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 "volumes.h"
  20 #include "space-info.h"
  21 #include "delalloc-space.h"
  22 #include "block-group.h"
  23 #include "discard.h"
  24
  25 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  26 #define MAX_CACHE_BYTES_PER_GIG SZ_64K
  27 #define FORCE_EXTENT_THRESHOLD  SZ_1M
  28
  29 struct btrfs_trim_range {
  30         u64 start;
  31         u64 bytes;
  32         struct list_head list;
  33 };
  34
  35 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  36                            struct btrfs_free_space *info);
  37 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  38                               struct btrfs_free_space *info);
  39 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
  40                          struct btrfs_free_space *bitmap_info, u64 *offset,
  41                          u64 *bytes, bool for_alloc);
  42 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
  43                         struct btrfs_free_space *bitmap_info);
  44 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
  45                               struct btrfs_free_space *info, u64 offset,
  46                               u64 bytes);
  47
  48 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  49                                                struct btrfs_path *path,
  50                                                u64 offset)
  51 {
  52         struct btrfs_fs_info *fs_info = root->fs_info;
  53         struct btrfs_key key;
  54         struct btrfs_key location;
  55         struct btrfs_disk_key disk_key;
  56         struct btrfs_free_space_header *header;
  57         struct extent_buffer *leaf;
  58         struct inode *inode = NULL;
  59         unsigned nofs_flag;
  60         int ret;
  61
  62         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  63         key.offset = offset;
  64         key.type = 0;
  65
  66         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  67         if (ret < 0)
  68                 return ERR_PTR(ret);
  69         if (ret > 0) {
  70                 btrfs_release_path(path);
  71                 return ERR_PTR(-ENOENT);
  72         }
  73
  74         leaf = path->nodes[0];
  75         header = btrfs_item_ptr(leaf, path->slots[0],
  76                                 struct btrfs_free_space_header);
  77         btrfs_free_space_key(leaf, header, &disk_key);
  78         btrfs_disk_key_to_cpu(&location, &disk_key);
  79         btrfs_release_path(path);
  80
  81         /*
  82          * We are often under a trans handle at this point, so we need to make
  83          * sure NOFS is set to keep us from deadlocking.
  84          */
  85         nofs_flag = memalloc_nofs_save();
  86         inode = btrfs_iget_path(fs_info->sb, location.objectid, root, path);
  87         btrfs_release_path(path);
  88         memalloc_nofs_restore(nofs_flag);
  89         if (IS_ERR(inode))
  90                 return inode;
  91
  92         mapping_set_gfp_mask(inode->i_mapping,
  93                         mapping_gfp_constraint(inode->i_mapping,
  94                         ~(__GFP_FS | __GFP_HIGHMEM)));
  95
  96         return inode;
  97 }
  98
  99 struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
 100                 struct btrfs_path *path)
 101 {
 102         struct btrfs_fs_info *fs_info = block_group->fs_info;
 103         struct inode *inode = NULL;
 104         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 105
 106         spin_lock(&block_group->lock);
 107         if (block_group->inode)
 108                 inode = igrab(block_group->inode);
 109         spin_unlock(&block_group->lock);
 110         if (inode)
 111                 return inode;
 112
 113         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 114                                           block_group->start);
 115         if (IS_ERR(inode))
 116                 return inode;
 117
 118         spin_lock(&block_group->lock);
 119         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 120                 btrfs_info(fs_info, "Old style space inode found, converting.");
 121                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 122                         BTRFS_INODE_NODATACOW;
 123                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 124         }
 125
 126         if (!block_group->iref) {
 127                 block_group->inode = igrab(inode);
 128                 block_group->iref = 1;
 129         }
 130         spin_unlock(&block_group->lock);
 131
 132         return inode;
 133 }
 134
 135 static int __create_free_space_inode(struct btrfs_root *root,
 136                                      struct btrfs_trans_handle *trans,
 137                                      struct btrfs_path *path,
 138                                      u64 ino, u64 offset)
 139 {
 140         struct btrfs_key key;
 141         struct btrfs_disk_key disk_key;
 142         struct btrfs_free_space_header *header;
 143         struct btrfs_inode_item *inode_item;
 144         struct extent_buffer *leaf;
 145         /* We inline CRCs for the free disk space cache */
 146         const u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC |
 147                           BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 148         int ret;
 149
 150         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 151         if (ret)
 152                 return ret;
 153
 154         leaf = path->nodes[0];
 155         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 156                                     struct btrfs_inode_item);
 157         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 158         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 159                              sizeof(*inode_item));
 160         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 161         btrfs_set_inode_size(leaf, inode_item, 0);
 162         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 163         btrfs_set_inode_uid(leaf, inode_item, 0);
 164         btrfs_set_inode_gid(leaf, inode_item, 0);
 165         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 166         btrfs_set_inode_flags(leaf, inode_item, flags);
 167         btrfs_set_inode_nlink(leaf, inode_item, 1);
 168         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 169         btrfs_set_inode_block_group(leaf, inode_item, offset);
 170         btrfs_mark_buffer_dirty(leaf);
 171         btrfs_release_path(path);
 172
 173         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 174         key.offset = offset;
 175         key.type = 0;
 176         ret = btrfs_insert_empty_item(trans, root, path, &key,
 177                                       sizeof(struct btrfs_free_space_header));
 178         if (ret < 0) {
 179                 btrfs_release_path(path);
 180                 return ret;
 181         }
 182
 183         leaf = path->nodes[0];
 184         header = btrfs_item_ptr(leaf, path->slots[0],
 185                                 struct btrfs_free_space_header);
 186         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 187         btrfs_set_free_space_key(leaf, header, &disk_key);
 188         btrfs_mark_buffer_dirty(leaf);
 189         btrfs_release_path(path);
 190
 191         return 0;
 192 }
 193
 194 int create_free_space_inode(struct btrfs_trans_handle *trans,
 195                             struct btrfs_block_group *block_group,
 196                             struct btrfs_path *path)
 197 {
 198         int ret;
 199         u64 ino;
 200
 201         ret = btrfs_find_free_objectid(trans->fs_info->tree_root, &ino);
 202         if (ret < 0)
 203                 return ret;
 204
 205         return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
 206                                          ino, block_group->start);
 207 }
 208
 209 /*
 210  * inode is an optional sink: if it is NULL, btrfs_remove_free_space_inode
 211  * handles lookup, otherwise it takes ownership and iputs the inode.
 212  * Don't reuse an inode pointer after passing it into this function.
 213  */
 214 int btrfs_remove_free_space_inode(struct btrfs_trans_handle *trans,
 215                                   struct inode *inode,
 216                                   struct btrfs_block_group *block_group)
 217 {
 218         struct btrfs_path *path;
 219         struct btrfs_key key;
 220         int ret = 0;
 221
 222         path = btrfs_alloc_path();
 223         if (!path)
 224                 return -ENOMEM;
 225
 226         if (!inode)
 227                 inode = lookup_free_space_inode(block_group, path);
 228         if (IS_ERR(inode)) {
 229                 if (PTR_ERR(inode) != -ENOENT)
 230                         ret = PTR_ERR(inode);
 231                 goto out;
 232         }
 233         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
 234         if (ret) {
 235                 btrfs_add_delayed_iput(inode);
 236                 goto out;
 237         }
 238         clear_nlink(inode);
 239         /* One for the block groups ref */
 240         spin_lock(&block_group->lock);
 241         if (block_group->iref) {
 242                 block_group->iref = 0;
 243                 block_group->inode = NULL;
 244                 spin_unlock(&block_group->lock);
 245                 iput(inode);
 246         } else {
 247                 spin_unlock(&block_group->lock);
 248         }
 249         /* One for the lookup ref */
 250         btrfs_add_delayed_iput(inode);
 251
 252         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 253         key.type = 0;
 254         key.offset = block_group->start;
 255         ret = btrfs_search_slot(trans, trans->fs_info->tree_root, &key, path,
 256                                 -1, 1);
 257         if (ret) {
 258                 if (ret > 0)
 259                         ret = 0;
 260                 goto out;
 261         }
 262         ret = btrfs_del_item(trans, trans->fs_info->tree_root, path);
 263 out:
 264         btrfs_free_path(path);
 265         return ret;
 266 }
 267
 268 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
 269                                        struct btrfs_block_rsv *rsv)
 270 {
 271         u64 needed_bytes;
 272         int ret;
 273
 274         /* 1 for slack space, 1 for updating the inode */
 275         needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
 276                 btrfs_calc_metadata_size(fs_info, 1);
 277
 278         spin_lock(&rsv->lock);
 279         if (rsv->reserved < needed_bytes)
 280                 ret = -ENOSPC;
 281         else
 282                 ret = 0;
 283         spin_unlock(&rsv->lock);
 284         return ret;
 285 }
 286
 287 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 288                                     struct btrfs_block_group *block_group,
 289                                     struct inode *inode)
 290 {
 291         struct btrfs_root *root = BTRFS_I(inode)->root;
 292         int ret = 0;
 293         bool locked = false;
 294
 295         if (block_group) {
 296                 struct btrfs_path *path = btrfs_alloc_path();
 297
 298                 if (!path) {
 299                         ret = -ENOMEM;
 300                         goto fail;
 301                 }
 302                 locked = true;
 303                 mutex_lock(&trans->transaction->cache_write_mutex);
 304                 if (!list_empty(&block_group->io_list)) {
 305                         list_del_init(&block_group->io_list);
 306
 307                         btrfs_wait_cache_io(trans, block_group, path);
 308                         btrfs_put_block_group(block_group);
 309                 }
 310
 311                 /*
 312                  * now that we've truncated the cache away, its no longer
 313                  * setup or written
 314                  */
 315                 spin_lock(&block_group->lock);
 316                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 317                 spin_unlock(&block_group->lock);
 318                 btrfs_free_path(path);
 319         }
 320
 321         btrfs_i_size_write(BTRFS_I(inode), 0);
 322         truncate_pagecache(inode, 0);
 323
 324         /*
 325          * We skip the throttling logic for free space cache inodes, so we don't
 326          * need to check for -EAGAIN.
 327          */
 328         ret = btrfs_truncate_inode_items(trans, root, BTRFS_I(inode),
 329                                          0, BTRFS_EXTENT_DATA_KEY);
 330         if (ret)
 331                 goto fail;
 332
 333         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
 334
 335 fail:
 336         if (locked)
 337                 mutex_unlock(&trans->transaction->cache_write_mutex);
 338         if (ret)
 339                 btrfs_abort_transaction(trans, ret);
 340
 341         return ret;
 342 }
 343
 344 static void readahead_cache(struct inode *inode)
 345 {
 346         struct file_ra_state *ra;
 347         unsigned long last_index;
 348
 349         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 350         if (!ra)
 351                 return;
 352
 353         file_ra_state_init(ra, inode->i_mapping);
 354         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 355
 356         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 357
 358         kfree(ra);
 359 }
 360
 361 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 362                        int write)
 363 {
 364         int num_pages;
 365
 366         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 367
 368         /* Make sure we can fit our crcs and generation into the first page */
 369         if (write && (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
 370                 return -ENOSPC;
 371
 372         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 373
 374         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 375         if (!io_ctl->pages)
 376                 return -ENOMEM;
 377
 378         io_ctl->num_pages = num_pages;
 379         io_ctl->fs_info = btrfs_sb(inode->i_sb);
 380         io_ctl->inode = inode;
 381
 382         return 0;
 383 }
 384 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 385
 386 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 387 {
 388         kfree(io_ctl->pages);
 389         io_ctl->pages = NULL;
 390 }
 391
 392 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 393 {
 394         if (io_ctl->cur) {
 395                 io_ctl->cur = NULL;
 396                 io_ctl->orig = NULL;
 397         }
 398 }
 399
 400 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 401 {
 402         ASSERT(io_ctl->index < io_ctl->num_pages);
 403         io_ctl->page = io_ctl->pages[io_ctl->index++];
 404         io_ctl->cur = page_address(io_ctl->page);
 405         io_ctl->orig = io_ctl->cur;
 406         io_ctl->size = PAGE_SIZE;
 407         if (clear)
 408                 clear_page(io_ctl->cur);
 409 }
 410
 411 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 412 {
 413         int i;
 414
 415         io_ctl_unmap_page(io_ctl);
 416
 417         for (i = 0; i < io_ctl->num_pages; i++) {
 418                 if (io_ctl->pages[i]) {
 419                         ClearPageChecked(io_ctl->pages[i]);
 420                         unlock_page(io_ctl->pages[i]);
 421                         put_page(io_ctl->pages[i]);
 422                 }
 423         }
 424 }
 425
 426 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
 427 {
 428         struct page *page;
 429         struct inode *inode = io_ctl->inode;
 430         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 431         int i;
 432
 433         for (i = 0; i < io_ctl->num_pages; i++) {
 434                 page = find_or_create_page(inode->i_mapping, i, mask);
 435                 if (!page) {
 436                         io_ctl_drop_pages(io_ctl);
 437                         return -ENOMEM;
 438                 }
 439                 io_ctl->pages[i] = page;
 440                 if (uptodate && !PageUptodate(page)) {
 441                         btrfs_readpage(NULL, page);
 442                         lock_page(page);
 443                         if (page->mapping != inode->i_mapping) {
 444                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 445                                           "free space cache page truncated");
 446                                 io_ctl_drop_pages(io_ctl);
 447                                 return -EIO;
 448                         }
 449                         if (!PageUptodate(page)) {
 450                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 451                                            "error reading free space cache");
 452                                 io_ctl_drop_pages(io_ctl);
 453                                 return -EIO;
 454                         }
 455                 }
 456         }
 457
 458         for (i = 0; i < io_ctl->num_pages; i++) {
 459                 clear_page_dirty_for_io(io_ctl->pages[i]);
 460                 set_page_extent_mapped(io_ctl->pages[i]);
 461         }
 462
 463         return 0;
 464 }
 465
 466 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 467 {
 468         io_ctl_map_page(io_ctl, 1);
 469
 470         /*
 471          * Skip the csum areas.  If we don't check crcs then we just have a
 472          * 64bit chunk at the front of the first page.
 473          */
 474         io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 475         io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 476
 477         put_unaligned_le64(generation, io_ctl->cur);
 478         io_ctl->cur += sizeof(u64);
 479 }
 480
 481 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 482 {
 483         u64 cache_gen;
 484
 485         /*
 486          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 487          * chunk at the front of the first page.
 488          */
 489         io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 490         io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 491
 492         cache_gen = get_unaligned_le64(io_ctl->cur);
 493         if (cache_gen != generation) {
 494                 btrfs_err_rl(io_ctl->fs_info,
 495                         "space cache generation (%llu) does not match inode (%llu)",
 496                                 cache_gen, generation);
 497                 io_ctl_unmap_page(io_ctl);
 498                 return -EIO;
 499         }
 500         io_ctl->cur += sizeof(u64);
 501         return 0;
 502 }
 503
 504 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 505 {
 506         u32 *tmp;
 507         u32 crc = ~(u32)0;
 508         unsigned offset = 0;
 509
 510         if (index == 0)
 511                 offset = sizeof(u32) * io_ctl->num_pages;
 512
 513         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 514         btrfs_crc32c_final(crc, (u8 *)&crc);
 515         io_ctl_unmap_page(io_ctl);
 516         tmp = page_address(io_ctl->pages[0]);
 517         tmp += index;
 518         *tmp = crc;
 519 }
 520
 521 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 522 {
 523         u32 *tmp, val;
 524         u32 crc = ~(u32)0;
 525         unsigned offset = 0;
 526
 527         if (index == 0)
 528                 offset = sizeof(u32) * io_ctl->num_pages;
 529
 530         tmp = page_address(io_ctl->pages[0]);
 531         tmp += index;
 532         val = *tmp;
 533
 534         io_ctl_map_page(io_ctl, 0);
 535         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 536         btrfs_crc32c_final(crc, (u8 *)&crc);
 537         if (val != crc) {
 538                 btrfs_err_rl(io_ctl->fs_info,
 539                         "csum mismatch on free space cache");
 540                 io_ctl_unmap_page(io_ctl);
 541                 return -EIO;
 542         }
 543
 544         return 0;
 545 }
 546
 547 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 548                             void *bitmap)
 549 {
 550         struct btrfs_free_space_entry *entry;
 551
 552         if (!io_ctl->cur)
 553                 return -ENOSPC;
 554
 555         entry = io_ctl->cur;
 556         put_unaligned_le64(offset, &entry->offset);
 557         put_unaligned_le64(bytes, &entry->bytes);
 558         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 559                 BTRFS_FREE_SPACE_EXTENT;
 560         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 561         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 562
 563         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 564                 return 0;
 565
 566         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 567
 568         /* No more pages to map */
 569         if (io_ctl->index >= io_ctl->num_pages)
 570                 return 0;
 571
 572         /* map the next page */
 573         io_ctl_map_page(io_ctl, 1);
 574         return 0;
 575 }
 576
 577 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 578 {
 579         if (!io_ctl->cur)
 580                 return -ENOSPC;
 581
 582         /*
 583          * If we aren't at the start of the current page, unmap this one and
 584          * map the next one if there is any left.
 585          */
 586         if (io_ctl->cur != io_ctl->orig) {
 587                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 588                 if (io_ctl->index >= io_ctl->num_pages)
 589                         return -ENOSPC;
 590                 io_ctl_map_page(io_ctl, 0);
 591         }
 592
 593         copy_page(io_ctl->cur, bitmap);
 594         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 595         if (io_ctl->index < io_ctl->num_pages)
 596                 io_ctl_map_page(io_ctl, 0);
 597         return 0;
 598 }
 599
 600 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 601 {
 602         /*
 603          * If we're not on the boundary we know we've modified the page and we
 604          * need to crc the page.
 605          */
 606         if (io_ctl->cur != io_ctl->orig)
 607                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 608         else
 609                 io_ctl_unmap_page(io_ctl);
 610
 611         while (io_ctl->index < io_ctl->num_pages) {
 612                 io_ctl_map_page(io_ctl, 1);
 613                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 614         }
 615 }
 616
 617 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 618                             struct btrfs_free_space *entry, u8 *type)
 619 {
 620         struct btrfs_free_space_entry *e;
 621         int ret;
 622
 623         if (!io_ctl->cur) {
 624                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 625                 if (ret)
 626                         return ret;
 627         }
 628
 629         e = io_ctl->cur;
 630         entry->offset = get_unaligned_le64(&e->offset);
 631         entry->bytes = get_unaligned_le64(&e->bytes);
 632         *type = e->type;
 633         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 634         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 635
 636         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 637                 return 0;
 638
 639         io_ctl_unmap_page(io_ctl);
 640
 641         return 0;
 642 }
 643
 644 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 645                               struct btrfs_free_space *entry)
 646 {
 647         int ret;
 648
 649         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 650         if (ret)
 651                 return ret;
 652
 653         copy_page(entry->bitmap, io_ctl->cur);
 654         io_ctl_unmap_page(io_ctl);
 655
 656         return 0;
 657 }
 658
 659 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 660 {
 661         struct btrfs_block_group *block_group = ctl->private;
 662         u64 max_bytes;
 663         u64 bitmap_bytes;
 664         u64 extent_bytes;
 665         u64 size = block_group->length;
 666         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
 667         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
 668
 669         max_bitmaps = max_t(u64, max_bitmaps, 1);
 670
 671         ASSERT(ctl->total_bitmaps <= max_bitmaps);
 672
 673         /*
 674          * We are trying to keep the total amount of memory used per 1GiB of
 675          * space to be MAX_CACHE_BYTES_PER_GIG.  However, with a reclamation
 676          * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
 677          * bitmaps, we may end up using more memory than this.
 678          */
 679         if (size < SZ_1G)
 680                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
 681         else
 682                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
 683
 684         bitmap_bytes = ctl->total_bitmaps * ctl->unit;
 685
 686         /*
 687          * we want the extent entry threshold to always be at most 1/2 the max
 688          * bytes we can have, or whatever is less than that.
 689          */
 690         extent_bytes = max_bytes - bitmap_bytes;
 691         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
 692
 693         ctl->extents_thresh =
 694                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
 695 }
 696
 697 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 698                                    struct btrfs_free_space_ctl *ctl,
 699                                    struct btrfs_path *path, u64 offset)
 700 {
 701         struct btrfs_fs_info *fs_info = root->fs_info;
 702         struct btrfs_free_space_header *header;
 703         struct extent_buffer *leaf;
 704         struct btrfs_io_ctl io_ctl;
 705         struct btrfs_key key;
 706         struct btrfs_free_space *e, *n;
 707         LIST_HEAD(bitmaps);
 708         u64 num_entries;
 709         u64 num_bitmaps;
 710         u64 generation;
 711         u8 type;
 712         int ret = 0;
 713
 714         /* Nothing in the space cache, goodbye */
 715         if (!i_size_read(inode))
 716                 return 0;
 717
 718         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 719         key.offset = offset;
 720         key.type = 0;
 721
 722         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 723         if (ret < 0)
 724                 return 0;
 725         else if (ret > 0) {
 726                 btrfs_release_path(path);
 727                 return 0;
 728         }
 729
 730         ret = -1;
 731
 732         leaf = path->nodes[0];
 733         header = btrfs_item_ptr(leaf, path->slots[0],
 734                                 struct btrfs_free_space_header);
 735         num_entries = btrfs_free_space_entries(leaf, header);
 736         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 737         generation = btrfs_free_space_generation(leaf, header);
 738         btrfs_release_path(path);
 739
 740         if (!BTRFS_I(inode)->generation) {
 741                 btrfs_info(fs_info,
 742                            "the free space cache file (%llu) is invalid, skip it",
 743                            offset);
 744                 return 0;
 745         }
 746
 747         if (BTRFS_I(inode)->generation != generation) {
 748                 btrfs_err(fs_info,
 749                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 750                           BTRFS_I(inode)->generation, generation);
 751                 return 0;
 752         }
 753
 754         if (!num_entries)
 755                 return 0;
 756
 757         ret = io_ctl_init(&io_ctl, inode, 0);
 758         if (ret)
 759                 return ret;
 760
 761         readahead_cache(inode);
 762
 763         ret = io_ctl_prepare_pages(&io_ctl, true);
 764         if (ret)
 765                 goto out;
 766
 767         ret = io_ctl_check_crc(&io_ctl, 0);
 768         if (ret)
 769                 goto free_cache;
 770
 771         ret = io_ctl_check_generation(&io_ctl, generation);
 772         if (ret)
 773                 goto free_cache;
 774
 775         while (num_entries) {
 776                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 777                                       GFP_NOFS);
 778                 if (!e)
 779                         goto free_cache;
 780
 781                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 782                 if (ret) {
 783                         kmem_cache_free(btrfs_free_space_cachep, e);
 784                         goto free_cache;
 785                 }
 786
 787                 if (!e->bytes) {
 788                         kmem_cache_free(btrfs_free_space_cachep, e);
 789                         goto free_cache;
 790                 }
 791
 792                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 793                         spin_lock(&ctl->tree_lock);
 794                         ret = link_free_space(ctl, e);
 795                         spin_unlock(&ctl->tree_lock);
 796                         if (ret) {
 797                                 btrfs_err(fs_info,
 798                                         "Duplicate entries in free space cache, dumping");
 799                                 kmem_cache_free(btrfs_free_space_cachep, e);
 800                                 goto free_cache;
 801                         }
 802                 } else {
 803                         ASSERT(num_bitmaps);
 804                         num_bitmaps--;
 805                         e->bitmap = kmem_cache_zalloc(
 806                                         btrfs_free_space_bitmap_cachep, GFP_NOFS);
 807                         if (!e->bitmap) {
 808                                 kmem_cache_free(
 809                                         btrfs_free_space_cachep, e);
 810                                 goto free_cache;
 811                         }
 812                         spin_lock(&ctl->tree_lock);
 813                         ret = link_free_space(ctl, e);
 814                         ctl->total_bitmaps++;
 815                         recalculate_thresholds(ctl);
 816                         spin_unlock(&ctl->tree_lock);
 817                         if (ret) {
 818                                 btrfs_err(fs_info,
 819                                         "Duplicate entries in free space cache, dumping");
 820                                 kmem_cache_free(btrfs_free_space_cachep, e);
 821                                 goto free_cache;
 822                         }
 823                         list_add_tail(&e->list, &bitmaps);
 824                 }
 825
 826                 num_entries--;
 827         }
 828
 829         io_ctl_unmap_page(&io_ctl);
 830
 831         /*
 832          * We add the bitmaps at the end of the entries in order that
 833          * the bitmap entries are added to the cache.
 834          */
 835         list_for_each_entry_safe(e, n, &bitmaps, list) {
 836                 list_del_init(&e->list);
 837                 ret = io_ctl_read_bitmap(&io_ctl, e);
 838                 if (ret)
 839                         goto free_cache;
 840         }
 841
 842         io_ctl_drop_pages(&io_ctl);
 843         ret = 1;
 844 out:
 845         io_ctl_free(&io_ctl);
 846         return ret;
 847 free_cache:
 848         io_ctl_drop_pages(&io_ctl);
 849         __btrfs_remove_free_space_cache(ctl);
 850         goto out;
 851 }
 852
 853 static int copy_free_space_cache(struct btrfs_block_group *block_group,
 854                                  struct btrfs_free_space_ctl *ctl)
 855 {
 856         struct btrfs_free_space *info;
 857         struct rb_node *n;
 858         int ret = 0;
 859
 860         while (!ret && (n = rb_first(&ctl->free_space_offset)) != NULL) {
 861                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 862                 if (!info->bitmap) {
 863                         unlink_free_space(ctl, info);
 864                         ret = btrfs_add_free_space(block_group, info->offset,
 865                                                    info->bytes);
 866                         kmem_cache_free(btrfs_free_space_cachep, info);
 867                 } else {
 868                         u64 offset = info->offset;
 869                         u64 bytes = ctl->unit;
 870
 871                         while (search_bitmap(ctl, info, &offset, &bytes,
 872                                              false) == 0) {
 873                                 ret = btrfs_add_free_space(block_group, offset,
 874                                                            bytes);
 875                                 if (ret)
 876                                         break;
 877                                 bitmap_clear_bits(ctl, info, offset, bytes);
 878                                 offset = info->offset;
 879                                 bytes = ctl->unit;
 880                         }
 881                         free_bitmap(ctl, info);
 882                 }
 883                 cond_resched();
 884         }
 885         return ret;
 886 }
 887
 888 int load_free_space_cache(struct btrfs_block_group *block_group)
 889 {
 890         struct btrfs_fs_info *fs_info = block_group->fs_info;
 891         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 892         struct btrfs_free_space_ctl tmp_ctl = {};
 893         struct inode *inode;
 894         struct btrfs_path *path;
 895         int ret = 0;
 896         bool matched;
 897         u64 used = block_group->used;
 898
 899         /*
 900          * Because we could potentially discard our loaded free space, we want
 901          * to load everything into a temporary structure first, and then if it's
 902          * valid copy it all into the actual free space ctl.
 903          */
 904         btrfs_init_free_space_ctl(block_group, &tmp_ctl);
 905
 906         /*
 907          * If this block group has been marked to be cleared for one reason or
 908          * another then we can't trust the on disk cache, so just return.
 909          */
 910         spin_lock(&block_group->lock);
 911         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 912                 spin_unlock(&block_group->lock);
 913                 return 0;
 914         }
 915         spin_unlock(&block_group->lock);
 916
 917         path = btrfs_alloc_path();
 918         if (!path)
 919                 return 0;
 920         path->search_commit_root = 1;
 921         path->skip_locking = 1;
 922
 923         /*
 924          * We must pass a path with search_commit_root set to btrfs_iget in
 925          * order to avoid a deadlock when allocating extents for the tree root.
 926          *
 927          * When we are COWing an extent buffer from the tree root, when looking
 928          * for a free extent, at extent-tree.c:find_free_extent(), we can find
 929          * block group without its free space cache loaded. When we find one
 930          * we must load its space cache which requires reading its free space
 931          * cache's inode item from the root tree. If this inode item is located
 932          * in the same leaf that we started COWing before, then we end up in
 933          * deadlock on the extent buffer (trying to read lock it when we
 934          * previously write locked it).
 935          *
 936          * It's safe to read the inode item using the commit root because
 937          * block groups, once loaded, stay in memory forever (until they are
 938          * removed) as well as their space caches once loaded. New block groups
 939          * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
 940          * we will never try to read their inode item while the fs is mounted.
 941          */
 942         inode = lookup_free_space_inode(block_group, path);
 943         if (IS_ERR(inode)) {
 944                 btrfs_free_path(path);
 945                 return 0;
 946         }
 947
 948         /* We may have converted the inode and made the cache invalid. */
 949         spin_lock(&block_group->lock);
 950         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 951                 spin_unlock(&block_group->lock);
 952                 btrfs_free_path(path);
 953                 goto out;
 954         }
 955         spin_unlock(&block_group->lock);
 956
 957         ret = __load_free_space_cache(fs_info->tree_root, inode, &tmp_ctl,
 958                                       path, block_group->start);
 959         btrfs_free_path(path);
 960         if (ret <= 0)
 961                 goto out;
 962
 963         matched = (tmp_ctl.free_space == (block_group->length - used -
 964                                           block_group->bytes_super));
 965
 966         if (matched) {
 967                 ret = copy_free_space_cache(block_group, &tmp_ctl);
 968                 /*
 969                  * ret == 1 means we successfully loaded the free space cache,
 970                  * so we need to re-set it here.
 971                  */
 972                 if (ret == 0)
 973                         ret = 1;
 974         } else {
 975                 __btrfs_remove_free_space_cache(&tmp_ctl);
 976                 btrfs_warn(fs_info,
 977                            "block group %llu has wrong amount of free space",
 978                            block_group->start);
 979                 ret = -1;
 980         }
 981 out:
 982         if (ret < 0) {
 983                 /* This cache is bogus, make sure it gets cleared */
 984                 spin_lock(&block_group->lock);
 985                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 986                 spin_unlock(&block_group->lock);
 987                 ret = 0;
 988
 989                 btrfs_warn(fs_info,
 990                            "failed to load free space cache for block group %llu, rebuilding it now",
 991                            block_group->start);
 992         }
 993
 994         spin_lock(&ctl->tree_lock);
 995         btrfs_discard_update_discardable(block_group);
 996         spin_unlock(&ctl->tree_lock);
 997         iput(inode);
 998         return ret;
 999 }
1000
1001 static noinline_for_stack
1002 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
1003                               struct btrfs_free_space_ctl *ctl,
1004                               struct btrfs_block_group *block_group,
1005                               int *entries, int *bitmaps,
1006                               struct list_head *bitmap_list)
1007 {
1008         int ret;
1009         struct btrfs_free_cluster *cluster = NULL;
1010         struct btrfs_free_cluster *cluster_locked = NULL;
1011         struct rb_node *node = rb_first(&ctl->free_space_offset);
1012         struct btrfs_trim_range *trim_entry;
1013
1014         /* Get the cluster for this block_group if it exists */
1015         if (block_group && !list_empty(&block_group->cluster_list)) {
1016                 cluster = list_entry(block_group->cluster_list.next,
1017                                      struct btrfs_free_cluster,
1018                                      block_group_list);
1019         }
1020
1021         if (!node && cluster) {
1022                 cluster_locked = cluster;
1023                 spin_lock(&cluster_locked->lock);
1024                 node = rb_first(&cluster->root);
1025                 cluster = NULL;
1026         }
1027
1028         /* Write out the extent entries */
1029         while (node) {
1030                 struct btrfs_free_space *e;
1031
1032                 e = rb_entry(node, struct btrfs_free_space, offset_index);
1033                 *entries += 1;
1034
1035                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
1036                                        e->bitmap);
1037                 if (ret)
1038                         goto fail;
1039
1040                 if (e->bitmap) {
1041                         list_add_tail(&e->list, bitmap_list);
1042                         *bitmaps += 1;
1043                 }
1044                 node = rb_next(node);
1045                 if (!node && cluster) {
1046                         node = rb_first(&cluster->root);
1047                         cluster_locked = cluster;
1048                         spin_lock(&cluster_locked->lock);
1049                         cluster = NULL;
1050                 }
1051         }
1052         if (cluster_locked) {
1053                 spin_unlock(&cluster_locked->lock);
1054                 cluster_locked = NULL;
1055         }
1056
1057         /*
1058          * Make sure we don't miss any range that was removed from our rbtree
1059          * because trimming is running. Otherwise after a umount+mount (or crash
1060          * after committing the transaction) we would leak free space and get
1061          * an inconsistent free space cache report from fsck.
1062          */
1063         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1064                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1065                                        trim_entry->bytes, NULL);
1066                 if (ret)
1067                         goto fail;
1068                 *entries += 1;
1069         }
1070
1071         return 0;
1072 fail:
1073         if (cluster_locked)
1074                 spin_unlock(&cluster_locked->lock);
1075         return -ENOSPC;
1076 }
1077
1078 static noinline_for_stack int
1079 update_cache_item(struct btrfs_trans_handle *trans,
1080                   struct btrfs_root *root,
1081                   struct inode *inode,
1082                   struct btrfs_path *path, u64 offset,
1083                   int entries, int bitmaps)
1084 {
1085         struct btrfs_key key;
1086         struct btrfs_free_space_header *header;
1087         struct extent_buffer *leaf;
1088         int ret;
1089
1090         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1091         key.offset = offset;
1092         key.type = 0;
1093
1094         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1095         if (ret < 0) {
1096                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1097                                  EXTENT_DELALLOC, 0, 0, NULL);
1098                 goto fail;
1099         }
1100         leaf = path->nodes[0];
1101         if (ret > 0) {
1102                 struct btrfs_key found_key;
1103                 ASSERT(path->slots[0]);
1104                 path->slots[0]--;
1105                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1106                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1107                     found_key.offset != offset) {
1108                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1109                                          inode->i_size - 1, EXTENT_DELALLOC, 0,
1110                                          0, NULL);
1111                         btrfs_release_path(path);
1112                         goto fail;
1113                 }
1114         }
1115
1116         BTRFS_I(inode)->generation = trans->transid;
1117         header = btrfs_item_ptr(leaf, path->slots[0],
1118                                 struct btrfs_free_space_header);
1119         btrfs_set_free_space_entries(leaf, header, entries);
1120         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1121         btrfs_set_free_space_generation(leaf, header, trans->transid);
1122         btrfs_mark_buffer_dirty(leaf);
1123         btrfs_release_path(path);
1124
1125         return 0;
1126
1127 fail:
1128         return -1;
1129 }
1130
1131 static noinline_for_stack int write_pinned_extent_entries(
1132                             struct btrfs_trans_handle *trans,
1133                             struct btrfs_block_group *block_group,
1134                             struct btrfs_io_ctl *io_ctl,
1135                             int *entries)
1136 {
1137         u64 start, extent_start, extent_end, len;
1138         struct extent_io_tree *unpin = NULL;
1139         int ret;
1140
1141         if (!block_group)
1142                 return 0;
1143
1144         /*
1145          * We want to add any pinned extents to our free space cache
1146          * so we don't leak the space
1147          *
1148          * We shouldn't have switched the pinned extents yet so this is the
1149          * right one
1150          */
1151         unpin = &trans->transaction->pinned_extents;
1152
1153         start = block_group->start;
1154
1155         while (start < block_group->start + block_group->length) {
1156                 ret = find_first_extent_bit(unpin, start,
1157                                             &extent_start, &extent_end,
1158                                             EXTENT_DIRTY, NULL);
1159                 if (ret)
1160                         return 0;
1161
1162                 /* This pinned extent is out of our range */
1163                 if (extent_start >= block_group->start + block_group->length)
1164                         return 0;
1165
1166                 extent_start = max(extent_start, start);
1167                 extent_end = min(block_group->start + block_group->length,
1168                                  extent_end + 1);
1169                 len = extent_end - extent_start;
1170
1171                 *entries += 1;
1172                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1173                 if (ret)
1174                         return -ENOSPC;
1175
1176                 start = extent_end;
1177         }
1178
1179         return 0;
1180 }
1181
1182 static noinline_for_stack int
1183 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1184 {
1185         struct btrfs_free_space *entry, *next;
1186         int ret;
1187
1188         /* Write out the bitmaps */
1189         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1190                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1191                 if (ret)
1192                         return -ENOSPC;
1193                 list_del_init(&entry->list);
1194         }
1195
1196         return 0;
1197 }
1198
1199 static int flush_dirty_cache(struct inode *inode)
1200 {
1201         int ret;
1202
1203         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1204         if (ret)
1205                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1206                                  EXTENT_DELALLOC, 0, 0, NULL);
1207
1208         return ret;
1209 }
1210
1211 static void noinline_for_stack
1212 cleanup_bitmap_list(struct list_head *bitmap_list)
1213 {
1214         struct btrfs_free_space *entry, *next;
1215
1216         list_for_each_entry_safe(entry, next, bitmap_list, list)
1217                 list_del_init(&entry->list);
1218 }
1219
1220 static void noinline_for_stack
1221 cleanup_write_cache_enospc(struct inode *inode,
1222                            struct btrfs_io_ctl *io_ctl,
1223                            struct extent_state **cached_state)
1224 {
1225         io_ctl_drop_pages(io_ctl);
1226         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1227                              i_size_read(inode) - 1, cached_state);
1228 }
1229
1230 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1231                                  struct btrfs_trans_handle *trans,
1232                                  struct btrfs_block_group *block_group,
1233                                  struct btrfs_io_ctl *io_ctl,
1234                                  struct btrfs_path *path, u64 offset)
1235 {
1236         int ret;
1237         struct inode *inode = io_ctl->inode;
1238
1239         if (!inode)
1240                 return 0;
1241
1242         /* Flush the dirty pages in the cache file. */
1243         ret = flush_dirty_cache(inode);
1244         if (ret)
1245                 goto out;
1246
1247         /* Update the cache item to tell everyone this cache file is valid. */
1248         ret = update_cache_item(trans, root, inode, path, offset,
1249                                 io_ctl->entries, io_ctl->bitmaps);
1250 out:
1251         if (ret) {
1252                 invalidate_inode_pages2(inode->i_mapping);
1253                 BTRFS_I(inode)->generation = 0;
1254                 if (block_group)
1255                         btrfs_debug(root->fs_info,
1256           "failed to write free space cache for block group %llu error %d",
1257                                   block_group->start, ret);
1258         }
1259         btrfs_update_inode(trans, root, BTRFS_I(inode));
1260
1261         if (block_group) {
1262                 /* the dirty list is protected by the dirty_bgs_lock */
1263                 spin_lock(&trans->transaction->dirty_bgs_lock);
1264
1265                 /* the disk_cache_state is protected by the block group lock */
1266                 spin_lock(&block_group->lock);
1267
1268                 /*
1269                  * only mark this as written if we didn't get put back on
1270                  * the dirty list while waiting for IO.   Otherwise our
1271                  * cache state won't be right, and we won't get written again
1272                  */
1273                 if (!ret && list_empty(&block_group->dirty_list))
1274                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1275                 else if (ret)
1276                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1277
1278                 spin_unlock(&block_group->lock);
1279                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1280                 io_ctl->inode = NULL;
1281                 iput(inode);
1282         }
1283
1284         return ret;
1285
1286 }
1287
1288 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1289                         struct btrfs_block_group *block_group,
1290                         struct btrfs_path *path)
1291 {
1292         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1293                                      block_group, &block_group->io_ctl,
1294                                      path, block_group->start);
1295 }
1296
1297 /**
1298  * __btrfs_write_out_cache - write out cached info to an inode
1299  * @root - the root the inode belongs to
1300  * @ctl - the free space cache we are going to write out
1301  * @block_group - the block_group for this cache if it belongs to a block_group
1302  * @trans - the trans handle
1303  *
1304  * This function writes out a free space cache struct to disk for quick recovery
1305  * on mount.  This will return 0 if it was successful in writing the cache out,
1306  * or an errno if it was not.
1307  */
1308 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1309                                    struct btrfs_free_space_ctl *ctl,
1310                                    struct btrfs_block_group *block_group,
1311                                    struct btrfs_io_ctl *io_ctl,
1312                                    struct btrfs_trans_handle *trans)
1313 {
1314         struct extent_state *cached_state = NULL;
1315         LIST_HEAD(bitmap_list);
1316         int entries = 0;
1317         int bitmaps = 0;
1318         int ret;
1319         int must_iput = 0;
1320
1321         if (!i_size_read(inode))
1322                 return -EIO;
1323
1324         WARN_ON(io_ctl->pages);
1325         ret = io_ctl_init(io_ctl, inode, 1);
1326         if (ret)
1327                 return ret;
1328
1329         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1330                 down_write(&block_group->data_rwsem);
1331                 spin_lock(&block_group->lock);
1332                 if (block_group->delalloc_bytes) {
1333                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1334                         spin_unlock(&block_group->lock);
1335                         up_write(&block_group->data_rwsem);
1336                         BTRFS_I(inode)->generation = 0;
1337                         ret = 0;
1338                         must_iput = 1;
1339                         goto out;
1340                 }
1341                 spin_unlock(&block_group->lock);
1342         }
1343
1344         /* Lock all pages first so we can lock the extent safely. */
1345         ret = io_ctl_prepare_pages(io_ctl, false);
1346         if (ret)
1347                 goto out_unlock;
1348
1349         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1350                          &cached_state);
1351
1352         io_ctl_set_generation(io_ctl, trans->transid);
1353
1354         mutex_lock(&ctl->cache_writeout_mutex);
1355         /* Write out the extent entries in the free space cache */
1356         spin_lock(&ctl->tree_lock);
1357         ret = write_cache_extent_entries(io_ctl, ctl,
1358                                          block_group, &entries, &bitmaps,
1359                                          &bitmap_list);
1360         if (ret)
1361                 goto out_nospc_locked;
1362
1363         /*
1364          * Some spaces that are freed in the current transaction are pinned,
1365          * they will be added into free space cache after the transaction is
1366          * committed, we shouldn't lose them.
1367          *
1368          * If this changes while we are working we'll get added back to
1369          * the dirty list and redo it.  No locking needed
1370          */
1371         ret = write_pinned_extent_entries(trans, block_group, io_ctl, &entries);
1372         if (ret)
1373                 goto out_nospc_locked;
1374
1375         /*
1376          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1377          * locked while doing it because a concurrent trim can be manipulating
1378          * or freeing the bitmap.
1379          */
1380         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1381         spin_unlock(&ctl->tree_lock);
1382         mutex_unlock(&ctl->cache_writeout_mutex);
1383         if (ret)
1384                 goto out_nospc;
1385
1386         /* Zero out the rest of the pages just to make sure */
1387         io_ctl_zero_remaining_pages(io_ctl);
1388
1389         /* Everything is written out, now we dirty the pages in the file. */
1390         ret = btrfs_dirty_pages(BTRFS_I(inode), io_ctl->pages,
1391                                 io_ctl->num_pages, 0, i_size_read(inode),
1392                                 &cached_state, false);
1393         if (ret)
1394                 goto out_nospc;
1395
1396         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1397                 up_write(&block_group->data_rwsem);
1398         /*
1399          * Release the pages and unlock the extent, we will flush
1400          * them out later
1401          */
1402         io_ctl_drop_pages(io_ctl);
1403         io_ctl_free(io_ctl);
1404
1405         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1406                              i_size_read(inode) - 1, &cached_state);
1407
1408         /*
1409          * at this point the pages are under IO and we're happy,
1410          * The caller is responsible for waiting on them and updating
1411          * the cache and the inode
1412          */
1413         io_ctl->entries = entries;
1414         io_ctl->bitmaps = bitmaps;
1415
1416         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1417         if (ret)
1418                 goto out;
1419
1420         return 0;
1421
1422 out_nospc_locked:
1423         cleanup_bitmap_list(&bitmap_list);
1424         spin_unlock(&ctl->tree_lock);
1425         mutex_unlock(&ctl->cache_writeout_mutex);
1426
1427 out_nospc:
1428         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1429
1430 out_unlock:
1431         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1432                 up_write(&block_group->data_rwsem);
1433
1434 out:
1435         io_ctl->inode = NULL;
1436         io_ctl_free(io_ctl);
1437         if (ret) {
1438                 invalidate_inode_pages2(inode->i_mapping);
1439                 BTRFS_I(inode)->generation = 0;
1440         }
1441         btrfs_update_inode(trans, root, BTRFS_I(inode));
1442         if (must_iput)
1443                 iput(inode);
1444         return ret;
1445 }
1446
1447 int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1448                           struct btrfs_block_group *block_group,
1449                           struct btrfs_path *path)
1450 {
1451         struct btrfs_fs_info *fs_info = trans->fs_info;
1452         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1453         struct inode *inode;
1454         int ret = 0;
1455
1456         spin_lock(&block_group->lock);
1457         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1458                 spin_unlock(&block_group->lock);
1459                 return 0;
1460         }
1461         spin_unlock(&block_group->lock);
1462
1463         inode = lookup_free_space_inode(block_group, path);
1464         if (IS_ERR(inode))
1465                 return 0;
1466
1467         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1468                                 block_group, &block_group->io_ctl, trans);
1469         if (ret) {
1470                 btrfs_debug(fs_info,
1471           "failed to write free space cache for block group %llu error %d",
1472                           block_group->start, ret);
1473                 spin_lock(&block_group->lock);
1474                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1475                 spin_unlock(&block_group->lock);
1476
1477                 block_group->io_ctl.inode = NULL;
1478                 iput(inode);
1479         }
1480
1481         /*
1482          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1483          * to wait for IO and put the inode
1484          */
1485
1486         return ret;
1487 }
1488
1489 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1490                                           u64 offset)
1491 {
1492         ASSERT(offset >= bitmap_start);
1493         offset -= bitmap_start;
1494         return (unsigned long)(div_u64(offset, unit));
1495 }
1496
1497 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1498 {
1499         return (unsigned long)(div_u64(bytes, unit));
1500 }
1501
1502 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1503                                    u64 offset)
1504 {
1505         u64 bitmap_start;
1506         u64 bytes_per_bitmap;
1507
1508         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1509         bitmap_start = offset - ctl->start;
1510         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1511         bitmap_start *= bytes_per_bitmap;
1512         bitmap_start += ctl->start;
1513
1514         return bitmap_start;
1515 }
1516
1517 static int tree_insert_offset(struct rb_root *root, u64 offset,
1518                               struct rb_node *node, int bitmap)
1519 {
1520         struct rb_node **p = &root->rb_node;
1521         struct rb_node *parent = NULL;
1522         struct btrfs_free_space *info;
1523
1524         while (*p) {
1525                 parent = *p;
1526                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1527
1528                 if (offset < info->offset) {
1529                         p = &(*p)->rb_left;
1530                 } else if (offset > info->offset) {
1531                         p = &(*p)->rb_right;
1532                 } else {
1533                         /*
1534                          * we could have a bitmap entry and an extent entry
1535                          * share the same offset.  If this is the case, we want
1536                          * the extent entry to always be found first if we do a
1537                          * linear search through the tree, since we want to have
1538                          * the quickest allocation time, and allocating from an
1539                          * extent is faster than allocating from a bitmap.  So
1540                          * if we're inserting a bitmap and we find an entry at
1541                          * this offset, we want to go right, or after this entry
1542                          * logically.  If we are inserting an extent and we've
1543                          * found a bitmap, we want to go left, or before
1544                          * logically.
1545                          */
1546                         if (bitmap) {
1547                                 if (info->bitmap) {
1548                                         WARN_ON_ONCE(1);
1549                                         return -EEXIST;
1550                                 }
1551                                 p = &(*p)->rb_right;
1552                         } else {
1553                                 if (!info->bitmap) {
1554                                         WARN_ON_ONCE(1);
1555                                         return -EEXIST;
1556                                 }
1557                                 p = &(*p)->rb_left;
1558                         }
1559                 }
1560         }
1561
1562         rb_link_node(node, parent, p);
1563         rb_insert_color(node, root);
1564
1565         return 0;
1566 }
1567
1568 /*
1569  * searches the tree for the given offset.
1570  *
1571  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1572  * want a section that has at least bytes size and comes at or after the given
1573  * offset.
1574  */
1575 static struct btrfs_free_space *
1576 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1577                    u64 offset, int bitmap_only, int fuzzy)
1578 {
1579         struct rb_node *n = ctl->free_space_offset.rb_node;
1580         struct btrfs_free_space *entry, *prev = NULL;
1581
1582         /* find entry that is closest to the 'offset' */
1583         while (1) {
1584                 if (!n) {
1585                         entry = NULL;
1586                         break;
1587                 }
1588
1589                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1590                 prev = entry;
1591
1592                 if (offset < entry->offset)
1593                         n = n->rb_left;
1594                 else if (offset > entry->offset)
1595                         n = n->rb_right;
1596                 else
1597                         break;
1598         }
1599
1600         if (bitmap_only) {
1601                 if (!entry)
1602                         return NULL;
1603                 if (entry->bitmap)
1604                         return entry;
1605
1606                 /*
1607                  * bitmap entry and extent entry may share same offset,
1608                  * in that case, bitmap entry comes after extent entry.
1609                  */
1610                 n = rb_next(n);
1611                 if (!n)
1612                         return NULL;
1613                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1614                 if (entry->offset != offset)
1615                         return NULL;
1616
1617                 WARN_ON(!entry->bitmap);
1618                 return entry;
1619         } else if (entry) {
1620                 if (entry->bitmap) {
1621                         /*
1622                          * if previous extent entry covers the offset,
1623                          * we should return it instead of the bitmap entry
1624                          */
1625                         n = rb_prev(&entry->offset_index);
1626                         if (n) {
1627                                 prev = rb_entry(n, struct btrfs_free_space,
1628                                                 offset_index);
1629                                 if (!prev->bitmap &&
1630                                     prev->offset + prev->bytes > offset)
1631                                         entry = prev;
1632                         }
1633                 }
1634                 return entry;
1635         }
1636
1637         if (!prev)
1638                 return NULL;
1639
1640         /* find last entry before the 'offset' */
1641         entry = prev;
1642         if (entry->offset > offset) {
1643                 n = rb_prev(&entry->offset_index);
1644                 if (n) {
1645                         entry = rb_entry(n, struct btrfs_free_space,
1646                                         offset_index);
1647                         ASSERT(entry->offset <= offset);
1648                 } else {
1649                         if (fuzzy)
1650                                 return entry;
1651                         else
1652                                 return NULL;
1653                 }
1654         }
1655
1656         if (entry->bitmap) {
1657                 n = rb_prev(&entry->offset_index);
1658                 if (n) {
1659                         prev = rb_entry(n, struct btrfs_free_space,
1660                                         offset_index);
1661                         if (!prev->bitmap &&
1662                             prev->offset + prev->bytes > offset)
1663                                 return prev;
1664                 }
1665                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1666                         return entry;
1667         } else if (entry->offset + entry->bytes > offset)
1668                 return entry;
1669
1670         if (!fuzzy)
1671                 return NULL;
1672
1673         while (1) {
1674                 if (entry->bitmap) {
1675                         if (entry->offset + BITS_PER_BITMAP *
1676                             ctl->unit > offset)
1677                                 break;
1678                 } else {
1679                         if (entry->offset + entry->bytes > offset)
1680                                 break;
1681                 }
1682
1683                 n = rb_next(&entry->offset_index);
1684                 if (!n)
1685                         return NULL;
1686                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1687         }
1688         return entry;
1689 }
1690
1691 static inline void
1692 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1693                     struct btrfs_free_space *info)
1694 {
1695         rb_erase(&info->offset_index, &ctl->free_space_offset);
1696         ctl->free_extents--;
1697
1698         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1699                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1700                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1701         }
1702 }
1703
1704 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1705                               struct btrfs_free_space *info)
1706 {
1707         __unlink_free_space(ctl, info);
1708         ctl->free_space -= info->bytes;
1709 }
1710
1711 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1712                            struct btrfs_free_space *info)
1713 {
1714         int ret = 0;
1715
1716         ASSERT(info->bytes || info->bitmap);
1717         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1718                                  &info->offset_index, (info->bitmap != NULL));
1719         if (ret)
1720                 return ret;
1721
1722         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1723                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1724                 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1725         }
1726
1727         ctl->free_space += info->bytes;
1728         ctl->free_extents++;
1729         return ret;
1730 }
1731
1732 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1733                                        struct btrfs_free_space *info,
1734                                        u64 offset, u64 bytes)
1735 {
1736         unsigned long start, count, end;
1737         int extent_delta = -1;
1738
1739         start = offset_to_bit(info->offset, ctl->unit, offset);
1740         count = bytes_to_bits(bytes, ctl->unit);
1741         end = start + count;
1742         ASSERT(end <= BITS_PER_BITMAP);
1743
1744         bitmap_clear(info->bitmap, start, count);
1745
1746         info->bytes -= bytes;
1747         if (info->max_extent_size > ctl->unit)
1748                 info->max_extent_size = 0;
1749
1750         if (start && test_bit(start - 1, info->bitmap))
1751                 extent_delta++;
1752
1753         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1754                 extent_delta++;
1755
1756         info->bitmap_extents += extent_delta;
1757         if (!btrfs_free_space_trimmed(info)) {
1758                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1759                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1760         }
1761 }
1762
1763 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1764                               struct btrfs_free_space *info, u64 offset,
1765                               u64 bytes)
1766 {
1767         __bitmap_clear_bits(ctl, info, offset, bytes);
1768         ctl->free_space -= bytes;
1769 }
1770
1771 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1772                             struct btrfs_free_space *info, u64 offset,
1773                             u64 bytes)
1774 {
1775         unsigned long start, count, end;
1776         int extent_delta = 1;
1777
1778         start = offset_to_bit(info->offset, ctl->unit, offset);
1779         count = bytes_to_bits(bytes, ctl->unit);
1780         end = start + count;
1781         ASSERT(end <= BITS_PER_BITMAP);
1782
1783         bitmap_set(info->bitmap, start, count);
1784
1785         info->bytes += bytes;
1786         ctl->free_space += bytes;
1787
1788         if (start && test_bit(start - 1, info->bitmap))
1789                 extent_delta--;
1790
1791         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1792                 extent_delta--;
1793
1794         info->bitmap_extents += extent_delta;
1795         if (!btrfs_free_space_trimmed(info)) {
1796                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1797                 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1798         }
1799 }
1800
1801 /*
1802  * If we can not find suitable extent, we will use bytes to record
1803  * the size of the max extent.
1804  */
1805 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1806                          struct btrfs_free_space *bitmap_info, u64 *offset,
1807                          u64 *bytes, bool for_alloc)
1808 {
1809         unsigned long found_bits = 0;
1810         unsigned long max_bits = 0;
1811         unsigned long bits, i;
1812         unsigned long next_zero;
1813         unsigned long extent_bits;
1814
1815         /*
1816          * Skip searching the bitmap if we don't have a contiguous section that
1817          * is large enough for this allocation.
1818          */
1819         if (for_alloc &&
1820             bitmap_info->max_extent_size &&
1821             bitmap_info->max_extent_size < *bytes) {
1822                 *bytes = bitmap_info->max_extent_size;
1823                 return -1;
1824         }
1825
1826         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1827                           max_t(u64, *offset, bitmap_info->offset));
1828         bits = bytes_to_bits(*bytes, ctl->unit);
1829
1830         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1831                 if (for_alloc && bits == 1) {
1832                         found_bits = 1;
1833                         break;
1834                 }
1835                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1836                                                BITS_PER_BITMAP, i);
1837                 extent_bits = next_zero - i;
1838                 if (extent_bits >= bits) {
1839                         found_bits = extent_bits;
1840                         break;
1841                 } else if (extent_bits > max_bits) {
1842                         max_bits = extent_bits;
1843                 }
1844                 i = next_zero;
1845         }
1846
1847         if (found_bits) {
1848                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1849                 *bytes = (u64)(found_bits) * ctl->unit;
1850                 return 0;
1851         }
1852
1853         *bytes = (u64)(max_bits) * ctl->unit;
1854         bitmap_info->max_extent_size = *bytes;
1855         return -1;
1856 }
1857
1858 static inline u64 get_max_extent_size(struct btrfs_free_space *entry)
1859 {
1860         if (entry->bitmap)
1861                 return entry->max_extent_size;
1862         return entry->bytes;
1863 }
1864
1865 /* Cache the size of the max extent in bytes */
1866 static struct btrfs_free_space *
1867 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1868                 unsigned long align, u64 *max_extent_size)
1869 {
1870         struct btrfs_free_space *entry;
1871         struct rb_node *node;
1872         u64 tmp;
1873         u64 align_off;
1874         int ret;
1875
1876         if (!ctl->free_space_offset.rb_node)
1877                 goto out;
1878
1879         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1880         if (!entry)
1881                 goto out;
1882
1883         for (node = &entry->offset_index; node; node = rb_next(node)) {
1884                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1885                 if (entry->bytes < *bytes) {
1886                         *max_extent_size = max(get_max_extent_size(entry),
1887                                                *max_extent_size);
1888                         continue;
1889                 }
1890
1891                 /* make sure the space returned is big enough
1892                  * to match our requested alignment
1893                  */
1894                 if (*bytes >= align) {
1895                         tmp = entry->offset - ctl->start + align - 1;
1896                         tmp = div64_u64(tmp, align);
1897                         tmp = tmp * align + ctl->start;
1898                         align_off = tmp - entry->offset;
1899                 } else {
1900                         align_off = 0;
1901                         tmp = entry->offset;
1902                 }
1903
1904                 if (entry->bytes < *bytes + align_off) {
1905                         *max_extent_size = max(get_max_extent_size(entry),
1906                                                *max_extent_size);
1907                         continue;
1908                 }
1909
1910                 if (entry->bitmap) {
1911                         u64 size = *bytes;
1912
1913                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1914                         if (!ret) {
1915                                 *offset = tmp;
1916                                 *bytes = size;
1917                                 return entry;
1918                         } else {
1919                                 *max_extent_size =
1920                                         max(get_max_extent_size(entry),
1921                                             *max_extent_size);
1922                         }
1923                         continue;
1924                 }
1925
1926                 *offset = tmp;
1927                 *bytes = entry->bytes - align_off;
1928                 return entry;
1929         }
1930 out:
1931         return NULL;
1932 }
1933
1934 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1935                            struct btrfs_free_space *info, u64 offset)
1936 {
1937         info->offset = offset_to_bitmap(ctl, offset);
1938         info->bytes = 0;
1939         info->bitmap_extents = 0;
1940         INIT_LIST_HEAD(&info->list);
1941         link_free_space(ctl, info);
1942         ctl->total_bitmaps++;
1943         recalculate_thresholds(ctl);
1944 }
1945
1946 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1947                         struct btrfs_free_space *bitmap_info)
1948 {
1949         /*
1950          * Normally when this is called, the bitmap is completely empty. However,
1951          * if we are blowing up the free space cache for one reason or another
1952          * via __btrfs_remove_free_space_cache(), then it may not be freed and
1953          * we may leave stats on the table.
1954          */
1955         if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
1956                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
1957                         bitmap_info->bitmap_extents;
1958                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
1959
1960         }
1961         unlink_free_space(ctl, bitmap_info);
1962         kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
1963         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1964         ctl->total_bitmaps--;
1965         recalculate_thresholds(ctl);
1966 }
1967
1968 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1969                               struct btrfs_free_space *bitmap_info,
1970                               u64 *offset, u64 *bytes)
1971 {
1972         u64 end;
1973         u64 search_start, search_bytes;
1974         int ret;
1975
1976 again:
1977         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1978
1979         /*
1980          * We need to search for bits in this bitmap.  We could only cover some
1981          * of the extent in this bitmap thanks to how we add space, so we need
1982          * to search for as much as it as we can and clear that amount, and then
1983          * go searching for the next bit.
1984          */
1985         search_start = *offset;
1986         search_bytes = ctl->unit;
1987         search_bytes = min(search_bytes, end - search_start + 1);
1988         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1989                             false);
1990         if (ret < 0 || search_start != *offset)
1991                 return -EINVAL;
1992
1993         /* We may have found more bits than what we need */
1994         search_bytes = min(search_bytes, *bytes);
1995
1996         /* Cannot clear past the end of the bitmap */
1997         search_bytes = min(search_bytes, end - search_start + 1);
1998
1999         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
2000         *offset += search_bytes;
2001         *bytes -= search_bytes;
2002
2003         if (*bytes) {
2004                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2005                 if (!bitmap_info->bytes)
2006                         free_bitmap(ctl, bitmap_info);
2007
2008                 /*
2009                  * no entry after this bitmap, but we still have bytes to
2010                  * remove, so something has gone wrong.
2011                  */
2012                 if (!next)
2013                         return -EINVAL;
2014
2015                 bitmap_info = rb_entry(next, struct btrfs_free_space,
2016                                        offset_index);
2017
2018                 /*
2019                  * if the next entry isn't a bitmap we need to return to let the
2020                  * extent stuff do its work.
2021                  */
2022                 if (!bitmap_info->bitmap)
2023                         return -EAGAIN;
2024
2025                 /*
2026                  * Ok the next item is a bitmap, but it may not actually hold
2027                  * the information for the rest of this free space stuff, so
2028                  * look for it, and if we don't find it return so we can try
2029                  * everything over again.
2030                  */
2031                 search_start = *offset;
2032                 search_bytes = ctl->unit;
2033                 ret = search_bitmap(ctl, bitmap_info, &search_start,
2034                                     &search_bytes, false);
2035                 if (ret < 0 || search_start != *offset)
2036                         return -EAGAIN;
2037
2038                 goto again;
2039         } else if (!bitmap_info->bytes)
2040                 free_bitmap(ctl, bitmap_info);
2041
2042         return 0;
2043 }
2044
2045 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2046                                struct btrfs_free_space *info, u64 offset,
2047                                u64 bytes, enum btrfs_trim_state trim_state)
2048 {
2049         u64 bytes_to_set = 0;
2050         u64 end;
2051
2052         /*
2053          * This is a tradeoff to make bitmap trim state minimal.  We mark the
2054          * whole bitmap untrimmed if at any point we add untrimmed regions.
2055          */
2056         if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2057                 if (btrfs_free_space_trimmed(info)) {
2058                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
2059                                 info->bitmap_extents;
2060                         ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2061                 }
2062                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2063         }
2064
2065         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2066
2067         bytes_to_set = min(end - offset, bytes);
2068
2069         bitmap_set_bits(ctl, info, offset, bytes_to_set);
2070
2071         /*
2072          * We set some bytes, we have no idea what the max extent size is
2073          * anymore.
2074          */
2075         info->max_extent_size = 0;
2076
2077         return bytes_to_set;
2078
2079 }
2080
2081 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2082                       struct btrfs_free_space *info)
2083 {
2084         struct btrfs_block_group *block_group = ctl->private;
2085         struct btrfs_fs_info *fs_info = block_group->fs_info;
2086         bool forced = false;
2087
2088 #ifdef CONFIG_BTRFS_DEBUG
2089         if (btrfs_should_fragment_free_space(block_group))
2090                 forced = true;
2091 #endif
2092
2093         /* This is a way to reclaim large regions from the bitmaps. */
2094         if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2095                 return false;
2096
2097         /*
2098          * If we are below the extents threshold then we can add this as an
2099          * extent, and don't have to deal with the bitmap
2100          */
2101         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2102                 /*
2103                  * If this block group has some small extents we don't want to
2104                  * use up all of our free slots in the cache with them, we want
2105                  * to reserve them to larger extents, however if we have plenty
2106                  * of cache left then go ahead an dadd them, no sense in adding
2107                  * the overhead of a bitmap if we don't have to.
2108                  */
2109                 if (info->bytes <= fs_info->sectorsize * 8) {
2110                         if (ctl->free_extents * 3 <= ctl->extents_thresh)
2111                                 return false;
2112                 } else {
2113                         return false;
2114                 }
2115         }
2116
2117         /*
2118          * The original block groups from mkfs can be really small, like 8
2119          * megabytes, so don't bother with a bitmap for those entries.  However
2120          * some block groups can be smaller than what a bitmap would cover but
2121          * are still large enough that they could overflow the 32k memory limit,
2122          * so allow those block groups to still be allowed to have a bitmap
2123          * entry.
2124          */
2125         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2126                 return false;
2127
2128         return true;
2129 }
2130
2131 static const struct btrfs_free_space_op free_space_op = {
2132         .use_bitmap             = use_bitmap,
2133 };
2134
2135 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2136                               struct btrfs_free_space *info)
2137 {
2138         struct btrfs_free_space *bitmap_info;
2139         struct btrfs_block_group *block_group = NULL;
2140         int added = 0;
2141         u64 bytes, offset, bytes_added;
2142         enum btrfs_trim_state trim_state;
2143         int ret;
2144
2145         bytes = info->bytes;
2146         offset = info->offset;
2147         trim_state = info->trim_state;
2148
2149         if (!ctl->op->use_bitmap(ctl, info))
2150                 return 0;
2151
2152         if (ctl->op == &free_space_op)
2153                 block_group = ctl->private;
2154 again:
2155         /*
2156          * Since we link bitmaps right into the cluster we need to see if we
2157          * have a cluster here, and if so and it has our bitmap we need to add
2158          * the free space to that bitmap.
2159          */
2160         if (block_group && !list_empty(&block_group->cluster_list)) {
2161                 struct btrfs_free_cluster *cluster;
2162                 struct rb_node *node;
2163                 struct btrfs_free_space *entry;
2164
2165                 cluster = list_entry(block_group->cluster_list.next,
2166                                      struct btrfs_free_cluster,
2167                                      block_group_list);
2168                 spin_lock(&cluster->lock);
2169                 node = rb_first(&cluster->root);
2170                 if (!node) {
2171                         spin_unlock(&cluster->lock);
2172                         goto no_cluster_bitmap;
2173                 }
2174
2175                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2176                 if (!entry->bitmap) {
2177                         spin_unlock(&cluster->lock);
2178                         goto no_cluster_bitmap;
2179                 }
2180
2181                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2182                         bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2183                                                           bytes, trim_state);
2184                         bytes -= bytes_added;
2185                         offset += bytes_added;
2186                 }
2187                 spin_unlock(&cluster->lock);
2188                 if (!bytes) {
2189                         ret = 1;
2190                         goto out;
2191                 }
2192         }
2193
2194 no_cluster_bitmap:
2195         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2196                                          1, 0);
2197         if (!bitmap_info) {
2198                 ASSERT(added == 0);
2199                 goto new_bitmap;
2200         }
2201
2202         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2203                                           trim_state);
2204         bytes -= bytes_added;
2205         offset += bytes_added;
2206         added = 0;
2207
2208         if (!bytes) {
2209                 ret = 1;
2210                 goto out;
2211         } else
2212                 goto again;
2213
2214 new_bitmap:
2215         if (info && info->bitmap) {
2216                 add_new_bitmap(ctl, info, offset);
2217                 added = 1;
2218                 info = NULL;
2219                 goto again;
2220         } else {
2221                 spin_unlock(&ctl->tree_lock);
2222
2223                 /* no pre-allocated info, allocate a new one */
2224                 if (!info) {
2225                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2226                                                  GFP_NOFS);
2227                         if (!info) {
2228                                 spin_lock(&ctl->tree_lock);
2229                                 ret = -ENOMEM;
2230                                 goto out;
2231                         }
2232                 }
2233
2234                 /* allocate the bitmap */
2235                 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2236                                                  GFP_NOFS);
2237                 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2238                 spin_lock(&ctl->tree_lock);
2239                 if (!info->bitmap) {
2240                         ret = -ENOMEM;
2241                         goto out;
2242                 }
2243                 goto again;
2244         }
2245
2246 out:
2247         if (info) {
2248                 if (info->bitmap)
2249                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2250                                         info->bitmap);
2251                 kmem_cache_free(btrfs_free_space_cachep, info);
2252         }
2253
2254         return ret;
2255 }
2256
2257 /*
2258  * Free space merging rules:
2259  *  1) Merge trimmed areas together
2260  *  2) Let untrimmed areas coalesce with trimmed areas
2261  *  3) Always pull neighboring regions from bitmaps
2262  *
2263  * The above rules are for when we merge free space based on btrfs_trim_state.
2264  * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2265  * same reason: to promote larger extent regions which makes life easier for
2266  * find_free_extent().  Rule 2 enables coalescing based on the common path
2267  * being returning free space from btrfs_finish_extent_commit().  So when free
2268  * space is trimmed, it will prevent aggregating trimmed new region and
2269  * untrimmed regions in the rb_tree.  Rule 3 is purely to obtain larger extents
2270  * and provide find_free_extent() with the largest extents possible hoping for
2271  * the reuse path.
2272  */
2273 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2274                           struct btrfs_free_space *info, bool update_stat)
2275 {
2276         struct btrfs_free_space *left_info = NULL;
2277         struct btrfs_free_space *right_info;
2278         bool merged = false;
2279         u64 offset = info->offset;
2280         u64 bytes = info->bytes;
2281         const bool is_trimmed = btrfs_free_space_trimmed(info);
2282
2283         /*
2284          * first we want to see if there is free space adjacent to the range we
2285          * are adding, if there is remove that struct and add a new one to
2286          * cover the entire range
2287          */
2288         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2289         if (right_info && rb_prev(&right_info->offset_index))
2290                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2291                                      struct btrfs_free_space, offset_index);
2292         else if (!right_info)
2293                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2294
2295         /* See try_merge_free_space() comment. */
2296         if (right_info && !right_info->bitmap &&
2297             (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2298                 if (update_stat)
2299                         unlink_free_space(ctl, right_info);
2300                 else
2301                         __unlink_free_space(ctl, right_info);
2302                 info->bytes += right_info->bytes;
2303                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2304                 merged = true;
2305         }
2306
2307         /* See try_merge_free_space() comment. */
2308         if (left_info && !left_info->bitmap &&
2309             left_info->offset + left_info->bytes == offset &&
2310             (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2311                 if (update_stat)
2312                         unlink_free_space(ctl, left_info);
2313                 else
2314                         __unlink_free_space(ctl, left_info);
2315                 info->offset = left_info->offset;
2316                 info->bytes += left_info->bytes;
2317                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2318                 merged = true;
2319         }
2320
2321         return merged;
2322 }
2323
2324 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2325                                      struct btrfs_free_space *info,
2326                                      bool update_stat)
2327 {
2328         struct btrfs_free_space *bitmap;
2329         unsigned long i;
2330         unsigned long j;
2331         const u64 end = info->offset + info->bytes;
2332         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2333         u64 bytes;
2334
2335         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2336         if (!bitmap)
2337                 return false;
2338
2339         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2340         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2341         if (j == i)
2342                 return false;
2343         bytes = (j - i) * ctl->unit;
2344         info->bytes += bytes;
2345
2346         /* See try_merge_free_space() comment. */
2347         if (!btrfs_free_space_trimmed(bitmap))
2348                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2349
2350         if (update_stat)
2351                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2352         else
2353                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2354
2355         if (!bitmap->bytes)
2356                 free_bitmap(ctl, bitmap);
2357
2358         return true;
2359 }
2360
2361 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2362                                        struct btrfs_free_space *info,
2363                                        bool update_stat)
2364 {
2365         struct btrfs_free_space *bitmap;
2366         u64 bitmap_offset;
2367         unsigned long i;
2368         unsigned long j;
2369         unsigned long prev_j;
2370         u64 bytes;
2371
2372         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2373         /* If we're on a boundary, try the previous logical bitmap. */
2374         if (bitmap_offset == info->offset) {
2375                 if (info->offset == 0)
2376                         return false;
2377                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2378         }
2379
2380         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2381         if (!bitmap)
2382                 return false;
2383
2384         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2385         j = 0;
2386         prev_j = (unsigned long)-1;
2387         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2388                 if (j > i)
2389                         break;
2390                 prev_j = j;
2391         }
2392         if (prev_j == i)
2393                 return false;
2394
2395         if (prev_j == (unsigned long)-1)
2396                 bytes = (i + 1) * ctl->unit;
2397         else
2398                 bytes = (i - prev_j) * ctl->unit;
2399
2400         info->offset -= bytes;
2401         info->bytes += bytes;
2402
2403         /* See try_merge_free_space() comment. */
2404         if (!btrfs_free_space_trimmed(bitmap))
2405                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2406
2407         if (update_stat)
2408                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2409         else
2410                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2411
2412         if (!bitmap->bytes)
2413                 free_bitmap(ctl, bitmap);
2414
2415         return true;
2416 }
2417
2418 /*
2419  * We prefer always to allocate from extent entries, both for clustered and
2420  * non-clustered allocation requests. So when attempting to add a new extent
2421  * entry, try to see if there's adjacent free space in bitmap entries, and if
2422  * there is, migrate that space from the bitmaps to the extent.
2423  * Like this we get better chances of satisfying space allocation requests
2424  * because we attempt to satisfy them based on a single cache entry, and never
2425  * on 2 or more entries - even if the entries represent a contiguous free space
2426  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2427  * ends).
2428  */
2429 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2430                               struct btrfs_free_space *info,
2431                               bool update_stat)
2432 {
2433         /*
2434          * Only work with disconnected entries, as we can change their offset,
2435          * and must be extent entries.
2436          */
2437         ASSERT(!info->bitmap);
2438         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2439
2440         if (ctl->total_bitmaps > 0) {
2441                 bool stole_end;
2442                 bool stole_front = false;
2443
2444                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2445                 if (ctl->total_bitmaps > 0)
2446                         stole_front = steal_from_bitmap_to_front(ctl, info,
2447                                                                  update_stat);
2448
2449                 if (stole_end || stole_front)
2450                         try_merge_free_space(ctl, info, update_stat);
2451         }
2452 }
2453
2454 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2455                            struct btrfs_free_space_ctl *ctl,
2456                            u64 offset, u64 bytes,
2457                            enum btrfs_trim_state trim_state)
2458 {
2459         struct btrfs_block_group *block_group = ctl->private;
2460         struct btrfs_free_space *info;
2461         int ret = 0;
2462         u64 filter_bytes = bytes;
2463
2464         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2465         if (!info)
2466                 return -ENOMEM;
2467
2468         info->offset = offset;
2469         info->bytes = bytes;
2470         info->trim_state = trim_state;
2471         RB_CLEAR_NODE(&info->offset_index);
2472
2473         spin_lock(&ctl->tree_lock);
2474
2475         if (try_merge_free_space(ctl, info, true))
2476                 goto link;
2477
2478         /*
2479          * There was no extent directly to the left or right of this new
2480          * extent then we know we're going to have to allocate a new extent, so
2481          * before we do that see if we need to drop this into a bitmap
2482          */
2483         ret = insert_into_bitmap(ctl, info);
2484         if (ret < 0) {
2485                 goto out;
2486         } else if (ret) {
2487                 ret = 0;
2488                 goto out;
2489         }
2490 link:
2491         /*
2492          * Only steal free space from adjacent bitmaps if we're sure we're not
2493          * going to add the new free space to existing bitmap entries - because
2494          * that would mean unnecessary work that would be reverted. Therefore
2495          * attempt to steal space from bitmaps if we're adding an extent entry.
2496          */
2497         steal_from_bitmap(ctl, info, true);
2498
2499         filter_bytes = max(filter_bytes, info->bytes);
2500
2501         ret = link_free_space(ctl, info);
2502         if (ret)
2503                 kmem_cache_free(btrfs_free_space_cachep, info);
2504 out:
2505         btrfs_discard_update_discardable(block_group);
2506         spin_unlock(&ctl->tree_lock);
2507
2508         if (ret) {
2509                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2510                 ASSERT(ret != -EEXIST);
2511         }
2512
2513         if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2514                 btrfs_discard_check_filter(block_group, filter_bytes);
2515                 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2516         }
2517
2518         return ret;
2519 }
2520
2521 int btrfs_add_free_space(struct btrfs_block_group *block_group,
2522                          u64 bytenr, u64 size)
2523 {
2524         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2525
2526         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2527                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2528
2529         return __btrfs_add_free_space(block_group->fs_info,
2530                                       block_group->free_space_ctl,
2531                                       bytenr, size, trim_state);
2532 }
2533
2534 /*
2535  * This is a subtle distinction because when adding free space back in general,
2536  * we want it to be added as untrimmed for async. But in the case where we add
2537  * it on loading of a block group, we want to consider it trimmed.
2538  */
2539 int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2540                                        u64 bytenr, u64 size)
2541 {
2542         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2543
2544         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2545             btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2546                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2547
2548         return __btrfs_add_free_space(block_group->fs_info,
2549                                       block_group->free_space_ctl,
2550                                       bytenr, size, trim_state);
2551 }
2552
2553 int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2554                             u64 offset, u64 bytes)
2555 {
2556         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2557         struct btrfs_free_space *info;
2558         int ret;
2559         bool re_search = false;
2560
2561         spin_lock(&ctl->tree_lock);
2562
2563 again:
2564         ret = 0;
2565         if (!bytes)
2566                 goto out_lock;
2567
2568         info = tree_search_offset(ctl, offset, 0, 0);
2569         if (!info) {
2570                 /*
2571                  * oops didn't find an extent that matched the space we wanted
2572                  * to remove, look for a bitmap instead
2573                  */
2574                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2575                                           1, 0);
2576                 if (!info) {
2577                         /*
2578                          * If we found a partial bit of our free space in a
2579                          * bitmap but then couldn't find the other part this may
2580                          * be a problem, so WARN about it.
2581                          */
2582                         WARN_ON(re_search);
2583                         goto out_lock;
2584                 }
2585         }
2586
2587         re_search = false;
2588         if (!info->bitmap) {
2589                 unlink_free_space(ctl, info);
2590                 if (offset == info->offset) {
2591                         u64 to_free = min(bytes, info->bytes);
2592
2593                         info->bytes -= to_free;
2594                         info->offset += to_free;
2595                         if (info->bytes) {
2596                                 ret = link_free_space(ctl, info);
2597                                 WARN_ON(ret);
2598                         } else {
2599                                 kmem_cache_free(btrfs_free_space_cachep, info);
2600                         }
2601
2602                         offset += to_free;
2603                         bytes -= to_free;
2604                         goto again;
2605                 } else {
2606                         u64 old_end = info->bytes + info->offset;
2607
2608                         info->bytes = offset - info->offset;
2609                         ret = link_free_space(ctl, info);
2610                         WARN_ON(ret);
2611                         if (ret)
2612                                 goto out_lock;
2613
2614                         /* Not enough bytes in this entry to satisfy us */
2615                         if (old_end < offset + bytes) {
2616                                 bytes -= old_end - offset;
2617                                 offset = old_end;
2618                                 goto again;
2619                         } else if (old_end == offset + bytes) {
2620                                 /* all done */
2621                                 goto out_lock;
2622                         }
2623                         spin_unlock(&ctl->tree_lock);
2624
2625                         ret = __btrfs_add_free_space(block_group->fs_info, ctl,
2626                                                      offset + bytes,
2627                                                      old_end - (offset + bytes),
2628                                                      info->trim_state);
2629                         WARN_ON(ret);
2630                         goto out;
2631                 }
2632         }
2633
2634         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2635         if (ret == -EAGAIN) {
2636                 re_search = true;
2637                 goto again;
2638         }
2639 out_lock:
2640         btrfs_discard_update_discardable(block_group);
2641         spin_unlock(&ctl->tree_lock);
2642 out:
2643         return ret;
2644 }
2645
2646 void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2647                            u64 bytes)
2648 {
2649         struct btrfs_fs_info *fs_info = block_group->fs_info;
2650         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2651         struct btrfs_free_space *info;
2652         struct rb_node *n;
2653         int count = 0;
2654
2655         spin_lock(&ctl->tree_lock);
2656         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2657                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2658                 if (info->bytes >= bytes && !block_group->ro)
2659                         count++;
2660                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2661                            info->offset, info->bytes,
2662                        (info->bitmap) ? "yes" : "no");
2663         }
2664         spin_unlock(&ctl->tree_lock);
2665         btrfs_info(fs_info, "block group has cluster?: %s",
2666                list_empty(&block_group->cluster_list) ? "no" : "yes");
2667         btrfs_info(fs_info,
2668                    "%d blocks of free space at or bigger than bytes is", count);
2669 }
2670
2671 void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group,
2672                                struct btrfs_free_space_ctl *ctl)
2673 {
2674         struct btrfs_fs_info *fs_info = block_group->fs_info;
2675
2676         spin_lock_init(&ctl->tree_lock);
2677         ctl->unit = fs_info->sectorsize;
2678         ctl->start = block_group->start;
2679         ctl->private = block_group;
2680         ctl->op = &free_space_op;
2681         INIT_LIST_HEAD(&ctl->trimming_ranges);
2682         mutex_init(&ctl->cache_writeout_mutex);
2683
2684         /*
2685          * we only want to have 32k of ram per block group for keeping
2686          * track of free space, and if we pass 1/2 of that we want to
2687          * start converting things over to using bitmaps
2688          */
2689         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2690 }
2691
2692 /*
2693  * for a given cluster, put all of its extents back into the free
2694  * space cache.  If the block group passed doesn't match the block group
2695  * pointed to by the cluster, someone else raced in and freed the
2696  * cluster already.  In that case, we just return without changing anything
2697  */
2698 static void __btrfs_return_cluster_to_free_space(
2699                              struct btrfs_block_group *block_group,
2700                              struct btrfs_free_cluster *cluster)
2701 {
2702         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2703         struct btrfs_free_space *entry;
2704         struct rb_node *node;
2705
2706         spin_lock(&cluster->lock);
2707         if (cluster->block_group != block_group)
2708                 goto out;
2709
2710         cluster->block_group = NULL;
2711         cluster->window_start = 0;
2712         list_del_init(&cluster->block_group_list);
2713
2714         node = rb_first(&cluster->root);
2715         while (node) {
2716                 bool bitmap;
2717
2718                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2719                 node = rb_next(&entry->offset_index);
2720                 rb_erase(&entry->offset_index, &cluster->root);
2721                 RB_CLEAR_NODE(&entry->offset_index);
2722
2723                 bitmap = (entry->bitmap != NULL);
2724                 if (!bitmap) {
2725                         /* Merging treats extents as if they were new */
2726                         if (!btrfs_free_space_trimmed(entry)) {
2727                                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
2728                                 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
2729                                         entry->bytes;
2730                         }
2731
2732                         try_merge_free_space(ctl, entry, false);
2733                         steal_from_bitmap(ctl, entry, false);
2734
2735                         /* As we insert directly, update these statistics */
2736                         if (!btrfs_free_space_trimmed(entry)) {
2737                                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
2738                                 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
2739                                         entry->bytes;
2740                         }
2741                 }
2742                 tree_insert_offset(&ctl->free_space_offset,
2743                                    entry->offset, &entry->offset_index, bitmap);
2744         }
2745         cluster->root = RB_ROOT;
2746
2747 out:
2748         spin_unlock(&cluster->lock);
2749         btrfs_put_block_group(block_group);
2750 }
2751
2752 static void __btrfs_remove_free_space_cache_locked(
2753                                 struct btrfs_free_space_ctl *ctl)
2754 {
2755         struct btrfs_free_space *info;
2756         struct rb_node *node;
2757
2758         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2759                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2760                 if (!info->bitmap) {
2761                         unlink_free_space(ctl, info);
2762                         kmem_cache_free(btrfs_free_space_cachep, info);
2763                 } else {
2764                         free_bitmap(ctl, info);
2765                 }
2766
2767                 cond_resched_lock(&ctl->tree_lock);
2768         }
2769 }
2770
2771 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2772 {
2773         spin_lock(&ctl->tree_lock);
2774         __btrfs_remove_free_space_cache_locked(ctl);
2775         if (ctl->private)
2776                 btrfs_discard_update_discardable(ctl->private);
2777         spin_unlock(&ctl->tree_lock);
2778 }
2779
2780 void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
2781 {
2782         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2783         struct btrfs_free_cluster *cluster;
2784         struct list_head *head;
2785
2786         spin_lock(&ctl->tree_lock);
2787         while ((head = block_group->cluster_list.next) !=
2788                &block_group->cluster_list) {
2789                 cluster = list_entry(head, struct btrfs_free_cluster,
2790                                      block_group_list);
2791
2792                 WARN_ON(cluster->block_group != block_group);
2793                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2794
2795                 cond_resched_lock(&ctl->tree_lock);
2796         }
2797         __btrfs_remove_free_space_cache_locked(ctl);
2798         btrfs_discard_update_discardable(block_group);
2799         spin_unlock(&ctl->tree_lock);
2800
2801 }
2802
2803 /**
2804  * btrfs_is_free_space_trimmed - see if everything is trimmed
2805  * @block_group: block_group of interest
2806  *
2807  * Walk @block_group's free space rb_tree to determine if everything is trimmed.
2808  */
2809 bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
2810 {
2811         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2812         struct btrfs_free_space *info;
2813         struct rb_node *node;
2814         bool ret = true;
2815
2816         spin_lock(&ctl->tree_lock);
2817         node = rb_first(&ctl->free_space_offset);
2818
2819         while (node) {
2820                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2821
2822                 if (!btrfs_free_space_trimmed(info)) {
2823                         ret = false;
2824                         break;
2825                 }
2826
2827                 node = rb_next(node);
2828         }
2829
2830         spin_unlock(&ctl->tree_lock);
2831         return ret;
2832 }
2833
2834 u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
2835                                u64 offset, u64 bytes, u64 empty_size,
2836                                u64 *max_extent_size)
2837 {
2838         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2839         struct btrfs_discard_ctl *discard_ctl =
2840                                         &block_group->fs_info->discard_ctl;
2841         struct btrfs_free_space *entry = NULL;
2842         u64 bytes_search = bytes + empty_size;
2843         u64 ret = 0;
2844         u64 align_gap = 0;
2845         u64 align_gap_len = 0;
2846         enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2847
2848         spin_lock(&ctl->tree_lock);
2849         entry = find_free_space(ctl, &offset, &bytes_search,
2850                                 block_group->full_stripe_len, max_extent_size);
2851         if (!entry)
2852                 goto out;
2853
2854         ret = offset;
2855         if (entry->bitmap) {
2856                 bitmap_clear_bits(ctl, entry, offset, bytes);
2857
2858                 if (!btrfs_free_space_trimmed(entry))
2859                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
2860
2861                 if (!entry->bytes)
2862                         free_bitmap(ctl, entry);
2863         } else {
2864                 unlink_free_space(ctl, entry);
2865                 align_gap_len = offset - entry->offset;
2866                 align_gap = entry->offset;
2867                 align_gap_trim_state = entry->trim_state;
2868
2869                 if (!btrfs_free_space_trimmed(entry))
2870                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
2871
2872                 entry->offset = offset + bytes;
2873                 WARN_ON(entry->bytes < bytes + align_gap_len);
2874
2875                 entry->bytes -= bytes + align_gap_len;
2876                 if (!entry->bytes)
2877                         kmem_cache_free(btrfs_free_space_cachep, entry);
2878                 else
2879                         link_free_space(ctl, entry);
2880         }
2881 out:
2882         btrfs_discard_update_discardable(block_group);
2883         spin_unlock(&ctl->tree_lock);
2884
2885         if (align_gap_len)
2886                 __btrfs_add_free_space(block_group->fs_info, ctl,
2887                                        align_gap, align_gap_len,
2888                                        align_gap_trim_state);
2889         return ret;
2890 }
2891
2892 /*
2893  * given a cluster, put all of its extents back into the free space
2894  * cache.  If a block group is passed, this function will only free
2895  * a cluster that belongs to the passed block group.
2896  *
2897  * Otherwise, it'll get a reference on the block group pointed to by the
2898  * cluster and remove the cluster from it.
2899  */
2900 void btrfs_return_cluster_to_free_space(
2901                                struct btrfs_block_group *block_group,
2902                                struct btrfs_free_cluster *cluster)
2903 {
2904         struct btrfs_free_space_ctl *ctl;
2905
2906         /* first, get a safe pointer to the block group */
2907         spin_lock(&cluster->lock);
2908         if (!block_group) {
2909                 block_group = cluster->block_group;
2910                 if (!block_group) {
2911                         spin_unlock(&cluster->lock);
2912                         return;
2913                 }
2914         } else if (cluster->block_group != block_group) {
2915                 /* someone else has already freed it don't redo their work */
2916                 spin_unlock(&cluster->lock);
2917                 return;
2918         }
2919         btrfs_get_block_group(block_group);
2920         spin_unlock(&cluster->lock);
2921
2922         ctl = block_group->free_space_ctl;
2923
2924         /* now return any extents the cluster had on it */
2925         spin_lock(&ctl->tree_lock);
2926         __btrfs_return_cluster_to_free_space(block_group, cluster);
2927         spin_unlock(&ctl->tree_lock);
2928
2929         btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
2930
2931         /* finally drop our ref */
2932         btrfs_put_block_group(block_group);
2933 }
2934
2935 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
2936                                    struct btrfs_free_cluster *cluster,
2937                                    struct btrfs_free_space *entry,
2938                                    u64 bytes, u64 min_start,
2939                                    u64 *max_extent_size)
2940 {
2941         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2942         int err;
2943         u64 search_start = cluster->window_start;
2944         u64 search_bytes = bytes;
2945         u64 ret = 0;
2946
2947         search_start = min_start;
2948         search_bytes = bytes;
2949
2950         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2951         if (err) {
2952                 *max_extent_size = max(get_max_extent_size(entry),
2953                                        *max_extent_size);
2954                 return 0;
2955         }
2956
2957         ret = search_start;
2958         __bitmap_clear_bits(ctl, entry, ret, bytes);
2959
2960         return ret;
2961 }
2962
2963 /*
2964  * given a cluster, try to allocate 'bytes' from it, returns 0
2965  * if it couldn't find anything suitably large, or a logical disk offset
2966  * if things worked out
2967  */
2968 u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
2969                              struct btrfs_free_cluster *cluster, u64 bytes,
2970                              u64 min_start, u64 *max_extent_size)
2971 {
2972         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2973         struct btrfs_discard_ctl *discard_ctl =
2974                                         &block_group->fs_info->discard_ctl;
2975         struct btrfs_free_space *entry = NULL;
2976         struct rb_node *node;
2977         u64 ret = 0;
2978
2979         spin_lock(&cluster->lock);
2980         if (bytes > cluster->max_size)
2981                 goto out;
2982
2983         if (cluster->block_group != block_group)
2984                 goto out;
2985
2986         node = rb_first(&cluster->root);
2987         if (!node)
2988                 goto out;
2989
2990         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2991         while (1) {
2992                 if (entry->bytes < bytes)
2993                         *max_extent_size = max(get_max_extent_size(entry),
2994                                                *max_extent_size);
2995
2996                 if (entry->bytes < bytes ||
2997                     (!entry->bitmap && entry->offset < min_start)) {
2998                         node = rb_next(&entry->offset_index);
2999                         if (!node)
3000                                 break;
3001                         entry = rb_entry(node, struct btrfs_free_space,
3002                                          offset_index);
3003                         continue;
3004                 }
3005
3006                 if (entry->bitmap) {
3007                         ret = btrfs_alloc_from_bitmap(block_group,
3008                                                       cluster, entry, bytes,
3009                                                       cluster->window_start,
3010                                                       max_extent_size);
3011                         if (ret == 0) {
3012                                 node = rb_next(&entry->offset_index);
3013                                 if (!node)
3014                                         break;
3015                                 entry = rb_entry(node, struct btrfs_free_space,
3016                                                  offset_index);
3017                                 continue;
3018                         }
3019                         cluster->window_start += bytes;
3020                 } else {
3021                         ret = entry->offset;
3022
3023                         entry->offset += bytes;
3024                         entry->bytes -= bytes;
3025                 }
3026
3027                 if (entry->bytes == 0)
3028                         rb_erase(&entry->offset_index, &cluster->root);
3029                 break;
3030         }
3031 out:
3032         spin_unlock(&cluster->lock);
3033
3034         if (!ret)
3035                 return 0;
3036
3037         spin_lock(&ctl->tree_lock);
3038
3039         if (!btrfs_free_space_trimmed(entry))
3040                 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3041
3042         ctl->free_space -= bytes;
3043         if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3044                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3045         if (entry->bytes == 0) {
3046                 ctl->free_extents--;
3047                 if (entry->bitmap) {
3048                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
3049                                         entry->bitmap);
3050                         ctl->total_bitmaps--;
3051                         recalculate_thresholds(ctl);
3052                 } else if (!btrfs_free_space_trimmed(entry)) {
3053                         ctl->discardable_extents[BTRFS_STAT_CURR]--;
3054                 }
3055                 kmem_cache_free(btrfs_free_space_cachep, entry);
3056         }
3057
3058         spin_unlock(&ctl->tree_lock);
3059
3060         return ret;
3061 }
3062
3063 static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3064                                 struct btrfs_free_space *entry,
3065                                 struct btrfs_free_cluster *cluster,
3066                                 u64 offset, u64 bytes,
3067                                 u64 cont1_bytes, u64 min_bytes)
3068 {
3069         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3070         unsigned long next_zero;
3071         unsigned long i;
3072         unsigned long want_bits;
3073         unsigned long min_bits;
3074         unsigned long found_bits;
3075         unsigned long max_bits = 0;
3076         unsigned long start = 0;
3077         unsigned long total_found = 0;
3078         int ret;
3079
3080         i = offset_to_bit(entry->offset, ctl->unit,
3081                           max_t(u64, offset, entry->offset));
3082         want_bits = bytes_to_bits(bytes, ctl->unit);
3083         min_bits = bytes_to_bits(min_bytes, ctl->unit);
3084
3085         /*
3086          * Don't bother looking for a cluster in this bitmap if it's heavily
3087          * fragmented.
3088          */
3089         if (entry->max_extent_size &&
3090             entry->max_extent_size < cont1_bytes)
3091                 return -ENOSPC;
3092 again:
3093         found_bits = 0;
3094         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3095                 next_zero = find_next_zero_bit(entry->bitmap,
3096                                                BITS_PER_BITMAP, i);
3097                 if (next_zero - i >= min_bits) {
3098                         found_bits = next_zero - i;
3099                         if (found_bits > max_bits)
3100                                 max_bits = found_bits;
3101                         break;
3102                 }
3103                 if (next_zero - i > max_bits)
3104                         max_bits = next_zero - i;
3105                 i = next_zero;
3106         }
3107
3108         if (!found_bits) {
3109                 entry->max_extent_size = (u64)max_bits * ctl->unit;
3110                 return -ENOSPC;
3111         }
3112
3113         if (!total_found) {
3114                 start = i;
3115                 cluster->max_size = 0;
3116         }
3117
3118         total_found += found_bits;
3119
3120         if (cluster->max_size < found_bits * ctl->unit)
3121                 cluster->max_size = found_bits * ctl->unit;
3122
3123         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3124                 i = next_zero + 1;
3125                 goto again;
3126         }
3127
3128         cluster->window_start = start * ctl->unit + entry->offset;
3129         rb_erase(&entry->offset_index, &ctl->free_space_offset);
3130         ret = tree_insert_offset(&cluster->root, entry->offset,
3131                                  &entry->offset_index, 1);
3132         ASSERT(!ret); /* -EEXIST; Logic error */
3133
3134         trace_btrfs_setup_cluster(block_group, cluster,
3135                                   total_found * ctl->unit, 1);
3136         return 0;
3137 }
3138
3139 /*
3140  * This searches the block group for just extents to fill the cluster with.
3141  * Try to find a cluster with at least bytes total bytes, at least one
3142  * extent of cont1_bytes, and other clusters of at least min_bytes.
3143  */
3144 static noinline int
3145 setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3146                         struct btrfs_free_cluster *cluster,
3147                         struct list_head *bitmaps, u64 offset, u64 bytes,
3148                         u64 cont1_bytes, u64 min_bytes)
3149 {
3150         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3151         struct btrfs_free_space *first = NULL;
3152         struct btrfs_free_space *entry = NULL;
3153         struct btrfs_free_space *last;
3154         struct rb_node *node;
3155         u64 window_free;
3156         u64 max_extent;
3157         u64 total_size = 0;
3158
3159         entry = tree_search_offset(ctl, offset, 0, 1);
3160         if (!entry)
3161                 return -ENOSPC;
3162
3163         /*
3164          * We don't want bitmaps, so just move along until we find a normal
3165          * extent entry.
3166          */
3167         while (entry->bitmap || entry->bytes < min_bytes) {
3168                 if (entry->bitmap && list_empty(&entry->list))
3169                         list_add_tail(&entry->list, bitmaps);
3170                 node = rb_next(&entry->offset_index);
3171                 if (!node)
3172                         return -ENOSPC;
3173                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3174         }
3175
3176         window_free = entry->bytes;
3177         max_extent = entry->bytes;
3178         first = entry;
3179         last = entry;
3180
3181         for (node = rb_next(&entry->offset_index); node;
3182              node = rb_next(&entry->offset_index)) {
3183                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3184
3185                 if (entry->bitmap) {
3186                         if (list_empty(&entry->list))
3187                                 list_add_tail(&entry->list, bitmaps);
3188                         continue;
3189                 }
3190
3191                 if (entry->bytes < min_bytes)
3192                         continue;
3193
3194                 last = entry;
3195                 window_free += entry->bytes;
3196                 if (entry->bytes > max_extent)
3197                         max_extent = entry->bytes;
3198         }
3199
3200         if (window_free < bytes || max_extent < cont1_bytes)
3201                 return -ENOSPC;
3202
3203         cluster->window_start = first->offset;
3204
3205         node = &first->offset_index;
3206
3207         /*
3208          * now we've found our entries, pull them out of the free space
3209          * cache and put them into the cluster rbtree
3210          */
3211         do {
3212                 int ret;
3213
3214                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3215                 node = rb_next(&entry->offset_index);
3216                 if (entry->bitmap || entry->bytes < min_bytes)
3217                         continue;
3218
3219                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3220                 ret = tree_insert_offset(&cluster->root, entry->offset,
3221                                          &entry->offset_index, 0);
3222                 total_size += entry->bytes;
3223                 ASSERT(!ret); /* -EEXIST; Logic error */
3224         } while (node && entry != last);
3225
3226         cluster->max_size = max_extent;
3227         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3228         return 0;
3229 }
3230
3231 /*
3232  * This specifically looks for bitmaps that may work in the cluster, we assume
3233  * that we have already failed to find extents that will work.
3234  */
3235 static noinline int
3236 setup_cluster_bitmap(struct btrfs_block_group *block_group,
3237                      struct btrfs_free_cluster *cluster,
3238                      struct list_head *bitmaps, u64 offset, u64 bytes,
3239                      u64 cont1_bytes, u64 min_bytes)
3240 {
3241         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3242         struct btrfs_free_space *entry = NULL;
3243         int ret = -ENOSPC;
3244         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3245
3246         if (ctl->total_bitmaps == 0)
3247                 return -ENOSPC;
3248
3249         /*
3250          * The bitmap that covers offset won't be in the list unless offset
3251          * is just its start offset.
3252          */
3253         if (!list_empty(bitmaps))
3254                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3255
3256         if (!entry || entry->offset != bitmap_offset) {
3257                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3258                 if (entry && list_empty(&entry->list))
3259                         list_add(&entry->list, bitmaps);
3260         }
3261
3262         list_for_each_entry(entry, bitmaps, list) {
3263                 if (entry->bytes < bytes)
3264                         continue;
3265                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3266                                            bytes, cont1_bytes, min_bytes);
3267                 if (!ret)
3268                         return 0;
3269         }
3270
3271         /*
3272          * The bitmaps list has all the bitmaps that record free space
3273          * starting after offset, so no more search is required.
3274          */
3275         return -ENOSPC;
3276 }
3277
3278 /*
3279  * here we try to find a cluster of blocks in a block group.  The goal
3280  * is to find at least bytes+empty_size.
3281  * We might not find them all in one contiguous area.
3282  *
3283  * returns zero and sets up cluster if things worked out, otherwise
3284  * it returns -enospc
3285  */
3286 int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3287                              struct btrfs_free_cluster *cluster,
3288                              u64 offset, u64 bytes, u64 empty_size)
3289 {
3290         struct btrfs_fs_info *fs_info = block_group->fs_info;
3291         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3292         struct btrfs_free_space *entry, *tmp;
3293         LIST_HEAD(bitmaps);
3294         u64 min_bytes;
3295         u64 cont1_bytes;
3296         int ret;
3297
3298         /*
3299          * Choose the minimum extent size we'll require for this
3300          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3301          * For metadata, allow allocates with smaller extents.  For
3302          * data, keep it dense.
3303          */
3304         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3305                 cont1_bytes = min_bytes = bytes + empty_size;
3306         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3307                 cont1_bytes = bytes;
3308                 min_bytes = fs_info->sectorsize;
3309         } else {
3310                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3311                 min_bytes = fs_info->sectorsize;
3312         }
3313
3314         spin_lock(&ctl->tree_lock);
3315
3316         /*
3317          * If we know we don't have enough space to make a cluster don't even
3318          * bother doing all the work to try and find one.
3319          */
3320         if (ctl->free_space < bytes) {
3321                 spin_unlock(&ctl->tree_lock);
3322                 return -ENOSPC;
3323         }
3324
3325         spin_lock(&cluster->lock);
3326
3327         /* someone already found a cluster, hooray */
3328         if (cluster->block_group) {
3329                 ret = 0;
3330                 goto out;
3331         }
3332
3333         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3334                                  min_bytes);
3335
3336         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3337                                       bytes + empty_size,
3338                                       cont1_bytes, min_bytes);
3339         if (ret)
3340                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3341                                            offset, bytes + empty_size,
3342                                            cont1_bytes, min_bytes);
3343
3344         /* Clear our temporary list */
3345         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3346                 list_del_init(&entry->list);
3347
3348         if (!ret) {
3349                 btrfs_get_block_group(block_group);
3350                 list_add_tail(&cluster->block_group_list,
3351                               &block_group->cluster_list);
3352                 cluster->block_group = block_group;
3353         } else {
3354                 trace_btrfs_failed_cluster_setup(block_group);
3355         }
3356 out:
3357         spin_unlock(&cluster->lock);
3358         spin_unlock(&ctl->tree_lock);
3359
3360         return ret;
3361 }
3362
3363 /*
3364  * simple code to zero out a cluster
3365  */
3366 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3367 {
3368         spin_lock_init(&cluster->lock);
3369         spin_lock_init(&cluster->refill_lock);
3370         cluster->root = RB_ROOT;
3371         cluster->max_size = 0;
3372         cluster->fragmented = false;
3373         INIT_LIST_HEAD(&cluster->block_group_list);
3374         cluster->block_group = NULL;
3375 }
3376
3377 static int do_trimming(struct btrfs_block_group *block_group,
3378                        u64 *total_trimmed, u64 start, u64 bytes,
3379                        u64 reserved_start, u64 reserved_bytes,
3380                        enum btrfs_trim_state reserved_trim_state,
3381                        struct btrfs_trim_range *trim_entry)
3382 {
3383         struct btrfs_space_info *space_info = block_group->space_info;
3384         struct btrfs_fs_info *fs_info = block_group->fs_info;
3385         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3386         int ret;
3387         int update = 0;
3388         const u64 end = start + bytes;
3389         const u64 reserved_end = reserved_start + reserved_bytes;
3390         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3391         u64 trimmed = 0;
3392
3393         spin_lock(&space_info->lock);
3394         spin_lock(&block_group->lock);
3395         if (!block_group->ro) {
3396                 block_group->reserved += reserved_bytes;
3397                 space_info->bytes_reserved += reserved_bytes;
3398                 update = 1;
3399         }
3400         spin_unlock(&block_group->lock);
3401         spin_unlock(&space_info->lock);
3402
3403         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3404         if (!ret) {
3405                 *total_trimmed += trimmed;
3406                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3407         }
3408
3409         mutex_lock(&ctl->cache_writeout_mutex);
3410         if (reserved_start < start)
3411                 __btrfs_add_free_space(fs_info, ctl, reserved_start,
3412                                        start - reserved_start,
3413                                        reserved_trim_state);
3414         if (start + bytes < reserved_start + reserved_bytes)
3415                 __btrfs_add_free_space(fs_info, ctl, end, reserved_end - end,
3416                                        reserved_trim_state);
3417         __btrfs_add_free_space(fs_info, ctl, start, bytes, trim_state);
3418         list_del(&trim_entry->list);
3419         mutex_unlock(&ctl->cache_writeout_mutex);
3420
3421         if (update) {
3422                 spin_lock(&space_info->lock);
3423                 spin_lock(&block_group->lock);
3424                 if (block_group->ro)
3425                         space_info->bytes_readonly += reserved_bytes;
3426                 block_group->reserved -= reserved_bytes;
3427                 space_info->bytes_reserved -= reserved_bytes;
3428                 spin_unlock(&block_group->lock);
3429                 spin_unlock(&space_info->lock);
3430         }
3431
3432         return ret;
3433 }
3434
3435 /*
3436  * If @async is set, then we will trim 1 region and return.
3437  */
3438 static int trim_no_bitmap(struct btrfs_block_group *block_group,
3439                           u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3440                           bool async)
3441 {
3442         struct btrfs_discard_ctl *discard_ctl =
3443                                         &block_group->fs_info->discard_ctl;
3444         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3445         struct btrfs_free_space *entry;
3446         struct rb_node *node;
3447         int ret = 0;
3448         u64 extent_start;
3449         u64 extent_bytes;
3450         enum btrfs_trim_state extent_trim_state;
3451         u64 bytes;
3452         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3453
3454         while (start < end) {
3455                 struct btrfs_trim_range trim_entry;
3456
3457                 mutex_lock(&ctl->cache_writeout_mutex);
3458                 spin_lock(&ctl->tree_lock);
3459
3460                 if (ctl->free_space < minlen)
3461                         goto out_unlock;
3462
3463                 entry = tree_search_offset(ctl, start, 0, 1);
3464                 if (!entry)
3465                         goto out_unlock;
3466
3467                 /* Skip bitmaps and if async, already trimmed entries */
3468                 while (entry->bitmap ||
3469                        (async && btrfs_free_space_trimmed(entry))) {
3470                         node = rb_next(&entry->offset_index);
3471                         if (!node)
3472                                 goto out_unlock;
3473                         entry = rb_entry(node, struct btrfs_free_space,
3474                                          offset_index);
3475                 }
3476
3477                 if (entry->offset >= end)
3478                         goto out_unlock;
3479
3480                 extent_start = entry->offset;
3481                 extent_bytes = entry->bytes;
3482                 extent_trim_state = entry->trim_state;
3483                 if (async) {
3484                         start = entry->offset;
3485                         bytes = entry->bytes;
3486                         if (bytes < minlen) {
3487                                 spin_unlock(&ctl->tree_lock);
3488                                 mutex_unlock(&ctl->cache_writeout_mutex);
3489                                 goto next;
3490                         }
3491                         unlink_free_space(ctl, entry);
3492                         /*
3493                          * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3494                          * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3495                          * X when we come back around.  So trim it now.
3496                          */
3497                         if (max_discard_size &&
3498                             bytes >= (max_discard_size +
3499                                       BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3500                                 bytes = max_discard_size;
3501                                 extent_bytes = max_discard_size;
3502                                 entry->offset += max_discard_size;
3503                                 entry->bytes -= max_discard_size;
3504                                 link_free_space(ctl, entry);
3505                         } else {
3506                                 kmem_cache_free(btrfs_free_space_cachep, entry);
3507                         }
3508                 } else {
3509                         start = max(start, extent_start);
3510                         bytes = min(extent_start + extent_bytes, end) - start;
3511                         if (bytes < minlen) {
3512                                 spin_unlock(&ctl->tree_lock);
3513                                 mutex_unlock(&ctl->cache_writeout_mutex);
3514                                 goto next;
3515                         }
3516
3517                         unlink_free_space(ctl, entry);
3518                         kmem_cache_free(btrfs_free_space_cachep, entry);
3519                 }
3520
3521                 spin_unlock(&ctl->tree_lock);
3522                 trim_entry.start = extent_start;
3523                 trim_entry.bytes = extent_bytes;
3524                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3525                 mutex_unlock(&ctl->cache_writeout_mutex);
3526
3527                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3528                                   extent_start, extent_bytes, extent_trim_state,
3529                                   &trim_entry);
3530                 if (ret) {
3531                         block_group->discard_cursor = start + bytes;
3532                         break;
3533                 }
3534 next:
3535                 start += bytes;
3536                 block_group->discard_cursor = start;
3537                 if (async && *total_trimmed)
3538                         break;
3539
3540                 if (fatal_signal_pending(current)) {
3541                         ret = -ERESTARTSYS;
3542                         break;
3543                 }
3544
3545                 cond_resched();
3546         }
3547
3548         return ret;
3549
3550 out_unlock:
3551         block_group->discard_cursor = btrfs_block_group_end(block_group);
3552         spin_unlock(&ctl->tree_lock);
3553         mutex_unlock(&ctl->cache_writeout_mutex);
3554
3555         return ret;
3556 }
3557
3558 /*
3559  * If we break out of trimming a bitmap prematurely, we should reset the
3560  * trimming bit.  In a rather contrieved case, it's possible to race here so
3561  * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3562  *
3563  * start = start of bitmap
3564  * end = near end of bitmap
3565  *
3566  * Thread 1:                    Thread 2:
3567  * trim_bitmaps(start)
3568  *                              trim_bitmaps(end)
3569  *                              end_trimming_bitmap()
3570  * reset_trimming_bitmap()
3571  */
3572 static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3573 {
3574         struct btrfs_free_space *entry;
3575
3576         spin_lock(&ctl->tree_lock);
3577         entry = tree_search_offset(ctl, offset, 1, 0);
3578         if (entry) {
3579                 if (btrfs_free_space_trimmed(entry)) {
3580                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
3581                                 entry->bitmap_extents;
3582                         ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3583                 }
3584                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3585         }
3586
3587         spin_unlock(&ctl->tree_lock);
3588 }
3589
3590 static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3591                                 struct btrfs_free_space *entry)
3592 {
3593         if (btrfs_free_space_trimming_bitmap(entry)) {
3594                 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3595                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3596                         entry->bitmap_extents;
3597                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3598         }
3599 }
3600
3601 /*
3602  * If @async is set, then we will trim 1 region and return.
3603  */
3604 static int trim_bitmaps(struct btrfs_block_group *block_group,
3605                         u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3606                         u64 maxlen, bool async)
3607 {
3608         struct btrfs_discard_ctl *discard_ctl =
3609                                         &block_group->fs_info->discard_ctl;
3610         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3611         struct btrfs_free_space *entry;
3612         int ret = 0;
3613         int ret2;
3614         u64 bytes;
3615         u64 offset = offset_to_bitmap(ctl, start);
3616         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3617
3618         while (offset < end) {
3619                 bool next_bitmap = false;
3620                 struct btrfs_trim_range trim_entry;
3621
3622                 mutex_lock(&ctl->cache_writeout_mutex);
3623                 spin_lock(&ctl->tree_lock);
3624
3625                 if (ctl->free_space < minlen) {
3626                         block_group->discard_cursor =
3627                                 btrfs_block_group_end(block_group);
3628                         spin_unlock(&ctl->tree_lock);
3629                         mutex_unlock(&ctl->cache_writeout_mutex);
3630                         break;
3631                 }
3632
3633                 entry = tree_search_offset(ctl, offset, 1, 0);
3634                 /*
3635                  * Bitmaps are marked trimmed lossily now to prevent constant
3636                  * discarding of the same bitmap (the reason why we are bound
3637                  * by the filters).  So, retrim the block group bitmaps when we
3638                  * are preparing to punt to the unused_bgs list.  This uses
3639                  * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3640                  * which is the only discard index which sets minlen to 0.
3641                  */
3642                 if (!entry || (async && minlen && start == offset &&
3643                                btrfs_free_space_trimmed(entry))) {
3644                         spin_unlock(&ctl->tree_lock);
3645                         mutex_unlock(&ctl->cache_writeout_mutex);
3646                         next_bitmap = true;
3647                         goto next;
3648                 }
3649
3650                 /*
3651                  * Async discard bitmap trimming begins at by setting the start
3652                  * to be key.objectid and the offset_to_bitmap() aligns to the
3653                  * start of the bitmap.  This lets us know we are fully
3654                  * scanning the bitmap rather than only some portion of it.
3655                  */
3656                 if (start == offset)
3657                         entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3658
3659                 bytes = minlen;
3660                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3661                 if (ret2 || start >= end) {
3662                         /*
3663                          * We lossily consider a bitmap trimmed if we only skip
3664                          * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3665                          */
3666                         if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3667                                 end_trimming_bitmap(ctl, entry);
3668                         else
3669                                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3670                         spin_unlock(&ctl->tree_lock);
3671                         mutex_unlock(&ctl->cache_writeout_mutex);
3672                         next_bitmap = true;
3673                         goto next;
3674                 }
3675
3676                 /*
3677                  * We already trimmed a region, but are using the locking above
3678                  * to reset the trim_state.
3679                  */
3680                 if (async && *total_trimmed) {
3681                         spin_unlock(&ctl->tree_lock);
3682                         mutex_unlock(&ctl->cache_writeout_mutex);
3683                         goto out;
3684                 }
3685
3686                 bytes = min(bytes, end - start);
3687                 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3688                         spin_unlock(&ctl->tree_lock);
3689                         mutex_unlock(&ctl->cache_writeout_mutex);
3690                         goto next;
3691                 }
3692
3693                 /*
3694                  * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3695                  * If X < @minlen, we won't trim X when we come back around.
3696                  * So trim it now.  We differ here from trimming extents as we
3697                  * don't keep individual state per bit.
3698                  */
3699                 if (async &&
3700                     max_discard_size &&
3701                     bytes > (max_discard_size + minlen))
3702                         bytes = max_discard_size;
3703
3704                 bitmap_clear_bits(ctl, entry, start, bytes);
3705                 if (entry->bytes == 0)
3706                         free_bitmap(ctl, entry);
3707
3708                 spin_unlock(&ctl->tree_lock);
3709                 trim_entry.start = start;
3710                 trim_entry.bytes = bytes;
3711                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3712                 mutex_unlock(&ctl->cache_writeout_mutex);
3713
3714                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3715                                   start, bytes, 0, &trim_entry);
3716                 if (ret) {
3717                         reset_trimming_bitmap(ctl, offset);
3718                         block_group->discard_cursor =
3719                                 btrfs_block_group_end(block_group);
3720                         break;
3721                 }
3722 next:
3723                 if (next_bitmap) {
3724                         offset += BITS_PER_BITMAP * ctl->unit;
3725                         start = offset;
3726                 } else {
3727                         start += bytes;
3728                 }
3729                 block_group->discard_cursor = start;
3730
3731                 if (fatal_signal_pending(current)) {
3732                         if (start != offset)
3733                                 reset_trimming_bitmap(ctl, offset);
3734                         ret = -ERESTARTSYS;
3735                         break;
3736                 }
3737
3738                 cond_resched();
3739         }
3740
3741         if (offset >= end)
3742                 block_group->discard_cursor = end;
3743
3744 out:
3745         return ret;
3746 }
3747
3748 int btrfs_trim_block_group(struct btrfs_block_group *block_group,
3749                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3750 {
3751         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3752         int ret;
3753         u64 rem = 0;
3754
3755         *trimmed = 0;
3756
3757         spin_lock(&block_group->lock);
3758         if (block_group->removed) {
3759                 spin_unlock(&block_group->lock);
3760                 return 0;
3761         }
3762         btrfs_freeze_block_group(block_group);
3763         spin_unlock(&block_group->lock);
3764
3765         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
3766         if (ret)
3767                 goto out;
3768
3769         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
3770         div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
3771         /* If we ended in the middle of a bitmap, reset the trimming flag */
3772         if (rem)
3773                 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
3774 out:
3775         btrfs_unfreeze_block_group(block_group);
3776         return ret;
3777 }
3778
3779 int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
3780                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
3781                                    bool async)
3782 {
3783         int ret;
3784
3785         *trimmed = 0;
3786
3787         spin_lock(&block_group->lock);
3788         if (block_group->removed) {
3789                 spin_unlock(&block_group->lock);
3790                 return 0;
3791         }
3792         btrfs_freeze_block_group(block_group);
3793         spin_unlock(&block_group->lock);
3794
3795         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
3796         btrfs_unfreeze_block_group(block_group);
3797
3798         return ret;
3799 }
3800
3801 int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
3802                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
3803                                    u64 maxlen, bool async)
3804 {
3805         int ret;
3806
3807         *trimmed = 0;
3808
3809         spin_lock(&block_group->lock);
3810         if (block_group->removed) {
3811                 spin_unlock(&block_group->lock);
3812                 return 0;
3813         }
3814         btrfs_freeze_block_group(block_group);
3815         spin_unlock(&block_group->lock);
3816
3817         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
3818                            async);
3819
3820         btrfs_unfreeze_block_group(block_group);
3821
3822         return ret;
3823 }
3824
3825 bool btrfs_free_space_cache_v1_active(struct btrfs_fs_info *fs_info)
3826 {
3827         return btrfs_super_cache_generation(fs_info->super_copy);
3828 }
3829
3830 static int cleanup_free_space_cache_v1(struct btrfs_fs_info *fs_info,
3831                                        struct btrfs_trans_handle *trans)
3832 {
3833         struct btrfs_block_group *block_group;
3834         struct rb_node *node;
3835         int ret;
3836
3837         btrfs_info(fs_info, "cleaning free space cache v1");
3838
3839         node = rb_first(&fs_info->block_group_cache_tree);
3840         while (node) {
3841                 block_group = rb_entry(node, struct btrfs_block_group, cache_node);
3842                 ret = btrfs_remove_free_space_inode(trans, NULL, block_group);
3843                 if (ret)
3844                         goto out;
3845                 node = rb_next(node);
3846         }
3847 out:
3848         return ret;
3849 }
3850
3851 int btrfs_set_free_space_cache_v1_active(struct btrfs_fs_info *fs_info, bool active)
3852 {
3853         struct btrfs_trans_handle *trans;
3854         int ret;
3855
3856         /*
3857          * update_super_roots will appropriately set or unset
3858          * super_copy->cache_generation based on SPACE_CACHE and
3859          * BTRFS_FS_CLEANUP_SPACE_CACHE_V1. For this reason, we need a
3860          * transaction commit whether we are enabling space cache v1 and don't
3861          * have any other work to do, or are disabling it and removing free
3862          * space inodes.
3863          */
3864         trans = btrfs_start_transaction(fs_info->tree_root, 0);
3865         if (IS_ERR(trans))
3866                 return PTR_ERR(trans);
3867
3868         if (!active) {
3869                 set_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
3870                 ret = cleanup_free_space_cache_v1(fs_info, trans);
3871                 if (ret) {
3872                         btrfs_abort_transaction(trans, ret);
3873                         btrfs_end_transaction(trans);
3874                         goto out;
3875                 }
3876         }
3877
3878         ret = btrfs_commit_transaction(trans);
3879 out:
3880         clear_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
3881
3882         return ret;
3883 }
3884
3885 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3886 /*
3887  * Use this if you need to make a bitmap or extent entry specifically, it
3888  * doesn't do any of the merging that add_free_space does, this acts a lot like
3889  * how the free space cache loading stuff works, so you can get really weird
3890  * configurations.
3891  */
3892 int test_add_free_space_entry(struct btrfs_block_group *cache,
3893                               u64 offset, u64 bytes, bool bitmap)
3894 {
3895         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3896         struct btrfs_free_space *info = NULL, *bitmap_info;
3897         void *map = NULL;
3898         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
3899         u64 bytes_added;
3900         int ret;
3901
3902 again:
3903         if (!info) {
3904                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3905                 if (!info)
3906                         return -ENOMEM;
3907         }
3908
3909         if (!bitmap) {
3910                 spin_lock(&ctl->tree_lock);
3911                 info->offset = offset;
3912                 info->bytes = bytes;
3913                 info->max_extent_size = 0;
3914                 ret = link_free_space(ctl, info);
3915                 spin_unlock(&ctl->tree_lock);
3916                 if (ret)
3917                         kmem_cache_free(btrfs_free_space_cachep, info);
3918                 return ret;
3919         }
3920
3921         if (!map) {
3922                 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
3923                 if (!map) {
3924                         kmem_cache_free(btrfs_free_space_cachep, info);
3925                         return -ENOMEM;
3926                 }
3927         }
3928
3929         spin_lock(&ctl->tree_lock);
3930         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3931                                          1, 0);
3932         if (!bitmap_info) {
3933                 info->bitmap = map;
3934                 map = NULL;
3935                 add_new_bitmap(ctl, info, offset);
3936                 bitmap_info = info;
3937                 info = NULL;
3938         }
3939
3940         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
3941                                           trim_state);
3942
3943         bytes -= bytes_added;
3944         offset += bytes_added;
3945         spin_unlock(&ctl->tree_lock);
3946
3947         if (bytes)
3948                 goto again;
3949
3950         if (info)
3951                 kmem_cache_free(btrfs_free_space_cachep, info);
3952         if (map)
3953                 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
3954         return 0;
3955 }
3956
3957 /*
3958  * Checks to see if the given range is in the free space cache.  This is really
3959  * just used to check the absence of space, so if there is free space in the
3960  * range at all we will return 1.
3961  */
3962 int test_check_exists(struct btrfs_block_group *cache,
3963                       u64 offset, u64 bytes)
3964 {
3965         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3966         struct btrfs_free_space *info;
3967         int ret = 0;
3968
3969         spin_lock(&ctl->tree_lock);
3970         info = tree_search_offset(ctl, offset, 0, 0);
3971         if (!info) {
3972                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3973                                           1, 0);
3974                 if (!info)
3975                         goto out;
3976         }
3977
3978 have_info:
3979         if (info->bitmap) {
3980                 u64 bit_off, bit_bytes;
3981                 struct rb_node *n;
3982                 struct btrfs_free_space *tmp;
3983
3984                 bit_off = offset;
3985                 bit_bytes = ctl->unit;
3986                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3987                 if (!ret) {
3988                         if (bit_off == offset) {
3989                                 ret = 1;
3990                                 goto out;
3991                         } else if (bit_off > offset &&
3992                                    offset + bytes > bit_off) {
3993                                 ret = 1;
3994                                 goto out;
3995                         }
3996                 }
3997
3998                 n = rb_prev(&info->offset_index);
3999                 while (n) {
4000                         tmp = rb_entry(n, struct btrfs_free_space,
4001                                        offset_index);
4002                         if (tmp->offset + tmp->bytes < offset)
4003                                 break;
4004                         if (offset + bytes < tmp->offset) {
4005                                 n = rb_prev(&tmp->offset_index);
4006                                 continue;
4007                         }
4008                         info = tmp;
4009                         goto have_info;
4010                 }
4011
4012                 n = rb_next(&info->offset_index);
4013                 while (n) {
4014                         tmp = rb_entry(n, struct btrfs_free_space,
4015                                        offset_index);
4016                         if (offset + bytes < tmp->offset)
4017                                 break;
4018                         if (tmp->offset + tmp->bytes < offset) {
4019                                 n = rb_next(&tmp->offset_index);
4020                                 continue;
4021                         }
4022                         info = tmp;
4023                         goto have_info;
4024                 }
4025
4026                 ret = 0;
4027                 goto out;
4028         }
4029
4030         if (info->offset == offset) {
4031                 ret = 1;
4032                 goto out;
4033         }
4034
4035         if (offset > info->offset && offset < info->offset + info->bytes)
4036                 ret = 1;
4037 out:
4038         spin_unlock(&ctl->tree_lock);
4039         return ret;
4040 }
4041 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */