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/sched/signal.h>
  22 #include <linux/slab.h>
  23 #include <linux/math64.h>
  24 #include <linux/ratelimit.h>
  25 #include "ctree.h"
  26 #include "free-space-cache.h"
  27 #include "transaction.h"
  28 #include "disk-io.h"
  29 #include "extent_io.h"
  30 #include "inode-map.h"
  31 #include "volumes.h"
  32
  33 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  34 #define MAX_CACHE_BYTES_PER_GIG SZ_32K
  35
  36 struct btrfs_trim_range {
  37         u64 start;
  38         u64 bytes;
  39         struct list_head list;
  40 };
  41
  42 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  43                            struct btrfs_free_space *info);
  44 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  45                               struct btrfs_free_space *info);
  46 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
  47                              struct btrfs_trans_handle *trans,
  48                              struct btrfs_io_ctl *io_ctl,
  49                              struct btrfs_path *path);
  50
  51 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  52                                                struct btrfs_path *path,
  53                                                u64 offset)
  54 {
  55         struct btrfs_fs_info *fs_info = root->fs_info;
  56         struct btrfs_key key;
  57         struct btrfs_key location;
  58         struct btrfs_disk_key disk_key;
  59         struct btrfs_free_space_header *header;
  60         struct extent_buffer *leaf;
  61         struct inode *inode = NULL;
  62         int ret;
  63
  64         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  65         key.offset = offset;
  66         key.type = 0;
  67
  68         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  69         if (ret < 0)
  70                 return ERR_PTR(ret);
  71         if (ret > 0) {
  72                 btrfs_release_path(path);
  73                 return ERR_PTR(-ENOENT);
  74         }
  75
  76         leaf = path->nodes[0];
  77         header = btrfs_item_ptr(leaf, path->slots[0],
  78                                 struct btrfs_free_space_header);
  79         btrfs_free_space_key(leaf, header, &disk_key);
  80         btrfs_disk_key_to_cpu(&location, &disk_key);
  81         btrfs_release_path(path);
  82
  83         inode = btrfs_iget(fs_info->sb, &location, root, NULL);
  84         if (IS_ERR(inode))
  85                 return inode;
  86         if (is_bad_inode(inode)) {
  87                 iput(inode);
  88                 return ERR_PTR(-ENOENT);
  89         }
  90
  91         mapping_set_gfp_mask(inode->i_mapping,
  92                         mapping_gfp_constraint(inode->i_mapping,
  93                         ~(__GFP_FS | __GFP_HIGHMEM)));
  94
  95         return inode;
  96 }
  97
  98 struct inode *lookup_free_space_inode(struct btrfs_fs_info *fs_info,
  99                                       struct btrfs_block_group_cache
 100                                       *block_group, struct btrfs_path *path)
 101 {
 102         struct inode *inode = NULL;
 103         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 104
 105         spin_lock(&block_group->lock);
 106         if (block_group->inode)
 107                 inode = igrab(block_group->inode);
 108         spin_unlock(&block_group->lock);
 109         if (inode)
 110                 return inode;
 111
 112         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 113                                           block_group->key.objectid);
 114         if (IS_ERR(inode))
 115                 return inode;
 116
 117         spin_lock(&block_group->lock);
 118         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 119                 btrfs_info(fs_info, "Old style space inode found, converting.");
 120                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 121                         BTRFS_INODE_NODATACOW;
 122                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 123         }
 124
 125         if (!block_group->iref) {
 126                 block_group->inode = igrab(inode);
 127                 block_group->iref = 1;
 128         }
 129         spin_unlock(&block_group->lock);
 130
 131         return inode;
 132 }
 133
 134 static int __create_free_space_inode(struct btrfs_root *root,
 135                                      struct btrfs_trans_handle *trans,
 136                                      struct btrfs_path *path,
 137                                      u64 ino, u64 offset)
 138 {
 139         struct btrfs_key key;
 140         struct btrfs_disk_key disk_key;
 141         struct btrfs_free_space_header *header;
 142         struct btrfs_inode_item *inode_item;
 143         struct extent_buffer *leaf;
 144         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 145         int ret;
 146
 147         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 148         if (ret)
 149                 return ret;
 150
 151         /* We inline crc's for the free disk space cache */
 152         if (ino != BTRFS_FREE_INO_OBJECTID)
 153                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 154
 155         leaf = path->nodes[0];
 156         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 157                                     struct btrfs_inode_item);
 158         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 159         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 160                              sizeof(*inode_item));
 161         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 162         btrfs_set_inode_size(leaf, inode_item, 0);
 163         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 164         btrfs_set_inode_uid(leaf, inode_item, 0);
 165         btrfs_set_inode_gid(leaf, inode_item, 0);
 166         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 167         btrfs_set_inode_flags(leaf, inode_item, flags);
 168         btrfs_set_inode_nlink(leaf, inode_item, 1);
 169         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 170         btrfs_set_inode_block_group(leaf, inode_item, offset);
 171         btrfs_mark_buffer_dirty(leaf);
 172         btrfs_release_path(path);
 173
 174         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 175         key.offset = offset;
 176         key.type = 0;
 177         ret = btrfs_insert_empty_item(trans, root, path, &key,
 178                                       sizeof(struct btrfs_free_space_header));
 179         if (ret < 0) {
 180                 btrfs_release_path(path);
 181                 return ret;
 182         }
 183
 184         leaf = path->nodes[0];
 185         header = btrfs_item_ptr(leaf, path->slots[0],
 186                                 struct btrfs_free_space_header);
 187         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 188         btrfs_set_free_space_key(leaf, header, &disk_key);
 189         btrfs_mark_buffer_dirty(leaf);
 190         btrfs_release_path(path);
 191
 192         return 0;
 193 }
 194
 195 int create_free_space_inode(struct btrfs_fs_info *fs_info,
 196                             struct btrfs_trans_handle *trans,
 197                             struct btrfs_block_group_cache *block_group,
 198                             struct btrfs_path *path)
 199 {
 200         int ret;
 201         u64 ino;
 202
 203         ret = btrfs_find_free_objectid(fs_info->tree_root, &ino);
 204         if (ret < 0)
 205                 return ret;
 206
 207         return __create_free_space_inode(fs_info->tree_root, trans, path, ino,
 208                                          block_group->key.objectid);
 209 }
 210
 211 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
 212                                        struct btrfs_block_rsv *rsv)
 213 {
 214         u64 needed_bytes;
 215         int ret;
 216
 217         /* 1 for slack space, 1 for updating the inode */
 218         needed_bytes = btrfs_calc_trunc_metadata_size(fs_info, 1) +
 219                 btrfs_calc_trans_metadata_size(fs_info, 1);
 220
 221         spin_lock(&rsv->lock);
 222         if (rsv->reserved < needed_bytes)
 223                 ret = -ENOSPC;
 224         else
 225                 ret = 0;
 226         spin_unlock(&rsv->lock);
 227         return ret;
 228 }
 229
 230 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 231                                     struct btrfs_block_group_cache *block_group,
 232                                     struct inode *inode)
 233 {
 234         struct btrfs_root *root = BTRFS_I(inode)->root;
 235         int ret = 0;
 236         bool locked = false;
 237
 238         if (block_group) {
 239                 struct btrfs_path *path = btrfs_alloc_path();
 240
 241                 if (!path) {
 242                         ret = -ENOMEM;
 243                         goto fail;
 244                 }
 245                 locked = true;
 246                 mutex_lock(&trans->transaction->cache_write_mutex);
 247                 if (!list_empty(&block_group->io_list)) {
 248                         list_del_init(&block_group->io_list);
 249
 250                         btrfs_wait_cache_io(trans, block_group, path);
 251                         btrfs_put_block_group(block_group);
 252                 }
 253
 254                 /*
 255                  * now that we've truncated the cache away, its no longer
 256                  * setup or written
 257                  */
 258                 spin_lock(&block_group->lock);
 259                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 260                 spin_unlock(&block_group->lock);
 261                 btrfs_free_path(path);
 262         }
 263
 264         btrfs_i_size_write(BTRFS_I(inode), 0);
 265         truncate_pagecache(inode, 0);
 266
 267         /*
 268          * We don't need an orphan item because truncating the free space cache
 269          * will never be split across transactions.
 270          * We don't need to check for -EAGAIN because we're a free space
 271          * cache inode
 272          */
 273         ret = btrfs_truncate_inode_items(trans, root, inode,
 274                                          0, BTRFS_EXTENT_DATA_KEY);
 275         if (ret)
 276                 goto fail;
 277
 278         ret = btrfs_update_inode(trans, root, inode);
 279
 280 fail:
 281         if (locked)
 282                 mutex_unlock(&trans->transaction->cache_write_mutex);
 283         if (ret)
 284                 btrfs_abort_transaction(trans, ret);
 285
 286         return ret;
 287 }
 288
 289 static void readahead_cache(struct inode *inode)
 290 {
 291         struct file_ra_state *ra;
 292         unsigned long last_index;
 293
 294         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 295         if (!ra)
 296                 return;
 297
 298         file_ra_state_init(ra, inode->i_mapping);
 299         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 300
 301         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 302
 303         kfree(ra);
 304 }
 305
 306 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 307                        int write)
 308 {
 309         int num_pages;
 310         int check_crcs = 0;
 311
 312         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 313
 314         if (btrfs_ino(BTRFS_I(inode)) != BTRFS_FREE_INO_OBJECTID)
 315                 check_crcs = 1;
 316
 317         /* Make sure we can fit our crcs into the first page */
 318         if (write && check_crcs &&
 319             (num_pages * sizeof(u32)) >= PAGE_SIZE)
 320                 return -ENOSPC;
 321
 322         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 323
 324         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 325         if (!io_ctl->pages)
 326                 return -ENOMEM;
 327
 328         io_ctl->num_pages = num_pages;
 329         io_ctl->fs_info = btrfs_sb(inode->i_sb);
 330         io_ctl->check_crcs = check_crcs;
 331         io_ctl->inode = inode;
 332
 333         return 0;
 334 }
 335
 336 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 337 {
 338         kfree(io_ctl->pages);
 339         io_ctl->pages = NULL;
 340 }
 341
 342 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 343 {
 344         if (io_ctl->cur) {
 345                 io_ctl->cur = NULL;
 346                 io_ctl->orig = NULL;
 347         }
 348 }
 349
 350 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 351 {
 352         ASSERT(io_ctl->index < io_ctl->num_pages);
 353         io_ctl->page = io_ctl->pages[io_ctl->index++];
 354         io_ctl->cur = page_address(io_ctl->page);
 355         io_ctl->orig = io_ctl->cur;
 356         io_ctl->size = PAGE_SIZE;
 357         if (clear)
 358                 clear_page(io_ctl->cur);
 359 }
 360
 361 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 362 {
 363         int i;
 364
 365         io_ctl_unmap_page(io_ctl);
 366
 367         for (i = 0; i < io_ctl->num_pages; i++) {
 368                 if (io_ctl->pages[i]) {
 369                         ClearPageChecked(io_ctl->pages[i]);
 370                         unlock_page(io_ctl->pages[i]);
 371                         put_page(io_ctl->pages[i]);
 372                 }
 373         }
 374 }
 375
 376 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 377                                 int uptodate)
 378 {
 379         struct page *page;
 380         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 381         int i;
 382
 383         for (i = 0; i < io_ctl->num_pages; i++) {
 384                 page = find_or_create_page(inode->i_mapping, i, mask);
 385                 if (!page) {
 386                         io_ctl_drop_pages(io_ctl);
 387                         return -ENOMEM;
 388                 }
 389                 io_ctl->pages[i] = page;
 390                 if (uptodate && !PageUptodate(page)) {
 391                         btrfs_readpage(NULL, page);
 392                         lock_page(page);
 393                         if (!PageUptodate(page)) {
 394                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 395                                            "error reading free space cache");
 396                                 io_ctl_drop_pages(io_ctl);
 397                                 return -EIO;
 398                         }
 399                 }
 400         }
 401
 402         for (i = 0; i < io_ctl->num_pages; i++) {
 403                 clear_page_dirty_for_io(io_ctl->pages[i]);
 404                 set_page_extent_mapped(io_ctl->pages[i]);
 405         }
 406
 407         return 0;
 408 }
 409
 410 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 411 {
 412         __le64 *val;
 413
 414         io_ctl_map_page(io_ctl, 1);
 415
 416         /*
 417          * Skip the csum areas.  If we don't check crcs then we just have a
 418          * 64bit chunk at the front of the first page.
 419          */
 420         if (io_ctl->check_crcs) {
 421                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 422                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 423         } else {
 424                 io_ctl->cur += sizeof(u64);
 425                 io_ctl->size -= sizeof(u64) * 2;
 426         }
 427
 428         val = io_ctl->cur;
 429         *val = cpu_to_le64(generation);
 430         io_ctl->cur += sizeof(u64);
 431 }
 432
 433 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 434 {
 435         __le64 *gen;
 436
 437         /*
 438          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 439          * chunk at the front of the first page.
 440          */
 441         if (io_ctl->check_crcs) {
 442                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 443                 io_ctl->size -= sizeof(u64) +
 444                         (sizeof(u32) * io_ctl->num_pages);
 445         } else {
 446                 io_ctl->cur += sizeof(u64);
 447                 io_ctl->size -= sizeof(u64) * 2;
 448         }
 449
 450         gen = io_ctl->cur;
 451         if (le64_to_cpu(*gen) != generation) {
 452                 btrfs_err_rl(io_ctl->fs_info,
 453                         "space cache generation (%llu) does not match inode (%llu)",
 454                                 *gen, generation);
 455                 io_ctl_unmap_page(io_ctl);
 456                 return -EIO;
 457         }
 458         io_ctl->cur += sizeof(u64);
 459         return 0;
 460 }
 461
 462 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 463 {
 464         u32 *tmp;
 465         u32 crc = ~(u32)0;
 466         unsigned offset = 0;
 467
 468         if (!io_ctl->check_crcs) {
 469                 io_ctl_unmap_page(io_ctl);
 470                 return;
 471         }
 472
 473         if (index == 0)
 474                 offset = sizeof(u32) * io_ctl->num_pages;
 475
 476         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 477                               PAGE_SIZE - offset);
 478         btrfs_csum_final(crc, (u8 *)&crc);
 479         io_ctl_unmap_page(io_ctl);
 480         tmp = page_address(io_ctl->pages[0]);
 481         tmp += index;
 482         *tmp = crc;
 483 }
 484
 485 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 486 {
 487         u32 *tmp, val;
 488         u32 crc = ~(u32)0;
 489         unsigned offset = 0;
 490
 491         if (!io_ctl->check_crcs) {
 492                 io_ctl_map_page(io_ctl, 0);
 493                 return 0;
 494         }
 495
 496         if (index == 0)
 497                 offset = sizeof(u32) * io_ctl->num_pages;
 498
 499         tmp = page_address(io_ctl->pages[0]);
 500         tmp += index;
 501         val = *tmp;
 502
 503         io_ctl_map_page(io_ctl, 0);
 504         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 505                               PAGE_SIZE - offset);
 506         btrfs_csum_final(crc, (u8 *)&crc);
 507         if (val != crc) {
 508                 btrfs_err_rl(io_ctl->fs_info,
 509                         "csum mismatch on free space cache");
 510                 io_ctl_unmap_page(io_ctl);
 511                 return -EIO;
 512         }
 513
 514         return 0;
 515 }
 516
 517 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 518                             void *bitmap)
 519 {
 520         struct btrfs_free_space_entry *entry;
 521
 522         if (!io_ctl->cur)
 523                 return -ENOSPC;
 524
 525         entry = io_ctl->cur;
 526         entry->offset = cpu_to_le64(offset);
 527         entry->bytes = cpu_to_le64(bytes);
 528         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 529                 BTRFS_FREE_SPACE_EXTENT;
 530         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 531         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 532
 533         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 534                 return 0;
 535
 536         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 537
 538         /* No more pages to map */
 539         if (io_ctl->index >= io_ctl->num_pages)
 540                 return 0;
 541
 542         /* map the next page */
 543         io_ctl_map_page(io_ctl, 1);
 544         return 0;
 545 }
 546
 547 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 548 {
 549         if (!io_ctl->cur)
 550                 return -ENOSPC;
 551
 552         /*
 553          * If we aren't at the start of the current page, unmap this one and
 554          * map the next one if there is any left.
 555          */
 556         if (io_ctl->cur != io_ctl->orig) {
 557                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 558                 if (io_ctl->index >= io_ctl->num_pages)
 559                         return -ENOSPC;
 560                 io_ctl_map_page(io_ctl, 0);
 561         }
 562
 563         memcpy(io_ctl->cur, bitmap, PAGE_SIZE);
 564         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 565         if (io_ctl->index < io_ctl->num_pages)
 566                 io_ctl_map_page(io_ctl, 0);
 567         return 0;
 568 }
 569
 570 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 571 {
 572         /*
 573          * If we're not on the boundary we know we've modified the page and we
 574          * need to crc the page.
 575          */
 576         if (io_ctl->cur != io_ctl->orig)
 577                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 578         else
 579                 io_ctl_unmap_page(io_ctl);
 580
 581         while (io_ctl->index < io_ctl->num_pages) {
 582                 io_ctl_map_page(io_ctl, 1);
 583                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 584         }
 585 }
 586
 587 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 588                             struct btrfs_free_space *entry, u8 *type)
 589 {
 590         struct btrfs_free_space_entry *e;
 591         int ret;
 592
 593         if (!io_ctl->cur) {
 594                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 595                 if (ret)
 596                         return ret;
 597         }
 598
 599         e = io_ctl->cur;
 600         entry->offset = le64_to_cpu(e->offset);
 601         entry->bytes = le64_to_cpu(e->bytes);
 602         *type = e->type;
 603         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 604         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 605
 606         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 607                 return 0;
 608
 609         io_ctl_unmap_page(io_ctl);
 610
 611         return 0;
 612 }
 613
 614 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 615                               struct btrfs_free_space *entry)
 616 {
 617         int ret;
 618
 619         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 620         if (ret)
 621                 return ret;
 622
 623         memcpy(entry->bitmap, io_ctl->cur, PAGE_SIZE);
 624         io_ctl_unmap_page(io_ctl);
 625
 626         return 0;
 627 }
 628
 629 /*
 630  * Since we attach pinned extents after the fact we can have contiguous sections
 631  * of free space that are split up in entries.  This poses a problem with the
 632  * tree logging stuff since it could have allocated across what appears to be 2
 633  * entries since we would have merged the entries when adding the pinned extents
 634  * back to the free space cache.  So run through the space cache that we just
 635  * loaded and merge contiguous entries.  This will make the log replay stuff not
 636  * blow up and it will make for nicer allocator behavior.
 637  */
 638 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 639 {
 640         struct btrfs_free_space *e, *prev = NULL;
 641         struct rb_node *n;
 642
 643 again:
 644         spin_lock(&ctl->tree_lock);
 645         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 646                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 647                 if (!prev)
 648                         goto next;
 649                 if (e->bitmap || prev->bitmap)
 650                         goto next;
 651                 if (prev->offset + prev->bytes == e->offset) {
 652                         unlink_free_space(ctl, prev);
 653                         unlink_free_space(ctl, e);
 654                         prev->bytes += e->bytes;
 655                         kmem_cache_free(btrfs_free_space_cachep, e);
 656                         link_free_space(ctl, prev);
 657                         prev = NULL;
 658                         spin_unlock(&ctl->tree_lock);
 659                         goto again;
 660                 }
 661 next:
 662                 prev = e;
 663         }
 664         spin_unlock(&ctl->tree_lock);
 665 }
 666
 667 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 668                                    struct btrfs_free_space_ctl *ctl,
 669                                    struct btrfs_path *path, u64 offset)
 670 {
 671         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 672         struct btrfs_free_space_header *header;
 673         struct extent_buffer *leaf;
 674         struct btrfs_io_ctl io_ctl;
 675         struct btrfs_key key;
 676         struct btrfs_free_space *e, *n;
 677         LIST_HEAD(bitmaps);
 678         u64 num_entries;
 679         u64 num_bitmaps;
 680         u64 generation;
 681         u8 type;
 682         int ret = 0;
 683
 684         /* Nothing in the space cache, goodbye */
 685         if (!i_size_read(inode))
 686                 return 0;
 687
 688         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 689         key.offset = offset;
 690         key.type = 0;
 691
 692         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 693         if (ret < 0)
 694                 return 0;
 695         else if (ret > 0) {
 696                 btrfs_release_path(path);
 697                 return 0;
 698         }
 699
 700         ret = -1;
 701
 702         leaf = path->nodes[0];
 703         header = btrfs_item_ptr(leaf, path->slots[0],
 704                                 struct btrfs_free_space_header);
 705         num_entries = btrfs_free_space_entries(leaf, header);
 706         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 707         generation = btrfs_free_space_generation(leaf, header);
 708         btrfs_release_path(path);
 709
 710         if (!BTRFS_I(inode)->generation) {
 711                 btrfs_info(fs_info,
 712                            "the free space cache file (%llu) is invalid, skip it",
 713                            offset);
 714                 return 0;
 715         }
 716
 717         if (BTRFS_I(inode)->generation != generation) {
 718                 btrfs_err(fs_info,
 719                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 720                           BTRFS_I(inode)->generation, generation);
 721                 return 0;
 722         }
 723
 724         if (!num_entries)
 725                 return 0;
 726
 727         ret = io_ctl_init(&io_ctl, inode, 0);
 728         if (ret)
 729                 return ret;
 730
 731         readahead_cache(inode);
 732
 733         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 734         if (ret)
 735                 goto out;
 736
 737         ret = io_ctl_check_crc(&io_ctl, 0);
 738         if (ret)
 739                 goto free_cache;
 740
 741         ret = io_ctl_check_generation(&io_ctl, generation);
 742         if (ret)
 743                 goto free_cache;
 744
 745         while (num_entries) {
 746                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 747                                       GFP_NOFS);
 748                 if (!e)
 749                         goto free_cache;
 750
 751                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 752                 if (ret) {
 753                         kmem_cache_free(btrfs_free_space_cachep, e);
 754                         goto free_cache;
 755                 }
 756
 757                 if (!e->bytes) {
 758                         kmem_cache_free(btrfs_free_space_cachep, e);
 759                         goto free_cache;
 760                 }
 761
 762                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 763                         spin_lock(&ctl->tree_lock);
 764                         ret = link_free_space(ctl, e);
 765                         spin_unlock(&ctl->tree_lock);
 766                         if (ret) {
 767                                 btrfs_err(fs_info,
 768                                         "Duplicate entries in free space cache, dumping");
 769                                 kmem_cache_free(btrfs_free_space_cachep, e);
 770                                 goto free_cache;
 771                         }
 772                 } else {
 773                         ASSERT(num_bitmaps);
 774                         num_bitmaps--;
 775                         e->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
 776                         if (!e->bitmap) {
 777                                 kmem_cache_free(
 778                                         btrfs_free_space_cachep, e);
 779                                 goto free_cache;
 780                         }
 781                         spin_lock(&ctl->tree_lock);
 782                         ret = link_free_space(ctl, e);
 783                         ctl->total_bitmaps++;
 784                         ctl->op->recalc_thresholds(ctl);
 785                         spin_unlock(&ctl->tree_lock);
 786                         if (ret) {
 787                                 btrfs_err(fs_info,
 788                                         "Duplicate entries in free space cache, dumping");
 789                                 kmem_cache_free(btrfs_free_space_cachep, e);
 790                                 goto free_cache;
 791                         }
 792                         list_add_tail(&e->list, &bitmaps);
 793                 }
 794
 795                 num_entries--;
 796         }
 797
 798         io_ctl_unmap_page(&io_ctl);
 799
 800         /*
 801          * We add the bitmaps at the end of the entries in order that
 802          * the bitmap entries are added to the cache.
 803          */
 804         list_for_each_entry_safe(e, n, &bitmaps, list) {
 805                 list_del_init(&e->list);
 806                 ret = io_ctl_read_bitmap(&io_ctl, e);
 807                 if (ret)
 808                         goto free_cache;
 809         }
 810
 811         io_ctl_drop_pages(&io_ctl);
 812         merge_space_tree(ctl);
 813         ret = 1;
 814 out:
 815         io_ctl_free(&io_ctl);
 816         return ret;
 817 free_cache:
 818         io_ctl_drop_pages(&io_ctl);
 819         __btrfs_remove_free_space_cache(ctl);
 820         goto out;
 821 }
 822
 823 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 824                           struct btrfs_block_group_cache *block_group)
 825 {
 826         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 827         struct inode *inode;
 828         struct btrfs_path *path;
 829         int ret = 0;
 830         bool matched;
 831         u64 used = btrfs_block_group_used(&block_group->item);
 832
 833         /*
 834          * If this block group has been marked to be cleared for one reason or
 835          * another then we can't trust the on disk cache, so just return.
 836          */
 837         spin_lock(&block_group->lock);
 838         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 839                 spin_unlock(&block_group->lock);
 840                 return 0;
 841         }
 842         spin_unlock(&block_group->lock);
 843
 844         path = btrfs_alloc_path();
 845         if (!path)
 846                 return 0;
 847         path->search_commit_root = 1;
 848         path->skip_locking = 1;
 849
 850         inode = lookup_free_space_inode(fs_info, block_group, path);
 851         if (IS_ERR(inode)) {
 852                 btrfs_free_path(path);
 853                 return 0;
 854         }
 855
 856         /* We may have converted the inode and made the cache invalid. */
 857         spin_lock(&block_group->lock);
 858         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 859                 spin_unlock(&block_group->lock);
 860                 btrfs_free_path(path);
 861                 goto out;
 862         }
 863         spin_unlock(&block_group->lock);
 864
 865         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 866                                       path, block_group->key.objectid);
 867         btrfs_free_path(path);
 868         if (ret <= 0)
 869                 goto out;
 870
 871         spin_lock(&ctl->tree_lock);
 872         matched = (ctl->free_space == (block_group->key.offset - used -
 873                                        block_group->bytes_super));
 874         spin_unlock(&ctl->tree_lock);
 875
 876         if (!matched) {
 877                 __btrfs_remove_free_space_cache(ctl);
 878                 btrfs_warn(fs_info,
 879                            "block group %llu has wrong amount of free space",
 880                            block_group->key.objectid);
 881                 ret = -1;
 882         }
 883 out:
 884         if (ret < 0) {
 885                 /* This cache is bogus, make sure it gets cleared */
 886                 spin_lock(&block_group->lock);
 887                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 888                 spin_unlock(&block_group->lock);
 889                 ret = 0;
 890
 891                 btrfs_warn(fs_info,
 892                            "failed to load free space cache for block group %llu, rebuilding it now",
 893                            block_group->key.objectid);
 894         }
 895
 896         iput(inode);
 897         return ret;
 898 }
 899
 900 static noinline_for_stack
 901 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
 902                               struct btrfs_free_space_ctl *ctl,
 903                               struct btrfs_block_group_cache *block_group,
 904                               int *entries, int *bitmaps,
 905                               struct list_head *bitmap_list)
 906 {
 907         int ret;
 908         struct btrfs_free_cluster *cluster = NULL;
 909         struct btrfs_free_cluster *cluster_locked = NULL;
 910         struct rb_node *node = rb_first(&ctl->free_space_offset);
 911         struct btrfs_trim_range *trim_entry;
 912
 913         /* Get the cluster for this block_group if it exists */
 914         if (block_group && !list_empty(&block_group->cluster_list)) {
 915                 cluster = list_entry(block_group->cluster_list.next,
 916                                      struct btrfs_free_cluster,
 917                                      block_group_list);
 918         }
 919
 920         if (!node && cluster) {
 921                 cluster_locked = cluster;
 922                 spin_lock(&cluster_locked->lock);
 923                 node = rb_first(&cluster->root);
 924                 cluster = NULL;
 925         }
 926
 927         /* Write out the extent entries */
 928         while (node) {
 929                 struct btrfs_free_space *e;
 930
 931                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 932                 *entries += 1;
 933
 934                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
 935                                        e->bitmap);
 936                 if (ret)
 937                         goto fail;
 938
 939                 if (e->bitmap) {
 940                         list_add_tail(&e->list, bitmap_list);
 941                         *bitmaps += 1;
 942                 }
 943                 node = rb_next(node);
 944                 if (!node && cluster) {
 945                         node = rb_first(&cluster->root);
 946                         cluster_locked = cluster;
 947                         spin_lock(&cluster_locked->lock);
 948                         cluster = NULL;
 949                 }
 950         }
 951         if (cluster_locked) {
 952                 spin_unlock(&cluster_locked->lock);
 953                 cluster_locked = NULL;
 954         }
 955
 956         /*
 957          * Make sure we don't miss any range that was removed from our rbtree
 958          * because trimming is running. Otherwise after a umount+mount (or crash
 959          * after committing the transaction) we would leak free space and get
 960          * an inconsistent free space cache report from fsck.
 961          */
 962         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
 963                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
 964                                        trim_entry->bytes, NULL);
 965                 if (ret)
 966                         goto fail;
 967                 *entries += 1;
 968         }
 969
 970         return 0;
 971 fail:
 972         if (cluster_locked)
 973                 spin_unlock(&cluster_locked->lock);
 974         return -ENOSPC;
 975 }
 976
 977 static noinline_for_stack int
 978 update_cache_item(struct btrfs_trans_handle *trans,
 979                   struct btrfs_root *root,
 980                   struct inode *inode,
 981                   struct btrfs_path *path, u64 offset,
 982                   int entries, int bitmaps)
 983 {
 984         struct btrfs_key key;
 985         struct btrfs_free_space_header *header;
 986         struct extent_buffer *leaf;
 987         int ret;
 988
 989         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 990         key.offset = offset;
 991         key.type = 0;
 992
 993         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
 994         if (ret < 0) {
 995                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
 996                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL);
 997                 goto fail;
 998         }
 999         leaf = path->nodes[0];
1000         if (ret > 0) {
1001                 struct btrfs_key found_key;
1002                 ASSERT(path->slots[0]);
1003                 path->slots[0]--;
1004                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1005                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1006                     found_key.offset != offset) {
1007                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1008                                          inode->i_size - 1,
1009                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1010                                          NULL);
1011                         btrfs_release_path(path);
1012                         goto fail;
1013                 }
1014         }
1015
1016         BTRFS_I(inode)->generation = trans->transid;
1017         header = btrfs_item_ptr(leaf, path->slots[0],
1018                                 struct btrfs_free_space_header);
1019         btrfs_set_free_space_entries(leaf, header, entries);
1020         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1021         btrfs_set_free_space_generation(leaf, header, trans->transid);
1022         btrfs_mark_buffer_dirty(leaf);
1023         btrfs_release_path(path);
1024
1025         return 0;
1026
1027 fail:
1028         return -1;
1029 }
1030
1031 static noinline_for_stack int
1032 write_pinned_extent_entries(struct btrfs_fs_info *fs_info,
1033                             struct btrfs_block_group_cache *block_group,
1034                             struct btrfs_io_ctl *io_ctl,
1035                             int *entries)
1036 {
1037         u64 start, extent_start, extent_end, len;
1038         struct extent_io_tree *unpin = NULL;
1039         int ret;
1040
1041         if (!block_group)
1042                 return 0;
1043
1044         /*
1045          * We want to add any pinned extents to our free space cache
1046          * so we don't leak the space
1047          *
1048          * We shouldn't have switched the pinned extents yet so this is the
1049          * right one
1050          */
1051         unpin = fs_info->pinned_extents;
1052
1053         start = block_group->key.objectid;
1054
1055         while (start < block_group->key.objectid + block_group->key.offset) {
1056                 ret = find_first_extent_bit(unpin, start,
1057                                             &extent_start, &extent_end,
1058                                             EXTENT_DIRTY, NULL);
1059                 if (ret)
1060                         return 0;
1061
1062                 /* This pinned extent is out of our range */
1063                 if (extent_start >= block_group->key.objectid +
1064                     block_group->key.offset)
1065                         return 0;
1066
1067                 extent_start = max(extent_start, start);
1068                 extent_end = min(block_group->key.objectid +
1069                                  block_group->key.offset, extent_end + 1);
1070                 len = extent_end - extent_start;
1071
1072                 *entries += 1;
1073                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1074                 if (ret)
1075                         return -ENOSPC;
1076
1077                 start = extent_end;
1078         }
1079
1080         return 0;
1081 }
1082
1083 static noinline_for_stack int
1084 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1085 {
1086         struct btrfs_free_space *entry, *next;
1087         int ret;
1088
1089         /* Write out the bitmaps */
1090         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1091                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1092                 if (ret)
1093                         return -ENOSPC;
1094                 list_del_init(&entry->list);
1095         }
1096
1097         return 0;
1098 }
1099
1100 static int flush_dirty_cache(struct inode *inode)
1101 {
1102         int ret;
1103
1104         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1105         if (ret)
1106                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1107                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL);
1108
1109         return ret;
1110 }
1111
1112 static void noinline_for_stack
1113 cleanup_bitmap_list(struct list_head *bitmap_list)
1114 {
1115         struct btrfs_free_space *entry, *next;
1116
1117         list_for_each_entry_safe(entry, next, bitmap_list, list)
1118                 list_del_init(&entry->list);
1119 }
1120
1121 static void noinline_for_stack
1122 cleanup_write_cache_enospc(struct inode *inode,
1123                            struct btrfs_io_ctl *io_ctl,
1124                            struct extent_state **cached_state)
1125 {
1126         io_ctl_drop_pages(io_ctl);
1127         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1128                              i_size_read(inode) - 1, cached_state);
1129 }
1130
1131 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1132                                  struct btrfs_trans_handle *trans,
1133                                  struct btrfs_block_group_cache *block_group,
1134                                  struct btrfs_io_ctl *io_ctl,
1135                                  struct btrfs_path *path, u64 offset)
1136 {
1137         int ret;
1138         struct inode *inode = io_ctl->inode;
1139         struct btrfs_fs_info *fs_info;
1140
1141         if (!inode)
1142                 return 0;
1143
1144         fs_info = btrfs_sb(inode->i_sb);
1145
1146         /* Flush the dirty pages in the cache file. */
1147         ret = flush_dirty_cache(inode);
1148         if (ret)
1149                 goto out;
1150
1151         /* Update the cache item to tell everyone this cache file is valid. */
1152         ret = update_cache_item(trans, root, inode, path, offset,
1153                                 io_ctl->entries, io_ctl->bitmaps);
1154 out:
1155         io_ctl_free(io_ctl);
1156         if (ret) {
1157                 invalidate_inode_pages2(inode->i_mapping);
1158                 BTRFS_I(inode)->generation = 0;
1159                 if (block_group) {
1160 #ifdef DEBUG
1161                         btrfs_err(fs_info,
1162                                   "failed to write free space cache for block group %llu",
1163                                   block_group->key.objectid);
1164 #endif
1165                 }
1166         }
1167         btrfs_update_inode(trans, root, inode);
1168
1169         if (block_group) {
1170                 /* the dirty list is protected by the dirty_bgs_lock */
1171                 spin_lock(&trans->transaction->dirty_bgs_lock);
1172
1173                 /* the disk_cache_state is protected by the block group lock */
1174                 spin_lock(&block_group->lock);
1175
1176                 /*
1177                  * only mark this as written if we didn't get put back on
1178                  * the dirty list while waiting for IO.   Otherwise our
1179                  * cache state won't be right, and we won't get written again
1180                  */
1181                 if (!ret && list_empty(&block_group->dirty_list))
1182                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1183                 else if (ret)
1184                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1185
1186                 spin_unlock(&block_group->lock);
1187                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1188                 io_ctl->inode = NULL;
1189                 iput(inode);
1190         }
1191
1192         return ret;
1193
1194 }
1195
1196 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
1197                                     struct btrfs_trans_handle *trans,
1198                                     struct btrfs_io_ctl *io_ctl,
1199                                     struct btrfs_path *path)
1200 {
1201         return __btrfs_wait_cache_io(root, trans, NULL, io_ctl, path, 0);
1202 }
1203
1204 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1205                         struct btrfs_block_group_cache *block_group,
1206                         struct btrfs_path *path)
1207 {
1208         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1209                                      block_group, &block_group->io_ctl,
1210                                      path, block_group->key.objectid);
1211 }
1212
1213 /**
1214  * __btrfs_write_out_cache - write out cached info to an inode
1215  * @root - the root the inode belongs to
1216  * @ctl - the free space cache we are going to write out
1217  * @block_group - the block_group for this cache if it belongs to a block_group
1218  * @trans - the trans handle
1219  *
1220  * This function writes out a free space cache struct to disk for quick recovery
1221  * on mount.  This will return 0 if it was successful in writing the cache out,
1222  * or an errno if it was not.
1223  */
1224 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1225                                    struct btrfs_free_space_ctl *ctl,
1226                                    struct btrfs_block_group_cache *block_group,
1227                                    struct btrfs_io_ctl *io_ctl,
1228                                    struct btrfs_trans_handle *trans)
1229 {
1230         struct btrfs_fs_info *fs_info = root->fs_info;
1231         struct extent_state *cached_state = NULL;
1232         LIST_HEAD(bitmap_list);
1233         int entries = 0;
1234         int bitmaps = 0;
1235         int ret;
1236         int must_iput = 0;
1237
1238         if (!i_size_read(inode))
1239                 return -EIO;
1240
1241         WARN_ON(io_ctl->pages);
1242         ret = io_ctl_init(io_ctl, inode, 1);
1243         if (ret)
1244                 return ret;
1245
1246         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1247                 down_write(&block_group->data_rwsem);
1248                 spin_lock(&block_group->lock);
1249                 if (block_group->delalloc_bytes) {
1250                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1251                         spin_unlock(&block_group->lock);
1252                         up_write(&block_group->data_rwsem);
1253                         BTRFS_I(inode)->generation = 0;
1254                         ret = 0;
1255                         must_iput = 1;
1256                         goto out;
1257                 }
1258                 spin_unlock(&block_group->lock);
1259         }
1260
1261         /* Lock all pages first so we can lock the extent safely. */
1262         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1263         if (ret)
1264                 goto out_unlock;
1265
1266         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1267                          &cached_state);
1268
1269         io_ctl_set_generation(io_ctl, trans->transid);
1270
1271         mutex_lock(&ctl->cache_writeout_mutex);
1272         /* Write out the extent entries in the free space cache */
1273         spin_lock(&ctl->tree_lock);
1274         ret = write_cache_extent_entries(io_ctl, ctl,
1275                                          block_group, &entries, &bitmaps,
1276                                          &bitmap_list);
1277         if (ret)
1278                 goto out_nospc_locked;
1279
1280         /*
1281          * Some spaces that are freed in the current transaction are pinned,
1282          * they will be added into free space cache after the transaction is
1283          * committed, we shouldn't lose them.
1284          *
1285          * If this changes while we are working we'll get added back to
1286          * the dirty list and redo it.  No locking needed
1287          */
1288         ret = write_pinned_extent_entries(fs_info, block_group,
1289                                           io_ctl, &entries);
1290         if (ret)
1291                 goto out_nospc_locked;
1292
1293         /*
1294          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1295          * locked while doing it because a concurrent trim can be manipulating
1296          * or freeing the bitmap.
1297          */
1298         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1299         spin_unlock(&ctl->tree_lock);
1300         mutex_unlock(&ctl->cache_writeout_mutex);
1301         if (ret)
1302                 goto out_nospc;
1303
1304         /* Zero out the rest of the pages just to make sure */
1305         io_ctl_zero_remaining_pages(io_ctl);
1306
1307         /* Everything is written out, now we dirty the pages in the file. */
1308         ret = btrfs_dirty_pages(inode, io_ctl->pages, io_ctl->num_pages, 0,
1309                                 i_size_read(inode), &cached_state);
1310         if (ret)
1311                 goto out_nospc;
1312
1313         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1314                 up_write(&block_group->data_rwsem);
1315         /*
1316          * Release the pages and unlock the extent, we will flush
1317          * them out later
1318          */
1319         io_ctl_drop_pages(io_ctl);
1320
1321         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1322                              i_size_read(inode) - 1, &cached_state);
1323
1324         /*
1325          * at this point the pages are under IO and we're happy,
1326          * The caller is responsible for waiting on them and updating the
1327          * the cache and the inode
1328          */
1329         io_ctl->entries = entries;
1330         io_ctl->bitmaps = bitmaps;
1331
1332         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1333         if (ret)
1334                 goto out;
1335
1336         return 0;
1337
1338 out:
1339         io_ctl->inode = NULL;
1340         io_ctl_free(io_ctl);
1341         if (ret) {
1342                 invalidate_inode_pages2(inode->i_mapping);
1343                 BTRFS_I(inode)->generation = 0;
1344         }
1345         btrfs_update_inode(trans, root, inode);
1346         if (must_iput)
1347                 iput(inode);
1348         return ret;
1349
1350 out_nospc_locked:
1351         cleanup_bitmap_list(&bitmap_list);
1352         spin_unlock(&ctl->tree_lock);
1353         mutex_unlock(&ctl->cache_writeout_mutex);
1354
1355 out_nospc:
1356         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1357
1358 out_unlock:
1359         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1360                 up_write(&block_group->data_rwsem);
1361
1362         goto out;
1363 }
1364
1365 int btrfs_write_out_cache(struct btrfs_fs_info *fs_info,
1366                           struct btrfs_trans_handle *trans,
1367                           struct btrfs_block_group_cache *block_group,
1368                           struct btrfs_path *path)
1369 {
1370         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1371         struct inode *inode;
1372         int ret = 0;
1373
1374         spin_lock(&block_group->lock);
1375         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1376                 spin_unlock(&block_group->lock);
1377                 return 0;
1378         }
1379         spin_unlock(&block_group->lock);
1380
1381         inode = lookup_free_space_inode(fs_info, block_group, path);
1382         if (IS_ERR(inode))
1383                 return 0;
1384
1385         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1386                                 block_group, &block_group->io_ctl, trans);
1387         if (ret) {
1388 #ifdef DEBUG
1389                 btrfs_err(fs_info,
1390                           "failed to write free space cache for block group %llu",
1391                           block_group->key.objectid);
1392 #endif
1393                 spin_lock(&block_group->lock);
1394                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1395                 spin_unlock(&block_group->lock);
1396
1397                 block_group->io_ctl.inode = NULL;
1398                 iput(inode);
1399         }
1400
1401         /*
1402          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1403          * to wait for IO and put the inode
1404          */
1405
1406         return ret;
1407 }
1408
1409 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1410                                           u64 offset)
1411 {
1412         ASSERT(offset >= bitmap_start);
1413         offset -= bitmap_start;
1414         return (unsigned long)(div_u64(offset, unit));
1415 }
1416
1417 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1418 {
1419         return (unsigned long)(div_u64(bytes, unit));
1420 }
1421
1422 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1423                                    u64 offset)
1424 {
1425         u64 bitmap_start;
1426         u64 bytes_per_bitmap;
1427
1428         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1429         bitmap_start = offset - ctl->start;
1430         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1431         bitmap_start *= bytes_per_bitmap;
1432         bitmap_start += ctl->start;
1433
1434         return bitmap_start;
1435 }
1436
1437 static int tree_insert_offset(struct rb_root *root, u64 offset,
1438                               struct rb_node *node, int bitmap)
1439 {
1440         struct rb_node **p = &root->rb_node;
1441         struct rb_node *parent = NULL;
1442         struct btrfs_free_space *info;
1443
1444         while (*p) {
1445                 parent = *p;
1446                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1447
1448                 if (offset < info->offset) {
1449                         p = &(*p)->rb_left;
1450                 } else if (offset > info->offset) {
1451                         p = &(*p)->rb_right;
1452                 } else {
1453                         /*
1454                          * we could have a bitmap entry and an extent entry
1455                          * share the same offset.  If this is the case, we want
1456                          * the extent entry to always be found first if we do a
1457                          * linear search through the tree, since we want to have
1458                          * the quickest allocation time, and allocating from an
1459                          * extent is faster than allocating from a bitmap.  So
1460                          * if we're inserting a bitmap and we find an entry at
1461                          * this offset, we want to go right, or after this entry
1462                          * logically.  If we are inserting an extent and we've
1463                          * found a bitmap, we want to go left, or before
1464                          * logically.
1465                          */
1466                         if (bitmap) {
1467                                 if (info->bitmap) {
1468                                         WARN_ON_ONCE(1);
1469                                         return -EEXIST;
1470                                 }
1471                                 p = &(*p)->rb_right;
1472                         } else {
1473                                 if (!info->bitmap) {
1474                                         WARN_ON_ONCE(1);
1475                                         return -EEXIST;
1476                                 }
1477                                 p = &(*p)->rb_left;
1478                         }
1479                 }
1480         }
1481
1482         rb_link_node(node, parent, p);
1483         rb_insert_color(node, root);
1484
1485         return 0;
1486 }
1487
1488 /*
1489  * searches the tree for the given offset.
1490  *
1491  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1492  * want a section that has at least bytes size and comes at or after the given
1493  * offset.
1494  */
1495 static struct btrfs_free_space *
1496 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1497                    u64 offset, int bitmap_only, int fuzzy)
1498 {
1499         struct rb_node *n = ctl->free_space_offset.rb_node;
1500         struct btrfs_free_space *entry, *prev = NULL;
1501
1502         /* find entry that is closest to the 'offset' */
1503         while (1) {
1504                 if (!n) {
1505                         entry = NULL;
1506                         break;
1507                 }
1508
1509                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1510                 prev = entry;
1511
1512                 if (offset < entry->offset)
1513                         n = n->rb_left;
1514                 else if (offset > entry->offset)
1515                         n = n->rb_right;
1516                 else
1517                         break;
1518         }
1519
1520         if (bitmap_only) {
1521                 if (!entry)
1522                         return NULL;
1523                 if (entry->bitmap)
1524                         return entry;
1525
1526                 /*
1527                  * bitmap entry and extent entry may share same offset,
1528                  * in that case, bitmap entry comes after extent entry.
1529                  */
1530                 n = rb_next(n);
1531                 if (!n)
1532                         return NULL;
1533                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1534                 if (entry->offset != offset)
1535                         return NULL;
1536
1537                 WARN_ON(!entry->bitmap);
1538                 return entry;
1539         } else if (entry) {
1540                 if (entry->bitmap) {
1541                         /*
1542                          * if previous extent entry covers the offset,
1543                          * we should return it instead of the bitmap entry
1544                          */
1545                         n = rb_prev(&entry->offset_index);
1546                         if (n) {
1547                                 prev = rb_entry(n, struct btrfs_free_space,
1548                                                 offset_index);
1549                                 if (!prev->bitmap &&
1550                                     prev->offset + prev->bytes > offset)
1551                                         entry = prev;
1552                         }
1553                 }
1554                 return entry;
1555         }
1556
1557         if (!prev)
1558                 return NULL;
1559
1560         /* find last entry before the 'offset' */
1561         entry = prev;
1562         if (entry->offset > offset) {
1563                 n = rb_prev(&entry->offset_index);
1564                 if (n) {
1565                         entry = rb_entry(n, struct btrfs_free_space,
1566                                         offset_index);
1567                         ASSERT(entry->offset <= offset);
1568                 } else {
1569                         if (fuzzy)
1570                                 return entry;
1571                         else
1572                                 return NULL;
1573                 }
1574         }
1575
1576         if (entry->bitmap) {
1577                 n = rb_prev(&entry->offset_index);
1578                 if (n) {
1579                         prev = rb_entry(n, struct btrfs_free_space,
1580                                         offset_index);
1581                         if (!prev->bitmap &&
1582                             prev->offset + prev->bytes > offset)
1583                                 return prev;
1584                 }
1585                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1586                         return entry;
1587         } else if (entry->offset + entry->bytes > offset)
1588                 return entry;
1589
1590         if (!fuzzy)
1591                 return NULL;
1592
1593         while (1) {
1594                 if (entry->bitmap) {
1595                         if (entry->offset + BITS_PER_BITMAP *
1596                             ctl->unit > offset)
1597                                 break;
1598                 } else {
1599                         if (entry->offset + entry->bytes > offset)
1600                                 break;
1601                 }
1602
1603                 n = rb_next(&entry->offset_index);
1604                 if (!n)
1605                         return NULL;
1606                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1607         }
1608         return entry;
1609 }
1610
1611 static inline void
1612 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1613                     struct btrfs_free_space *info)
1614 {
1615         rb_erase(&info->offset_index, &ctl->free_space_offset);
1616         ctl->free_extents--;
1617 }
1618
1619 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1620                               struct btrfs_free_space *info)
1621 {
1622         __unlink_free_space(ctl, info);
1623         ctl->free_space -= info->bytes;
1624 }
1625
1626 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1627                            struct btrfs_free_space *info)
1628 {
1629         int ret = 0;
1630
1631         ASSERT(info->bytes || info->bitmap);
1632         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1633                                  &info->offset_index, (info->bitmap != NULL));
1634         if (ret)
1635                 return ret;
1636
1637         ctl->free_space += info->bytes;
1638         ctl->free_extents++;
1639         return ret;
1640 }
1641
1642 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1643 {
1644         struct btrfs_block_group_cache *block_group = ctl->private;
1645         u64 max_bytes;
1646         u64 bitmap_bytes;
1647         u64 extent_bytes;
1648         u64 size = block_group->key.offset;
1649         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1650         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1651
1652         max_bitmaps = max_t(u64, max_bitmaps, 1);
1653
1654         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1655
1656         /*
1657          * The goal is to keep the total amount of memory used per 1gb of space
1658          * at or below 32k, so we need to adjust how much memory we allow to be
1659          * used by extent based free space tracking
1660          */
1661         if (size < SZ_1G)
1662                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1663         else
1664                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1665
1666         /*
1667          * we want to account for 1 more bitmap than what we have so we can make
1668          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1669          * we add more bitmaps.
1670          */
1671         bitmap_bytes = (ctl->total_bitmaps + 1) * ctl->unit;
1672
1673         if (bitmap_bytes >= max_bytes) {
1674                 ctl->extents_thresh = 0;
1675                 return;
1676         }
1677
1678         /*
1679          * we want the extent entry threshold to always be at most 1/2 the max
1680          * bytes we can have, or whatever is less than that.
1681          */
1682         extent_bytes = max_bytes - bitmap_bytes;
1683         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1684
1685         ctl->extents_thresh =
1686                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1687 }
1688
1689 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1690                                        struct btrfs_free_space *info,
1691                                        u64 offset, u64 bytes)
1692 {
1693         unsigned long start, count;
1694
1695         start = offset_to_bit(info->offset, ctl->unit, offset);
1696         count = bytes_to_bits(bytes, ctl->unit);
1697         ASSERT(start + count <= BITS_PER_BITMAP);
1698
1699         bitmap_clear(info->bitmap, start, count);
1700
1701         info->bytes -= bytes;
1702 }
1703
1704 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1705                               struct btrfs_free_space *info, u64 offset,
1706                               u64 bytes)
1707 {
1708         __bitmap_clear_bits(ctl, info, offset, bytes);
1709         ctl->free_space -= bytes;
1710 }
1711
1712 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1713                             struct btrfs_free_space *info, u64 offset,
1714                             u64 bytes)
1715 {
1716         unsigned long start, count;
1717
1718         start = offset_to_bit(info->offset, ctl->unit, offset);
1719         count = bytes_to_bits(bytes, ctl->unit);
1720         ASSERT(start + count <= BITS_PER_BITMAP);
1721
1722         bitmap_set(info->bitmap, start, count);
1723
1724         info->bytes += bytes;
1725         ctl->free_space += bytes;
1726 }
1727
1728 /*
1729  * If we can not find suitable extent, we will use bytes to record
1730  * the size of the max extent.
1731  */
1732 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1733                          struct btrfs_free_space *bitmap_info, u64 *offset,
1734                          u64 *bytes, bool for_alloc)
1735 {
1736         unsigned long found_bits = 0;
1737         unsigned long max_bits = 0;
1738         unsigned long bits, i;
1739         unsigned long next_zero;
1740         unsigned long extent_bits;
1741
1742         /*
1743          * Skip searching the bitmap if we don't have a contiguous section that
1744          * is large enough for this allocation.
1745          */
1746         if (for_alloc &&
1747             bitmap_info->max_extent_size &&
1748             bitmap_info->max_extent_size < *bytes) {
1749                 *bytes = bitmap_info->max_extent_size;
1750                 return -1;
1751         }
1752
1753         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1754                           max_t(u64, *offset, bitmap_info->offset));
1755         bits = bytes_to_bits(*bytes, ctl->unit);
1756
1757         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1758                 if (for_alloc && bits == 1) {
1759                         found_bits = 1;
1760                         break;
1761                 }
1762                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1763                                                BITS_PER_BITMAP, i);
1764                 extent_bits = next_zero - i;
1765                 if (extent_bits >= bits) {
1766                         found_bits = extent_bits;
1767                         break;
1768                 } else if (extent_bits > max_bits) {
1769                         max_bits = extent_bits;
1770                 }
1771                 i = next_zero;
1772         }
1773
1774         if (found_bits) {
1775                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1776                 *bytes = (u64)(found_bits) * ctl->unit;
1777                 return 0;
1778         }
1779
1780         *bytes = (u64)(max_bits) * ctl->unit;
1781         bitmap_info->max_extent_size = *bytes;
1782         return -1;
1783 }
1784
1785 /* Cache the size of the max extent in bytes */
1786 static struct btrfs_free_space *
1787 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1788                 unsigned long align, u64 *max_extent_size)
1789 {
1790         struct btrfs_free_space *entry;
1791         struct rb_node *node;
1792         u64 tmp;
1793         u64 align_off;
1794         int ret;
1795
1796         if (!ctl->free_space_offset.rb_node)
1797                 goto out;
1798
1799         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1800         if (!entry)
1801                 goto out;
1802
1803         for (node = &entry->offset_index; node; node = rb_next(node)) {
1804                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1805                 if (entry->bytes < *bytes) {
1806                         if (entry->bytes > *max_extent_size)
1807                                 *max_extent_size = entry->bytes;
1808                         continue;
1809                 }
1810
1811                 /* make sure the space returned is big enough
1812                  * to match our requested alignment
1813                  */
1814                 if (*bytes >= align) {
1815                         tmp = entry->offset - ctl->start + align - 1;
1816                         tmp = div64_u64(tmp, align);
1817                         tmp = tmp * align + ctl->start;
1818                         align_off = tmp - entry->offset;
1819                 } else {
1820                         align_off = 0;
1821                         tmp = entry->offset;
1822                 }
1823
1824                 if (entry->bytes < *bytes + align_off) {
1825                         if (entry->bytes > *max_extent_size)
1826                                 *max_extent_size = entry->bytes;
1827                         continue;
1828                 }
1829
1830                 if (entry->bitmap) {
1831                         u64 size = *bytes;
1832
1833                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1834                         if (!ret) {
1835                                 *offset = tmp;
1836                                 *bytes = size;
1837                                 return entry;
1838                         } else if (size > *max_extent_size) {
1839                                 *max_extent_size = size;
1840                         }
1841                         continue;
1842                 }
1843
1844                 *offset = tmp;
1845                 *bytes = entry->bytes - align_off;
1846                 return entry;
1847         }
1848 out:
1849         return NULL;
1850 }
1851
1852 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1853                            struct btrfs_free_space *info, u64 offset)
1854 {
1855         info->offset = offset_to_bitmap(ctl, offset);
1856         info->bytes = 0;
1857         INIT_LIST_HEAD(&info->list);
1858         link_free_space(ctl, info);
1859         ctl->total_bitmaps++;
1860
1861         ctl->op->recalc_thresholds(ctl);
1862 }
1863
1864 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1865                         struct btrfs_free_space *bitmap_info)
1866 {
1867         unlink_free_space(ctl, bitmap_info);
1868         kfree(bitmap_info->bitmap);
1869         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1870         ctl->total_bitmaps--;
1871         ctl->op->recalc_thresholds(ctl);
1872 }
1873
1874 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1875                               struct btrfs_free_space *bitmap_info,
1876                               u64 *offset, u64 *bytes)
1877 {
1878         u64 end;
1879         u64 search_start, search_bytes;
1880         int ret;
1881
1882 again:
1883         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1884
1885         /*
1886          * We need to search for bits in this bitmap.  We could only cover some
1887          * of the extent in this bitmap thanks to how we add space, so we need
1888          * to search for as much as it as we can and clear that amount, and then
1889          * go searching for the next bit.
1890          */
1891         search_start = *offset;
1892         search_bytes = ctl->unit;
1893         search_bytes = min(search_bytes, end - search_start + 1);
1894         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1895                             false);
1896         if (ret < 0 || search_start != *offset)
1897                 return -EINVAL;
1898
1899         /* We may have found more bits than what we need */
1900         search_bytes = min(search_bytes, *bytes);
1901
1902         /* Cannot clear past the end of the bitmap */
1903         search_bytes = min(search_bytes, end - search_start + 1);
1904
1905         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1906         *offset += search_bytes;
1907         *bytes -= search_bytes;
1908
1909         if (*bytes) {
1910                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1911                 if (!bitmap_info->bytes)
1912                         free_bitmap(ctl, bitmap_info);
1913
1914                 /*
1915                  * no entry after this bitmap, but we still have bytes to
1916                  * remove, so something has gone wrong.
1917                  */
1918                 if (!next)
1919                         return -EINVAL;
1920
1921                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1922                                        offset_index);
1923
1924                 /*
1925                  * if the next entry isn't a bitmap we need to return to let the
1926                  * extent stuff do its work.
1927                  */
1928                 if (!bitmap_info->bitmap)
1929                         return -EAGAIN;
1930
1931                 /*
1932                  * Ok the next item is a bitmap, but it may not actually hold
1933                  * the information for the rest of this free space stuff, so
1934                  * look for it, and if we don't find it return so we can try
1935                  * everything over again.
1936                  */
1937                 search_start = *offset;
1938                 search_bytes = ctl->unit;
1939                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1940                                     &search_bytes, false);
1941                 if (ret < 0 || search_start != *offset)
1942                         return -EAGAIN;
1943
1944                 goto again;
1945         } else if (!bitmap_info->bytes)
1946                 free_bitmap(ctl, bitmap_info);
1947
1948         return 0;
1949 }
1950
1951 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1952                                struct btrfs_free_space *info, u64 offset,
1953                                u64 bytes)
1954 {
1955         u64 bytes_to_set = 0;
1956         u64 end;
1957
1958         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1959
1960         bytes_to_set = min(end - offset, bytes);
1961
1962         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1963
1964         /*
1965          * We set some bytes, we have no idea what the max extent size is
1966          * anymore.
1967          */
1968         info->max_extent_size = 0;
1969
1970         return bytes_to_set;
1971
1972 }
1973
1974 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1975                       struct btrfs_free_space *info)
1976 {
1977         struct btrfs_block_group_cache *block_group = ctl->private;
1978         struct btrfs_fs_info *fs_info = block_group->fs_info;
1979         bool forced = false;
1980
1981 #ifdef CONFIG_BTRFS_DEBUG
1982         if (btrfs_should_fragment_free_space(block_group))
1983                 forced = true;
1984 #endif
1985
1986         /*
1987          * If we are below the extents threshold then we can add this as an
1988          * extent, and don't have to deal with the bitmap
1989          */
1990         if (!forced && ctl->free_extents < ctl->extents_thresh) {
1991                 /*
1992                  * If this block group has some small extents we don't want to
1993                  * use up all of our free slots in the cache with them, we want
1994                  * to reserve them to larger extents, however if we have plenty
1995                  * of cache left then go ahead an dadd them, no sense in adding
1996                  * the overhead of a bitmap if we don't have to.
1997                  */
1998                 if (info->bytes <= fs_info->sectorsize * 4) {
1999                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
2000                                 return false;
2001                 } else {
2002                         return false;
2003                 }
2004         }
2005
2006         /*
2007          * The original block groups from mkfs can be really small, like 8
2008          * megabytes, so don't bother with a bitmap for those entries.  However
2009          * some block groups can be smaller than what a bitmap would cover but
2010          * are still large enough that they could overflow the 32k memory limit,
2011          * so allow those block groups to still be allowed to have a bitmap
2012          * entry.
2013          */
2014         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
2015                 return false;
2016
2017         return true;
2018 }
2019
2020 static const struct btrfs_free_space_op free_space_op = {
2021         .recalc_thresholds      = recalculate_thresholds,
2022         .use_bitmap             = use_bitmap,
2023 };
2024
2025 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2026                               struct btrfs_free_space *info)
2027 {
2028         struct btrfs_free_space *bitmap_info;
2029         struct btrfs_block_group_cache *block_group = NULL;
2030         int added = 0;
2031         u64 bytes, offset, bytes_added;
2032         int ret;
2033
2034         bytes = info->bytes;
2035         offset = info->offset;
2036
2037         if (!ctl->op->use_bitmap(ctl, info))
2038                 return 0;
2039
2040         if (ctl->op == &free_space_op)
2041                 block_group = ctl->private;
2042 again:
2043         /*
2044          * Since we link bitmaps right into the cluster we need to see if we
2045          * have a cluster here, and if so and it has our bitmap we need to add
2046          * the free space to that bitmap.
2047          */
2048         if (block_group && !list_empty(&block_group->cluster_list)) {
2049                 struct btrfs_free_cluster *cluster;
2050                 struct rb_node *node;
2051                 struct btrfs_free_space *entry;
2052
2053                 cluster = list_entry(block_group->cluster_list.next,
2054                                      struct btrfs_free_cluster,
2055                                      block_group_list);
2056                 spin_lock(&cluster->lock);
2057                 node = rb_first(&cluster->root);
2058                 if (!node) {
2059                         spin_unlock(&cluster->lock);
2060                         goto no_cluster_bitmap;
2061                 }
2062
2063                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2064                 if (!entry->bitmap) {
2065                         spin_unlock(&cluster->lock);
2066                         goto no_cluster_bitmap;
2067                 }
2068
2069                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2070                         bytes_added = add_bytes_to_bitmap(ctl, entry,
2071                                                           offset, bytes);
2072                         bytes -= bytes_added;
2073                         offset += bytes_added;
2074                 }
2075                 spin_unlock(&cluster->lock);
2076                 if (!bytes) {
2077                         ret = 1;
2078                         goto out;
2079                 }
2080         }
2081
2082 no_cluster_bitmap:
2083         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2084                                          1, 0);
2085         if (!bitmap_info) {
2086                 ASSERT(added == 0);
2087                 goto new_bitmap;
2088         }
2089
2090         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2091         bytes -= bytes_added;
2092         offset += bytes_added;
2093         added = 0;
2094
2095         if (!bytes) {
2096                 ret = 1;
2097                 goto out;
2098         } else
2099                 goto again;
2100
2101 new_bitmap:
2102         if (info && info->bitmap) {
2103                 add_new_bitmap(ctl, info, offset);
2104                 added = 1;
2105                 info = NULL;
2106                 goto again;
2107         } else {
2108                 spin_unlock(&ctl->tree_lock);
2109
2110                 /* no pre-allocated info, allocate a new one */
2111                 if (!info) {
2112                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2113                                                  GFP_NOFS);
2114                         if (!info) {
2115                                 spin_lock(&ctl->tree_lock);
2116                                 ret = -ENOMEM;
2117                                 goto out;
2118                         }
2119                 }
2120
2121                 /* allocate the bitmap */
2122                 info->bitmap = kzalloc(PAGE_SIZE, GFP_NOFS);
2123                 spin_lock(&ctl->tree_lock);
2124                 if (!info->bitmap) {
2125                         ret = -ENOMEM;
2126                         goto out;
2127                 }
2128                 goto again;
2129         }
2130
2131 out:
2132         if (info) {
2133                 if (info->bitmap)
2134                         kfree(info->bitmap);
2135                 kmem_cache_free(btrfs_free_space_cachep, info);
2136         }
2137
2138         return ret;
2139 }
2140
2141 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2142                           struct btrfs_free_space *info, bool update_stat)
2143 {
2144         struct btrfs_free_space *left_info;
2145         struct btrfs_free_space *right_info;
2146         bool merged = false;
2147         u64 offset = info->offset;
2148         u64 bytes = info->bytes;
2149
2150         /*
2151          * first we want to see if there is free space adjacent to the range we
2152          * are adding, if there is remove that struct and add a new one to
2153          * cover the entire range
2154          */
2155         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2156         if (right_info && rb_prev(&right_info->offset_index))
2157                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2158                                      struct btrfs_free_space, offset_index);
2159         else
2160                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2161
2162         if (right_info && !right_info->bitmap) {
2163                 if (update_stat)
2164                         unlink_free_space(ctl, right_info);
2165                 else
2166                         __unlink_free_space(ctl, right_info);
2167                 info->bytes += right_info->bytes;
2168                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2169                 merged = true;
2170         }
2171
2172         if (left_info && !left_info->bitmap &&
2173             left_info->offset + left_info->bytes == offset) {
2174                 if (update_stat)
2175                         unlink_free_space(ctl, left_info);
2176                 else
2177                         __unlink_free_space(ctl, left_info);
2178                 info->offset = left_info->offset;
2179                 info->bytes += left_info->bytes;
2180                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2181                 merged = true;
2182         }
2183
2184         return merged;
2185 }
2186
2187 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2188                                      struct btrfs_free_space *info,
2189                                      bool update_stat)
2190 {
2191         struct btrfs_free_space *bitmap;
2192         unsigned long i;
2193         unsigned long j;
2194         const u64 end = info->offset + info->bytes;
2195         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2196         u64 bytes;
2197
2198         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2199         if (!bitmap)
2200                 return false;
2201
2202         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2203         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2204         if (j == i)
2205                 return false;
2206         bytes = (j - i) * ctl->unit;
2207         info->bytes += bytes;
2208
2209         if (update_stat)
2210                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2211         else
2212                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2213
2214         if (!bitmap->bytes)
2215                 free_bitmap(ctl, bitmap);
2216
2217         return true;
2218 }
2219
2220 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2221                                        struct btrfs_free_space *info,
2222                                        bool update_stat)
2223 {
2224         struct btrfs_free_space *bitmap;
2225         u64 bitmap_offset;
2226         unsigned long i;
2227         unsigned long j;
2228         unsigned long prev_j;
2229         u64 bytes;
2230
2231         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2232         /* If we're on a boundary, try the previous logical bitmap. */
2233         if (bitmap_offset == info->offset) {
2234                 if (info->offset == 0)
2235                         return false;
2236                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2237         }
2238
2239         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2240         if (!bitmap)
2241                 return false;
2242
2243         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2244         j = 0;
2245         prev_j = (unsigned long)-1;
2246         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2247                 if (j > i)
2248                         break;
2249                 prev_j = j;
2250         }
2251         if (prev_j == i)
2252                 return false;
2253
2254         if (prev_j == (unsigned long)-1)
2255                 bytes = (i + 1) * ctl->unit;
2256         else
2257                 bytes = (i - prev_j) * ctl->unit;
2258
2259         info->offset -= bytes;
2260         info->bytes += bytes;
2261
2262         if (update_stat)
2263                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2264         else
2265                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2266
2267         if (!bitmap->bytes)
2268                 free_bitmap(ctl, bitmap);
2269
2270         return true;
2271 }
2272
2273 /*
2274  * We prefer always to allocate from extent entries, both for clustered and
2275  * non-clustered allocation requests. So when attempting to add a new extent
2276  * entry, try to see if there's adjacent free space in bitmap entries, and if
2277  * there is, migrate that space from the bitmaps to the extent.
2278  * Like this we get better chances of satisfying space allocation requests
2279  * because we attempt to satisfy them based on a single cache entry, and never
2280  * on 2 or more entries - even if the entries represent a contiguous free space
2281  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2282  * ends).
2283  */
2284 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2285                               struct btrfs_free_space *info,
2286                               bool update_stat)
2287 {
2288         /*
2289          * Only work with disconnected entries, as we can change their offset,
2290          * and must be extent entries.
2291          */
2292         ASSERT(!info->bitmap);
2293         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2294
2295         if (ctl->total_bitmaps > 0) {
2296                 bool stole_end;
2297                 bool stole_front = false;
2298
2299                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2300                 if (ctl->total_bitmaps > 0)
2301                         stole_front = steal_from_bitmap_to_front(ctl, info,
2302                                                                  update_stat);
2303
2304                 if (stole_end || stole_front)
2305                         try_merge_free_space(ctl, info, update_stat);
2306         }
2307 }
2308
2309 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2310                            struct btrfs_free_space_ctl *ctl,
2311                            u64 offset, u64 bytes)
2312 {
2313         struct btrfs_free_space *info;
2314         int ret = 0;
2315
2316         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2317         if (!info)
2318                 return -ENOMEM;
2319
2320         info->offset = offset;
2321         info->bytes = bytes;
2322         RB_CLEAR_NODE(&info->offset_index);
2323
2324         spin_lock(&ctl->tree_lock);
2325
2326         if (try_merge_free_space(ctl, info, true))
2327                 goto link;
2328
2329         /*
2330          * There was no extent directly to the left or right of this new
2331          * extent then we know we're going to have to allocate a new extent, so
2332          * before we do that see if we need to drop this into a bitmap
2333          */
2334         ret = insert_into_bitmap(ctl, info);
2335         if (ret < 0) {
2336                 goto out;
2337         } else if (ret) {
2338                 ret = 0;
2339                 goto out;
2340         }
2341 link:
2342         /*
2343          * Only steal free space from adjacent bitmaps if we're sure we're not
2344          * going to add the new free space to existing bitmap entries - because
2345          * that would mean unnecessary work that would be reverted. Therefore
2346          * attempt to steal space from bitmaps if we're adding an extent entry.
2347          */
2348         steal_from_bitmap(ctl, info, true);
2349
2350         ret = link_free_space(ctl, info);
2351         if (ret)
2352                 kmem_cache_free(btrfs_free_space_cachep, info);
2353 out:
2354         spin_unlock(&ctl->tree_lock);
2355
2356         if (ret) {
2357                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2358                 ASSERT(ret != -EEXIST);
2359         }
2360
2361         return ret;
2362 }
2363
2364 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
2365                             u64 offset, u64 bytes)
2366 {
2367         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2368         struct btrfs_free_space *info;
2369         int ret;
2370         bool re_search = false;
2371
2372         spin_lock(&ctl->tree_lock);
2373
2374 again:
2375         ret = 0;
2376         if (!bytes)
2377                 goto out_lock;
2378
2379         info = tree_search_offset(ctl, offset, 0, 0);
2380         if (!info) {
2381                 /*
2382                  * oops didn't find an extent that matched the space we wanted
2383                  * to remove, look for a bitmap instead
2384                  */
2385                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2386                                           1, 0);
2387                 if (!info) {
2388                         /*
2389                          * If we found a partial bit of our free space in a
2390                          * bitmap but then couldn't find the other part this may
2391                          * be a problem, so WARN about it.
2392                          */
2393                         WARN_ON(re_search);
2394                         goto out_lock;
2395                 }
2396         }
2397
2398         re_search = false;
2399         if (!info->bitmap) {
2400                 unlink_free_space(ctl, info);
2401                 if (offset == info->offset) {
2402                         u64 to_free = min(bytes, info->bytes);
2403
2404                         info->bytes -= to_free;
2405                         info->offset += to_free;
2406                         if (info->bytes) {
2407                                 ret = link_free_space(ctl, info);
2408                                 WARN_ON(ret);
2409                         } else {
2410                                 kmem_cache_free(btrfs_free_space_cachep, info);
2411                         }
2412
2413                         offset += to_free;
2414                         bytes -= to_free;
2415                         goto again;
2416                 } else {
2417                         u64 old_end = info->bytes + info->offset;
2418
2419                         info->bytes = offset - info->offset;
2420                         ret = link_free_space(ctl, info);
2421                         WARN_ON(ret);
2422                         if (ret)
2423                                 goto out_lock;
2424
2425                         /* Not enough bytes in this entry to satisfy us */
2426                         if (old_end < offset + bytes) {
2427                                 bytes -= old_end - offset;
2428                                 offset = old_end;
2429                                 goto again;
2430                         } else if (old_end == offset + bytes) {
2431                                 /* all done */
2432                                 goto out_lock;
2433                         }
2434                         spin_unlock(&ctl->tree_lock);
2435
2436                         ret = btrfs_add_free_space(block_group, offset + bytes,
2437                                                    old_end - (offset + bytes));
2438                         WARN_ON(ret);
2439                         goto out;
2440                 }
2441         }
2442
2443         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2444         if (ret == -EAGAIN) {
2445                 re_search = true;
2446                 goto again;
2447         }
2448 out_lock:
2449         spin_unlock(&ctl->tree_lock);
2450 out:
2451         return ret;
2452 }
2453
2454 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
2455                            u64 bytes)
2456 {
2457         struct btrfs_fs_info *fs_info = block_group->fs_info;
2458         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2459         struct btrfs_free_space *info;
2460         struct rb_node *n;
2461         int count = 0;
2462
2463         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2464                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2465                 if (info->bytes >= bytes && !block_group->ro)
2466                         count++;
2467                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2468                            info->offset, info->bytes,
2469                        (info->bitmap) ? "yes" : "no");
2470         }
2471         btrfs_info(fs_info, "block group has cluster?: %s",
2472                list_empty(&block_group->cluster_list) ? "no" : "yes");
2473         btrfs_info(fs_info,
2474                    "%d blocks of free space at or bigger than bytes is", count);
2475 }
2476
2477 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2478 {
2479         struct btrfs_fs_info *fs_info = block_group->fs_info;
2480         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2481
2482         spin_lock_init(&ctl->tree_lock);
2483         ctl->unit = fs_info->sectorsize;
2484         ctl->start = block_group->key.objectid;
2485         ctl->private = block_group;
2486         ctl->op = &free_space_op;
2487         INIT_LIST_HEAD(&ctl->trimming_ranges);
2488         mutex_init(&ctl->cache_writeout_mutex);
2489
2490         /*
2491          * we only want to have 32k of ram per block group for keeping
2492          * track of free space, and if we pass 1/2 of that we want to
2493          * start converting things over to using bitmaps
2494          */
2495         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2496 }
2497
2498 /*
2499  * for a given cluster, put all of its extents back into the free
2500  * space cache.  If the block group passed doesn't match the block group
2501  * pointed to by the cluster, someone else raced in and freed the
2502  * cluster already.  In that case, we just return without changing anything
2503  */
2504 static int
2505 __btrfs_return_cluster_to_free_space(
2506                              struct btrfs_block_group_cache *block_group,
2507                              struct btrfs_free_cluster *cluster)
2508 {
2509         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2510         struct btrfs_free_space *entry;
2511         struct rb_node *node;
2512
2513         spin_lock(&cluster->lock);
2514         if (cluster->block_group != block_group)
2515                 goto out;
2516
2517         cluster->block_group = NULL;
2518         cluster->window_start = 0;
2519         list_del_init(&cluster->block_group_list);
2520
2521         node = rb_first(&cluster->root);
2522         while (node) {
2523                 bool bitmap;
2524
2525                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2526                 node = rb_next(&entry->offset_index);
2527                 rb_erase(&entry->offset_index, &cluster->root);
2528                 RB_CLEAR_NODE(&entry->offset_index);
2529
2530                 bitmap = (entry->bitmap != NULL);
2531                 if (!bitmap) {
2532                         try_merge_free_space(ctl, entry, false);
2533                         steal_from_bitmap(ctl, entry, false);
2534                 }
2535                 tree_insert_offset(&ctl->free_space_offset,
2536                                    entry->offset, &entry->offset_index, bitmap);
2537         }
2538         cluster->root = RB_ROOT;
2539
2540 out:
2541         spin_unlock(&cluster->lock);
2542         btrfs_put_block_group(block_group);
2543         return 0;
2544 }
2545
2546 static void __btrfs_remove_free_space_cache_locked(
2547                                 struct btrfs_free_space_ctl *ctl)
2548 {
2549         struct btrfs_free_space *info;
2550         struct rb_node *node;
2551
2552         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2553                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2554                 if (!info->bitmap) {
2555                         unlink_free_space(ctl, info);
2556                         kmem_cache_free(btrfs_free_space_cachep, info);
2557                 } else {
2558                         free_bitmap(ctl, info);
2559                 }
2560
2561                 cond_resched_lock(&ctl->tree_lock);
2562         }
2563 }
2564
2565 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2566 {
2567         spin_lock(&ctl->tree_lock);
2568         __btrfs_remove_free_space_cache_locked(ctl);
2569         spin_unlock(&ctl->tree_lock);
2570 }
2571
2572 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2573 {
2574         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2575         struct btrfs_free_cluster *cluster;
2576         struct list_head *head;
2577
2578         spin_lock(&ctl->tree_lock);
2579         while ((head = block_group->cluster_list.next) !=
2580                &block_group->cluster_list) {
2581                 cluster = list_entry(head, struct btrfs_free_cluster,
2582                                      block_group_list);
2583
2584                 WARN_ON(cluster->block_group != block_group);
2585                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2586
2587                 cond_resched_lock(&ctl->tree_lock);
2588         }
2589         __btrfs_remove_free_space_cache_locked(ctl);
2590         spin_unlock(&ctl->tree_lock);
2591
2592 }
2593
2594 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2595                                u64 offset, u64 bytes, u64 empty_size,
2596                                u64 *max_extent_size)
2597 {
2598         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2599         struct btrfs_free_space *entry = NULL;
2600         u64 bytes_search = bytes + empty_size;
2601         u64 ret = 0;
2602         u64 align_gap = 0;
2603         u64 align_gap_len = 0;
2604
2605         spin_lock(&ctl->tree_lock);
2606         entry = find_free_space(ctl, &offset, &bytes_search,
2607                                 block_group->full_stripe_len, max_extent_size);
2608         if (!entry)
2609                 goto out;
2610
2611         ret = offset;
2612         if (entry->bitmap) {
2613                 bitmap_clear_bits(ctl, entry, offset, bytes);
2614                 if (!entry->bytes)
2615                         free_bitmap(ctl, entry);
2616         } else {
2617                 unlink_free_space(ctl, entry);
2618                 align_gap_len = offset - entry->offset;
2619                 align_gap = entry->offset;
2620
2621                 entry->offset = offset + bytes;
2622                 WARN_ON(entry->bytes < bytes + align_gap_len);
2623
2624                 entry->bytes -= bytes + align_gap_len;
2625                 if (!entry->bytes)
2626                         kmem_cache_free(btrfs_free_space_cachep, entry);
2627                 else
2628                         link_free_space(ctl, entry);
2629         }
2630 out:
2631         spin_unlock(&ctl->tree_lock);
2632
2633         if (align_gap_len)
2634                 __btrfs_add_free_space(block_group->fs_info, ctl,
2635                                        align_gap, align_gap_len);
2636         return ret;
2637 }
2638
2639 /*
2640  * given a cluster, put all of its extents back into the free space
2641  * cache.  If a block group is passed, this function will only free
2642  * a cluster that belongs to the passed block group.
2643  *
2644  * Otherwise, it'll get a reference on the block group pointed to by the
2645  * cluster and remove the cluster from it.
2646  */
2647 int btrfs_return_cluster_to_free_space(
2648                                struct btrfs_block_group_cache *block_group,
2649                                struct btrfs_free_cluster *cluster)
2650 {
2651         struct btrfs_free_space_ctl *ctl;
2652         int ret;
2653
2654         /* first, get a safe pointer to the block group */
2655         spin_lock(&cluster->lock);
2656         if (!block_group) {
2657                 block_group = cluster->block_group;
2658                 if (!block_group) {
2659                         spin_unlock(&cluster->lock);
2660                         return 0;
2661                 }
2662         } else if (cluster->block_group != block_group) {
2663                 /* someone else has already freed it don't redo their work */
2664                 spin_unlock(&cluster->lock);
2665                 return 0;
2666         }
2667         atomic_inc(&block_group->count);
2668         spin_unlock(&cluster->lock);
2669
2670         ctl = block_group->free_space_ctl;
2671
2672         /* now return any extents the cluster had on it */
2673         spin_lock(&ctl->tree_lock);
2674         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2675         spin_unlock(&ctl->tree_lock);
2676
2677         /* finally drop our ref */
2678         btrfs_put_block_group(block_group);
2679         return ret;
2680 }
2681
2682 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2683                                    struct btrfs_free_cluster *cluster,
2684                                    struct btrfs_free_space *entry,
2685                                    u64 bytes, u64 min_start,
2686                                    u64 *max_extent_size)
2687 {
2688         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2689         int err;
2690         u64 search_start = cluster->window_start;
2691         u64 search_bytes = bytes;
2692         u64 ret = 0;
2693
2694         search_start = min_start;
2695         search_bytes = bytes;
2696
2697         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2698         if (err) {
2699                 if (search_bytes > *max_extent_size)
2700                         *max_extent_size = search_bytes;
2701                 return 0;
2702         }
2703
2704         ret = search_start;
2705         __bitmap_clear_bits(ctl, entry, ret, bytes);
2706
2707         return ret;
2708 }
2709
2710 /*
2711  * given a cluster, try to allocate 'bytes' from it, returns 0
2712  * if it couldn't find anything suitably large, or a logical disk offset
2713  * if things worked out
2714  */
2715 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2716                              struct btrfs_free_cluster *cluster, u64 bytes,
2717                              u64 min_start, u64 *max_extent_size)
2718 {
2719         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2720         struct btrfs_free_space *entry = NULL;
2721         struct rb_node *node;
2722         u64 ret = 0;
2723
2724         spin_lock(&cluster->lock);
2725         if (bytes > cluster->max_size)
2726                 goto out;
2727
2728         if (cluster->block_group != block_group)
2729                 goto out;
2730
2731         node = rb_first(&cluster->root);
2732         if (!node)
2733                 goto out;
2734
2735         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2736         while (1) {
2737                 if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2738                         *max_extent_size = entry->bytes;
2739
2740                 if (entry->bytes < bytes ||
2741                     (!entry->bitmap && entry->offset < min_start)) {
2742                         node = rb_next(&entry->offset_index);
2743                         if (!node)
2744                                 break;
2745                         entry = rb_entry(node, struct btrfs_free_space,
2746                                          offset_index);
2747                         continue;
2748                 }
2749
2750                 if (entry->bitmap) {
2751                         ret = btrfs_alloc_from_bitmap(block_group,
2752                                                       cluster, entry, bytes,
2753                                                       cluster->window_start,
2754                                                       max_extent_size);
2755                         if (ret == 0) {
2756                                 node = rb_next(&entry->offset_index);
2757                                 if (!node)
2758                                         break;
2759                                 entry = rb_entry(node, struct btrfs_free_space,
2760                                                  offset_index);
2761                                 continue;
2762                         }
2763                         cluster->window_start += bytes;
2764                 } else {
2765                         ret = entry->offset;
2766
2767                         entry->offset += bytes;
2768                         entry->bytes -= bytes;
2769                 }
2770
2771                 if (entry->bytes == 0)
2772                         rb_erase(&entry->offset_index, &cluster->root);
2773                 break;
2774         }
2775 out:
2776         spin_unlock(&cluster->lock);
2777
2778         if (!ret)
2779                 return 0;
2780
2781         spin_lock(&ctl->tree_lock);
2782
2783         ctl->free_space -= bytes;
2784         if (entry->bytes == 0) {
2785                 ctl->free_extents--;
2786                 if (entry->bitmap) {
2787                         kfree(entry->bitmap);
2788                         ctl->total_bitmaps--;
2789                         ctl->op->recalc_thresholds(ctl);
2790                 }
2791                 kmem_cache_free(btrfs_free_space_cachep, entry);
2792         }
2793
2794         spin_unlock(&ctl->tree_lock);
2795
2796         return ret;
2797 }
2798
2799 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2800                                 struct btrfs_free_space *entry,
2801                                 struct btrfs_free_cluster *cluster,
2802                                 u64 offset, u64 bytes,
2803                                 u64 cont1_bytes, u64 min_bytes)
2804 {
2805         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2806         unsigned long next_zero;
2807         unsigned long i;
2808         unsigned long want_bits;
2809         unsigned long min_bits;
2810         unsigned long found_bits;
2811         unsigned long max_bits = 0;
2812         unsigned long start = 0;
2813         unsigned long total_found = 0;
2814         int ret;
2815
2816         i = offset_to_bit(entry->offset, ctl->unit,
2817                           max_t(u64, offset, entry->offset));
2818         want_bits = bytes_to_bits(bytes, ctl->unit);
2819         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2820
2821         /*
2822          * Don't bother looking for a cluster in this bitmap if it's heavily
2823          * fragmented.
2824          */
2825         if (entry->max_extent_size &&
2826             entry->max_extent_size < cont1_bytes)
2827                 return -ENOSPC;
2828 again:
2829         found_bits = 0;
2830         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2831                 next_zero = find_next_zero_bit(entry->bitmap,
2832                                                BITS_PER_BITMAP, i);
2833                 if (next_zero - i >= min_bits) {
2834                         found_bits = next_zero - i;
2835                         if (found_bits > max_bits)
2836                                 max_bits = found_bits;
2837                         break;
2838                 }
2839                 if (next_zero - i > max_bits)
2840                         max_bits = next_zero - i;
2841                 i = next_zero;
2842         }
2843
2844         if (!found_bits) {
2845                 entry->max_extent_size = (u64)max_bits * ctl->unit;
2846                 return -ENOSPC;
2847         }
2848
2849         if (!total_found) {
2850                 start = i;
2851                 cluster->max_size = 0;
2852         }
2853
2854         total_found += found_bits;
2855
2856         if (cluster->max_size < found_bits * ctl->unit)
2857                 cluster->max_size = found_bits * ctl->unit;
2858
2859         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2860                 i = next_zero + 1;
2861                 goto again;
2862         }
2863
2864         cluster->window_start = start * ctl->unit + entry->offset;
2865         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2866         ret = tree_insert_offset(&cluster->root, entry->offset,
2867                                  &entry->offset_index, 1);
2868         ASSERT(!ret); /* -EEXIST; Logic error */
2869
2870         trace_btrfs_setup_cluster(block_group, cluster,
2871                                   total_found * ctl->unit, 1);
2872         return 0;
2873 }
2874
2875 /*
2876  * This searches the block group for just extents to fill the cluster with.
2877  * Try to find a cluster with at least bytes total bytes, at least one
2878  * extent of cont1_bytes, and other clusters of at least min_bytes.
2879  */
2880 static noinline int
2881 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2882                         struct btrfs_free_cluster *cluster,
2883                         struct list_head *bitmaps, u64 offset, u64 bytes,
2884                         u64 cont1_bytes, u64 min_bytes)
2885 {
2886         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2887         struct btrfs_free_space *first = NULL;
2888         struct btrfs_free_space *entry = NULL;
2889         struct btrfs_free_space *last;
2890         struct rb_node *node;
2891         u64 window_free;
2892         u64 max_extent;
2893         u64 total_size = 0;
2894
2895         entry = tree_search_offset(ctl, offset, 0, 1);
2896         if (!entry)
2897                 return -ENOSPC;
2898
2899         /*
2900          * We don't want bitmaps, so just move along until we find a normal
2901          * extent entry.
2902          */
2903         while (entry->bitmap || entry->bytes < min_bytes) {
2904                 if (entry->bitmap && list_empty(&entry->list))
2905                         list_add_tail(&entry->list, bitmaps);
2906                 node = rb_next(&entry->offset_index);
2907                 if (!node)
2908                         return -ENOSPC;
2909                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2910         }
2911
2912         window_free = entry->bytes;
2913         max_extent = entry->bytes;
2914         first = entry;
2915         last = entry;
2916
2917         for (node = rb_next(&entry->offset_index); node;
2918              node = rb_next(&entry->offset_index)) {
2919                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2920
2921                 if (entry->bitmap) {
2922                         if (list_empty(&entry->list))
2923                                 list_add_tail(&entry->list, bitmaps);
2924                         continue;
2925                 }
2926
2927                 if (entry->bytes < min_bytes)
2928                         continue;
2929
2930                 last = entry;
2931                 window_free += entry->bytes;
2932                 if (entry->bytes > max_extent)
2933                         max_extent = entry->bytes;
2934         }
2935
2936         if (window_free < bytes || max_extent < cont1_bytes)
2937                 return -ENOSPC;
2938
2939         cluster->window_start = first->offset;
2940
2941         node = &first->offset_index;
2942
2943         /*
2944          * now we've found our entries, pull them out of the free space
2945          * cache and put them into the cluster rbtree
2946          */
2947         do {
2948                 int ret;
2949
2950                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2951                 node = rb_next(&entry->offset_index);
2952                 if (entry->bitmap || entry->bytes < min_bytes)
2953                         continue;
2954
2955                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2956                 ret = tree_insert_offset(&cluster->root, entry->offset,
2957                                          &entry->offset_index, 0);
2958                 total_size += entry->bytes;
2959                 ASSERT(!ret); /* -EEXIST; Logic error */
2960         } while (node && entry != last);
2961
2962         cluster->max_size = max_extent;
2963         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2964         return 0;
2965 }
2966
2967 /*
2968  * This specifically looks for bitmaps that may work in the cluster, we assume
2969  * that we have already failed to find extents that will work.
2970  */
2971 static noinline int
2972 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2973                      struct btrfs_free_cluster *cluster,
2974                      struct list_head *bitmaps, u64 offset, u64 bytes,
2975                      u64 cont1_bytes, u64 min_bytes)
2976 {
2977         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2978         struct btrfs_free_space *entry = NULL;
2979         int ret = -ENOSPC;
2980         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2981
2982         if (ctl->total_bitmaps == 0)
2983                 return -ENOSPC;
2984
2985         /*
2986          * The bitmap that covers offset won't be in the list unless offset
2987          * is just its start offset.
2988          */
2989         if (!list_empty(bitmaps))
2990                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2991
2992         if (!entry || entry->offset != bitmap_offset) {
2993                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2994                 if (entry && list_empty(&entry->list))
2995                         list_add(&entry->list, bitmaps);
2996         }
2997
2998         list_for_each_entry(entry, bitmaps, list) {
2999                 if (entry->bytes < bytes)
3000                         continue;
3001                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3002                                            bytes, cont1_bytes, min_bytes);
3003                 if (!ret)
3004                         return 0;
3005         }
3006
3007         /*
3008          * The bitmaps list has all the bitmaps that record free space
3009          * starting after offset, so no more search is required.
3010          */
3011         return -ENOSPC;
3012 }
3013
3014 /*
3015  * here we try to find a cluster of blocks in a block group.  The goal
3016  * is to find at least bytes+empty_size.
3017  * We might not find them all in one contiguous area.
3018  *
3019  * returns zero and sets up cluster if things worked out, otherwise
3020  * it returns -enospc
3021  */
3022 int btrfs_find_space_cluster(struct btrfs_fs_info *fs_info,
3023                              struct btrfs_block_group_cache *block_group,
3024                              struct btrfs_free_cluster *cluster,
3025                              u64 offset, u64 bytes, u64 empty_size)
3026 {
3027         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3028         struct btrfs_free_space *entry, *tmp;
3029         LIST_HEAD(bitmaps);
3030         u64 min_bytes;
3031         u64 cont1_bytes;
3032         int ret;
3033
3034         /*
3035          * Choose the minimum extent size we'll require for this
3036          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3037          * For metadata, allow allocates with smaller extents.  For
3038          * data, keep it dense.
3039          */
3040         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3041                 cont1_bytes = min_bytes = bytes + empty_size;
3042         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3043                 cont1_bytes = bytes;
3044                 min_bytes = fs_info->sectorsize;
3045         } else {
3046                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3047                 min_bytes = fs_info->sectorsize;
3048         }
3049
3050         spin_lock(&ctl->tree_lock);
3051
3052         /*
3053          * If we know we don't have enough space to make a cluster don't even
3054          * bother doing all the work to try and find one.
3055          */
3056         if (ctl->free_space < bytes) {
3057                 spin_unlock(&ctl->tree_lock);
3058                 return -ENOSPC;
3059         }
3060
3061         spin_lock(&cluster->lock);
3062
3063         /* someone already found a cluster, hooray */
3064         if (cluster->block_group) {
3065                 ret = 0;
3066                 goto out;
3067         }
3068
3069         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3070                                  min_bytes);
3071
3072         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3073                                       bytes + empty_size,
3074                                       cont1_bytes, min_bytes);
3075         if (ret)
3076                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3077                                            offset, bytes + empty_size,
3078                                            cont1_bytes, min_bytes);
3079
3080         /* Clear our temporary list */
3081         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3082                 list_del_init(&entry->list);
3083
3084         if (!ret) {
3085                 atomic_inc(&block_group->count);
3086                 list_add_tail(&cluster->block_group_list,
3087                               &block_group->cluster_list);
3088                 cluster->block_group = block_group;
3089         } else {
3090                 trace_btrfs_failed_cluster_setup(block_group);
3091         }
3092 out:
3093         spin_unlock(&cluster->lock);
3094         spin_unlock(&ctl->tree_lock);
3095
3096         return ret;
3097 }
3098
3099 /*
3100  * simple code to zero out a cluster
3101  */
3102 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3103 {
3104         spin_lock_init(&cluster->lock);
3105         spin_lock_init(&cluster->refill_lock);
3106         cluster->root = RB_ROOT;
3107         cluster->max_size = 0;
3108         cluster->fragmented = false;
3109         INIT_LIST_HEAD(&cluster->block_group_list);
3110         cluster->block_group = NULL;
3111 }
3112
3113 static int do_trimming(struct btrfs_block_group_cache *block_group,
3114                        u64 *total_trimmed, u64 start, u64 bytes,
3115                        u64 reserved_start, u64 reserved_bytes,
3116                        struct btrfs_trim_range *trim_entry)
3117 {
3118         struct btrfs_space_info *space_info = block_group->space_info;
3119         struct btrfs_fs_info *fs_info = block_group->fs_info;
3120         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3121         int ret;
3122         int update = 0;
3123         u64 trimmed = 0;
3124
3125         spin_lock(&space_info->lock);
3126         spin_lock(&block_group->lock);
3127         if (!block_group->ro) {
3128                 block_group->reserved += reserved_bytes;
3129                 space_info->bytes_reserved += reserved_bytes;
3130                 update = 1;
3131         }
3132         spin_unlock(&block_group->lock);
3133         spin_unlock(&space_info->lock);
3134
3135         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3136         if (!ret)
3137                 *total_trimmed += trimmed;
3138
3139         mutex_lock(&ctl->cache_writeout_mutex);
3140         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3141         list_del(&trim_entry->list);
3142         mutex_unlock(&ctl->cache_writeout_mutex);
3143
3144         if (update) {
3145                 spin_lock(&space_info->lock);
3146                 spin_lock(&block_group->lock);
3147                 if (block_group->ro)
3148                         space_info->bytes_readonly += reserved_bytes;
3149                 block_group->reserved -= reserved_bytes;
3150                 space_info->bytes_reserved -= reserved_bytes;
3151                 spin_unlock(&space_info->lock);
3152                 spin_unlock(&block_group->lock);
3153         }
3154
3155         return ret;
3156 }
3157
3158 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
3159                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3160 {
3161         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3162         struct btrfs_free_space *entry;
3163         struct rb_node *node;
3164         int ret = 0;
3165         u64 extent_start;
3166         u64 extent_bytes;
3167         u64 bytes;
3168
3169         while (start < end) {
3170                 struct btrfs_trim_range trim_entry;
3171
3172                 mutex_lock(&ctl->cache_writeout_mutex);
3173                 spin_lock(&ctl->tree_lock);
3174
3175                 if (ctl->free_space < minlen) {
3176                         spin_unlock(&ctl->tree_lock);
3177                         mutex_unlock(&ctl->cache_writeout_mutex);
3178                         break;
3179                 }
3180
3181                 entry = tree_search_offset(ctl, start, 0, 1);
3182                 if (!entry) {
3183                         spin_unlock(&ctl->tree_lock);
3184                         mutex_unlock(&ctl->cache_writeout_mutex);
3185                         break;
3186                 }
3187
3188                 /* skip bitmaps */
3189                 while (entry->bitmap) {
3190                         node = rb_next(&entry->offset_index);
3191                         if (!node) {
3192                                 spin_unlock(&ctl->tree_lock);
3193                                 mutex_unlock(&ctl->cache_writeout_mutex);
3194                                 goto out;
3195                         }
3196                         entry = rb_entry(node, struct btrfs_free_space,
3197                                          offset_index);
3198                 }
3199
3200                 if (entry->offset >= end) {
3201                         spin_unlock(&ctl->tree_lock);
3202                         mutex_unlock(&ctl->cache_writeout_mutex);
3203                         break;
3204                 }
3205
3206                 extent_start = entry->offset;
3207                 extent_bytes = entry->bytes;
3208                 start = max(start, extent_start);
3209                 bytes = min(extent_start + extent_bytes, end) - start;
3210                 if (bytes < minlen) {
3211                         spin_unlock(&ctl->tree_lock);
3212                         mutex_unlock(&ctl->cache_writeout_mutex);
3213                         goto next;
3214                 }
3215
3216                 unlink_free_space(ctl, entry);
3217                 kmem_cache_free(btrfs_free_space_cachep, entry);
3218
3219                 spin_unlock(&ctl->tree_lock);
3220                 trim_entry.start = extent_start;
3221                 trim_entry.bytes = extent_bytes;
3222                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3223                 mutex_unlock(&ctl->cache_writeout_mutex);
3224
3225                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3226                                   extent_start, extent_bytes, &trim_entry);
3227                 if (ret)
3228                         break;
3229 next:
3230                 start += bytes;
3231
3232                 if (fatal_signal_pending(current)) {
3233                         ret = -ERESTARTSYS;
3234                         break;
3235                 }
3236
3237                 cond_resched();
3238         }
3239 out:
3240         return ret;
3241 }
3242
3243 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
3244                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3245 {
3246         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3247         struct btrfs_free_space *entry;
3248         int ret = 0;
3249         int ret2;
3250         u64 bytes;
3251         u64 offset = offset_to_bitmap(ctl, start);
3252
3253         while (offset < end) {
3254                 bool next_bitmap = false;
3255                 struct btrfs_trim_range trim_entry;
3256
3257                 mutex_lock(&ctl->cache_writeout_mutex);
3258                 spin_lock(&ctl->tree_lock);
3259
3260                 if (ctl->free_space < minlen) {
3261                         spin_unlock(&ctl->tree_lock);
3262                         mutex_unlock(&ctl->cache_writeout_mutex);
3263                         break;
3264                 }
3265
3266                 entry = tree_search_offset(ctl, offset, 1, 0);
3267                 if (!entry) {
3268                         spin_unlock(&ctl->tree_lock);
3269                         mutex_unlock(&ctl->cache_writeout_mutex);
3270                         next_bitmap = true;
3271                         goto next;
3272                 }
3273
3274                 bytes = minlen;
3275                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3276                 if (ret2 || start >= end) {
3277                         spin_unlock(&ctl->tree_lock);
3278                         mutex_unlock(&ctl->cache_writeout_mutex);
3279                         next_bitmap = true;
3280                         goto next;
3281                 }
3282
3283                 bytes = min(bytes, end - start);
3284                 if (bytes < minlen) {
3285                         spin_unlock(&ctl->tree_lock);
3286                         mutex_unlock(&ctl->cache_writeout_mutex);
3287                         goto next;
3288                 }
3289
3290                 bitmap_clear_bits(ctl, entry, start, bytes);
3291                 if (entry->bytes == 0)
3292                         free_bitmap(ctl, entry);
3293
3294                 spin_unlock(&ctl->tree_lock);
3295                 trim_entry.start = start;
3296                 trim_entry.bytes = bytes;
3297                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3298                 mutex_unlock(&ctl->cache_writeout_mutex);
3299
3300                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3301                                   start, bytes, &trim_entry);
3302                 if (ret)
3303                         break;
3304 next:
3305                 if (next_bitmap) {
3306                         offset += BITS_PER_BITMAP * ctl->unit;
3307                 } else {
3308                         start += bytes;
3309                         if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3310                                 offset += BITS_PER_BITMAP * ctl->unit;
3311                 }
3312
3313                 if (fatal_signal_pending(current)) {
3314                         ret = -ERESTARTSYS;
3315                         break;
3316                 }
3317
3318                 cond_resched();
3319         }
3320
3321         return ret;
3322 }
3323
3324 void btrfs_get_block_group_trimming(struct btrfs_block_group_cache *cache)
3325 {
3326         atomic_inc(&cache->trimming);
3327 }
3328
3329 void btrfs_put_block_group_trimming(struct btrfs_block_group_cache *block_group)
3330 {
3331         struct btrfs_fs_info *fs_info = block_group->fs_info;
3332         struct extent_map_tree *em_tree;
3333         struct extent_map *em;
3334         bool cleanup;
3335
3336         spin_lock(&block_group->lock);
3337         cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3338                    block_group->removed);
3339         spin_unlock(&block_group->lock);
3340
3341         if (cleanup) {
3342                 mutex_lock(&fs_info->chunk_mutex);
3343                 em_tree = &fs_info->mapping_tree.map_tree;
3344                 write_lock(&em_tree->lock);
3345                 em = lookup_extent_mapping(em_tree, block_group->key.objectid,
3346                                            1);
3347                 BUG_ON(!em); /* logic error, can't happen */
3348                 /*
3349                  * remove_extent_mapping() will delete us from the pinned_chunks
3350                  * list, which is protected by the chunk mutex.
3351                  */
3352                 remove_extent_mapping(em_tree, em);
3353                 write_unlock(&em_tree->lock);
3354                 mutex_unlock(&fs_info->chunk_mutex);
3355
3356                 /* once for us and once for the tree */
3357                 free_extent_map(em);
3358                 free_extent_map(em);
3359
3360                 /*
3361                  * We've left one free space entry and other tasks trimming
3362                  * this block group have left 1 entry each one. Free them.
3363                  */
3364                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3365         }
3366 }
3367
3368 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
3369                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3370 {
3371         int ret;
3372
3373         *trimmed = 0;
3374
3375         spin_lock(&block_group->lock);
3376         if (block_group->removed) {
3377                 spin_unlock(&block_group->lock);
3378                 return 0;
3379         }
3380         btrfs_get_block_group_trimming(block_group);
3381         spin_unlock(&block_group->lock);
3382
3383         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3384         if (ret)
3385                 goto out;
3386
3387         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3388 out:
3389         btrfs_put_block_group_trimming(block_group);
3390         return ret;
3391 }
3392
3393 /*
3394  * Find the left-most item in the cache tree, and then return the
3395  * smallest inode number in the item.
3396  *
3397  * Note: the returned inode number may not be the smallest one in
3398  * the tree, if the left-most item is a bitmap.
3399  */
3400 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3401 {
3402         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3403         struct btrfs_free_space *entry = NULL;
3404         u64 ino = 0;
3405
3406         spin_lock(&ctl->tree_lock);
3407
3408         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3409                 goto out;
3410
3411         entry = rb_entry(rb_first(&ctl->free_space_offset),
3412                          struct btrfs_free_space, offset_index);
3413
3414         if (!entry->bitmap) {
3415                 ino = entry->offset;
3416
3417                 unlink_free_space(ctl, entry);
3418                 entry->offset++;
3419                 entry->bytes--;
3420                 if (!entry->bytes)
3421                         kmem_cache_free(btrfs_free_space_cachep, entry);
3422                 else
3423                         link_free_space(ctl, entry);
3424         } else {
3425                 u64 offset = 0;
3426                 u64 count = 1;
3427                 int ret;
3428
3429                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3430                 /* Logic error; Should be empty if it can't find anything */
3431                 ASSERT(!ret);
3432
3433                 ino = offset;
3434                 bitmap_clear_bits(ctl, entry, offset, 1);
3435                 if (entry->bytes == 0)
3436                         free_bitmap(ctl, entry);
3437         }
3438 out:
3439         spin_unlock(&ctl->tree_lock);
3440
3441         return ino;
3442 }
3443
3444 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3445                                     struct btrfs_path *path)
3446 {
3447         struct inode *inode = NULL;
3448
3449         spin_lock(&root->ino_cache_lock);
3450         if (root->ino_cache_inode)
3451                 inode = igrab(root->ino_cache_inode);
3452         spin_unlock(&root->ino_cache_lock);
3453         if (inode)
3454                 return inode;
3455
3456         inode = __lookup_free_space_inode(root, path, 0);
3457         if (IS_ERR(inode))
3458                 return inode;
3459
3460         spin_lock(&root->ino_cache_lock);
3461         if (!btrfs_fs_closing(root->fs_info))
3462                 root->ino_cache_inode = igrab(inode);
3463         spin_unlock(&root->ino_cache_lock);
3464
3465         return inode;
3466 }
3467
3468 int create_free_ino_inode(struct btrfs_root *root,
3469                           struct btrfs_trans_handle *trans,
3470                           struct btrfs_path *path)
3471 {
3472         return __create_free_space_inode(root, trans, path,
3473                                          BTRFS_FREE_INO_OBJECTID, 0);
3474 }
3475
3476 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3477 {
3478         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3479         struct btrfs_path *path;
3480         struct inode *inode;
3481         int ret = 0;
3482         u64 root_gen = btrfs_root_generation(&root->root_item);
3483
3484         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3485                 return 0;
3486
3487         /*
3488          * If we're unmounting then just return, since this does a search on the
3489          * normal root and not the commit root and we could deadlock.
3490          */
3491         if (btrfs_fs_closing(fs_info))
3492                 return 0;
3493
3494         path = btrfs_alloc_path();
3495         if (!path)
3496                 return 0;
3497
3498         inode = lookup_free_ino_inode(root, path);
3499         if (IS_ERR(inode))
3500                 goto out;
3501
3502         if (root_gen != BTRFS_I(inode)->generation)
3503                 goto out_put;
3504
3505         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3506
3507         if (ret < 0)
3508                 btrfs_err(fs_info,
3509                         "failed to load free ino cache for root %llu",
3510                         root->root_key.objectid);
3511 out_put:
3512         iput(inode);
3513 out:
3514         btrfs_free_path(path);
3515         return ret;
3516 }
3517
3518 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3519                               struct btrfs_trans_handle *trans,
3520                               struct btrfs_path *path,
3521                               struct inode *inode)
3522 {
3523         struct btrfs_fs_info *fs_info = root->fs_info;
3524         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3525         int ret;
3526         struct btrfs_io_ctl io_ctl;
3527         bool release_metadata = true;
3528
3529         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3530                 return 0;
3531
3532         memset(&io_ctl, 0, sizeof(io_ctl));
3533         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl, trans);
3534         if (!ret) {
3535                 /*
3536                  * At this point writepages() didn't error out, so our metadata
3537                  * reservation is released when the writeback finishes, at
3538                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3539                  * with or without an error.
3540                  */
3541                 release_metadata = false;
3542                 ret = btrfs_wait_cache_io_root(root, trans, &io_ctl, path);
3543         }
3544
3545         if (ret) {
3546                 if (release_metadata)
3547                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3548                                         inode->i_size);
3549 #ifdef DEBUG
3550                 btrfs_err(fs_info,
3551                           "failed to write free ino cache for root %llu",
3552                           root->root_key.objectid);
3553 #endif
3554         }
3555
3556         return ret;
3557 }
3558
3559 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3560 /*
3561  * Use this if you need to make a bitmap or extent entry specifically, it
3562  * doesn't do any of the merging that add_free_space does, this acts a lot like
3563  * how the free space cache loading stuff works, so you can get really weird
3564  * configurations.
3565  */
3566 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3567                               u64 offset, u64 bytes, bool bitmap)
3568 {
3569         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3570         struct btrfs_free_space *info = NULL, *bitmap_info;
3571         void *map = NULL;
3572         u64 bytes_added;
3573         int ret;
3574
3575 again:
3576         if (!info) {
3577                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3578                 if (!info)
3579                         return -ENOMEM;
3580         }
3581
3582         if (!bitmap) {
3583                 spin_lock(&ctl->tree_lock);
3584                 info->offset = offset;
3585                 info->bytes = bytes;
3586                 info->max_extent_size = 0;
3587                 ret = link_free_space(ctl, info);
3588                 spin_unlock(&ctl->tree_lock);
3589                 if (ret)
3590                         kmem_cache_free(btrfs_free_space_cachep, info);
3591                 return ret;
3592         }
3593
3594         if (!map) {
3595                 map = kzalloc(PAGE_SIZE, GFP_NOFS);
3596                 if (!map) {
3597                         kmem_cache_free(btrfs_free_space_cachep, info);
3598                         return -ENOMEM;
3599                 }
3600         }
3601
3602         spin_lock(&ctl->tree_lock);
3603         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3604                                          1, 0);
3605         if (!bitmap_info) {
3606                 info->bitmap = map;
3607                 map = NULL;
3608                 add_new_bitmap(ctl, info, offset);
3609                 bitmap_info = info;
3610                 info = NULL;
3611         }
3612
3613         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3614
3615         bytes -= bytes_added;
3616         offset += bytes_added;
3617         spin_unlock(&ctl->tree_lock);
3618
3619         if (bytes)
3620                 goto again;
3621
3622         if (info)
3623                 kmem_cache_free(btrfs_free_space_cachep, info);
3624         if (map)
3625                 kfree(map);
3626         return 0;
3627 }
3628
3629 /*
3630  * Checks to see if the given range is in the free space cache.  This is really
3631  * just used to check the absence of space, so if there is free space in the
3632  * range at all we will return 1.
3633  */
3634 int test_check_exists(struct btrfs_block_group_cache *cache,
3635                       u64 offset, u64 bytes)
3636 {
3637         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3638         struct btrfs_free_space *info;
3639         int ret = 0;
3640
3641         spin_lock(&ctl->tree_lock);
3642         info = tree_search_offset(ctl, offset, 0, 0);
3643         if (!info) {
3644                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3645                                           1, 0);
3646                 if (!info)
3647                         goto out;
3648         }
3649
3650 have_info:
3651         if (info->bitmap) {
3652                 u64 bit_off, bit_bytes;
3653                 struct rb_node *n;
3654                 struct btrfs_free_space *tmp;
3655
3656                 bit_off = offset;
3657                 bit_bytes = ctl->unit;
3658                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3659                 if (!ret) {
3660                         if (bit_off == offset) {
3661                                 ret = 1;
3662                                 goto out;
3663                         } else if (bit_off > offset &&
3664                                    offset + bytes > bit_off) {
3665                                 ret = 1;
3666                                 goto out;
3667                         }
3668                 }
3669
3670                 n = rb_prev(&info->offset_index);
3671                 while (n) {
3672                         tmp = rb_entry(n, struct btrfs_free_space,
3673                                        offset_index);
3674                         if (tmp->offset + tmp->bytes < offset)
3675                                 break;
3676                         if (offset + bytes < tmp->offset) {
3677                                 n = rb_prev(&tmp->offset_index);
3678                                 continue;
3679                         }
3680                         info = tmp;
3681                         goto have_info;
3682                 }
3683
3684                 n = rb_next(&info->offset_index);
3685                 while (n) {
3686                         tmp = rb_entry(n, struct btrfs_free_space,
3687                                        offset_index);
3688                         if (offset + bytes < tmp->offset)
3689                                 break;
3690                         if (tmp->offset + tmp->bytes < offset) {
3691                                 n = rb_next(&tmp->offset_index);
3692                                 continue;
3693                         }
3694                         info = tmp;
3695                         goto have_info;
3696                 }
3697
3698                 ret = 0;
3699                 goto out;
3700         }
3701
3702         if (info->offset == offset) {
3703                 ret = 1;
3704                 goto out;
3705         }
3706
3707         if (offset > info->offset && offset < info->offset + info->bytes)
3708                 ret = 1;
3709 out:
3710         spin_unlock(&ctl->tree_lock);
3711         return ret;
3712 }
3713 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */