fs/btrfs/free-space-cache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2008 Red Hat.  All rights reserved.
   4  */
   5
   6 #include <linux/pagemap.h>
   7 #include <linux/sched.h>
   8 #include <linux/sched/signal.h>
   9 #include <linux/slab.h>
  10 #include <linux/math64.h>
  11 #include <linux/ratelimit.h>
  12 #include <linux/error-injection.h>
  13 #include <linux/sched/mm.h>
  14 #include "ctree.h"
  15 #include "free-space-cache.h"
  16 #include "transaction.h"
  17 #include "disk-io.h"
  18 #include "extent_io.h"
  19 #include "inode-map.h"
  20 #include "volumes.h"
  21 #include "space-info.h"
  22 #include "delalloc-space.h"
  23 #include "block-group.h"
  24
  25 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  26 #define MAX_CACHE_BYTES_PER_GIG SZ_32K
  27
  28 struct btrfs_trim_range {
  29         u64 start;
  30         u64 bytes;
  31         struct list_head list;
  32 };
  33
  34 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  35                            struct btrfs_free_space *info);
  36 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  37                               struct btrfs_free_space *info);
  38 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
  39                              struct btrfs_trans_handle *trans,
  40                              struct btrfs_io_ctl *io_ctl,
  41                              struct btrfs_path *path);
  42
  43 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  44                                                struct btrfs_path *path,
  45                                                u64 offset)
  46 {
  47         struct btrfs_fs_info *fs_info = root->fs_info;
  48         struct btrfs_key key;
  49         struct btrfs_key location;
  50         struct btrfs_disk_key disk_key;
  51         struct btrfs_free_space_header *header;
  52         struct extent_buffer *leaf;
  53         struct inode *inode = NULL;
  54         unsigned nofs_flag;
  55         int ret;
  56
  57         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  58         key.offset = offset;
  59         key.type = 0;
  60
  61         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  62         if (ret < 0)
  63                 return ERR_PTR(ret);
  64         if (ret > 0) {
  65                 btrfs_release_path(path);
  66                 return ERR_PTR(-ENOENT);
  67         }
  68
  69         leaf = path->nodes[0];
  70         header = btrfs_item_ptr(leaf, path->slots[0],
  71                                 struct btrfs_free_space_header);
  72         btrfs_free_space_key(leaf, header, &disk_key);
  73         btrfs_disk_key_to_cpu(&location, &disk_key);
  74         btrfs_release_path(path);
  75
  76         /*
  77          * We are often under a trans handle at this point, so we need to make
  78          * sure NOFS is set to keep us from deadlocking.
  79          */
  80         nofs_flag = memalloc_nofs_save();
  81         inode = btrfs_iget_path(fs_info->sb, &location, root, path);
  82         btrfs_release_path(path);
  83         memalloc_nofs_restore(nofs_flag);
  84         if (IS_ERR(inode))
  85                 return inode;
  86
  87         mapping_set_gfp_mask(inode->i_mapping,
  88                         mapping_gfp_constraint(inode->i_mapping,
  89                         ~(__GFP_FS | __GFP_HIGHMEM)));
  90
  91         return inode;
  92 }
  93
  94 struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
  95                 struct btrfs_path *path)
  96 {
  97         struct btrfs_fs_info *fs_info = block_group->fs_info;
  98         struct inode *inode = NULL;
  99         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 100
 101         spin_lock(&block_group->lock);
 102         if (block_group->inode)
 103                 inode = igrab(block_group->inode);
 104         spin_unlock(&block_group->lock);
 105         if (inode)
 106                 return inode;
 107
 108         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 109                                           block_group->start);
 110         if (IS_ERR(inode))
 111                 return inode;
 112
 113         spin_lock(&block_group->lock);
 114         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 115                 btrfs_info(fs_info, "Old style space inode found, converting.");
 116                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 117                         BTRFS_INODE_NODATACOW;
 118                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 119         }
 120
 121         if (!block_group->iref) {
 122                 block_group->inode = igrab(inode);
 123                 block_group->iref = 1;
 124         }
 125         spin_unlock(&block_group->lock);
 126
 127         return inode;
 128 }
 129
 130 static int __create_free_space_inode(struct btrfs_root *root,
 131                                      struct btrfs_trans_handle *trans,
 132                                      struct btrfs_path *path,
 133                                      u64 ino, u64 offset)
 134 {
 135         struct btrfs_key key;
 136         struct btrfs_disk_key disk_key;
 137         struct btrfs_free_space_header *header;
 138         struct btrfs_inode_item *inode_item;
 139         struct extent_buffer *leaf;
 140         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 141         int ret;
 142
 143         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 144         if (ret)
 145                 return ret;
 146
 147         /* We inline crc's for the free disk space cache */
 148         if (ino != BTRFS_FREE_INO_OBJECTID)
 149                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 150
 151         leaf = path->nodes[0];
 152         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 153                                     struct btrfs_inode_item);
 154         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 155         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 156                              sizeof(*inode_item));
 157         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 158         btrfs_set_inode_size(leaf, inode_item, 0);
 159         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 160         btrfs_set_inode_uid(leaf, inode_item, 0);
 161         btrfs_set_inode_gid(leaf, inode_item, 0);
 162         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 163         btrfs_set_inode_flags(leaf, inode_item, flags);
 164         btrfs_set_inode_nlink(leaf, inode_item, 1);
 165         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 166         btrfs_set_inode_block_group(leaf, inode_item, offset);
 167         btrfs_mark_buffer_dirty(leaf);
 168         btrfs_release_path(path);
 169
 170         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 171         key.offset = offset;
 172         key.type = 0;
 173         ret = btrfs_insert_empty_item(trans, root, path, &key,
 174                                       sizeof(struct btrfs_free_space_header));
 175         if (ret < 0) {
 176                 btrfs_release_path(path);
 177                 return ret;
 178         }
 179
 180         leaf = path->nodes[0];
 181         header = btrfs_item_ptr(leaf, path->slots[0],
 182                                 struct btrfs_free_space_header);
 183         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 184         btrfs_set_free_space_key(leaf, header, &disk_key);
 185         btrfs_mark_buffer_dirty(leaf);
 186         btrfs_release_path(path);
 187
 188         return 0;
 189 }
 190
 191 int create_free_space_inode(struct btrfs_trans_handle *trans,
 192                             struct btrfs_block_group *block_group,
 193                             struct btrfs_path *path)
 194 {
 195         int ret;
 196         u64 ino;
 197
 198         ret = btrfs_find_free_objectid(trans->fs_info->tree_root, &ino);
 199         if (ret < 0)
 200                 return ret;
 201
 202         return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
 203                                          ino, block_group->start);
 204 }
 205
 206 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
 207                                        struct btrfs_block_rsv *rsv)
 208 {
 209         u64 needed_bytes;
 210         int ret;
 211
 212         /* 1 for slack space, 1 for updating the inode */
 213         needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
 214                 btrfs_calc_metadata_size(fs_info, 1);
 215
 216         spin_lock(&rsv->lock);
 217         if (rsv->reserved < needed_bytes)
 218                 ret = -ENOSPC;
 219         else
 220                 ret = 0;
 221         spin_unlock(&rsv->lock);
 222         return ret;
 223 }
 224
 225 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 226                                     struct btrfs_block_group *block_group,
 227                                     struct inode *inode)
 228 {
 229         struct btrfs_root *root = BTRFS_I(inode)->root;
 230         int ret = 0;
 231         bool locked = false;
 232
 233         if (block_group) {
 234                 struct btrfs_path *path = btrfs_alloc_path();
 235
 236                 if (!path) {
 237                         ret = -ENOMEM;
 238                         goto fail;
 239                 }
 240                 locked = true;
 241                 mutex_lock(&trans->transaction->cache_write_mutex);
 242                 if (!list_empty(&block_group->io_list)) {
 243                         list_del_init(&block_group->io_list);
 244
 245                         btrfs_wait_cache_io(trans, block_group, path);
 246                         btrfs_put_block_group(block_group);
 247                 }
 248
 249                 /*
 250                  * now that we've truncated the cache away, its no longer
 251                  * setup or written
 252                  */
 253                 spin_lock(&block_group->lock);
 254                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 255                 spin_unlock(&block_group->lock);
 256                 btrfs_free_path(path);
 257         }
 258
 259         btrfs_i_size_write(BTRFS_I(inode), 0);
 260         truncate_pagecache(inode, 0);
 261
 262         /*
 263          * We skip the throttling logic for free space cache inodes, so we don't
 264          * need to check for -EAGAIN.
 265          */
 266         ret = btrfs_truncate_inode_items(trans, root, inode,
 267                                          0, BTRFS_EXTENT_DATA_KEY);
 268         if (ret)
 269                 goto fail;
 270
 271         ret = btrfs_update_inode(trans, root, inode);
 272
 273 fail:
 274         if (locked)
 275                 mutex_unlock(&trans->transaction->cache_write_mutex);
 276         if (ret)
 277                 btrfs_abort_transaction(trans, ret);
 278
 279         return ret;
 280 }
 281
 282 static void readahead_cache(struct inode *inode)
 283 {
 284         struct file_ra_state *ra;
 285         unsigned long last_index;
 286
 287         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 288         if (!ra)
 289                 return;
 290
 291         file_ra_state_init(ra, inode->i_mapping);
 292         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 293
 294         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 295
 296         kfree(ra);
 297 }
 298
 299 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 300                        int write)
 301 {
 302         int num_pages;
 303         int check_crcs = 0;
 304
 305         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 306
 307         if (btrfs_ino(BTRFS_I(inode)) != BTRFS_FREE_INO_OBJECTID)
 308                 check_crcs = 1;
 309
 310         /* Make sure we can fit our crcs and generation into the first page */
 311         if (write && check_crcs &&
 312             (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
 313                 return -ENOSPC;
 314
 315         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 316
 317         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 318         if (!io_ctl->pages)
 319                 return -ENOMEM;
 320
 321         io_ctl->num_pages = num_pages;
 322         io_ctl->fs_info = btrfs_sb(inode->i_sb);
 323         io_ctl->check_crcs = check_crcs;
 324         io_ctl->inode = inode;
 325
 326         return 0;
 327 }
 328 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 329
 330 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 331 {
 332         kfree(io_ctl->pages);
 333         io_ctl->pages = NULL;
 334 }
 335
 336 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 337 {
 338         if (io_ctl->cur) {
 339                 io_ctl->cur = NULL;
 340                 io_ctl->orig = NULL;
 341         }
 342 }
 343
 344 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 345 {
 346         ASSERT(io_ctl->index < io_ctl->num_pages);
 347         io_ctl->page = io_ctl->pages[io_ctl->index++];
 348         io_ctl->cur = page_address(io_ctl->page);
 349         io_ctl->orig = io_ctl->cur;
 350         io_ctl->size = PAGE_SIZE;
 351         if (clear)
 352                 clear_page(io_ctl->cur);
 353 }
 354
 355 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 356 {
 357         int i;
 358
 359         io_ctl_unmap_page(io_ctl);
 360
 361         for (i = 0; i < io_ctl->num_pages; i++) {
 362                 if (io_ctl->pages[i]) {
 363                         ClearPageChecked(io_ctl->pages[i]);
 364                         unlock_page(io_ctl->pages[i]);
 365                         put_page(io_ctl->pages[i]);
 366                 }
 367         }
 368 }
 369
 370 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 371                                 int uptodate)
 372 {
 373         struct page *page;
 374         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 375         int i;
 376
 377         for (i = 0; i < io_ctl->num_pages; i++) {
 378                 page = find_or_create_page(inode->i_mapping, i, mask);
 379                 if (!page) {
 380                         io_ctl_drop_pages(io_ctl);
 381                         return -ENOMEM;
 382                 }
 383                 io_ctl->pages[i] = page;
 384                 if (uptodate && !PageUptodate(page)) {
 385                         btrfs_readpage(NULL, page);
 386                         lock_page(page);
 387                         if (page->mapping != inode->i_mapping) {
 388                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 389                                           "free space cache page truncated");
 390                                 io_ctl_drop_pages(io_ctl);
 391                                 return -EIO;
 392                         }
 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_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 477         btrfs_crc32c_final(crc, (u8 *)&crc);
 478         io_ctl_unmap_page(io_ctl);
 479         tmp = page_address(io_ctl->pages[0]);
 480         tmp += index;
 481         *tmp = crc;
 482 }
 483
 484 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 485 {
 486         u32 *tmp, val;
 487         u32 crc = ~(u32)0;
 488         unsigned offset = 0;
 489
 490         if (!io_ctl->check_crcs) {
 491                 io_ctl_map_page(io_ctl, 0);
 492                 return 0;
 493         }
 494
 495         if (index == 0)
 496                 offset = sizeof(u32) * io_ctl->num_pages;
 497
 498         tmp = page_address(io_ctl->pages[0]);
 499         tmp += index;
 500         val = *tmp;
 501
 502         io_ctl_map_page(io_ctl, 0);
 503         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 504         btrfs_crc32c_final(crc, (u8 *)&crc);
 505         if (val != crc) {
 506                 btrfs_err_rl(io_ctl->fs_info,
 507                         "csum mismatch on free space cache");
 508                 io_ctl_unmap_page(io_ctl);
 509                 return -EIO;
 510         }
 511
 512         return 0;
 513 }
 514
 515 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 516                             void *bitmap)
 517 {
 518         struct btrfs_free_space_entry *entry;
 519
 520         if (!io_ctl->cur)
 521                 return -ENOSPC;
 522
 523         entry = io_ctl->cur;
 524         entry->offset = cpu_to_le64(offset);
 525         entry->bytes = cpu_to_le64(bytes);
 526         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 527                 BTRFS_FREE_SPACE_EXTENT;
 528         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 529         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 530
 531         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 532                 return 0;
 533
 534         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 535
 536         /* No more pages to map */
 537         if (io_ctl->index >= io_ctl->num_pages)
 538                 return 0;
 539
 540         /* map the next page */
 541         io_ctl_map_page(io_ctl, 1);
 542         return 0;
 543 }
 544
 545 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 546 {
 547         if (!io_ctl->cur)
 548                 return -ENOSPC;
 549
 550         /*
 551          * If we aren't at the start of the current page, unmap this one and
 552          * map the next one if there is any left.
 553          */
 554         if (io_ctl->cur != io_ctl->orig) {
 555                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 556                 if (io_ctl->index >= io_ctl->num_pages)
 557                         return -ENOSPC;
 558                 io_ctl_map_page(io_ctl, 0);
 559         }
 560
 561         copy_page(io_ctl->cur, bitmap);
 562         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 563         if (io_ctl->index < io_ctl->num_pages)
 564                 io_ctl_map_page(io_ctl, 0);
 565         return 0;
 566 }
 567
 568 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 569 {
 570         /*
 571          * If we're not on the boundary we know we've modified the page and we
 572          * need to crc the page.
 573          */
 574         if (io_ctl->cur != io_ctl->orig)
 575                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 576         else
 577                 io_ctl_unmap_page(io_ctl);
 578
 579         while (io_ctl->index < io_ctl->num_pages) {
 580                 io_ctl_map_page(io_ctl, 1);
 581                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 582         }
 583 }
 584
 585 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 586                             struct btrfs_free_space *entry, u8 *type)
 587 {
 588         struct btrfs_free_space_entry *e;
 589         int ret;
 590
 591         if (!io_ctl->cur) {
 592                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 593                 if (ret)
 594                         return ret;
 595         }
 596
 597         e = io_ctl->cur;
 598         entry->offset = le64_to_cpu(e->offset);
 599         entry->bytes = le64_to_cpu(e->bytes);
 600         *type = e->type;
 601         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 602         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 603
 604         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 605                 return 0;
 606
 607         io_ctl_unmap_page(io_ctl);
 608
 609         return 0;
 610 }
 611
 612 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 613                               struct btrfs_free_space *entry)
 614 {
 615         int ret;
 616
 617         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 618         if (ret)
 619                 return ret;
 620
 621         copy_page(entry->bitmap, io_ctl->cur);
 622         io_ctl_unmap_page(io_ctl);
 623
 624         return 0;
 625 }
 626
 627 /*
 628  * Since we attach pinned extents after the fact we can have contiguous sections
 629  * of free space that are split up in entries.  This poses a problem with the
 630  * tree logging stuff since it could have allocated across what appears to be 2
 631  * entries since we would have merged the entries when adding the pinned extents
 632  * back to the free space cache.  So run through the space cache that we just
 633  * loaded and merge contiguous entries.  This will make the log replay stuff not
 634  * blow up and it will make for nicer allocator behavior.
 635  */
 636 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 637 {
 638         struct btrfs_free_space *e, *prev = NULL;
 639         struct rb_node *n;
 640
 641 again:
 642         spin_lock(&ctl->tree_lock);
 643         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 644                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 645                 if (!prev)
 646                         goto next;
 647                 if (e->bitmap || prev->bitmap)
 648                         goto next;
 649                 if (prev->offset + prev->bytes == e->offset) {
 650                         unlink_free_space(ctl, prev);
 651                         unlink_free_space(ctl, e);
 652                         prev->bytes += e->bytes;
 653                         kmem_cache_free(btrfs_free_space_cachep, e);
 654                         link_free_space(ctl, prev);
 655                         prev = NULL;
 656                         spin_unlock(&ctl->tree_lock);
 657                         goto again;
 658                 }
 659 next:
 660                 prev = e;
 661         }
 662         spin_unlock(&ctl->tree_lock);
 663 }
 664
 665 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 666                                    struct btrfs_free_space_ctl *ctl,
 667                                    struct btrfs_path *path, u64 offset)
 668 {
 669         struct btrfs_fs_info *fs_info = root->fs_info;
 670         struct btrfs_free_space_header *header;
 671         struct extent_buffer *leaf;
 672         struct btrfs_io_ctl io_ctl;
 673         struct btrfs_key key;
 674         struct btrfs_free_space *e, *n;
 675         LIST_HEAD(bitmaps);
 676         u64 num_entries;
 677         u64 num_bitmaps;
 678         u64 generation;
 679         u8 type;
 680         int ret = 0;
 681
 682         /* Nothing in the space cache, goodbye */
 683         if (!i_size_read(inode))
 684                 return 0;
 685
 686         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 687         key.offset = offset;
 688         key.type = 0;
 689
 690         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 691         if (ret < 0)
 692                 return 0;
 693         else if (ret > 0) {
 694                 btrfs_release_path(path);
 695                 return 0;
 696         }
 697
 698         ret = -1;
 699
 700         leaf = path->nodes[0];
 701         header = btrfs_item_ptr(leaf, path->slots[0],
 702                                 struct btrfs_free_space_header);
 703         num_entries = btrfs_free_space_entries(leaf, header);
 704         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 705         generation = btrfs_free_space_generation(leaf, header);
 706         btrfs_release_path(path);
 707
 708         if (!BTRFS_I(inode)->generation) {
 709                 btrfs_info(fs_info,
 710                            "the free space cache file (%llu) is invalid, skip it",
 711                            offset);
 712                 return 0;
 713         }
 714
 715         if (BTRFS_I(inode)->generation != generation) {
 716                 btrfs_err(fs_info,
 717                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 718                           BTRFS_I(inode)->generation, generation);
 719                 return 0;
 720         }
 721
 722         if (!num_entries)
 723                 return 0;
 724
 725         ret = io_ctl_init(&io_ctl, inode, 0);
 726         if (ret)
 727                 return ret;
 728
 729         readahead_cache(inode);
 730
 731         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 732         if (ret)
 733                 goto out;
 734
 735         ret = io_ctl_check_crc(&io_ctl, 0);
 736         if (ret)
 737                 goto free_cache;
 738
 739         ret = io_ctl_check_generation(&io_ctl, generation);
 740         if (ret)
 741                 goto free_cache;
 742
 743         while (num_entries) {
 744                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 745                                       GFP_NOFS);
 746                 if (!e)
 747                         goto free_cache;
 748
 749                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 750                 if (ret) {
 751                         kmem_cache_free(btrfs_free_space_cachep, e);
 752                         goto free_cache;
 753                 }
 754
 755                 if (!e->bytes) {
 756                         kmem_cache_free(btrfs_free_space_cachep, e);
 757                         goto free_cache;
 758                 }
 759
 760                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 761                         spin_lock(&ctl->tree_lock);
 762                         ret = link_free_space(ctl, e);
 763                         spin_unlock(&ctl->tree_lock);
 764                         if (ret) {
 765                                 btrfs_err(fs_info,
 766                                         "Duplicate entries in free space cache, dumping");
 767                                 kmem_cache_free(btrfs_free_space_cachep, e);
 768                                 goto free_cache;
 769                         }
 770                 } else {
 771                         ASSERT(num_bitmaps);
 772                         num_bitmaps--;
 773                         e->bitmap = kmem_cache_zalloc(
 774                                         btrfs_free_space_bitmap_cachep, GFP_NOFS);
 775                         if (!e->bitmap) {
 776                                 kmem_cache_free(
 777                                         btrfs_free_space_cachep, e);
 778                                 goto free_cache;
 779                         }
 780                         spin_lock(&ctl->tree_lock);
 781                         ret = link_free_space(ctl, e);
 782                         ctl->total_bitmaps++;
 783                         ctl->op->recalc_thresholds(ctl);
 784                         spin_unlock(&ctl->tree_lock);
 785                         if (ret) {
 786                                 btrfs_err(fs_info,
 787                                         "Duplicate entries in free space cache, dumping");
 788                                 kmem_cache_free(btrfs_free_space_cachep, e);
 789                                 goto free_cache;
 790                         }
 791                         list_add_tail(&e->list, &bitmaps);
 792                 }
 793
 794                 num_entries--;
 795         }
 796
 797         io_ctl_unmap_page(&io_ctl);
 798
 799         /*
 800          * We add the bitmaps at the end of the entries in order that
 801          * the bitmap entries are added to the cache.
 802          */
 803         list_for_each_entry_safe(e, n, &bitmaps, list) {
 804                 list_del_init(&e->list);
 805                 ret = io_ctl_read_bitmap(&io_ctl, e);
 806                 if (ret)
 807                         goto free_cache;
 808         }
 809
 810         io_ctl_drop_pages(&io_ctl);
 811         merge_space_tree(ctl);
 812         ret = 1;
 813 out:
 814         io_ctl_free(&io_ctl);
 815         return ret;
 816 free_cache:
 817         io_ctl_drop_pages(&io_ctl);
 818         __btrfs_remove_free_space_cache(ctl);
 819         goto out;
 820 }
 821
 822 int load_free_space_cache(struct btrfs_block_group *block_group)
 823 {
 824         struct btrfs_fs_info *fs_info = block_group->fs_info;
 825         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 826         struct inode *inode;
 827         struct btrfs_path *path;
 828         int ret = 0;
 829         bool matched;
 830         u64 used = block_group->used;
 831
 832         /*
 833          * If this block group has been marked to be cleared for one reason or
 834          * another then we can't trust the on disk cache, so just return.
 835          */
 836         spin_lock(&block_group->lock);
 837         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 838                 spin_unlock(&block_group->lock);
 839                 return 0;
 840         }
 841         spin_unlock(&block_group->lock);
 842
 843         path = btrfs_alloc_path();
 844         if (!path)
 845                 return 0;
 846         path->search_commit_root = 1;
 847         path->skip_locking = 1;
 848
 849         /*
 850          * We must pass a path with search_commit_root set to btrfs_iget in
 851          * order to avoid a deadlock when allocating extents for the tree root.
 852          *
 853          * When we are COWing an extent buffer from the tree root, when looking
 854          * for a free extent, at extent-tree.c:find_free_extent(), we can find
 855          * block group without its free space cache loaded. When we find one
 856          * we must load its space cache which requires reading its free space
 857          * cache's inode item from the root tree. If this inode item is located
 858          * in the same leaf that we started COWing before, then we end up in
 859          * deadlock on the extent buffer (trying to read lock it when we
 860          * previously write locked it).
 861          *
 862          * It's safe to read the inode item using the commit root because
 863          * block groups, once loaded, stay in memory forever (until they are
 864          * removed) as well as their space caches once loaded. New block groups
 865          * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
 866          * we will never try to read their inode item while the fs is mounted.
 867          */
 868         inode = lookup_free_space_inode(block_group, path);
 869         if (IS_ERR(inode)) {
 870                 btrfs_free_path(path);
 871                 return 0;
 872         }
 873
 874         /* We may have converted the inode and made the cache invalid. */
 875         spin_lock(&block_group->lock);
 876         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 877                 spin_unlock(&block_group->lock);
 878                 btrfs_free_path(path);
 879                 goto out;
 880         }
 881         spin_unlock(&block_group->lock);
 882
 883         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 884                                       path, block_group->start);
 885         btrfs_free_path(path);
 886         if (ret <= 0)
 887                 goto out;
 888
 889         spin_lock(&ctl->tree_lock);
 890         matched = (ctl->free_space == (block_group->length - used -
 891                                        block_group->bytes_super));
 892         spin_unlock(&ctl->tree_lock);
 893
 894         if (!matched) {
 895                 __btrfs_remove_free_space_cache(ctl);
 896                 btrfs_warn(fs_info,
 897                            "block group %llu has wrong amount of free space",
 898                            block_group->start);
 899                 ret = -1;
 900         }
 901 out:
 902         if (ret < 0) {
 903                 /* This cache is bogus, make sure it gets cleared */
 904                 spin_lock(&block_group->lock);
 905                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 906                 spin_unlock(&block_group->lock);
 907                 ret = 0;
 908
 909                 btrfs_warn(fs_info,
 910                            "failed to load free space cache for block group %llu, rebuilding it now",
 911                            block_group->start);
 912         }
 913
 914         iput(inode);
 915         return ret;
 916 }
 917
 918 static noinline_for_stack
 919 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
 920                               struct btrfs_free_space_ctl *ctl,
 921                               struct btrfs_block_group *block_group,
 922                               int *entries, int *bitmaps,
 923                               struct list_head *bitmap_list)
 924 {
 925         int ret;
 926         struct btrfs_free_cluster *cluster = NULL;
 927         struct btrfs_free_cluster *cluster_locked = NULL;
 928         struct rb_node *node = rb_first(&ctl->free_space_offset);
 929         struct btrfs_trim_range *trim_entry;
 930
 931         /* Get the cluster for this block_group if it exists */
 932         if (block_group && !list_empty(&block_group->cluster_list)) {
 933                 cluster = list_entry(block_group->cluster_list.next,
 934                                      struct btrfs_free_cluster,
 935                                      block_group_list);
 936         }
 937
 938         if (!node && cluster) {
 939                 cluster_locked = cluster;
 940                 spin_lock(&cluster_locked->lock);
 941                 node = rb_first(&cluster->root);
 942                 cluster = NULL;
 943         }
 944
 945         /* Write out the extent entries */
 946         while (node) {
 947                 struct btrfs_free_space *e;
 948
 949                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 950                 *entries += 1;
 951
 952                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
 953                                        e->bitmap);
 954                 if (ret)
 955                         goto fail;
 956
 957                 if (e->bitmap) {
 958                         list_add_tail(&e->list, bitmap_list);
 959                         *bitmaps += 1;
 960                 }
 961                 node = rb_next(node);
 962                 if (!node && cluster) {
 963                         node = rb_first(&cluster->root);
 964                         cluster_locked = cluster;
 965                         spin_lock(&cluster_locked->lock);
 966                         cluster = NULL;
 967                 }
 968         }
 969         if (cluster_locked) {
 970                 spin_unlock(&cluster_locked->lock);
 971                 cluster_locked = NULL;
 972         }
 973
 974         /*
 975          * Make sure we don't miss any range that was removed from our rbtree
 976          * because trimming is running. Otherwise after a umount+mount (or crash
 977          * after committing the transaction) we would leak free space and get
 978          * an inconsistent free space cache report from fsck.
 979          */
 980         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
 981                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
 982                                        trim_entry->bytes, NULL);
 983                 if (ret)
 984                         goto fail;
 985                 *entries += 1;
 986         }
 987
 988         return 0;
 989 fail:
 990         if (cluster_locked)
 991                 spin_unlock(&cluster_locked->lock);
 992         return -ENOSPC;
 993 }
 994
 995 static noinline_for_stack int
 996 update_cache_item(struct btrfs_trans_handle *trans,
 997                   struct btrfs_root *root,
 998                   struct inode *inode,
 999                   struct btrfs_path *path, u64 offset,
1000                   int entries, int bitmaps)
1001 {
1002         struct btrfs_key key;
1003         struct btrfs_free_space_header *header;
1004         struct extent_buffer *leaf;
1005         int ret;
1006
1007         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1008         key.offset = offset;
1009         key.type = 0;
1010
1011         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1012         if (ret < 0) {
1013                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1014                                  EXTENT_DELALLOC, 0, 0, NULL);
1015                 goto fail;
1016         }
1017         leaf = path->nodes[0];
1018         if (ret > 0) {
1019                 struct btrfs_key found_key;
1020                 ASSERT(path->slots[0]);
1021                 path->slots[0]--;
1022                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1023                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1024                     found_key.offset != offset) {
1025                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1026                                          inode->i_size - 1, EXTENT_DELALLOC, 0,
1027                                          0, NULL);
1028                         btrfs_release_path(path);
1029                         goto fail;
1030                 }
1031         }
1032
1033         BTRFS_I(inode)->generation = trans->transid;
1034         header = btrfs_item_ptr(leaf, path->slots[0],
1035                                 struct btrfs_free_space_header);
1036         btrfs_set_free_space_entries(leaf, header, entries);
1037         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1038         btrfs_set_free_space_generation(leaf, header, trans->transid);
1039         btrfs_mark_buffer_dirty(leaf);
1040         btrfs_release_path(path);
1041
1042         return 0;
1043
1044 fail:
1045         return -1;
1046 }
1047
1048 static noinline_for_stack int write_pinned_extent_entries(
1049                             struct btrfs_block_group *block_group,
1050                             struct btrfs_io_ctl *io_ctl,
1051                             int *entries)
1052 {
1053         u64 start, extent_start, extent_end, len;
1054         struct extent_io_tree *unpin = NULL;
1055         int ret;
1056
1057         if (!block_group)
1058                 return 0;
1059
1060         /*
1061          * We want to add any pinned extents to our free space cache
1062          * so we don't leak the space
1063          *
1064          * We shouldn't have switched the pinned extents yet so this is the
1065          * right one
1066          */
1067         unpin = block_group->fs_info->pinned_extents;
1068
1069         start = block_group->start;
1070
1071         while (start < block_group->start + block_group->length) {
1072                 ret = find_first_extent_bit(unpin, start,
1073                                             &extent_start, &extent_end,
1074                                             EXTENT_DIRTY, NULL);
1075                 if (ret)
1076                         return 0;
1077
1078                 /* This pinned extent is out of our range */
1079                 if (extent_start >= block_group->start + block_group->length)
1080                         return 0;
1081
1082                 extent_start = max(extent_start, start);
1083                 extent_end = min(block_group->start + block_group->length,
1084                                  extent_end + 1);
1085                 len = extent_end - extent_start;
1086
1087                 *entries += 1;
1088                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1089                 if (ret)
1090                         return -ENOSPC;
1091
1092                 start = extent_end;
1093         }
1094
1095         return 0;
1096 }
1097
1098 static noinline_for_stack int
1099 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1100 {
1101         struct btrfs_free_space *entry, *next;
1102         int ret;
1103
1104         /* Write out the bitmaps */
1105         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1106                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1107                 if (ret)
1108                         return -ENOSPC;
1109                 list_del_init(&entry->list);
1110         }
1111
1112         return 0;
1113 }
1114
1115 static int flush_dirty_cache(struct inode *inode)
1116 {
1117         int ret;
1118
1119         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1120         if (ret)
1121                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1122                                  EXTENT_DELALLOC, 0, 0, NULL);
1123
1124         return ret;
1125 }
1126
1127 static void noinline_for_stack
1128 cleanup_bitmap_list(struct list_head *bitmap_list)
1129 {
1130         struct btrfs_free_space *entry, *next;
1131
1132         list_for_each_entry_safe(entry, next, bitmap_list, list)
1133                 list_del_init(&entry->list);
1134 }
1135
1136 static void noinline_for_stack
1137 cleanup_write_cache_enospc(struct inode *inode,
1138                            struct btrfs_io_ctl *io_ctl,
1139                            struct extent_state **cached_state)
1140 {
1141         io_ctl_drop_pages(io_ctl);
1142         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1143                              i_size_read(inode) - 1, cached_state);
1144 }
1145
1146 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1147                                  struct btrfs_trans_handle *trans,
1148                                  struct btrfs_block_group *block_group,
1149                                  struct btrfs_io_ctl *io_ctl,
1150                                  struct btrfs_path *path, u64 offset)
1151 {
1152         int ret;
1153         struct inode *inode = io_ctl->inode;
1154
1155         if (!inode)
1156                 return 0;
1157
1158         /* Flush the dirty pages in the cache file. */
1159         ret = flush_dirty_cache(inode);
1160         if (ret)
1161                 goto out;
1162
1163         /* Update the cache item to tell everyone this cache file is valid. */
1164         ret = update_cache_item(trans, root, inode, path, offset,
1165                                 io_ctl->entries, io_ctl->bitmaps);
1166 out:
1167         io_ctl_free(io_ctl);
1168         if (ret) {
1169                 invalidate_inode_pages2(inode->i_mapping);
1170                 BTRFS_I(inode)->generation = 0;
1171                 if (block_group) {
1172 #ifdef DEBUG
1173                         btrfs_err(root->fs_info,
1174                                   "failed to write free space cache for block group %llu",
1175                                   block_group->start);
1176 #endif
1177                 }
1178         }
1179         btrfs_update_inode(trans, root, inode);
1180
1181         if (block_group) {
1182                 /* the dirty list is protected by the dirty_bgs_lock */
1183                 spin_lock(&trans->transaction->dirty_bgs_lock);
1184
1185                 /* the disk_cache_state is protected by the block group lock */
1186                 spin_lock(&block_group->lock);
1187
1188                 /*
1189                  * only mark this as written if we didn't get put back on
1190                  * the dirty list while waiting for IO.   Otherwise our
1191                  * cache state won't be right, and we won't get written again
1192                  */
1193                 if (!ret && list_empty(&block_group->dirty_list))
1194                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1195                 else if (ret)
1196                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1197
1198                 spin_unlock(&block_group->lock);
1199                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1200                 io_ctl->inode = NULL;
1201                 iput(inode);
1202         }
1203
1204         return ret;
1205
1206 }
1207
1208 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
1209                                     struct btrfs_trans_handle *trans,
1210                                     struct btrfs_io_ctl *io_ctl,
1211                                     struct btrfs_path *path)
1212 {
1213         return __btrfs_wait_cache_io(root, trans, NULL, io_ctl, path, 0);
1214 }
1215
1216 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1217                         struct btrfs_block_group *block_group,
1218                         struct btrfs_path *path)
1219 {
1220         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1221                                      block_group, &block_group->io_ctl,
1222                                      path, block_group->start);
1223 }
1224
1225 /**
1226  * __btrfs_write_out_cache - write out cached info to an inode
1227  * @root - the root the inode belongs to
1228  * @ctl - the free space cache we are going to write out
1229  * @block_group - the block_group for this cache if it belongs to a block_group
1230  * @trans - the trans handle
1231  *
1232  * This function writes out a free space cache struct to disk for quick recovery
1233  * on mount.  This will return 0 if it was successful in writing the cache out,
1234  * or an errno if it was not.
1235  */
1236 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1237                                    struct btrfs_free_space_ctl *ctl,
1238                                    struct btrfs_block_group *block_group,
1239                                    struct btrfs_io_ctl *io_ctl,
1240                                    struct btrfs_trans_handle *trans)
1241 {
1242         struct extent_state *cached_state = NULL;
1243         LIST_HEAD(bitmap_list);
1244         int entries = 0;
1245         int bitmaps = 0;
1246         int ret;
1247         int must_iput = 0;
1248
1249         if (!i_size_read(inode))
1250                 return -EIO;
1251
1252         WARN_ON(io_ctl->pages);
1253         ret = io_ctl_init(io_ctl, inode, 1);
1254         if (ret)
1255                 return ret;
1256
1257         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1258                 down_write(&block_group->data_rwsem);
1259                 spin_lock(&block_group->lock);
1260                 if (block_group->delalloc_bytes) {
1261                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1262                         spin_unlock(&block_group->lock);
1263                         up_write(&block_group->data_rwsem);
1264                         BTRFS_I(inode)->generation = 0;
1265                         ret = 0;
1266                         must_iput = 1;
1267                         goto out;
1268                 }
1269                 spin_unlock(&block_group->lock);
1270         }
1271
1272         /* Lock all pages first so we can lock the extent safely. */
1273         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1274         if (ret)
1275                 goto out_unlock;
1276
1277         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1278                          &cached_state);
1279
1280         io_ctl_set_generation(io_ctl, trans->transid);
1281
1282         mutex_lock(&ctl->cache_writeout_mutex);
1283         /* Write out the extent entries in the free space cache */
1284         spin_lock(&ctl->tree_lock);
1285         ret = write_cache_extent_entries(io_ctl, ctl,
1286                                          block_group, &entries, &bitmaps,
1287                                          &bitmap_list);
1288         if (ret)
1289                 goto out_nospc_locked;
1290
1291         /*
1292          * Some spaces that are freed in the current transaction are pinned,
1293          * they will be added into free space cache after the transaction is
1294          * committed, we shouldn't lose them.
1295          *
1296          * If this changes while we are working we'll get added back to
1297          * the dirty list and redo it.  No locking needed
1298          */
1299         ret = write_pinned_extent_entries(block_group, io_ctl, &entries);
1300         if (ret)
1301                 goto out_nospc_locked;
1302
1303         /*
1304          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1305          * locked while doing it because a concurrent trim can be manipulating
1306          * or freeing the bitmap.
1307          */
1308         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1309         spin_unlock(&ctl->tree_lock);
1310         mutex_unlock(&ctl->cache_writeout_mutex);
1311         if (ret)
1312                 goto out_nospc;
1313
1314         /* Zero out the rest of the pages just to make sure */
1315         io_ctl_zero_remaining_pages(io_ctl);
1316
1317         /* Everything is written out, now we dirty the pages in the file. */
1318         ret = btrfs_dirty_pages(inode, io_ctl->pages, io_ctl->num_pages, 0,
1319                                 i_size_read(inode), &cached_state);
1320         if (ret)
1321                 goto out_nospc;
1322
1323         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1324                 up_write(&block_group->data_rwsem);
1325         /*
1326          * Release the pages and unlock the extent, we will flush
1327          * them out later
1328          */
1329         io_ctl_drop_pages(io_ctl);
1330
1331         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1332                              i_size_read(inode) - 1, &cached_state);
1333
1334         /*
1335          * at this point the pages are under IO and we're happy,
1336          * The caller is responsible for waiting on them and updating the
1337          * the cache and the inode
1338          */
1339         io_ctl->entries = entries;
1340         io_ctl->bitmaps = bitmaps;
1341
1342         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1343         if (ret)
1344                 goto out;
1345
1346         return 0;
1347
1348 out:
1349         io_ctl->inode = NULL;
1350         io_ctl_free(io_ctl);
1351         if (ret) {
1352                 invalidate_inode_pages2(inode->i_mapping);
1353                 BTRFS_I(inode)->generation = 0;
1354         }
1355         btrfs_update_inode(trans, root, inode);
1356         if (must_iput)
1357                 iput(inode);
1358         return ret;
1359
1360 out_nospc_locked:
1361         cleanup_bitmap_list(&bitmap_list);
1362         spin_unlock(&ctl->tree_lock);
1363         mutex_unlock(&ctl->cache_writeout_mutex);
1364
1365 out_nospc:
1366         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1367
1368 out_unlock:
1369         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1370                 up_write(&block_group->data_rwsem);
1371
1372         goto out;
1373 }
1374
1375 int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1376                           struct btrfs_block_group *block_group,
1377                           struct btrfs_path *path)
1378 {
1379         struct btrfs_fs_info *fs_info = trans->fs_info;
1380         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1381         struct inode *inode;
1382         int ret = 0;
1383
1384         spin_lock(&block_group->lock);
1385         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1386                 spin_unlock(&block_group->lock);
1387                 return 0;
1388         }
1389         spin_unlock(&block_group->lock);
1390
1391         inode = lookup_free_space_inode(block_group, path);
1392         if (IS_ERR(inode))
1393                 return 0;
1394
1395         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1396                                 block_group, &block_group->io_ctl, trans);
1397         if (ret) {
1398 #ifdef DEBUG
1399                 btrfs_err(fs_info,
1400                           "failed to write free space cache for block group %llu",
1401                           block_group->start);
1402 #endif
1403                 spin_lock(&block_group->lock);
1404                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1405                 spin_unlock(&block_group->lock);
1406
1407                 block_group->io_ctl.inode = NULL;
1408                 iput(inode);
1409         }
1410
1411         /*
1412          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1413          * to wait for IO and put the inode
1414          */
1415
1416         return ret;
1417 }
1418
1419 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1420                                           u64 offset)
1421 {
1422         ASSERT(offset >= bitmap_start);
1423         offset -= bitmap_start;
1424         return (unsigned long)(div_u64(offset, unit));
1425 }
1426
1427 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1428 {
1429         return (unsigned long)(div_u64(bytes, unit));
1430 }
1431
1432 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1433                                    u64 offset)
1434 {
1435         u64 bitmap_start;
1436         u64 bytes_per_bitmap;
1437
1438         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1439         bitmap_start = offset - ctl->start;
1440         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1441         bitmap_start *= bytes_per_bitmap;
1442         bitmap_start += ctl->start;
1443
1444         return bitmap_start;
1445 }
1446
1447 static int tree_insert_offset(struct rb_root *root, u64 offset,
1448                               struct rb_node *node, int bitmap)
1449 {
1450         struct rb_node **p = &root->rb_node;
1451         struct rb_node *parent = NULL;
1452         struct btrfs_free_space *info;
1453
1454         while (*p) {
1455                 parent = *p;
1456                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1457
1458                 if (offset < info->offset) {
1459                         p = &(*p)->rb_left;
1460                 } else if (offset > info->offset) {
1461                         p = &(*p)->rb_right;
1462                 } else {
1463                         /*
1464                          * we could have a bitmap entry and an extent entry
1465                          * share the same offset.  If this is the case, we want
1466                          * the extent entry to always be found first if we do a
1467                          * linear search through the tree, since we want to have
1468                          * the quickest allocation time, and allocating from an
1469                          * extent is faster than allocating from a bitmap.  So
1470                          * if we're inserting a bitmap and we find an entry at
1471                          * this offset, we want to go right, or after this entry
1472                          * logically.  If we are inserting an extent and we've
1473                          * found a bitmap, we want to go left, or before
1474                          * logically.
1475                          */
1476                         if (bitmap) {
1477                                 if (info->bitmap) {
1478                                         WARN_ON_ONCE(1);
1479                                         return -EEXIST;
1480                                 }
1481                                 p = &(*p)->rb_right;
1482                         } else {
1483                                 if (!info->bitmap) {
1484                                         WARN_ON_ONCE(1);
1485                                         return -EEXIST;
1486                                 }
1487                                 p = &(*p)->rb_left;
1488                         }
1489                 }
1490         }
1491
1492         rb_link_node(node, parent, p);
1493         rb_insert_color(node, root);
1494
1495         return 0;
1496 }
1497
1498 /*
1499  * searches the tree for the given offset.
1500  *
1501  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1502  * want a section that has at least bytes size and comes at or after the given
1503  * offset.
1504  */
1505 static struct btrfs_free_space *
1506 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1507                    u64 offset, int bitmap_only, int fuzzy)
1508 {
1509         struct rb_node *n = ctl->free_space_offset.rb_node;
1510         struct btrfs_free_space *entry, *prev = NULL;
1511
1512         /* find entry that is closest to the 'offset' */
1513         while (1) {
1514                 if (!n) {
1515                         entry = NULL;
1516                         break;
1517                 }
1518
1519                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1520                 prev = entry;
1521
1522                 if (offset < entry->offset)
1523                         n = n->rb_left;
1524                 else if (offset > entry->offset)
1525                         n = n->rb_right;
1526                 else
1527                         break;
1528         }
1529
1530         if (bitmap_only) {
1531                 if (!entry)
1532                         return NULL;
1533                 if (entry->bitmap)
1534                         return entry;
1535
1536                 /*
1537                  * bitmap entry and extent entry may share same offset,
1538                  * in that case, bitmap entry comes after extent entry.
1539                  */
1540                 n = rb_next(n);
1541                 if (!n)
1542                         return NULL;
1543                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1544                 if (entry->offset != offset)
1545                         return NULL;
1546
1547                 WARN_ON(!entry->bitmap);
1548                 return entry;
1549         } else if (entry) {
1550                 if (entry->bitmap) {
1551                         /*
1552                          * if previous extent entry covers the offset,
1553                          * we should return it instead of the bitmap entry
1554                          */
1555                         n = rb_prev(&entry->offset_index);
1556                         if (n) {
1557                                 prev = rb_entry(n, struct btrfs_free_space,
1558                                                 offset_index);
1559                                 if (!prev->bitmap &&
1560                                     prev->offset + prev->bytes > offset)
1561                                         entry = prev;
1562                         }
1563                 }
1564                 return entry;
1565         }
1566
1567         if (!prev)
1568                 return NULL;
1569
1570         /* find last entry before the 'offset' */
1571         entry = prev;
1572         if (entry->offset > offset) {
1573                 n = rb_prev(&entry->offset_index);
1574                 if (n) {
1575                         entry = rb_entry(n, struct btrfs_free_space,
1576                                         offset_index);
1577                         ASSERT(entry->offset <= offset);
1578                 } else {
1579                         if (fuzzy)
1580                                 return entry;
1581                         else
1582                                 return NULL;
1583                 }
1584         }
1585
1586         if (entry->bitmap) {
1587                 n = rb_prev(&entry->offset_index);
1588                 if (n) {
1589                         prev = rb_entry(n, struct btrfs_free_space,
1590                                         offset_index);
1591                         if (!prev->bitmap &&
1592                             prev->offset + prev->bytes > offset)
1593                                 return prev;
1594                 }
1595                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1596                         return entry;
1597         } else if (entry->offset + entry->bytes > offset)
1598                 return entry;
1599
1600         if (!fuzzy)
1601                 return NULL;
1602
1603         while (1) {
1604                 if (entry->bitmap) {
1605                         if (entry->offset + BITS_PER_BITMAP *
1606                             ctl->unit > offset)
1607                                 break;
1608                 } else {
1609                         if (entry->offset + entry->bytes > offset)
1610                                 break;
1611                 }
1612
1613                 n = rb_next(&entry->offset_index);
1614                 if (!n)
1615                         return NULL;
1616                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1617         }
1618         return entry;
1619 }
1620
1621 static inline void
1622 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1623                     struct btrfs_free_space *info)
1624 {
1625         rb_erase(&info->offset_index, &ctl->free_space_offset);
1626         ctl->free_extents--;
1627 }
1628
1629 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1630                               struct btrfs_free_space *info)
1631 {
1632         __unlink_free_space(ctl, info);
1633         ctl->free_space -= info->bytes;
1634 }
1635
1636 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1637                            struct btrfs_free_space *info)
1638 {
1639         int ret = 0;
1640
1641         ASSERT(info->bytes || info->bitmap);
1642         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1643                                  &info->offset_index, (info->bitmap != NULL));
1644         if (ret)
1645                 return ret;
1646
1647         ctl->free_space += info->bytes;
1648         ctl->free_extents++;
1649         return ret;
1650 }
1651
1652 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1653 {
1654         struct btrfs_block_group *block_group = ctl->private;
1655         u64 max_bytes;
1656         u64 bitmap_bytes;
1657         u64 extent_bytes;
1658         u64 size = block_group->length;
1659         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1660         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1661
1662         max_bitmaps = max_t(u64, max_bitmaps, 1);
1663
1664         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1665
1666         /*
1667          * The goal is to keep the total amount of memory used per 1gb of space
1668          * at or below 32k, so we need to adjust how much memory we allow to be
1669          * used by extent based free space tracking
1670          */
1671         if (size < SZ_1G)
1672                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1673         else
1674                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1675
1676         /*
1677          * we want to account for 1 more bitmap than what we have so we can make
1678          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1679          * we add more bitmaps.
1680          */
1681         bitmap_bytes = (ctl->total_bitmaps + 1) * ctl->unit;
1682
1683         if (bitmap_bytes >= max_bytes) {
1684                 ctl->extents_thresh = 0;
1685                 return;
1686         }
1687
1688         /*
1689          * we want the extent entry threshold to always be at most 1/2 the max
1690          * bytes we can have, or whatever is less than that.
1691          */
1692         extent_bytes = max_bytes - bitmap_bytes;
1693         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1694
1695         ctl->extents_thresh =
1696                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1697 }
1698
1699 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1700                                        struct btrfs_free_space *info,
1701                                        u64 offset, u64 bytes)
1702 {
1703         unsigned long start, count;
1704
1705         start = offset_to_bit(info->offset, ctl->unit, offset);
1706         count = bytes_to_bits(bytes, ctl->unit);
1707         ASSERT(start + count <= BITS_PER_BITMAP);
1708
1709         bitmap_clear(info->bitmap, start, count);
1710
1711         info->bytes -= bytes;
1712         if (info->max_extent_size > ctl->unit)
1713                 info->max_extent_size = 0;
1714 }
1715
1716 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1717                               struct btrfs_free_space *info, u64 offset,
1718                               u64 bytes)
1719 {
1720         __bitmap_clear_bits(ctl, info, offset, bytes);
1721         ctl->free_space -= bytes;
1722 }
1723
1724 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1725                             struct btrfs_free_space *info, u64 offset,
1726                             u64 bytes)
1727 {
1728         unsigned long start, count;
1729
1730         start = offset_to_bit(info->offset, ctl->unit, offset);
1731         count = bytes_to_bits(bytes, ctl->unit);
1732         ASSERT(start + count <= BITS_PER_BITMAP);
1733
1734         bitmap_set(info->bitmap, start, count);
1735
1736         info->bytes += bytes;
1737         ctl->free_space += bytes;
1738 }
1739
1740 /*
1741  * If we can not find suitable extent, we will use bytes to record
1742  * the size of the max extent.
1743  */
1744 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1745                          struct btrfs_free_space *bitmap_info, u64 *offset,
1746                          u64 *bytes, bool for_alloc)
1747 {
1748         unsigned long found_bits = 0;
1749         unsigned long max_bits = 0;
1750         unsigned long bits, i;
1751         unsigned long next_zero;
1752         unsigned long extent_bits;
1753
1754         /*
1755          * Skip searching the bitmap if we don't have a contiguous section that
1756          * is large enough for this allocation.
1757          */
1758         if (for_alloc &&
1759             bitmap_info->max_extent_size &&
1760             bitmap_info->max_extent_size < *bytes) {
1761                 *bytes = bitmap_info->max_extent_size;
1762                 return -1;
1763         }
1764
1765         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1766                           max_t(u64, *offset, bitmap_info->offset));
1767         bits = bytes_to_bits(*bytes, ctl->unit);
1768
1769         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1770                 if (for_alloc && bits == 1) {
1771                         found_bits = 1;
1772                         break;
1773                 }
1774                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1775                                                BITS_PER_BITMAP, i);
1776                 extent_bits = next_zero - i;
1777                 if (extent_bits >= bits) {
1778                         found_bits = extent_bits;
1779                         break;
1780                 } else if (extent_bits > max_bits) {
1781                         max_bits = extent_bits;
1782                 }
1783                 i = next_zero;
1784         }
1785
1786         if (found_bits) {
1787                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1788                 *bytes = (u64)(found_bits) * ctl->unit;
1789                 return 0;
1790         }
1791
1792         *bytes = (u64)(max_bits) * ctl->unit;
1793         bitmap_info->max_extent_size = *bytes;
1794         return -1;
1795 }
1796
1797 static inline u64 get_max_extent_size(struct btrfs_free_space *entry)
1798 {
1799         if (entry->bitmap)
1800                 return entry->max_extent_size;
1801         return entry->bytes;
1802 }
1803
1804 /* Cache the size of the max extent in bytes */
1805 static struct btrfs_free_space *
1806 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1807                 unsigned long align, u64 *max_extent_size)
1808 {
1809         struct btrfs_free_space *entry;
1810         struct rb_node *node;
1811         u64 tmp;
1812         u64 align_off;
1813         int ret;
1814
1815         if (!ctl->free_space_offset.rb_node)
1816                 goto out;
1817
1818         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1819         if (!entry)
1820                 goto out;
1821
1822         for (node = &entry->offset_index; node; node = rb_next(node)) {
1823                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1824                 if (entry->bytes < *bytes) {
1825                         *max_extent_size = max(get_max_extent_size(entry),
1826                                                *max_extent_size);
1827                         continue;
1828                 }
1829
1830                 /* make sure the space returned is big enough
1831                  * to match our requested alignment
1832                  */
1833                 if (*bytes >= align) {
1834                         tmp = entry->offset - ctl->start + align - 1;
1835                         tmp = div64_u64(tmp, align);
1836                         tmp = tmp * align + ctl->start;
1837                         align_off = tmp - entry->offset;
1838                 } else {
1839                         align_off = 0;
1840                         tmp = entry->offset;
1841                 }
1842
1843                 if (entry->bytes < *bytes + align_off) {
1844                         *max_extent_size = max(get_max_extent_size(entry),
1845                                                *max_extent_size);
1846                         continue;
1847                 }
1848
1849                 if (entry->bitmap) {
1850                         u64 size = *bytes;
1851
1852                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1853                         if (!ret) {
1854                                 *offset = tmp;
1855                                 *bytes = size;
1856                                 return entry;
1857                         } else {
1858                                 *max_extent_size =
1859                                         max(get_max_extent_size(entry),
1860                                             *max_extent_size);
1861                         }
1862                         continue;
1863                 }
1864
1865                 *offset = tmp;
1866                 *bytes = entry->bytes - align_off;
1867                 return entry;
1868         }
1869 out:
1870         return NULL;
1871 }
1872
1873 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1874                            struct btrfs_free_space *info, u64 offset)
1875 {
1876         info->offset = offset_to_bitmap(ctl, offset);
1877         info->bytes = 0;
1878         INIT_LIST_HEAD(&info->list);
1879         link_free_space(ctl, info);
1880         ctl->total_bitmaps++;
1881
1882         ctl->op->recalc_thresholds(ctl);
1883 }
1884
1885 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1886                         struct btrfs_free_space *bitmap_info)
1887 {
1888         unlink_free_space(ctl, bitmap_info);
1889         kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
1890         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1891         ctl->total_bitmaps--;
1892         ctl->op->recalc_thresholds(ctl);
1893 }
1894
1895 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1896                               struct btrfs_free_space *bitmap_info,
1897                               u64 *offset, u64 *bytes)
1898 {
1899         u64 end;
1900         u64 search_start, search_bytes;
1901         int ret;
1902
1903 again:
1904         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1905
1906         /*
1907          * We need to search for bits in this bitmap.  We could only cover some
1908          * of the extent in this bitmap thanks to how we add space, so we need
1909          * to search for as much as it as we can and clear that amount, and then
1910          * go searching for the next bit.
1911          */
1912         search_start = *offset;
1913         search_bytes = ctl->unit;
1914         search_bytes = min(search_bytes, end - search_start + 1);
1915         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1916                             false);
1917         if (ret < 0 || search_start != *offset)
1918                 return -EINVAL;
1919
1920         /* We may have found more bits than what we need */
1921         search_bytes = min(search_bytes, *bytes);
1922
1923         /* Cannot clear past the end of the bitmap */
1924         search_bytes = min(search_bytes, end - search_start + 1);
1925
1926         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1927         *offset += search_bytes;
1928         *bytes -= search_bytes;
1929
1930         if (*bytes) {
1931                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1932                 if (!bitmap_info->bytes)
1933                         free_bitmap(ctl, bitmap_info);
1934
1935                 /*
1936                  * no entry after this bitmap, but we still have bytes to
1937                  * remove, so something has gone wrong.
1938                  */
1939                 if (!next)
1940                         return -EINVAL;
1941
1942                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1943                                        offset_index);
1944
1945                 /*
1946                  * if the next entry isn't a bitmap we need to return to let the
1947                  * extent stuff do its work.
1948                  */
1949                 if (!bitmap_info->bitmap)
1950                         return -EAGAIN;
1951
1952                 /*
1953                  * Ok the next item is a bitmap, but it may not actually hold
1954                  * the information for the rest of this free space stuff, so
1955                  * look for it, and if we don't find it return so we can try
1956                  * everything over again.
1957                  */
1958                 search_start = *offset;
1959                 search_bytes = ctl->unit;
1960                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1961                                     &search_bytes, false);
1962                 if (ret < 0 || search_start != *offset)
1963                         return -EAGAIN;
1964
1965                 goto again;
1966         } else if (!bitmap_info->bytes)
1967                 free_bitmap(ctl, bitmap_info);
1968
1969         return 0;
1970 }
1971
1972 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1973                                struct btrfs_free_space *info, u64 offset,
1974                                u64 bytes)
1975 {
1976         u64 bytes_to_set = 0;
1977         u64 end;
1978
1979         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1980
1981         bytes_to_set = min(end - offset, bytes);
1982
1983         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1984
1985         /*
1986          * We set some bytes, we have no idea what the max extent size is
1987          * anymore.
1988          */
1989         info->max_extent_size = 0;
1990
1991         return bytes_to_set;
1992
1993 }
1994
1995 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1996                       struct btrfs_free_space *info)
1997 {
1998         struct btrfs_block_group *block_group = ctl->private;
1999         struct btrfs_fs_info *fs_info = block_group->fs_info;
2000         bool forced = false;
2001
2002 #ifdef CONFIG_BTRFS_DEBUG
2003         if (btrfs_should_fragment_free_space(block_group))
2004                 forced = true;
2005 #endif
2006
2007         /*
2008          * If we are below the extents threshold then we can add this as an
2009          * extent, and don't have to deal with the bitmap
2010          */
2011         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2012                 /*
2013                  * If this block group has some small extents we don't want to
2014                  * use up all of our free slots in the cache with them, we want
2015                  * to reserve them to larger extents, however if we have plenty
2016                  * of cache left then go ahead an dadd them, no sense in adding
2017                  * the overhead of a bitmap if we don't have to.
2018                  */
2019                 if (info->bytes <= fs_info->sectorsize * 4) {
2020                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
2021                                 return false;
2022                 } else {
2023                         return false;
2024                 }
2025         }
2026
2027         /*
2028          * The original block groups from mkfs can be really small, like 8
2029          * megabytes, so don't bother with a bitmap for those entries.  However
2030          * some block groups can be smaller than what a bitmap would cover but
2031          * are still large enough that they could overflow the 32k memory limit,
2032          * so allow those block groups to still be allowed to have a bitmap
2033          * entry.
2034          */
2035         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2036                 return false;
2037
2038         return true;
2039 }
2040
2041 static const struct btrfs_free_space_op free_space_op = {
2042         .recalc_thresholds      = recalculate_thresholds,
2043         .use_bitmap             = use_bitmap,
2044 };
2045
2046 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2047                               struct btrfs_free_space *info)
2048 {
2049         struct btrfs_free_space *bitmap_info;
2050         struct btrfs_block_group *block_group = NULL;
2051         int added = 0;
2052         u64 bytes, offset, bytes_added;
2053         int ret;
2054
2055         bytes = info->bytes;
2056         offset = info->offset;
2057
2058         if (!ctl->op->use_bitmap(ctl, info))
2059                 return 0;
2060
2061         if (ctl->op == &free_space_op)
2062                 block_group = ctl->private;
2063 again:
2064         /*
2065          * Since we link bitmaps right into the cluster we need to see if we
2066          * have a cluster here, and if so and it has our bitmap we need to add
2067          * the free space to that bitmap.
2068          */
2069         if (block_group && !list_empty(&block_group->cluster_list)) {
2070                 struct btrfs_free_cluster *cluster;
2071                 struct rb_node *node;
2072                 struct btrfs_free_space *entry;
2073
2074                 cluster = list_entry(block_group->cluster_list.next,
2075                                      struct btrfs_free_cluster,
2076                                      block_group_list);
2077                 spin_lock(&cluster->lock);
2078                 node = rb_first(&cluster->root);
2079                 if (!node) {
2080                         spin_unlock(&cluster->lock);
2081                         goto no_cluster_bitmap;
2082                 }
2083
2084                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2085                 if (!entry->bitmap) {
2086                         spin_unlock(&cluster->lock);
2087                         goto no_cluster_bitmap;
2088                 }
2089
2090                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2091                         bytes_added = add_bytes_to_bitmap(ctl, entry,
2092                                                           offset, bytes);
2093                         bytes -= bytes_added;
2094                         offset += bytes_added;
2095                 }
2096                 spin_unlock(&cluster->lock);
2097                 if (!bytes) {
2098                         ret = 1;
2099                         goto out;
2100                 }
2101         }
2102
2103 no_cluster_bitmap:
2104         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2105                                          1, 0);
2106         if (!bitmap_info) {
2107                 ASSERT(added == 0);
2108                 goto new_bitmap;
2109         }
2110
2111         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
2112         bytes -= bytes_added;
2113         offset += bytes_added;
2114         added = 0;
2115
2116         if (!bytes) {
2117                 ret = 1;
2118                 goto out;
2119         } else
2120                 goto again;
2121
2122 new_bitmap:
2123         if (info && info->bitmap) {
2124                 add_new_bitmap(ctl, info, offset);
2125                 added = 1;
2126                 info = NULL;
2127                 goto again;
2128         } else {
2129                 spin_unlock(&ctl->tree_lock);
2130
2131                 /* no pre-allocated info, allocate a new one */
2132                 if (!info) {
2133                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2134                                                  GFP_NOFS);
2135                         if (!info) {
2136                                 spin_lock(&ctl->tree_lock);
2137                                 ret = -ENOMEM;
2138                                 goto out;
2139                         }
2140                 }
2141
2142                 /* allocate the bitmap */
2143                 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2144                                                  GFP_NOFS);
2145                 spin_lock(&ctl->tree_lock);
2146                 if (!info->bitmap) {
2147                         ret = -ENOMEM;
2148                         goto out;
2149                 }
2150                 goto again;
2151         }
2152
2153 out:
2154         if (info) {
2155                 if (info->bitmap)
2156                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2157                                         info->bitmap);
2158                 kmem_cache_free(btrfs_free_space_cachep, info);
2159         }
2160
2161         return ret;
2162 }
2163
2164 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2165                           struct btrfs_free_space *info, bool update_stat)
2166 {
2167         struct btrfs_free_space *left_info;
2168         struct btrfs_free_space *right_info;
2169         bool merged = false;
2170         u64 offset = info->offset;
2171         u64 bytes = info->bytes;
2172
2173         /*
2174          * first we want to see if there is free space adjacent to the range we
2175          * are adding, if there is remove that struct and add a new one to
2176          * cover the entire range
2177          */
2178         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2179         if (right_info && rb_prev(&right_info->offset_index))
2180                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2181                                      struct btrfs_free_space, offset_index);
2182         else
2183                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2184
2185         if (right_info && !right_info->bitmap) {
2186                 if (update_stat)
2187                         unlink_free_space(ctl, right_info);
2188                 else
2189                         __unlink_free_space(ctl, right_info);
2190                 info->bytes += right_info->bytes;
2191                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2192                 merged = true;
2193         }
2194
2195         if (left_info && !left_info->bitmap &&
2196             left_info->offset + left_info->bytes == offset) {
2197                 if (update_stat)
2198                         unlink_free_space(ctl, left_info);
2199                 else
2200                         __unlink_free_space(ctl, left_info);
2201                 info->offset = left_info->offset;
2202                 info->bytes += left_info->bytes;
2203                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2204                 merged = true;
2205         }
2206
2207         return merged;
2208 }
2209
2210 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2211                                      struct btrfs_free_space *info,
2212                                      bool update_stat)
2213 {
2214         struct btrfs_free_space *bitmap;
2215         unsigned long i;
2216         unsigned long j;
2217         const u64 end = info->offset + info->bytes;
2218         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2219         u64 bytes;
2220
2221         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2222         if (!bitmap)
2223                 return false;
2224
2225         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2226         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2227         if (j == i)
2228                 return false;
2229         bytes = (j - i) * ctl->unit;
2230         info->bytes += bytes;
2231
2232         if (update_stat)
2233                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2234         else
2235                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2236
2237         if (!bitmap->bytes)
2238                 free_bitmap(ctl, bitmap);
2239
2240         return true;
2241 }
2242
2243 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2244                                        struct btrfs_free_space *info,
2245                                        bool update_stat)
2246 {
2247         struct btrfs_free_space *bitmap;
2248         u64 bitmap_offset;
2249         unsigned long i;
2250         unsigned long j;
2251         unsigned long prev_j;
2252         u64 bytes;
2253
2254         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2255         /* If we're on a boundary, try the previous logical bitmap. */
2256         if (bitmap_offset == info->offset) {
2257                 if (info->offset == 0)
2258                         return false;
2259                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2260         }
2261
2262         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2263         if (!bitmap)
2264                 return false;
2265
2266         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2267         j = 0;
2268         prev_j = (unsigned long)-1;
2269         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2270                 if (j > i)
2271                         break;
2272                 prev_j = j;
2273         }
2274         if (prev_j == i)
2275                 return false;
2276
2277         if (prev_j == (unsigned long)-1)
2278                 bytes = (i + 1) * ctl->unit;
2279         else
2280                 bytes = (i - prev_j) * ctl->unit;
2281
2282         info->offset -= bytes;
2283         info->bytes += bytes;
2284
2285         if (update_stat)
2286                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2287         else
2288                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2289
2290         if (!bitmap->bytes)
2291                 free_bitmap(ctl, bitmap);
2292
2293         return true;
2294 }
2295
2296 /*
2297  * We prefer always to allocate from extent entries, both for clustered and
2298  * non-clustered allocation requests. So when attempting to add a new extent
2299  * entry, try to see if there's adjacent free space in bitmap entries, and if
2300  * there is, migrate that space from the bitmaps to the extent.
2301  * Like this we get better chances of satisfying space allocation requests
2302  * because we attempt to satisfy them based on a single cache entry, and never
2303  * on 2 or more entries - even if the entries represent a contiguous free space
2304  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2305  * ends).
2306  */
2307 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2308                               struct btrfs_free_space *info,
2309                               bool update_stat)
2310 {
2311         /*
2312          * Only work with disconnected entries, as we can change their offset,
2313          * and must be extent entries.
2314          */
2315         ASSERT(!info->bitmap);
2316         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2317
2318         if (ctl->total_bitmaps > 0) {
2319                 bool stole_end;
2320                 bool stole_front = false;
2321
2322                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2323                 if (ctl->total_bitmaps > 0)
2324                         stole_front = steal_from_bitmap_to_front(ctl, info,
2325                                                                  update_stat);
2326
2327                 if (stole_end || stole_front)
2328                         try_merge_free_space(ctl, info, update_stat);
2329         }
2330 }
2331
2332 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2333                            struct btrfs_free_space_ctl *ctl,
2334                            u64 offset, u64 bytes)
2335 {
2336         struct btrfs_free_space *info;
2337         int ret = 0;
2338
2339         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2340         if (!info)
2341                 return -ENOMEM;
2342
2343         info->offset = offset;
2344         info->bytes = bytes;
2345         RB_CLEAR_NODE(&info->offset_index);
2346
2347         spin_lock(&ctl->tree_lock);
2348
2349         if (try_merge_free_space(ctl, info, true))
2350                 goto link;
2351
2352         /*
2353          * There was no extent directly to the left or right of this new
2354          * extent then we know we're going to have to allocate a new extent, so
2355          * before we do that see if we need to drop this into a bitmap
2356          */
2357         ret = insert_into_bitmap(ctl, info);
2358         if (ret < 0) {
2359                 goto out;
2360         } else if (ret) {
2361                 ret = 0;
2362                 goto out;
2363         }
2364 link:
2365         /*
2366          * Only steal free space from adjacent bitmaps if we're sure we're not
2367          * going to add the new free space to existing bitmap entries - because
2368          * that would mean unnecessary work that would be reverted. Therefore
2369          * attempt to steal space from bitmaps if we're adding an extent entry.
2370          */
2371         steal_from_bitmap(ctl, info, true);
2372
2373         ret = link_free_space(ctl, info);
2374         if (ret)
2375                 kmem_cache_free(btrfs_free_space_cachep, info);
2376 out:
2377         spin_unlock(&ctl->tree_lock);
2378
2379         if (ret) {
2380                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2381                 ASSERT(ret != -EEXIST);
2382         }
2383
2384         return ret;
2385 }
2386
2387 int btrfs_add_free_space(struct btrfs_block_group *block_group,
2388                          u64 bytenr, u64 size)
2389 {
2390         return __btrfs_add_free_space(block_group->fs_info,
2391                                       block_group->free_space_ctl,
2392                                       bytenr, size);
2393 }
2394
2395 int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2396                             u64 offset, u64 bytes)
2397 {
2398         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2399         struct btrfs_free_space *info;
2400         int ret;
2401         bool re_search = false;
2402
2403         spin_lock(&ctl->tree_lock);
2404
2405 again:
2406         ret = 0;
2407         if (!bytes)
2408                 goto out_lock;
2409
2410         info = tree_search_offset(ctl, offset, 0, 0);
2411         if (!info) {
2412                 /*
2413                  * oops didn't find an extent that matched the space we wanted
2414                  * to remove, look for a bitmap instead
2415                  */
2416                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2417                                           1, 0);
2418                 if (!info) {
2419                         /*
2420                          * If we found a partial bit of our free space in a
2421                          * bitmap but then couldn't find the other part this may
2422                          * be a problem, so WARN about it.
2423                          */
2424                         WARN_ON(re_search);
2425                         goto out_lock;
2426                 }
2427         }
2428
2429         re_search = false;
2430         if (!info->bitmap) {
2431                 unlink_free_space(ctl, info);
2432                 if (offset == info->offset) {
2433                         u64 to_free = min(bytes, info->bytes);
2434
2435                         info->bytes -= to_free;
2436                         info->offset += to_free;
2437                         if (info->bytes) {
2438                                 ret = link_free_space(ctl, info);
2439                                 WARN_ON(ret);
2440                         } else {
2441                                 kmem_cache_free(btrfs_free_space_cachep, info);
2442                         }
2443
2444                         offset += to_free;
2445                         bytes -= to_free;
2446                         goto again;
2447                 } else {
2448                         u64 old_end = info->bytes + info->offset;
2449
2450                         info->bytes = offset - info->offset;
2451                         ret = link_free_space(ctl, info);
2452                         WARN_ON(ret);
2453                         if (ret)
2454                                 goto out_lock;
2455
2456                         /* Not enough bytes in this entry to satisfy us */
2457                         if (old_end < offset + bytes) {
2458                                 bytes -= old_end - offset;
2459                                 offset = old_end;
2460                                 goto again;
2461                         } else if (old_end == offset + bytes) {
2462                                 /* all done */
2463                                 goto out_lock;
2464                         }
2465                         spin_unlock(&ctl->tree_lock);
2466
2467                         ret = btrfs_add_free_space(block_group, offset + bytes,
2468                                                    old_end - (offset + bytes));
2469                         WARN_ON(ret);
2470                         goto out;
2471                 }
2472         }
2473
2474         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2475         if (ret == -EAGAIN) {
2476                 re_search = true;
2477                 goto again;
2478         }
2479 out_lock:
2480         spin_unlock(&ctl->tree_lock);
2481 out:
2482         return ret;
2483 }
2484
2485 void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2486                            u64 bytes)
2487 {
2488         struct btrfs_fs_info *fs_info = block_group->fs_info;
2489         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2490         struct btrfs_free_space *info;
2491         struct rb_node *n;
2492         int count = 0;
2493
2494         spin_lock(&ctl->tree_lock);
2495         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2496                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2497                 if (info->bytes >= bytes && !block_group->ro)
2498                         count++;
2499                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2500                            info->offset, info->bytes,
2501                        (info->bitmap) ? "yes" : "no");
2502         }
2503         spin_unlock(&ctl->tree_lock);
2504         btrfs_info(fs_info, "block group has cluster?: %s",
2505                list_empty(&block_group->cluster_list) ? "no" : "yes");
2506         btrfs_info(fs_info,
2507                    "%d blocks of free space at or bigger than bytes is", count);
2508 }
2509
2510 void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group)
2511 {
2512         struct btrfs_fs_info *fs_info = block_group->fs_info;
2513         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2514
2515         spin_lock_init(&ctl->tree_lock);
2516         ctl->unit = fs_info->sectorsize;
2517         ctl->start = block_group->start;
2518         ctl->private = block_group;
2519         ctl->op = &free_space_op;
2520         INIT_LIST_HEAD(&ctl->trimming_ranges);
2521         mutex_init(&ctl->cache_writeout_mutex);
2522
2523         /*
2524          * we only want to have 32k of ram per block group for keeping
2525          * track of free space, and if we pass 1/2 of that we want to
2526          * start converting things over to using bitmaps
2527          */
2528         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2529 }
2530
2531 /*
2532  * for a given cluster, put all of its extents back into the free
2533  * space cache.  If the block group passed doesn't match the block group
2534  * pointed to by the cluster, someone else raced in and freed the
2535  * cluster already.  In that case, we just return without changing anything
2536  */
2537 static int
2538 __btrfs_return_cluster_to_free_space(
2539                              struct btrfs_block_group *block_group,
2540                              struct btrfs_free_cluster *cluster)
2541 {
2542         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2543         struct btrfs_free_space *entry;
2544         struct rb_node *node;
2545
2546         spin_lock(&cluster->lock);
2547         if (cluster->block_group != block_group)
2548                 goto out;
2549
2550         cluster->block_group = NULL;
2551         cluster->window_start = 0;
2552         list_del_init(&cluster->block_group_list);
2553
2554         node = rb_first(&cluster->root);
2555         while (node) {
2556                 bool bitmap;
2557
2558                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2559                 node = rb_next(&entry->offset_index);
2560                 rb_erase(&entry->offset_index, &cluster->root);
2561                 RB_CLEAR_NODE(&entry->offset_index);
2562
2563                 bitmap = (entry->bitmap != NULL);
2564                 if (!bitmap) {
2565                         try_merge_free_space(ctl, entry, false);
2566                         steal_from_bitmap(ctl, entry, false);
2567                 }
2568                 tree_insert_offset(&ctl->free_space_offset,
2569                                    entry->offset, &entry->offset_index, bitmap);
2570         }
2571         cluster->root = RB_ROOT;
2572
2573 out:
2574         spin_unlock(&cluster->lock);
2575         btrfs_put_block_group(block_group);
2576         return 0;
2577 }
2578
2579 static void __btrfs_remove_free_space_cache_locked(
2580                                 struct btrfs_free_space_ctl *ctl)
2581 {
2582         struct btrfs_free_space *info;
2583         struct rb_node *node;
2584
2585         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2586                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2587                 if (!info->bitmap) {
2588                         unlink_free_space(ctl, info);
2589                         kmem_cache_free(btrfs_free_space_cachep, info);
2590                 } else {
2591                         free_bitmap(ctl, info);
2592                 }
2593
2594                 cond_resched_lock(&ctl->tree_lock);
2595         }
2596 }
2597
2598 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2599 {
2600         spin_lock(&ctl->tree_lock);
2601         __btrfs_remove_free_space_cache_locked(ctl);
2602         spin_unlock(&ctl->tree_lock);
2603 }
2604
2605 void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
2606 {
2607         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2608         struct btrfs_free_cluster *cluster;
2609         struct list_head *head;
2610
2611         spin_lock(&ctl->tree_lock);
2612         while ((head = block_group->cluster_list.next) !=
2613                &block_group->cluster_list) {
2614                 cluster = list_entry(head, struct btrfs_free_cluster,
2615                                      block_group_list);
2616
2617                 WARN_ON(cluster->block_group != block_group);
2618                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2619
2620                 cond_resched_lock(&ctl->tree_lock);
2621         }
2622         __btrfs_remove_free_space_cache_locked(ctl);
2623         spin_unlock(&ctl->tree_lock);
2624
2625 }
2626
2627 u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
2628                                u64 offset, u64 bytes, u64 empty_size,
2629                                u64 *max_extent_size)
2630 {
2631         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2632         struct btrfs_free_space *entry = NULL;
2633         u64 bytes_search = bytes + empty_size;
2634         u64 ret = 0;
2635         u64 align_gap = 0;
2636         u64 align_gap_len = 0;
2637
2638         spin_lock(&ctl->tree_lock);
2639         entry = find_free_space(ctl, &offset, &bytes_search,
2640                                 block_group->full_stripe_len, max_extent_size);
2641         if (!entry)
2642                 goto out;
2643
2644         ret = offset;
2645         if (entry->bitmap) {
2646                 bitmap_clear_bits(ctl, entry, offset, bytes);
2647                 if (!entry->bytes)
2648                         free_bitmap(ctl, entry);
2649         } else {
2650                 unlink_free_space(ctl, entry);
2651                 align_gap_len = offset - entry->offset;
2652                 align_gap = entry->offset;
2653
2654                 entry->offset = offset + bytes;
2655                 WARN_ON(entry->bytes < bytes + align_gap_len);
2656
2657                 entry->bytes -= bytes + align_gap_len;
2658                 if (!entry->bytes)
2659                         kmem_cache_free(btrfs_free_space_cachep, entry);
2660                 else
2661                         link_free_space(ctl, entry);
2662         }
2663 out:
2664         spin_unlock(&ctl->tree_lock);
2665
2666         if (align_gap_len)
2667                 __btrfs_add_free_space(block_group->fs_info, ctl,
2668                                        align_gap, align_gap_len);
2669         return ret;
2670 }
2671
2672 /*
2673  * given a cluster, put all of its extents back into the free space
2674  * cache.  If a block group is passed, this function will only free
2675  * a cluster that belongs to the passed block group.
2676  *
2677  * Otherwise, it'll get a reference on the block group pointed to by the
2678  * cluster and remove the cluster from it.
2679  */
2680 int btrfs_return_cluster_to_free_space(
2681                                struct btrfs_block_group *block_group,
2682                                struct btrfs_free_cluster *cluster)
2683 {
2684         struct btrfs_free_space_ctl *ctl;
2685         int ret;
2686
2687         /* first, get a safe pointer to the block group */
2688         spin_lock(&cluster->lock);
2689         if (!block_group) {
2690                 block_group = cluster->block_group;
2691                 if (!block_group) {
2692                         spin_unlock(&cluster->lock);
2693                         return 0;
2694                 }
2695         } else if (cluster->block_group != block_group) {
2696                 /* someone else has already freed it don't redo their work */
2697                 spin_unlock(&cluster->lock);
2698                 return 0;
2699         }
2700         atomic_inc(&block_group->count);
2701         spin_unlock(&cluster->lock);
2702
2703         ctl = block_group->free_space_ctl;
2704
2705         /* now return any extents the cluster had on it */
2706         spin_lock(&ctl->tree_lock);
2707         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2708         spin_unlock(&ctl->tree_lock);
2709
2710         /* finally drop our ref */
2711         btrfs_put_block_group(block_group);
2712         return ret;
2713 }
2714
2715 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
2716                                    struct btrfs_free_cluster *cluster,
2717                                    struct btrfs_free_space *entry,
2718                                    u64 bytes, u64 min_start,
2719                                    u64 *max_extent_size)
2720 {
2721         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2722         int err;
2723         u64 search_start = cluster->window_start;
2724         u64 search_bytes = bytes;
2725         u64 ret = 0;
2726
2727         search_start = min_start;
2728         search_bytes = bytes;
2729
2730         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2731         if (err) {
2732                 *max_extent_size = max(get_max_extent_size(entry),
2733                                        *max_extent_size);
2734                 return 0;
2735         }
2736
2737         ret = search_start;
2738         __bitmap_clear_bits(ctl, entry, ret, bytes);
2739
2740         return ret;
2741 }
2742
2743 /*
2744  * given a cluster, try to allocate 'bytes' from it, returns 0
2745  * if it couldn't find anything suitably large, or a logical disk offset
2746  * if things worked out
2747  */
2748 u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
2749                              struct btrfs_free_cluster *cluster, u64 bytes,
2750                              u64 min_start, u64 *max_extent_size)
2751 {
2752         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2753         struct btrfs_free_space *entry = NULL;
2754         struct rb_node *node;
2755         u64 ret = 0;
2756
2757         spin_lock(&cluster->lock);
2758         if (bytes > cluster->max_size)
2759                 goto out;
2760
2761         if (cluster->block_group != block_group)
2762                 goto out;
2763
2764         node = rb_first(&cluster->root);
2765         if (!node)
2766                 goto out;
2767
2768         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2769         while (1) {
2770                 if (entry->bytes < bytes)
2771                         *max_extent_size = max(get_max_extent_size(entry),
2772                                                *max_extent_size);
2773
2774                 if (entry->bytes < bytes ||
2775                     (!entry->bitmap && entry->offset < min_start)) {
2776                         node = rb_next(&entry->offset_index);
2777                         if (!node)
2778                                 break;
2779                         entry = rb_entry(node, struct btrfs_free_space,
2780                                          offset_index);
2781                         continue;
2782                 }
2783
2784                 if (entry->bitmap) {
2785                         ret = btrfs_alloc_from_bitmap(block_group,
2786                                                       cluster, entry, bytes,
2787                                                       cluster->window_start,
2788                                                       max_extent_size);
2789                         if (ret == 0) {
2790                                 node = rb_next(&entry->offset_index);
2791                                 if (!node)
2792                                         break;
2793                                 entry = rb_entry(node, struct btrfs_free_space,
2794                                                  offset_index);
2795                                 continue;
2796                         }
2797                         cluster->window_start += bytes;
2798                 } else {
2799                         ret = entry->offset;
2800
2801                         entry->offset += bytes;
2802                         entry->bytes -= bytes;
2803                 }
2804
2805                 if (entry->bytes == 0)
2806                         rb_erase(&entry->offset_index, &cluster->root);
2807                 break;
2808         }
2809 out:
2810         spin_unlock(&cluster->lock);
2811
2812         if (!ret)
2813                 return 0;
2814
2815         spin_lock(&ctl->tree_lock);
2816
2817         ctl->free_space -= bytes;
2818         if (entry->bytes == 0) {
2819                 ctl->free_extents--;
2820                 if (entry->bitmap) {
2821                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2822                                         entry->bitmap);
2823                         ctl->total_bitmaps--;
2824                         ctl->op->recalc_thresholds(ctl);
2825                 }
2826                 kmem_cache_free(btrfs_free_space_cachep, entry);
2827         }
2828
2829         spin_unlock(&ctl->tree_lock);
2830
2831         return ret;
2832 }
2833
2834 static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
2835                                 struct btrfs_free_space *entry,
2836                                 struct btrfs_free_cluster *cluster,
2837                                 u64 offset, u64 bytes,
2838                                 u64 cont1_bytes, u64 min_bytes)
2839 {
2840         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2841         unsigned long next_zero;
2842         unsigned long i;
2843         unsigned long want_bits;
2844         unsigned long min_bits;
2845         unsigned long found_bits;
2846         unsigned long max_bits = 0;
2847         unsigned long start = 0;
2848         unsigned long total_found = 0;
2849         int ret;
2850
2851         i = offset_to_bit(entry->offset, ctl->unit,
2852                           max_t(u64, offset, entry->offset));
2853         want_bits = bytes_to_bits(bytes, ctl->unit);
2854         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2855
2856         /*
2857          * Don't bother looking for a cluster in this bitmap if it's heavily
2858          * fragmented.
2859          */
2860         if (entry->max_extent_size &&
2861             entry->max_extent_size < cont1_bytes)
2862                 return -ENOSPC;
2863 again:
2864         found_bits = 0;
2865         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2866                 next_zero = find_next_zero_bit(entry->bitmap,
2867                                                BITS_PER_BITMAP, i);
2868                 if (next_zero - i >= min_bits) {
2869                         found_bits = next_zero - i;
2870                         if (found_bits > max_bits)
2871                                 max_bits = found_bits;
2872                         break;
2873                 }
2874                 if (next_zero - i > max_bits)
2875                         max_bits = next_zero - i;
2876                 i = next_zero;
2877         }
2878
2879         if (!found_bits) {
2880                 entry->max_extent_size = (u64)max_bits * ctl->unit;
2881                 return -ENOSPC;
2882         }
2883
2884         if (!total_found) {
2885                 start = i;
2886                 cluster->max_size = 0;
2887         }
2888
2889         total_found += found_bits;
2890
2891         if (cluster->max_size < found_bits * ctl->unit)
2892                 cluster->max_size = found_bits * ctl->unit;
2893
2894         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2895                 i = next_zero + 1;
2896                 goto again;
2897         }
2898
2899         cluster->window_start = start * ctl->unit + entry->offset;
2900         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2901         ret = tree_insert_offset(&cluster->root, entry->offset,
2902                                  &entry->offset_index, 1);
2903         ASSERT(!ret); /* -EEXIST; Logic error */
2904
2905         trace_btrfs_setup_cluster(block_group, cluster,
2906                                   total_found * ctl->unit, 1);
2907         return 0;
2908 }
2909
2910 /*
2911  * This searches the block group for just extents to fill the cluster with.
2912  * Try to find a cluster with at least bytes total bytes, at least one
2913  * extent of cont1_bytes, and other clusters of at least min_bytes.
2914  */
2915 static noinline int
2916 setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
2917                         struct btrfs_free_cluster *cluster,
2918                         struct list_head *bitmaps, u64 offset, u64 bytes,
2919                         u64 cont1_bytes, u64 min_bytes)
2920 {
2921         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2922         struct btrfs_free_space *first = NULL;
2923         struct btrfs_free_space *entry = NULL;
2924         struct btrfs_free_space *last;
2925         struct rb_node *node;
2926         u64 window_free;
2927         u64 max_extent;
2928         u64 total_size = 0;
2929
2930         entry = tree_search_offset(ctl, offset, 0, 1);
2931         if (!entry)
2932                 return -ENOSPC;
2933
2934         /*
2935          * We don't want bitmaps, so just move along until we find a normal
2936          * extent entry.
2937          */
2938         while (entry->bitmap || entry->bytes < min_bytes) {
2939                 if (entry->bitmap && list_empty(&entry->list))
2940                         list_add_tail(&entry->list, bitmaps);
2941                 node = rb_next(&entry->offset_index);
2942                 if (!node)
2943                         return -ENOSPC;
2944                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2945         }
2946
2947         window_free = entry->bytes;
2948         max_extent = entry->bytes;
2949         first = entry;
2950         last = entry;
2951
2952         for (node = rb_next(&entry->offset_index); node;
2953              node = rb_next(&entry->offset_index)) {
2954                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2955
2956                 if (entry->bitmap) {
2957                         if (list_empty(&entry->list))
2958                                 list_add_tail(&entry->list, bitmaps);
2959                         continue;
2960                 }
2961
2962                 if (entry->bytes < min_bytes)
2963                         continue;
2964
2965                 last = entry;
2966                 window_free += entry->bytes;
2967                 if (entry->bytes > max_extent)
2968                         max_extent = entry->bytes;
2969         }
2970
2971         if (window_free < bytes || max_extent < cont1_bytes)
2972                 return -ENOSPC;
2973
2974         cluster->window_start = first->offset;
2975
2976         node = &first->offset_index;
2977
2978         /*
2979          * now we've found our entries, pull them out of the free space
2980          * cache and put them into the cluster rbtree
2981          */
2982         do {
2983                 int ret;
2984
2985                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2986                 node = rb_next(&entry->offset_index);
2987                 if (entry->bitmap || entry->bytes < min_bytes)
2988                         continue;
2989
2990                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2991                 ret = tree_insert_offset(&cluster->root, entry->offset,
2992                                          &entry->offset_index, 0);
2993                 total_size += entry->bytes;
2994                 ASSERT(!ret); /* -EEXIST; Logic error */
2995         } while (node && entry != last);
2996
2997         cluster->max_size = max_extent;
2998         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2999         return 0;
3000 }
3001
3002 /*
3003  * This specifically looks for bitmaps that may work in the cluster, we assume
3004  * that we have already failed to find extents that will work.
3005  */
3006 static noinline int
3007 setup_cluster_bitmap(struct btrfs_block_group *block_group,
3008                      struct btrfs_free_cluster *cluster,
3009                      struct list_head *bitmaps, u64 offset, u64 bytes,
3010                      u64 cont1_bytes, u64 min_bytes)
3011 {
3012         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3013         struct btrfs_free_space *entry = NULL;
3014         int ret = -ENOSPC;
3015         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3016
3017         if (ctl->total_bitmaps == 0)
3018                 return -ENOSPC;
3019
3020         /*
3021          * The bitmap that covers offset won't be in the list unless offset
3022          * is just its start offset.
3023          */
3024         if (!list_empty(bitmaps))
3025                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3026
3027         if (!entry || entry->offset != bitmap_offset) {
3028                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3029                 if (entry && list_empty(&entry->list))
3030                         list_add(&entry->list, bitmaps);
3031         }
3032
3033         list_for_each_entry(entry, bitmaps, list) {
3034                 if (entry->bytes < bytes)
3035                         continue;
3036                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3037                                            bytes, cont1_bytes, min_bytes);
3038                 if (!ret)
3039                         return 0;
3040         }
3041
3042         /*
3043          * The bitmaps list has all the bitmaps that record free space
3044          * starting after offset, so no more search is required.
3045          */
3046         return -ENOSPC;
3047 }
3048
3049 /*
3050  * here we try to find a cluster of blocks in a block group.  The goal
3051  * is to find at least bytes+empty_size.
3052  * We might not find them all in one contiguous area.
3053  *
3054  * returns zero and sets up cluster if things worked out, otherwise
3055  * it returns -enospc
3056  */
3057 int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3058                              struct btrfs_free_cluster *cluster,
3059                              u64 offset, u64 bytes, u64 empty_size)
3060 {
3061         struct btrfs_fs_info *fs_info = block_group->fs_info;
3062         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3063         struct btrfs_free_space *entry, *tmp;
3064         LIST_HEAD(bitmaps);
3065         u64 min_bytes;
3066         u64 cont1_bytes;
3067         int ret;
3068
3069         /*
3070          * Choose the minimum extent size we'll require for this
3071          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3072          * For metadata, allow allocates with smaller extents.  For
3073          * data, keep it dense.
3074          */
3075         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3076                 cont1_bytes = min_bytes = bytes + empty_size;
3077         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3078                 cont1_bytes = bytes;
3079                 min_bytes = fs_info->sectorsize;
3080         } else {
3081                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3082                 min_bytes = fs_info->sectorsize;
3083         }
3084
3085         spin_lock(&ctl->tree_lock);
3086
3087         /*
3088          * If we know we don't have enough space to make a cluster don't even
3089          * bother doing all the work to try and find one.
3090          */
3091         if (ctl->free_space < bytes) {
3092                 spin_unlock(&ctl->tree_lock);
3093                 return -ENOSPC;
3094         }
3095
3096         spin_lock(&cluster->lock);
3097
3098         /* someone already found a cluster, hooray */
3099         if (cluster->block_group) {
3100                 ret = 0;
3101                 goto out;
3102         }
3103
3104         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3105                                  min_bytes);
3106
3107         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3108                                       bytes + empty_size,
3109                                       cont1_bytes, min_bytes);
3110         if (ret)
3111                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3112                                            offset, bytes + empty_size,
3113                                            cont1_bytes, min_bytes);
3114
3115         /* Clear our temporary list */
3116         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3117                 list_del_init(&entry->list);
3118
3119         if (!ret) {
3120                 atomic_inc(&block_group->count);
3121                 list_add_tail(&cluster->block_group_list,
3122                               &block_group->cluster_list);
3123                 cluster->block_group = block_group;
3124         } else {
3125                 trace_btrfs_failed_cluster_setup(block_group);
3126         }
3127 out:
3128         spin_unlock(&cluster->lock);
3129         spin_unlock(&ctl->tree_lock);
3130
3131         return ret;
3132 }
3133
3134 /*
3135  * simple code to zero out a cluster
3136  */
3137 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3138 {
3139         spin_lock_init(&cluster->lock);
3140         spin_lock_init(&cluster->refill_lock);
3141         cluster->root = RB_ROOT;
3142         cluster->max_size = 0;
3143         cluster->fragmented = false;
3144         INIT_LIST_HEAD(&cluster->block_group_list);
3145         cluster->block_group = NULL;
3146 }
3147
3148 static int do_trimming(struct btrfs_block_group *block_group,
3149                        u64 *total_trimmed, u64 start, u64 bytes,
3150                        u64 reserved_start, u64 reserved_bytes,
3151                        struct btrfs_trim_range *trim_entry)
3152 {
3153         struct btrfs_space_info *space_info = block_group->space_info;
3154         struct btrfs_fs_info *fs_info = block_group->fs_info;
3155         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3156         int ret;
3157         int update = 0;
3158         u64 trimmed = 0;
3159
3160         spin_lock(&space_info->lock);
3161         spin_lock(&block_group->lock);
3162         if (!block_group->ro) {
3163                 block_group->reserved += reserved_bytes;
3164                 space_info->bytes_reserved += reserved_bytes;
3165                 update = 1;
3166         }
3167         spin_unlock(&block_group->lock);
3168         spin_unlock(&space_info->lock);
3169
3170         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3171         if (!ret)
3172                 *total_trimmed += trimmed;
3173
3174         mutex_lock(&ctl->cache_writeout_mutex);
3175         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
3176         list_del(&trim_entry->list);
3177         mutex_unlock(&ctl->cache_writeout_mutex);
3178
3179         if (update) {
3180                 spin_lock(&space_info->lock);
3181                 spin_lock(&block_group->lock);
3182                 if (block_group->ro)
3183                         space_info->bytes_readonly += reserved_bytes;
3184                 block_group->reserved -= reserved_bytes;
3185                 space_info->bytes_reserved -= reserved_bytes;
3186                 spin_unlock(&block_group->lock);
3187                 spin_unlock(&space_info->lock);
3188         }
3189
3190         return ret;
3191 }
3192
3193 static int trim_no_bitmap(struct btrfs_block_group *block_group,
3194                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3195 {
3196         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3197         struct btrfs_free_space *entry;
3198         struct rb_node *node;
3199         int ret = 0;
3200         u64 extent_start;
3201         u64 extent_bytes;
3202         u64 bytes;
3203
3204         while (start < end) {
3205                 struct btrfs_trim_range trim_entry;
3206
3207                 mutex_lock(&ctl->cache_writeout_mutex);
3208                 spin_lock(&ctl->tree_lock);
3209
3210                 if (ctl->free_space < minlen) {
3211                         spin_unlock(&ctl->tree_lock);
3212                         mutex_unlock(&ctl->cache_writeout_mutex);
3213                         break;
3214                 }
3215
3216                 entry = tree_search_offset(ctl, start, 0, 1);
3217                 if (!entry) {
3218                         spin_unlock(&ctl->tree_lock);
3219                         mutex_unlock(&ctl->cache_writeout_mutex);
3220                         break;
3221                 }
3222
3223                 /* skip bitmaps */
3224                 while (entry->bitmap) {
3225                         node = rb_next(&entry->offset_index);
3226                         if (!node) {
3227                                 spin_unlock(&ctl->tree_lock);
3228                                 mutex_unlock(&ctl->cache_writeout_mutex);
3229                                 goto out;
3230                         }
3231                         entry = rb_entry(node, struct btrfs_free_space,
3232                                          offset_index);
3233                 }
3234
3235                 if (entry->offset >= end) {
3236                         spin_unlock(&ctl->tree_lock);
3237                         mutex_unlock(&ctl->cache_writeout_mutex);
3238                         break;
3239                 }
3240
3241                 extent_start = entry->offset;
3242                 extent_bytes = entry->bytes;
3243                 start = max(start, extent_start);
3244                 bytes = min(extent_start + extent_bytes, end) - start;
3245                 if (bytes < minlen) {
3246                         spin_unlock(&ctl->tree_lock);
3247                         mutex_unlock(&ctl->cache_writeout_mutex);
3248                         goto next;
3249                 }
3250
3251                 unlink_free_space(ctl, entry);
3252                 kmem_cache_free(btrfs_free_space_cachep, entry);
3253
3254                 spin_unlock(&ctl->tree_lock);
3255                 trim_entry.start = extent_start;
3256                 trim_entry.bytes = extent_bytes;
3257                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3258                 mutex_unlock(&ctl->cache_writeout_mutex);
3259
3260                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3261                                   extent_start, extent_bytes, &trim_entry);
3262                 if (ret)
3263                         break;
3264 next:
3265                 start += bytes;
3266
3267                 if (fatal_signal_pending(current)) {
3268                         ret = -ERESTARTSYS;
3269                         break;
3270                 }
3271
3272                 cond_resched();
3273         }
3274 out:
3275         return ret;
3276 }
3277
3278 static int trim_bitmaps(struct btrfs_block_group *block_group,
3279                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3280 {
3281         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3282         struct btrfs_free_space *entry;
3283         int ret = 0;
3284         int ret2;
3285         u64 bytes;
3286         u64 offset = offset_to_bitmap(ctl, start);
3287
3288         while (offset < end) {
3289                 bool next_bitmap = false;
3290                 struct btrfs_trim_range trim_entry;
3291
3292                 mutex_lock(&ctl->cache_writeout_mutex);
3293                 spin_lock(&ctl->tree_lock);
3294
3295                 if (ctl->free_space < minlen) {
3296                         spin_unlock(&ctl->tree_lock);
3297                         mutex_unlock(&ctl->cache_writeout_mutex);
3298                         break;
3299                 }
3300
3301                 entry = tree_search_offset(ctl, offset, 1, 0);
3302                 if (!entry) {
3303                         spin_unlock(&ctl->tree_lock);
3304                         mutex_unlock(&ctl->cache_writeout_mutex);
3305                         next_bitmap = true;
3306                         goto next;
3307                 }
3308
3309                 bytes = minlen;
3310                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3311                 if (ret2 || start >= end) {
3312                         spin_unlock(&ctl->tree_lock);
3313                         mutex_unlock(&ctl->cache_writeout_mutex);
3314                         next_bitmap = true;
3315                         goto next;
3316                 }
3317
3318                 bytes = min(bytes, end - start);
3319                 if (bytes < minlen) {
3320                         spin_unlock(&ctl->tree_lock);
3321                         mutex_unlock(&ctl->cache_writeout_mutex);
3322                         goto next;
3323                 }
3324
3325                 bitmap_clear_bits(ctl, entry, start, bytes);
3326                 if (entry->bytes == 0)
3327                         free_bitmap(ctl, entry);
3328
3329                 spin_unlock(&ctl->tree_lock);
3330                 trim_entry.start = start;
3331                 trim_entry.bytes = bytes;
3332                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3333                 mutex_unlock(&ctl->cache_writeout_mutex);
3334
3335                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3336                                   start, bytes, &trim_entry);
3337                 if (ret)
3338                         break;
3339 next:
3340                 if (next_bitmap) {
3341                         offset += BITS_PER_BITMAP * ctl->unit;
3342                 } else {
3343                         start += bytes;
3344                         if (start >= offset + BITS_PER_BITMAP * ctl->unit)
3345                                 offset += BITS_PER_BITMAP * ctl->unit;
3346                 }
3347
3348                 if (fatal_signal_pending(current)) {
3349                         ret = -ERESTARTSYS;
3350                         break;
3351                 }
3352
3353                 cond_resched();
3354         }
3355
3356         return ret;
3357 }
3358
3359 void btrfs_get_block_group_trimming(struct btrfs_block_group *cache)
3360 {
3361         atomic_inc(&cache->trimming);
3362 }
3363
3364 void btrfs_put_block_group_trimming(struct btrfs_block_group *block_group)
3365 {
3366         struct btrfs_fs_info *fs_info = block_group->fs_info;
3367         struct extent_map_tree *em_tree;
3368         struct extent_map *em;
3369         bool cleanup;
3370
3371         spin_lock(&block_group->lock);
3372         cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3373                    block_group->removed);
3374         spin_unlock(&block_group->lock);
3375
3376         if (cleanup) {
3377                 mutex_lock(&fs_info->chunk_mutex);
3378                 em_tree = &fs_info->mapping_tree;
3379                 write_lock(&em_tree->lock);
3380                 em = lookup_extent_mapping(em_tree, block_group->start,
3381                                            1);
3382                 BUG_ON(!em); /* logic error, can't happen */
3383                 remove_extent_mapping(em_tree, em);
3384                 write_unlock(&em_tree->lock);
3385                 mutex_unlock(&fs_info->chunk_mutex);
3386
3387                 /* once for us and once for the tree */
3388                 free_extent_map(em);
3389                 free_extent_map(em);
3390
3391                 /*
3392                  * We've left one free space entry and other tasks trimming
3393                  * this block group have left 1 entry each one. Free them.
3394                  */
3395                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3396         }
3397 }
3398
3399 int btrfs_trim_block_group(struct btrfs_block_group *block_group,
3400                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3401 {
3402         int ret;
3403
3404         *trimmed = 0;
3405
3406         spin_lock(&block_group->lock);
3407         if (block_group->removed) {
3408                 spin_unlock(&block_group->lock);
3409                 return 0;
3410         }
3411         btrfs_get_block_group_trimming(block_group);
3412         spin_unlock(&block_group->lock);
3413
3414         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3415         if (ret)
3416                 goto out;
3417
3418         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3419 out:
3420         btrfs_put_block_group_trimming(block_group);
3421         return ret;
3422 }
3423
3424 /*
3425  * Find the left-most item in the cache tree, and then return the
3426  * smallest inode number in the item.
3427  *
3428  * Note: the returned inode number may not be the smallest one in
3429  * the tree, if the left-most item is a bitmap.
3430  */
3431 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3432 {
3433         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3434         struct btrfs_free_space *entry = NULL;
3435         u64 ino = 0;
3436
3437         spin_lock(&ctl->tree_lock);
3438
3439         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3440                 goto out;
3441
3442         entry = rb_entry(rb_first(&ctl->free_space_offset),
3443                          struct btrfs_free_space, offset_index);
3444
3445         if (!entry->bitmap) {
3446                 ino = entry->offset;
3447
3448                 unlink_free_space(ctl, entry);
3449                 entry->offset++;
3450                 entry->bytes--;
3451                 if (!entry->bytes)
3452                         kmem_cache_free(btrfs_free_space_cachep, entry);
3453                 else
3454                         link_free_space(ctl, entry);
3455         } else {
3456                 u64 offset = 0;
3457                 u64 count = 1;
3458                 int ret;
3459
3460                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3461                 /* Logic error; Should be empty if it can't find anything */
3462                 ASSERT(!ret);
3463
3464                 ino = offset;
3465                 bitmap_clear_bits(ctl, entry, offset, 1);
3466                 if (entry->bytes == 0)
3467                         free_bitmap(ctl, entry);
3468         }
3469 out:
3470         spin_unlock(&ctl->tree_lock);
3471
3472         return ino;
3473 }
3474
3475 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3476                                     struct btrfs_path *path)
3477 {
3478         struct inode *inode = NULL;
3479
3480         spin_lock(&root->ino_cache_lock);
3481         if (root->ino_cache_inode)
3482                 inode = igrab(root->ino_cache_inode);
3483         spin_unlock(&root->ino_cache_lock);
3484         if (inode)
3485                 return inode;
3486
3487         inode = __lookup_free_space_inode(root, path, 0);
3488         if (IS_ERR(inode))
3489                 return inode;
3490
3491         spin_lock(&root->ino_cache_lock);
3492         if (!btrfs_fs_closing(root->fs_info))
3493                 root->ino_cache_inode = igrab(inode);
3494         spin_unlock(&root->ino_cache_lock);
3495
3496         return inode;
3497 }
3498
3499 int create_free_ino_inode(struct btrfs_root *root,
3500                           struct btrfs_trans_handle *trans,
3501                           struct btrfs_path *path)
3502 {
3503         return __create_free_space_inode(root, trans, path,
3504                                          BTRFS_FREE_INO_OBJECTID, 0);
3505 }
3506
3507 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3508 {
3509         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3510         struct btrfs_path *path;
3511         struct inode *inode;
3512         int ret = 0;
3513         u64 root_gen = btrfs_root_generation(&root->root_item);
3514
3515         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3516                 return 0;
3517
3518         /*
3519          * If we're unmounting then just return, since this does a search on the
3520          * normal root and not the commit root and we could deadlock.
3521          */
3522         if (btrfs_fs_closing(fs_info))
3523                 return 0;
3524
3525         path = btrfs_alloc_path();
3526         if (!path)
3527                 return 0;
3528
3529         inode = lookup_free_ino_inode(root, path);
3530         if (IS_ERR(inode))
3531                 goto out;
3532
3533         if (root_gen != BTRFS_I(inode)->generation)
3534                 goto out_put;
3535
3536         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3537
3538         if (ret < 0)
3539                 btrfs_err(fs_info,
3540                         "failed to load free ino cache for root %llu",
3541                         root->root_key.objectid);
3542 out_put:
3543         iput(inode);
3544 out:
3545         btrfs_free_path(path);
3546         return ret;
3547 }
3548
3549 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3550                               struct btrfs_trans_handle *trans,
3551                               struct btrfs_path *path,
3552                               struct inode *inode)
3553 {
3554         struct btrfs_fs_info *fs_info = root->fs_info;
3555         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3556         int ret;
3557         struct btrfs_io_ctl io_ctl;
3558         bool release_metadata = true;
3559
3560         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3561                 return 0;
3562
3563         memset(&io_ctl, 0, sizeof(io_ctl));
3564         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl, trans);
3565         if (!ret) {
3566                 /*
3567                  * At this point writepages() didn't error out, so our metadata
3568                  * reservation is released when the writeback finishes, at
3569                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3570                  * with or without an error.
3571                  */
3572                 release_metadata = false;
3573                 ret = btrfs_wait_cache_io_root(root, trans, &io_ctl, path);
3574         }
3575
3576         if (ret) {
3577                 if (release_metadata)
3578                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3579                                         inode->i_size, true);
3580 #ifdef DEBUG
3581                 btrfs_err(fs_info,
3582                           "failed to write free ino cache for root %llu",
3583                           root->root_key.objectid);
3584 #endif
3585         }
3586
3587         return ret;
3588 }
3589
3590 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3591 /*
3592  * Use this if you need to make a bitmap or extent entry specifically, it
3593  * doesn't do any of the merging that add_free_space does, this acts a lot like
3594  * how the free space cache loading stuff works, so you can get really weird
3595  * configurations.
3596  */
3597 int test_add_free_space_entry(struct btrfs_block_group *cache,
3598                               u64 offset, u64 bytes, bool bitmap)
3599 {
3600         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3601         struct btrfs_free_space *info = NULL, *bitmap_info;
3602         void *map = NULL;
3603         u64 bytes_added;
3604         int ret;
3605
3606 again:
3607         if (!info) {
3608                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3609                 if (!info)
3610                         return -ENOMEM;
3611         }
3612
3613         if (!bitmap) {
3614                 spin_lock(&ctl->tree_lock);
3615                 info->offset = offset;
3616                 info->bytes = bytes;
3617                 info->max_extent_size = 0;
3618                 ret = link_free_space(ctl, info);
3619                 spin_unlock(&ctl->tree_lock);
3620                 if (ret)
3621                         kmem_cache_free(btrfs_free_space_cachep, info);
3622                 return ret;
3623         }
3624
3625         if (!map) {
3626                 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
3627                 if (!map) {
3628                         kmem_cache_free(btrfs_free_space_cachep, info);
3629                         return -ENOMEM;
3630                 }
3631         }
3632
3633         spin_lock(&ctl->tree_lock);
3634         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3635                                          1, 0);
3636         if (!bitmap_info) {
3637                 info->bitmap = map;
3638                 map = NULL;
3639                 add_new_bitmap(ctl, info, offset);
3640                 bitmap_info = info;
3641                 info = NULL;
3642         }
3643
3644         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3645
3646         bytes -= bytes_added;
3647         offset += bytes_added;
3648         spin_unlock(&ctl->tree_lock);
3649
3650         if (bytes)
3651                 goto again;
3652
3653         if (info)
3654                 kmem_cache_free(btrfs_free_space_cachep, info);
3655         if (map)
3656                 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
3657         return 0;
3658 }
3659
3660 /*
3661  * Checks to see if the given range is in the free space cache.  This is really
3662  * just used to check the absence of space, so if there is free space in the
3663  * range at all we will return 1.
3664  */
3665 int test_check_exists(struct btrfs_block_group *cache,
3666                       u64 offset, u64 bytes)
3667 {
3668         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3669         struct btrfs_free_space *info;
3670         int ret = 0;
3671
3672         spin_lock(&ctl->tree_lock);
3673         info = tree_search_offset(ctl, offset, 0, 0);
3674         if (!info) {
3675                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3676                                           1, 0);
3677                 if (!info)
3678                         goto out;
3679         }
3680
3681 have_info:
3682         if (info->bitmap) {
3683                 u64 bit_off, bit_bytes;
3684                 struct rb_node *n;
3685                 struct btrfs_free_space *tmp;
3686
3687                 bit_off = offset;
3688                 bit_bytes = ctl->unit;
3689                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3690                 if (!ret) {
3691                         if (bit_off == offset) {
3692                                 ret = 1;
3693                                 goto out;
3694                         } else if (bit_off > offset &&
3695                                    offset + bytes > bit_off) {
3696                                 ret = 1;
3697                                 goto out;
3698                         }
3699                 }
3700
3701                 n = rb_prev(&info->offset_index);
3702                 while (n) {
3703                         tmp = rb_entry(n, struct btrfs_free_space,
3704                                        offset_index);
3705                         if (tmp->offset + tmp->bytes < offset)
3706                                 break;
3707                         if (offset + bytes < tmp->offset) {
3708                                 n = rb_prev(&tmp->offset_index);
3709                                 continue;
3710                         }
3711                         info = tmp;
3712                         goto have_info;
3713                 }
3714
3715                 n = rb_next(&info->offset_index);
3716                 while (n) {
3717                         tmp = rb_entry(n, struct btrfs_free_space,
3718                                        offset_index);
3719                         if (offset + bytes < tmp->offset)
3720                                 break;
3721                         if (tmp->offset + tmp->bytes < offset) {
3722                                 n = rb_next(&tmp->offset_index);
3723                                 continue;
3724                         }
3725                         info = tmp;
3726                         goto have_info;
3727                 }
3728
3729                 ret = 0;
3730                 goto out;
3731         }
3732
3733         if (info->offset == offset) {
3734                 ret = 1;
3735                 goto out;
3736         }
3737
3738         if (offset > info->offset && offset < info->offset + info->bytes)
3739                 ret = 1;
3740 out:
3741         spin_unlock(&ctl->tree_lock);
3742         return ret;
3743 }
3744 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */