fs/btrfs/free-space-cache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2008 Red Hat.  All rights reserved.
   4  */
   5
   6 #include <linux/pagemap.h>
   7 #include <linux/sched.h>
   8 #include <linux/sched/signal.h>
   9 #include <linux/slab.h>
  10 #include <linux/math64.h>
  11 #include <linux/ratelimit.h>
  12 #include <linux/error-injection.h>
  13 #include "ctree.h"
  14 #include "free-space-cache.h"
  15 #include "transaction.h"
  16 #include "disk-io.h"
  17 #include "extent_io.h"
  18 #include "inode-map.h"
  19 #include "volumes.h"
  20
  21 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  22 #define MAX_CACHE_BYTES_PER_GIG SZ_32K
  23
  24 struct btrfs_trim_range {
  25         u64 start;
  26         u64 bytes;
  27         struct list_head list;
  28 };
  29
  30 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  31                            struct btrfs_free_space *info);
  32 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  33                               struct btrfs_free_space *info);
  34 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
  35                              struct btrfs_trans_handle *trans,
  36                              struct btrfs_io_ctl *io_ctl,
  37                              struct btrfs_path *path);
  38
  39 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  40                                                struct btrfs_path *path,
  41                                                u64 offset)
  42 {
  43         struct btrfs_fs_info *fs_info = root->fs_info;
  44         struct btrfs_key key;
  45         struct btrfs_key location;
  46         struct btrfs_disk_key disk_key;
  47         struct btrfs_free_space_header *header;
  48         struct extent_buffer *leaf;
  49         struct inode *inode = NULL;
  50         int ret;
  51
  52         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  53         key.offset = offset;
  54         key.type = 0;
  55
  56         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  57         if (ret < 0)
  58                 return ERR_PTR(ret);
  59         if (ret > 0) {
  60                 btrfs_release_path(path);
  61                 return ERR_PTR(-ENOENT);
  62         }
  63
  64         leaf = path->nodes[0];
  65         header = btrfs_item_ptr(leaf, path->slots[0],
  66                                 struct btrfs_free_space_header);
  67         btrfs_free_space_key(leaf, header, &disk_key);
  68         btrfs_disk_key_to_cpu(&location, &disk_key);
  69         btrfs_release_path(path);
  70
  71         inode = btrfs_iget(fs_info->sb, &location, root, NULL);
  72         if (IS_ERR(inode))
  73                 return inode;
  74         if (is_bad_inode(inode)) {
  75                 iput(inode);
  76                 return ERR_PTR(-ENOENT);
  77         }
  78
  79         mapping_set_gfp_mask(inode->i_mapping,
  80                         mapping_gfp_constraint(inode->i_mapping,
  81                         ~(__GFP_FS | __GFP_HIGHMEM)));
  82
  83         return inode;
  84 }
  85
  86 struct inode *lookup_free_space_inode(struct btrfs_fs_info *fs_info,
  87                                       struct btrfs_block_group_cache
  88                                       *block_group, struct btrfs_path *path)
  89 {
  90         struct inode *inode = NULL;
  91         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
  92
  93         spin_lock(&block_group->lock);
  94         if (block_group->inode)
  95                 inode = igrab(block_group->inode);
  96         spin_unlock(&block_group->lock);
  97         if (inode)
  98                 return inode;
  99
 100         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 101                                           block_group->key.objectid);
 102         if (IS_ERR(inode))
 103                 return inode;
 104
 105         spin_lock(&block_group->lock);
 106         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 107                 btrfs_info(fs_info, "Old style space inode found, converting.");
 108                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 109                         BTRFS_INODE_NODATACOW;
 110                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 111         }
 112
 113         if (!block_group->iref) {
 114                 block_group->inode = igrab(inode);
 115                 block_group->iref = 1;
 116         }
 117         spin_unlock(&block_group->lock);
 118
 119         return inode;
 120 }
 121
 122 static int __create_free_space_inode(struct btrfs_root *root,
 123                                      struct btrfs_trans_handle *trans,
 124                                      struct btrfs_path *path,
 125                                      u64 ino, u64 offset)
 126 {
 127         struct btrfs_key key;
 128         struct btrfs_disk_key disk_key;
 129         struct btrfs_free_space_header *header;
 130         struct btrfs_inode_item *inode_item;
 131         struct extent_buffer *leaf;
 132         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 133         int ret;
 134
 135         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 136         if (ret)
 137                 return ret;
 138
 139         /* We inline crc's for the free disk space cache */
 140         if (ino != BTRFS_FREE_INO_OBJECTID)
 141                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 142
 143         leaf = path->nodes[0];
 144         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 145                                     struct btrfs_inode_item);
 146         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 147         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 148                              sizeof(*inode_item));
 149         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 150         btrfs_set_inode_size(leaf, inode_item, 0);
 151         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 152         btrfs_set_inode_uid(leaf, inode_item, 0);
 153         btrfs_set_inode_gid(leaf, inode_item, 0);
 154         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 155         btrfs_set_inode_flags(leaf, inode_item, flags);
 156         btrfs_set_inode_nlink(leaf, inode_item, 1);
 157         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 158         btrfs_set_inode_block_group(leaf, inode_item, offset);
 159         btrfs_mark_buffer_dirty(leaf);
 160         btrfs_release_path(path);
 161
 162         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 163         key.offset = offset;
 164         key.type = 0;
 165         ret = btrfs_insert_empty_item(trans, root, path, &key,
 166                                       sizeof(struct btrfs_free_space_header));
 167         if (ret < 0) {
 168                 btrfs_release_path(path);
 169                 return ret;
 170         }
 171
 172         leaf = path->nodes[0];
 173         header = btrfs_item_ptr(leaf, path->slots[0],
 174                                 struct btrfs_free_space_header);
 175         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 176         btrfs_set_free_space_key(leaf, header, &disk_key);
 177         btrfs_mark_buffer_dirty(leaf);
 178         btrfs_release_path(path);
 179
 180         return 0;
 181 }
 182
 183 int create_free_space_inode(struct btrfs_fs_info *fs_info,
 184                             struct btrfs_trans_handle *trans,
 185                             struct btrfs_block_group_cache *block_group,
 186                             struct btrfs_path *path)
 187 {
 188         int ret;
 189         u64 ino;
 190
 191         ret = btrfs_find_free_objectid(fs_info->tree_root, &ino);
 192         if (ret < 0)
 193                 return ret;
 194
 195         return __create_free_space_inode(fs_info->tree_root, trans, path, ino,
 196                                          block_group->key.objectid);
 197 }
 198
 199 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
 200                                        struct btrfs_block_rsv *rsv)
 201 {
 202         u64 needed_bytes;
 203         int ret;
 204
 205         /* 1 for slack space, 1 for updating the inode */
 206         needed_bytes = btrfs_calc_trunc_metadata_size(fs_info, 1) +
 207                 btrfs_calc_trans_metadata_size(fs_info, 1);
 208
 209         spin_lock(&rsv->lock);
 210         if (rsv->reserved < needed_bytes)
 211                 ret = -ENOSPC;
 212         else
 213                 ret = 0;
 214         spin_unlock(&rsv->lock);
 215         return ret;
 216 }
 217
 218 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 219                                     struct btrfs_block_group_cache *block_group,
 220                                     struct inode *inode)
 221 {
 222         struct btrfs_root *root = BTRFS_I(inode)->root;
 223         int ret = 0;
 224         bool locked = false;
 225
 226         if (block_group) {
 227                 struct btrfs_path *path = btrfs_alloc_path();
 228
 229                 if (!path) {
 230                         ret = -ENOMEM;
 231                         goto fail;
 232                 }
 233                 locked = true;
 234                 mutex_lock(&trans->transaction->cache_write_mutex);
 235                 if (!list_empty(&block_group->io_list)) {
 236                         list_del_init(&block_group->io_list);
 237
 238                         btrfs_wait_cache_io(trans, block_group, path);
 239                         btrfs_put_block_group(block_group);
 240                 }
 241
 242                 /*
 243                  * now that we've truncated the cache away, its no longer
 244                  * setup or written
 245                  */
 246                 spin_lock(&block_group->lock);
 247                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 248                 spin_unlock(&block_group->lock);
 249                 btrfs_free_path(path);
 250         }
 251
 252         btrfs_i_size_write(BTRFS_I(inode), 0);
 253         truncate_pagecache(inode, 0);
 254
 255         /*
 256          * We don't need an orphan item because truncating the free space cache
 257          * will never be split across transactions.
 258          * We don't need to check for -EAGAIN because we're a free space
 259          * cache inode
 260          */
 261         ret = btrfs_truncate_inode_items(trans, root, inode,
 262                                          0, BTRFS_EXTENT_DATA_KEY);
 263         if (ret)
 264                 goto fail;
 265
 266         ret = btrfs_update_inode(trans, root, inode);
 267
 268 fail:
 269         if (locked)
 270                 mutex_unlock(&trans->transaction->cache_write_mutex);
 271         if (ret)
 272                 btrfs_abort_transaction(trans, ret);
 273
 274         return ret;
 275 }
 276
 277 static void readahead_cache(struct inode *inode)
 278 {
 279         struct file_ra_state *ra;
 280         unsigned long last_index;
 281
 282         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 283         if (!ra)
 284                 return;
 285
 286         file_ra_state_init(ra, inode->i_mapping);
 287         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 288
 289         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 290
 291         kfree(ra);
 292 }
 293
 294 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 295                        int write)
 296 {
 297         int num_pages;
 298         int check_crcs = 0;
 299
 300         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 301
 302         if (btrfs_ino(BTRFS_I(inode)) != BTRFS_FREE_INO_OBJECTID)
 303                 check_crcs = 1;
 304
 305         /* Make sure we can fit our crcs into the first page */
 306         if (write && check_crcs &&
 307             (num_pages * sizeof(u32)) >= PAGE_SIZE)
 308                 return -ENOSPC;
 309
 310         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 311
 312         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 313         if (!io_ctl->pages)
 314                 return -ENOMEM;
 315
 316         io_ctl->num_pages = num_pages;
 317         io_ctl->fs_info = btrfs_sb(inode->i_sb);
 318         io_ctl->check_crcs = check_crcs;
 319         io_ctl->inode = inode;
 320
 321         return 0;
 322 }
 323 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 324
 325 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 326 {
 327         kfree(io_ctl->pages);
 328         io_ctl->pages = NULL;
 329 }
 330
 331 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 332 {
 333         if (io_ctl->cur) {
 334                 io_ctl->cur = NULL;
 335                 io_ctl->orig = NULL;
 336         }
 337 }
 338
 339 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 340 {
 341         ASSERT(io_ctl->index < io_ctl->num_pages);
 342         io_ctl->page = io_ctl->pages[io_ctl->index++];
 343         io_ctl->cur = page_address(io_ctl->page);
 344         io_ctl->orig = io_ctl->cur;
 345         io_ctl->size = PAGE_SIZE;
 346         if (clear)
 347                 clear_page(io_ctl->cur);
 348 }
 349
 350 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 351 {
 352         int i;
 353
 354         io_ctl_unmap_page(io_ctl);
 355
 356         for (i = 0; i < io_ctl->num_pages; i++) {
 357                 if (io_ctl->pages[i]) {
 358                         ClearPageChecked(io_ctl->pages[i]);
 359                         unlock_page(io_ctl->pages[i]);
 360                         put_page(io_ctl->pages[i]);
 361                 }
 362         }
 363 }
 364
 365 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 366                                 int uptodate)
 367 {
 368         struct page *page;
 369         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 370         int i;
 371
 372         for (i = 0; i < io_ctl->num_pages; i++) {
 373                 page = find_or_create_page(inode->i_mapping, i, mask);
 374                 if (!page) {
 375                         io_ctl_drop_pages(io_ctl);
 376                         return -ENOMEM;
 377                 }
 378                 io_ctl->pages[i] = page;
 379                 if (uptodate && !PageUptodate(page)) {
 380                         btrfs_readpage(NULL, page);
 381                         lock_page(page);
 382                         if (!PageUptodate(page)) {
 383                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 384                                            "error reading free space cache");
 385                                 io_ctl_drop_pages(io_ctl);
 386                                 return -EIO;
 387                         }
 388                 }
 389         }
 390
 391         for (i = 0; i < io_ctl->num_pages; i++) {
 392                 clear_page_dirty_for_io(io_ctl->pages[i]);
 393                 set_page_extent_mapped(io_ctl->pages[i]);
 394         }
 395
 396         return 0;
 397 }
 398
 399 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 400 {
 401         __le64 *val;
 402
 403         io_ctl_map_page(io_ctl, 1);
 404
 405         /*
 406          * Skip the csum areas.  If we don't check crcs then we just have a
 407          * 64bit chunk at the front of the first page.
 408          */
 409         if (io_ctl->check_crcs) {
 410                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 411                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 412         } else {
 413                 io_ctl->cur += sizeof(u64);
 414                 io_ctl->size -= sizeof(u64) * 2;
 415         }
 416
 417         val = io_ctl->cur;
 418         *val = cpu_to_le64(generation);
 419         io_ctl->cur += sizeof(u64);
 420 }
 421
 422 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 423 {
 424         __le64 *gen;
 425
 426         /*
 427          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 428          * chunk at the front of the first page.
 429          */
 430         if (io_ctl->check_crcs) {
 431                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 432                 io_ctl->size -= sizeof(u64) +
 433                         (sizeof(u32) * io_ctl->num_pages);
 434         } else {
 435                 io_ctl->cur += sizeof(u64);
 436                 io_ctl->size -= sizeof(u64) * 2;
 437         }
 438
 439         gen = io_ctl->cur;
 440         if (le64_to_cpu(*gen) != generation) {
 441                 btrfs_err_rl(io_ctl->fs_info,
 442                         "space cache generation (%llu) does not match inode (%llu)",
 443                                 *gen, generation);
 444                 io_ctl_unmap_page(io_ctl);
 445                 return -EIO;
 446         }
 447         io_ctl->cur += sizeof(u64);
 448         return 0;
 449 }
 450
 451 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 452 {
 453         u32 *tmp;
 454         u32 crc = ~(u32)0;
 455         unsigned offset = 0;
 456
 457         if (!io_ctl->check_crcs) {
 458                 io_ctl_unmap_page(io_ctl);
 459                 return;
 460         }
 461
 462         if (index == 0)
 463                 offset = sizeof(u32) * io_ctl->num_pages;
 464
 465         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 466                               PAGE_SIZE - offset);
 467         btrfs_csum_final(crc, (u8 *)&crc);
 468         io_ctl_unmap_page(io_ctl);
 469         tmp = page_address(io_ctl->pages[0]);
 470         tmp += index;
 471         *tmp = crc;
 472 }
 473
 474 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 475 {
 476         u32 *tmp, val;
 477         u32 crc = ~(u32)0;
 478         unsigned offset = 0;
 479
 480         if (!io_ctl->check_crcs) {
 481                 io_ctl_map_page(io_ctl, 0);
 482                 return 0;
 483         }
 484
 485         if (index == 0)
 486                 offset = sizeof(u32) * io_ctl->num_pages;
 487
 488         tmp = page_address(io_ctl->pages[0]);
 489         tmp += index;
 490         val = *tmp;
 491
 492         io_ctl_map_page(io_ctl, 0);
 493         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 494                               PAGE_SIZE - offset);
 495         btrfs_csum_final(crc, (u8 *)&crc);
 496         if (val != crc) {
 497                 btrfs_err_rl(io_ctl->fs_info,
 498                         "csum mismatch on free space cache");
 499                 io_ctl_unmap_page(io_ctl);
 500                 return -EIO;
 501         }
 502
 503         return 0;
 504 }
 505
 506 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 507                             void *bitmap)
 508 {
 509         struct btrfs_free_space_entry *entry;
 510
 511         if (!io_ctl->cur)
 512                 return -ENOSPC;
 513
 514         entry = io_ctl->cur;
 515         entry->offset = cpu_to_le64(offset);
 516         entry->bytes = cpu_to_le64(bytes);
 517         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 518                 BTRFS_FREE_SPACE_EXTENT;
 519         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 520         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 521
 522         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 523                 return 0;
 524
 525         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 526
 527         /* No more pages to map */
 528         if (io_ctl->index >= io_ctl->num_pages)
 529                 return 0;
 530
 531         /* map the next page */
 532         io_ctl_map_page(io_ctl, 1);
 533         return 0;
 534 }
 535
 536 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 537 {
 538         if (!io_ctl->cur)
 539                 return -ENOSPC;
 540
 541         /*
 542          * If we aren't at the start of the current page, unmap this one and
 543          * map the next one if there is any left.
 544          */
 545         if (io_ctl->cur != io_ctl->orig) {
 546                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 547                 if (io_ctl->index >= io_ctl->num_pages)
 548                         return -ENOSPC;
 549                 io_ctl_map_page(io_ctl, 0);
 550         }
 551
 552         memcpy(io_ctl->cur, bitmap, PAGE_SIZE);
 553         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 554         if (io_ctl->index < io_ctl->num_pages)
 555                 io_ctl_map_page(io_ctl, 0);
 556         return 0;
 557 }
 558
 559 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 560 {
 561         /*
 562          * If we're not on the boundary we know we've modified the page and we
 563          * need to crc the page.
 564          */
 565         if (io_ctl->cur != io_ctl->orig)
 566                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 567         else
 568                 io_ctl_unmap_page(io_ctl);
 569
 570         while (io_ctl->index < io_ctl->num_pages) {
 571                 io_ctl_map_page(io_ctl, 1);
 572                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 573         }
 574 }
 575
 576 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 577                             struct btrfs_free_space *entry, u8 *type)
 578 {
 579         struct btrfs_free_space_entry *e;
 580         int ret;
 581
 582         if (!io_ctl->cur) {
 583                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 584                 if (ret)
 585                         return ret;
 586         }
 587
 588         e = io_ctl->cur;
 589         entry->offset = le64_to_cpu(e->offset);
 590         entry->bytes = le64_to_cpu(e->bytes);
 591         *type = e->type;
 592         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 593         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 594
 595         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 596                 return 0;
 597
 598         io_ctl_unmap_page(io_ctl);
 599
 600         return 0;
 601 }
 602
 603 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 604                               struct btrfs_free_space *entry)
 605 {
 606         int ret;
 607
 608         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 609         if (ret)
 610                 return ret;
 611
 612         memcpy(entry->bitmap, io_ctl->cur, PAGE_SIZE);
 613         io_ctl_unmap_page(io_ctl);
 614
 615         return 0;
 616 }
 617
 618 /*
 619  * Since we attach pinned extents after the fact we can have contiguous sections
 620  * of free space that are split up in entries.  This poses a problem with the
 621  * tree logging stuff since it could have allocated across what appears to be 2
 622  * entries since we would have merged the entries when adding the pinned extents
 623  * back to the free space cache.  So run through the space cache that we just
 624  * loaded and merge contiguous entries.  This will make the log replay stuff not
 625  * blow up and it will make for nicer allocator behavior.
 626  */
 627 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 628 {
 629         struct btrfs_free_space *e, *prev = NULL;
 630         struct rb_node *n;
 631
 632 again:
 633         spin_lock(&ctl->tree_lock);
 634         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 635                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 636                 if (!prev)
 637                         goto next;
 638                 if (e->bitmap || prev->bitmap)
 639                         goto next;
 640                 if (prev->offset + prev->bytes == e->offset) {
 641                         unlink_free_space(ctl, prev);
 642                         unlink_free_space(ctl, e);
 643                         prev->bytes += e->bytes;
 644                         kmem_cache_free(btrfs_free_space_cachep, e);
 645                         link_free_space(ctl, prev);
 646                         prev = NULL;
 647                         spin_unlock(&ctl->tree_lock);
 648                         goto again;
 649                 }
 650 next:
 651                 prev = e;
 652         }
 653         spin_unlock(&ctl->tree_lock);
 654 }
 655
 656 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 657                                    struct btrfs_free_space_ctl *ctl,
 658                                    struct btrfs_path *path, u64 offset)
 659 {
 660         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 661         struct btrfs_free_space_header *header;
 662         struct extent_buffer *leaf;
 663         struct btrfs_io_ctl io_ctl;
 664         struct btrfs_key key;
 665         struct btrfs_free_space *e, *n;
 666         LIST_HEAD(bitmaps);
 667         u64 num_entries;
 668         u64 num_bitmaps;
 669         u64 generation;
 670         u8 type;
 671         int ret = 0;
 672
 673         /* Nothing in the space cache, goodbye */
 674         if (!i_size_read(inode))
 675                 return 0;
 676
 677         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 678         key.offset = offset;
 679         key.type = 0;
 680
 681         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 682         if (ret < 0)
 683                 return 0;
 684         else if (ret > 0) {
 685                 btrfs_release_path(path);
 686                 return 0;
 687         }
 688
 689         ret = -1;
 690
 691         leaf = path->nodes[0];
 692         header = btrfs_item_ptr(leaf, path->slots[0],
 693                                 struct btrfs_free_space_header);
 694         num_entries = btrfs_free_space_entries(leaf, header);
 695         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 696         generation = btrfs_free_space_generation(leaf, header);
 697         btrfs_release_path(path);
 698
 699         if (!BTRFS_I(inode)->generation) {
 700                 btrfs_info(fs_info,
 701                            "the free space cache file (%llu) is invalid, skip it",
 702                            offset);
 703                 return 0;
 704         }
 705
 706         if (BTRFS_I(inode)->generation != generation) {
 707                 btrfs_err(fs_info,
 708                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 709                           BTRFS_I(inode)->generation, generation);
 710                 return 0;
 711         }
 712
 713         if (!num_entries)
 714                 return 0;
 715
 716         ret = io_ctl_init(&io_ctl, inode, 0);
 717         if (ret)
 718                 return ret;
 719
 720         readahead_cache(inode);
 721
 722         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 723         if (ret)
 724                 goto out;
 725
 726         ret = io_ctl_check_crc(&io_ctl, 0);
 727         if (ret)
 728                 goto free_cache;
 729
 730         ret = io_ctl_check_generation(&io_ctl, generation);
 731         if (ret)
 732                 goto free_cache;
 733
 734         while (num_entries) {
 735                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 736                                       GFP_NOFS);
 737                 if (!e)
 738                         goto free_cache;
 739
 740                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 741                 if (ret) {
 742                         kmem_cache_free(btrfs_free_space_cachep, e);
 743                         goto free_cache;
 744                 }
 745
 746                 if (!e->bytes) {
 747                         kmem_cache_free(btrfs_free_space_cachep, e);
 748                         goto free_cache;
 749                 }
 750
 751                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 752                         spin_lock(&ctl->tree_lock);
 753                         ret = link_free_space(ctl, e);
 754                         spin_unlock(&ctl->tree_lock);
 755                         if (ret) {
 756                                 btrfs_err(fs_info,
 757                                         "Duplicate entries in free space cache, dumping");
 758                                 kmem_cache_free(btrfs_free_space_cachep, e);
 759                                 goto free_cache;
 760                         }
 761                 } else {
 762                         ASSERT(num_bitmaps);
 763                         num_bitmaps--;
 764                         e->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
 765                         if (!e->bitmap) {
 766                                 kmem_cache_free(
 767                                         btrfs_free_space_cachep, e);
 768                                 goto free_cache;
 769                         }
 770                         spin_lock(&ctl->tree_lock);
 771                         ret = link_free_space(ctl, e);
 772                         ctl->total_bitmaps++;
 773                         ctl->op->recalc_thresholds(ctl);
 774                         spin_unlock(&ctl->tree_lock);
 775                         if (ret) {
 776                                 btrfs_err(fs_info,
 777                                         "Duplicate entries in free space cache, dumping");
 778                                 kmem_cache_free(btrfs_free_space_cachep, e);
 779                                 goto free_cache;
 780                         }
 781                         list_add_tail(&e->list, &bitmaps);
 782                 }
 783
 784                 num_entries--;
 785         }
 786
 787         io_ctl_unmap_page(&io_ctl);
 788
 789         /*
 790          * We add the bitmaps at the end of the entries in order that
 791          * the bitmap entries are added to the cache.
 792          */
 793         list_for_each_entry_safe(e, n, &bitmaps, list) {
 794                 list_del_init(&e->list);
 795                 ret = io_ctl_read_bitmap(&io_ctl, e);
 796                 if (ret)
 797                         goto free_cache;
 798         }
 799
 800         io_ctl_drop_pages(&io_ctl);
 801         merge_space_tree(ctl);
 802         ret = 1;
 803 out:
 804         io_ctl_free(&io_ctl);
 805         return ret;
 806 free_cache:
 807         io_ctl_drop_pages(&io_ctl);
 808         __btrfs_remove_free_space_cache(ctl);
 809         goto out;
 810 }
 811
 812 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 813                           struct btrfs_block_group_cache *block_group)
 814 {
 815         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 816         struct inode *inode;
 817         struct btrfs_path *path;
 818         int ret = 0;
 819         bool matched;
 820         u64 used = btrfs_block_group_used(&block_group->item);
 821
 822         /*
 823          * If this block group has been marked to be cleared for one reason or
 824          * another then we can't trust the on disk cache, so just return.
 825          */
 826         spin_lock(&block_group->lock);
 827         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 828                 spin_unlock(&block_group->lock);
 829                 return 0;
 830         }
 831         spin_unlock(&block_group->lock);
 832
 833         path = btrfs_alloc_path();
 834         if (!path)
 835                 return 0;
 836         path->search_commit_root = 1;
 837         path->skip_locking = 1;
 838
 839         inode = lookup_free_space_inode(fs_info, block_group, path);
 840         if (IS_ERR(inode)) {
 841                 btrfs_free_path(path);
 842                 return 0;
 843         }
 844
 845         /* We may have converted the inode and made the cache invalid. */
 846         spin_lock(&block_group->lock);
 847         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 848                 spin_unlock(&block_group->lock);
 849                 btrfs_free_path(path);
 850                 goto out;
 851         }
 852         spin_unlock(&block_group->lock);
 853
 854         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 855                                       path, block_group->key.objectid);
 856         btrfs_free_path(path);
 857         if (ret <= 0)
 858                 goto out;
 859
 860         spin_lock(&ctl->tree_lock);
 861         matched = (ctl->free_space == (block_group->key.offset - used -
 862                                        block_group->bytes_super));
 863         spin_unlock(&ctl->tree_lock);
 864
 865         if (!matched) {
 866                 __btrfs_remove_free_space_cache(ctl);
 867                 btrfs_warn(fs_info,
 868                            "block group %llu has wrong amount of free space",
 869                            block_group->key.objectid);
 870                 ret = -1;
 871         }
 872 out:
 873         if (ret < 0) {
 874                 /* This cache is bogus, make sure it gets cleared */
 875                 spin_lock(&block_group->lock);
 876                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 877                 spin_unlock(&block_group->lock);
 878                 ret = 0;
 879
 880                 btrfs_warn(fs_info,
 881                            "failed to load free space cache for block group %llu, rebuilding it now",
 882                            block_group->key.objectid);
 883         }
 884
 885         iput(inode);
 886         return ret;
 887 }
 888
 889 static noinline_for_stack
 890 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
 891                               struct btrfs_free_space_ctl *ctl,
 892                               struct btrfs_block_group_cache *block_group,
 893                               int *entries, int *bitmaps,
 894                               struct list_head *bitmap_list)
 895 {
 896         int ret;
 897         struct btrfs_free_cluster *cluster = NULL;
 898         struct btrfs_free_cluster *cluster_locked = NULL;
 899         struct rb_node *node = rb_first(&ctl->free_space_offset);
 900         struct btrfs_trim_range *trim_entry;
 901
 902         /* Get the cluster for this block_group if it exists */
 903         if (block_group && !list_empty(&block_group->cluster_list)) {
 904                 cluster = list_entry(block_group->cluster_list.next,
 905                                      struct btrfs_free_cluster,
 906                                      block_group_list);
 907         }
 908
 909         if (!node && cluster) {
 910                 cluster_locked = cluster;
 911                 spin_lock(&cluster_locked->lock);
 912                 node = rb_first(&cluster->root);
 913                 cluster = NULL;
 914         }
 915
 916         /* Write out the extent entries */
 917         while (node) {
 918                 struct btrfs_free_space *e;
 919
 920                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 921                 *entries += 1;
 922
 923                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
 924                                        e->bitmap);
 925                 if (ret)
 926                         goto fail;
 927
 928                 if (e->bitmap) {
 929                         list_add_tail(&e->list, bitmap_list);
 930                         *bitmaps += 1;
 931                 }
 932                 node = rb_next(node);
 933                 if (!node && cluster) {
 934                         node = rb_first(&cluster->root);
 935                         cluster_locked = cluster;
 936                         spin_lock(&cluster_locked->lock);
 937                         cluster = NULL;
 938                 }
 939         }
 940         if (cluster_locked) {
 941                 spin_unlock(&cluster_locked->lock);
 942                 cluster_locked = NULL;
 943         }
 944
 945         /*
 946          * Make sure we don't miss any range that was removed from our rbtree
 947          * because trimming is running. Otherwise after a umount+mount (or crash
 948          * after committing the transaction) we would leak free space and get
 949          * an inconsistent free space cache report from fsck.
 950          */
 951         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
 952                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
 953                                        trim_entry->bytes, NULL);
 954                 if (ret)
 955                         goto fail;
 956                 *entries += 1;
 957         }
 958
 959         return 0;
 960 fail:
 961         if (cluster_locked)
 962                 spin_unlock(&cluster_locked->lock);
 963         return -ENOSPC;
 964 }
 965
 966 static noinline_for_stack int
 967 update_cache_item(struct btrfs_trans_handle *trans,
 968                   struct btrfs_root *root,
 969                   struct inode *inode,
 970                   struct btrfs_path *path, u64 offset,
 971                   int entries, int bitmaps)
 972 {
 973         struct btrfs_key key;
 974         struct btrfs_free_space_header *header;
 975         struct extent_buffer *leaf;
 976         int ret;
 977
 978         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 979         key.offset = offset;
 980         key.type = 0;
 981
 982         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 983         if (ret < 0) {
 984                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
 985                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL);
 986                 goto fail;
 987         }
 988         leaf = path->nodes[0];
 989         if (ret > 0) {
 990                 struct btrfs_key found_key;
 991                 ASSERT(path->slots[0]);
 992                 path->slots[0]--;
 993                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
 994                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
 995                     found_key.offset != offset) {
 996                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
 997                                          inode->i_size - 1,
 998                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
 999                                          NULL);
1000                         btrfs_release_path(path);
1001                         goto fail;
1002                 }
1003         }
1004
1005         BTRFS_I(inode)->generation = trans->transid;
1006         header = btrfs_item_ptr(leaf, path->slots[0],
1007                                 struct btrfs_free_space_header);
1008         btrfs_set_free_space_entries(leaf, header, entries);
1009         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1010         btrfs_set_free_space_generation(leaf, header, trans->transid);
1011         btrfs_mark_buffer_dirty(leaf);
1012         btrfs_release_path(path);
1013
1014         return 0;
1015
1016 fail:
1017         return -1;
1018 }
1019
1020 static noinline_for_stack int
1021 write_pinned_extent_entries(struct btrfs_fs_info *fs_info,
1022                             struct btrfs_block_group_cache *block_group,
1023                             struct btrfs_io_ctl *io_ctl,
1024                             int *entries)
1025 {
1026         u64 start, extent_start, extent_end, len;
1027         struct extent_io_tree *unpin = NULL;
1028         int ret;
1029
1030         if (!block_group)
1031                 return 0;
1032
1033         /*
1034          * We want to add any pinned extents to our free space cache
1035          * so we don't leak the space
1036          *
1037          * We shouldn't have switched the pinned extents yet so this is the
1038          * right one
1039          */
1040         unpin = fs_info->pinned_extents;
1041
1042         start = block_group->key.objectid;
1043
1044         while (start < block_group->key.objectid + block_group->key.offset) {
1045                 ret = find_first_extent_bit(unpin, start,
1046                                             &extent_start, &extent_end,
1047                                             EXTENT_DIRTY, NULL);
1048                 if (ret)
1049                         return 0;
1050
1051                 /* This pinned extent is out of our range */
1052                 if (extent_start >= block_group->key.objectid +
1053                     block_group->key.offset)
1054                         return 0;
1055
1056                 extent_start = max(extent_start, start);
1057                 extent_end = min(block_group->key.objectid +
1058                                  block_group->key.offset, extent_end + 1);
1059                 len = extent_end - extent_start;
1060
1061                 *entries += 1;
1062                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1063                 if (ret)
1064                         return -ENOSPC;
1065
1066                 start = extent_end;
1067         }
1068
1069         return 0;
1070 }
1071
1072 static noinline_for_stack int
1073 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1074 {
1075         struct btrfs_free_space *entry, *next;
1076         int ret;
1077
1078         /* Write out the bitmaps */
1079         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1080                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1081                 if (ret)
1082                         return -ENOSPC;
1083                 list_del_init(&entry->list);
1084         }
1085
1086         return 0;
1087 }
1088
1089 static int flush_dirty_cache(struct inode *inode)
1090 {
1091         int ret;
1092
1093         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1094         if (ret)
1095                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1096                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL);
1097
1098         return ret;
1099 }
1100
1101 static void noinline_for_stack
1102 cleanup_bitmap_list(struct list_head *bitmap_list)
1103 {
1104         struct btrfs_free_space *entry, *next;
1105
1106         list_for_each_entry_safe(entry, next, bitmap_list, list)
1107                 list_del_init(&entry->list);
1108 }
1109
1110 static void noinline_for_stack
1111 cleanup_write_cache_enospc(struct inode *inode,
1112                            struct btrfs_io_ctl *io_ctl,
1113                            struct extent_state **cached_state)
1114 {
1115         io_ctl_drop_pages(io_ctl);
1116         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1117                              i_size_read(inode) - 1, cached_state);
1118 }
1119
1120 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1121                                  struct btrfs_trans_handle *trans,
1122                                  struct btrfs_block_group_cache *block_group,
1123                                  struct btrfs_io_ctl *io_ctl,
1124                                  struct btrfs_path *path, u64 offset)
1125 {
1126         int ret;
1127         struct inode *inode = io_ctl->inode;
1128         struct btrfs_fs_info *fs_info;
1129
1130         if (!inode)
1131                 return 0;
1132
1133         fs_info = btrfs_sb(inode->i_sb);
1134
1135         /* Flush the dirty pages in the cache file. */
1136         ret = flush_dirty_cache(inode);
1137         if (ret)
1138                 goto out;
1139
1140         /* Update the cache item to tell everyone this cache file is valid. */
1141         ret = update_cache_item(trans, root, inode, path, offset,
1142                                 io_ctl->entries, io_ctl->bitmaps);
1143 out:
1144         io_ctl_free(io_ctl);
1145         if (ret) {
1146                 invalidate_inode_pages2(inode->i_mapping);
1147                 BTRFS_I(inode)->generation = 0;
1148                 if (block_group) {
1149 #ifdef DEBUG
1150                         btrfs_err(fs_info,
1151                                   "failed to write free space cache for block group %llu",
1152                                   block_group->key.objectid);
1153 #endif
1154                 }
1155         }
1156         btrfs_update_inode(trans, root, inode);
1157
1158         if (block_group) {
1159                 /* the dirty list is protected by the dirty_bgs_lock */
1160                 spin_lock(&trans->transaction->dirty_bgs_lock);
1161
1162                 /* the disk_cache_state is protected by the block group lock */
1163                 spin_lock(&block_group->lock);
1164
1165                 /*
1166                  * only mark this as written if we didn't get put back on
1167                  * the dirty list while waiting for IO.   Otherwise our
1168                  * cache state won't be right, and we won't get written again
1169                  */
1170                 if (!ret && list_empty(&block_group->dirty_list))
1171                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1172                 else if (ret)
1173                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1174
1175                 spin_unlock(&block_group->lock);
1176                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1177                 io_ctl->inode = NULL;
1178                 iput(inode);
1179         }
1180
1181         return ret;
1182
1183 }
1184
1185 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
1186                                     struct btrfs_trans_handle *trans,
1187                                     struct btrfs_io_ctl *io_ctl,
1188                                     struct btrfs_path *path)
1189 {
1190         return __btrfs_wait_cache_io(root, trans, NULL, io_ctl, path, 0);
1191 }
1192
1193 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1194                         struct btrfs_block_group_cache *block_group,
1195                         struct btrfs_path *path)
1196 {
1197         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1198                                      block_group, &block_group->io_ctl,
1199                                      path, block_group->key.objectid);
1200 }
1201
1202 /**
1203  * __btrfs_write_out_cache - write out cached info to an inode
1204  * @root - the root the inode belongs to
1205  * @ctl - the free space cache we are going to write out
1206  * @block_group - the block_group for this cache if it belongs to a block_group
1207  * @trans - the trans handle
1208  *
1209  * This function writes out a free space cache struct to disk for quick recovery
1210  * on mount.  This will return 0 if it was successful in writing the cache out,
1211  * or an errno if it was not.
1212  */
1213 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1214                                    struct btrfs_free_space_ctl *ctl,
1215                                    struct btrfs_block_group_cache *block_group,
1216                                    struct btrfs_io_ctl *io_ctl,
1217                                    struct btrfs_trans_handle *trans)
1218 {
1219         struct btrfs_fs_info *fs_info = root->fs_info;
1220         struct extent_state *cached_state = NULL;
1221         LIST_HEAD(bitmap_list);
1222         int entries = 0;
1223         int bitmaps = 0;
1224         int ret;
1225         int must_iput = 0;
1226
1227         if (!i_size_read(inode))
1228                 return -EIO;
1229
1230         WARN_ON(io_ctl->pages);
1231         ret = io_ctl_init(io_ctl, inode, 1);
1232         if (ret)
1233                 return ret;
1234
1235         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1236                 down_write(&block_group->data_rwsem);
1237                 spin_lock(&block_group->lock);
1238                 if (block_group->delalloc_bytes) {
1239                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1240                         spin_unlock(&block_group->lock);
1241                         up_write(&block_group->data_rwsem);
1242                         BTRFS_I(inode)->generation = 0;
1243                         ret = 0;
1244                         must_iput = 1;
1245                         goto out;
1246                 }
1247                 spin_unlock(&block_group->lock);
1248         }
1249
1250         /* Lock all pages first so we can lock the extent safely. */
1251         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1252         if (ret)
1253                 goto out_unlock;
1254
1255         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1256                          &cached_state);
1257
1258         io_ctl_set_generation(io_ctl, trans->transid);
1259
1260         mutex_lock(&ctl->cache_writeout_mutex);
1261         /* Write out the extent entries in the free space cache */
1262         spin_lock(&ctl->tree_lock);
1263         ret = write_cache_extent_entries(io_ctl, ctl,
1264                                          block_group, &entries, &bitmaps,
1265                                          &bitmap_list);
1266         if (ret)
1267                 goto out_nospc_locked;
1268
1269         /*
1270          * Some spaces that are freed in the current transaction are pinned,
1271          * they will be added into free space cache after the transaction is
1272          * committed, we shouldn't lose them.
1273          *
1274          * If this changes while we are working we'll get added back to
1275          * the dirty list and redo it.  No locking needed
1276          */
1277         ret = write_pinned_extent_entries(fs_info, block_group,
1278                                           io_ctl, &entries);
1279         if (ret)
1280                 goto out_nospc_locked;
1281
1282         /*
1283          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1284          * locked while doing it because a concurrent trim can be manipulating
1285          * or freeing the bitmap.
1286          */
1287         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1288         spin_unlock(&ctl->tree_lock);
1289         mutex_unlock(&ctl->cache_writeout_mutex);
1290         if (ret)
1291                 goto out_nospc;
1292
1293         /* Zero out the rest of the pages just to make sure */
1294         io_ctl_zero_remaining_pages(io_ctl);
1295
1296         /* Everything is written out, now we dirty the pages in the file. */
1297         ret = btrfs_dirty_pages(inode, io_ctl->pages, io_ctl->num_pages, 0,
1298                                 i_size_read(inode), &cached_state);
1299         if (ret)
1300                 goto out_nospc;
1301
1302         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1303                 up_write(&block_group->data_rwsem);
1304         /*
1305          * Release the pages and unlock the extent, we will flush
1306          * them out later
1307          */
1308         io_ctl_drop_pages(io_ctl);
1309
1310         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1311                              i_size_read(inode) - 1, &cached_state);
1312
1313         /*
1314          * at this point the pages are under IO and we're happy,
1315          * The caller is responsible for waiting on them and updating the
1316          * the cache and the inode
1317          */
1318         io_ctl->entries = entries;
1319         io_ctl->bitmaps = bitmaps;
1320
1321         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1322         if (ret)
1323                 goto out;
1324
1325         return 0;
1326
1327 out:
1328         io_ctl->inode = NULL;
1329         io_ctl_free(io_ctl);
1330         if (ret) {
1331                 invalidate_inode_pages2(inode->i_mapping);
1332                 BTRFS_I(inode)->generation = 0;
1333         }
1334         btrfs_update_inode(trans, root, inode);
1335         if (must_iput)
1336                 iput(inode);
1337         return ret;
1338
1339 out_nospc_locked:
1340         cleanup_bitmap_list(&bitmap_list);
1341         spin_unlock(&ctl->tree_lock);
1342         mutex_unlock(&ctl->cache_writeout_mutex);
1343
1344 out_nospc:
1345         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1346
1347 out_unlock:
1348         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1349                 up_write(&block_group->data_rwsem);
1350
1351         goto out;
1352 }
1353
1354 int btrfs_write_out_cache(struct btrfs_fs_info *fs_info,
1355                           struct btrfs_trans_handle *trans,
1356                           struct btrfs_block_group_cache *block_group,
1357                           struct btrfs_path *path)
1358 {
1359         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1360         struct inode *inode;
1361         int ret = 0;
1362
1363         spin_lock(&block_group->lock);
1364         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1365                 spin_unlock(&block_group->lock);
1366                 return 0;
1367         }
1368         spin_unlock(&block_group->lock);
1369
1370         inode = lookup_free_space_inode(fs_info, block_group, path);
1371         if (IS_ERR(inode))
1372                 return 0;
1373
1374         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1375                                 block_group, &block_group->io_ctl, trans);
1376         if (ret) {
1377 #ifdef DEBUG
1378                 btrfs_err(fs_info,
1379                           "failed to write free space cache for block group %llu",
1380                           block_group->key.objectid);
1381 #endif
1382                 spin_lock(&block_group->lock);
1383                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1384                 spin_unlock(&block_group->lock);
1385
1386                 block_group->io_ctl.inode = NULL;
1387                 iput(inode);
1388         }
1389
1390         /*
1391          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1392          * to wait for IO and put the inode
1393          */
1394
1395         return ret;
1396 }
1397
1398 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1399                                           u64 offset)
1400 {
1401         ASSERT(offset >= bitmap_start);
1402         offset -= bitmap_start;
1403         return (unsigned long)(div_u64(offset, unit));
1404 }
1405
1406 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1407 {
1408         return (unsigned long)(div_u64(bytes, unit));
1409 }
1410
1411 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1412                                    u64 offset)
1413 {
1414         u64 bitmap_start;
1415         u64 bytes_per_bitmap;
1416
1417         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1418         bitmap_start = offset - ctl->start;
1419         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1420         bitmap_start *= bytes_per_bitmap;
1421         bitmap_start += ctl->start;
1422
1423         return bitmap_start;
1424 }
1425
1426 static int tree_insert_offset(struct rb_root *root, u64 offset,
1427                               struct rb_node *node, int bitmap)
1428 {
1429         struct rb_node **p = &root->rb_node;
1430         struct rb_node *parent = NULL;
1431         struct btrfs_free_space *info;
1432
1433         while (*p) {
1434                 parent = *p;
1435                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1436
1437                 if (offset < info->offset) {
1438                         p = &(*p)->rb_left;
1439                 } else if (offset > info->offset) {
1440                         p = &(*p)->rb_right;
1441                 } else {
1442                         /*
1443                          * we could have a bitmap entry and an extent entry
1444                          * share the same offset.  If this is the case, we want
1445                          * the extent entry to always be found first if we do a
1446                          * linear search through the tree, since we want to have
1447                          * the quickest allocation time, and allocating from an
1448                          * extent is faster than allocating from a bitmap.  So
1449                          * if we're inserting a bitmap and we find an entry at
1450                          * this offset, we want to go right, or after this entry
1451                          * logically.  If we are inserting an extent and we've
1452                          * found a bitmap, we want to go left, or before
1453                          * logically.
1454                          */
1455                         if (bitmap) {
1456                                 if (info->bitmap) {
1457                                         WARN_ON_ONCE(1);
1458                                         return -EEXIST;
1459                                 }
1460                                 p = &(*p)->rb_right;
1461                         } else {
1462                                 if (!info->bitmap) {
1463                                         WARN_ON_ONCE(1);
1464                                         return -EEXIST;
1465                                 }
1466                                 p = &(*p)->rb_left;
1467                         }
1468                 }
1469         }
1470
1471         rb_link_node(node, parent, p);
1472         rb_insert_color(node, root);
1473
1474         return 0;
1475 }
1476
1477 /*
1478  * searches the tree for the given offset.
1479  *
1480  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1481  * want a section that has at least bytes size and comes at or after the given
1482  * offset.
1483  */
1484 static struct btrfs_free_space *
1485 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1486                    u64 offset, int bitmap_only, int fuzzy)
1487 {
1488         struct rb_node *n = ctl->free_space_offset.rb_node;
1489         struct btrfs_free_space *entry, *prev = NULL;
1490
1491         /* find entry that is closest to the 'offset' */
1492         while (1) {
1493                 if (!n) {
1494                         entry = NULL;
1495                         break;
1496                 }
1497
1498                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1499                 prev = entry;
1500
1501                 if (offset < entry->offset)
1502                         n = n->rb_left;
1503                 else if (offset > entry->offset)
1504                         n = n->rb_right;
1505                 else
1506                         break;
1507         }
1508
1509         if (bitmap_only) {
1510                 if (!entry)
1511                         return NULL;
1512                 if (entry->bitmap)
1513                         return entry;
1514
1515                 /*
1516                  * bitmap entry and extent entry may share same offset,
1517                  * in that case, bitmap entry comes after extent entry.
1518                  */
1519                 n = rb_next(n);
1520                 if (!n)
1521                         return NULL;
1522                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1523                 if (entry->offset != offset)
1524                         return NULL;
1525
1526                 WARN_ON(!entry->bitmap);
1527                 return entry;
1528         } else if (entry) {
1529                 if (entry->bitmap) {
1530                         /*
1531                          * if previous extent entry covers the offset,
1532                          * we should return it instead of the bitmap entry
1533                          */
1534                         n = rb_prev(&entry->offset_index);
1535                         if (n) {
1536                                 prev = rb_entry(n, struct btrfs_free_space,
1537                                                 offset_index);
1538                                 if (!prev->bitmap &&
1539                                     prev->offset + prev->bytes > offset)
1540                                         entry = prev;
1541                         }
1542                 }
1543                 return entry;
1544         }
1545
1546         if (!prev)
1547                 return NULL;
1548
1549         /* find last entry before the 'offset' */
1550         entry = prev;
1551         if (entry->offset > offset) {
1552                 n = rb_prev(&entry->offset_index);
1553                 if (n) {
1554                         entry = rb_entry(n, struct btrfs_free_space,
1555                                         offset_index);
1556                         ASSERT(entry->offset <= offset);
1557                 } else {
1558                         if (fuzzy)
1559                                 return entry;
1560                         else
1561                                 return NULL;
1562                 }
1563         }
1564
1565         if (entry->bitmap) {
1566                 n = rb_prev(&entry->offset_index);
1567                 if (n) {
1568                         prev = rb_entry(n, struct btrfs_free_space,
1569                                         offset_index);
1570                         if (!prev->bitmap &&
1571                             prev->offset + prev->bytes > offset)
1572                                 return prev;
1573                 }
1574                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1575                         return entry;
1576         } else if (entry->offset + entry->bytes > offset)
1577                 return entry;
1578
1579         if (!fuzzy)
1580                 return NULL;
1581
1582         while (1) {
1583                 if (entry->bitmap) {
1584                         if (entry->offset + BITS_PER_BITMAP *
1585                             ctl->unit > offset)
1586                                 break;
1587                 } else {
1588                         if (entry->offset + entry->bytes > offset)
1589                                 break;
1590                 }
1591
1592                 n = rb_next(&entry->offset_index);
1593                 if (!n)
1594                         return NULL;
1595                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1596         }
1597         return entry;
1598 }
1599
1600 static inline void
1601 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1602                     struct btrfs_free_space *info)
1603 {
1604         rb_erase(&info->offset_index, &ctl->free_space_offset);
1605         ctl->free_extents--;
1606 }
1607
1608 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1609                               struct btrfs_free_space *info)
1610 {
1611         __unlink_free_space(ctl, info);
1612         ctl->free_space -= info->bytes;
1613 }
1614
1615 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1616                            struct btrfs_free_space *info)
1617 {
1618         int ret = 0;
1619
1620         ASSERT(info->bytes || info->bitmap);
1621         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1622                                  &info->offset_index, (info->bitmap != NULL));
1623         if (ret)
1624                 return ret;
1625
1626         ctl->free_space += info->bytes;
1627         ctl->free_extents++;
1628         return ret;
1629 }
1630
1631 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1632 {
1633         struct btrfs_block_group_cache *block_group = ctl->private;
1634         u64 max_bytes;
1635         u64 bitmap_bytes;
1636         u64 extent_bytes;
1637         u64 size = block_group->key.offset;
1638         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1639         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1640
1641         max_bitmaps = max_t(u64, max_bitmaps, 1);
1642
1643         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1644
1645         /*
1646          * The goal is to keep the total amount of memory used per 1gb of space
1647          * at or below 32k, so we need to adjust how much memory we allow to be
1648          * used by extent based free space tracking
1649          */
1650         if (size < SZ_1G)
1651                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1652         else
1653                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1654
1655         /*
1656          * we want to account for 1 more bitmap than what we have so we can make
1657          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1658          * we add more bitmaps.
1659          */
1660         bitmap_bytes = (ctl->total_bitmaps + 1) * ctl->unit;
1661
1662         if (bitmap_bytes >= max_bytes) {
1663                 ctl->extents_thresh = 0;
1664                 return;
1665         }
1666
1667         /*
1668          * we want the extent entry threshold to always be at most 1/2 the max
1669          * bytes we can have, or whatever is less than that.
1670          */
1671         extent_bytes = max_bytes - bitmap_bytes;
1672         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1673
1674         ctl->extents_thresh =
1675                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1676 }
1677
1678 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1679                                        struct btrfs_free_space *info,
1680                                        u64 offset, u64 bytes)
1681 {
1682         unsigned long start, count;
1683
1684         start = offset_to_bit(info->offset, ctl->unit, offset);
1685         count = bytes_to_bits(bytes, ctl->unit);
1686         ASSERT(start + count <= BITS_PER_BITMAP);
1687
1688         bitmap_clear(info->bitmap, start, count);
1689
1690         info->bytes -= bytes;
1691 }
1692
1693 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1694                               struct btrfs_free_space *info, u64 offset,
1695                               u64 bytes)
1696 {
1697         __bitmap_clear_bits(ctl, info, offset, bytes);
1698         ctl->free_space -= bytes;
1699 }
1700
1701 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1702                             struct btrfs_free_space *info, u64 offset,
1703                             u64 bytes)
1704 {
1705         unsigned long start, count;
1706
1707         start = offset_to_bit(info->offset, ctl->unit, offset);
1708         count = bytes_to_bits(bytes, ctl->unit);
1709         ASSERT(start + count <= BITS_PER_BITMAP);
1710
1711         bitmap_set(info->bitmap, start, count);
1712
1713         info->bytes += bytes;
1714         ctl->free_space += bytes;
1715 }
1716
1717 /*
1718  * If we can not find suitable extent, we will use bytes to record
1719  * the size of the max extent.
1720  */
1721 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1722                          struct btrfs_free_space *bitmap_info, u64 *offset,
1723                          u64 *bytes, bool for_alloc)
1724 {
1725         unsigned long found_bits = 0;
1726         unsigned long max_bits = 0;
1727         unsigned long bits, i;
1728         unsigned long next_zero;
1729         unsigned long extent_bits;
1730
1731         /*
1732          * Skip searching the bitmap if we don't have a contiguous section that
1733          * is large enough for this allocation.
1734          */
1735         if (for_alloc &&
1736             bitmap_info->max_extent_size &&
1737             bitmap_info->max_extent_size < *bytes) {
1738                 *bytes = bitmap_info->max_extent_size;
1739                 return -1;
1740         }
1741
1742         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1743                           max_t(u64, *offset, bitmap_info->offset));
1744         bits = bytes_to_bits(*bytes, ctl->unit);
1745
1746         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1747                 if (for_alloc && bits == 1) {
1748                         found_bits = 1;
1749                         break;
1750                 }
1751                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1752                                                BITS_PER_BITMAP, i);
1753                 extent_bits = next_zero - i;
1754                 if (extent_bits >= bits) {
1755                         found_bits = extent_bits;
1756                         break;
1757                 } else if (extent_bits > max_bits) {
1758                         max_bits = extent_bits;
1759                 }
1760                 i = next_zero;
1761         }
1762
1763         if (found_bits) {
1764                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1765                 *bytes = (u64)(found_bits) * ctl->unit;
1766                 return 0;
1767         }
1768
1769         *bytes = (u64)(max_bits) * ctl->unit;
1770         bitmap_info->max_extent_size = *bytes;
1771         return -1;
1772 }
1773
1774 /* Cache the size of the max extent in bytes */
1775 static struct btrfs_free_space *
1776 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1777                 unsigned long align, u64 *max_extent_size)
1778 {
1779         struct btrfs_free_space *entry;
1780         struct rb_node *node;
1781         u64 tmp;
1782         u64 align_off;
1783         int ret;
1784
1785         if (!ctl->free_space_offset.rb_node)
1786                 goto out;
1787
1788         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1789         if (!entry)
1790                 goto out;
1791
1792         for (node = &entry->offset_index; node; node = rb_next(node)) {
1793                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1794                 if (entry->bytes < *bytes) {
1795                         if (entry->bytes > *max_extent_size)
1796                                 *max_extent_size = entry->bytes;
1797                         continue;
1798                 }
1799
1800                 /* make sure the space returned is big enough
1801                  * to match our requested alignment
1802                  */
1803                 if (*bytes >= align) {
1804                         tmp = entry->offset - ctl->start + align - 1;
1805                         tmp = div64_u64(tmp, align);
1806                         tmp = tmp * align + ctl->start;
1807                         align_off = tmp - entry->offset;
1808                 } else {
1809                         align_off = 0;
1810                         tmp = entry->offset;
1811                 }
1812
1813                 if (entry->bytes < *bytes + align_off) {
1814                         if (entry->bytes > *max_extent_size)
1815                                 *max_extent_size = entry->bytes;
1816                         continue;
1817                 }
1818
1819                 if (entry->bitmap) {
1820                         u64 size = *bytes;
1821
1822                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1823                         if (!ret) {
1824                                 *offset = tmp;
1825                                 *bytes = size;
1826                                 return entry;
1827                         } else if (size > *max_extent_size) {
1828                                 *max_extent_size = size;
1829                         }
1830                         continue;
1831                 }
1832
1833                 *offset = tmp;
1834                 *bytes = entry->bytes - align_off;
1835                 return entry;
1836         }
1837 out:
1838         return NULL;
1839 }
1840
1841 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1842                            struct btrfs_free_space *info, u64 offset)
1843 {
1844         info->offset = offset_to_bitmap(ctl, offset);
1845         info->bytes = 0;
1846         INIT_LIST_HEAD(&info->list);
1847         link_free_space(ctl, info);
1848         ctl->total_bitmaps++;
1849
1850         ctl->op->recalc_thresholds(ctl);
1851 }
1852
1853 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1854                         struct btrfs_free_space *bitmap_info)
1855 {
1856         unlink_free_space(ctl, bitmap_info);
1857         kfree(bitmap_info->bitmap);
1858         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1859         ctl->total_bitmaps--;
1860         ctl->op->recalc_thresholds(ctl);
1861 }
1862
1863 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1864                               struct btrfs_free_space *bitmap_info,
1865                               u64 *offset, u64 *bytes)
1866 {
1867         u64 end;
1868         u64 search_start, search_bytes;
1869         int ret;
1870
1871 again:
1872         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1873
1874         /*
1875          * We need to search for bits in this bitmap.  We could only cover some
1876          * of the extent in this bitmap thanks to how we add space, so we need
1877          * to search for as much as it as we can and clear that amount, and then
1878          * go searching for the next bit.
1879          */
1880         search_start = *offset;
1881         search_bytes = ctl->unit;
1882         search_bytes = min(search_bytes, end - search_start + 1);
1883         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1884                             false);
1885         if (ret < 0 || search_start != *offset)
1886                 return -EINVAL;
1887
1888         /* We may have found more bits than what we need */
1889         search_bytes = min(search_bytes, *bytes);
1890
1891         /* Cannot clear past the end of the bitmap */
1892         search_bytes = min(search_bytes, end - search_start + 1);
1893
1894         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1895         *offset += search_bytes;
1896         *bytes -= search_bytes;
1897
1898         if (*bytes) {
1899                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1900                 if (!bitmap_info->bytes)
1901                         free_bitmap(ctl, bitmap_info);
1902
1903                 /*
1904                  * no entry after this bitmap, but we still have bytes to
1905                  * remove, so something has gone wrong.
1906                  */
1907                 if (!next)
1908                         return -EINVAL;
1909
1910                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1911                                        offset_index);
1912
1913                 /*
1914                  * if the next entry isn't a bitmap we need to return to let the
1915                  * extent stuff do its work.
1916                  */
1917                 if (!bitmap_info->bitmap)
1918                         return -EAGAIN;
1919
1920                 /*
1921                  * Ok the next item is a bitmap, but it may not actually hold
1922                  * the information for the rest of this free space stuff, so
1923                  * look for it, and if we don't find it return so we can try
1924                  * everything over again.
1925                  */
1926                 search_start = *offset;
1927                 search_bytes = ctl->unit;
1928                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1929                                     &search_bytes, false);
1930                 if (ret < 0 || search_start != *offset)
1931                         return -EAGAIN;
1932
1933                 goto again;
1934         } else if (!bitmap_info->bytes)
1935                 free_bitmap(ctl, bitmap_info);
1936
1937         return 0;
1938 }
1939
1940 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1941                                struct btrfs_free_space *info, u64 offset,
1942                                u64 bytes)
1943 {
1944         u64 bytes_to_set = 0;
1945         u64 end;
1946
1947         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1948
1949         bytes_to_set = min(end - offset, bytes);
1950
1951         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1952
1953         /*
1954          * We set some bytes, we have no idea what the max extent size is
1955          * anymore.
1956          */
1957         info->max_extent_size = 0;
1958
1959         return bytes_to_set;
1960
1961 }
1962
1963 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1964                       struct btrfs_free_space *info)
1965 {
1966         struct btrfs_block_group_cache *block_group = ctl->private;
1967         struct btrfs_fs_info *fs_info = block_group->fs_info;
1968         bool forced = false;
1969
1970 #ifdef CONFIG_BTRFS_DEBUG
1971         if (btrfs_should_fragment_free_space(block_group))
1972                 forced = true;
1973 #endif
1974
1975         /*
1976          * If we are below the extents threshold then we can add this as an
1977          * extent, and don't have to deal with the bitmap
1978          */
1979         if (!forced && ctl->free_extents < ctl->extents_thresh) {
1980                 /*
1981                  * If this block group has some small extents we don't want to
1982                  * use up all of our free slots in the cache with them, we want
1983                  * to reserve them to larger extents, however if we have plenty
1984                  * of cache left then go ahead an dadd them, no sense in adding
1985                  * the overhead of a bitmap if we don't have to.
1986                  */
1987                 if (info->bytes <= fs_info->sectorsize * 4) {
1988                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
1989                                 return false;
1990                 } else {
1991                         return false;
1992                 }
1993         }
1994
1995         /*
1996          * The original block groups from mkfs can be really small, like 8
1997          * megabytes, so don't bother with a bitmap for those entries.  However
1998          * some block groups can be smaller than what a bitmap would cover but
1999          * are still large enough that they could overflow the 32k memory limit,
2000          * so allow those block groups to still be allowed to have a bitmap
2001          * entry.
2002          */
2003         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
2004                 return false;
2005
2006         return true;
2007 }
2008
2009 static const struct btrfs_free_space_op free_space_op = {
2010         .recalc_thresholds      = recalculate_thresholds,
2011         .use_bitmap             = use_bitmap,
2012 };
2013
2014 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2015                               struct btrfs_free_space *info)
2016 {
2017         struct btrfs_free_space *bitmap_info;
2018         struct btrfs_block_group_cache *block_group = NULL;
2019         int added = 0;
2020         u64 bytes, offset, bytes_added;
2021         int ret;
2022
2023         bytes = info->bytes;
2024         offset = info->offset;
2025
2026         if (!ctl->op->use_bitmap(ctl, info))
2027                 return 0;
2028
2029         if (ctl->op == &free_space_op)
2030                 block_group = ctl->private;
2031 again:
2032         /*
2033          * Since we link bitmaps right into the cluster we need to see if we
2034          * have a cluster here, and if so and it has our bitmap we need to add
2035          * the free space to that bitmap.
2036          */
2037         if (block_group && !list_empty(&block_group->cluster_list)) {
2038                 struct btrfs_free_cluster *cluster;
2039                 struct rb_node *node;
2040                 struct btrfs_free_space *entry;
2041
2042                 cluster = list_entry(block_group->cluster_list.next,
2043                                      struct btrfs_free_cluster,
2044                                      block_group_list);
2045                 spin_lock(&cluster->lock);
2046                 node = rb_first(&cluster->root);
2047                 if (!node) {
2048                         spin_unlock(&cluster->lock);
2049                         goto no_cluster_bitmap;
2050                 }
2051
2052                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2053                 if (!entry->bitmap) {
2054                         spin_unlock(&cluster->lock);
2055                         goto no_cluster_bitmap;
2056                 }
2057
2058                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2059                         bytes_added = add_bytes_to_bitmap(ctl, entry,
2060                                                           offset, bytes);
2061                         bytes -= bytes_added;
2062                         offset += bytes_added;
2063                 }
2064                 spin_unlock(&cluster->lock);
2065                 if (!bytes) {
2066                         ret = 1;
2067                         goto out;
2068                 }
2069         }
2070
2071 no_cluster_bitmap:
2072         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2073                                          1, 0);
2074         if (!bitmap_info) {
2075                 ASSERT(added == 0);
2076                 goto new_bitmap;
2077         }
2078
2079         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2080         bytes -= bytes_added;
2081         offset += bytes_added;
2082         added = 0;
2083
2084         if (!bytes) {
2085                 ret = 1;
2086                 goto out;
2087         } else
2088                 goto again;
2089
2090 new_bitmap:
2091         if (info && info->bitmap) {
2092                 add_new_bitmap(ctl, info, offset);
2093                 added = 1;
2094                 info = NULL;
2095                 goto again;
2096         } else {
2097                 spin_unlock(&ctl->tree_lock);
2098
2099                 /* no pre-allocated info, allocate a new one */
2100                 if (!info) {
2101                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2102                                                  GFP_NOFS);
2103                         if (!info) {
2104                                 spin_lock(&ctl->tree_lock);
2105                                 ret = -ENOMEM;
2106                                 goto out;
2107                         }
2108                 }
2109
2110                 /* allocate the bitmap */
2111                 info->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
2112                 spin_lock(&ctl->tree_lock);
2113                 if (!info->bitmap) {
2114                         ret = -ENOMEM;
2115                         goto out;
2116                 }
2117                 goto again;
2118         }
2119
2120 out:
2121         if (info) {
2122                 if (info->bitmap)
2123                         kfree(info->bitmap);
2124                 kmem_cache_free(btrfs_free_space_cachep, info);
2125         }
2126
2127         return ret;
2128 }
2129
2130 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2131                           struct btrfs_free_space *info, bool update_stat)
2132 {
2133         struct btrfs_free_space *left_info;
2134         struct btrfs_free_space *right_info;
2135         bool merged = false;
2136         u64 offset = info->offset;
2137         u64 bytes = info->bytes;
2138
2139         /*
2140          * first we want to see if there is free space adjacent to the range we
2141          * are adding, if there is remove that struct and add a new one to
2142          * cover the entire range
2143          */
2144         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2145         if (right_info && rb_prev(&right_info->offset_index))
2146                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2147                                      struct btrfs_free_space, offset_index);
2148         else
2149                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2150
2151         if (right_info && !right_info->bitmap) {
2152                 if (update_stat)
2153                         unlink_free_space(ctl, right_info);
2154                 else
2155                         __unlink_free_space(ctl, right_info);
2156                 info->bytes += right_info->bytes;
2157                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2158                 merged = true;
2159         }
2160
2161         if (left_info && !left_info->bitmap &&
2162             left_info->offset + left_info->bytes == offset) {
2163                 if (update_stat)
2164                         unlink_free_space(ctl, left_info);
2165                 else
2166                         __unlink_free_space(ctl, left_info);
2167                 info->offset = left_info->offset;
2168                 info->bytes += left_info->bytes;
2169                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2170                 merged = true;
2171         }
2172
2173         return merged;
2174 }
2175
2176 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2177                                      struct btrfs_free_space *info,
2178                                      bool update_stat)
2179 {
2180         struct btrfs_free_space *bitmap;
2181         unsigned long i;
2182         unsigned long j;
2183         const u64 end = info->offset + info->bytes;
2184         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2185         u64 bytes;
2186
2187         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2188         if (!bitmap)
2189                 return false;
2190
2191         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2192         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2193         if (j == i)
2194                 return false;
2195         bytes = (j - i) * ctl->unit;
2196         info->bytes += bytes;
2197
2198         if (update_stat)
2199                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2200         else
2201                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2202
2203         if (!bitmap->bytes)
2204                 free_bitmap(ctl, bitmap);
2205
2206         return true;
2207 }
2208
2209 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2210                                        struct btrfs_free_space *info,
2211                                        bool update_stat)
2212 {
2213         struct btrfs_free_space *bitmap;
2214         u64 bitmap_offset;
2215         unsigned long i;
2216         unsigned long j;
2217         unsigned long prev_j;
2218         u64 bytes;
2219
2220         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2221         /* If we're on a boundary, try the previous logical bitmap. */
2222         if (bitmap_offset == info->offset) {
2223                 if (info->offset == 0)
2224                         return false;
2225                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2226         }
2227
2228         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2229         if (!bitmap)
2230                 return false;
2231
2232         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2233         j = 0;
2234         prev_j = (unsigned long)-1;
2235         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2236                 if (j > i)
2237                         break;
2238                 prev_j = j;
2239         }
2240         if (prev_j == i)
2241                 return false;
2242
2243         if (prev_j == (unsigned long)-1)
2244                 bytes = (i + 1) * ctl->unit;
2245         else
2246                 bytes = (i - prev_j) * ctl->unit;
2247
2248         info->offset -= bytes;
2249         info->bytes += bytes;
2250
2251         if (update_stat)
2252                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2253         else
2254                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2255
2256         if (!bitmap->bytes)
2257                 free_bitmap(ctl, bitmap);
2258
2259         return true;
2260 }
2261
2262 /*
2263  * We prefer always to allocate from extent entries, both for clustered and
2264  * non-clustered allocation requests. So when attempting to add a new extent
2265  * entry, try to see if there's adjacent free space in bitmap entries, and if
2266  * there is, migrate that space from the bitmaps to the extent.
2267  * Like this we get better chances of satisfying space allocation requests
2268  * because we attempt to satisfy them based on a single cache entry, and never
2269  * on 2 or more entries - even if the entries represent a contiguous free space
2270  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2271  * ends).
2272  */
2273 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2274                               struct btrfs_free_space *info,
2275                               bool update_stat)
2276 {
2277         /*
2278          * Only work with disconnected entries, as we can change their offset,
2279          * and must be extent entries.
2280          */
2281         ASSERT(!info->bitmap);
2282         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2283
2284         if (ctl->total_bitmaps > 0) {
2285                 bool stole_end;
2286                 bool stole_front = false;
2287
2288                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2289                 if (ctl->total_bitmaps > 0)
2290                         stole_front = steal_from_bitmap_to_front(ctl, info,
2291                                                                  update_stat);
2292
2293                 if (stole_end || stole_front)
2294                         try_merge_free_space(ctl, info, update_stat);
2295         }
2296 }
2297
2298 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2299                            struct btrfs_free_space_ctl *ctl,
2300                            u64 offset, u64 bytes)
2301 {
2302         struct btrfs_free_space *info;
2303         int ret = 0;
2304
2305         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2306         if (!info)
2307                 return -ENOMEM;
2308
2309         info->offset = offset;
2310         info->bytes = bytes;
2311         RB_CLEAR_NODE(&info->offset_index);
2312
2313         spin_lock(&ctl->tree_lock);
2314
2315         if (try_merge_free_space(ctl, info, true))
2316                 goto link;
2317
2318         /*
2319          * There was no extent directly to the left or right of this new
2320          * extent then we know we're going to have to allocate a new extent, so
2321          * before we do that see if we need to drop this into a bitmap
2322          */
2323         ret = insert_into_bitmap(ctl, info);
2324         if (ret < 0) {
2325                 goto out;
2326         } else if (ret) {
2327                 ret = 0;
2328                 goto out;
2329         }
2330 link:
2331         /*
2332          * Only steal free space from adjacent bitmaps if we're sure we're not
2333          * going to add the new free space to existing bitmap entries - because
2334          * that would mean unnecessary work that would be reverted. Therefore
2335          * attempt to steal space from bitmaps if we're adding an extent entry.
2336          */
2337         steal_from_bitmap(ctl, info, true);
2338
2339         ret = link_free_space(ctl, info);
2340         if (ret)
2341                 kmem_cache_free(btrfs_free_space_cachep, info);
2342 out:
2343         spin_unlock(&ctl->tree_lock);
2344
2345         if (ret) {
2346                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2347                 ASSERT(ret != -EEXIST);
2348         }
2349
2350         return ret;
2351 }
2352
2353 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2354                             u64 offset, u64 bytes)
2355 {
2356         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2357         struct btrfs_free_space *info;
2358         int ret;
2359         bool re_search = false;
2360
2361         spin_lock(&ctl->tree_lock);
2362
2363 again:
2364         ret = 0;
2365         if (!bytes)
2366                 goto out_lock;
2367
2368         info = tree_search_offset(ctl, offset, 0, 0);
2369         if (!info) {
2370                 /*
2371                  * oops didn't find an extent that matched the space we wanted
2372                  * to remove, look for a bitmap instead
2373                  */
2374                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2375                                           1, 0);
2376                 if (!info) {
2377                         /*
2378                          * If we found a partial bit of our free space in a
2379                          * bitmap but then couldn't find the other part this may
2380                          * be a problem, so WARN about it.
2381                          */
2382                         WARN_ON(re_search);
2383                         goto out_lock;
2384                 }
2385         }
2386
2387         re_search = false;
2388         if (!info->bitmap) {
2389                 unlink_free_space(ctl, info);
2390                 if (offset == info->offset) {
2391                         u64 to_free = min(bytes, info->bytes);
2392
2393                         info->bytes -= to_free;
2394                         info->offset += to_free;
2395                         if (info->bytes) {
2396                                 ret = link_free_space(ctl, info);
2397                                 WARN_ON(ret);
2398                         } else {
2399                                 kmem_cache_free(btrfs_free_space_cachep, info);
2400                         }
2401
2402                         offset += to_free;
2403                         bytes -= to_free;
2404                         goto again;
2405                 } else {
2406                         u64 old_end = info->bytes + info->offset;
2407
2408                         info->bytes = offset - info->offset;
2409                         ret = link_free_space(ctl, info);
2410                         WARN_ON(ret);
2411                         if (ret)
2412                                 goto out_lock;
2413
2414                         /* Not enough bytes in this entry to satisfy us */
2415                         if (old_end < offset + bytes) {
2416                                 bytes -= old_end - offset;
2417                                 offset = old_end;
2418                                 goto again;
2419                         } else if (old_end == offset + bytes) {
2420                                 /* all done */
2421                                 goto out_lock;
2422                         }
2423                         spin_unlock(&ctl->tree_lock);
2424
2425                         ret = btrfs_add_free_space(block_group, offset + bytes,
2426                                                    old_end - (offset + bytes));
2427                         WARN_ON(ret);
2428                         goto out;
2429                 }
2430         }
2431
2432         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2433         if (ret == -EAGAIN) {
2434                 re_search = true;
2435                 goto again;
2436         }
2437 out_lock:
2438         spin_unlock(&ctl->tree_lock);
2439 out:
2440         return ret;
2441 }
2442
2443 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2444                            u64 bytes)
2445 {
2446         struct btrfs_fs_info *fs_info = block_group->fs_info;
2447         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2448         struct btrfs_free_space *info;
2449         struct rb_node *n;
2450         int count = 0;
2451
2452         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2453                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2454                 if (info->bytes >= bytes && !block_group->ro)
2455                         count++;
2456                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2457                            info->offset, info->bytes,
2458                        (info->bitmap) ? "yes" : "no");
2459         }
2460         btrfs_info(fs_info, "block group has cluster?: %s",
2461                list_empty(&block_group->cluster_list) ? "no" : "yes");
2462         btrfs_info(fs_info,
2463                    "%d blocks of free space at or bigger than bytes is", count);
2464 }
2465
2466 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2467 {
2468         struct btrfs_fs_info *fs_info = block_group->fs_info;
2469         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2470
2471         spin_lock_init(&ctl->tree_lock);
2472         ctl->unit = fs_info->sectorsize;
2473         ctl->start = block_group->key.objectid;
2474         ctl->private = block_group;
2475         ctl->op = &free_space_op;
2476         INIT_LIST_HEAD(&ctl->trimming_ranges);
2477         mutex_init(&ctl->cache_writeout_mutex);
2478
2479         /*
2480          * we only want to have 32k of ram per block group for keeping
2481          * track of free space, and if we pass 1/2 of that we want to
2482          * start converting things over to using bitmaps
2483          */
2484         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2485 }
2486
2487 /*
2488  * for a given cluster, put all of its extents back into the free
2489  * space cache.  If the block group passed doesn't match the block group
2490  * pointed to by the cluster, someone else raced in and freed the
2491  * cluster already.  In that case, we just return without changing anything
2492  */
2493 static int
2494 __btrfs_return_cluster_to_free_space(
2495                              struct btrfs_block_group_cache *block_group,
2496                              struct btrfs_free_cluster *cluster)
2497 {
2498         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2499         struct btrfs_free_space *entry;
2500         struct rb_node *node;
2501
2502         spin_lock(&cluster->lock);
2503         if (cluster->block_group != block_group)
2504                 goto out;
2505
2506         cluster->block_group = NULL;
2507         cluster->window_start = 0;
2508         list_del_init(&cluster->block_group_list);
2509
2510         node = rb_first(&cluster->root);
2511         while (node) {
2512                 bool bitmap;
2513
2514                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2515                 node = rb_next(&entry->offset_index);
2516                 rb_erase(&entry->offset_index, &cluster->root);
2517                 RB_CLEAR_NODE(&entry->offset_index);
2518
2519                 bitmap = (entry->bitmap != NULL);
2520                 if (!bitmap) {
2521                         try_merge_free_space(ctl, entry, false);
2522                         steal_from_bitmap(ctl, entry, false);
2523                 }
2524                 tree_insert_offset(&ctl->free_space_offset,
2525                                    entry->offset, &entry->offset_index, bitmap);
2526         }
2527         cluster->root = RB_ROOT;
2528
2529 out:
2530         spin_unlock(&cluster->lock);
2531         btrfs_put_block_group(block_group);
2532         return 0;
2533 }
2534
2535 static void __btrfs_remove_free_space_cache_locked(
2536                                 struct btrfs_free_space_ctl *ctl)
2537 {
2538         struct btrfs_free_space *info;
2539         struct rb_node *node;
2540
2541         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2542                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2543                 if (!info->bitmap) {
2544                         unlink_free_space(ctl, info);
2545                         kmem_cache_free(btrfs_free_space_cachep, info);
2546                 } else {
2547                         free_bitmap(ctl, info);
2548                 }
2549
2550                 cond_resched_lock(&ctl->tree_lock);
2551         }
2552 }
2553
2554 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2555 {
2556         spin_lock(&ctl->tree_lock);
2557         __btrfs_remove_free_space_cache_locked(ctl);
2558         spin_unlock(&ctl->tree_lock);
2559 }
2560
2561 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2562 {
2563         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2564         struct btrfs_free_cluster *cluster;
2565         struct list_head *head;
2566
2567         spin_lock(&ctl->tree_lock);
2568         while ((head = block_group->cluster_list.next) !=
2569                &block_group->cluster_list) {
2570                 cluster = list_entry(head, struct btrfs_free_cluster,
2571                                      block_group_list);
2572
2573                 WARN_ON(cluster->block_group != block_group);
2574                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2575
2576                 cond_resched_lock(&ctl->tree_lock);
2577         }
2578         __btrfs_remove_free_space_cache_locked(ctl);
2579         spin_unlock(&ctl->tree_lock);
2580
2581 }
2582
2583 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2584                                u64 offset, u64 bytes, u64 empty_size,
2585                                u64 *max_extent_size)
2586 {
2587         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2588         struct btrfs_free_space *entry = NULL;
2589         u64 bytes_search = bytes + empty_size;
2590         u64 ret = 0;
2591         u64 align_gap = 0;
2592         u64 align_gap_len = 0;
2593
2594         spin_lock(&ctl->tree_lock);
2595         entry = find_free_space(ctl, &offset, &bytes_search,
2596                                 block_group->full_stripe_len, max_extent_size);
2597         if (!entry)
2598                 goto out;
2599
2600         ret = offset;
2601         if (entry->bitmap) {
2602                 bitmap_clear_bits(ctl, entry, offset, bytes);
2603                 if (!entry->bytes)
2604                         free_bitmap(ctl, entry);
2605         } else {
2606                 unlink_free_space(ctl, entry);
2607                 align_gap_len = offset - entry->offset;
2608                 align_gap = entry->offset;
2609
2610                 entry->offset = offset + bytes;
2611                 WARN_ON(entry->bytes < bytes + align_gap_len);
2612
2613                 entry->bytes -= bytes + align_gap_len;
2614                 if (!entry->bytes)
2615                         kmem_cache_free(btrfs_free_space_cachep, entry);
2616                 else
2617                         link_free_space(ctl, entry);
2618         }
2619 out:
2620         spin_unlock(&ctl->tree_lock);
2621
2622         if (align_gap_len)
2623                 __btrfs_add_free_space(block_group->fs_info, ctl,
2624                                        align_gap, align_gap_len);
2625         return ret;
2626 }
2627
2628 /*
2629  * given a cluster, put all of its extents back into the free space
2630  * cache.  If a block group is passed, this function will only free
2631  * a cluster that belongs to the passed block group.
2632  *
2633  * Otherwise, it'll get a reference on the block group pointed to by the
2634  * cluster and remove the cluster from it.
2635  */
2636 int btrfs_return_cluster_to_free_space(
2637                                struct btrfs_block_group_cache *block_group,
2638                                struct btrfs_free_cluster *cluster)
2639 {
2640         struct btrfs_free_space_ctl *ctl;
2641         int ret;
2642
2643         /* first, get a safe pointer to the block group */
2644         spin_lock(&cluster->lock);
2645         if (!block_group) {
2646                 block_group = cluster->block_group;
2647                 if (!block_group) {
2648                         spin_unlock(&cluster->lock);
2649                         return 0;
2650                 }
2651         } else if (cluster->block_group != block_group) {
2652                 /* someone else has already freed it don't redo their work */
2653                 spin_unlock(&cluster->lock);
2654                 return 0;
2655         }
2656         atomic_inc(&block_group->count);
2657         spin_unlock(&cluster->lock);
2658
2659         ctl = block_group->free_space_ctl;
2660
2661         /* now return any extents the cluster had on it */
2662         spin_lock(&ctl->tree_lock);
2663         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2664         spin_unlock(&ctl->tree_lock);
2665
2666         /* finally drop our ref */
2667         btrfs_put_block_group(block_group);
2668         return ret;
2669 }
2670
2671 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2672                                    struct btrfs_free_cluster *cluster,
2673                                    struct btrfs_free_space *entry,
2674                                    u64 bytes, u64 min_start,
2675                                    u64 *max_extent_size)
2676 {
2677         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2678         int err;
2679         u64 search_start = cluster->window_start;
2680         u64 search_bytes = bytes;
2681         u64 ret = 0;
2682
2683         search_start = min_start;
2684         search_bytes = bytes;
2685
2686         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2687         if (err) {
2688                 if (search_bytes > *max_extent_size)
2689                         *max_extent_size = search_bytes;
2690                 return 0;
2691         }
2692
2693         ret = search_start;
2694         __bitmap_clear_bits(ctl, entry, ret, bytes);
2695
2696         return ret;
2697 }
2698
2699 /*
2700  * given a cluster, try to allocate 'bytes' from it, returns 0
2701  * if it couldn't find anything suitably large, or a logical disk offset
2702  * if things worked out
2703  */
2704 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2705                              struct btrfs_free_cluster *cluster, u64 bytes,
2706                              u64 min_start, u64 *max_extent_size)
2707 {
2708         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2709         struct btrfs_free_space *entry = NULL;
2710         struct rb_node *node;
2711         u64 ret = 0;
2712
2713         spin_lock(&cluster->lock);
2714         if (bytes > cluster->max_size)
2715                 goto out;
2716
2717         if (cluster->block_group != block_group)
2718                 goto out;
2719
2720         node = rb_first(&cluster->root);
2721         if (!node)
2722                 goto out;
2723
2724         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2725         while (1) {
2726                 if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2727                         *max_extent_size = entry->bytes;
2728
2729                 if (entry->bytes < bytes ||
2730                     (!entry->bitmap && entry->offset < min_start)) {
2731                         node = rb_next(&entry->offset_index);
2732                         if (!node)
2733                                 break;
2734                         entry = rb_entry(node, struct btrfs_free_space,
2735                                          offset_index);
2736                         continue;
2737                 }
2738
2739                 if (entry->bitmap) {
2740                         ret = btrfs_alloc_from_bitmap(block_group,
2741                                                       cluster, entry, bytes,
2742                                                       cluster->window_start,
2743                                                       max_extent_size);
2744                         if (ret == 0) {
2745                                 node = rb_next(&entry->offset_index);
2746                                 if (!node)
2747                                         break;
2748                                 entry = rb_entry(node, struct btrfs_free_space,
2749                                                  offset_index);
2750                                 continue;
2751                         }
2752                         cluster->window_start += bytes;
2753                 } else {
2754                         ret = entry->offset;
2755
2756                         entry->offset += bytes;
2757                         entry->bytes -= bytes;
2758                 }
2759
2760                 if (entry->bytes == 0)
2761                         rb_erase(&entry->offset_index, &cluster->root);
2762                 break;
2763         }
2764 out:
2765         spin_unlock(&cluster->lock);
2766
2767         if (!ret)
2768                 return 0;
2769
2770         spin_lock(&ctl->tree_lock);
2771
2772         ctl->free_space -= bytes;
2773         if (entry->bytes == 0) {
2774                 ctl->free_extents--;
2775                 if (entry->bitmap) {
2776                         kfree(entry->bitmap);
2777                         ctl->total_bitmaps--;
2778                         ctl->op->recalc_thresholds(ctl);
2779                 }
2780                 kmem_cache_free(btrfs_free_space_cachep, entry);
2781         }
2782
2783         spin_unlock(&ctl->tree_lock);
2784
2785         return ret;
2786 }
2787
2788 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2789                                 struct btrfs_free_space *entry,
2790                                 struct btrfs_free_cluster *cluster,
2791                                 u64 offset, u64 bytes,
2792                                 u64 cont1_bytes, u64 min_bytes)
2793 {
2794         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2795         unsigned long next_zero;
2796         unsigned long i;
2797         unsigned long want_bits;
2798         unsigned long min_bits;
2799         unsigned long found_bits;
2800         unsigned long max_bits = 0;
2801         unsigned long start = 0;
2802         unsigned long total_found = 0;
2803         int ret;
2804
2805         i = offset_to_bit(entry->offset, ctl->unit,
2806                           max_t(u64, offset, entry->offset));
2807         want_bits = bytes_to_bits(bytes, ctl->unit);
2808         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2809
2810         /*
2811          * Don't bother looking for a cluster in this bitmap if it's heavily
2812          * fragmented.
2813          */
2814         if (entry->max_extent_size &&
2815             entry->max_extent_size < cont1_bytes)
2816                 return -ENOSPC;
2817 again:
2818         found_bits = 0;
2819         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2820                 next_zero = find_next_zero_bit(entry->bitmap,
2821                                                BITS_PER_BITMAP, i);
2822                 if (next_zero - i >= min_bits) {
2823                         found_bits = next_zero - i;
2824                         if (found_bits > max_bits)
2825                                 max_bits = found_bits;
2826                         break;
2827                 }
2828                 if (next_zero - i > max_bits)
2829                         max_bits = next_zero - i;
2830                 i = next_zero;
2831         }
2832
2833         if (!found_bits) {
2834                 entry->max_extent_size = (u64)max_bits * ctl->unit;
2835                 return -ENOSPC;
2836         }
2837
2838         if (!total_found) {
2839                 start = i;
2840                 cluster->max_size = 0;
2841         }
2842
2843         total_found += found_bits;
2844
2845         if (cluster->max_size < found_bits * ctl->unit)
2846                 cluster->max_size = found_bits * ctl->unit;
2847
2848         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2849                 i = next_zero + 1;
2850                 goto again;
2851         }
2852
2853         cluster->window_start = start * ctl->unit + entry->offset;
2854         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2855         ret = tree_insert_offset(&cluster->root, entry->offset,
2856                                  &entry->offset_index, 1);
2857         ASSERT(!ret); /* -EEXIST; Logic error */
2858
2859         trace_btrfs_setup_cluster(block_group, cluster,
2860                                   total_found * ctl->unit, 1);
2861         return 0;
2862 }
2863
2864 /*
2865  * This searches the block group for just extents to fill the cluster with.
2866  * Try to find a cluster with at least bytes total bytes, at least one
2867  * extent of cont1_bytes, and other clusters of at least min_bytes.
2868  */
2869 static noinline int
2870 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2871                         struct btrfs_free_cluster *cluster,
2872                         struct list_head *bitmaps, u64 offset, u64 bytes,
2873                         u64 cont1_bytes, u64 min_bytes)
2874 {
2875         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2876         struct btrfs_free_space *first = NULL;
2877         struct btrfs_free_space *entry = NULL;
2878         struct btrfs_free_space *last;
2879         struct rb_node *node;
2880         u64 window_free;
2881         u64 max_extent;
2882         u64 total_size = 0;
2883
2884         entry = tree_search_offset(ctl, offset, 0, 1);
2885         if (!entry)
2886                 return -ENOSPC;
2887
2888         /*
2889          * We don't want bitmaps, so just move along until we find a normal
2890          * extent entry.
2891          */
2892         while (entry->bitmap || entry->bytes < min_bytes) {
2893                 if (entry->bitmap && list_empty(&entry->list))
2894                         list_add_tail(&entry->list, bitmaps);
2895                 node = rb_next(&entry->offset_index);
2896                 if (!node)
2897                         return -ENOSPC;
2898                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2899         }
2900
2901         window_free = entry->bytes;
2902         max_extent = entry->bytes;
2903         first = entry;
2904         last = entry;
2905
2906         for (node = rb_next(&entry->offset_index); node;
2907              node = rb_next(&entry->offset_index)) {
2908                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2909
2910                 if (entry->bitmap) {
2911                         if (list_empty(&entry->list))
2912                                 list_add_tail(&entry->list, bitmaps);
2913                         continue;
2914                 }
2915
2916                 if (entry->bytes < min_bytes)
2917                         continue;
2918
2919                 last = entry;
2920                 window_free += entry->bytes;
2921                 if (entry->bytes > max_extent)
2922                         max_extent = entry->bytes;
2923         }
2924
2925         if (window_free < bytes || max_extent < cont1_bytes)
2926                 return -ENOSPC;
2927
2928         cluster->window_start = first->offset;
2929
2930         node = &first->offset_index;
2931
2932         /*
2933          * now we've found our entries, pull them out of the free space
2934          * cache and put them into the cluster rbtree
2935          */
2936         do {
2937                 int ret;
2938
2939                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2940                 node = rb_next(&entry->offset_index);
2941                 if (entry->bitmap || entry->bytes < min_bytes)
2942                         continue;
2943
2944                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2945                 ret = tree_insert_offset(&cluster->root, entry->offset,
2946                                          &entry->offset_index, 0);
2947                 total_size += entry->bytes;
2948                 ASSERT(!ret); /* -EEXIST; Logic error */
2949         } while (node && entry != last);
2950
2951         cluster->max_size = max_extent;
2952         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2953         return 0;
2954 }
2955
2956 /*
2957  * This specifically looks for bitmaps that may work in the cluster, we assume
2958  * that we have already failed to find extents that will work.
2959  */
2960 static noinline int
2961 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2962                      struct btrfs_free_cluster *cluster,
2963                      struct list_head *bitmaps, u64 offset, u64 bytes,
2964                      u64 cont1_bytes, u64 min_bytes)
2965 {
2966         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2967         struct btrfs_free_space *entry = NULL;
2968         int ret = -ENOSPC;
2969         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2970
2971         if (ctl->total_bitmaps == 0)
2972                 return -ENOSPC;
2973
2974         /*
2975          * The bitmap that covers offset won't be in the list unless offset
2976          * is just its start offset.
2977          */
2978         if (!list_empty(bitmaps))
2979                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2980
2981         if (!entry || entry->offset != bitmap_offset) {
2982                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2983                 if (entry && list_empty(&entry->list))
2984                         list_add(&entry->list, bitmaps);
2985         }
2986
2987         list_for_each_entry(entry, bitmaps, list) {
2988                 if (entry->bytes < bytes)
2989                         continue;
2990                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2991                                            bytes, cont1_bytes, min_bytes);
2992                 if (!ret)
2993                         return 0;
2994         }
2995
2996         /*
2997          * The bitmaps list has all the bitmaps that record free space
2998          * starting after offset, so no more search is required.
2999          */
3000         return -ENOSPC;
3001 }
3002
3003 /*
3004  * here we try to find a cluster of blocks in a block group.  The goal
3005  * is to find at least bytes+empty_size.
3006  * We might not find them all in one contiguous area.
3007  *
3008  * returns zero and sets up cluster if things worked out, otherwise
3009  * it returns -enospc
3010  */
3011 int btrfs_find_space_cluster(struct btrfs_fs_info *fs_info,
3012                              struct btrfs_block_group_cache *block_group,
3013                              struct btrfs_free_cluster *cluster,
3014                              u64 offset, u64 bytes, u64 empty_size)
3015 {
3016         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3017         struct btrfs_free_space *entry, *tmp;
3018         LIST_HEAD(bitmaps);
3019         u64 min_bytes;
3020         u64 cont1_bytes;
3021         int ret;
3022
3023         /*
3024          * Choose the minimum extent size we'll require for this
3025          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3026          * For metadata, allow allocates with smaller extents.  For
3027          * data, keep it dense.
3028          */
3029         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3030                 cont1_bytes = min_bytes = bytes + empty_size;
3031         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3032                 cont1_bytes = bytes;
3033                 min_bytes = fs_info->sectorsize;
3034         } else {
3035                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3036                 min_bytes = fs_info->sectorsize;
3037         }
3038
3039         spin_lock(&ctl->tree_lock);
3040
3041         /*
3042          * If we know we don't have enough space to make a cluster don't even
3043          * bother doing all the work to try and find one.
3044          */
3045         if (ctl->free_space < bytes) {
3046                 spin_unlock(&ctl->tree_lock);
3047                 return -ENOSPC;
3048         }
3049
3050         spin_lock(&cluster->lock);
3051
3052         /* someone already found a cluster, hooray */
3053         if (cluster->block_group) {
3054                 ret = 0;
3055                 goto out;
3056         }
3057
3058         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3059                                  min_bytes);
3060
3061         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3062                                       bytes + empty_size,
3063                                       cont1_bytes, min_bytes);
3064         if (ret)
3065                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3066                                            offset, bytes + empty_size,
3067                                            cont1_bytes, min_bytes);
3068
3069         /* Clear our temporary list */
3070         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3071                 list_del_init(&entry->list);
3072
3073         if (!ret) {
3074                 atomic_inc(&block_group->count);
3075                 list_add_tail(&cluster->block_group_list,
3076                               &block_group->cluster_list);
3077                 cluster->block_group = block_group;
3078         } else {
3079                 trace_btrfs_failed_cluster_setup(block_group);
3080         }
3081 out:
3082         spin_unlock(&cluster->lock);
3083         spin_unlock(&ctl->tree_lock);
3084
3085         return ret;
3086 }
3087
3088 /*
3089  * simple code to zero out a cluster
3090  */
3091 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3092 {
3093         spin_lock_init(&cluster->lock);
3094         spin_lock_init(&cluster->refill_lock);
3095         cluster->root = RB_ROOT;
3096         cluster->max_size = 0;
3097         cluster->fragmented = false;
3098         INIT_LIST_HEAD(&cluster->block_group_list);
3099         cluster->block_group = NULL;
3100 }
3101
3102 static int do_trimming(struct btrfs_block_group_cache *block_group,
3103                        u64 *total_trimmed, u64 start, u64 bytes,
3104                        u64 reserved_start, u64 reserved_bytes,
3105                        struct btrfs_trim_range *trim_entry)
3106 {
3107         struct btrfs_space_info *space_info = block_group->space_info;
3108         struct btrfs_fs_info *fs_info = block_group->fs_info;
3109         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3110         int ret;
3111         int update = 0;
3112         u64 trimmed = 0;
3113
3114         spin_lock(&space_info->lock);
3115         spin_lock(&block_group->lock);
3116         if (!block_group->ro) {
3117                 block_group->reserved += reserved_bytes;
3118                 space_info->bytes_reserved += reserved_bytes;
3119                 update = 1;
3120         }
3121         spin_unlock(&block_group->lock);
3122         spin_unlock(&space_info->lock);
3123
3124         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3125         if (!ret)
3126                 *total_trimmed += trimmed;
3127
3128         mutex_lock(&ctl->cache_writeout_mutex);
3129         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3130         list_del(&trim_entry->list);
3131         mutex_unlock(&ctl->cache_writeout_mutex);
3132
3133         if (update) {
3134                 spin_lock(&space_info->lock);
3135                 spin_lock(&block_group->lock);
3136                 if (block_group->ro)
3137                         space_info->bytes_readonly += reserved_bytes;
3138                 block_group->reserved -= reserved_bytes;
3139                 space_info->bytes_reserved -= reserved_bytes;
3140                 spin_unlock(&space_info->lock);
3141                 spin_unlock(&block_group->lock);
3142         }
3143
3144         return ret;
3145 }
3146
3147 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3148                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3149 {
3150         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3151         struct btrfs_free_space *entry;
3152         struct rb_node *node;
3153         int ret = 0;
3154         u64 extent_start;
3155         u64 extent_bytes;
3156         u64 bytes;
3157
3158         while (start < end) {
3159                 struct btrfs_trim_range trim_entry;
3160
3161                 mutex_lock(&ctl->cache_writeout_mutex);
3162                 spin_lock(&ctl->tree_lock);
3163
3164                 if (ctl->free_space < minlen) {
3165                         spin_unlock(&ctl->tree_lock);
3166                         mutex_unlock(&ctl->cache_writeout_mutex);
3167                         break;
3168                 }
3169
3170                 entry = tree_search_offset(ctl, start, 0, 1);
3171                 if (!entry) {
3172                         spin_unlock(&ctl->tree_lock);
3173                         mutex_unlock(&ctl->cache_writeout_mutex);
3174                         break;
3175                 }
3176
3177                 /* skip bitmaps */
3178                 while (entry->bitmap) {
3179                         node = rb_next(&entry->offset_index);
3180                         if (!node) {
3181                                 spin_unlock(&ctl->tree_lock);
3182                                 mutex_unlock(&ctl->cache_writeout_mutex);
3183                                 goto out;
3184                         }
3185                         entry = rb_entry(node, struct btrfs_free_space,
3186                                          offset_index);
3187                 }
3188
3189                 if (entry->offset >= end) {
3190                         spin_unlock(&ctl->tree_lock);
3191                         mutex_unlock(&ctl->cache_writeout_mutex);
3192                         break;
3193                 }
3194
3195                 extent_start = entry->offset;
3196                 extent_bytes = entry->bytes;
3197                 start = max(start, extent_start);
3198                 bytes = min(extent_start + extent_bytes, end) - start;
3199                 if (bytes < minlen) {
3200                         spin_unlock(&ctl->tree_lock);
3201                         mutex_unlock(&ctl->cache_writeout_mutex);
3202                         goto next;
3203                 }
3204
3205                 unlink_free_space(ctl, entry);
3206                 kmem_cache_free(btrfs_free_space_cachep, entry);
3207
3208                 spin_unlock(&ctl->tree_lock);
3209                 trim_entry.start = extent_start;
3210                 trim_entry.bytes = extent_bytes;
3211                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3212                 mutex_unlock(&ctl->cache_writeout_mutex);
3213
3214                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3215                                   extent_start, extent_bytes, &trim_entry);
3216                 if (ret)
3217                         break;
3218 next:
3219                 start += bytes;
3220
3221                 if (fatal_signal_pending(current)) {
3222                         ret = -ERESTARTSYS;
3223                         break;
3224                 }
3225
3226                 cond_resched();
3227         }
3228 out:
3229         return ret;
3230 }
3231
3232 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3233                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3234 {
3235         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3236         struct btrfs_free_space *entry;
3237         int ret = 0;
3238         int ret2;
3239         u64 bytes;
3240         u64 offset = offset_to_bitmap(ctl, start);
3241
3242         while (offset < end) {
3243                 bool next_bitmap = false;
3244                 struct btrfs_trim_range trim_entry;
3245
3246                 mutex_lock(&ctl->cache_writeout_mutex);
3247                 spin_lock(&ctl->tree_lock);
3248
3249                 if (ctl->free_space < minlen) {
3250                         spin_unlock(&ctl->tree_lock);
3251                         mutex_unlock(&ctl->cache_writeout_mutex);
3252                         break;
3253                 }
3254
3255                 entry = tree_search_offset(ctl, offset, 1, 0);
3256                 if (!entry) {
3257                         spin_unlock(&ctl->tree_lock);
3258                         mutex_unlock(&ctl->cache_writeout_mutex);
3259                         next_bitmap = true;
3260                         goto next;
3261                 }
3262
3263                 bytes = minlen;
3264                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3265                 if (ret2 || start >= end) {
3266                         spin_unlock(&ctl->tree_lock);
3267                         mutex_unlock(&ctl->cache_writeout_mutex);
3268                         next_bitmap = true;
3269                         goto next;
3270                 }
3271
3272                 bytes = min(bytes, end - start);
3273                 if (bytes < minlen) {
3274                         spin_unlock(&ctl->tree_lock);
3275                         mutex_unlock(&ctl->cache_writeout_mutex);
3276                         goto next;
3277                 }
3278
3279                 bitmap_clear_bits(ctl, entry, start, bytes);
3280                 if (entry->bytes == 0)
3281                         free_bitmap(ctl, entry);
3282
3283                 spin_unlock(&ctl->tree_lock);
3284                 trim_entry.start = start;
3285                 trim_entry.bytes = bytes;
3286                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3287                 mutex_unlock(&ctl->cache_writeout_mutex);
3288
3289                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3290                                   start, bytes, &trim_entry);
3291                 if (ret)
3292                         break;
3293 next:
3294                 if (next_bitmap) {
3295                         offset += BITS_PER_BITMAP * ctl->unit;
3296                 } else {
3297                         start += bytes;
3298                         if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3299                                 offset += BITS_PER_BITMAP * ctl->unit;
3300                 }
3301
3302                 if (fatal_signal_pending(current)) {
3303                         ret = -ERESTARTSYS;
3304                         break;
3305                 }
3306
3307                 cond_resched();
3308         }
3309
3310         return ret;
3311 }
3312
3313 void btrfs_get_block_group_trimming(struct btrfs_block_group_cache *cache)
3314 {
3315         atomic_inc(&cache->trimming);
3316 }
3317
3318 void btrfs_put_block_group_trimming(struct btrfs_block_group_cache *block_group)
3319 {
3320         struct btrfs_fs_info *fs_info = block_group->fs_info;
3321         struct extent_map_tree *em_tree;
3322         struct extent_map *em;
3323         bool cleanup;
3324
3325         spin_lock(&block_group->lock);
3326         cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3327                    block_group->removed);
3328         spin_unlock(&block_group->lock);
3329
3330         if (cleanup) {
3331                 mutex_lock(&fs_info->chunk_mutex);
3332                 em_tree = &fs_info->mapping_tree.map_tree;
3333                 write_lock(&em_tree->lock);
3334                 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3335                                            1);
3336                 BUG_ON(!em); /* logic error, can't happen */
3337                 /*
3338                  * remove_extent_mapping() will delete us from the pinned_chunks
3339                  * list, which is protected by the chunk mutex.
3340                  */
3341                 remove_extent_mapping(em_tree, em);
3342                 write_unlock(&em_tree->lock);
3343                 mutex_unlock(&fs_info->chunk_mutex);
3344
3345                 /* once for us and once for the tree */
3346                 free_extent_map(em);
3347                 free_extent_map(em);
3348
3349                 /*
3350                  * We've left one free space entry and other tasks trimming
3351                  * this block group have left 1 entry each one. Free them.
3352                  */
3353                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3354         }
3355 }
3356
3357 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3358                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3359 {
3360         int ret;
3361
3362         *trimmed = 0;
3363
3364         spin_lock(&block_group->lock);
3365         if (block_group->removed) {
3366                 spin_unlock(&block_group->lock);
3367                 return 0;
3368         }
3369         btrfs_get_block_group_trimming(block_group);
3370         spin_unlock(&block_group->lock);
3371
3372         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3373         if (ret)
3374                 goto out;
3375
3376         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3377 out:
3378         btrfs_put_block_group_trimming(block_group);
3379         return ret;
3380 }
3381
3382 /*
3383  * Find the left-most item in the cache tree, and then return the
3384  * smallest inode number in the item.
3385  *
3386  * Note: the returned inode number may not be the smallest one in
3387  * the tree, if the left-most item is a bitmap.
3388  */
3389 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3390 {
3391         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3392         struct btrfs_free_space *entry = NULL;
3393         u64 ino = 0;
3394
3395         spin_lock(&ctl->tree_lock);
3396
3397         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3398                 goto out;
3399
3400         entry = rb_entry(rb_first(&ctl->free_space_offset),
3401                          struct btrfs_free_space, offset_index);
3402
3403         if (!entry->bitmap) {
3404                 ino = entry->offset;
3405
3406                 unlink_free_space(ctl, entry);
3407                 entry->offset++;
3408                 entry->bytes--;
3409                 if (!entry->bytes)
3410                         kmem_cache_free(btrfs_free_space_cachep, entry);
3411                 else
3412                         link_free_space(ctl, entry);
3413         } else {
3414                 u64 offset = 0;
3415                 u64 count = 1;
3416                 int ret;
3417
3418                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3419                 /* Logic error; Should be empty if it can't find anything */
3420                 ASSERT(!ret);
3421
3422                 ino = offset;
3423                 bitmap_clear_bits(ctl, entry, offset, 1);
3424                 if (entry->bytes == 0)
3425                         free_bitmap(ctl, entry);
3426         }
3427 out:
3428         spin_unlock(&ctl->tree_lock);
3429
3430         return ino;
3431 }
3432
3433 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3434                                     struct btrfs_path *path)
3435 {
3436         struct inode *inode = NULL;
3437
3438         spin_lock(&root->ino_cache_lock);
3439         if (root->ino_cache_inode)
3440                 inode = igrab(root->ino_cache_inode);
3441         spin_unlock(&root->ino_cache_lock);
3442         if (inode)
3443                 return inode;
3444
3445         inode = __lookup_free_space_inode(root, path, 0);
3446         if (IS_ERR(inode))
3447                 return inode;
3448
3449         spin_lock(&root->ino_cache_lock);
3450         if (!btrfs_fs_closing(root->fs_info))
3451                 root->ino_cache_inode = igrab(inode);
3452         spin_unlock(&root->ino_cache_lock);
3453
3454         return inode;
3455 }
3456
3457 int create_free_ino_inode(struct btrfs_root *root,
3458                           struct btrfs_trans_handle *trans,
3459                           struct btrfs_path *path)
3460 {
3461         return __create_free_space_inode(root, trans, path,
3462                                          BTRFS_FREE_INO_OBJECTID, 0);
3463 }
3464
3465 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3466 {
3467         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3468         struct btrfs_path *path;
3469         struct inode *inode;
3470         int ret = 0;
3471         u64 root_gen = btrfs_root_generation(&root->root_item);
3472
3473         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3474                 return 0;
3475
3476         /*
3477          * If we're unmounting then just return, since this does a search on the
3478          * normal root and not the commit root and we could deadlock.
3479          */
3480         if (btrfs_fs_closing(fs_info))
3481                 return 0;
3482
3483         path = btrfs_alloc_path();
3484         if (!path)
3485                 return 0;
3486
3487         inode = lookup_free_ino_inode(root, path);
3488         if (IS_ERR(inode))
3489                 goto out;
3490
3491         if (root_gen != BTRFS_I(inode)->generation)
3492                 goto out_put;
3493
3494         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3495
3496         if (ret < 0)
3497                 btrfs_err(fs_info,
3498                         "failed to load free ino cache for root %llu",
3499                         root->root_key.objectid);
3500 out_put:
3501         iput(inode);
3502 out:
3503         btrfs_free_path(path);
3504         return ret;
3505 }
3506
3507 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3508                               struct btrfs_trans_handle *trans,
3509                               struct btrfs_path *path,
3510                               struct inode *inode)
3511 {
3512         struct btrfs_fs_info *fs_info = root->fs_info;
3513         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3514         int ret;
3515         struct btrfs_io_ctl io_ctl;
3516         bool release_metadata = true;
3517
3518         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3519                 return 0;
3520
3521         memset(&io_ctl, 0, sizeof(io_ctl));
3522         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl, trans);
3523         if (!ret) {
3524                 /*
3525                  * At this point writepages() didn't error out, so our metadata
3526                  * reservation is released when the writeback finishes, at
3527                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3528                  * with or without an error.
3529                  */
3530                 release_metadata = false;
3531                 ret = btrfs_wait_cache_io_root(root, trans, &io_ctl, path);
3532         }
3533
3534         if (ret) {
3535                 if (release_metadata)
3536                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3537                                         inode->i_size, true);
3538 #ifdef DEBUG
3539                 btrfs_err(fs_info,
3540                           "failed to write free ino cache for root %llu",
3541                           root->root_key.objectid);
3542 #endif
3543         }
3544
3545         return ret;
3546 }
3547
3548 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3549 /*
3550  * Use this if you need to make a bitmap or extent entry specifically, it
3551  * doesn't do any of the merging that add_free_space does, this acts a lot like
3552  * how the free space cache loading stuff works, so you can get really weird
3553  * configurations.
3554  */
3555 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3556                               u64 offset, u64 bytes, bool bitmap)
3557 {
3558         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3559         struct btrfs_free_space *info = NULL, *bitmap_info;
3560         void *map = NULL;
3561         u64 bytes_added;
3562         int ret;
3563
3564 again:
3565         if (!info) {
3566                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3567                 if (!info)
3568                         return -ENOMEM;
3569         }
3570
3571         if (!bitmap) {
3572                 spin_lock(&ctl->tree_lock);
3573                 info->offset = offset;
3574                 info->bytes = bytes;
3575                 info->max_extent_size = 0;
3576                 ret = link_free_space(ctl, info);
3577                 spin_unlock(&ctl->tree_lock);
3578                 if (ret)
3579                         kmem_cache_free(btrfs_free_space_cachep, info);
3580                 return ret;
3581         }
3582
3583         if (!map) {
3584                 map = kzalloc(PAGE_SIZE, GFP_NOFS);
3585                 if (!map) {
3586                         kmem_cache_free(btrfs_free_space_cachep, info);
3587                         return -ENOMEM;
3588                 }
3589         }
3590
3591         spin_lock(&ctl->tree_lock);
3592         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3593                                          1, 0);
3594         if (!bitmap_info) {
3595                 info->bitmap = map;
3596                 map = NULL;
3597                 add_new_bitmap(ctl, info, offset);
3598                 bitmap_info = info;
3599                 info = NULL;
3600         }
3601
3602         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3603
3604         bytes -= bytes_added;
3605         offset += bytes_added;
3606         spin_unlock(&ctl->tree_lock);
3607
3608         if (bytes)
3609                 goto again;
3610
3611         if (info)
3612                 kmem_cache_free(btrfs_free_space_cachep, info);
3613         if (map)
3614                 kfree(map);
3615         return 0;
3616 }
3617
3618 /*
3619  * Checks to see if the given range is in the free space cache.  This is really
3620  * just used to check the absence of space, so if there is free space in the
3621  * range at all we will return 1.
3622  */
3623 int test_check_exists(struct btrfs_block_group_cache *cache,
3624                       u64 offset, u64 bytes)
3625 {
3626         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3627         struct btrfs_free_space *info;
3628         int ret = 0;
3629
3630         spin_lock(&ctl->tree_lock);
3631         info = tree_search_offset(ctl, offset, 0, 0);
3632         if (!info) {
3633                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3634                                           1, 0);
3635                 if (!info)
3636                         goto out;
3637         }
3638
3639 have_info:
3640         if (info->bitmap) {
3641                 u64 bit_off, bit_bytes;
3642                 struct rb_node *n;
3643                 struct btrfs_free_space *tmp;
3644
3645                 bit_off = offset;
3646                 bit_bytes = ctl->unit;
3647                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3648                 if (!ret) {
3649                         if (bit_off == offset) {
3650                                 ret = 1;
3651                                 goto out;
3652                         } else if (bit_off > offset &&
3653                                    offset + bytes > bit_off) {
3654                                 ret = 1;
3655                                 goto out;
3656                         }
3657                 }
3658
3659                 n = rb_prev(&info->offset_index);
3660                 while (n) {
3661                         tmp = rb_entry(n, struct btrfs_free_space,
3662                                        offset_index);
3663                         if (tmp->offset + tmp->bytes < offset)
3664                                 break;
3665                         if (offset + bytes < tmp->offset) {
3666                                 n = rb_prev(&tmp->offset_index);
3667                                 continue;
3668                         }
3669                         info = tmp;
3670                         goto have_info;
3671                 }
3672
3673                 n = rb_next(&info->offset_index);
3674                 while (n) {
3675                         tmp = rb_entry(n, struct btrfs_free_space,
3676                                        offset_index);
3677                         if (offset + bytes < tmp->offset)
3678                                 break;
3679                         if (tmp->offset + tmp->bytes < offset) {
3680                                 n = rb_next(&tmp->offset_index);
3681                                 continue;
3682                         }
3683                         info = tmp;
3684                         goto have_info;
3685                 }
3686
3687                 ret = 0;
3688                 goto out;
3689         }
3690
3691         if (info->offset == offset) {
3692                 ret = 1;
3693                 goto out;
3694         }
3695
3696         if (offset > info->offset && offset < info->offset + info->bytes)
3697                 ret = 1;
3698 out:
3699         spin_unlock(&ctl->tree_lock);
3700         return ret;
3701 }
3702 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */