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