fs/btrfs/free-space-cache.c

   1 /*
   2  * Copyright (C) 2008 Red Hat.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/pagemap.h>
  20 #include <linux/sched.h>
  21 #include <linux/slab.h>
  22 #include <linux/math64.h>
  23 #include <linux/ratelimit.h>
  24 #include "ctree.h"
  25 #include "free-space-cache.h"
  26 #include "transaction.h"
  27 #include "disk-io.h"
  28 #include "extent_io.h"
  29 #include "inode-map.h"
  30 #include "volumes.h"
  31
  32 #define BITS_PER_BITMAP         (PAGE_CACHE_SIZE * 8)
  33 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
  34
  35 struct btrfs_trim_range {
  36         u64 start;
  37         u64 bytes;
  38         struct list_head list;
  39 };
  40
  41 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  42                            struct btrfs_free_space *info);
  43 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  44                               struct btrfs_free_space *info);
  45
  46 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  47                                                struct btrfs_path *path,
  48                                                u64 offset)
  49 {
  50         struct btrfs_key key;
  51         struct btrfs_key location;
  52         struct btrfs_disk_key disk_key;
  53         struct btrfs_free_space_header *header;
  54         struct extent_buffer *leaf;
  55         struct inode *inode = NULL;
  56         int ret;
  57
  58         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  59         key.offset = offset;
  60         key.type = 0;
  61
  62         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  63         if (ret < 0)
  64                 return ERR_PTR(ret);
  65         if (ret > 0) {
  66                 btrfs_release_path(path);
  67                 return ERR_PTR(-ENOENT);
  68         }
  69
  70         leaf = path->nodes[0];
  71         header = btrfs_item_ptr(leaf, path->slots[0],
  72                                 struct btrfs_free_space_header);
  73         btrfs_free_space_key(leaf, header, &disk_key);
  74         btrfs_disk_key_to_cpu(&location, &disk_key);
  75         btrfs_release_path(path);
  76
  77         inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
  78         if (!inode)
  79                 return ERR_PTR(-ENOENT);
  80         if (IS_ERR(inode))
  81                 return inode;
  82         if (is_bad_inode(inode)) {
  83                 iput(inode);
  84                 return ERR_PTR(-ENOENT);
  85         }
  86
  87         mapping_set_gfp_mask(inode->i_mapping,
  88                         mapping_gfp_mask(inode->i_mapping) &
  89                         ~(GFP_NOFS & ~__GFP_HIGHMEM));
  90
  91         return inode;
  92 }
  93
  94 struct inode *lookup_free_space_inode(struct btrfs_root *root,
  95                                       struct btrfs_block_group_cache
  96                                       *block_group, struct btrfs_path *path)
  97 {
  98         struct inode *inode = NULL;
  99         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 100
 101         spin_lock(&block_group->lock);
 102         if (block_group->inode)
 103                 inode = igrab(block_group->inode);
 104         spin_unlock(&block_group->lock);
 105         if (inode)
 106                 return inode;
 107
 108         inode = __lookup_free_space_inode(root, path,
 109                                           block_group->key.objectid);
 110         if (IS_ERR(inode))
 111                 return inode;
 112
 113         spin_lock(&block_group->lock);
 114         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 115                 btrfs_info(root->fs_info,
 116                         "Old style space inode found, converting.");
 117                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 118                         BTRFS_INODE_NODATACOW;
 119                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 120         }
 121
 122         if (!block_group->iref) {
 123                 block_group->inode = igrab(inode);
 124                 block_group->iref = 1;
 125         }
 126         spin_unlock(&block_group->lock);
 127
 128         return inode;
 129 }
 130
 131 static int __create_free_space_inode(struct btrfs_root *root,
 132                                      struct btrfs_trans_handle *trans,
 133                                      struct btrfs_path *path,
 134                                      u64 ino, u64 offset)
 135 {
 136         struct btrfs_key key;
 137         struct btrfs_disk_key disk_key;
 138         struct btrfs_free_space_header *header;
 139         struct btrfs_inode_item *inode_item;
 140         struct extent_buffer *leaf;
 141         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 142         int ret;
 143
 144         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 145         if (ret)
 146                 return ret;
 147
 148         /* We inline crc's for the free disk space cache */
 149         if (ino != BTRFS_FREE_INO_OBJECTID)
 150                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 151
 152         leaf = path->nodes[0];
 153         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 154                                     struct btrfs_inode_item);
 155         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 156         memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
 157                              sizeof(*inode_item));
 158         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 159         btrfs_set_inode_size(leaf, inode_item, 0);
 160         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 161         btrfs_set_inode_uid(leaf, inode_item, 0);
 162         btrfs_set_inode_gid(leaf, inode_item, 0);
 163         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 164         btrfs_set_inode_flags(leaf, inode_item, flags);
 165         btrfs_set_inode_nlink(leaf, inode_item, 1);
 166         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 167         btrfs_set_inode_block_group(leaf, inode_item, offset);
 168         btrfs_mark_buffer_dirty(leaf);
 169         btrfs_release_path(path);
 170
 171         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 172         key.offset = offset;
 173         key.type = 0;
 174         ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                       sizeof(struct btrfs_free_space_header));
 176         if (ret < 0) {
 177                 btrfs_release_path(path);
 178                 return ret;
 179         }
 180
 181         leaf = path->nodes[0];
 182         header = btrfs_item_ptr(leaf, path->slots[0],
 183                                 struct btrfs_free_space_header);
 184         memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
 185         btrfs_set_free_space_key(leaf, header, &disk_key);
 186         btrfs_mark_buffer_dirty(leaf);
 187         btrfs_release_path(path);
 188
 189         return 0;
 190 }
 191
 192 int create_free_space_inode(struct btrfs_root *root,
 193                             struct btrfs_trans_handle *trans,
 194                             struct btrfs_block_group_cache *block_group,
 195                             struct btrfs_path *path)
 196 {
 197         int ret;
 198         u64 ino;
 199
 200         ret = btrfs_find_free_objectid(root, &ino);
 201         if (ret < 0)
 202                 return ret;
 203
 204         return __create_free_space_inode(root, trans, path, ino,
 205                                          block_group->key.objectid);
 206 }
 207
 208 int btrfs_check_trunc_cache_free_space(struct btrfs_root *root,
 209                                        struct btrfs_block_rsv *rsv)
 210 {
 211         u64 needed_bytes;
 212         int ret;
 213
 214         /* 1 for slack space, 1 for updating the inode */
 215         needed_bytes = btrfs_calc_trunc_metadata_size(root, 1) +
 216                 btrfs_calc_trans_metadata_size(root, 1);
 217
 218         spin_lock(&rsv->lock);
 219         if (rsv->reserved < needed_bytes)
 220                 ret = -ENOSPC;
 221         else
 222                 ret = 0;
 223         spin_unlock(&rsv->lock);
 224         return ret;
 225 }
 226
 227 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
 228                                     struct btrfs_trans_handle *trans,
 229                                     struct btrfs_block_group_cache *block_group,
 230                                     struct inode *inode)
 231 {
 232         int ret = 0;
 233         struct btrfs_path *path = btrfs_alloc_path();
 234
 235         if (!path) {
 236                 ret = -ENOMEM;
 237                 goto fail;
 238         }
 239
 240         if (block_group) {
 241                 mutex_lock(&trans->transaction->cache_write_mutex);
 242                 if (!list_empty(&block_group->io_list)) {
 243                         list_del_init(&block_group->io_list);
 244
 245                         btrfs_wait_cache_io(root, trans, block_group,
 246                                             &block_group->io_ctl, path,
 247                                             block_group->key.objectid);
 248                         btrfs_put_block_group(block_group);
 249                 }
 250
 251                 /*
 252                  * now that we've truncated the cache away, its no longer
 253                  * setup or written
 254                  */
 255                 spin_lock(&block_group->lock);
 256                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 257                 spin_unlock(&block_group->lock);
 258         }
 259         btrfs_free_path(path);
 260
 261         btrfs_i_size_write(inode, 0);
 262         truncate_pagecache(inode, 0);
 263
 264         /*
 265          * We don't need an orphan item because truncating the free space cache
 266          * will never be split across transactions.
 267          * We don't need to check for -EAGAIN because we're a free space
 268          * cache inode
 269          */
 270         ret = btrfs_truncate_inode_items(trans, root, inode,
 271                                          0, BTRFS_EXTENT_DATA_KEY);
 272         if (ret) {
 273                 mutex_unlock(&trans->transaction->cache_write_mutex);
 274                 btrfs_abort_transaction(trans, root, ret);
 275                 return ret;
 276         }
 277
 278         ret = btrfs_update_inode(trans, root, inode);
 279
 280         if (block_group)
 281                 mutex_unlock(&trans->transaction->cache_write_mutex);
 282
 283 fail:
 284         if (ret)
 285                 btrfs_abort_transaction(trans, root, ret);
 286
 287         return ret;
 288 }
 289
 290 static int readahead_cache(struct inode *inode)
 291 {
 292         struct file_ra_state *ra;
 293         unsigned long last_index;
 294
 295         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 296         if (!ra)
 297                 return -ENOMEM;
 298
 299         file_ra_state_init(ra, inode->i_mapping);
 300         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 301
 302         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 303
 304         kfree(ra);
 305
 306         return 0;
 307 }
 308
 309 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 310                        struct btrfs_root *root, int write)
 311 {
 312         int num_pages;
 313         int check_crcs = 0;
 314
 315         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_CACHE_SIZE);
 316
 317         if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
 318                 check_crcs = 1;
 319
 320         /* Make sure we can fit our crcs into the first page */
 321         if (write && check_crcs &&
 322             (num_pages * sizeof(u32)) >= PAGE_CACHE_SIZE)
 323                 return -ENOSPC;
 324
 325         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 326
 327         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 328         if (!io_ctl->pages)
 329                 return -ENOMEM;
 330
 331         io_ctl->num_pages = num_pages;
 332         io_ctl->root = root;
 333         io_ctl->check_crcs = check_crcs;
 334         io_ctl->inode = inode;
 335
 336         return 0;
 337 }
 338
 339 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 340 {
 341         kfree(io_ctl->pages);
 342         io_ctl->pages = NULL;
 343 }
 344
 345 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 346 {
 347         if (io_ctl->cur) {
 348                 io_ctl->cur = NULL;
 349                 io_ctl->orig = NULL;
 350         }
 351 }
 352
 353 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 354 {
 355         ASSERT(io_ctl->index < io_ctl->num_pages);
 356         io_ctl->page = io_ctl->pages[io_ctl->index++];
 357         io_ctl->cur = page_address(io_ctl->page);
 358         io_ctl->orig = io_ctl->cur;
 359         io_ctl->size = PAGE_CACHE_SIZE;
 360         if (clear)
 361                 memset(io_ctl->cur, 0, PAGE_CACHE_SIZE);
 362 }
 363
 364 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 365 {
 366         int i;
 367
 368         io_ctl_unmap_page(io_ctl);
 369
 370         for (i = 0; i < io_ctl->num_pages; i++) {
 371                 if (io_ctl->pages[i]) {
 372                         ClearPageChecked(io_ctl->pages[i]);
 373                         unlock_page(io_ctl->pages[i]);
 374                         page_cache_release(io_ctl->pages[i]);
 375                 }
 376         }
 377 }
 378
 379 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 380                                 int uptodate)
 381 {
 382         struct page *page;
 383         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 384         int i;
 385
 386         for (i = 0; i < io_ctl->num_pages; i++) {
 387                 page = find_or_create_page(inode->i_mapping, i, mask);
 388                 if (!page) {
 389                         io_ctl_drop_pages(io_ctl);
 390                         return -ENOMEM;
 391                 }
 392                 io_ctl->pages[i] = page;
 393                 if (uptodate && !PageUptodate(page)) {
 394                         btrfs_readpage(NULL, page);
 395                         lock_page(page);
 396                         if (!PageUptodate(page)) {
 397                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 398                                            "error reading free space cache");
 399                                 io_ctl_drop_pages(io_ctl);
 400                                 return -EIO;
 401                         }
 402                 }
 403         }
 404
 405         for (i = 0; i < io_ctl->num_pages; i++) {
 406                 clear_page_dirty_for_io(io_ctl->pages[i]);
 407                 set_page_extent_mapped(io_ctl->pages[i]);
 408         }
 409
 410         return 0;
 411 }
 412
 413 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 414 {
 415         __le64 *val;
 416
 417         io_ctl_map_page(io_ctl, 1);
 418
 419         /*
 420          * Skip the csum areas.  If we don't check crcs then we just have a
 421          * 64bit chunk at the front of the first page.
 422          */
 423         if (io_ctl->check_crcs) {
 424                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 425                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 426         } else {
 427                 io_ctl->cur += sizeof(u64);
 428                 io_ctl->size -= sizeof(u64) * 2;
 429         }
 430
 431         val = io_ctl->cur;
 432         *val = cpu_to_le64(generation);
 433         io_ctl->cur += sizeof(u64);
 434 }
 435
 436 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 437 {
 438         __le64 *gen;
 439
 440         /*
 441          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 442          * chunk at the front of the first page.
 443          */
 444         if (io_ctl->check_crcs) {
 445                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 446                 io_ctl->size -= sizeof(u64) +
 447                         (sizeof(u32) * io_ctl->num_pages);
 448         } else {
 449                 io_ctl->cur += sizeof(u64);
 450                 io_ctl->size -= sizeof(u64) * 2;
 451         }
 452
 453         gen = io_ctl->cur;
 454         if (le64_to_cpu(*gen) != generation) {
 455                 printk_ratelimited(KERN_ERR "BTRFS: space cache generation "
 456                                    "(%Lu) does not match inode (%Lu)\n", *gen,
 457                                    generation);
 458                 io_ctl_unmap_page(io_ctl);
 459                 return -EIO;
 460         }
 461         io_ctl->cur += sizeof(u64);
 462         return 0;
 463 }
 464
 465 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 466 {
 467         u32 *tmp;
 468         u32 crc = ~(u32)0;
 469         unsigned offset = 0;
 470
 471         if (!io_ctl->check_crcs) {
 472                 io_ctl_unmap_page(io_ctl);
 473                 return;
 474         }
 475
 476         if (index == 0)
 477                 offset = sizeof(u32) * io_ctl->num_pages;
 478
 479         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 480                               PAGE_CACHE_SIZE - offset);
 481         btrfs_csum_final(crc, (char *)&crc);
 482         io_ctl_unmap_page(io_ctl);
 483         tmp = page_address(io_ctl->pages[0]);
 484         tmp += index;
 485         *tmp = crc;
 486 }
 487
 488 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 489 {
 490         u32 *tmp, val;
 491         u32 crc = ~(u32)0;
 492         unsigned offset = 0;
 493
 494         if (!io_ctl->check_crcs) {
 495                 io_ctl_map_page(io_ctl, 0);
 496                 return 0;
 497         }
 498
 499         if (index == 0)
 500                 offset = sizeof(u32) * io_ctl->num_pages;
 501
 502         tmp = page_address(io_ctl->pages[0]);
 503         tmp += index;
 504         val = *tmp;
 505
 506         io_ctl_map_page(io_ctl, 0);
 507         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 508                               PAGE_CACHE_SIZE - offset);
 509         btrfs_csum_final(crc, (char *)&crc);
 510         if (val != crc) {
 511                 printk_ratelimited(KERN_ERR "BTRFS: csum mismatch on free "
 512                                    "space cache\n");
 513                 io_ctl_unmap_page(io_ctl);
 514                 return -EIO;
 515         }
 516
 517         return 0;
 518 }
 519
 520 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 521                             void *bitmap)
 522 {
 523         struct btrfs_free_space_entry *entry;
 524
 525         if (!io_ctl->cur)
 526                 return -ENOSPC;
 527
 528         entry = io_ctl->cur;
 529         entry->offset = cpu_to_le64(offset);
 530         entry->bytes = cpu_to_le64(bytes);
 531         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 532                 BTRFS_FREE_SPACE_EXTENT;
 533         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 534         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 535
 536         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 537                 return 0;
 538
 539         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 540
 541         /* No more pages to map */
 542         if (io_ctl->index >= io_ctl->num_pages)
 543                 return 0;
 544
 545         /* map the next page */
 546         io_ctl_map_page(io_ctl, 1);
 547         return 0;
 548 }
 549
 550 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 551 {
 552         if (!io_ctl->cur)
 553                 return -ENOSPC;
 554
 555         /*
 556          * If we aren't at the start of the current page, unmap this one and
 557          * map the next one if there is any left.
 558          */
 559         if (io_ctl->cur != io_ctl->orig) {
 560                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 561                 if (io_ctl->index >= io_ctl->num_pages)
 562                         return -ENOSPC;
 563                 io_ctl_map_page(io_ctl, 0);
 564         }
 565
 566         memcpy(io_ctl->cur, bitmap, PAGE_CACHE_SIZE);
 567         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 568         if (io_ctl->index < io_ctl->num_pages)
 569                 io_ctl_map_page(io_ctl, 0);
 570         return 0;
 571 }
 572
 573 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 574 {
 575         /*
 576          * If we're not on the boundary we know we've modified the page and we
 577          * need to crc the page.
 578          */
 579         if (io_ctl->cur != io_ctl->orig)
 580                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 581         else
 582                 io_ctl_unmap_page(io_ctl);
 583
 584         while (io_ctl->index < io_ctl->num_pages) {
 585                 io_ctl_map_page(io_ctl, 1);
 586                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 587         }
 588 }
 589
 590 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 591                             struct btrfs_free_space *entry, u8 *type)
 592 {
 593         struct btrfs_free_space_entry *e;
 594         int ret;
 595
 596         if (!io_ctl->cur) {
 597                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 598                 if (ret)
 599                         return ret;
 600         }
 601
 602         e = io_ctl->cur;
 603         entry->offset = le64_to_cpu(e->offset);
 604         entry->bytes = le64_to_cpu(e->bytes);
 605         *type = e->type;
 606         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 607         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 608
 609         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 610                 return 0;
 611
 612         io_ctl_unmap_page(io_ctl);
 613
 614         return 0;
 615 }
 616
 617 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 618                               struct btrfs_free_space *entry)
 619 {
 620         int ret;
 621
 622         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 623         if (ret)
 624                 return ret;
 625
 626         memcpy(entry->bitmap, io_ctl->cur, PAGE_CACHE_SIZE);
 627         io_ctl_unmap_page(io_ctl);
 628
 629         return 0;
 630 }
 631
 632 /*
 633  * Since we attach pinned extents after the fact we can have contiguous sections
 634  * of free space that are split up in entries.  This poses a problem with the
 635  * tree logging stuff since it could have allocated across what appears to be 2
 636  * entries since we would have merged the entries when adding the pinned extents
 637  * back to the free space cache.  So run through the space cache that we just
 638  * loaded and merge contiguous entries.  This will make the log replay stuff not
 639  * blow up and it will make for nicer allocator behavior.
 640  */
 641 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 642 {
 643         struct btrfs_free_space *e, *prev = NULL;
 644         struct rb_node *n;
 645
 646 again:
 647         spin_lock(&ctl->tree_lock);
 648         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 649                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 650                 if (!prev)
 651                         goto next;
 652                 if (e->bitmap || prev->bitmap)
 653                         goto next;
 654                 if (prev->offset + prev->bytes == e->offset) {
 655                         unlink_free_space(ctl, prev);
 656                         unlink_free_space(ctl, e);
 657                         prev->bytes += e->bytes;
 658                         kmem_cache_free(btrfs_free_space_cachep, e);
 659                         link_free_space(ctl, prev);
 660                         prev = NULL;
 661                         spin_unlock(&ctl->tree_lock);
 662                         goto again;
 663                 }
 664 next:
 665                 prev = e;
 666         }
 667         spin_unlock(&ctl->tree_lock);
 668 }
 669
 670 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 671                                    struct btrfs_free_space_ctl *ctl,
 672                                    struct btrfs_path *path, u64 offset)
 673 {
 674         struct btrfs_free_space_header *header;
 675         struct extent_buffer *leaf;
 676         struct btrfs_io_ctl io_ctl;
 677         struct btrfs_key key;
 678         struct btrfs_free_space *e, *n;
 679         LIST_HEAD(bitmaps);
 680         u64 num_entries;
 681         u64 num_bitmaps;
 682         u64 generation;
 683         u8 type;
 684         int ret = 0;
 685
 686         /* Nothing in the space cache, goodbye */
 687         if (!i_size_read(inode))
 688                 return 0;
 689
 690         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 691         key.offset = offset;
 692         key.type = 0;
 693
 694         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 695         if (ret < 0)
 696                 return 0;
 697         else if (ret > 0) {
 698                 btrfs_release_path(path);
 699                 return 0;
 700         }
 701
 702         ret = -1;
 703
 704         leaf = path->nodes[0];
 705         header = btrfs_item_ptr(leaf, path->slots[0],
 706                                 struct btrfs_free_space_header);
 707         num_entries = btrfs_free_space_entries(leaf, header);
 708         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 709         generation = btrfs_free_space_generation(leaf, header);
 710         btrfs_release_path(path);
 711
 712         if (!BTRFS_I(inode)->generation) {
 713                 btrfs_info(root->fs_info,
 714                            "The free space cache file (%llu) is invalid. skip it\n",
 715                            offset);
 716                 return 0;
 717         }
 718
 719         if (BTRFS_I(inode)->generation != generation) {
 720                 btrfs_err(root->fs_info,
 721                         "free space inode generation (%llu) "
 722                         "did not match free space cache generation (%llu)",
 723                         BTRFS_I(inode)->generation, generation);
 724                 return 0;
 725         }
 726
 727         if (!num_entries)
 728                 return 0;
 729
 730         ret = io_ctl_init(&io_ctl, inode, root, 0);
 731         if (ret)
 732                 return ret;
 733
 734         ret = readahead_cache(inode);
 735         if (ret)
 736                 goto out;
 737
 738         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 739         if (ret)
 740                 goto out;
 741
 742         ret = io_ctl_check_crc(&io_ctl, 0);
 743         if (ret)
 744                 goto free_cache;
 745
 746         ret = io_ctl_check_generation(&io_ctl, generation);
 747         if (ret)
 748                 goto free_cache;
 749
 750         while (num_entries) {
 751                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 752                                       GFP_NOFS);
 753                 if (!e)
 754                         goto free_cache;
 755
 756                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 757                 if (ret) {
 758                         kmem_cache_free(btrfs_free_space_cachep, e);
 759                         goto free_cache;
 760                 }
 761
 762                 if (!e->bytes) {
 763                         kmem_cache_free(btrfs_free_space_cachep, e);
 764                         goto free_cache;
 765                 }
 766
 767                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 768                         spin_lock(&ctl->tree_lock);
 769                         ret = link_free_space(ctl, e);
 770                         spin_unlock(&ctl->tree_lock);
 771                         if (ret) {
 772                                 btrfs_err(root->fs_info,
 773                                         "Duplicate entries in free space cache, dumping");
 774                                 kmem_cache_free(btrfs_free_space_cachep, e);
 775                                 goto free_cache;
 776                         }
 777                 } else {
 778                         ASSERT(num_bitmaps);
 779                         num_bitmaps--;
 780                         e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
 781                         if (!e->bitmap) {
 782                                 kmem_cache_free(
 783                                         btrfs_free_space_cachep, e);
 784                                 goto free_cache;
 785                         }
 786                         spin_lock(&ctl->tree_lock);
 787                         ret = link_free_space(ctl, e);
 788                         ctl->total_bitmaps++;
 789                         ctl->op->recalc_thresholds(ctl);
 790                         spin_unlock(&ctl->tree_lock);
 791                         if (ret) {
 792                                 btrfs_err(root->fs_info,
 793                                         "Duplicate entries in free space cache, dumping");
 794                                 kmem_cache_free(btrfs_free_space_cachep, e);
 795                                 goto free_cache;
 796                         }
 797                         list_add_tail(&e->list, &bitmaps);
 798                 }
 799
 800                 num_entries--;
 801         }
 802
 803         io_ctl_unmap_page(&io_ctl);
 804
 805         /*
 806          * We add the bitmaps at the end of the entries in order that
 807          * the bitmap entries are added to the cache.
 808          */
 809         list_for_each_entry_safe(e, n, &bitmaps, list) {
 810                 list_del_init(&e->list);
 811                 ret = io_ctl_read_bitmap(&io_ctl, e);
 812                 if (ret)
 813                         goto free_cache;
 814         }
 815
 816         io_ctl_drop_pages(&io_ctl);
 817         merge_space_tree(ctl);
 818         ret = 1;
 819 out:
 820         io_ctl_free(&io_ctl);
 821         return ret;
 822 free_cache:
 823         io_ctl_drop_pages(&io_ctl);
 824         __btrfs_remove_free_space_cache(ctl);
 825         goto out;
 826 }
 827
 828 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 829                           struct btrfs_block_group_cache *block_group)
 830 {
 831         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 832         struct btrfs_root *root = fs_info->tree_root;
 833         struct inode *inode;
 834         struct btrfs_path *path;
 835         int ret = 0;
 836         bool matched;
 837         u64 used = btrfs_block_group_used(&block_group->item);
 838
 839         /*
 840          * If this block group has been marked to be cleared for one reason or
 841          * another then we can't trust the on disk cache, so just return.
 842          */
 843         spin_lock(&block_group->lock);
 844         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 845                 spin_unlock(&block_group->lock);
 846                 return 0;
 847         }
 848         spin_unlock(&block_group->lock);
 849
 850         path = btrfs_alloc_path();
 851         if (!path)
 852                 return 0;
 853         path->search_commit_root = 1;
 854         path->skip_locking = 1;
 855
 856         inode = lookup_free_space_inode(root, block_group, path);
 857         if (IS_ERR(inode)) {
 858                 btrfs_free_path(path);
 859                 return 0;
 860         }
 861
 862         /* We may have converted the inode and made the cache invalid. */
 863         spin_lock(&block_group->lock);
 864         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 865                 spin_unlock(&block_group->lock);
 866                 btrfs_free_path(path);
 867                 goto out;
 868         }
 869         spin_unlock(&block_group->lock);
 870
 871         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 872                                       path, block_group->key.objectid);
 873         btrfs_free_path(path);
 874         if (ret <= 0)
 875                 goto out;
 876
 877         spin_lock(&ctl->tree_lock);
 878         matched = (ctl->free_space == (block_group->key.offset - used -
 879                                        block_group->bytes_super));
 880         spin_unlock(&ctl->tree_lock);
 881
 882         if (!matched) {
 883                 __btrfs_remove_free_space_cache(ctl);
 884                 btrfs_warn(fs_info, "block group %llu has wrong amount of free space",
 885                         block_group->key.objectid);
 886                 ret = -1;
 887         }
 888 out:
 889         if (ret < 0) {
 890                 /* This cache is bogus, make sure it gets cleared */
 891                 spin_lock(&block_group->lock);
 892                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 893                 spin_unlock(&block_group->lock);
 894                 ret = 0;
 895
 896                 btrfs_warn(fs_info, "failed to load free space cache for block group %llu, rebuild it now",
 897                         block_group->key.objectid);
 898         }
 899
 900         iput(inode);
 901         return ret;
 902 }
 903
 904 static noinline_for_stack
 905 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
 906                               struct btrfs_free_space_ctl *ctl,
 907                               struct btrfs_block_group_cache *block_group,
 908                               int *entries, int *bitmaps,
 909                               struct list_head *bitmap_list)
 910 {
 911         int ret;
 912         struct btrfs_free_cluster *cluster = NULL;
 913         struct btrfs_free_cluster *cluster_locked = NULL;
 914         struct rb_node *node = rb_first(&ctl->free_space_offset);
 915         struct btrfs_trim_range *trim_entry;
 916
 917         /* Get the cluster for this block_group if it exists */
 918         if (block_group && !list_empty(&block_group->cluster_list)) {
 919                 cluster = list_entry(block_group->cluster_list.next,
 920                                      struct btrfs_free_cluster,
 921                                      block_group_list);
 922         }
 923
 924         if (!node && cluster) {
 925                 cluster_locked = cluster;
 926                 spin_lock(&cluster_locked->lock);
 927                 node = rb_first(&cluster->root);
 928                 cluster = NULL;
 929         }
 930
 931         /* Write out the extent entries */
 932         while (node) {
 933                 struct btrfs_free_space *e;
 934
 935                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 936                 *entries += 1;
 937
 938                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
 939                                        e->bitmap);
 940                 if (ret)
 941                         goto fail;
 942
 943                 if (e->bitmap) {
 944                         list_add_tail(&e->list, bitmap_list);
 945                         *bitmaps += 1;
 946                 }
 947                 node = rb_next(node);
 948                 if (!node && cluster) {
 949                         node = rb_first(&cluster->root);
 950                         cluster_locked = cluster;
 951                         spin_lock(&cluster_locked->lock);
 952                         cluster = NULL;
 953                 }
 954         }
 955         if (cluster_locked) {
 956                 spin_unlock(&cluster_locked->lock);
 957                 cluster_locked = NULL;
 958         }
 959
 960         /*
 961          * Make sure we don't miss any range that was removed from our rbtree
 962          * because trimming is running. Otherwise after a umount+mount (or crash
 963          * after committing the transaction) we would leak free space and get
 964          * an inconsistent free space cache report from fsck.
 965          */
 966         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
 967                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
 968                                        trim_entry->bytes, NULL);
 969                 if (ret)
 970                         goto fail;
 971                 *entries += 1;
 972         }
 973
 974         return 0;
 975 fail:
 976         if (cluster_locked)
 977                 spin_unlock(&cluster_locked->lock);
 978         return -ENOSPC;
 979 }
 980
 981 static noinline_for_stack int
 982 update_cache_item(struct btrfs_trans_handle *trans,
 983                   struct btrfs_root *root,
 984                   struct inode *inode,
 985                   struct btrfs_path *path, u64 offset,
 986                   int entries, int bitmaps)
 987 {
 988         struct btrfs_key key;
 989         struct btrfs_free_space_header *header;
 990         struct extent_buffer *leaf;
 991         int ret;
 992
 993         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 994         key.offset = offset;
 995         key.type = 0;
 996
 997         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 998         if (ret < 0) {
 999                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1000                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1001                                  GFP_NOFS);
1002                 goto fail;
1003         }
1004         leaf = path->nodes[0];
1005         if (ret > 0) {
1006                 struct btrfs_key found_key;
1007                 ASSERT(path->slots[0]);
1008                 path->slots[0]--;
1009                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1010                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1011                     found_key.offset != offset) {
1012                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1013                                          inode->i_size - 1,
1014                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1015                                          NULL, GFP_NOFS);
1016                         btrfs_release_path(path);
1017                         goto fail;
1018                 }
1019         }
1020
1021         BTRFS_I(inode)->generation = trans->transid;
1022         header = btrfs_item_ptr(leaf, path->slots[0],
1023                                 struct btrfs_free_space_header);
1024         btrfs_set_free_space_entries(leaf, header, entries);
1025         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1026         btrfs_set_free_space_generation(leaf, header, trans->transid);
1027         btrfs_mark_buffer_dirty(leaf);
1028         btrfs_release_path(path);
1029
1030         return 0;
1031
1032 fail:
1033         return -1;
1034 }
1035
1036 static noinline_for_stack int
1037 write_pinned_extent_entries(struct btrfs_root *root,
1038                             struct btrfs_block_group_cache *block_group,
1039                             struct btrfs_io_ctl *io_ctl,
1040                             int *entries)
1041 {
1042         u64 start, extent_start, extent_end, len;
1043         struct extent_io_tree *unpin = NULL;
1044         int ret;
1045
1046         if (!block_group)
1047                 return 0;
1048
1049         /*
1050          * We want to add any pinned extents to our free space cache
1051          * so we don't leak the space
1052          *
1053          * We shouldn't have switched the pinned extents yet so this is the
1054          * right one
1055          */
1056         unpin = root->fs_info->pinned_extents;
1057
1058         start = block_group->key.objectid;
1059
1060         while (start < block_group->key.objectid + block_group->key.offset) {
1061                 ret = find_first_extent_bit(unpin, start,
1062                                             &extent_start, &extent_end,
1063                                             EXTENT_DIRTY, NULL);
1064                 if (ret)
1065                         return 0;
1066
1067                 /* This pinned extent is out of our range */
1068                 if (extent_start >= block_group->key.objectid +
1069                     block_group->key.offset)
1070                         return 0;
1071
1072                 extent_start = max(extent_start, start);
1073                 extent_end = min(block_group->key.objectid +
1074                                  block_group->key.offset, extent_end + 1);
1075                 len = extent_end - extent_start;
1076
1077                 *entries += 1;
1078                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1079                 if (ret)
1080                         return -ENOSPC;
1081
1082                 start = extent_end;
1083         }
1084
1085         return 0;
1086 }
1087
1088 static noinline_for_stack int
1089 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1090 {
1091         struct list_head *pos, *n;
1092         int ret;
1093
1094         /* Write out the bitmaps */
1095         list_for_each_safe(pos, n, bitmap_list) {
1096                 struct btrfs_free_space *entry =
1097                         list_entry(pos, struct btrfs_free_space, list);
1098
1099                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1100                 if (ret)
1101                         return -ENOSPC;
1102                 list_del_init(&entry->list);
1103         }
1104
1105         return 0;
1106 }
1107
1108 static int flush_dirty_cache(struct inode *inode)
1109 {
1110         int ret;
1111
1112         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1113         if (ret)
1114                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1115                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1116                                  GFP_NOFS);
1117
1118         return ret;
1119 }
1120
1121 static void noinline_for_stack
1122 cleanup_bitmap_list(struct list_head *bitmap_list)
1123 {
1124         struct list_head *pos, *n;
1125
1126         list_for_each_safe(pos, n, bitmap_list) {
1127                 struct btrfs_free_space *entry =
1128                         list_entry(pos, struct btrfs_free_space, list);
1129                 list_del_init(&entry->list);
1130         }
1131 }
1132
1133 static void noinline_for_stack
1134 cleanup_write_cache_enospc(struct inode *inode,
1135                            struct btrfs_io_ctl *io_ctl,
1136                            struct extent_state **cached_state,
1137                            struct list_head *bitmap_list)
1138 {
1139         io_ctl_drop_pages(io_ctl);
1140         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1141                              i_size_read(inode) - 1, cached_state,
1142                              GFP_NOFS);
1143 }
1144
1145 int btrfs_wait_cache_io(struct btrfs_root *root,
1146                         struct btrfs_trans_handle *trans,
1147                         struct btrfs_block_group_cache *block_group,
1148                         struct btrfs_io_ctl *io_ctl,
1149                         struct btrfs_path *path, u64 offset)
1150 {
1151         int ret;
1152         struct inode *inode = io_ctl->inode;
1153
1154         if (!inode)
1155                 return 0;
1156
1157         if (block_group)
1158                 root = root->fs_info->tree_root;
1159
1160         /* Flush the dirty pages in the cache file. */
1161         ret = flush_dirty_cache(inode);
1162         if (ret)
1163                 goto out;
1164
1165         /* Update the cache item to tell everyone this cache file is valid. */
1166         ret = update_cache_item(trans, root, inode, path, offset,
1167                                 io_ctl->entries, io_ctl->bitmaps);
1168 out:
1169         io_ctl_free(io_ctl);
1170         if (ret) {
1171                 invalidate_inode_pages2(inode->i_mapping);
1172                 BTRFS_I(inode)->generation = 0;
1173                 if (block_group) {
1174 #ifdef DEBUG
1175                         btrfs_err(root->fs_info,
1176                                 "failed to write free space cache for block group %llu",
1177                                 block_group->key.objectid);
1178 #endif
1179                 }
1180         }
1181         btrfs_update_inode(trans, root, inode);
1182
1183         if (block_group) {
1184                 /* the dirty list is protected by the dirty_bgs_lock */
1185                 spin_lock(&trans->transaction->dirty_bgs_lock);
1186
1187                 /* the disk_cache_state is protected by the block group lock */
1188                 spin_lock(&block_group->lock);
1189
1190                 /*
1191                  * only mark this as written if we didn't get put back on
1192                  * the dirty list while waiting for IO.   Otherwise our
1193                  * cache state won't be right, and we won't get written again
1194                  */
1195                 if (!ret && list_empty(&block_group->dirty_list))
1196                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1197                 else if (ret)
1198                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1199
1200                 spin_unlock(&block_group->lock);
1201                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1202                 io_ctl->inode = NULL;
1203                 iput(inode);
1204         }
1205
1206         return ret;
1207
1208 }
1209
1210 /**
1211  * __btrfs_write_out_cache - write out cached info to an inode
1212  * @root - the root the inode belongs to
1213  * @ctl - the free space cache we are going to write out
1214  * @block_group - the block_group for this cache if it belongs to a block_group
1215  * @trans - the trans handle
1216  * @path - the path to use
1217  * @offset - the offset for the key we'll insert
1218  *
1219  * This function writes out a free space cache struct to disk for quick recovery
1220  * on mount.  This will return 0 if it was successfull in writing the cache out,
1221  * and -1 if it was not.
1222  */
1223 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1224                                    struct btrfs_free_space_ctl *ctl,
1225                                    struct btrfs_block_group_cache *block_group,
1226                                    struct btrfs_io_ctl *io_ctl,
1227                                    struct btrfs_trans_handle *trans,
1228                                    struct btrfs_path *path, u64 offset)
1229 {
1230         struct extent_state *cached_state = NULL;
1231         LIST_HEAD(bitmap_list);
1232         int entries = 0;
1233         int bitmaps = 0;
1234         int ret;
1235         int must_iput = 0;
1236
1237         if (!i_size_read(inode))
1238                 return -1;
1239
1240         WARN_ON(io_ctl->pages);
1241         ret = io_ctl_init(io_ctl, inode, root, 1);
1242         if (ret)
1243                 return -1;
1244
1245         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1246                 down_write(&block_group->data_rwsem);
1247                 spin_lock(&block_group->lock);
1248                 if (block_group->delalloc_bytes) {
1249                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1250                         spin_unlock(&block_group->lock);
1251                         up_write(&block_group->data_rwsem);
1252                         BTRFS_I(inode)->generation = 0;
1253                         ret = 0;
1254                         must_iput = 1;
1255                         goto out;
1256                 }
1257                 spin_unlock(&block_group->lock);
1258         }
1259
1260         /* Lock all pages first so we can lock the extent safely. */
1261         io_ctl_prepare_pages(io_ctl, inode, 0);
1262
1263         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1264                          0, &cached_state);
1265
1266         io_ctl_set_generation(io_ctl, trans->transid);
1267
1268         mutex_lock(&ctl->cache_writeout_mutex);
1269         /* Write out the extent entries in the free space cache */
1270         spin_lock(&ctl->tree_lock);
1271         ret = write_cache_extent_entries(io_ctl, ctl,
1272                                          block_group, &entries, &bitmaps,
1273                                          &bitmap_list);
1274         if (ret)
1275                 goto out_nospc_locked;
1276
1277         /*
1278          * Some spaces that are freed in the current transaction are pinned,
1279          * they will be added into free space cache after the transaction is
1280          * committed, we shouldn't lose them.
1281          *
1282          * If this changes while we are working we'll get added back to
1283          * the dirty list and redo it.  No locking needed
1284          */
1285         ret = write_pinned_extent_entries(root, block_group, io_ctl, &entries);
1286         if (ret)
1287                 goto out_nospc_locked;
1288
1289         /*
1290          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1291          * locked while doing it because a concurrent trim can be manipulating
1292          * or freeing the bitmap.
1293          */
1294         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1295         spin_unlock(&ctl->tree_lock);
1296         mutex_unlock(&ctl->cache_writeout_mutex);
1297         if (ret)
1298                 goto out_nospc;
1299
1300         /* Zero out the rest of the pages just to make sure */
1301         io_ctl_zero_remaining_pages(io_ctl);
1302
1303         /* Everything is written out, now we dirty the pages in the file. */
1304         ret = btrfs_dirty_pages(root, inode, io_ctl->pages, io_ctl->num_pages,
1305                                 0, i_size_read(inode), &cached_state);
1306         if (ret)
1307                 goto out_nospc;
1308
1309         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1310                 up_write(&block_group->data_rwsem);
1311         /*
1312          * Release the pages and unlock the extent, we will flush
1313          * them out later
1314          */
1315         io_ctl_drop_pages(io_ctl);
1316
1317         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1318                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1319
1320         /*
1321          * at this point the pages are under IO and we're happy,
1322          * The caller is responsible for waiting on them and updating the
1323          * the cache and the inode
1324          */
1325         io_ctl->entries = entries;
1326         io_ctl->bitmaps = bitmaps;
1327
1328         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1329         if (ret)
1330                 goto out;
1331
1332         return 0;
1333
1334 out:
1335         io_ctl->inode = NULL;
1336         io_ctl_free(io_ctl);
1337         if (ret) {
1338                 invalidate_inode_pages2(inode->i_mapping);
1339                 BTRFS_I(inode)->generation = 0;
1340         }
1341         btrfs_update_inode(trans, root, inode);
1342         if (must_iput)
1343                 iput(inode);
1344         return ret;
1345
1346 out_nospc_locked:
1347         cleanup_bitmap_list(&bitmap_list);
1348         spin_unlock(&ctl->tree_lock);
1349         mutex_unlock(&ctl->cache_writeout_mutex);
1350
1351 out_nospc:
1352         cleanup_write_cache_enospc(inode, io_ctl, &cached_state, &bitmap_list);
1353
1354         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1355                 up_write(&block_group->data_rwsem);
1356
1357         goto out;
1358 }
1359
1360 int btrfs_write_out_cache(struct btrfs_root *root,
1361                           struct btrfs_trans_handle *trans,
1362                           struct btrfs_block_group_cache *block_group,
1363                           struct btrfs_path *path)
1364 {
1365         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1366         struct inode *inode;
1367         int ret = 0;
1368
1369         root = root->fs_info->tree_root;
1370
1371         spin_lock(&block_group->lock);
1372         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1373                 spin_unlock(&block_group->lock);
1374                 return 0;
1375         }
1376         spin_unlock(&block_group->lock);
1377
1378         inode = lookup_free_space_inode(root, block_group, path);
1379         if (IS_ERR(inode))
1380                 return 0;
1381
1382         ret = __btrfs_write_out_cache(root, inode, ctl, block_group,
1383                                       &block_group->io_ctl, trans,
1384                                       path, block_group->key.objectid);
1385         if (ret) {
1386 #ifdef DEBUG
1387                 btrfs_err(root->fs_info,
1388                         "failed to write free space cache for block group %llu",
1389                         block_group->key.objectid);
1390 #endif
1391                 spin_lock(&block_group->lock);
1392                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1393                 spin_unlock(&block_group->lock);
1394
1395                 block_group->io_ctl.inode = NULL;
1396                 iput(inode);
1397         }
1398
1399         /*
1400          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1401          * to wait for IO and put the inode
1402          */
1403
1404         return ret;
1405 }
1406
1407 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1408                                           u64 offset)
1409 {
1410         ASSERT(offset >= bitmap_start);
1411         offset -= bitmap_start;
1412         return (unsigned long)(div_u64(offset, unit));
1413 }
1414
1415 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1416 {
1417         return (unsigned long)(div_u64(bytes, unit));
1418 }
1419
1420 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1421                                    u64 offset)
1422 {
1423         u64 bitmap_start;
1424         u32 bytes_per_bitmap;
1425
1426         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1427         bitmap_start = offset - ctl->start;
1428         bitmap_start = div_u64(bitmap_start, bytes_per_bitmap);
1429         bitmap_start *= bytes_per_bitmap;
1430         bitmap_start += ctl->start;
1431
1432         return bitmap_start;
1433 }
1434
1435 static int tree_insert_offset(struct rb_root *root, u64 offset,
1436                               struct rb_node *node, int bitmap)
1437 {
1438         struct rb_node **p = &root->rb_node;
1439         struct rb_node *parent = NULL;
1440         struct btrfs_free_space *info;
1441
1442         while (*p) {
1443                 parent = *p;
1444                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1445
1446                 if (offset < info->offset) {
1447                         p = &(*p)->rb_left;
1448                 } else if (offset > info->offset) {
1449                         p = &(*p)->rb_right;
1450                 } else {
1451                         /*
1452                          * we could have a bitmap entry and an extent entry
1453                          * share the same offset.  If this is the case, we want
1454                          * the extent entry to always be found first if we do a
1455                          * linear search through the tree, since we want to have
1456                          * the quickest allocation time, and allocating from an
1457                          * extent is faster than allocating from a bitmap.  So
1458                          * if we're inserting a bitmap and we find an entry at
1459                          * this offset, we want to go right, or after this entry
1460                          * logically.  If we are inserting an extent and we've
1461                          * found a bitmap, we want to go left, or before
1462                          * logically.
1463                          */
1464                         if (bitmap) {
1465                                 if (info->bitmap) {
1466                                         WARN_ON_ONCE(1);
1467                                         return -EEXIST;
1468                                 }
1469                                 p = &(*p)->rb_right;
1470                         } else {
1471                                 if (!info->bitmap) {
1472                                         WARN_ON_ONCE(1);
1473                                         return -EEXIST;
1474                                 }
1475                                 p = &(*p)->rb_left;
1476                         }
1477                 }
1478         }
1479
1480         rb_link_node(node, parent, p);
1481         rb_insert_color(node, root);
1482
1483         return 0;
1484 }
1485
1486 /*
1487  * searches the tree for the given offset.
1488  *
1489  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1490  * want a section that has at least bytes size and comes at or after the given
1491  * offset.
1492  */
1493 static struct btrfs_free_space *
1494 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1495                    u64 offset, int bitmap_only, int fuzzy)
1496 {
1497         struct rb_node *n = ctl->free_space_offset.rb_node;
1498         struct btrfs_free_space *entry, *prev = NULL;
1499
1500         /* find entry that is closest to the 'offset' */
1501         while (1) {
1502                 if (!n) {
1503                         entry = NULL;
1504                         break;
1505                 }
1506
1507                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1508                 prev = entry;
1509
1510                 if (offset < entry->offset)
1511                         n = n->rb_left;
1512                 else if (offset > entry->offset)
1513                         n = n->rb_right;
1514                 else
1515                         break;
1516         }
1517
1518         if (bitmap_only) {
1519                 if (!entry)
1520                         return NULL;
1521                 if (entry->bitmap)
1522                         return entry;
1523
1524                 /*
1525                  * bitmap entry and extent entry may share same offset,
1526                  * in that case, bitmap entry comes after extent entry.
1527                  */
1528                 n = rb_next(n);
1529                 if (!n)
1530                         return NULL;
1531                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1532                 if (entry->offset != offset)
1533                         return NULL;
1534
1535                 WARN_ON(!entry->bitmap);
1536                 return entry;
1537         } else if (entry) {
1538                 if (entry->bitmap) {
1539                         /*
1540                          * if previous extent entry covers the offset,
1541                          * we should return it instead of the bitmap entry
1542                          */
1543                         n = rb_prev(&entry->offset_index);
1544                         if (n) {
1545                                 prev = rb_entry(n, struct btrfs_free_space,
1546                                                 offset_index);
1547                                 if (!prev->bitmap &&
1548                                     prev->offset + prev->bytes > offset)
1549                                         entry = prev;
1550                         }
1551                 }
1552                 return entry;
1553         }
1554
1555         if (!prev)
1556                 return NULL;
1557
1558         /* find last entry before the 'offset' */
1559         entry = prev;
1560         if (entry->offset > offset) {
1561                 n = rb_prev(&entry->offset_index);
1562                 if (n) {
1563                         entry = rb_entry(n, struct btrfs_free_space,
1564                                         offset_index);
1565                         ASSERT(entry->offset <= offset);
1566                 } else {
1567                         if (fuzzy)
1568                                 return entry;
1569                         else
1570                                 return NULL;
1571                 }
1572         }
1573
1574         if (entry->bitmap) {
1575                 n = rb_prev(&entry->offset_index);
1576                 if (n) {
1577                         prev = rb_entry(n, struct btrfs_free_space,
1578                                         offset_index);
1579                         if (!prev->bitmap &&
1580                             prev->offset + prev->bytes > offset)
1581                                 return prev;
1582                 }
1583                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1584                         return entry;
1585         } else if (entry->offset + entry->bytes > offset)
1586                 return entry;
1587
1588         if (!fuzzy)
1589                 return NULL;
1590
1591         while (1) {
1592                 if (entry->bitmap) {
1593                         if (entry->offset + BITS_PER_BITMAP *
1594                             ctl->unit > offset)
1595                                 break;
1596                 } else {
1597                         if (entry->offset + entry->bytes > offset)
1598                                 break;
1599                 }
1600
1601                 n = rb_next(&entry->offset_index);
1602                 if (!n)
1603                         return NULL;
1604                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1605         }
1606         return entry;
1607 }
1608
1609 static inline void
1610 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1611                     struct btrfs_free_space *info)
1612 {
1613         rb_erase(&info->offset_index, &ctl->free_space_offset);
1614         ctl->free_extents--;
1615 }
1616
1617 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1618                               struct btrfs_free_space *info)
1619 {
1620         __unlink_free_space(ctl, info);
1621         ctl->free_space -= info->bytes;
1622 }
1623
1624 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1625                            struct btrfs_free_space *info)
1626 {
1627         int ret = 0;
1628
1629         ASSERT(info->bytes || info->bitmap);
1630         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1631                                  &info->offset_index, (info->bitmap != NULL));
1632         if (ret)
1633                 return ret;
1634
1635         ctl->free_space += info->bytes;
1636         ctl->free_extents++;
1637         return ret;
1638 }
1639
1640 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1641 {
1642         struct btrfs_block_group_cache *block_group = ctl->private;
1643         u64 max_bytes;
1644         u64 bitmap_bytes;
1645         u64 extent_bytes;
1646         u64 size = block_group->key.offset;
1647         u32 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1648         u32 max_bitmaps = div_u64(size + bytes_per_bg - 1, bytes_per_bg);
1649
1650         max_bitmaps = max_t(u32, max_bitmaps, 1);
1651
1652         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1653
1654         /*
1655          * The goal is to keep the total amount of memory used per 1gb of space
1656          * at or below 32k, so we need to adjust how much memory we allow to be
1657          * used by extent based free space tracking
1658          */
1659         if (size < 1024 * 1024 * 1024)
1660                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1661         else
1662                 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1663                         div_u64(size, 1024 * 1024 * 1024);
1664
1665         /*
1666          * we want to account for 1 more bitmap than what we have so we can make
1667          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1668          * we add more bitmaps.
1669          */
1670         bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1671
1672         if (bitmap_bytes >= max_bytes) {
1673                 ctl->extents_thresh = 0;
1674                 return;
1675         }
1676
1677         /*
1678          * we want the extent entry threshold to always be at most 1/2 the max
1679          * bytes we can have, or whatever is less than that.
1680          */
1681         extent_bytes = max_bytes - bitmap_bytes;
1682         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1683
1684         ctl->extents_thresh =
1685                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1686 }
1687
1688 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1689                                        struct btrfs_free_space *info,
1690                                        u64 offset, u64 bytes)
1691 {
1692         unsigned long start, count;
1693
1694         start = offset_to_bit(info->offset, ctl->unit, offset);
1695         count = bytes_to_bits(bytes, ctl->unit);
1696         ASSERT(start + count <= BITS_PER_BITMAP);
1697
1698         bitmap_clear(info->bitmap, start, count);
1699
1700         info->bytes -= bytes;
1701 }
1702
1703 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1704                               struct btrfs_free_space *info, u64 offset,
1705                               u64 bytes)
1706 {
1707         __bitmap_clear_bits(ctl, info, offset, bytes);
1708         ctl->free_space -= bytes;
1709 }
1710
1711 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1712                             struct btrfs_free_space *info, u64 offset,
1713                             u64 bytes)
1714 {
1715         unsigned long start, count;
1716
1717         start = offset_to_bit(info->offset, ctl->unit, offset);
1718         count = bytes_to_bits(bytes, ctl->unit);
1719         ASSERT(start + count <= BITS_PER_BITMAP);
1720
1721         bitmap_set(info->bitmap, start, count);
1722
1723         info->bytes += bytes;
1724         ctl->free_space += bytes;
1725 }
1726
1727 /*
1728  * If we can not find suitable extent, we will use bytes to record
1729  * the size of the max extent.
1730  */
1731 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1732                          struct btrfs_free_space *bitmap_info, u64 *offset,
1733                          u64 *bytes)
1734 {
1735         unsigned long found_bits = 0;
1736         unsigned long max_bits = 0;
1737         unsigned long bits, i;
1738         unsigned long next_zero;
1739         unsigned long extent_bits;
1740
1741         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1742                           max_t(u64, *offset, bitmap_info->offset));
1743         bits = bytes_to_bits(*bytes, ctl->unit);
1744
1745         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1746                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1747                                                BITS_PER_BITMAP, i);
1748                 extent_bits = next_zero - i;
1749                 if (extent_bits >= bits) {
1750                         found_bits = extent_bits;
1751                         break;
1752                 } else if (extent_bits > max_bits) {
1753                         max_bits = extent_bits;
1754                 }
1755                 i = next_zero;
1756         }
1757
1758         if (found_bits) {
1759                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1760                 *bytes = (u64)(found_bits) * ctl->unit;
1761                 return 0;
1762         }
1763
1764         *bytes = (u64)(max_bits) * ctl->unit;
1765         return -1;
1766 }
1767
1768 /* Cache the size of the max extent in bytes */
1769 static struct btrfs_free_space *
1770 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1771                 unsigned long align, u64 *max_extent_size)
1772 {
1773         struct btrfs_free_space *entry;
1774         struct rb_node *node;
1775         u64 tmp;
1776         u64 align_off;
1777         int ret;
1778
1779         if (!ctl->free_space_offset.rb_node)
1780                 goto out;
1781
1782         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1783         if (!entry)
1784                 goto out;
1785
1786         for (node = &entry->offset_index; node; node = rb_next(node)) {
1787                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1788                 if (entry->bytes < *bytes) {
1789                         if (entry->bytes > *max_extent_size)
1790                                 *max_extent_size = entry->bytes;
1791                         continue;
1792                 }
1793
1794                 /* make sure the space returned is big enough
1795                  * to match our requested alignment
1796                  */
1797                 if (*bytes >= align) {
1798                         tmp = entry->offset - ctl->start + align - 1;
1799                         tmp = div64_u64(tmp, align);
1800                         tmp = tmp * align + ctl->start;
1801                         align_off = tmp - entry->offset;
1802                 } else {
1803                         align_off = 0;
1804                         tmp = entry->offset;
1805                 }
1806
1807                 if (entry->bytes < *bytes + align_off) {
1808                         if (entry->bytes > *max_extent_size)
1809                                 *max_extent_size = entry->bytes;
1810                         continue;
1811                 }
1812
1813                 if (entry->bitmap) {
1814                         u64 size = *bytes;
1815
1816                         ret = search_bitmap(ctl, entry, &tmp, &size);
1817                         if (!ret) {
1818                                 *offset = tmp;
1819                                 *bytes = size;
1820                                 return entry;
1821                         } else if (size > *max_extent_size) {
1822                                 *max_extent_size = size;
1823                         }
1824                         continue;
1825                 }
1826
1827                 *offset = tmp;
1828                 *bytes = entry->bytes - align_off;
1829                 return entry;
1830         }
1831 out:
1832         return NULL;
1833 }
1834
1835 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1836                            struct btrfs_free_space *info, u64 offset)
1837 {
1838         info->offset = offset_to_bitmap(ctl, offset);
1839         info->bytes = 0;
1840         INIT_LIST_HEAD(&info->list);
1841         link_free_space(ctl, info);
1842         ctl->total_bitmaps++;
1843
1844         ctl->op->recalc_thresholds(ctl);
1845 }
1846
1847 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1848                         struct btrfs_free_space *bitmap_info)
1849 {
1850         unlink_free_space(ctl, bitmap_info);
1851         kfree(bitmap_info->bitmap);
1852         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1853         ctl->total_bitmaps--;
1854         ctl->op->recalc_thresholds(ctl);
1855 }
1856
1857 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1858                               struct btrfs_free_space *bitmap_info,
1859                               u64 *offset, u64 *bytes)
1860 {
1861         u64 end;
1862         u64 search_start, search_bytes;
1863         int ret;
1864
1865 again:
1866         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1867
1868         /*
1869          * We need to search for bits in this bitmap.  We could only cover some
1870          * of the extent in this bitmap thanks to how we add space, so we need
1871          * to search for as much as it as we can and clear that amount, and then
1872          * go searching for the next bit.
1873          */
1874         search_start = *offset;
1875         search_bytes = ctl->unit;
1876         search_bytes = min(search_bytes, end - search_start + 1);
1877         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1878         if (ret < 0 || search_start != *offset)
1879                 return -EINVAL;
1880
1881         /* We may have found more bits than what we need */
1882         search_bytes = min(search_bytes, *bytes);
1883
1884         /* Cannot clear past the end of the bitmap */
1885         search_bytes = min(search_bytes, end - search_start + 1);
1886
1887         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1888         *offset += search_bytes;
1889         *bytes -= search_bytes;
1890
1891         if (*bytes) {
1892                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1893                 if (!bitmap_info->bytes)
1894                         free_bitmap(ctl, bitmap_info);
1895
1896                 /*
1897                  * no entry after this bitmap, but we still have bytes to
1898                  * remove, so something has gone wrong.
1899                  */
1900                 if (!next)
1901                         return -EINVAL;
1902
1903                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1904                                        offset_index);
1905
1906                 /*
1907                  * if the next entry isn't a bitmap we need to return to let the
1908                  * extent stuff do its work.
1909                  */
1910                 if (!bitmap_info->bitmap)
1911                         return -EAGAIN;
1912
1913                 /*
1914                  * Ok the next item is a bitmap, but it may not actually hold
1915                  * the information for the rest of this free space stuff, so
1916                  * look for it, and if we don't find it return so we can try
1917                  * everything over again.
1918                  */
1919                 search_start = *offset;
1920                 search_bytes = ctl->unit;
1921                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1922                                     &search_bytes);
1923                 if (ret < 0 || search_start != *offset)
1924                         return -EAGAIN;
1925
1926                 goto again;
1927         } else if (!bitmap_info->bytes)
1928                 free_bitmap(ctl, bitmap_info);
1929
1930         return 0;
1931 }
1932
1933 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1934                                struct btrfs_free_space *info, u64 offset,
1935                                u64 bytes)
1936 {
1937         u64 bytes_to_set = 0;
1938         u64 end;
1939
1940         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1941
1942         bytes_to_set = min(end - offset, bytes);
1943
1944         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1945
1946         return bytes_to_set;
1947
1948 }
1949
1950 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1951                       struct btrfs_free_space *info)
1952 {
1953         struct btrfs_block_group_cache *block_group = ctl->private;
1954
1955         /*
1956          * If we are below the extents threshold then we can add this as an
1957          * extent, and don't have to deal with the bitmap
1958          */
1959         if (ctl->free_extents < ctl->extents_thresh) {
1960                 /*
1961                  * If this block group has some small extents we don't want to
1962                  * use up all of our free slots in the cache with them, we want
1963                  * to reserve them to larger extents, however if we have plent
1964                  * of cache left then go ahead an dadd them, no sense in adding
1965                  * the overhead of a bitmap if we don't have to.
1966                  */
1967                 if (info->bytes <= block_group->sectorsize * 4) {
1968                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
1969                                 return false;
1970                 } else {
1971                         return false;
1972                 }
1973         }
1974
1975         /*
1976          * The original block groups from mkfs can be really small, like 8
1977          * megabytes, so don't bother with a bitmap for those entries.  However
1978          * some block groups can be smaller than what a bitmap would cover but
1979          * are still large enough that they could overflow the 32k memory limit,
1980          * so allow those block groups to still be allowed to have a bitmap
1981          * entry.
1982          */
1983         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
1984                 return false;
1985
1986         return true;
1987 }
1988
1989 static struct btrfs_free_space_op free_space_op = {
1990         .recalc_thresholds      = recalculate_thresholds,
1991         .use_bitmap             = use_bitmap,
1992 };
1993
1994 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1995                               struct btrfs_free_space *info)
1996 {
1997         struct btrfs_free_space *bitmap_info;
1998         struct btrfs_block_group_cache *block_group = NULL;
1999         int added = 0;
2000         u64 bytes, offset, bytes_added;
2001         int ret;
2002
2003         bytes = info->bytes;
2004         offset = info->offset;
2005
2006         if (!ctl->op->use_bitmap(ctl, info))
2007                 return 0;
2008
2009         if (ctl->op == &free_space_op)
2010                 block_group = ctl->private;
2011 again:
2012         /*
2013          * Since we link bitmaps right into the cluster we need to see if we
2014          * have a cluster here, and if so and it has our bitmap we need to add
2015          * the free space to that bitmap.
2016          */
2017         if (block_group && !list_empty(&block_group->cluster_list)) {
2018                 struct btrfs_free_cluster *cluster;
2019                 struct rb_node *node;
2020                 struct btrfs_free_space *entry;
2021
2022                 cluster = list_entry(block_group->cluster_list.next,
2023                                      struct btrfs_free_cluster,
2024                                      block_group_list);
2025                 spin_lock(&cluster->lock);
2026                 node = rb_first(&cluster->root);
2027                 if (!node) {
2028                         spin_unlock(&cluster->lock);
2029                         goto no_cluster_bitmap;
2030                 }
2031
2032                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2033                 if (!entry->bitmap) {
2034                         spin_unlock(&cluster->lock);
2035                         goto no_cluster_bitmap;
2036                 }
2037
2038                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2039                         bytes_added = add_bytes_to_bitmap(ctl, entry,
2040                                                           offset, bytes);
2041                         bytes -= bytes_added;
2042                         offset += bytes_added;
2043                 }
2044                 spin_unlock(&cluster->lock);
2045                 if (!bytes) {
2046                         ret = 1;
2047                         goto out;
2048                 }
2049         }
2050
2051 no_cluster_bitmap:
2052         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2053                                          1, 0);
2054         if (!bitmap_info) {
2055                 ASSERT(added == 0);
2056                 goto new_bitmap;
2057         }
2058
2059         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2060         bytes -= bytes_added;
2061         offset += bytes_added;
2062         added = 0;
2063
2064         if (!bytes) {
2065                 ret = 1;
2066                 goto out;
2067         } else
2068                 goto again;
2069
2070 new_bitmap:
2071         if (info && info->bitmap) {
2072                 add_new_bitmap(ctl, info, offset);
2073                 added = 1;
2074                 info = NULL;
2075                 goto again;
2076         } else {
2077                 spin_unlock(&ctl->tree_lock);
2078
2079                 /* no pre-allocated info, allocate a new one */
2080                 if (!info) {
2081                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2082                                                  GFP_NOFS);
2083                         if (!info) {
2084                                 spin_lock(&ctl->tree_lock);
2085                                 ret = -ENOMEM;
2086                                 goto out;
2087                         }
2088                 }
2089
2090                 /* allocate the bitmap */
2091                 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
2092                 spin_lock(&ctl->tree_lock);
2093                 if (!info->bitmap) {
2094                         ret = -ENOMEM;
2095                         goto out;
2096                 }
2097                 goto again;
2098         }
2099
2100 out:
2101         if (info) {
2102                 if (info->bitmap)
2103                         kfree(info->bitmap);
2104                 kmem_cache_free(btrfs_free_space_cachep, info);
2105         }
2106
2107         return ret;
2108 }
2109
2110 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2111                           struct btrfs_free_space *info, bool update_stat)
2112 {
2113         struct btrfs_free_space *left_info;
2114         struct btrfs_free_space *right_info;
2115         bool merged = false;
2116         u64 offset = info->offset;
2117         u64 bytes = info->bytes;
2118
2119         /*
2120          * first we want to see if there is free space adjacent to the range we
2121          * are adding, if there is remove that struct and add a new one to
2122          * cover the entire range
2123          */
2124         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2125         if (right_info && rb_prev(&right_info->offset_index))
2126                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2127                                      struct btrfs_free_space, offset_index);
2128         else
2129                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2130
2131         if (right_info && !right_info->bitmap) {
2132                 if (update_stat)
2133                         unlink_free_space(ctl, right_info);
2134                 else
2135                         __unlink_free_space(ctl, right_info);
2136                 info->bytes += right_info->bytes;
2137                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2138                 merged = true;
2139         }
2140
2141         if (left_info && !left_info->bitmap &&
2142             left_info->offset + left_info->bytes == offset) {
2143                 if (update_stat)
2144                         unlink_free_space(ctl, left_info);
2145                 else
2146                         __unlink_free_space(ctl, left_info);
2147                 info->offset = left_info->offset;
2148                 info->bytes += left_info->bytes;
2149                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2150                 merged = true;
2151         }
2152
2153         return merged;
2154 }
2155
2156 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2157                                      struct btrfs_free_space *info,
2158                                      bool update_stat)
2159 {
2160         struct btrfs_free_space *bitmap;
2161         unsigned long i;
2162         unsigned long j;
2163         const u64 end = info->offset + info->bytes;
2164         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2165         u64 bytes;
2166
2167         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2168         if (!bitmap)
2169                 return false;
2170
2171         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2172         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2173         if (j == i)
2174                 return false;
2175         bytes = (j - i) * ctl->unit;
2176         info->bytes += bytes;
2177
2178         if (update_stat)
2179                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2180         else
2181                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2182
2183         if (!bitmap->bytes)
2184                 free_bitmap(ctl, bitmap);
2185
2186         return true;
2187 }
2188
2189 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2190                                        struct btrfs_free_space *info,
2191                                        bool update_stat)
2192 {
2193         struct btrfs_free_space *bitmap;
2194         u64 bitmap_offset;
2195         unsigned long i;
2196         unsigned long j;
2197         unsigned long prev_j;
2198         u64 bytes;
2199
2200         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2201         /* If we're on a boundary, try the previous logical bitmap. */
2202         if (bitmap_offset == info->offset) {
2203                 if (info->offset == 0)
2204                         return false;
2205                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2206         }
2207
2208         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2209         if (!bitmap)
2210                 return false;
2211
2212         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2213         j = 0;
2214         prev_j = (unsigned long)-1;
2215         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2216                 if (j > i)
2217                         break;
2218                 prev_j = j;
2219         }
2220         if (prev_j == i)
2221                 return false;
2222
2223         if (prev_j == (unsigned long)-1)
2224                 bytes = (i + 1) * ctl->unit;
2225         else
2226                 bytes = (i - prev_j) * ctl->unit;
2227
2228         info->offset -= bytes;
2229         info->bytes += bytes;
2230
2231         if (update_stat)
2232                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2233         else
2234                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2235
2236         if (!bitmap->bytes)
2237                 free_bitmap(ctl, bitmap);
2238
2239         return true;
2240 }
2241
2242 /*
2243  * We prefer always to allocate from extent entries, both for clustered and
2244  * non-clustered allocation requests. So when attempting to add a new extent
2245  * entry, try to see if there's adjacent free space in bitmap entries, and if
2246  * there is, migrate that space from the bitmaps to the extent.
2247  * Like this we get better chances of satisfying space allocation requests
2248  * because we attempt to satisfy them based on a single cache entry, and never
2249  * on 2 or more entries - even if the entries represent a contiguous free space
2250  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2251  * ends).
2252  */
2253 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2254                               struct btrfs_free_space *info,
2255                               bool update_stat)
2256 {
2257         /*
2258          * Only work with disconnected entries, as we can change their offset,
2259          * and must be extent entries.
2260          */
2261         ASSERT(!info->bitmap);
2262         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2263
2264         if (ctl->total_bitmaps > 0) {
2265                 bool stole_end;
2266                 bool stole_front = false;
2267
2268                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2269                 if (ctl->total_bitmaps > 0)
2270                         stole_front = steal_from_bitmap_to_front(ctl, info,
2271                                                                  update_stat);
2272
2273                 if (stole_end || stole_front)
2274                         try_merge_free_space(ctl, info, update_stat);
2275         }
2276 }
2277
2278 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
2279                            u64 offset, u64 bytes)
2280 {
2281         struct btrfs_free_space *info;
2282         int ret = 0;
2283
2284         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2285         if (!info)
2286                 return -ENOMEM;
2287
2288         info->offset = offset;
2289         info->bytes = bytes;
2290         RB_CLEAR_NODE(&info->offset_index);
2291
2292         spin_lock(&ctl->tree_lock);
2293
2294         if (try_merge_free_space(ctl, info, true))
2295                 goto link;
2296
2297         /*
2298          * There was no extent directly to the left or right of this new
2299          * extent then we know we're going to have to allocate a new extent, so
2300          * before we do that see if we need to drop this into a bitmap
2301          */
2302         ret = insert_into_bitmap(ctl, info);
2303         if (ret < 0) {
2304                 goto out;
2305         } else if (ret) {
2306                 ret = 0;
2307                 goto out;
2308         }
2309 link:
2310         /*
2311          * Only steal free space from adjacent bitmaps if we're sure we're not
2312          * going to add the new free space to existing bitmap entries - because
2313          * that would mean unnecessary work that would be reverted. Therefore
2314          * attempt to steal space from bitmaps if we're adding an extent entry.
2315          */
2316         steal_from_bitmap(ctl, info, true);
2317
2318         ret = link_free_space(ctl, info);
2319         if (ret)
2320                 kmem_cache_free(btrfs_free_space_cachep, info);
2321 out:
2322         spin_unlock(&ctl->tree_lock);
2323
2324         if (ret) {
2325                 printk(KERN_CRIT "BTRFS: unable to add free space :%d\n", ret);
2326                 ASSERT(ret != -EEXIST);
2327         }
2328
2329         return ret;
2330 }
2331
2332 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2333                             u64 offset, u64 bytes)
2334 {
2335         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2336         struct btrfs_free_space *info;
2337         int ret;
2338         bool re_search = false;
2339
2340         spin_lock(&ctl->tree_lock);
2341
2342 again:
2343         ret = 0;
2344         if (!bytes)
2345                 goto out_lock;
2346
2347         info = tree_search_offset(ctl, offset, 0, 0);
2348         if (!info) {
2349                 /*
2350                  * oops didn't find an extent that matched the space we wanted
2351                  * to remove, look for a bitmap instead
2352                  */
2353                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2354                                           1, 0);
2355                 if (!info) {
2356                         /*
2357                          * If we found a partial bit of our free space in a
2358                          * bitmap but then couldn't find the other part this may
2359                          * be a problem, so WARN about it.
2360                          */
2361                         WARN_ON(re_search);
2362                         goto out_lock;
2363                 }
2364         }
2365
2366         re_search = false;
2367         if (!info->bitmap) {
2368                 unlink_free_space(ctl, info);
2369                 if (offset == info->offset) {
2370                         u64 to_free = min(bytes, info->bytes);
2371
2372                         info->bytes -= to_free;
2373                         info->offset += to_free;
2374                         if (info->bytes) {
2375                                 ret = link_free_space(ctl, info);
2376                                 WARN_ON(ret);
2377                         } else {
2378                                 kmem_cache_free(btrfs_free_space_cachep, info);
2379                         }
2380
2381                         offset += to_free;
2382                         bytes -= to_free;
2383                         goto again;
2384                 } else {
2385                         u64 old_end = info->bytes + info->offset;
2386
2387                         info->bytes = offset - info->offset;
2388                         ret = link_free_space(ctl, info);
2389                         WARN_ON(ret);
2390                         if (ret)
2391                                 goto out_lock;
2392
2393                         /* Not enough bytes in this entry to satisfy us */
2394                         if (old_end < offset + bytes) {
2395                                 bytes -= old_end - offset;
2396                                 offset = old_end;
2397                                 goto again;
2398                         } else if (old_end == offset + bytes) {
2399                                 /* all done */
2400                                 goto out_lock;
2401                         }
2402                         spin_unlock(&ctl->tree_lock);
2403
2404                         ret = btrfs_add_free_space(block_group, offset + bytes,
2405                                                    old_end - (offset + bytes));
2406                         WARN_ON(ret);
2407                         goto out;
2408                 }
2409         }
2410
2411         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2412         if (ret == -EAGAIN) {
2413                 re_search = true;
2414                 goto again;
2415         }
2416 out_lock:
2417         spin_unlock(&ctl->tree_lock);
2418 out:
2419         return ret;
2420 }
2421
2422 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2423                            u64 bytes)
2424 {
2425         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2426         struct btrfs_free_space *info;
2427         struct rb_node *n;
2428         int count = 0;
2429
2430         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2431                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2432                 if (info->bytes >= bytes && !block_group->ro)
2433                         count++;
2434                 btrfs_crit(block_group->fs_info,
2435                            "entry offset %llu, bytes %llu, bitmap %s",
2436                            info->offset, info->bytes,
2437                        (info->bitmap) ? "yes" : "no");
2438         }
2439         btrfs_info(block_group->fs_info, "block group has cluster?: %s",
2440                list_empty(&block_group->cluster_list) ? "no" : "yes");
2441         btrfs_info(block_group->fs_info,
2442                    "%d blocks of free space at or bigger than bytes is", count);
2443 }
2444
2445 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2446 {
2447         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2448
2449         spin_lock_init(&ctl->tree_lock);
2450         ctl->unit = block_group->sectorsize;
2451         ctl->start = block_group->key.objectid;
2452         ctl->private = block_group;
2453         ctl->op = &free_space_op;
2454         INIT_LIST_HEAD(&ctl->trimming_ranges);
2455         mutex_init(&ctl->cache_writeout_mutex);
2456
2457         /*
2458          * we only want to have 32k of ram per block group for keeping
2459          * track of free space, and if we pass 1/2 of that we want to
2460          * start converting things over to using bitmaps
2461          */
2462         ctl->extents_thresh = ((1024 * 32) / 2) /
2463                                 sizeof(struct btrfs_free_space);
2464 }
2465
2466 /*
2467  * for a given cluster, put all of its extents back into the free
2468  * space cache.  If the block group passed doesn't match the block group
2469  * pointed to by the cluster, someone else raced in and freed the
2470  * cluster already.  In that case, we just return without changing anything
2471  */
2472 static int
2473 __btrfs_return_cluster_to_free_space(
2474                              struct btrfs_block_group_cache *block_group,
2475                              struct btrfs_free_cluster *cluster)
2476 {
2477         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2478         struct btrfs_free_space *entry;
2479         struct rb_node *node;
2480
2481         spin_lock(&cluster->lock);
2482         if (cluster->block_group != block_group)
2483                 goto out;
2484
2485         cluster->block_group = NULL;
2486         cluster->window_start = 0;
2487         list_del_init(&cluster->block_group_list);
2488
2489         node = rb_first(&cluster->root);
2490         while (node) {
2491                 bool bitmap;
2492
2493                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2494                 node = rb_next(&entry->offset_index);
2495                 rb_erase(&entry->offset_index, &cluster->root);
2496                 RB_CLEAR_NODE(&entry->offset_index);
2497
2498                 bitmap = (entry->bitmap != NULL);
2499                 if (!bitmap) {
2500                         try_merge_free_space(ctl, entry, false);
2501                         steal_from_bitmap(ctl, entry, false);
2502                 }
2503                 tree_insert_offset(&ctl->free_space_offset,
2504                                    entry->offset, &entry->offset_index, bitmap);
2505         }
2506         cluster->root = RB_ROOT;
2507
2508 out:
2509         spin_unlock(&cluster->lock);
2510         btrfs_put_block_group(block_group);
2511         return 0;
2512 }
2513
2514 static void __btrfs_remove_free_space_cache_locked(
2515                                 struct btrfs_free_space_ctl *ctl)
2516 {
2517         struct btrfs_free_space *info;
2518         struct rb_node *node;
2519
2520         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2521                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2522                 if (!info->bitmap) {
2523                         unlink_free_space(ctl, info);
2524                         kmem_cache_free(btrfs_free_space_cachep, info);
2525                 } else {
2526                         free_bitmap(ctl, info);
2527                 }
2528
2529                 cond_resched_lock(&ctl->tree_lock);
2530         }
2531 }
2532
2533 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2534 {
2535         spin_lock(&ctl->tree_lock);
2536         __btrfs_remove_free_space_cache_locked(ctl);
2537         spin_unlock(&ctl->tree_lock);
2538 }
2539
2540 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2541 {
2542         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2543         struct btrfs_free_cluster *cluster;
2544         struct list_head *head;
2545
2546         spin_lock(&ctl->tree_lock);
2547         while ((head = block_group->cluster_list.next) !=
2548                &block_group->cluster_list) {
2549                 cluster = list_entry(head, struct btrfs_free_cluster,
2550                                      block_group_list);
2551
2552                 WARN_ON(cluster->block_group != block_group);
2553                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2554
2555                 cond_resched_lock(&ctl->tree_lock);
2556         }
2557         __btrfs_remove_free_space_cache_locked(ctl);
2558         spin_unlock(&ctl->tree_lock);
2559
2560 }
2561
2562 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2563                                u64 offset, u64 bytes, u64 empty_size,
2564                                u64 *max_extent_size)
2565 {
2566         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2567         struct btrfs_free_space *entry = NULL;
2568         u64 bytes_search = bytes + empty_size;
2569         u64 ret = 0;
2570         u64 align_gap = 0;
2571         u64 align_gap_len = 0;
2572
2573         spin_lock(&ctl->tree_lock);
2574         entry = find_free_space(ctl, &offset, &bytes_search,
2575                                 block_group->full_stripe_len, max_extent_size);
2576         if (!entry)
2577                 goto out;
2578
2579         ret = offset;
2580         if (entry->bitmap) {
2581                 bitmap_clear_bits(ctl, entry, offset, bytes);
2582                 if (!entry->bytes)
2583                         free_bitmap(ctl, entry);
2584         } else {
2585                 unlink_free_space(ctl, entry);
2586                 align_gap_len = offset - entry->offset;
2587                 align_gap = entry->offset;
2588
2589                 entry->offset = offset + bytes;
2590                 WARN_ON(entry->bytes < bytes + align_gap_len);
2591
2592                 entry->bytes -= bytes + align_gap_len;
2593                 if (!entry->bytes)
2594                         kmem_cache_free(btrfs_free_space_cachep, entry);
2595                 else
2596                         link_free_space(ctl, entry);
2597         }
2598 out:
2599         spin_unlock(&ctl->tree_lock);
2600
2601         if (align_gap_len)
2602                 __btrfs_add_free_space(ctl, align_gap, align_gap_len);
2603         return ret;
2604 }
2605
2606 /*
2607  * given a cluster, put all of its extents back into the free space
2608  * cache.  If a block group is passed, this function will only free
2609  * a cluster that belongs to the passed block group.
2610  *
2611  * Otherwise, it'll get a reference on the block group pointed to by the
2612  * cluster and remove the cluster from it.
2613  */
2614 int btrfs_return_cluster_to_free_space(
2615                                struct btrfs_block_group_cache *block_group,
2616                                struct btrfs_free_cluster *cluster)
2617 {
2618         struct btrfs_free_space_ctl *ctl;
2619         int ret;
2620
2621         /* first, get a safe pointer to the block group */
2622         spin_lock(&cluster->lock);
2623         if (!block_group) {
2624                 block_group = cluster->block_group;
2625                 if (!block_group) {
2626                         spin_unlock(&cluster->lock);
2627                         return 0;
2628                 }
2629         } else if (cluster->block_group != block_group) {
2630                 /* someone else has already freed it don't redo their work */
2631                 spin_unlock(&cluster->lock);
2632                 return 0;
2633         }
2634         atomic_inc(&block_group->count);
2635         spin_unlock(&cluster->lock);
2636
2637         ctl = block_group->free_space_ctl;
2638
2639         /* now return any extents the cluster had on it */
2640         spin_lock(&ctl->tree_lock);
2641         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2642         spin_unlock(&ctl->tree_lock);
2643
2644         /* finally drop our ref */
2645         btrfs_put_block_group(block_group);
2646         return ret;
2647 }
2648
2649 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2650                                    struct btrfs_free_cluster *cluster,
2651                                    struct btrfs_free_space *entry,
2652                                    u64 bytes, u64 min_start,
2653                                    u64 *max_extent_size)
2654 {
2655         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2656         int err;
2657         u64 search_start = cluster->window_start;
2658         u64 search_bytes = bytes;
2659         u64 ret = 0;
2660
2661         search_start = min_start;
2662         search_bytes = bytes;
2663
2664         err = search_bitmap(ctl, entry, &search_start, &search_bytes);
2665         if (err) {
2666                 if (search_bytes > *max_extent_size)
2667                         *max_extent_size = search_bytes;
2668                 return 0;
2669         }
2670
2671         ret = search_start;
2672         __bitmap_clear_bits(ctl, entry, ret, bytes);
2673
2674         return ret;
2675 }
2676
2677 /*
2678  * given a cluster, try to allocate 'bytes' from it, returns 0
2679  * if it couldn't find anything suitably large, or a logical disk offset
2680  * if things worked out
2681  */
2682 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2683                              struct btrfs_free_cluster *cluster, u64 bytes,
2684                              u64 min_start, u64 *max_extent_size)
2685 {
2686         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2687         struct btrfs_free_space *entry = NULL;
2688         struct rb_node *node;
2689         u64 ret = 0;
2690
2691         spin_lock(&cluster->lock);
2692         if (bytes > cluster->max_size)
2693                 goto out;
2694
2695         if (cluster->block_group != block_group)
2696                 goto out;
2697
2698         node = rb_first(&cluster->root);
2699         if (!node)
2700                 goto out;
2701
2702         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2703         while (1) {
2704                 if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2705                         *max_extent_size = entry->bytes;
2706
2707                 if (entry->bytes < bytes ||
2708                     (!entry->bitmap && entry->offset < min_start)) {
2709                         node = rb_next(&entry->offset_index);
2710                         if (!node)
2711                                 break;
2712                         entry = rb_entry(node, struct btrfs_free_space,
2713                                          offset_index);
2714                         continue;
2715                 }
2716
2717                 if (entry->bitmap) {
2718                         ret = btrfs_alloc_from_bitmap(block_group,
2719                                                       cluster, entry, bytes,
2720                                                       cluster->window_start,
2721                                                       max_extent_size);
2722                         if (ret == 0) {
2723                                 node = rb_next(&entry->offset_index);
2724                                 if (!node)
2725                                         break;
2726                                 entry = rb_entry(node, struct btrfs_free_space,
2727                                                  offset_index);
2728                                 continue;
2729                         }
2730                         cluster->window_start += bytes;
2731                 } else {
2732                         ret = entry->offset;
2733
2734                         entry->offset += bytes;
2735                         entry->bytes -= bytes;
2736                 }
2737
2738                 if (entry->bytes == 0)
2739                         rb_erase(&entry->offset_index, &cluster->root);
2740                 break;
2741         }
2742 out:
2743         spin_unlock(&cluster->lock);
2744
2745         if (!ret)
2746                 return 0;
2747
2748         spin_lock(&ctl->tree_lock);
2749
2750         ctl->free_space -= bytes;
2751         if (entry->bytes == 0) {
2752                 ctl->free_extents--;
2753                 if (entry->bitmap) {
2754                         kfree(entry->bitmap);
2755                         ctl->total_bitmaps--;
2756                         ctl->op->recalc_thresholds(ctl);
2757                 }
2758                 kmem_cache_free(btrfs_free_space_cachep, entry);
2759         }
2760
2761         spin_unlock(&ctl->tree_lock);
2762
2763         return ret;
2764 }
2765
2766 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2767                                 struct btrfs_free_space *entry,
2768                                 struct btrfs_free_cluster *cluster,
2769                                 u64 offset, u64 bytes,
2770                                 u64 cont1_bytes, u64 min_bytes)
2771 {
2772         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2773         unsigned long next_zero;
2774         unsigned long i;
2775         unsigned long want_bits;
2776         unsigned long min_bits;
2777         unsigned long found_bits;
2778         unsigned long start = 0;
2779         unsigned long total_found = 0;
2780         int ret;
2781
2782         i = offset_to_bit(entry->offset, ctl->unit,
2783                           max_t(u64, offset, entry->offset));
2784         want_bits = bytes_to_bits(bytes, ctl->unit);
2785         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2786
2787 again:
2788         found_bits = 0;
2789         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2790                 next_zero = find_next_zero_bit(entry->bitmap,
2791                                                BITS_PER_BITMAP, i);
2792                 if (next_zero - i >= min_bits) {
2793                         found_bits = next_zero - i;
2794                         break;
2795                 }
2796                 i = next_zero;
2797         }
2798
2799         if (!found_bits)
2800                 return -ENOSPC;
2801
2802         if (!total_found) {
2803                 start = i;
2804                 cluster->max_size = 0;
2805         }
2806
2807         total_found += found_bits;
2808
2809         if (cluster->max_size < found_bits * ctl->unit)
2810                 cluster->max_size = found_bits * ctl->unit;
2811
2812         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2813                 i = next_zero + 1;
2814                 goto again;
2815         }
2816
2817         cluster->window_start = start * ctl->unit + entry->offset;
2818         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2819         ret = tree_insert_offset(&cluster->root, entry->offset,
2820                                  &entry->offset_index, 1);
2821         ASSERT(!ret); /* -EEXIST; Logic error */
2822
2823         trace_btrfs_setup_cluster(block_group, cluster,
2824                                   total_found * ctl->unit, 1);
2825         return 0;
2826 }
2827
2828 /*
2829  * This searches the block group for just extents to fill the cluster with.
2830  * Try to find a cluster with at least bytes total bytes, at least one
2831  * extent of cont1_bytes, and other clusters of at least min_bytes.
2832  */
2833 static noinline int
2834 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2835                         struct btrfs_free_cluster *cluster,
2836                         struct list_head *bitmaps, u64 offset, u64 bytes,
2837                         u64 cont1_bytes, u64 min_bytes)
2838 {
2839         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2840         struct btrfs_free_space *first = NULL;
2841         struct btrfs_free_space *entry = NULL;
2842         struct btrfs_free_space *last;
2843         struct rb_node *node;
2844         u64 window_free;
2845         u64 max_extent;
2846         u64 total_size = 0;
2847
2848         entry = tree_search_offset(ctl, offset, 0, 1);
2849         if (!entry)
2850                 return -ENOSPC;
2851
2852         /*
2853          * We don't want bitmaps, so just move along until we find a normal
2854          * extent entry.
2855          */
2856         while (entry->bitmap || entry->bytes < min_bytes) {
2857                 if (entry->bitmap && list_empty(&entry->list))
2858                         list_add_tail(&entry->list, bitmaps);
2859                 node = rb_next(&entry->offset_index);
2860                 if (!node)
2861                         return -ENOSPC;
2862                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2863         }
2864
2865         window_free = entry->bytes;
2866         max_extent = entry->bytes;
2867         first = entry;
2868         last = entry;
2869
2870         for (node = rb_next(&entry->offset_index); node;
2871              node = rb_next(&entry->offset_index)) {
2872                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2873
2874                 if (entry->bitmap) {
2875                         if (list_empty(&entry->list))
2876                                 list_add_tail(&entry->list, bitmaps);
2877                         continue;
2878                 }
2879
2880                 if (entry->bytes < min_bytes)
2881                         continue;
2882
2883                 last = entry;
2884                 window_free += entry->bytes;
2885                 if (entry->bytes > max_extent)
2886                         max_extent = entry->bytes;
2887         }
2888
2889         if (window_free < bytes || max_extent < cont1_bytes)
2890                 return -ENOSPC;
2891
2892         cluster->window_start = first->offset;
2893
2894         node = &first->offset_index;
2895
2896         /*
2897          * now we've found our entries, pull them out of the free space
2898          * cache and put them into the cluster rbtree
2899          */
2900         do {
2901                 int ret;
2902
2903                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2904                 node = rb_next(&entry->offset_index);
2905                 if (entry->bitmap || entry->bytes < min_bytes)
2906                         continue;
2907
2908                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2909                 ret = tree_insert_offset(&cluster->root, entry->offset,
2910                                          &entry->offset_index, 0);
2911                 total_size += entry->bytes;
2912                 ASSERT(!ret); /* -EEXIST; Logic error */
2913         } while (node && entry != last);
2914
2915         cluster->max_size = max_extent;
2916         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2917         return 0;
2918 }
2919
2920 /*
2921  * This specifically looks for bitmaps that may work in the cluster, we assume
2922  * that we have already failed to find extents that will work.
2923  */
2924 static noinline int
2925 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2926                      struct btrfs_free_cluster *cluster,
2927                      struct list_head *bitmaps, u64 offset, u64 bytes,
2928                      u64 cont1_bytes, u64 min_bytes)
2929 {
2930         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2931         struct btrfs_free_space *entry;
2932         int ret = -ENOSPC;
2933         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2934
2935         if (ctl->total_bitmaps == 0)
2936                 return -ENOSPC;
2937
2938         /*
2939          * The bitmap that covers offset won't be in the list unless offset
2940          * is just its start offset.
2941          */
2942         entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2943         if (entry->offset != bitmap_offset) {
2944                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2945                 if (entry && list_empty(&entry->list))
2946                         list_add(&entry->list, bitmaps);
2947         }
2948
2949         list_for_each_entry(entry, bitmaps, list) {
2950                 if (entry->bytes < bytes)
2951                         continue;
2952                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2953                                            bytes, cont1_bytes, min_bytes);
2954                 if (!ret)
2955                         return 0;
2956         }
2957
2958         /*
2959          * The bitmaps list has all the bitmaps that record free space
2960          * starting after offset, so no more search is required.
2961          */
2962         return -ENOSPC;
2963 }
2964
2965 /*
2966  * here we try to find a cluster of blocks in a block group.  The goal
2967  * is to find at least bytes+empty_size.
2968  * We might not find them all in one contiguous area.
2969  *
2970  * returns zero and sets up cluster if things worked out, otherwise
2971  * it returns -enospc
2972  */
2973 int btrfs_find_space_cluster(struct btrfs_root *root,
2974                              struct btrfs_block_group_cache *block_group,
2975                              struct btrfs_free_cluster *cluster,
2976                              u64 offset, u64 bytes, u64 empty_size)
2977 {
2978         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2979         struct btrfs_free_space *entry, *tmp;
2980         LIST_HEAD(bitmaps);
2981         u64 min_bytes;
2982         u64 cont1_bytes;
2983         int ret;
2984
2985         /*
2986          * Choose the minimum extent size we'll require for this
2987          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
2988          * For metadata, allow allocates with smaller extents.  For
2989          * data, keep it dense.
2990          */
2991         if (btrfs_test_opt(root, SSD_SPREAD)) {
2992                 cont1_bytes = min_bytes = bytes + empty_size;
2993         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2994                 cont1_bytes = bytes;
2995                 min_bytes = block_group->sectorsize;
2996         } else {
2997                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
2998                 min_bytes = block_group->sectorsize;
2999         }
3000
3001         spin_lock(&ctl->tree_lock);
3002
3003         /*
3004          * If we know we don't have enough space to make a cluster don't even
3005          * bother doing all the work to try and find one.
3006          */
3007         if (ctl->free_space < bytes) {
3008                 spin_unlock(&ctl->tree_lock);
3009                 return -ENOSPC;
3010         }
3011
3012         spin_lock(&cluster->lock);
3013
3014         /* someone already found a cluster, hooray */
3015         if (cluster->block_group) {
3016                 ret = 0;
3017                 goto out;
3018         }
3019
3020         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3021                                  min_bytes);
3022
3023         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3024                                       bytes + empty_size,
3025                                       cont1_bytes, min_bytes);
3026         if (ret)
3027                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3028                                            offset, bytes + empty_size,
3029                                            cont1_bytes, min_bytes);
3030
3031         /* Clear our temporary list */
3032         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3033                 list_del_init(&entry->list);
3034
3035         if (!ret) {
3036                 atomic_inc(&block_group->count);
3037                 list_add_tail(&cluster->block_group_list,
3038                               &block_group->cluster_list);
3039                 cluster->block_group = block_group;
3040         } else {
3041                 trace_btrfs_failed_cluster_setup(block_group);
3042         }
3043 out:
3044         spin_unlock(&cluster->lock);
3045         spin_unlock(&ctl->tree_lock);
3046
3047         return ret;
3048 }
3049
3050 /*
3051  * simple code to zero out a cluster
3052  */
3053 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3054 {
3055         spin_lock_init(&cluster->lock);
3056         spin_lock_init(&cluster->refill_lock);
3057         cluster->root = RB_ROOT;
3058         cluster->max_size = 0;
3059         INIT_LIST_HEAD(&cluster->block_group_list);
3060         cluster->block_group = NULL;
3061 }
3062
3063 static int do_trimming(struct btrfs_block_group_cache *block_group,
3064                        u64 *total_trimmed, u64 start, u64 bytes,
3065                        u64 reserved_start, u64 reserved_bytes,
3066                        struct btrfs_trim_range *trim_entry)
3067 {
3068         struct btrfs_space_info *space_info = block_group->space_info;
3069         struct btrfs_fs_info *fs_info = block_group->fs_info;
3070         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3071         int ret;
3072         int update = 0;
3073         u64 trimmed = 0;
3074
3075         spin_lock(&space_info->lock);
3076         spin_lock(&block_group->lock);
3077         if (!block_group->ro) {
3078                 block_group->reserved += reserved_bytes;
3079                 space_info->bytes_reserved += reserved_bytes;
3080                 update = 1;
3081         }
3082         spin_unlock(&block_group->lock);
3083         spin_unlock(&space_info->lock);
3084
3085         ret = btrfs_discard_extent(fs_info->extent_root,
3086                                    start, bytes, &trimmed);
3087         if (!ret)
3088                 *total_trimmed += trimmed;
3089
3090         mutex_lock(&ctl->cache_writeout_mutex);
3091         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3092         list_del(&trim_entry->list);
3093         mutex_unlock(&ctl->cache_writeout_mutex);
3094
3095         if (update) {
3096                 spin_lock(&space_info->lock);
3097                 spin_lock(&block_group->lock);
3098                 if (block_group->ro)
3099                         space_info->bytes_readonly += reserved_bytes;
3100                 block_group->reserved -= reserved_bytes;
3101                 space_info->bytes_reserved -= reserved_bytes;
3102                 spin_unlock(&space_info->lock);
3103                 spin_unlock(&block_group->lock);
3104         }
3105
3106         return ret;
3107 }
3108
3109 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3110                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3111 {
3112         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3113         struct btrfs_free_space *entry;
3114         struct rb_node *node;
3115         int ret = 0;
3116         u64 extent_start;
3117         u64 extent_bytes;
3118         u64 bytes;
3119
3120         while (start < end) {
3121                 struct btrfs_trim_range trim_entry;
3122
3123                 mutex_lock(&ctl->cache_writeout_mutex);
3124                 spin_lock(&ctl->tree_lock);
3125
3126                 if (ctl->free_space < minlen) {
3127                         spin_unlock(&ctl->tree_lock);
3128                         mutex_unlock(&ctl->cache_writeout_mutex);
3129                         break;
3130                 }
3131
3132                 entry = tree_search_offset(ctl, start, 0, 1);
3133                 if (!entry) {
3134                         spin_unlock(&ctl->tree_lock);
3135                         mutex_unlock(&ctl->cache_writeout_mutex);
3136                         break;
3137                 }
3138
3139                 /* skip bitmaps */
3140                 while (entry->bitmap) {
3141                         node = rb_next(&entry->offset_index);
3142                         if (!node) {
3143                                 spin_unlock(&ctl->tree_lock);
3144                                 mutex_unlock(&ctl->cache_writeout_mutex);
3145                                 goto out;
3146                         }
3147                         entry = rb_entry(node, struct btrfs_free_space,
3148                                          offset_index);
3149                 }
3150
3151                 if (entry->offset >= end) {
3152                         spin_unlock(&ctl->tree_lock);
3153                         mutex_unlock(&ctl->cache_writeout_mutex);
3154                         break;
3155                 }
3156
3157                 extent_start = entry->offset;
3158                 extent_bytes = entry->bytes;
3159                 start = max(start, extent_start);
3160                 bytes = min(extent_start + extent_bytes, end) - start;
3161                 if (bytes < minlen) {
3162                         spin_unlock(&ctl->tree_lock);
3163                         mutex_unlock(&ctl->cache_writeout_mutex);
3164                         goto next;
3165                 }
3166
3167                 unlink_free_space(ctl, entry);
3168                 kmem_cache_free(btrfs_free_space_cachep, entry);
3169
3170                 spin_unlock(&ctl->tree_lock);
3171                 trim_entry.start = extent_start;
3172                 trim_entry.bytes = extent_bytes;
3173                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3174                 mutex_unlock(&ctl->cache_writeout_mutex);
3175
3176                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3177                                   extent_start, extent_bytes, &trim_entry);
3178                 if (ret)
3179                         break;
3180 next:
3181                 start += bytes;
3182
3183                 if (fatal_signal_pending(current)) {
3184                         ret = -ERESTARTSYS;
3185                         break;
3186                 }
3187
3188                 cond_resched();
3189         }
3190 out:
3191         return ret;
3192 }
3193
3194 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3195                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3196 {
3197         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3198         struct btrfs_free_space *entry;
3199         int ret = 0;
3200         int ret2;
3201         u64 bytes;
3202         u64 offset = offset_to_bitmap(ctl, start);
3203
3204         while (offset < end) {
3205                 bool next_bitmap = false;
3206                 struct btrfs_trim_range trim_entry;
3207
3208                 mutex_lock(&ctl->cache_writeout_mutex);
3209                 spin_lock(&ctl->tree_lock);
3210
3211                 if (ctl->free_space < minlen) {
3212                         spin_unlock(&ctl->tree_lock);
3213                         mutex_unlock(&ctl->cache_writeout_mutex);
3214                         break;
3215                 }
3216
3217                 entry = tree_search_offset(ctl, offset, 1, 0);
3218                 if (!entry) {
3219                         spin_unlock(&ctl->tree_lock);
3220                         mutex_unlock(&ctl->cache_writeout_mutex);
3221                         next_bitmap = true;
3222                         goto next;
3223                 }
3224
3225                 bytes = minlen;
3226                 ret2 = search_bitmap(ctl, entry, &start, &bytes);
3227                 if (ret2 || start >= end) {
3228                         spin_unlock(&ctl->tree_lock);
3229                         mutex_unlock(&ctl->cache_writeout_mutex);
3230                         next_bitmap = true;
3231                         goto next;
3232                 }
3233
3234                 bytes = min(bytes, end - start);
3235                 if (bytes < minlen) {
3236                         spin_unlock(&ctl->tree_lock);
3237                         mutex_unlock(&ctl->cache_writeout_mutex);
3238                         goto next;
3239                 }
3240
3241                 bitmap_clear_bits(ctl, entry, start, bytes);
3242                 if (entry->bytes == 0)
3243                         free_bitmap(ctl, entry);
3244
3245                 spin_unlock(&ctl->tree_lock);
3246                 trim_entry.start = start;
3247                 trim_entry.bytes = bytes;
3248                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3249                 mutex_unlock(&ctl->cache_writeout_mutex);
3250
3251                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3252                                   start, bytes, &trim_entry);
3253                 if (ret)
3254                         break;
3255 next:
3256                 if (next_bitmap) {
3257                         offset += BITS_PER_BITMAP * ctl->unit;
3258                 } else {
3259                         start += bytes;
3260                         if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3261                                 offset += BITS_PER_BITMAP * ctl->unit;
3262                 }
3263
3264                 if (fatal_signal_pending(current)) {
3265                         ret = -ERESTARTSYS;
3266                         break;
3267                 }
3268
3269                 cond_resched();
3270         }
3271
3272         return ret;
3273 }
3274
3275 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3276                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3277 {
3278         int ret;
3279
3280         *trimmed = 0;
3281
3282         spin_lock(&block_group->lock);
3283         if (block_group->removed) {
3284                 spin_unlock(&block_group->lock);
3285                 return 0;
3286         }
3287         atomic_inc(&block_group->trimming);
3288         spin_unlock(&block_group->lock);
3289
3290         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3291         if (ret)
3292                 goto out;
3293
3294         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3295 out:
3296         spin_lock(&block_group->lock);
3297         if (atomic_dec_and_test(&block_group->trimming) &&
3298             block_group->removed) {
3299                 struct extent_map_tree *em_tree;
3300                 struct extent_map *em;
3301
3302                 spin_unlock(&block_group->lock);
3303
3304                 lock_chunks(block_group->fs_info->chunk_root);
3305                 em_tree = &block_group->fs_info->mapping_tree.map_tree;
3306                 write_lock(&em_tree->lock);
3307                 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3308                                            1);
3309                 BUG_ON(!em); /* logic error, can't happen */
3310                 /*
3311                  * remove_extent_mapping() will delete us from the pinned_chunks
3312                  * list, which is protected by the chunk mutex.
3313                  */
3314                 remove_extent_mapping(em_tree, em);
3315                 write_unlock(&em_tree->lock);
3316                 unlock_chunks(block_group->fs_info->chunk_root);
3317
3318                 /* once for us and once for the tree */
3319                 free_extent_map(em);
3320                 free_extent_map(em);
3321
3322                 /*
3323                  * We've left one free space entry and other tasks trimming
3324                  * this block group have left 1 entry each one. Free them.
3325                  */
3326                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3327         } else {
3328                 spin_unlock(&block_group->lock);
3329         }
3330
3331         return ret;
3332 }
3333
3334 /*
3335  * Find the left-most item in the cache tree, and then return the
3336  * smallest inode number in the item.
3337  *
3338  * Note: the returned inode number may not be the smallest one in
3339  * the tree, if the left-most item is a bitmap.
3340  */
3341 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3342 {
3343         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3344         struct btrfs_free_space *entry = NULL;
3345         u64 ino = 0;
3346
3347         spin_lock(&ctl->tree_lock);
3348
3349         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3350                 goto out;
3351
3352         entry = rb_entry(rb_first(&ctl->free_space_offset),
3353                          struct btrfs_free_space, offset_index);
3354
3355         if (!entry->bitmap) {
3356                 ino = entry->offset;
3357
3358                 unlink_free_space(ctl, entry);
3359                 entry->offset++;
3360                 entry->bytes--;
3361                 if (!entry->bytes)
3362                         kmem_cache_free(btrfs_free_space_cachep, entry);
3363                 else
3364                         link_free_space(ctl, entry);
3365         } else {
3366                 u64 offset = 0;
3367                 u64 count = 1;
3368                 int ret;
3369
3370                 ret = search_bitmap(ctl, entry, &offset, &count);
3371                 /* Logic error; Should be empty if it can't find anything */
3372                 ASSERT(!ret);
3373
3374                 ino = offset;
3375                 bitmap_clear_bits(ctl, entry, offset, 1);
3376                 if (entry->bytes == 0)
3377                         free_bitmap(ctl, entry);
3378         }
3379 out:
3380         spin_unlock(&ctl->tree_lock);
3381
3382         return ino;
3383 }
3384
3385 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3386                                     struct btrfs_path *path)
3387 {
3388         struct inode *inode = NULL;
3389
3390         spin_lock(&root->ino_cache_lock);
3391         if (root->ino_cache_inode)
3392                 inode = igrab(root->ino_cache_inode);
3393         spin_unlock(&root->ino_cache_lock);
3394         if (inode)
3395                 return inode;
3396
3397         inode = __lookup_free_space_inode(root, path, 0);
3398         if (IS_ERR(inode))
3399                 return inode;
3400
3401         spin_lock(&root->ino_cache_lock);
3402         if (!btrfs_fs_closing(root->fs_info))
3403                 root->ino_cache_inode = igrab(inode);
3404         spin_unlock(&root->ino_cache_lock);
3405
3406         return inode;
3407 }
3408
3409 int create_free_ino_inode(struct btrfs_root *root,
3410                           struct btrfs_trans_handle *trans,
3411                           struct btrfs_path *path)
3412 {
3413         return __create_free_space_inode(root, trans, path,
3414                                          BTRFS_FREE_INO_OBJECTID, 0);
3415 }
3416
3417 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3418 {
3419         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3420         struct btrfs_path *path;
3421         struct inode *inode;
3422         int ret = 0;
3423         u64 root_gen = btrfs_root_generation(&root->root_item);
3424
3425         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3426                 return 0;
3427
3428         /*
3429          * If we're unmounting then just return, since this does a search on the
3430          * normal root and not the commit root and we could deadlock.
3431          */
3432         if (btrfs_fs_closing(fs_info))
3433                 return 0;
3434
3435         path = btrfs_alloc_path();
3436         if (!path)
3437                 return 0;
3438
3439         inode = lookup_free_ino_inode(root, path);
3440         if (IS_ERR(inode))
3441                 goto out;
3442
3443         if (root_gen != BTRFS_I(inode)->generation)
3444                 goto out_put;
3445
3446         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3447
3448         if (ret < 0)
3449                 btrfs_err(fs_info,
3450                         "failed to load free ino cache for root %llu",
3451                         root->root_key.objectid);
3452 out_put:
3453         iput(inode);
3454 out:
3455         btrfs_free_path(path);
3456         return ret;
3457 }
3458
3459 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3460                               struct btrfs_trans_handle *trans,
3461                               struct btrfs_path *path,
3462                               struct inode *inode)
3463 {
3464         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3465         int ret;
3466         struct btrfs_io_ctl io_ctl;
3467
3468         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
3469                 return 0;
3470
3471         memset(&io_ctl, 0, sizeof(io_ctl));
3472         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl,
3473                                       trans, path, 0);
3474         if (!ret)
3475                 ret = btrfs_wait_cache_io(root, trans, NULL, &io_ctl, path, 0);
3476
3477         if (ret) {
3478                 btrfs_delalloc_release_metadata(inode, inode->i_size);
3479 #ifdef DEBUG
3480                 btrfs_err(root->fs_info,
3481                         "failed to write free ino cache for root %llu",
3482                         root->root_key.objectid);
3483 #endif
3484         }
3485
3486         return ret;
3487 }
3488
3489 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3490 /*
3491  * Use this if you need to make a bitmap or extent entry specifically, it
3492  * doesn't do any of the merging that add_free_space does, this acts a lot like
3493  * how the free space cache loading stuff works, so you can get really weird
3494  * configurations.
3495  */
3496 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3497                               u64 offset, u64 bytes, bool bitmap)
3498 {
3499         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3500         struct btrfs_free_space *info = NULL, *bitmap_info;
3501         void *map = NULL;
3502         u64 bytes_added;
3503         int ret;
3504
3505 again:
3506         if (!info) {
3507                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3508                 if (!info)
3509                         return -ENOMEM;
3510         }
3511
3512         if (!bitmap) {
3513                 spin_lock(&ctl->tree_lock);
3514                 info->offset = offset;
3515                 info->bytes = bytes;
3516                 ret = link_free_space(ctl, info);
3517                 spin_unlock(&ctl->tree_lock);
3518                 if (ret)
3519                         kmem_cache_free(btrfs_free_space_cachep, info);
3520                 return ret;
3521         }
3522
3523         if (!map) {
3524                 map = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
3525                 if (!map) {
3526                         kmem_cache_free(btrfs_free_space_cachep, info);
3527                         return -ENOMEM;
3528                 }
3529         }
3530
3531         spin_lock(&ctl->tree_lock);
3532         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3533                                          1, 0);
3534         if (!bitmap_info) {
3535                 info->bitmap = map;
3536                 map = NULL;
3537                 add_new_bitmap(ctl, info, offset);
3538                 bitmap_info = info;
3539                 info = NULL;
3540         }
3541
3542         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3543         bytes -= bytes_added;
3544         offset += bytes_added;
3545         spin_unlock(&ctl->tree_lock);
3546
3547         if (bytes)
3548                 goto again;
3549
3550         if (info)
3551                 kmem_cache_free(btrfs_free_space_cachep, info);
3552         if (map)
3553                 kfree(map);
3554         return 0;
3555 }
3556
3557 /*
3558  * Checks to see if the given range is in the free space cache.  This is really
3559  * just used to check the absence of space, so if there is free space in the
3560  * range at all we will return 1.
3561  */
3562 int test_check_exists(struct btrfs_block_group_cache *cache,
3563                       u64 offset, u64 bytes)
3564 {
3565         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3566         struct btrfs_free_space *info;
3567         int ret = 0;
3568
3569         spin_lock(&ctl->tree_lock);
3570         info = tree_search_offset(ctl, offset, 0, 0);
3571         if (!info) {
3572                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3573                                           1, 0);
3574                 if (!info)
3575                         goto out;
3576         }
3577
3578 have_info:
3579         if (info->bitmap) {
3580                 u64 bit_off, bit_bytes;
3581                 struct rb_node *n;
3582                 struct btrfs_free_space *tmp;
3583
3584                 bit_off = offset;
3585                 bit_bytes = ctl->unit;
3586                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes);
3587                 if (!ret) {
3588                         if (bit_off == offset) {
3589                                 ret = 1;
3590                                 goto out;
3591                         } else if (bit_off > offset &&
3592                                    offset + bytes > bit_off) {
3593                                 ret = 1;
3594                                 goto out;
3595                         }
3596                 }
3597
3598                 n = rb_prev(&info->offset_index);
3599                 while (n) {
3600                         tmp = rb_entry(n, struct btrfs_free_space,
3601                                        offset_index);
3602                         if (tmp->offset + tmp->bytes < offset)
3603                                 break;
3604                         if (offset + bytes < tmp->offset) {
3605                                 n = rb_prev(&info->offset_index);
3606                                 continue;
3607                         }
3608                         info = tmp;
3609                         goto have_info;
3610                 }
3611
3612                 n = rb_next(&info->offset_index);
3613                 while (n) {
3614                         tmp = rb_entry(n, struct btrfs_free_space,
3615                                        offset_index);
3616                         if (offset + bytes < tmp->offset)
3617                                 break;
3618                         if (tmp->offset + tmp->bytes < offset) {
3619                                 n = rb_next(&info->offset_index);
3620                                 continue;
3621                         }
3622                         info = tmp;
3623                         goto have_info;
3624                 }
3625
3626                 ret = 0;
3627                 goto out;
3628         }
3629
3630         if (info->offset == offset) {
3631                 ret = 1;
3632                 goto out;
3633         }
3634
3635         if (offset > info->offset && offset < info->offset + info->bytes)
3636                 ret = 1;
3637 out:
3638         spin_unlock(&ctl->tree_lock);
3639         return ret;
3640 }
3641 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */