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 #include "discard.h"
  25
  26 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  27 #define MAX_CACHE_BYTES_PER_GIG SZ_64K
  28 #define FORCE_EXTENT_THRESHOLD  SZ_1M
  29
  30 struct btrfs_trim_range {
  31         u64 start;
  32         u64 bytes;
  33         struct list_head list;
  34 };
  35
  36 static int count_bitmap_extents(struct btrfs_free_space_ctl *ctl,
  37                                 struct btrfs_free_space *bitmap_info);
  38 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  39                            struct btrfs_free_space *info);
  40 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  41                               struct btrfs_free_space *info);
  42 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
  43                              struct btrfs_trans_handle *trans,
  44                              struct btrfs_io_ctl *io_ctl,
  45                              struct btrfs_path *path);
  46
  47 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  48                                                struct btrfs_path *path,
  49                                                u64 offset)
  50 {
  51         struct btrfs_fs_info *fs_info = root->fs_info;
  52         struct btrfs_key key;
  53         struct btrfs_key location;
  54         struct btrfs_disk_key disk_key;
  55         struct btrfs_free_space_header *header;
  56         struct extent_buffer *leaf;
  57         struct inode *inode = NULL;
  58         unsigned nofs_flag;
  59         int ret;
  60
  61         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  62         key.offset = offset;
  63         key.type = 0;
  64
  65         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  66         if (ret < 0)
  67                 return ERR_PTR(ret);
  68         if (ret > 0) {
  69                 btrfs_release_path(path);
  70                 return ERR_PTR(-ENOENT);
  71         }
  72
  73         leaf = path->nodes[0];
  74         header = btrfs_item_ptr(leaf, path->slots[0],
  75                                 struct btrfs_free_space_header);
  76         btrfs_free_space_key(leaf, header, &disk_key);
  77         btrfs_disk_key_to_cpu(&location, &disk_key);
  78         btrfs_release_path(path);
  79
  80         /*
  81          * We are often under a trans handle at this point, so we need to make
  82          * sure NOFS is set to keep us from deadlocking.
  83          */
  84         nofs_flag = memalloc_nofs_save();
  85         inode = btrfs_iget_path(fs_info->sb, location.objectid, root, path);
  86         btrfs_release_path(path);
  87         memalloc_nofs_restore(nofs_flag);
  88         if (IS_ERR(inode))
  89                 return inode;
  90
  91         mapping_set_gfp_mask(inode->i_mapping,
  92                         mapping_gfp_constraint(inode->i_mapping,
  93                         ~(__GFP_FS | __GFP_HIGHMEM)));
  94
  95         return inode;
  96 }
  97
  98 struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
  99                 struct btrfs_path *path)
 100 {
 101         struct btrfs_fs_info *fs_info = block_group->fs_info;
 102         struct inode *inode = NULL;
 103         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 104
 105         spin_lock(&block_group->lock);
 106         if (block_group->inode)
 107                 inode = igrab(block_group->inode);
 108         spin_unlock(&block_group->lock);
 109         if (inode)
 110                 return inode;
 111
 112         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 113                                           block_group->start);
 114         if (IS_ERR(inode))
 115                 return inode;
 116
 117         spin_lock(&block_group->lock);
 118         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 119                 btrfs_info(fs_info, "Old style space inode found, converting.");
 120                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 121                         BTRFS_INODE_NODATACOW;
 122                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 123         }
 124
 125         if (!block_group->iref) {
 126                 block_group->inode = igrab(inode);
 127                 block_group->iref = 1;
 128         }
 129         spin_unlock(&block_group->lock);
 130
 131         return inode;
 132 }
 133
 134 static int __create_free_space_inode(struct btrfs_root *root,
 135                                      struct btrfs_trans_handle *trans,
 136                                      struct btrfs_path *path,
 137                                      u64 ino, u64 offset)
 138 {
 139         struct btrfs_key key;
 140         struct btrfs_disk_key disk_key;
 141         struct btrfs_free_space_header *header;
 142         struct btrfs_inode_item *inode_item;
 143         struct extent_buffer *leaf;
 144         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 145         int ret;
 146
 147         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 148         if (ret)
 149                 return ret;
 150
 151         /* We inline crc's for the free disk space cache */
 152         if (ino != BTRFS_FREE_INO_OBJECTID)
 153                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 154
 155         leaf = path->nodes[0];
 156         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 157                                     struct btrfs_inode_item);
 158         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 159         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 160                              sizeof(*inode_item));
 161         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 162         btrfs_set_inode_size(leaf, inode_item, 0);
 163         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 164         btrfs_set_inode_uid(leaf, inode_item, 0);
 165         btrfs_set_inode_gid(leaf, inode_item, 0);
 166         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 167         btrfs_set_inode_flags(leaf, inode_item, flags);
 168         btrfs_set_inode_nlink(leaf, inode_item, 1);
 169         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 170         btrfs_set_inode_block_group(leaf, inode_item, offset);
 171         btrfs_mark_buffer_dirty(leaf);
 172         btrfs_release_path(path);
 173
 174         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 175         key.offset = offset;
 176         key.type = 0;
 177         ret = btrfs_insert_empty_item(trans, root, path, &key,
 178                                       sizeof(struct btrfs_free_space_header));
 179         if (ret < 0) {
 180                 btrfs_release_path(path);
 181                 return ret;
 182         }
 183
 184         leaf = path->nodes[0];
 185         header = btrfs_item_ptr(leaf, path->slots[0],
 186                                 struct btrfs_free_space_header);
 187         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 188         btrfs_set_free_space_key(leaf, header, &disk_key);
 189         btrfs_mark_buffer_dirty(leaf);
 190         btrfs_release_path(path);
 191
 192         return 0;
 193 }
 194
 195 int create_free_space_inode(struct btrfs_trans_handle *trans,
 196                             struct btrfs_block_group *block_group,
 197                             struct btrfs_path *path)
 198 {
 199         int ret;
 200         u64 ino;
 201
 202         ret = btrfs_find_free_objectid(trans->fs_info->tree_root, &ino);
 203         if (ret < 0)
 204                 return ret;
 205
 206         return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
 207                                          ino, block_group->start);
 208 }
 209
 210 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
 211                                        struct btrfs_block_rsv *rsv)
 212 {
 213         u64 needed_bytes;
 214         int ret;
 215
 216         /* 1 for slack space, 1 for updating the inode */
 217         needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
 218                 btrfs_calc_metadata_size(fs_info, 1);
 219
 220         spin_lock(&rsv->lock);
 221         if (rsv->reserved < needed_bytes)
 222                 ret = -ENOSPC;
 223         else
 224                 ret = 0;
 225         spin_unlock(&rsv->lock);
 226         return ret;
 227 }
 228
 229 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 230                                     struct btrfs_block_group *block_group,
 231                                     struct inode *inode)
 232 {
 233         struct btrfs_root *root = BTRFS_I(inode)->root;
 234         int ret = 0;
 235         bool locked = false;
 236
 237         if (block_group) {
 238                 struct btrfs_path *path = btrfs_alloc_path();
 239
 240                 if (!path) {
 241                         ret = -ENOMEM;
 242                         goto fail;
 243                 }
 244                 locked = true;
 245                 mutex_lock(&trans->transaction->cache_write_mutex);
 246                 if (!list_empty(&block_group->io_list)) {
 247                         list_del_init(&block_group->io_list);
 248
 249                         btrfs_wait_cache_io(trans, block_group, path);
 250                         btrfs_put_block_group(block_group);
 251                 }
 252
 253                 /*
 254                  * now that we've truncated the cache away, its no longer
 255                  * setup or written
 256                  */
 257                 spin_lock(&block_group->lock);
 258                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 259                 spin_unlock(&block_group->lock);
 260                 btrfs_free_path(path);
 261         }
 262
 263         btrfs_i_size_write(BTRFS_I(inode), 0);
 264         truncate_pagecache(inode, 0);
 265
 266         /*
 267          * We skip the throttling logic for free space cache inodes, so we don't
 268          * need to check for -EAGAIN.
 269          */
 270         ret = btrfs_truncate_inode_items(trans, root, inode,
 271                                          0, BTRFS_EXTENT_DATA_KEY);
 272         if (ret)
 273                 goto fail;
 274
 275         ret = btrfs_update_inode(trans, root, inode);
 276
 277 fail:
 278         if (locked)
 279                 mutex_unlock(&trans->transaction->cache_write_mutex);
 280         if (ret)
 281                 btrfs_abort_transaction(trans, ret);
 282
 283         return ret;
 284 }
 285
 286 static void readahead_cache(struct inode *inode)
 287 {
 288         struct file_ra_state *ra;
 289         unsigned long last_index;
 290
 291         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 292         if (!ra)
 293                 return;
 294
 295         file_ra_state_init(ra, inode->i_mapping);
 296         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 297
 298         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 299
 300         kfree(ra);
 301 }
 302
 303 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 304                        int write)
 305 {
 306         int num_pages;
 307         int check_crcs = 0;
 308
 309         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 310
 311         if (btrfs_ino(BTRFS_I(inode)) != BTRFS_FREE_INO_OBJECTID)
 312                 check_crcs = 1;
 313
 314         /* Make sure we can fit our crcs and generation into the first page */
 315         if (write && check_crcs &&
 316             (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
 317                 return -ENOSPC;
 318
 319         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 320
 321         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 322         if (!io_ctl->pages)
 323                 return -ENOMEM;
 324
 325         io_ctl->num_pages = num_pages;
 326         io_ctl->fs_info = btrfs_sb(inode->i_sb);
 327         io_ctl->check_crcs = check_crcs;
 328         io_ctl->inode = inode;
 329
 330         return 0;
 331 }
 332 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 333
 334 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 335 {
 336         kfree(io_ctl->pages);
 337         io_ctl->pages = NULL;
 338 }
 339
 340 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 341 {
 342         if (io_ctl->cur) {
 343                 io_ctl->cur = NULL;
 344                 io_ctl->orig = NULL;
 345         }
 346 }
 347
 348 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 349 {
 350         ASSERT(io_ctl->index < io_ctl->num_pages);
 351         io_ctl->page = io_ctl->pages[io_ctl->index++];
 352         io_ctl->cur = page_address(io_ctl->page);
 353         io_ctl->orig = io_ctl->cur;
 354         io_ctl->size = PAGE_SIZE;
 355         if (clear)
 356                 clear_page(io_ctl->cur);
 357 }
 358
 359 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 360 {
 361         int i;
 362
 363         io_ctl_unmap_page(io_ctl);
 364
 365         for (i = 0; i < io_ctl->num_pages; i++) {
 366                 if (io_ctl->pages[i]) {
 367                         ClearPageChecked(io_ctl->pages[i]);
 368                         unlock_page(io_ctl->pages[i]);
 369                         put_page(io_ctl->pages[i]);
 370                 }
 371         }
 372 }
 373
 374 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
 375 {
 376         struct page *page;
 377         struct inode *inode = io_ctl->inode;
 378         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 379         int i;
 380
 381         for (i = 0; i < io_ctl->num_pages; i++) {
 382                 page = find_or_create_page(inode->i_mapping, i, mask);
 383                 if (!page) {
 384                         io_ctl_drop_pages(io_ctl);
 385                         return -ENOMEM;
 386                 }
 387                 io_ctl->pages[i] = page;
 388                 if (uptodate && !PageUptodate(page)) {
 389                         btrfs_readpage(NULL, page);
 390                         lock_page(page);
 391                         if (page->mapping != inode->i_mapping) {
 392                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 393                                           "free space cache page truncated");
 394                                 io_ctl_drop_pages(io_ctl);
 395                                 return -EIO;
 396                         }
 397                         if (!PageUptodate(page)) {
 398                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 399                                            "error reading free space cache");
 400                                 io_ctl_drop_pages(io_ctl);
 401                                 return -EIO;
 402                         }
 403                 }
 404         }
 405
 406         for (i = 0; i < io_ctl->num_pages; i++) {
 407                 clear_page_dirty_for_io(io_ctl->pages[i]);
 408                 set_page_extent_mapped(io_ctl->pages[i]);
 409         }
 410
 411         return 0;
 412 }
 413
 414 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 415 {
 416         __le64 *val;
 417
 418         io_ctl_map_page(io_ctl, 1);
 419
 420         /*
 421          * Skip the csum areas.  If we don't check crcs then we just have a
 422          * 64bit chunk at the front of the first page.
 423          */
 424         if (io_ctl->check_crcs) {
 425                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 426                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 427         } else {
 428                 io_ctl->cur += sizeof(u64);
 429                 io_ctl->size -= sizeof(u64) * 2;
 430         }
 431
 432         val = io_ctl->cur;
 433         *val = cpu_to_le64(generation);
 434         io_ctl->cur += sizeof(u64);
 435 }
 436
 437 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 438 {
 439         __le64 *gen;
 440
 441         /*
 442          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 443          * chunk at the front of the first page.
 444          */
 445         if (io_ctl->check_crcs) {
 446                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 447                 io_ctl->size -= sizeof(u64) +
 448                         (sizeof(u32) * io_ctl->num_pages);
 449         } else {
 450                 io_ctl->cur += sizeof(u64);
 451                 io_ctl->size -= sizeof(u64) * 2;
 452         }
 453
 454         gen = io_ctl->cur;
 455         if (le64_to_cpu(*gen) != generation) {
 456                 btrfs_err_rl(io_ctl->fs_info,
 457                         "space cache generation (%llu) does not match inode (%llu)",
 458                                 *gen, generation);
 459                 io_ctl_unmap_page(io_ctl);
 460                 return -EIO;
 461         }
 462         io_ctl->cur += sizeof(u64);
 463         return 0;
 464 }
 465
 466 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 467 {
 468         u32 *tmp;
 469         u32 crc = ~(u32)0;
 470         unsigned offset = 0;
 471
 472         if (!io_ctl->check_crcs) {
 473                 io_ctl_unmap_page(io_ctl);
 474                 return;
 475         }
 476
 477         if (index == 0)
 478                 offset = sizeof(u32) * io_ctl->num_pages;
 479
 480         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 481         btrfs_crc32c_final(crc, (u8 *)&crc);
 482         io_ctl_unmap_page(io_ctl);
 483         tmp = page_address(io_ctl->pages[0]);
 484         tmp += index;
 485         *tmp = crc;
 486 }
 487
 488 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 489 {
 490         u32 *tmp, val;
 491         u32 crc = ~(u32)0;
 492         unsigned offset = 0;
 493
 494         if (!io_ctl->check_crcs) {
 495                 io_ctl_map_page(io_ctl, 0);
 496                 return 0;
 497         }
 498
 499         if (index == 0)
 500                 offset = sizeof(u32) * io_ctl->num_pages;
 501
 502         tmp = page_address(io_ctl->pages[0]);
 503         tmp += index;
 504         val = *tmp;
 505
 506         io_ctl_map_page(io_ctl, 0);
 507         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 508         btrfs_crc32c_final(crc, (u8 *)&crc);
 509         if (val != crc) {
 510                 btrfs_err_rl(io_ctl->fs_info,
 511                         "csum mismatch on free space cache");
 512                 io_ctl_unmap_page(io_ctl);
 513                 return -EIO;
 514         }
 515
 516         return 0;
 517 }
 518
 519 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 520                             void *bitmap)
 521 {
 522         struct btrfs_free_space_entry *entry;
 523
 524         if (!io_ctl->cur)
 525                 return -ENOSPC;
 526
 527         entry = io_ctl->cur;
 528         entry->offset = cpu_to_le64(offset);
 529         entry->bytes = cpu_to_le64(bytes);
 530         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 531                 BTRFS_FREE_SPACE_EXTENT;
 532         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 533         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 534
 535         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 536                 return 0;
 537
 538         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 539
 540         /* No more pages to map */
 541         if (io_ctl->index >= io_ctl->num_pages)
 542                 return 0;
 543
 544         /* map the next page */
 545         io_ctl_map_page(io_ctl, 1);
 546         return 0;
 547 }
 548
 549 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 550 {
 551         if (!io_ctl->cur)
 552                 return -ENOSPC;
 553
 554         /*
 555          * If we aren't at the start of the current page, unmap this one and
 556          * map the next one if there is any left.
 557          */
 558         if (io_ctl->cur != io_ctl->orig) {
 559                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 560                 if (io_ctl->index >= io_ctl->num_pages)
 561                         return -ENOSPC;
 562                 io_ctl_map_page(io_ctl, 0);
 563         }
 564
 565         copy_page(io_ctl->cur, bitmap);
 566         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 567         if (io_ctl->index < io_ctl->num_pages)
 568                 io_ctl_map_page(io_ctl, 0);
 569         return 0;
 570 }
 571
 572 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 573 {
 574         /*
 575          * If we're not on the boundary we know we've modified the page and we
 576          * need to crc the page.
 577          */
 578         if (io_ctl->cur != io_ctl->orig)
 579                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 580         else
 581                 io_ctl_unmap_page(io_ctl);
 582
 583         while (io_ctl->index < io_ctl->num_pages) {
 584                 io_ctl_map_page(io_ctl, 1);
 585                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 586         }
 587 }
 588
 589 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 590                             struct btrfs_free_space *entry, u8 *type)
 591 {
 592         struct btrfs_free_space_entry *e;
 593         int ret;
 594
 595         if (!io_ctl->cur) {
 596                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 597                 if (ret)
 598                         return ret;
 599         }
 600
 601         e = io_ctl->cur;
 602         entry->offset = le64_to_cpu(e->offset);
 603         entry->bytes = le64_to_cpu(e->bytes);
 604         *type = e->type;
 605         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 606         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 607
 608         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 609                 return 0;
 610
 611         io_ctl_unmap_page(io_ctl);
 612
 613         return 0;
 614 }
 615
 616 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 617                               struct btrfs_free_space *entry)
 618 {
 619         int ret;
 620
 621         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 622         if (ret)
 623                 return ret;
 624
 625         copy_page(entry->bitmap, io_ctl->cur);
 626         io_ctl_unmap_page(io_ctl);
 627
 628         return 0;
 629 }
 630
 631 /*
 632  * Since we attach pinned extents after the fact we can have contiguous sections
 633  * of free space that are split up in entries.  This poses a problem with the
 634  * tree logging stuff since it could have allocated across what appears to be 2
 635  * entries since we would have merged the entries when adding the pinned extents
 636  * back to the free space cache.  So run through the space cache that we just
 637  * loaded and merge contiguous entries.  This will make the log replay stuff not
 638  * blow up and it will make for nicer allocator behavior.
 639  */
 640 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 641 {
 642         struct btrfs_free_space *e, *prev = NULL;
 643         struct rb_node *n;
 644
 645 again:
 646         spin_lock(&ctl->tree_lock);
 647         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 648                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 649                 if (!prev)
 650                         goto next;
 651                 if (e->bitmap || prev->bitmap)
 652                         goto next;
 653                 if (prev->offset + prev->bytes == e->offset) {
 654                         unlink_free_space(ctl, prev);
 655                         unlink_free_space(ctl, e);
 656                         prev->bytes += e->bytes;
 657                         kmem_cache_free(btrfs_free_space_cachep, e);
 658                         link_free_space(ctl, prev);
 659                         prev = NULL;
 660                         spin_unlock(&ctl->tree_lock);
 661                         goto again;
 662                 }
 663 next:
 664                 prev = e;
 665         }
 666         spin_unlock(&ctl->tree_lock);
 667 }
 668
 669 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 670                                    struct btrfs_free_space_ctl *ctl,
 671                                    struct btrfs_path *path, u64 offset)
 672 {
 673         struct btrfs_fs_info *fs_info = root->fs_info;
 674         struct btrfs_free_space_header *header;
 675         struct extent_buffer *leaf;
 676         struct btrfs_io_ctl io_ctl;
 677         struct btrfs_key key;
 678         struct btrfs_free_space *e, *n;
 679         LIST_HEAD(bitmaps);
 680         u64 num_entries;
 681         u64 num_bitmaps;
 682         u64 generation;
 683         u8 type;
 684         int ret = 0;
 685
 686         /* Nothing in the space cache, goodbye */
 687         if (!i_size_read(inode))
 688                 return 0;
 689
 690         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 691         key.offset = offset;
 692         key.type = 0;
 693
 694         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 695         if (ret < 0)
 696                 return 0;
 697         else if (ret > 0) {
 698                 btrfs_release_path(path);
 699                 return 0;
 700         }
 701
 702         ret = -1;
 703
 704         leaf = path->nodes[0];
 705         header = btrfs_item_ptr(leaf, path->slots[0],
 706                                 struct btrfs_free_space_header);
 707         num_entries = btrfs_free_space_entries(leaf, header);
 708         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 709         generation = btrfs_free_space_generation(leaf, header);
 710         btrfs_release_path(path);
 711
 712         if (!BTRFS_I(inode)->generation) {
 713                 btrfs_info(fs_info,
 714                            "the free space cache file (%llu) is invalid, skip it",
 715                            offset);
 716                 return 0;
 717         }
 718
 719         if (BTRFS_I(inode)->generation != generation) {
 720                 btrfs_err(fs_info,
 721                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 722                           BTRFS_I(inode)->generation, generation);
 723                 return 0;
 724         }
 725
 726         if (!num_entries)
 727                 return 0;
 728
 729         ret = io_ctl_init(&io_ctl, inode, 0);
 730         if (ret)
 731                 return ret;
 732
 733         readahead_cache(inode);
 734
 735         ret = io_ctl_prepare_pages(&io_ctl, true);
 736         if (ret)
 737                 goto out;
 738
 739         ret = io_ctl_check_crc(&io_ctl, 0);
 740         if (ret)
 741                 goto free_cache;
 742
 743         ret = io_ctl_check_generation(&io_ctl, generation);
 744         if (ret)
 745                 goto free_cache;
 746
 747         while (num_entries) {
 748                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 749                                       GFP_NOFS);
 750                 if (!e)
 751                         goto free_cache;
 752
 753                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 754                 if (ret) {
 755                         kmem_cache_free(btrfs_free_space_cachep, e);
 756                         goto free_cache;
 757                 }
 758
 759                 /*
 760                  * Sync discard ensures that the free space cache is always
 761                  * trimmed.  So when reading this in, the state should reflect
 762                  * that.  We also do this for async as a stop gap for lack of
 763                  * persistence.
 764                  */
 765                 if (btrfs_test_opt(fs_info, DISCARD_SYNC) ||
 766                     btrfs_test_opt(fs_info, DISCARD_ASYNC))
 767                         e->trim_state = BTRFS_TRIM_STATE_TRIMMED;
 768
 769                 if (!e->bytes) {
 770                         kmem_cache_free(btrfs_free_space_cachep, e);
 771                         goto free_cache;
 772                 }
 773
 774                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 775                         spin_lock(&ctl->tree_lock);
 776                         ret = link_free_space(ctl, e);
 777                         spin_unlock(&ctl->tree_lock);
 778                         if (ret) {
 779                                 btrfs_err(fs_info,
 780                                         "Duplicate entries in free space cache, dumping");
 781                                 kmem_cache_free(btrfs_free_space_cachep, e);
 782                                 goto free_cache;
 783                         }
 784                 } else {
 785                         ASSERT(num_bitmaps);
 786                         num_bitmaps--;
 787                         e->bitmap = kmem_cache_zalloc(
 788                                         btrfs_free_space_bitmap_cachep, GFP_NOFS);
 789                         if (!e->bitmap) {
 790                                 kmem_cache_free(
 791                                         btrfs_free_space_cachep, e);
 792                                 goto free_cache;
 793                         }
 794                         spin_lock(&ctl->tree_lock);
 795                         ret = link_free_space(ctl, e);
 796                         ctl->total_bitmaps++;
 797                         ctl->op->recalc_thresholds(ctl);
 798                         spin_unlock(&ctl->tree_lock);
 799                         if (ret) {
 800                                 btrfs_err(fs_info,
 801                                         "Duplicate entries in free space cache, dumping");
 802                                 kmem_cache_free(btrfs_free_space_cachep, e);
 803                                 goto free_cache;
 804                         }
 805                         list_add_tail(&e->list, &bitmaps);
 806                 }
 807
 808                 num_entries--;
 809         }
 810
 811         io_ctl_unmap_page(&io_ctl);
 812
 813         /*
 814          * We add the bitmaps at the end of the entries in order that
 815          * the bitmap entries are added to the cache.
 816          */
 817         list_for_each_entry_safe(e, n, &bitmaps, list) {
 818                 list_del_init(&e->list);
 819                 ret = io_ctl_read_bitmap(&io_ctl, e);
 820                 if (ret)
 821                         goto free_cache;
 822                 e->bitmap_extents = count_bitmap_extents(ctl, e);
 823                 if (!btrfs_free_space_trimmed(e)) {
 824                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
 825                                 e->bitmap_extents;
 826                         ctl->discardable_bytes[BTRFS_STAT_CURR] += e->bytes;
 827                 }
 828         }
 829
 830         io_ctl_drop_pages(&io_ctl);
 831         merge_space_tree(ctl);
 832         ret = 1;
 833 out:
 834         btrfs_discard_update_discardable(ctl->private, ctl);
 835         io_ctl_free(&io_ctl);
 836         return ret;
 837 free_cache:
 838         io_ctl_drop_pages(&io_ctl);
 839         __btrfs_remove_free_space_cache(ctl);
 840         goto out;
 841 }
 842
 843 int load_free_space_cache(struct btrfs_block_group *block_group)
 844 {
 845         struct btrfs_fs_info *fs_info = block_group->fs_info;
 846         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 847         struct inode *inode;
 848         struct btrfs_path *path;
 849         int ret = 0;
 850         bool matched;
 851         u64 used = block_group->used;
 852
 853         /*
 854          * If this block group has been marked to be cleared for one reason or
 855          * another then we can't trust the on disk cache, so just return.
 856          */
 857         spin_lock(&block_group->lock);
 858         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 859                 spin_unlock(&block_group->lock);
 860                 return 0;
 861         }
 862         spin_unlock(&block_group->lock);
 863
 864         path = btrfs_alloc_path();
 865         if (!path)
 866                 return 0;
 867         path->search_commit_root = 1;
 868         path->skip_locking = 1;
 869
 870         /*
 871          * We must pass a path with search_commit_root set to btrfs_iget in
 872          * order to avoid a deadlock when allocating extents for the tree root.
 873          *
 874          * When we are COWing an extent buffer from the tree root, when looking
 875          * for a free extent, at extent-tree.c:find_free_extent(), we can find
 876          * block group without its free space cache loaded. When we find one
 877          * we must load its space cache which requires reading its free space
 878          * cache's inode item from the root tree. If this inode item is located
 879          * in the same leaf that we started COWing before, then we end up in
 880          * deadlock on the extent buffer (trying to read lock it when we
 881          * previously write locked it).
 882          *
 883          * It's safe to read the inode item using the commit root because
 884          * block groups, once loaded, stay in memory forever (until they are
 885          * removed) as well as their space caches once loaded. New block groups
 886          * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
 887          * we will never try to read their inode item while the fs is mounted.
 888          */
 889         inode = lookup_free_space_inode(block_group, path);
 890         if (IS_ERR(inode)) {
 891                 btrfs_free_path(path);
 892                 return 0;
 893         }
 894
 895         /* We may have converted the inode and made the cache invalid. */
 896         spin_lock(&block_group->lock);
 897         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 898                 spin_unlock(&block_group->lock);
 899                 btrfs_free_path(path);
 900                 goto out;
 901         }
 902         spin_unlock(&block_group->lock);
 903
 904         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 905                                       path, block_group->start);
 906         btrfs_free_path(path);
 907         if (ret <= 0)
 908                 goto out;
 909
 910         spin_lock(&ctl->tree_lock);
 911         matched = (ctl->free_space == (block_group->length - used -
 912                                        block_group->bytes_super));
 913         spin_unlock(&ctl->tree_lock);
 914
 915         if (!matched) {
 916                 __btrfs_remove_free_space_cache(ctl);
 917                 btrfs_warn(fs_info,
 918                            "block group %llu has wrong amount of free space",
 919                            block_group->start);
 920                 ret = -1;
 921         }
 922 out:
 923         if (ret < 0) {
 924                 /* This cache is bogus, make sure it gets cleared */
 925                 spin_lock(&block_group->lock);
 926                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 927                 spin_unlock(&block_group->lock);
 928                 ret = 0;
 929
 930                 btrfs_warn(fs_info,
 931                            "failed to load free space cache for block group %llu, rebuilding it now",
 932                            block_group->start);
 933         }
 934
 935         iput(inode);
 936         return ret;
 937 }
 938
 939 static noinline_for_stack
 940 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
 941                               struct btrfs_free_space_ctl *ctl,
 942                               struct btrfs_block_group *block_group,
 943                               int *entries, int *bitmaps,
 944                               struct list_head *bitmap_list)
 945 {
 946         int ret;
 947         struct btrfs_free_cluster *cluster = NULL;
 948         struct btrfs_free_cluster *cluster_locked = NULL;
 949         struct rb_node *node = rb_first(&ctl->free_space_offset);
 950         struct btrfs_trim_range *trim_entry;
 951
 952         /* Get the cluster for this block_group if it exists */
 953         if (block_group && !list_empty(&block_group->cluster_list)) {
 954                 cluster = list_entry(block_group->cluster_list.next,
 955                                      struct btrfs_free_cluster,
 956                                      block_group_list);
 957         }
 958
 959         if (!node && cluster) {
 960                 cluster_locked = cluster;
 961                 spin_lock(&cluster_locked->lock);
 962                 node = rb_first(&cluster->root);
 963                 cluster = NULL;
 964         }
 965
 966         /* Write out the extent entries */
 967         while (node) {
 968                 struct btrfs_free_space *e;
 969
 970                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 971                 *entries += 1;
 972
 973                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
 974                                        e->bitmap);
 975                 if (ret)
 976                         goto fail;
 977
 978                 if (e->bitmap) {
 979                         list_add_tail(&e->list, bitmap_list);
 980                         *bitmaps += 1;
 981                 }
 982                 node = rb_next(node);
 983                 if (!node && cluster) {
 984                         node = rb_first(&cluster->root);
 985                         cluster_locked = cluster;
 986                         spin_lock(&cluster_locked->lock);
 987                         cluster = NULL;
 988                 }
 989         }
 990         if (cluster_locked) {
 991                 spin_unlock(&cluster_locked->lock);
 992                 cluster_locked = NULL;
 993         }
 994
 995         /*
 996          * Make sure we don't miss any range that was removed from our rbtree
 997          * because trimming is running. Otherwise after a umount+mount (or crash
 998          * after committing the transaction) we would leak free space and get
 999          * an inconsistent free space cache report from fsck.
1000          */
1001         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1002                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1003                                        trim_entry->bytes, NULL);
1004                 if (ret)
1005                         goto fail;
1006                 *entries += 1;
1007         }
1008
1009         return 0;
1010 fail:
1011         if (cluster_locked)
1012                 spin_unlock(&cluster_locked->lock);
1013         return -ENOSPC;
1014 }
1015
1016 static noinline_for_stack int
1017 update_cache_item(struct btrfs_trans_handle *trans,
1018                   struct btrfs_root *root,
1019                   struct inode *inode,
1020                   struct btrfs_path *path, u64 offset,
1021                   int entries, int bitmaps)
1022 {
1023         struct btrfs_key key;
1024         struct btrfs_free_space_header *header;
1025         struct extent_buffer *leaf;
1026         int ret;
1027
1028         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1029         key.offset = offset;
1030         key.type = 0;
1031
1032         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1033         if (ret < 0) {
1034                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1035                                  EXTENT_DELALLOC, 0, 0, NULL);
1036                 goto fail;
1037         }
1038         leaf = path->nodes[0];
1039         if (ret > 0) {
1040                 struct btrfs_key found_key;
1041                 ASSERT(path->slots[0]);
1042                 path->slots[0]--;
1043                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1044                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1045                     found_key.offset != offset) {
1046                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1047                                          inode->i_size - 1, EXTENT_DELALLOC, 0,
1048                                          0, NULL);
1049                         btrfs_release_path(path);
1050                         goto fail;
1051                 }
1052         }
1053
1054         BTRFS_I(inode)->generation = trans->transid;
1055         header = btrfs_item_ptr(leaf, path->slots[0],
1056                                 struct btrfs_free_space_header);
1057         btrfs_set_free_space_entries(leaf, header, entries);
1058         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1059         btrfs_set_free_space_generation(leaf, header, trans->transid);
1060         btrfs_mark_buffer_dirty(leaf);
1061         btrfs_release_path(path);
1062
1063         return 0;
1064
1065 fail:
1066         return -1;
1067 }
1068
1069 static noinline_for_stack int write_pinned_extent_entries(
1070                             struct btrfs_trans_handle *trans,
1071                             struct btrfs_block_group *block_group,
1072                             struct btrfs_io_ctl *io_ctl,
1073                             int *entries)
1074 {
1075         u64 start, extent_start, extent_end, len;
1076         struct extent_io_tree *unpin = NULL;
1077         int ret;
1078
1079         if (!block_group)
1080                 return 0;
1081
1082         /*
1083          * We want to add any pinned extents to our free space cache
1084          * so we don't leak the space
1085          *
1086          * We shouldn't have switched the pinned extents yet so this is the
1087          * right one
1088          */
1089         unpin = &trans->transaction->pinned_extents;
1090
1091         start = block_group->start;
1092
1093         while (start < block_group->start + block_group->length) {
1094                 ret = find_first_extent_bit(unpin, start,
1095                                             &extent_start, &extent_end,
1096                                             EXTENT_DIRTY, NULL);
1097                 if (ret)
1098                         return 0;
1099
1100                 /* This pinned extent is out of our range */
1101                 if (extent_start >= block_group->start + block_group->length)
1102                         return 0;
1103
1104                 extent_start = max(extent_start, start);
1105                 extent_end = min(block_group->start + block_group->length,
1106                                  extent_end + 1);
1107                 len = extent_end - extent_start;
1108
1109                 *entries += 1;
1110                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1111                 if (ret)
1112                         return -ENOSPC;
1113
1114                 start = extent_end;
1115         }
1116
1117         return 0;
1118 }
1119
1120 static noinline_for_stack int
1121 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1122 {
1123         struct btrfs_free_space *entry, *next;
1124         int ret;
1125
1126         /* Write out the bitmaps */
1127         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1128                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1129                 if (ret)
1130                         return -ENOSPC;
1131                 list_del_init(&entry->list);
1132         }
1133
1134         return 0;
1135 }
1136
1137 static int flush_dirty_cache(struct inode *inode)
1138 {
1139         int ret;
1140
1141         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1142         if (ret)
1143                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1144                                  EXTENT_DELALLOC, 0, 0, NULL);
1145
1146         return ret;
1147 }
1148
1149 static void noinline_for_stack
1150 cleanup_bitmap_list(struct list_head *bitmap_list)
1151 {
1152         struct btrfs_free_space *entry, *next;
1153
1154         list_for_each_entry_safe(entry, next, bitmap_list, list)
1155                 list_del_init(&entry->list);
1156 }
1157
1158 static void noinline_for_stack
1159 cleanup_write_cache_enospc(struct inode *inode,
1160                            struct btrfs_io_ctl *io_ctl,
1161                            struct extent_state **cached_state)
1162 {
1163         io_ctl_drop_pages(io_ctl);
1164         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1165                              i_size_read(inode) - 1, cached_state);
1166 }
1167
1168 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1169                                  struct btrfs_trans_handle *trans,
1170                                  struct btrfs_block_group *block_group,
1171                                  struct btrfs_io_ctl *io_ctl,
1172                                  struct btrfs_path *path, u64 offset)
1173 {
1174         int ret;
1175         struct inode *inode = io_ctl->inode;
1176
1177         if (!inode)
1178                 return 0;
1179
1180         /* Flush the dirty pages in the cache file. */
1181         ret = flush_dirty_cache(inode);
1182         if (ret)
1183                 goto out;
1184
1185         /* Update the cache item to tell everyone this cache file is valid. */
1186         ret = update_cache_item(trans, root, inode, path, offset,
1187                                 io_ctl->entries, io_ctl->bitmaps);
1188 out:
1189         io_ctl_free(io_ctl);
1190         if (ret) {
1191                 invalidate_inode_pages2(inode->i_mapping);
1192                 BTRFS_I(inode)->generation = 0;
1193                 if (block_group)
1194                         btrfs_debug(root->fs_info,
1195           "failed to write free space cache for block group %llu error %d",
1196                                   block_group->start, ret);
1197         }
1198         btrfs_update_inode(trans, root, inode);
1199
1200         if (block_group) {
1201                 /* the dirty list is protected by the dirty_bgs_lock */
1202                 spin_lock(&trans->transaction->dirty_bgs_lock);
1203
1204                 /* the disk_cache_state is protected by the block group lock */
1205                 spin_lock(&block_group->lock);
1206
1207                 /*
1208                  * only mark this as written if we didn't get put back on
1209                  * the dirty list while waiting for IO.   Otherwise our
1210                  * cache state won't be right, and we won't get written again
1211                  */
1212                 if (!ret && list_empty(&block_group->dirty_list))
1213                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1214                 else if (ret)
1215                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1216
1217                 spin_unlock(&block_group->lock);
1218                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1219                 io_ctl->inode = NULL;
1220                 iput(inode);
1221         }
1222
1223         return ret;
1224
1225 }
1226
1227 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
1228                                     struct btrfs_trans_handle *trans,
1229                                     struct btrfs_io_ctl *io_ctl,
1230                                     struct btrfs_path *path)
1231 {
1232         return __btrfs_wait_cache_io(root, trans, NULL, io_ctl, path, 0);
1233 }
1234
1235 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1236                         struct btrfs_block_group *block_group,
1237                         struct btrfs_path *path)
1238 {
1239         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1240                                      block_group, &block_group->io_ctl,
1241                                      path, block_group->start);
1242 }
1243
1244 /**
1245  * __btrfs_write_out_cache - write out cached info to an inode
1246  * @root - the root the inode belongs to
1247  * @ctl - the free space cache we are going to write out
1248  * @block_group - the block_group for this cache if it belongs to a block_group
1249  * @trans - the trans handle
1250  *
1251  * This function writes out a free space cache struct to disk for quick recovery
1252  * on mount.  This will return 0 if it was successful in writing the cache out,
1253  * or an errno if it was not.
1254  */
1255 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1256                                    struct btrfs_free_space_ctl *ctl,
1257                                    struct btrfs_block_group *block_group,
1258                                    struct btrfs_io_ctl *io_ctl,
1259                                    struct btrfs_trans_handle *trans)
1260 {
1261         struct extent_state *cached_state = NULL;
1262         LIST_HEAD(bitmap_list);
1263         int entries = 0;
1264         int bitmaps = 0;
1265         int ret;
1266         int must_iput = 0;
1267
1268         if (!i_size_read(inode))
1269                 return -EIO;
1270
1271         WARN_ON(io_ctl->pages);
1272         ret = io_ctl_init(io_ctl, inode, 1);
1273         if (ret)
1274                 return ret;
1275
1276         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1277                 down_write(&block_group->data_rwsem);
1278                 spin_lock(&block_group->lock);
1279                 if (block_group->delalloc_bytes) {
1280                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1281                         spin_unlock(&block_group->lock);
1282                         up_write(&block_group->data_rwsem);
1283                         BTRFS_I(inode)->generation = 0;
1284                         ret = 0;
1285                         must_iput = 1;
1286                         goto out;
1287                 }
1288                 spin_unlock(&block_group->lock);
1289         }
1290
1291         /* Lock all pages first so we can lock the extent safely. */
1292         ret = io_ctl_prepare_pages(io_ctl, false);
1293         if (ret)
1294                 goto out_unlock;
1295
1296         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1297                          &cached_state);
1298
1299         io_ctl_set_generation(io_ctl, trans->transid);
1300
1301         mutex_lock(&ctl->cache_writeout_mutex);
1302         /* Write out the extent entries in the free space cache */
1303         spin_lock(&ctl->tree_lock);
1304         ret = write_cache_extent_entries(io_ctl, ctl,
1305                                          block_group, &entries, &bitmaps,
1306                                          &bitmap_list);
1307         if (ret)
1308                 goto out_nospc_locked;
1309
1310         /*
1311          * Some spaces that are freed in the current transaction are pinned,
1312          * they will be added into free space cache after the transaction is
1313          * committed, we shouldn't lose them.
1314          *
1315          * If this changes while we are working we'll get added back to
1316          * the dirty list and redo it.  No locking needed
1317          */
1318         ret = write_pinned_extent_entries(trans, block_group, io_ctl, &entries);
1319         if (ret)
1320                 goto out_nospc_locked;
1321
1322         /*
1323          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1324          * locked while doing it because a concurrent trim can be manipulating
1325          * or freeing the bitmap.
1326          */
1327         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1328         spin_unlock(&ctl->tree_lock);
1329         mutex_unlock(&ctl->cache_writeout_mutex);
1330         if (ret)
1331                 goto out_nospc;
1332
1333         /* Zero out the rest of the pages just to make sure */
1334         io_ctl_zero_remaining_pages(io_ctl);
1335
1336         /* Everything is written out, now we dirty the pages in the file. */
1337         ret = btrfs_dirty_pages(inode, io_ctl->pages, io_ctl->num_pages, 0,
1338                                 i_size_read(inode), &cached_state);
1339         if (ret)
1340                 goto out_nospc;
1341
1342         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1343                 up_write(&block_group->data_rwsem);
1344         /*
1345          * Release the pages and unlock the extent, we will flush
1346          * them out later
1347          */
1348         io_ctl_drop_pages(io_ctl);
1349
1350         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1351                              i_size_read(inode) - 1, &cached_state);
1352
1353         /*
1354          * at this point the pages are under IO and we're happy,
1355          * The caller is responsible for waiting on them and updating the
1356          * the cache and the inode
1357          */
1358         io_ctl->entries = entries;
1359         io_ctl->bitmaps = bitmaps;
1360
1361         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1362         if (ret)
1363                 goto out;
1364
1365         return 0;
1366
1367 out_nospc_locked:
1368         cleanup_bitmap_list(&bitmap_list);
1369         spin_unlock(&ctl->tree_lock);
1370         mutex_unlock(&ctl->cache_writeout_mutex);
1371
1372 out_nospc:
1373         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1374
1375 out_unlock:
1376         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1377                 up_write(&block_group->data_rwsem);
1378
1379 out:
1380         io_ctl->inode = NULL;
1381         io_ctl_free(io_ctl);
1382         if (ret) {
1383                 invalidate_inode_pages2(inode->i_mapping);
1384                 BTRFS_I(inode)->generation = 0;
1385         }
1386         btrfs_update_inode(trans, root, inode);
1387         if (must_iput)
1388                 iput(inode);
1389         return ret;
1390 }
1391
1392 int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1393                           struct btrfs_block_group *block_group,
1394                           struct btrfs_path *path)
1395 {
1396         struct btrfs_fs_info *fs_info = trans->fs_info;
1397         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1398         struct inode *inode;
1399         int ret = 0;
1400
1401         spin_lock(&block_group->lock);
1402         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1403                 spin_unlock(&block_group->lock);
1404                 return 0;
1405         }
1406         spin_unlock(&block_group->lock);
1407
1408         inode = lookup_free_space_inode(block_group, path);
1409         if (IS_ERR(inode))
1410                 return 0;
1411
1412         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1413                                 block_group, &block_group->io_ctl, trans);
1414         if (ret) {
1415                 btrfs_debug(fs_info,
1416           "failed to write free space cache for block group %llu error %d",
1417                           block_group->start, ret);
1418                 spin_lock(&block_group->lock);
1419                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1420                 spin_unlock(&block_group->lock);
1421
1422                 block_group->io_ctl.inode = NULL;
1423                 iput(inode);
1424         }
1425
1426         /*
1427          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1428          * to wait for IO and put the inode
1429          */
1430
1431         return ret;
1432 }
1433
1434 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1435                                           u64 offset)
1436 {
1437         ASSERT(offset >= bitmap_start);
1438         offset -= bitmap_start;
1439         return (unsigned long)(div_u64(offset, unit));
1440 }
1441
1442 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1443 {
1444         return (unsigned long)(div_u64(bytes, unit));
1445 }
1446
1447 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1448                                    u64 offset)
1449 {
1450         u64 bitmap_start;
1451         u64 bytes_per_bitmap;
1452
1453         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1454         bitmap_start = offset - ctl->start;
1455         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1456         bitmap_start *= bytes_per_bitmap;
1457         bitmap_start += ctl->start;
1458
1459         return bitmap_start;
1460 }
1461
1462 static int tree_insert_offset(struct rb_root *root, u64 offset,
1463                               struct rb_node *node, int bitmap)
1464 {
1465         struct rb_node **p = &root->rb_node;
1466         struct rb_node *parent = NULL;
1467         struct btrfs_free_space *info;
1468
1469         while (*p) {
1470                 parent = *p;
1471                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1472
1473                 if (offset < info->offset) {
1474                         p = &(*p)->rb_left;
1475                 } else if (offset > info->offset) {
1476                         p = &(*p)->rb_right;
1477                 } else {
1478                         /*
1479                          * we could have a bitmap entry and an extent entry
1480                          * share the same offset.  If this is the case, we want
1481                          * the extent entry to always be found first if we do a
1482                          * linear search through the tree, since we want to have
1483                          * the quickest allocation time, and allocating from an
1484                          * extent is faster than allocating from a bitmap.  So
1485                          * if we're inserting a bitmap and we find an entry at
1486                          * this offset, we want to go right, or after this entry
1487                          * logically.  If we are inserting an extent and we've
1488                          * found a bitmap, we want to go left, or before
1489                          * logically.
1490                          */
1491                         if (bitmap) {
1492                                 if (info->bitmap) {
1493                                         WARN_ON_ONCE(1);
1494                                         return -EEXIST;
1495                                 }
1496                                 p = &(*p)->rb_right;
1497                         } else {
1498                                 if (!info->bitmap) {
1499                                         WARN_ON_ONCE(1);
1500                                         return -EEXIST;
1501                                 }
1502                                 p = &(*p)->rb_left;
1503                         }
1504                 }
1505         }
1506
1507         rb_link_node(node, parent, p);
1508         rb_insert_color(node, root);
1509
1510         return 0;
1511 }
1512
1513 /*
1514  * searches the tree for the given offset.
1515  *
1516  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1517  * want a section that has at least bytes size and comes at or after the given
1518  * offset.
1519  */
1520 static struct btrfs_free_space *
1521 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1522                    u64 offset, int bitmap_only, int fuzzy)
1523 {
1524         struct rb_node *n = ctl->free_space_offset.rb_node;
1525         struct btrfs_free_space *entry, *prev = NULL;
1526
1527         /* find entry that is closest to the 'offset' */
1528         while (1) {
1529                 if (!n) {
1530                         entry = NULL;
1531                         break;
1532                 }
1533
1534                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1535                 prev = entry;
1536
1537                 if (offset < entry->offset)
1538                         n = n->rb_left;
1539                 else if (offset > entry->offset)
1540                         n = n->rb_right;
1541                 else
1542                         break;
1543         }
1544
1545         if (bitmap_only) {
1546                 if (!entry)
1547                         return NULL;
1548                 if (entry->bitmap)
1549                         return entry;
1550
1551                 /*
1552                  * bitmap entry and extent entry may share same offset,
1553                  * in that case, bitmap entry comes after extent entry.
1554                  */
1555                 n = rb_next(n);
1556                 if (!n)
1557                         return NULL;
1558                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1559                 if (entry->offset != offset)
1560                         return NULL;
1561
1562                 WARN_ON(!entry->bitmap);
1563                 return entry;
1564         } else if (entry) {
1565                 if (entry->bitmap) {
1566                         /*
1567                          * if previous extent entry covers the offset,
1568                          * we should return it instead of the bitmap entry
1569                          */
1570                         n = rb_prev(&entry->offset_index);
1571                         if (n) {
1572                                 prev = rb_entry(n, struct btrfs_free_space,
1573                                                 offset_index);
1574                                 if (!prev->bitmap &&
1575                                     prev->offset + prev->bytes > offset)
1576                                         entry = prev;
1577                         }
1578                 }
1579                 return entry;
1580         }
1581
1582         if (!prev)
1583                 return NULL;
1584
1585         /* find last entry before the 'offset' */
1586         entry = prev;
1587         if (entry->offset > offset) {
1588                 n = rb_prev(&entry->offset_index);
1589                 if (n) {
1590                         entry = rb_entry(n, struct btrfs_free_space,
1591                                         offset_index);
1592                         ASSERT(entry->offset <= offset);
1593                 } else {
1594                         if (fuzzy)
1595                                 return entry;
1596                         else
1597                                 return NULL;
1598                 }
1599         }
1600
1601         if (entry->bitmap) {
1602                 n = rb_prev(&entry->offset_index);
1603                 if (n) {
1604                         prev = rb_entry(n, struct btrfs_free_space,
1605                                         offset_index);
1606                         if (!prev->bitmap &&
1607                             prev->offset + prev->bytes > offset)
1608                                 return prev;
1609                 }
1610                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1611                         return entry;
1612         } else if (entry->offset + entry->bytes > offset)
1613                 return entry;
1614
1615         if (!fuzzy)
1616                 return NULL;
1617
1618         while (1) {
1619                 if (entry->bitmap) {
1620                         if (entry->offset + BITS_PER_BITMAP *
1621                             ctl->unit > offset)
1622                                 break;
1623                 } else {
1624                         if (entry->offset + entry->bytes > offset)
1625                                 break;
1626                 }
1627
1628                 n = rb_next(&entry->offset_index);
1629                 if (!n)
1630                         return NULL;
1631                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1632         }
1633         return entry;
1634 }
1635
1636 static inline void
1637 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1638                     struct btrfs_free_space *info)
1639 {
1640         rb_erase(&info->offset_index, &ctl->free_space_offset);
1641         ctl->free_extents--;
1642
1643         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1644                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1645                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1646         }
1647 }
1648
1649 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1650                               struct btrfs_free_space *info)
1651 {
1652         __unlink_free_space(ctl, info);
1653         ctl->free_space -= info->bytes;
1654 }
1655
1656 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1657                            struct btrfs_free_space *info)
1658 {
1659         int ret = 0;
1660
1661         ASSERT(info->bytes || info->bitmap);
1662         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1663                                  &info->offset_index, (info->bitmap != NULL));
1664         if (ret)
1665                 return ret;
1666
1667         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1668                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1669                 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1670         }
1671
1672         ctl->free_space += info->bytes;
1673         ctl->free_extents++;
1674         return ret;
1675 }
1676
1677 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1678 {
1679         struct btrfs_block_group *block_group = ctl->private;
1680         u64 max_bytes;
1681         u64 bitmap_bytes;
1682         u64 extent_bytes;
1683         u64 size = block_group->length;
1684         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1685         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1686
1687         max_bitmaps = max_t(u64, max_bitmaps, 1);
1688
1689         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1690
1691         /*
1692          * We are trying to keep the total amount of memory used per 1GiB of
1693          * space to be MAX_CACHE_BYTES_PER_GIG.  However, with a reclamation
1694          * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
1695          * bitmaps, we may end up using more memory than this.
1696          */
1697         if (size < SZ_1G)
1698                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1699         else
1700                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1701
1702         bitmap_bytes = ctl->total_bitmaps * ctl->unit;
1703
1704         /*
1705          * we want the extent entry threshold to always be at most 1/2 the max
1706          * bytes we can have, or whatever is less than that.
1707          */
1708         extent_bytes = max_bytes - bitmap_bytes;
1709         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1710
1711         ctl->extents_thresh =
1712                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1713 }
1714
1715 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1716                                        struct btrfs_free_space *info,
1717                                        u64 offset, u64 bytes)
1718 {
1719         unsigned long start, count, end;
1720         int extent_delta = -1;
1721
1722         start = offset_to_bit(info->offset, ctl->unit, offset);
1723         count = bytes_to_bits(bytes, ctl->unit);
1724         end = start + count;
1725         ASSERT(end <= BITS_PER_BITMAP);
1726
1727         bitmap_clear(info->bitmap, start, count);
1728
1729         info->bytes -= bytes;
1730         if (info->max_extent_size > ctl->unit)
1731                 info->max_extent_size = 0;
1732
1733         if (start && test_bit(start - 1, info->bitmap))
1734                 extent_delta++;
1735
1736         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1737                 extent_delta++;
1738
1739         info->bitmap_extents += extent_delta;
1740         if (!btrfs_free_space_trimmed(info)) {
1741                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1742                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1743         }
1744 }
1745
1746 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1747                               struct btrfs_free_space *info, u64 offset,
1748                               u64 bytes)
1749 {
1750         __bitmap_clear_bits(ctl, info, offset, bytes);
1751         ctl->free_space -= bytes;
1752 }
1753
1754 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1755                             struct btrfs_free_space *info, u64 offset,
1756                             u64 bytes)
1757 {
1758         unsigned long start, count, end;
1759         int extent_delta = 1;
1760
1761         start = offset_to_bit(info->offset, ctl->unit, offset);
1762         count = bytes_to_bits(bytes, ctl->unit);
1763         end = start + count;
1764         ASSERT(end <= BITS_PER_BITMAP);
1765
1766         bitmap_set(info->bitmap, start, count);
1767
1768         info->bytes += bytes;
1769         ctl->free_space += bytes;
1770
1771         if (start && test_bit(start - 1, info->bitmap))
1772                 extent_delta--;
1773
1774         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1775                 extent_delta--;
1776
1777         info->bitmap_extents += extent_delta;
1778         if (!btrfs_free_space_trimmed(info)) {
1779                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1780                 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1781         }
1782 }
1783
1784 /*
1785  * If we can not find suitable extent, we will use bytes to record
1786  * the size of the max extent.
1787  */
1788 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1789                          struct btrfs_free_space *bitmap_info, u64 *offset,
1790                          u64 *bytes, bool for_alloc)
1791 {
1792         unsigned long found_bits = 0;
1793         unsigned long max_bits = 0;
1794         unsigned long bits, i;
1795         unsigned long next_zero;
1796         unsigned long extent_bits;
1797
1798         /*
1799          * Skip searching the bitmap if we don't have a contiguous section that
1800          * is large enough for this allocation.
1801          */
1802         if (for_alloc &&
1803             bitmap_info->max_extent_size &&
1804             bitmap_info->max_extent_size < *bytes) {
1805                 *bytes = bitmap_info->max_extent_size;
1806                 return -1;
1807         }
1808
1809         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1810                           max_t(u64, *offset, bitmap_info->offset));
1811         bits = bytes_to_bits(*bytes, ctl->unit);
1812
1813         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1814                 if (for_alloc && bits == 1) {
1815                         found_bits = 1;
1816                         break;
1817                 }
1818                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1819                                                BITS_PER_BITMAP, i);
1820                 extent_bits = next_zero - i;
1821                 if (extent_bits >= bits) {
1822                         found_bits = extent_bits;
1823                         break;
1824                 } else if (extent_bits > max_bits) {
1825                         max_bits = extent_bits;
1826                 }
1827                 i = next_zero;
1828         }
1829
1830         if (found_bits) {
1831                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1832                 *bytes = (u64)(found_bits) * ctl->unit;
1833                 return 0;
1834         }
1835
1836         *bytes = (u64)(max_bits) * ctl->unit;
1837         bitmap_info->max_extent_size = *bytes;
1838         return -1;
1839 }
1840
1841 static inline u64 get_max_extent_size(struct btrfs_free_space *entry)
1842 {
1843         if (entry->bitmap)
1844                 return entry->max_extent_size;
1845         return entry->bytes;
1846 }
1847
1848 /* Cache the size of the max extent in bytes */
1849 static struct btrfs_free_space *
1850 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1851                 unsigned long align, u64 *max_extent_size)
1852 {
1853         struct btrfs_free_space *entry;
1854         struct rb_node *node;
1855         u64 tmp;
1856         u64 align_off;
1857         int ret;
1858
1859         if (!ctl->free_space_offset.rb_node)
1860                 goto out;
1861
1862         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1863         if (!entry)
1864                 goto out;
1865
1866         for (node = &entry->offset_index; node; node = rb_next(node)) {
1867                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1868                 if (entry->bytes < *bytes) {
1869                         *max_extent_size = max(get_max_extent_size(entry),
1870                                                *max_extent_size);
1871                         continue;
1872                 }
1873
1874                 /* make sure the space returned is big enough
1875                  * to match our requested alignment
1876                  */
1877                 if (*bytes >= align) {
1878                         tmp = entry->offset - ctl->start + align - 1;
1879                         tmp = div64_u64(tmp, align);
1880                         tmp = tmp * align + ctl->start;
1881                         align_off = tmp - entry->offset;
1882                 } else {
1883                         align_off = 0;
1884                         tmp = entry->offset;
1885                 }
1886
1887                 if (entry->bytes < *bytes + align_off) {
1888                         *max_extent_size = max(get_max_extent_size(entry),
1889                                                *max_extent_size);
1890                         continue;
1891                 }
1892
1893                 if (entry->bitmap) {
1894                         u64 size = *bytes;
1895
1896                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1897                         if (!ret) {
1898                                 *offset = tmp;
1899                                 *bytes = size;
1900                                 return entry;
1901                         } else {
1902                                 *max_extent_size =
1903                                         max(get_max_extent_size(entry),
1904                                             *max_extent_size);
1905                         }
1906                         continue;
1907                 }
1908
1909                 *offset = tmp;
1910                 *bytes = entry->bytes - align_off;
1911                 return entry;
1912         }
1913 out:
1914         return NULL;
1915 }
1916
1917 static int count_bitmap_extents(struct btrfs_free_space_ctl *ctl,
1918                                 struct btrfs_free_space *bitmap_info)
1919 {
1920         struct btrfs_block_group *block_group = ctl->private;
1921         u64 bytes = bitmap_info->bytes;
1922         unsigned int rs, re;
1923         int count = 0;
1924
1925         if (!block_group || !bytes)
1926                 return count;
1927
1928         bitmap_for_each_set_region(bitmap_info->bitmap, rs, re, 0,
1929                                    BITS_PER_BITMAP) {
1930                 bytes -= (rs - re) * ctl->unit;
1931                 count++;
1932
1933                 if (!bytes)
1934                         break;
1935         }
1936
1937         return count;
1938 }
1939
1940 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1941                            struct btrfs_free_space *info, u64 offset)
1942 {
1943         info->offset = offset_to_bitmap(ctl, offset);
1944         info->bytes = 0;
1945         info->bitmap_extents = 0;
1946         INIT_LIST_HEAD(&info->list);
1947         link_free_space(ctl, info);
1948         ctl->total_bitmaps++;
1949
1950         ctl->op->recalc_thresholds(ctl);
1951 }
1952
1953 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1954                         struct btrfs_free_space *bitmap_info)
1955 {
1956         /*
1957          * Normally when this is called, the bitmap is completely empty. However,
1958          * if we are blowing up the free space cache for one reason or another
1959          * via __btrfs_remove_free_space_cache(), then it may not be freed and
1960          * we may leave stats on the table.
1961          */
1962         if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
1963                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
1964                         bitmap_info->bitmap_extents;
1965                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
1966
1967         }
1968         unlink_free_space(ctl, bitmap_info);
1969         kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
1970         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1971         ctl->total_bitmaps--;
1972         ctl->op->recalc_thresholds(ctl);
1973 }
1974
1975 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1976                               struct btrfs_free_space *bitmap_info,
1977                               u64 *offset, u64 *bytes)
1978 {
1979         u64 end;
1980         u64 search_start, search_bytes;
1981         int ret;
1982
1983 again:
1984         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1985
1986         /*
1987          * We need to search for bits in this bitmap.  We could only cover some
1988          * of the extent in this bitmap thanks to how we add space, so we need
1989          * to search for as much as it as we can and clear that amount, and then
1990          * go searching for the next bit.
1991          */
1992         search_start = *offset;
1993         search_bytes = ctl->unit;
1994         search_bytes = min(search_bytes, end - search_start + 1);
1995         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1996                             false);
1997         if (ret < 0 || search_start != *offset)
1998                 return -EINVAL;
1999
2000         /* We may have found more bits than what we need */
2001         search_bytes = min(search_bytes, *bytes);
2002
2003         /* Cannot clear past the end of the bitmap */
2004         search_bytes = min(search_bytes, end - search_start + 1);
2005
2006         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
2007         *offset += search_bytes;
2008         *bytes -= search_bytes;
2009
2010         if (*bytes) {
2011                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2012                 if (!bitmap_info->bytes)
2013                         free_bitmap(ctl, bitmap_info);
2014
2015                 /*
2016                  * no entry after this bitmap, but we still have bytes to
2017                  * remove, so something has gone wrong.
2018                  */
2019                 if (!next)
2020                         return -EINVAL;
2021
2022                 bitmap_info = rb_entry(next, struct btrfs_free_space,
2023                                        offset_index);
2024
2025                 /*
2026                  * if the next entry isn't a bitmap we need to return to let the
2027                  * extent stuff do its work.
2028                  */
2029                 if (!bitmap_info->bitmap)
2030                         return -EAGAIN;
2031
2032                 /*
2033                  * Ok the next item is a bitmap, but it may not actually hold
2034                  * the information for the rest of this free space stuff, so
2035                  * look for it, and if we don't find it return so we can try
2036                  * everything over again.
2037                  */
2038                 search_start = *offset;
2039                 search_bytes = ctl->unit;
2040                 ret = search_bitmap(ctl, bitmap_info, &search_start,
2041                                     &search_bytes, false);
2042                 if (ret < 0 || search_start != *offset)
2043                         return -EAGAIN;
2044
2045                 goto again;
2046         } else if (!bitmap_info->bytes)
2047                 free_bitmap(ctl, bitmap_info);
2048
2049         return 0;
2050 }
2051
2052 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2053                                struct btrfs_free_space *info, u64 offset,
2054                                u64 bytes, enum btrfs_trim_state trim_state)
2055 {
2056         u64 bytes_to_set = 0;
2057         u64 end;
2058
2059         /*
2060          * This is a tradeoff to make bitmap trim state minimal.  We mark the
2061          * whole bitmap untrimmed if at any point we add untrimmed regions.
2062          */
2063         if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2064                 if (btrfs_free_space_trimmed(info)) {
2065                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
2066                                 info->bitmap_extents;
2067                         ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2068                 }
2069                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2070         }
2071
2072         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2073
2074         bytes_to_set = min(end - offset, bytes);
2075
2076         bitmap_set_bits(ctl, info, offset, bytes_to_set);
2077
2078         /*
2079          * We set some bytes, we have no idea what the max extent size is
2080          * anymore.
2081          */
2082         info->max_extent_size = 0;
2083
2084         return bytes_to_set;
2085
2086 }
2087
2088 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2089                       struct btrfs_free_space *info)
2090 {
2091         struct btrfs_block_group *block_group = ctl->private;
2092         struct btrfs_fs_info *fs_info = block_group->fs_info;
2093         bool forced = false;
2094
2095 #ifdef CONFIG_BTRFS_DEBUG
2096         if (btrfs_should_fragment_free_space(block_group))
2097                 forced = true;
2098 #endif
2099
2100         /* This is a way to reclaim large regions from the bitmaps. */
2101         if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2102                 return false;
2103
2104         /*
2105          * If we are below the extents threshold then we can add this as an
2106          * extent, and don't have to deal with the bitmap
2107          */
2108         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2109                 /*
2110                  * If this block group has some small extents we don't want to
2111                  * use up all of our free slots in the cache with them, we want
2112                  * to reserve them to larger extents, however if we have plenty
2113                  * of cache left then go ahead an dadd them, no sense in adding
2114                  * the overhead of a bitmap if we don't have to.
2115                  */
2116                 if (info->bytes <= fs_info->sectorsize * 8) {
2117                         if (ctl->free_extents * 3 <= ctl->extents_thresh)
2118                                 return false;
2119                 } else {
2120                         return false;
2121                 }
2122         }
2123
2124         /*
2125          * The original block groups from mkfs can be really small, like 8
2126          * megabytes, so don't bother with a bitmap for those entries.  However
2127          * some block groups can be smaller than what a bitmap would cover but
2128          * are still large enough that they could overflow the 32k memory limit,
2129          * so allow those block groups to still be allowed to have a bitmap
2130          * entry.
2131          */
2132         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2133                 return false;
2134
2135         return true;
2136 }
2137
2138 static const struct btrfs_free_space_op free_space_op = {
2139         .recalc_thresholds      = recalculate_thresholds,
2140         .use_bitmap             = use_bitmap,
2141 };
2142
2143 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2144                               struct btrfs_free_space *info)
2145 {
2146         struct btrfs_free_space *bitmap_info;
2147         struct btrfs_block_group *block_group = NULL;
2148         int added = 0;
2149         u64 bytes, offset, bytes_added;
2150         enum btrfs_trim_state trim_state;
2151         int ret;
2152
2153         bytes = info->bytes;
2154         offset = info->offset;
2155         trim_state = info->trim_state;
2156
2157         if (!ctl->op->use_bitmap(ctl, info))
2158                 return 0;
2159
2160         if (ctl->op == &free_space_op)
2161                 block_group = ctl->private;
2162 again:
2163         /*
2164          * Since we link bitmaps right into the cluster we need to see if we
2165          * have a cluster here, and if so and it has our bitmap we need to add
2166          * the free space to that bitmap.
2167          */
2168         if (block_group && !list_empty(&block_group->cluster_list)) {
2169                 struct btrfs_free_cluster *cluster;
2170                 struct rb_node *node;
2171                 struct btrfs_free_space *entry;
2172
2173                 cluster = list_entry(block_group->cluster_list.next,
2174                                      struct btrfs_free_cluster,
2175                                      block_group_list);
2176                 spin_lock(&cluster->lock);
2177                 node = rb_first(&cluster->root);
2178                 if (!node) {
2179                         spin_unlock(&cluster->lock);
2180                         goto no_cluster_bitmap;
2181                 }
2182
2183                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2184                 if (!entry->bitmap) {
2185                         spin_unlock(&cluster->lock);
2186                         goto no_cluster_bitmap;
2187                 }
2188
2189                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2190                         bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2191                                                           bytes, trim_state);
2192                         bytes -= bytes_added;
2193                         offset += bytes_added;
2194                 }
2195                 spin_unlock(&cluster->lock);
2196                 if (!bytes) {
2197                         ret = 1;
2198                         goto out;
2199                 }
2200         }
2201
2202 no_cluster_bitmap:
2203         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2204                                          1, 0);
2205         if (!bitmap_info) {
2206                 ASSERT(added == 0);
2207                 goto new_bitmap;
2208         }
2209
2210         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2211                                           trim_state);
2212         bytes -= bytes_added;
2213         offset += bytes_added;
2214         added = 0;
2215
2216         if (!bytes) {
2217                 ret = 1;
2218                 goto out;
2219         } else
2220                 goto again;
2221
2222 new_bitmap:
2223         if (info && info->bitmap) {
2224                 add_new_bitmap(ctl, info, offset);
2225                 added = 1;
2226                 info = NULL;
2227                 goto again;
2228         } else {
2229                 spin_unlock(&ctl->tree_lock);
2230
2231                 /* no pre-allocated info, allocate a new one */
2232                 if (!info) {
2233                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2234                                                  GFP_NOFS);
2235                         if (!info) {
2236                                 spin_lock(&ctl->tree_lock);
2237                                 ret = -ENOMEM;
2238                                 goto out;
2239                         }
2240                 }
2241
2242                 /* allocate the bitmap */
2243                 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2244                                                  GFP_NOFS);
2245                 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2246                 spin_lock(&ctl->tree_lock);
2247                 if (!info->bitmap) {
2248                         ret = -ENOMEM;
2249                         goto out;
2250                 }
2251                 goto again;
2252         }
2253
2254 out:
2255         if (info) {
2256                 if (info->bitmap)
2257                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2258                                         info->bitmap);
2259                 kmem_cache_free(btrfs_free_space_cachep, info);
2260         }
2261
2262         return ret;
2263 }
2264
2265 /*
2266  * Free space merging rules:
2267  *  1) Merge trimmed areas together
2268  *  2) Let untrimmed areas coalesce with trimmed areas
2269  *  3) Always pull neighboring regions from bitmaps
2270  *
2271  * The above rules are for when we merge free space based on btrfs_trim_state.
2272  * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2273  * same reason: to promote larger extent regions which makes life easier for
2274  * find_free_extent().  Rule 2 enables coalescing based on the common path
2275  * being returning free space from btrfs_finish_extent_commit().  So when free
2276  * space is trimmed, it will prevent aggregating trimmed new region and
2277  * untrimmed regions in the rb_tree.  Rule 3 is purely to obtain larger extents
2278  * and provide find_free_extent() with the largest extents possible hoping for
2279  * the reuse path.
2280  */
2281 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2282                           struct btrfs_free_space *info, bool update_stat)
2283 {
2284         struct btrfs_free_space *left_info;
2285         struct btrfs_free_space *right_info;
2286         bool merged = false;
2287         u64 offset = info->offset;
2288         u64 bytes = info->bytes;
2289         const bool is_trimmed = btrfs_free_space_trimmed(info);
2290
2291         /*
2292          * first we want to see if there is free space adjacent to the range we
2293          * are adding, if there is remove that struct and add a new one to
2294          * cover the entire range
2295          */
2296         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2297         if (right_info && rb_prev(&right_info->offset_index))
2298                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2299                                      struct btrfs_free_space, offset_index);
2300         else
2301                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2302
2303         /* See try_merge_free_space() comment. */
2304         if (right_info && !right_info->bitmap &&
2305             (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2306                 if (update_stat)
2307                         unlink_free_space(ctl, right_info);
2308                 else
2309                         __unlink_free_space(ctl, right_info);
2310                 info->bytes += right_info->bytes;
2311                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2312                 merged = true;
2313         }
2314
2315         /* See try_merge_free_space() comment. */
2316         if (left_info && !left_info->bitmap &&
2317             left_info->offset + left_info->bytes == offset &&
2318             (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2319                 if (update_stat)
2320                         unlink_free_space(ctl, left_info);
2321                 else
2322                         __unlink_free_space(ctl, left_info);
2323                 info->offset = left_info->offset;
2324                 info->bytes += left_info->bytes;
2325                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2326                 merged = true;
2327         }
2328
2329         return merged;
2330 }
2331
2332 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2333                                      struct btrfs_free_space *info,
2334                                      bool update_stat)
2335 {
2336         struct btrfs_free_space *bitmap;
2337         unsigned long i;
2338         unsigned long j;
2339         const u64 end = info->offset + info->bytes;
2340         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2341         u64 bytes;
2342
2343         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2344         if (!bitmap)
2345                 return false;
2346
2347         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2348         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2349         if (j == i)
2350                 return false;
2351         bytes = (j - i) * ctl->unit;
2352         info->bytes += bytes;
2353
2354         /* See try_merge_free_space() comment. */
2355         if (!btrfs_free_space_trimmed(bitmap))
2356                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2357
2358         if (update_stat)
2359                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2360         else
2361                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2362
2363         if (!bitmap->bytes)
2364                 free_bitmap(ctl, bitmap);
2365
2366         return true;
2367 }
2368
2369 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2370                                        struct btrfs_free_space *info,
2371                                        bool update_stat)
2372 {
2373         struct btrfs_free_space *bitmap;
2374         u64 bitmap_offset;
2375         unsigned long i;
2376         unsigned long j;
2377         unsigned long prev_j;
2378         u64 bytes;
2379
2380         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2381         /* If we're on a boundary, try the previous logical bitmap. */
2382         if (bitmap_offset == info->offset) {
2383                 if (info->offset == 0)
2384                         return false;
2385                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2386         }
2387
2388         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2389         if (!bitmap)
2390                 return false;
2391
2392         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2393         j = 0;
2394         prev_j = (unsigned long)-1;
2395         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2396                 if (j > i)
2397                         break;
2398                 prev_j = j;
2399         }
2400         if (prev_j == i)
2401                 return false;
2402
2403         if (prev_j == (unsigned long)-1)
2404                 bytes = (i + 1) * ctl->unit;
2405         else
2406                 bytes = (i - prev_j) * ctl->unit;
2407
2408         info->offset -= bytes;
2409         info->bytes += bytes;
2410
2411         /* See try_merge_free_space() comment. */
2412         if (!btrfs_free_space_trimmed(bitmap))
2413                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2414
2415         if (update_stat)
2416                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2417         else
2418                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2419
2420         if (!bitmap->bytes)
2421                 free_bitmap(ctl, bitmap);
2422
2423         return true;
2424 }
2425
2426 /*
2427  * We prefer always to allocate from extent entries, both for clustered and
2428  * non-clustered allocation requests. So when attempting to add a new extent
2429  * entry, try to see if there's adjacent free space in bitmap entries, and if
2430  * there is, migrate that space from the bitmaps to the extent.
2431  * Like this we get better chances of satisfying space allocation requests
2432  * because we attempt to satisfy them based on a single cache entry, and never
2433  * on 2 or more entries - even if the entries represent a contiguous free space
2434  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2435  * ends).
2436  */
2437 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2438                               struct btrfs_free_space *info,
2439                               bool update_stat)
2440 {
2441         /*
2442          * Only work with disconnected entries, as we can change their offset,
2443          * and must be extent entries.
2444          */
2445         ASSERT(!info->bitmap);
2446         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2447
2448         if (ctl->total_bitmaps > 0) {
2449                 bool stole_end;
2450                 bool stole_front = false;
2451
2452                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2453                 if (ctl->total_bitmaps > 0)
2454                         stole_front = steal_from_bitmap_to_front(ctl, info,
2455                                                                  update_stat);
2456
2457                 if (stole_end || stole_front)
2458                         try_merge_free_space(ctl, info, update_stat);
2459         }
2460 }
2461
2462 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2463                            struct btrfs_free_space_ctl *ctl,
2464                            u64 offset, u64 bytes,
2465                            enum btrfs_trim_state trim_state)
2466 {
2467         struct btrfs_block_group *block_group = ctl->private;
2468         struct btrfs_free_space *info;
2469         int ret = 0;
2470         u64 filter_bytes = bytes;
2471
2472         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2473         if (!info)
2474                 return -ENOMEM;
2475
2476         info->offset = offset;
2477         info->bytes = bytes;
2478         info->trim_state = trim_state;
2479         RB_CLEAR_NODE(&info->offset_index);
2480
2481         spin_lock(&ctl->tree_lock);
2482
2483         if (try_merge_free_space(ctl, info, true))
2484                 goto link;
2485
2486         /*
2487          * There was no extent directly to the left or right of this new
2488          * extent then we know we're going to have to allocate a new extent, so
2489          * before we do that see if we need to drop this into a bitmap
2490          */
2491         ret = insert_into_bitmap(ctl, info);
2492         if (ret < 0) {
2493                 goto out;
2494         } else if (ret) {
2495                 ret = 0;
2496                 goto out;
2497         }
2498 link:
2499         /*
2500          * Only steal free space from adjacent bitmaps if we're sure we're not
2501          * going to add the new free space to existing bitmap entries - because
2502          * that would mean unnecessary work that would be reverted. Therefore
2503          * attempt to steal space from bitmaps if we're adding an extent entry.
2504          */
2505         steal_from_bitmap(ctl, info, true);
2506
2507         filter_bytes = max(filter_bytes, info->bytes);
2508
2509         ret = link_free_space(ctl, info);
2510         if (ret)
2511                 kmem_cache_free(btrfs_free_space_cachep, info);
2512 out:
2513         btrfs_discard_update_discardable(block_group, ctl);
2514         spin_unlock(&ctl->tree_lock);
2515
2516         if (ret) {
2517                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2518                 ASSERT(ret != -EEXIST);
2519         }
2520
2521         if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2522                 btrfs_discard_check_filter(block_group, filter_bytes);
2523                 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2524         }
2525
2526         return ret;
2527 }
2528
2529 int btrfs_add_free_space(struct btrfs_block_group *block_group,
2530                          u64 bytenr, u64 size)
2531 {
2532         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2533
2534         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2535                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2536
2537         return __btrfs_add_free_space(block_group->fs_info,
2538                                       block_group->free_space_ctl,
2539                                       bytenr, size, trim_state);
2540 }
2541
2542 /*
2543  * This is a subtle distinction because when adding free space back in general,
2544  * we want it to be added as untrimmed for async. But in the case where we add
2545  * it on loading of a block group, we want to consider it trimmed.
2546  */
2547 int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2548                                        u64 bytenr, u64 size)
2549 {
2550         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2551
2552         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2553             btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2554                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2555
2556         return __btrfs_add_free_space(block_group->fs_info,
2557                                       block_group->free_space_ctl,
2558                                       bytenr, size, trim_state);
2559 }
2560
2561 int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2562                             u64 offset, u64 bytes)
2563 {
2564         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2565         struct btrfs_free_space *info;
2566         int ret;
2567         bool re_search = false;
2568
2569         spin_lock(&ctl->tree_lock);
2570
2571 again:
2572         ret = 0;
2573         if (!bytes)
2574                 goto out_lock;
2575
2576         info = tree_search_offset(ctl, offset, 0, 0);
2577         if (!info) {
2578                 /*
2579                  * oops didn't find an extent that matched the space we wanted
2580                  * to remove, look for a bitmap instead
2581                  */
2582                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2583                                           1, 0);
2584                 if (!info) {
2585                         /*
2586                          * If we found a partial bit of our free space in a
2587                          * bitmap but then couldn't find the other part this may
2588                          * be a problem, so WARN about it.
2589                          */
2590                         WARN_ON(re_search);
2591                         goto out_lock;
2592                 }
2593         }
2594
2595         re_search = false;
2596         if (!info->bitmap) {
2597                 unlink_free_space(ctl, info);
2598                 if (offset == info->offset) {
2599                         u64 to_free = min(bytes, info->bytes);
2600
2601                         info->bytes -= to_free;
2602                         info->offset += to_free;
2603                         if (info->bytes) {
2604                                 ret = link_free_space(ctl, info);
2605                                 WARN_ON(ret);
2606                         } else {
2607                                 kmem_cache_free(btrfs_free_space_cachep, info);
2608                         }
2609
2610                         offset += to_free;
2611                         bytes -= to_free;
2612                         goto again;
2613                 } else {
2614                         u64 old_end = info->bytes + info->offset;
2615
2616                         info->bytes = offset - info->offset;
2617                         ret = link_free_space(ctl, info);
2618                         WARN_ON(ret);
2619                         if (ret)
2620                                 goto out_lock;
2621
2622                         /* Not enough bytes in this entry to satisfy us */
2623                         if (old_end < offset + bytes) {
2624                                 bytes -= old_end - offset;
2625                                 offset = old_end;
2626                                 goto again;
2627                         } else if (old_end == offset + bytes) {
2628                                 /* all done */
2629                                 goto out_lock;
2630                         }
2631                         spin_unlock(&ctl->tree_lock);
2632
2633                         ret = __btrfs_add_free_space(block_group->fs_info, ctl,
2634                                                      offset + bytes,
2635                                                      old_end - (offset + bytes),
2636                                                      info->trim_state);
2637                         WARN_ON(ret);
2638                         goto out;
2639                 }
2640         }
2641
2642         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2643         if (ret == -EAGAIN) {
2644                 re_search = true;
2645                 goto again;
2646         }
2647 out_lock:
2648         btrfs_discard_update_discardable(block_group, ctl);
2649         spin_unlock(&ctl->tree_lock);
2650 out:
2651         return ret;
2652 }
2653
2654 void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2655                            u64 bytes)
2656 {
2657         struct btrfs_fs_info *fs_info = block_group->fs_info;
2658         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2659         struct btrfs_free_space *info;
2660         struct rb_node *n;
2661         int count = 0;
2662
2663         spin_lock(&ctl->tree_lock);
2664         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2665                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2666                 if (info->bytes >= bytes && !block_group->ro)
2667                         count++;
2668                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2669                            info->offset, info->bytes,
2670                        (info->bitmap) ? "yes" : "no");
2671         }
2672         spin_unlock(&ctl->tree_lock);
2673         btrfs_info(fs_info, "block group has cluster?: %s",
2674                list_empty(&block_group->cluster_list) ? "no" : "yes");
2675         btrfs_info(fs_info,
2676                    "%d blocks of free space at or bigger than bytes is", count);
2677 }
2678
2679 void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group)
2680 {
2681         struct btrfs_fs_info *fs_info = block_group->fs_info;
2682         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2683
2684         spin_lock_init(&ctl->tree_lock);
2685         ctl->unit = fs_info->sectorsize;
2686         ctl->start = block_group->start;
2687         ctl->private = block_group;
2688         ctl->op = &free_space_op;
2689         INIT_LIST_HEAD(&ctl->trimming_ranges);
2690         mutex_init(&ctl->cache_writeout_mutex);
2691
2692         /*
2693          * we only want to have 32k of ram per block group for keeping
2694          * track of free space, and if we pass 1/2 of that we want to
2695          * start converting things over to using bitmaps
2696          */
2697         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2698 }
2699
2700 /*
2701  * for a given cluster, put all of its extents back into the free
2702  * space cache.  If the block group passed doesn't match the block group
2703  * pointed to by the cluster, someone else raced in and freed the
2704  * cluster already.  In that case, we just return without changing anything
2705  */
2706 static int
2707 __btrfs_return_cluster_to_free_space(
2708                              struct btrfs_block_group *block_group,
2709                              struct btrfs_free_cluster *cluster)
2710 {
2711         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2712         struct btrfs_free_space *entry;
2713         struct rb_node *node;
2714
2715         spin_lock(&cluster->lock);
2716         if (cluster->block_group != block_group)
2717                 goto out;
2718
2719         cluster->block_group = NULL;
2720         cluster->window_start = 0;
2721         list_del_init(&cluster->block_group_list);
2722
2723         node = rb_first(&cluster->root);
2724         while (node) {
2725                 bool bitmap;
2726
2727                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2728                 node = rb_next(&entry->offset_index);
2729                 rb_erase(&entry->offset_index, &cluster->root);
2730                 RB_CLEAR_NODE(&entry->offset_index);
2731
2732                 bitmap = (entry->bitmap != NULL);
2733                 if (!bitmap) {
2734                         /* Merging treats extents as if they were new */
2735                         if (!btrfs_free_space_trimmed(entry)) {
2736                                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
2737                                 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
2738                                         entry->bytes;
2739                         }
2740
2741                         try_merge_free_space(ctl, entry, false);
2742                         steal_from_bitmap(ctl, entry, false);
2743
2744                         /* As we insert directly, update these statistics */
2745                         if (!btrfs_free_space_trimmed(entry)) {
2746                                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
2747                                 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
2748                                         entry->bytes;
2749                         }
2750                 }
2751                 tree_insert_offset(&ctl->free_space_offset,
2752                                    entry->offset, &entry->offset_index, bitmap);
2753         }
2754         cluster->root = RB_ROOT;
2755
2756 out:
2757         spin_unlock(&cluster->lock);
2758         btrfs_put_block_group(block_group);
2759         return 0;
2760 }
2761
2762 static void __btrfs_remove_free_space_cache_locked(
2763                                 struct btrfs_free_space_ctl *ctl)
2764 {
2765         struct btrfs_free_space *info;
2766         struct rb_node *node;
2767
2768         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2769                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2770                 if (!info->bitmap) {
2771                         unlink_free_space(ctl, info);
2772                         kmem_cache_free(btrfs_free_space_cachep, info);
2773                 } else {
2774                         free_bitmap(ctl, info);
2775                 }
2776
2777                 cond_resched_lock(&ctl->tree_lock);
2778         }
2779 }
2780
2781 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2782 {
2783         spin_lock(&ctl->tree_lock);
2784         __btrfs_remove_free_space_cache_locked(ctl);
2785         if (ctl->private)
2786                 btrfs_discard_update_discardable(ctl->private, ctl);
2787         spin_unlock(&ctl->tree_lock);
2788 }
2789
2790 void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
2791 {
2792         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2793         struct btrfs_free_cluster *cluster;
2794         struct list_head *head;
2795
2796         spin_lock(&ctl->tree_lock);
2797         while ((head = block_group->cluster_list.next) !=
2798                &block_group->cluster_list) {
2799                 cluster = list_entry(head, struct btrfs_free_cluster,
2800                                      block_group_list);
2801
2802                 WARN_ON(cluster->block_group != block_group);
2803                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2804
2805                 cond_resched_lock(&ctl->tree_lock);
2806         }
2807         __btrfs_remove_free_space_cache_locked(ctl);
2808         btrfs_discard_update_discardable(block_group, ctl);
2809         spin_unlock(&ctl->tree_lock);
2810
2811 }
2812
2813 /**
2814  * btrfs_is_free_space_trimmed - see if everything is trimmed
2815  * @block_group: block_group of interest
2816  *
2817  * Walk @block_group's free space rb_tree to determine if everything is trimmed.
2818  */
2819 bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
2820 {
2821         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2822         struct btrfs_free_space *info;
2823         struct rb_node *node;
2824         bool ret = true;
2825
2826         spin_lock(&ctl->tree_lock);
2827         node = rb_first(&ctl->free_space_offset);
2828
2829         while (node) {
2830                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2831
2832                 if (!btrfs_free_space_trimmed(info)) {
2833                         ret = false;
2834                         break;
2835                 }
2836
2837                 node = rb_next(node);
2838         }
2839
2840         spin_unlock(&ctl->tree_lock);
2841         return ret;
2842 }
2843
2844 u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
2845                                u64 offset, u64 bytes, u64 empty_size,
2846                                u64 *max_extent_size)
2847 {
2848         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2849         struct btrfs_discard_ctl *discard_ctl =
2850                                         &block_group->fs_info->discard_ctl;
2851         struct btrfs_free_space *entry = NULL;
2852         u64 bytes_search = bytes + empty_size;
2853         u64 ret = 0;
2854         u64 align_gap = 0;
2855         u64 align_gap_len = 0;
2856         enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2857
2858         spin_lock(&ctl->tree_lock);
2859         entry = find_free_space(ctl, &offset, &bytes_search,
2860                                 block_group->full_stripe_len, max_extent_size);
2861         if (!entry)
2862                 goto out;
2863
2864         ret = offset;
2865         if (entry->bitmap) {
2866                 bitmap_clear_bits(ctl, entry, offset, bytes);
2867
2868                 if (!btrfs_free_space_trimmed(entry))
2869                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
2870
2871                 if (!entry->bytes)
2872                         free_bitmap(ctl, entry);
2873         } else {
2874                 unlink_free_space(ctl, entry);
2875                 align_gap_len = offset - entry->offset;
2876                 align_gap = entry->offset;
2877                 align_gap_trim_state = entry->trim_state;
2878
2879                 if (!btrfs_free_space_trimmed(entry))
2880                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
2881
2882                 entry->offset = offset + bytes;
2883                 WARN_ON(entry->bytes < bytes + align_gap_len);
2884
2885                 entry->bytes -= bytes + align_gap_len;
2886                 if (!entry->bytes)
2887                         kmem_cache_free(btrfs_free_space_cachep, entry);
2888                 else
2889                         link_free_space(ctl, entry);
2890         }
2891 out:
2892         btrfs_discard_update_discardable(block_group, ctl);
2893         spin_unlock(&ctl->tree_lock);
2894
2895         if (align_gap_len)
2896                 __btrfs_add_free_space(block_group->fs_info, ctl,
2897                                        align_gap, align_gap_len,
2898                                        align_gap_trim_state);
2899         return ret;
2900 }
2901
2902 /*
2903  * given a cluster, put all of its extents back into the free space
2904  * cache.  If a block group is passed, this function will only free
2905  * a cluster that belongs to the passed block group.
2906  *
2907  * Otherwise, it'll get a reference on the block group pointed to by the
2908  * cluster and remove the cluster from it.
2909  */
2910 int btrfs_return_cluster_to_free_space(
2911                                struct btrfs_block_group *block_group,
2912                                struct btrfs_free_cluster *cluster)
2913 {
2914         struct btrfs_free_space_ctl *ctl;
2915         int ret;
2916
2917         /* first, get a safe pointer to the block group */
2918         spin_lock(&cluster->lock);
2919         if (!block_group) {
2920                 block_group = cluster->block_group;
2921                 if (!block_group) {
2922                         spin_unlock(&cluster->lock);
2923                         return 0;
2924                 }
2925         } else if (cluster->block_group != block_group) {
2926                 /* someone else has already freed it don't redo their work */
2927                 spin_unlock(&cluster->lock);
2928                 return 0;
2929         }
2930         atomic_inc(&block_group->count);
2931         spin_unlock(&cluster->lock);
2932
2933         ctl = block_group->free_space_ctl;
2934
2935         /* now return any extents the cluster had on it */
2936         spin_lock(&ctl->tree_lock);
2937         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2938         spin_unlock(&ctl->tree_lock);
2939
2940         btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
2941
2942         /* finally drop our ref */
2943         btrfs_put_block_group(block_group);
2944         return ret;
2945 }
2946
2947 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
2948                                    struct btrfs_free_cluster *cluster,
2949                                    struct btrfs_free_space *entry,
2950                                    u64 bytes, u64 min_start,
2951                                    u64 *max_extent_size)
2952 {
2953         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2954         int err;
2955         u64 search_start = cluster->window_start;
2956         u64 search_bytes = bytes;
2957         u64 ret = 0;
2958
2959         search_start = min_start;
2960         search_bytes = bytes;
2961
2962         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2963         if (err) {
2964                 *max_extent_size = max(get_max_extent_size(entry),
2965                                        *max_extent_size);
2966                 return 0;
2967         }
2968
2969         ret = search_start;
2970         __bitmap_clear_bits(ctl, entry, ret, bytes);
2971
2972         return ret;
2973 }
2974
2975 /*
2976  * given a cluster, try to allocate 'bytes' from it, returns 0
2977  * if it couldn't find anything suitably large, or a logical disk offset
2978  * if things worked out
2979  */
2980 u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
2981                              struct btrfs_free_cluster *cluster, u64 bytes,
2982                              u64 min_start, u64 *max_extent_size)
2983 {
2984         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2985         struct btrfs_discard_ctl *discard_ctl =
2986                                         &block_group->fs_info->discard_ctl;
2987         struct btrfs_free_space *entry = NULL;
2988         struct rb_node *node;
2989         u64 ret = 0;
2990
2991         spin_lock(&cluster->lock);
2992         if (bytes > cluster->max_size)
2993                 goto out;
2994
2995         if (cluster->block_group != block_group)
2996                 goto out;
2997
2998         node = rb_first(&cluster->root);
2999         if (!node)
3000                 goto out;
3001
3002         entry = rb_entry(node, struct btrfs_free_space, offset_index);
3003         while (1) {
3004                 if (entry->bytes < bytes)
3005                         *max_extent_size = max(get_max_extent_size(entry),
3006                                                *max_extent_size);
3007
3008                 if (entry->bytes < bytes ||
3009                     (!entry->bitmap && entry->offset < min_start)) {
3010                         node = rb_next(&entry->offset_index);
3011                         if (!node)
3012                                 break;
3013                         entry = rb_entry(node, struct btrfs_free_space,
3014                                          offset_index);
3015                         continue;
3016                 }
3017
3018                 if (entry->bitmap) {
3019                         ret = btrfs_alloc_from_bitmap(block_group,
3020                                                       cluster, entry, bytes,
3021                                                       cluster->window_start,
3022                                                       max_extent_size);
3023                         if (ret == 0) {
3024                                 node = rb_next(&entry->offset_index);
3025                                 if (!node)
3026                                         break;
3027                                 entry = rb_entry(node, struct btrfs_free_space,
3028                                                  offset_index);
3029                                 continue;
3030                         }
3031                         cluster->window_start += bytes;
3032                 } else {
3033                         ret = entry->offset;
3034
3035                         entry->offset += bytes;
3036                         entry->bytes -= bytes;
3037                 }
3038
3039                 if (entry->bytes == 0)
3040                         rb_erase(&entry->offset_index, &cluster->root);
3041                 break;
3042         }
3043 out:
3044         spin_unlock(&cluster->lock);
3045
3046         if (!ret)
3047                 return 0;
3048
3049         spin_lock(&ctl->tree_lock);
3050
3051         if (!btrfs_free_space_trimmed(entry))
3052                 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3053
3054         ctl->free_space -= bytes;
3055         if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3056                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3057         if (entry->bytes == 0) {
3058                 ctl->free_extents--;
3059                 if (entry->bitmap) {
3060                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
3061                                         entry->bitmap);
3062                         ctl->total_bitmaps--;
3063                         ctl->op->recalc_thresholds(ctl);
3064                 } else if (!btrfs_free_space_trimmed(entry)) {
3065                         ctl->discardable_extents[BTRFS_STAT_CURR]--;
3066                 }
3067                 kmem_cache_free(btrfs_free_space_cachep, entry);
3068         }
3069
3070         spin_unlock(&ctl->tree_lock);
3071
3072         return ret;
3073 }
3074
3075 static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3076                                 struct btrfs_free_space *entry,
3077                                 struct btrfs_free_cluster *cluster,
3078                                 u64 offset, u64 bytes,
3079                                 u64 cont1_bytes, u64 min_bytes)
3080 {
3081         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3082         unsigned long next_zero;
3083         unsigned long i;
3084         unsigned long want_bits;
3085         unsigned long min_bits;
3086         unsigned long found_bits;
3087         unsigned long max_bits = 0;
3088         unsigned long start = 0;
3089         unsigned long total_found = 0;
3090         int ret;
3091
3092         i = offset_to_bit(entry->offset, ctl->unit,
3093                           max_t(u64, offset, entry->offset));
3094         want_bits = bytes_to_bits(bytes, ctl->unit);
3095         min_bits = bytes_to_bits(min_bytes, ctl->unit);
3096
3097         /*
3098          * Don't bother looking for a cluster in this bitmap if it's heavily
3099          * fragmented.
3100          */
3101         if (entry->max_extent_size &&
3102             entry->max_extent_size < cont1_bytes)
3103                 return -ENOSPC;
3104 again:
3105         found_bits = 0;
3106         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3107                 next_zero = find_next_zero_bit(entry->bitmap,
3108                                                BITS_PER_BITMAP, i);
3109                 if (next_zero - i >= min_bits) {
3110                         found_bits = next_zero - i;
3111                         if (found_bits > max_bits)
3112                                 max_bits = found_bits;
3113                         break;
3114                 }
3115                 if (next_zero - i > max_bits)
3116                         max_bits = next_zero - i;
3117                 i = next_zero;
3118         }
3119
3120         if (!found_bits) {
3121                 entry->max_extent_size = (u64)max_bits * ctl->unit;
3122                 return -ENOSPC;
3123         }
3124
3125         if (!total_found) {
3126                 start = i;
3127                 cluster->max_size = 0;
3128         }
3129
3130         total_found += found_bits;
3131
3132         if (cluster->max_size < found_bits * ctl->unit)
3133                 cluster->max_size = found_bits * ctl->unit;
3134
3135         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3136                 i = next_zero + 1;
3137                 goto again;
3138         }
3139
3140         cluster->window_start = start * ctl->unit + entry->offset;
3141         rb_erase(&entry->offset_index, &ctl->free_space_offset);
3142         ret = tree_insert_offset(&cluster->root, entry->offset,
3143                                  &entry->offset_index, 1);
3144         ASSERT(!ret); /* -EEXIST; Logic error */
3145
3146         trace_btrfs_setup_cluster(block_group, cluster,
3147                                   total_found * ctl->unit, 1);
3148         return 0;
3149 }
3150
3151 /*
3152  * This searches the block group for just extents to fill the cluster with.
3153  * Try to find a cluster with at least bytes total bytes, at least one
3154  * extent of cont1_bytes, and other clusters of at least min_bytes.
3155  */
3156 static noinline int
3157 setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3158                         struct btrfs_free_cluster *cluster,
3159                         struct list_head *bitmaps, u64 offset, u64 bytes,
3160                         u64 cont1_bytes, u64 min_bytes)
3161 {
3162         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3163         struct btrfs_free_space *first = NULL;
3164         struct btrfs_free_space *entry = NULL;
3165         struct btrfs_free_space *last;
3166         struct rb_node *node;
3167         u64 window_free;
3168         u64 max_extent;
3169         u64 total_size = 0;
3170
3171         entry = tree_search_offset(ctl, offset, 0, 1);
3172         if (!entry)
3173                 return -ENOSPC;
3174
3175         /*
3176          * We don't want bitmaps, so just move along until we find a normal
3177          * extent entry.
3178          */
3179         while (entry->bitmap || entry->bytes < min_bytes) {
3180                 if (entry->bitmap && list_empty(&entry->list))
3181                         list_add_tail(&entry->list, bitmaps);
3182                 node = rb_next(&entry->offset_index);
3183                 if (!node)
3184                         return -ENOSPC;
3185                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3186         }
3187
3188         window_free = entry->bytes;
3189         max_extent = entry->bytes;
3190         first = entry;
3191         last = entry;
3192
3193         for (node = rb_next(&entry->offset_index); node;
3194              node = rb_next(&entry->offset_index)) {
3195                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3196
3197                 if (entry->bitmap) {
3198                         if (list_empty(&entry->list))
3199                                 list_add_tail(&entry->list, bitmaps);
3200                         continue;
3201                 }
3202
3203                 if (entry->bytes < min_bytes)
3204                         continue;
3205
3206                 last = entry;
3207                 window_free += entry->bytes;
3208                 if (entry->bytes > max_extent)
3209                         max_extent = entry->bytes;
3210         }
3211
3212         if (window_free < bytes || max_extent < cont1_bytes)
3213                 return -ENOSPC;
3214
3215         cluster->window_start = first->offset;
3216
3217         node = &first->offset_index;
3218
3219         /*
3220          * now we've found our entries, pull them out of the free space
3221          * cache and put them into the cluster rbtree
3222          */
3223         do {
3224                 int ret;
3225
3226                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3227                 node = rb_next(&entry->offset_index);
3228                 if (entry->bitmap || entry->bytes < min_bytes)
3229                         continue;
3230
3231                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3232                 ret = tree_insert_offset(&cluster->root, entry->offset,
3233                                          &entry->offset_index, 0);
3234                 total_size += entry->bytes;
3235                 ASSERT(!ret); /* -EEXIST; Logic error */
3236         } while (node && entry != last);
3237
3238         cluster->max_size = max_extent;
3239         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3240         return 0;
3241 }
3242
3243 /*
3244  * This specifically looks for bitmaps that may work in the cluster, we assume
3245  * that we have already failed to find extents that will work.
3246  */
3247 static noinline int
3248 setup_cluster_bitmap(struct btrfs_block_group *block_group,
3249                      struct btrfs_free_cluster *cluster,
3250                      struct list_head *bitmaps, u64 offset, u64 bytes,
3251                      u64 cont1_bytes, u64 min_bytes)
3252 {
3253         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3254         struct btrfs_free_space *entry = NULL;
3255         int ret = -ENOSPC;
3256         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3257
3258         if (ctl->total_bitmaps == 0)
3259                 return -ENOSPC;
3260
3261         /*
3262          * The bitmap that covers offset won't be in the list unless offset
3263          * is just its start offset.
3264          */
3265         if (!list_empty(bitmaps))
3266                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3267
3268         if (!entry || entry->offset != bitmap_offset) {
3269                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3270                 if (entry && list_empty(&entry->list))
3271                         list_add(&entry->list, bitmaps);
3272         }
3273
3274         list_for_each_entry(entry, bitmaps, list) {
3275                 if (entry->bytes < bytes)
3276                         continue;
3277                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3278                                            bytes, cont1_bytes, min_bytes);
3279                 if (!ret)
3280                         return 0;
3281         }
3282
3283         /*
3284          * The bitmaps list has all the bitmaps that record free space
3285          * starting after offset, so no more search is required.
3286          */
3287         return -ENOSPC;
3288 }
3289
3290 /*
3291  * here we try to find a cluster of blocks in a block group.  The goal
3292  * is to find at least bytes+empty_size.
3293  * We might not find them all in one contiguous area.
3294  *
3295  * returns zero and sets up cluster if things worked out, otherwise
3296  * it returns -enospc
3297  */
3298 int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3299                              struct btrfs_free_cluster *cluster,
3300                              u64 offset, u64 bytes, u64 empty_size)
3301 {
3302         struct btrfs_fs_info *fs_info = block_group->fs_info;
3303         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3304         struct btrfs_free_space *entry, *tmp;
3305         LIST_HEAD(bitmaps);
3306         u64 min_bytes;
3307         u64 cont1_bytes;
3308         int ret;
3309
3310         /*
3311          * Choose the minimum extent size we'll require for this
3312          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3313          * For metadata, allow allocates with smaller extents.  For
3314          * data, keep it dense.
3315          */
3316         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3317                 cont1_bytes = min_bytes = bytes + empty_size;
3318         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3319                 cont1_bytes = bytes;
3320                 min_bytes = fs_info->sectorsize;
3321         } else {
3322                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3323                 min_bytes = fs_info->sectorsize;
3324         }
3325
3326         spin_lock(&ctl->tree_lock);
3327
3328         /*
3329          * If we know we don't have enough space to make a cluster don't even
3330          * bother doing all the work to try and find one.
3331          */
3332         if (ctl->free_space < bytes) {
3333                 spin_unlock(&ctl->tree_lock);
3334                 return -ENOSPC;
3335         }
3336
3337         spin_lock(&cluster->lock);
3338
3339         /* someone already found a cluster, hooray */
3340         if (cluster->block_group) {
3341                 ret = 0;
3342                 goto out;
3343         }
3344
3345         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3346                                  min_bytes);
3347
3348         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3349                                       bytes + empty_size,
3350                                       cont1_bytes, min_bytes);
3351         if (ret)
3352                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3353                                            offset, bytes + empty_size,
3354                                            cont1_bytes, min_bytes);
3355
3356         /* Clear our temporary list */
3357         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3358                 list_del_init(&entry->list);
3359
3360         if (!ret) {
3361                 atomic_inc(&block_group->count);
3362                 list_add_tail(&cluster->block_group_list,
3363                               &block_group->cluster_list);
3364                 cluster->block_group = block_group;
3365         } else {
3366                 trace_btrfs_failed_cluster_setup(block_group);
3367         }
3368 out:
3369         spin_unlock(&cluster->lock);
3370         spin_unlock(&ctl->tree_lock);
3371
3372         return ret;
3373 }
3374
3375 /*
3376  * simple code to zero out a cluster
3377  */
3378 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3379 {
3380         spin_lock_init(&cluster->lock);
3381         spin_lock_init(&cluster->refill_lock);
3382         cluster->root = RB_ROOT;
3383         cluster->max_size = 0;
3384         cluster->fragmented = false;
3385         INIT_LIST_HEAD(&cluster->block_group_list);
3386         cluster->block_group = NULL;
3387 }
3388
3389 static int do_trimming(struct btrfs_block_group *block_group,
3390                        u64 *total_trimmed, u64 start, u64 bytes,
3391                        u64 reserved_start, u64 reserved_bytes,
3392                        enum btrfs_trim_state reserved_trim_state,
3393                        struct btrfs_trim_range *trim_entry)
3394 {
3395         struct btrfs_space_info *space_info = block_group->space_info;
3396         struct btrfs_fs_info *fs_info = block_group->fs_info;
3397         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3398         int ret;
3399         int update = 0;
3400         const u64 end = start + bytes;
3401         const u64 reserved_end = reserved_start + reserved_bytes;
3402         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3403         u64 trimmed = 0;
3404
3405         spin_lock(&space_info->lock);
3406         spin_lock(&block_group->lock);
3407         if (!block_group->ro) {
3408                 block_group->reserved += reserved_bytes;
3409                 space_info->bytes_reserved += reserved_bytes;
3410                 update = 1;
3411         }
3412         spin_unlock(&block_group->lock);
3413         spin_unlock(&space_info->lock);
3414
3415         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3416         if (!ret) {
3417                 *total_trimmed += trimmed;
3418                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3419         }
3420
3421         mutex_lock(&ctl->cache_writeout_mutex);
3422         if (reserved_start < start)
3423                 __btrfs_add_free_space(fs_info, ctl, reserved_start,
3424                                        start - reserved_start,
3425                                        reserved_trim_state);
3426         if (start + bytes < reserved_start + reserved_bytes)
3427                 __btrfs_add_free_space(fs_info, ctl, end, reserved_end - end,
3428                                        reserved_trim_state);
3429         __btrfs_add_free_space(fs_info, ctl, start, bytes, trim_state);
3430         list_del(&trim_entry->list);
3431         mutex_unlock(&ctl->cache_writeout_mutex);
3432
3433         if (update) {
3434                 spin_lock(&space_info->lock);
3435                 spin_lock(&block_group->lock);
3436                 if (block_group->ro)
3437                         space_info->bytes_readonly += reserved_bytes;
3438                 block_group->reserved -= reserved_bytes;
3439                 space_info->bytes_reserved -= reserved_bytes;
3440                 spin_unlock(&block_group->lock);
3441                 spin_unlock(&space_info->lock);
3442         }
3443
3444         return ret;
3445 }
3446
3447 /*
3448  * If @async is set, then we will trim 1 region and return.
3449  */
3450 static int trim_no_bitmap(struct btrfs_block_group *block_group,
3451                           u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3452                           bool async)
3453 {
3454         struct btrfs_discard_ctl *discard_ctl =
3455                                         &block_group->fs_info->discard_ctl;
3456         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3457         struct btrfs_free_space *entry;
3458         struct rb_node *node;
3459         int ret = 0;
3460         u64 extent_start;
3461         u64 extent_bytes;
3462         enum btrfs_trim_state extent_trim_state;
3463         u64 bytes;
3464         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3465
3466         while (start < end) {
3467                 struct btrfs_trim_range trim_entry;
3468
3469                 mutex_lock(&ctl->cache_writeout_mutex);
3470                 spin_lock(&ctl->tree_lock);
3471
3472                 if (ctl->free_space < minlen)
3473                         goto out_unlock;
3474
3475                 entry = tree_search_offset(ctl, start, 0, 1);
3476                 if (!entry)
3477                         goto out_unlock;
3478
3479                 /* Skip bitmaps and if async, already trimmed entries */
3480                 while (entry->bitmap ||
3481                        (async && btrfs_free_space_trimmed(entry))) {
3482                         node = rb_next(&entry->offset_index);
3483                         if (!node)
3484                                 goto out_unlock;
3485                         entry = rb_entry(node, struct btrfs_free_space,
3486                                          offset_index);
3487                 }
3488
3489                 if (entry->offset >= end)
3490                         goto out_unlock;
3491
3492                 extent_start = entry->offset;
3493                 extent_bytes = entry->bytes;
3494                 extent_trim_state = entry->trim_state;
3495                 if (async) {
3496                         start = entry->offset;
3497                         bytes = entry->bytes;
3498                         if (bytes < minlen) {
3499                                 spin_unlock(&ctl->tree_lock);
3500                                 mutex_unlock(&ctl->cache_writeout_mutex);
3501                                 goto next;
3502                         }
3503                         unlink_free_space(ctl, entry);
3504                         /*
3505                          * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3506                          * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3507                          * X when we come back around.  So trim it now.
3508                          */
3509                         if (max_discard_size &&
3510                             bytes >= (max_discard_size +
3511                                       BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3512                                 bytes = max_discard_size;
3513                                 extent_bytes = max_discard_size;
3514                                 entry->offset += max_discard_size;
3515                                 entry->bytes -= max_discard_size;
3516                                 link_free_space(ctl, entry);
3517                         } else {
3518                                 kmem_cache_free(btrfs_free_space_cachep, entry);
3519                         }
3520                 } else {
3521                         start = max(start, extent_start);
3522                         bytes = min(extent_start + extent_bytes, end) - start;
3523                         if (bytes < minlen) {
3524                                 spin_unlock(&ctl->tree_lock);
3525                                 mutex_unlock(&ctl->cache_writeout_mutex);
3526                                 goto next;
3527                         }
3528
3529                         unlink_free_space(ctl, entry);
3530                         kmem_cache_free(btrfs_free_space_cachep, entry);
3531                 }
3532
3533                 spin_unlock(&ctl->tree_lock);
3534                 trim_entry.start = extent_start;
3535                 trim_entry.bytes = extent_bytes;
3536                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3537                 mutex_unlock(&ctl->cache_writeout_mutex);
3538
3539                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3540                                   extent_start, extent_bytes, extent_trim_state,
3541                                   &trim_entry);
3542                 if (ret) {
3543                         block_group->discard_cursor = start + bytes;
3544                         break;
3545                 }
3546 next:
3547                 start += bytes;
3548                 block_group->discard_cursor = start;
3549                 if (async && *total_trimmed)
3550                         break;
3551
3552                 if (fatal_signal_pending(current)) {
3553                         ret = -ERESTARTSYS;
3554                         break;
3555                 }
3556
3557                 cond_resched();
3558         }
3559
3560         return ret;
3561
3562 out_unlock:
3563         block_group->discard_cursor = btrfs_block_group_end(block_group);
3564         spin_unlock(&ctl->tree_lock);
3565         mutex_unlock(&ctl->cache_writeout_mutex);
3566
3567         return ret;
3568 }
3569
3570 /*
3571  * If we break out of trimming a bitmap prematurely, we should reset the
3572  * trimming bit.  In a rather contrieved case, it's possible to race here so
3573  * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3574  *
3575  * start = start of bitmap
3576  * end = near end of bitmap
3577  *
3578  * Thread 1:                    Thread 2:
3579  * trim_bitmaps(start)
3580  *                              trim_bitmaps(end)
3581  *                              end_trimming_bitmap()
3582  * reset_trimming_bitmap()
3583  */
3584 static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3585 {
3586         struct btrfs_free_space *entry;
3587
3588         spin_lock(&ctl->tree_lock);
3589         entry = tree_search_offset(ctl, offset, 1, 0);
3590         if (entry) {
3591                 if (btrfs_free_space_trimmed(entry)) {
3592                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
3593                                 entry->bitmap_extents;
3594                         ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3595                 }
3596                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3597         }
3598
3599         spin_unlock(&ctl->tree_lock);
3600 }
3601
3602 static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3603                                 struct btrfs_free_space *entry)
3604 {
3605         if (btrfs_free_space_trimming_bitmap(entry)) {
3606                 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3607                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3608                         entry->bitmap_extents;
3609                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3610         }
3611 }
3612
3613 /*
3614  * If @async is set, then we will trim 1 region and return.
3615  */
3616 static int trim_bitmaps(struct btrfs_block_group *block_group,
3617                         u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3618                         u64 maxlen, bool async)
3619 {
3620         struct btrfs_discard_ctl *discard_ctl =
3621                                         &block_group->fs_info->discard_ctl;
3622         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3623         struct btrfs_free_space *entry;
3624         int ret = 0;
3625         int ret2;
3626         u64 bytes;
3627         u64 offset = offset_to_bitmap(ctl, start);
3628         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3629
3630         while (offset < end) {
3631                 bool next_bitmap = false;
3632                 struct btrfs_trim_range trim_entry;
3633
3634                 mutex_lock(&ctl->cache_writeout_mutex);
3635                 spin_lock(&ctl->tree_lock);
3636
3637                 if (ctl->free_space < minlen) {
3638                         block_group->discard_cursor =
3639                                 btrfs_block_group_end(block_group);
3640                         spin_unlock(&ctl->tree_lock);
3641                         mutex_unlock(&ctl->cache_writeout_mutex);
3642                         break;
3643                 }
3644
3645                 entry = tree_search_offset(ctl, offset, 1, 0);
3646                 /*
3647                  * Bitmaps are marked trimmed lossily now to prevent constant
3648                  * discarding of the same bitmap (the reason why we are bound
3649                  * by the filters).  So, retrim the block group bitmaps when we
3650                  * are preparing to punt to the unused_bgs list.  This uses
3651                  * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3652                  * which is the only discard index which sets minlen to 0.
3653                  */
3654                 if (!entry || (async && minlen && start == offset &&
3655                                btrfs_free_space_trimmed(entry))) {
3656                         spin_unlock(&ctl->tree_lock);
3657                         mutex_unlock(&ctl->cache_writeout_mutex);
3658                         next_bitmap = true;
3659                         goto next;
3660                 }
3661
3662                 /*
3663                  * Async discard bitmap trimming begins at by setting the start
3664                  * to be key.objectid and the offset_to_bitmap() aligns to the
3665                  * start of the bitmap.  This lets us know we are fully
3666                  * scanning the bitmap rather than only some portion of it.
3667                  */
3668                 if (start == offset)
3669                         entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3670
3671                 bytes = minlen;
3672                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3673                 if (ret2 || start >= end) {
3674                         /*
3675                          * We lossily consider a bitmap trimmed if we only skip
3676                          * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3677                          */
3678                         if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3679                                 end_trimming_bitmap(ctl, entry);
3680                         else
3681                                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3682                         spin_unlock(&ctl->tree_lock);
3683                         mutex_unlock(&ctl->cache_writeout_mutex);
3684                         next_bitmap = true;
3685                         goto next;
3686                 }
3687
3688                 /*
3689                  * We already trimmed a region, but are using the locking above
3690                  * to reset the trim_state.
3691                  */
3692                 if (async && *total_trimmed) {
3693                         spin_unlock(&ctl->tree_lock);
3694                         mutex_unlock(&ctl->cache_writeout_mutex);
3695                         goto out;
3696                 }
3697
3698                 bytes = min(bytes, end - start);
3699                 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3700                         spin_unlock(&ctl->tree_lock);
3701                         mutex_unlock(&ctl->cache_writeout_mutex);
3702                         goto next;
3703                 }
3704
3705                 /*
3706                  * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3707                  * If X < @minlen, we won't trim X when we come back around.
3708                  * So trim it now.  We differ here from trimming extents as we
3709                  * don't keep individual state per bit.
3710                  */
3711                 if (async &&
3712                     max_discard_size &&
3713                     bytes > (max_discard_size + minlen))
3714                         bytes = max_discard_size;
3715
3716                 bitmap_clear_bits(ctl, entry, start, bytes);
3717                 if (entry->bytes == 0)
3718                         free_bitmap(ctl, entry);
3719
3720                 spin_unlock(&ctl->tree_lock);
3721                 trim_entry.start = start;
3722                 trim_entry.bytes = bytes;
3723                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3724                 mutex_unlock(&ctl->cache_writeout_mutex);
3725
3726                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3727                                   start, bytes, 0, &trim_entry);
3728                 if (ret) {
3729                         reset_trimming_bitmap(ctl, offset);
3730                         block_group->discard_cursor =
3731                                 btrfs_block_group_end(block_group);
3732                         break;
3733                 }
3734 next:
3735                 if (next_bitmap) {
3736                         offset += BITS_PER_BITMAP * ctl->unit;
3737                         start = offset;
3738                 } else {
3739                         start += bytes;
3740                 }
3741                 block_group->discard_cursor = start;
3742
3743                 if (fatal_signal_pending(current)) {
3744                         if (start != offset)
3745                                 reset_trimming_bitmap(ctl, offset);
3746                         ret = -ERESTARTSYS;
3747                         break;
3748                 }
3749
3750                 cond_resched();
3751         }
3752
3753         if (offset >= end)
3754                 block_group->discard_cursor = end;
3755
3756 out:
3757         return ret;
3758 }
3759
3760 int btrfs_trim_block_group(struct btrfs_block_group *block_group,
3761                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3762 {
3763         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3764         int ret;
3765         u64 rem = 0;
3766
3767         *trimmed = 0;
3768
3769         spin_lock(&block_group->lock);
3770         if (block_group->removed) {
3771                 spin_unlock(&block_group->lock);
3772                 return 0;
3773         }
3774         btrfs_freeze_block_group(block_group);
3775         spin_unlock(&block_group->lock);
3776
3777         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
3778         if (ret)
3779                 goto out;
3780
3781         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
3782         div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
3783         /* If we ended in the middle of a bitmap, reset the trimming flag */
3784         if (rem)
3785                 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
3786 out:
3787         btrfs_unfreeze_block_group(block_group);
3788         return ret;
3789 }
3790
3791 int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
3792                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
3793                                    bool async)
3794 {
3795         int ret;
3796
3797         *trimmed = 0;
3798
3799         spin_lock(&block_group->lock);
3800         if (block_group->removed) {
3801                 spin_unlock(&block_group->lock);
3802                 return 0;
3803         }
3804         btrfs_freeze_block_group(block_group);
3805         spin_unlock(&block_group->lock);
3806
3807         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
3808         btrfs_unfreeze_block_group(block_group);
3809
3810         return ret;
3811 }
3812
3813 int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
3814                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
3815                                    u64 maxlen, bool async)
3816 {
3817         int ret;
3818
3819         *trimmed = 0;
3820
3821         spin_lock(&block_group->lock);
3822         if (block_group->removed) {
3823                 spin_unlock(&block_group->lock);
3824                 return 0;
3825         }
3826         btrfs_freeze_block_group(block_group);
3827         spin_unlock(&block_group->lock);
3828
3829         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
3830                            async);
3831
3832         btrfs_unfreeze_block_group(block_group);
3833
3834         return ret;
3835 }
3836
3837 /*
3838  * Find the left-most item in the cache tree, and then return the
3839  * smallest inode number in the item.
3840  *
3841  * Note: the returned inode number may not be the smallest one in
3842  * the tree, if the left-most item is a bitmap.
3843  */
3844 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3845 {
3846         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3847         struct btrfs_free_space *entry = NULL;
3848         u64 ino = 0;
3849
3850         spin_lock(&ctl->tree_lock);
3851
3852         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3853                 goto out;
3854
3855         entry = rb_entry(rb_first(&ctl->free_space_offset),
3856                          struct btrfs_free_space, offset_index);
3857
3858         if (!entry->bitmap) {
3859                 ino = entry->offset;
3860
3861                 unlink_free_space(ctl, entry);
3862                 entry->offset++;
3863                 entry->bytes--;
3864                 if (!entry->bytes)
3865                         kmem_cache_free(btrfs_free_space_cachep, entry);
3866                 else
3867                         link_free_space(ctl, entry);
3868         } else {
3869                 u64 offset = 0;
3870                 u64 count = 1;
3871                 int ret;
3872
3873                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3874                 /* Logic error; Should be empty if it can't find anything */
3875                 ASSERT(!ret);
3876
3877                 ino = offset;
3878                 bitmap_clear_bits(ctl, entry, offset, 1);
3879                 if (entry->bytes == 0)
3880                         free_bitmap(ctl, entry);
3881         }
3882 out:
3883         spin_unlock(&ctl->tree_lock);
3884
3885         return ino;
3886 }
3887
3888 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3889                                     struct btrfs_path *path)
3890 {
3891         struct inode *inode = NULL;
3892
3893         spin_lock(&root->ino_cache_lock);
3894         if (root->ino_cache_inode)
3895                 inode = igrab(root->ino_cache_inode);
3896         spin_unlock(&root->ino_cache_lock);
3897         if (inode)
3898                 return inode;
3899
3900         inode = __lookup_free_space_inode(root, path, 0);
3901         if (IS_ERR(inode))
3902                 return inode;
3903
3904         spin_lock(&root->ino_cache_lock);
3905         if (!btrfs_fs_closing(root->fs_info))
3906                 root->ino_cache_inode = igrab(inode);
3907         spin_unlock(&root->ino_cache_lock);
3908
3909         return inode;
3910 }
3911
3912 int create_free_ino_inode(struct btrfs_root *root,
3913                           struct btrfs_trans_handle *trans,
3914                           struct btrfs_path *path)
3915 {
3916         return __create_free_space_inode(root, trans, path,
3917                                          BTRFS_FREE_INO_OBJECTID, 0);
3918 }
3919
3920 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3921 {
3922         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3923         struct btrfs_path *path;
3924         struct inode *inode;
3925         int ret = 0;
3926         u64 root_gen = btrfs_root_generation(&root->root_item);
3927
3928         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3929                 return 0;
3930
3931         /*
3932          * If we're unmounting then just return, since this does a search on the
3933          * normal root and not the commit root and we could deadlock.
3934          */
3935         if (btrfs_fs_closing(fs_info))
3936                 return 0;
3937
3938         path = btrfs_alloc_path();
3939         if (!path)
3940                 return 0;
3941
3942         inode = lookup_free_ino_inode(root, path);
3943         if (IS_ERR(inode))
3944                 goto out;
3945
3946         if (root_gen != BTRFS_I(inode)->generation)
3947                 goto out_put;
3948
3949         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3950
3951         if (ret < 0)
3952                 btrfs_err(fs_info,
3953                         "failed to load free ino cache for root %llu",
3954                         root->root_key.objectid);
3955 out_put:
3956         iput(inode);
3957 out:
3958         btrfs_free_path(path);
3959         return ret;
3960 }
3961
3962 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3963                               struct btrfs_trans_handle *trans,
3964                               struct btrfs_path *path,
3965                               struct inode *inode)
3966 {
3967         struct btrfs_fs_info *fs_info = root->fs_info;
3968         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3969         int ret;
3970         struct btrfs_io_ctl io_ctl;
3971         bool release_metadata = true;
3972
3973         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3974                 return 0;
3975
3976         memset(&io_ctl, 0, sizeof(io_ctl));
3977         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl, trans);
3978         if (!ret) {
3979                 /*
3980                  * At this point writepages() didn't error out, so our metadata
3981                  * reservation is released when the writeback finishes, at
3982                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3983                  * with or without an error.
3984                  */
3985                 release_metadata = false;
3986                 ret = btrfs_wait_cache_io_root(root, trans, &io_ctl, path);
3987         }
3988
3989         if (ret) {
3990                 if (release_metadata)
3991                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3992                                         inode->i_size, true);
3993                 btrfs_debug(fs_info,
3994                           "failed to write free ino cache for root %llu error %d",
3995                           root->root_key.objectid, ret);
3996         }
3997
3998         return ret;
3999 }
4000
4001 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4002 /*
4003  * Use this if you need to make a bitmap or extent entry specifically, it
4004  * doesn't do any of the merging that add_free_space does, this acts a lot like
4005  * how the free space cache loading stuff works, so you can get really weird
4006  * configurations.
4007  */
4008 int test_add_free_space_entry(struct btrfs_block_group *cache,
4009                               u64 offset, u64 bytes, bool bitmap)
4010 {
4011         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4012         struct btrfs_free_space *info = NULL, *bitmap_info;
4013         void *map = NULL;
4014         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4015         u64 bytes_added;
4016         int ret;
4017
4018 again:
4019         if (!info) {
4020                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4021                 if (!info)
4022                         return -ENOMEM;
4023         }
4024
4025         if (!bitmap) {
4026                 spin_lock(&ctl->tree_lock);
4027                 info->offset = offset;
4028                 info->bytes = bytes;
4029                 info->max_extent_size = 0;
4030                 ret = link_free_space(ctl, info);
4031                 spin_unlock(&ctl->tree_lock);
4032                 if (ret)
4033                         kmem_cache_free(btrfs_free_space_cachep, info);
4034                 return ret;
4035         }
4036
4037         if (!map) {
4038                 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4039                 if (!map) {
4040                         kmem_cache_free(btrfs_free_space_cachep, info);
4041                         return -ENOMEM;
4042                 }
4043         }
4044
4045         spin_lock(&ctl->tree_lock);
4046         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4047                                          1, 0);
4048         if (!bitmap_info) {
4049                 info->bitmap = map;
4050                 map = NULL;
4051                 add_new_bitmap(ctl, info, offset);
4052                 bitmap_info = info;
4053                 info = NULL;
4054         }
4055
4056         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4057                                           trim_state);
4058
4059         bytes -= bytes_added;
4060         offset += bytes_added;
4061         spin_unlock(&ctl->tree_lock);
4062
4063         if (bytes)
4064                 goto again;
4065
4066         if (info)
4067                 kmem_cache_free(btrfs_free_space_cachep, info);
4068         if (map)
4069                 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4070         return 0;
4071 }
4072
4073 /*
4074  * Checks to see if the given range is in the free space cache.  This is really
4075  * just used to check the absence of space, so if there is free space in the
4076  * range at all we will return 1.
4077  */
4078 int test_check_exists(struct btrfs_block_group *cache,
4079                       u64 offset, u64 bytes)
4080 {
4081         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4082         struct btrfs_free_space *info;
4083         int ret = 0;
4084
4085         spin_lock(&ctl->tree_lock);
4086         info = tree_search_offset(ctl, offset, 0, 0);
4087         if (!info) {
4088                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4089                                           1, 0);
4090                 if (!info)
4091                         goto out;
4092         }
4093
4094 have_info:
4095         if (info->bitmap) {
4096                 u64 bit_off, bit_bytes;
4097                 struct rb_node *n;
4098                 struct btrfs_free_space *tmp;
4099
4100                 bit_off = offset;
4101                 bit_bytes = ctl->unit;
4102                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4103                 if (!ret) {
4104                         if (bit_off == offset) {
4105                                 ret = 1;
4106                                 goto out;
4107                         } else if (bit_off > offset &&
4108                                    offset + bytes > bit_off) {
4109                                 ret = 1;
4110                                 goto out;
4111                         }
4112                 }
4113
4114                 n = rb_prev(&info->offset_index);
4115                 while (n) {
4116                         tmp = rb_entry(n, struct btrfs_free_space,
4117                                        offset_index);
4118                         if (tmp->offset + tmp->bytes < offset)
4119                                 break;
4120                         if (offset + bytes < tmp->offset) {
4121                                 n = rb_prev(&tmp->offset_index);
4122                                 continue;
4123                         }
4124                         info = tmp;
4125                         goto have_info;
4126                 }
4127
4128                 n = rb_next(&info->offset_index);
4129                 while (n) {
4130                         tmp = rb_entry(n, struct btrfs_free_space,
4131                                        offset_index);
4132                         if (offset + bytes < tmp->offset)
4133                                 break;
4134                         if (tmp->offset + tmp->bytes < offset) {
4135                                 n = rb_next(&tmp->offset_index);
4136                                 continue;
4137                         }
4138                         info = tmp;
4139                         goto have_info;
4140                 }
4141
4142                 ret = 0;
4143                 goto out;
4144         }
4145
4146         if (info->offset == offset) {
4147                 ret = 1;
4148                 goto out;
4149         }
4150
4151         if (offset > info->offset && offset < info->offset + info->bytes)
4152                 ret = 1;
4153 out:
4154         spin_unlock(&ctl->tree_lock);
4155         return ret;
4156 }
4157 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */