
3696                 while (n) {
3697                         tmp = rb_entry(n, struct btrfs_free_space,
3698                                        offset_index);
3699                         if (offset + bytes < tmp->offset)
3700                                 break;
3701                         if (tmp->offset + tmp->bytes < offset) {
3702                                 n = rb_next(&info->offset_index);
3703                                 continue;
3704                         }
3705                         info = tmp;
3706                         goto have_info;
3707                 }
3708
3709                 ret = 0;
3710                 goto out;
3711         }
3712
3713         if (info->offset == offset) {
3714                 ret = 1;
3715                 goto out;
3716         }
3717
3718         if (offset > info->offset && offset < info->offset + info->bytes)
3719                 ret = 1;
3720 out:
3721         spin_unlock(&ctl->tree_lock);
3722         return ret;
3723 }
3724 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */