fs/btrfs/free-space-cache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2008 Red Hat.  All rights reserved.
   4  */
   5
   6 #include <linux/pagemap.h>
   7 #include <linux/sched.h>
   8 #include <linux/sched/signal.h>
   9 #include <linux/slab.h>
  10 #include <linux/math64.h>
  11 #include <linux/ratelimit.h>
  12 #include <linux/error-injection.h>
  13 #include <linux/sched/mm.h>
  14 #include "ctree.h"
  15 #include "free-space-cache.h"
  16 #include "transaction.h"
  17 #include "disk-io.h"
  18 #include "extent_io.h"
  19 #include "inode-map.h"
  20 #include "volumes.h"
  21 #include "space-info.h"
  22 #include "delalloc-space.h"
  23 #include "block-group.h"
  24 #include "discard.h"
  25
  26 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  27 #define MAX_CACHE_BYTES_PER_GIG SZ_64K
  28 #define FORCE_EXTENT_THRESHOLD  SZ_1M
  29
  30 struct btrfs_trim_range {
  31         u64 start;
  32         u64 bytes;
  33         struct list_head list;
  34 };
  35
  36 static int count_bitmap_extents(struct btrfs_free_space_ctl *ctl,
  37                                 struct btrfs_free_space *bitmap_info);
  38 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  39                            struct btrfs_free_space *info);
  40 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  41                               struct btrfs_free_space *info);
  42 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
  43                              struct btrfs_trans_handle *trans,
  44                              struct btrfs_io_ctl *io_ctl,
  45                              struct btrfs_path *path);
  46
  47 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  48                                                struct btrfs_path *path,
  49                                                u64 offset)
  50 {
  51         struct btrfs_fs_info *fs_info = root->fs_info;
  52         struct btrfs_key key;
  53         struct btrfs_key location;
  54         struct btrfs_disk_key disk_key;
  55         struct btrfs_free_space_header *header;
  56         struct extent_buffer *leaf;
  57         struct inode *inode = NULL;
  58         unsigned nofs_flag;
  59         int ret;
  60
  61         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  62         key.offset = offset;
  63         key.type = 0;
  64
  65         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  66         if (ret < 0)
  67                 return ERR_PTR(ret);
  68         if (ret > 0) {
  69                 btrfs_release_path(path);
  70                 return ERR_PTR(-ENOENT);
  71         }
  72
  73         leaf = path->nodes[0];
  74         header = btrfs_item_ptr(leaf, path->slots[0],
  75                                 struct btrfs_free_space_header);
  76         btrfs_free_space_key(leaf, header, &disk_key);
  77         btrfs_disk_key_to_cpu(&location, &disk_key);
  78         btrfs_release_path(path);
  79
  80         /*
  81          * We are often under a trans handle at this point, so we need to make
  82          * sure NOFS is set to keep us from deadlocking.
  83          */
  84         nofs_flag = memalloc_nofs_save();
  85         inode = btrfs_iget_path(fs_info->sb, &location, root, path);
  86         btrfs_release_path(path);
  87         memalloc_nofs_restore(nofs_flag);
  88         if (IS_ERR(inode))
  89                 return inode;
  90
  91         mapping_set_gfp_mask(inode->i_mapping,
  92                         mapping_gfp_constraint(inode->i_mapping,
  93                         ~(__GFP_FS | __GFP_HIGHMEM)));
  94
  95         return inode;
  96 }
  97
  98 struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
  99                 struct btrfs_path *path)
 100 {
 101         struct btrfs_fs_info *fs_info = block_group->fs_info;
 102         struct inode *inode = NULL;
 103         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 104
 105         spin_lock(&block_group->lock);
 106         if (block_group->inode)
 107                 inode = igrab(block_group->inode);
 108         spin_unlock(&block_group->lock);
 109         if (inode)
 110                 return inode;
 111
 112         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 113                                           block_group->start);
 114         if (IS_ERR(inode))
 115                 return inode;
 116
 117         spin_lock(&block_group->lock);
 118         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 119                 btrfs_info(fs_info, "Old style space inode found, converting.");
 120                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 121                         BTRFS_INODE_NODATACOW;
 122                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 123         }
 124
 125         if (!block_group->iref) {
 126                 block_group->inode = igrab(inode);
 127                 block_group->iref = 1;
 128         }
 129         spin_unlock(&block_group->lock);
 130
 131         return inode;
 132 }
 133
 134 static int __create_free_space_inode(struct btrfs_root *root,
 135                                      struct btrfs_trans_handle *trans,
 136                                      struct btrfs_path *path,
 137                                      u64 ino, u64 offset)
 138 {
 139         struct btrfs_key key;
 140         struct btrfs_disk_key disk_key;
 141         struct btrfs_free_space_header *header;
 142         struct btrfs_inode_item *inode_item;
 143         struct extent_buffer *leaf;
 144         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 145         int ret;
 146
 147         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 148         if (ret)
 149                 return ret;
 150
 151         /* We inline crc's for the free disk space cache */
 152         if (ino != BTRFS_FREE_INO_OBJECTID)
 153                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 154
 155         leaf = path->nodes[0];
 156         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 157                                     struct btrfs_inode_item);
 158         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 159         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 160                              sizeof(*inode_item));
 161         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 162         btrfs_set_inode_size(leaf, inode_item, 0);
 163         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 164         btrfs_set_inode_uid(leaf, inode_item, 0);
 165         btrfs_set_inode_gid(leaf, inode_item, 0);
 166         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 167         btrfs_set_inode_flags(leaf, inode_item, flags);
 168         btrfs_set_inode_nlink(leaf, inode_item, 1);
 169         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 170         btrfs_set_inode_block_group(leaf, inode_item, offset);
 171         btrfs_mark_buffer_dirty(leaf);
 172         btrfs_release_path(path);
 173
 174         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 175         key.offset = offset;
 176         key.type = 0;
 177         ret = btrfs_insert_empty_item(trans, root, path, &key,
 178                                       sizeof(struct btrfs_free_space_header));
 179         if (ret < 0) {
 180                 btrfs_release_path(path);
 181                 return ret;
 182         }
 183
 184         leaf = path->nodes[0];
 185         header = btrfs_item_ptr(leaf, path->slots[0],
 186                                 struct btrfs_free_space_header);
 187         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 188         btrfs_set_free_space_key(leaf, header, &disk_key);
 189         btrfs_mark_buffer_dirty(leaf);
 190         btrfs_release_path(path);
 191
 192         return 0;
 193 }
 194
 195 int create_free_space_inode(struct btrfs_trans_handle *trans,
 196                             struct btrfs_block_group *block_group,
 197                             struct btrfs_path *path)
 198 {
 199         int ret;
 200         u64 ino;
 201
 202         ret = btrfs_find_free_objectid(trans->fs_info->tree_root, &ino);
 203         if (ret < 0)
 204                 return ret;
 205
 206         return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
 207                                          ino, block_group->start);
 208 }
 209
 210 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
 211                                        struct btrfs_block_rsv *rsv)
 212 {
 213         u64 needed_bytes;
 214         int ret;
 215
 216         /* 1 for slack space, 1 for updating the inode */
 217         needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
 218                 btrfs_calc_metadata_size(fs_info, 1);
 219
 220         spin_lock(&rsv->lock);
 221         if (rsv->reserved < needed_bytes)
 222                 ret = -ENOSPC;
 223         else
 224                 ret = 0;
 225         spin_unlock(&rsv->lock);
 226         return ret;
 227 }
 228
 229 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 230                                     struct btrfs_block_group *block_group,
 231                                     struct inode *inode)
 232 {
 233         struct btrfs_root *root = BTRFS_I(inode)->root;
 234         int ret = 0;
 235         bool locked = false;
 236
 237         if (block_group) {
 238                 struct btrfs_path *path = btrfs_alloc_path();
 239
 240                 if (!path) {
 241                         ret = -ENOMEM;
 242                         goto fail;
 243                 }
 244                 locked = true;
 245                 mutex_lock(&trans->transaction->cache_write_mutex);
 246                 if (!list_empty(&block_group->io_list)) {
 247                         list_del_init(&block_group->io_list);
 248
 249                         btrfs_wait_cache_io(trans, block_group, path);
 250                         btrfs_put_block_group(block_group);
 251                 }
 252
 253                 /*
 254                  * now that we've truncated the cache away, its no longer
 255                  * setup or written
 256                  */
 257                 spin_lock(&block_group->lock);
 258                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 259                 spin_unlock(&block_group->lock);
 260                 btrfs_free_path(path);
 261         }
 262
 263         btrfs_i_size_write(BTRFS_I(inode), 0);
 264         truncate_pagecache(inode, 0);
 265
 266         /*
 267          * We skip the throttling logic for free space cache inodes, so we don't
 268          * need to check for -EAGAIN.
 269          */
 270         ret = btrfs_truncate_inode_items(trans, root, inode,
 271                                          0, BTRFS_EXTENT_DATA_KEY);
 272         if (ret)
 273                 goto fail;
 274
 275         ret = btrfs_update_inode(trans, root, inode);
 276
 277 fail:
 278         if (locked)
 279                 mutex_unlock(&trans->transaction->cache_write_mutex);
 280         if (ret)
 281                 btrfs_abort_transaction(trans, ret);
 282
 283         return ret;
 284 }
 285
 286 static void readahead_cache(struct inode *inode)
 287 {
 288         struct file_ra_state *ra;
 289         unsigned long last_index;
 290
 291         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 292         if (!ra)
 293                 return;
 294
 295         file_ra_state_init(ra, inode->i_mapping);
 296         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 297
 298         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 299
 300         kfree(ra);
 301 }
 302
 303 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 304                        int write)
 305 {
 306         int num_pages;
 307         int check_crcs = 0;
 308
 309         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 310
 311         if (btrfs_ino(BTRFS_I(inode)) != BTRFS_FREE_INO_OBJECTID)
 312                 check_crcs = 1;
 313
 314         /* Make sure we can fit our crcs and generation into the first page */
 315         if (write && check_crcs &&
 316             (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
 317                 return -ENOSPC;
 318
 319         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 320
 321         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 322         if (!io_ctl->pages)
 323                 return -ENOMEM;
 324
 325         io_ctl->num_pages = num_pages;
 326         io_ctl->fs_info = btrfs_sb(inode->i_sb);
 327         io_ctl->check_crcs = check_crcs;
 328         io_ctl->inode = inode;
 329
 330         return 0;
 331 }
 332 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 333
 334 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 335 {
 336         kfree(io_ctl->pages);
 337         io_ctl->pages = NULL;
 338 }
 339
 340 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 341 {
 342         if (io_ctl->cur) {
 343                 io_ctl->cur = NULL;
 344                 io_ctl->orig = NULL;
 345         }
 346 }
 347
 348 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 349 {
 350         ASSERT(io_ctl->index < io_ctl->num_pages);
 351         io_ctl->page = io_ctl->pages[io_ctl->index++];
 352         io_ctl->cur = page_address(io_ctl->page);
 353         io_ctl->orig = io_ctl->cur;
 354         io_ctl->size = PAGE_SIZE;
 355         if (clear)
 356                 clear_page(io_ctl->cur);
 357 }
 358
 359 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 360 {
 361         int i;
 362
 363         io_ctl_unmap_page(io_ctl);
 364
 365         for (i = 0; i < io_ctl->num_pages; i++) {
 366                 if (io_ctl->pages[i]) {
 367                         ClearPageChecked(io_ctl->pages[i]);
 368                         unlock_page(io_ctl->pages[i]);
 369                         put_page(io_ctl->pages[i]);
 370                 }
 371         }
 372 }
 373
 374 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 375                                 int uptodate)
 376 {
 377         struct page *page;
 378         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 379         int i;
 380
 381         for (i = 0; i < io_ctl->num_pages; i++) {
 382                 page = find_or_create_page(inode->i_mapping, i, mask);
 383                 if (!page) {
 384                         io_ctl_drop_pages(io_ctl);
 385                         return -ENOMEM;
 386                 }
 387                 io_ctl->pages[i] = page;
 388                 if (uptodate && !PageUptodate(page)) {
 389                         btrfs_readpage(NULL, page);
 390                         lock_page(page);
 391                         if (page->mapping != inode->i_mapping) {
 392                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 393                                           "free space cache page truncated");
 394                                 io_ctl_drop_pages(io_ctl);
 395                                 return -EIO;
 396                         }
 397                         if (!PageUptodate(page)) {
 398                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 399                                            "error reading free space cache");
 400                                 io_ctl_drop_pages(io_ctl);
 401                                 return -EIO;
 402                         }
 403                 }
 404         }
 405
 406         for (i = 0; i < io_ctl->num_pages; i++) {
 407                 clear_page_dirty_for_io(io_ctl->pages[i]);
 408                 set_page_extent_mapped(io_ctl->pages[i]);
 409         }
 410
 411         return 0;
 412 }
 413
 414 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 415 {
 416         __le64 *val;
 417
 418         io_ctl_map_page(io_ctl, 1);
 419
 420         /*
 421          * Skip the csum areas.  If we don't check crcs then we just have a
 422          * 64bit chunk at the front of the first page.
 423          */
 424         if (io_ctl->check_crcs) {
 425                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 426                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 427         } else {
 428                 io_ctl->cur += sizeof(u64);
 429                 io_ctl->size -= sizeof(u64) * 2;
 430         }
 431
 432         val = io_ctl->cur;
 433         *val = cpu_to_le64(generation);
 434         io_ctl->cur += sizeof(u64);
 435 }
 436
 437 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 438 {
 439         __le64 *gen;
 440
 441         /*
 442          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 443          * chunk at the front of the first page.
 444          */
 445         if (io_ctl->check_crcs) {
 446                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 447                 io_ctl->size -= sizeof(u64) +
 448                         (sizeof(u32) * io_ctl->num_pages);
 449         } else {
 450                 io_ctl->cur += sizeof(u64);
 451                 io_ctl->size -= sizeof(u64) * 2;
 452         }
 453
 454         gen = io_ctl->cur;
 455         if (le64_to_cpu(*gen) != generation) {
 456                 btrfs_err_rl(io_ctl->fs_info,
 457                         "space cache generation (%llu) does not match inode (%llu)",
 458                                 *gen, generation);
 459                 io_ctl_unmap_page(io_ctl);
 460                 return -EIO;
 461         }
 462         io_ctl->cur += sizeof(u64);
 463         return 0;
 464 }
 465
 466 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 467 {
 468         u32 *tmp;
 469         u32 crc = ~(u32)0;
 470         unsigned offset = 0;
 471
 472         if (!io_ctl->check_crcs) {
 473                 io_ctl_unmap_page(io_ctl);
 474                 return;
 475         }
 476
 477         if (index == 0)
 478                 offset = sizeof(u32) * io_ctl->num_pages;
 479
 480         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 481         btrfs_crc32c_final(crc, (u8 *)&crc);
 482         io_ctl_unmap_page(io_ctl);
 483         tmp = page_address(io_ctl->pages[0]);
 484         tmp += index;
 485         *tmp = crc;
 486 }
 487
 488 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 489 {
 490         u32 *tmp, val;
 491         u32 crc = ~(u32)0;
 492         unsigned offset = 0;
 493
 494         if (!io_ctl->check_crcs) {
 495                 io_ctl_map_page(io_ctl, 0);
 496                 return 0;
 497         }
 498
 499         if (index == 0)
 500                 offset = sizeof(u32) * io_ctl->num_pages;
 501
 502         tmp = page_address(io_ctl->pages[0]);
 503         tmp += index;
 504         val = *tmp;
 505
 506         io_ctl_map_page(io_ctl, 0);
 507         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 508         btrfs_crc32c_final(crc, (u8 *)&crc);
 509         if (val != crc) {
 510                 btrfs_err_rl(io_ctl->fs_info,
 511                         "csum mismatch on free space cache");
 512                 io_ctl_unmap_page(io_ctl);
 513                 return -EIO;
 514         }
 515
 516         return 0;
 517 }
 518
 519 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 520                             void *bitmap)
 521 {
 522         struct btrfs_free_space_entry *entry;
 523
 524         if (!io_ctl->cur)
 525                 return -ENOSPC;
 526
 527         entry = io_ctl->cur;
 528         entry->offset = cpu_to_le64(offset);
 529         entry->bytes = cpu_to_le64(bytes);
 530         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 531                 BTRFS_FREE_SPACE_EXTENT;
 532         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 533         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 534
 535         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 536                 return 0;
 537
 538         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 539
 540         /* No more pages to map */
 541         if (io_ctl->index >= io_ctl->num_pages)
 542                 return 0;
 543
 544         /* map the next page */
 545         io_ctl_map_page(io_ctl, 1);
 546         return 0;
 547 }
 548
 549 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 550 {
 551         if (!io_ctl->cur)
 552                 return -ENOSPC;
 553
 554         /*
 555          * If we aren't at the start of the current page, unmap this one and
 556          * map the next one if there is any left.
 557          */
 558         if (io_ctl->cur != io_ctl->orig) {
 559                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 560                 if (io_ctl->index >= io_ctl->num_pages)
 561                         return -ENOSPC;
 562                 io_ctl_map_page(io_ctl, 0);
 563         }
 564
 565         copy_page(io_ctl->cur, bitmap);
 566         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 567         if (io_ctl->index < io_ctl->num_pages)
 568                 io_ctl_map_page(io_ctl, 0);
 569         return 0;
 570 }
 571
 572 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 573 {
 574         /*
 575          * If we're not on the boundary we know we've modified the page and we
 576          * need to crc the page.
 577          */
 578         if (io_ctl->cur != io_ctl->orig)
 579                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 580         else
 581                 io_ctl_unmap_page(io_ctl);
 582
 583         while (io_ctl->index < io_ctl->num_pages) {
 584                 io_ctl_map_page(io_ctl, 1);
 585                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 586         }
 587 }
 588
 589 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 590                             struct btrfs_free_space *entry, u8 *type)
 591 {
 592         struct btrfs_free_space_entry *e;
 593         int ret;
 594
 595         if (!io_ctl->cur) {
 596                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 597                 if (ret)
 598                         return ret;
 599         }
 600
 601         e = io_ctl->cur;
 602         entry->offset = le64_to_cpu(e->offset);
 603         entry->bytes = le64_to_cpu(e->bytes);
 604         *type = e->type;
 605         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 606         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 607
 608         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 609                 return 0;
 610
 611         io_ctl_unmap_page(io_ctl);
 612
 613         return 0;
 614 }
 615
 616 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 617                               struct btrfs_free_space *entry)
 618 {
 619         int ret;
 620
 621         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 622         if (ret)
 623                 return ret;
 624
 625         copy_page(entry->bitmap, io_ctl->cur);
 626         io_ctl_unmap_page(io_ctl);
 627
 628         return 0;
 629 }
 630
 631 /*
 632  * Since we attach pinned extents after the fact we can have contiguous sections
 633  * of free space that are split up in entries.  This poses a problem with the
 634  * tree logging stuff since it could have allocated across what appears to be 2
 635  * entries since we would have merged the entries when adding the pinned extents
 636  * back to the free space cache.  So run through the space cache that we just
 637  * loaded and merge contiguous entries.  This will make the log replay stuff not
 638  * blow up and it will make for nicer allocator behavior.
 639  */
 640 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 641 {
 642         struct btrfs_free_space *e, *prev = NULL;
 643         struct rb_node *n;
 644
 645 again:
 646         spin_lock(&ctl->tree_lock);
 647         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 648                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 649                 if (!prev)
 650                         goto next;
 651                 if (e->bitmap || prev->bitmap)
 652                         goto next;
 653                 if (prev->offset + prev->bytes == e->offset) {
 654                         unlink_free_space(ctl, prev);
 655                         unlink_free_space(ctl, e);
 656                         prev->bytes += e->bytes;
 657                         kmem_cache_free(btrfs_free_space_cachep, e);
 658                         link_free_space(ctl, prev);
 659                         prev = NULL;
 660                         spin_unlock(&ctl->tree_lock);
 661                         goto again;
 662                 }
 663 next:
 664                 prev = e;
 665         }
 666         spin_unlock(&ctl->tree_lock);
 667 }
 668
 669 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 670                                    struct btrfs_free_space_ctl *ctl,
 671                                    struct btrfs_path *path, u64 offset)
 672 {
 673         struct btrfs_fs_info *fs_info = root->fs_info;
 674         struct btrfs_free_space_header *header;
 675         struct extent_buffer *leaf;
 676         struct btrfs_io_ctl io_ctl;
 677         struct btrfs_key key;
 678         struct btrfs_free_space *e, *n;
 679         LIST_HEAD(bitmaps);
 680         u64 num_entries;
 681         u64 num_bitmaps;
 682         u64 generation;
 683         u8 type;
 684         int ret = 0;
 685
 686         /* Nothing in the space cache, goodbye */
 687         if (!i_size_read(inode))
 688                 return 0;
 689
 690         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 691         key.offset = offset;
 692         key.type = 0;
 693
 694         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 695         if (ret < 0)
 696                 return 0;
 697         else if (ret > 0) {
 698                 btrfs_release_path(path);
 699                 return 0;
 700         }
 701
 702         ret = -1;
 703
 704         leaf = path->nodes[0];
 705         header = btrfs_item_ptr(leaf, path->slots[0],
 706                                 struct btrfs_free_space_header);
 707         num_entries = btrfs_free_space_entries(leaf, header);
 708         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 709         generation = btrfs_free_space_generation(leaf, header);
 710         btrfs_release_path(path);
 711
 712         if (!BTRFS_I(inode)->generation) {
 713                 btrfs_info(fs_info,
 714                            "the free space cache file (%llu) is invalid, skip it",
 715                            offset);
 716                 return 0;
 717         }
 718
 719         if (BTRFS_I(inode)->generation != generation) {
 720                 btrfs_err(fs_info,
 721                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 722                           BTRFS_I(inode)->generation, generation);
 723                 return 0;
 724         }
 725
 726         if (!num_entries)
 727                 return 0;
 728
 729         ret = io_ctl_init(&io_ctl, inode, 0);
 730         if (ret)
 731                 return ret;
 732
 733         readahead_cache(inode);
 734
 735         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 736         if (ret)
 737                 goto out;
 738
 739         ret = io_ctl_check_crc(&io_ctl, 0);
 740         if (ret)
 741                 goto free_cache;
 742
 743         ret = io_ctl_check_generation(&io_ctl, generation);
 744         if (ret)
 745                 goto free_cache;
 746
 747         while (num_entries) {
 748                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 749                                       GFP_NOFS);
 750                 if (!e)
 751                         goto free_cache;
 752
 753                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 754                 if (ret) {
 755                         kmem_cache_free(btrfs_free_space_cachep, e);
 756                         goto free_cache;
 757                 }
 758
 759                 /*
 760                  * Sync discard ensures that the free space cache is always
 761                  * trimmed.  So when reading this in, the state should reflect
 762                  * that.  We also do this for async as a stop gap for lack of
 763                  * persistence.
 764                  */
 765                 if (btrfs_test_opt(fs_info, DISCARD_SYNC) ||
 766                     btrfs_test_opt(fs_info, DISCARD_ASYNC))
 767                         e->trim_state = BTRFS_TRIM_STATE_TRIMMED;
 768
 769                 if (!e->bytes) {
 770                         kmem_cache_free(btrfs_free_space_cachep, e);
 771                         goto free_cache;
 772                 }
 773
 774                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 775                         spin_lock(&ctl->tree_lock);
 776                         ret = link_free_space(ctl, e);
 777                         spin_unlock(&ctl->tree_lock);
 778                         if (ret) {
 779                                 btrfs_err(fs_info,
 780                                         "Duplicate entries in free space cache, dumping");
 781                                 kmem_cache_free(btrfs_free_space_cachep, e);
 782                                 goto free_cache;
 783                         }
 784                 } else {
 785                         ASSERT(num_bitmaps);
 786                         num_bitmaps--;
 787                         e->bitmap = kmem_cache_zalloc(
 788                                         btrfs_free_space_bitmap_cachep, GFP_NOFS);
 789                         if (!e->bitmap) {
 790                                 kmem_cache_free(
 791                                         btrfs_free_space_cachep, e);
 792                                 goto free_cache;
 793                         }
 794                         spin_lock(&ctl->tree_lock);
 795                         ret = link_free_space(ctl, e);
 796                         ctl->total_bitmaps++;
 797                         ctl->op->recalc_thresholds(ctl);
 798                         spin_unlock(&ctl->tree_lock);
 799                         if (ret) {
 800                                 btrfs_err(fs_info,
 801                                         "Duplicate entries in free space cache, dumping");
 802                                 kmem_cache_free(btrfs_free_space_cachep, e);
 803                                 goto free_cache;
 804                         }
 805                         list_add_tail(&e->list, &bitmaps);
 806                 }
 807
 808                 num_entries--;
 809         }
 810
 811         io_ctl_unmap_page(&io_ctl);
 812
 813         /*
 814          * We add the bitmaps at the end of the entries in order that
 815          * the bitmap entries are added to the cache.
 816          */
 817         list_for_each_entry_safe(e, n, &bitmaps, list) {
 818                 list_del_init(&e->list);
 819                 ret = io_ctl_read_bitmap(&io_ctl, e);
 820                 if (ret)
 821                         goto free_cache;
 822                 e->bitmap_extents = count_bitmap_extents(ctl, e);
 823                 if (!btrfs_free_space_trimmed(e)) {
 824                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
 825                                 e->bitmap_extents;
 826                         ctl->discardable_bytes[BTRFS_STAT_CURR] += e->bytes;
 827                 }
 828         }
 829
 830         io_ctl_drop_pages(&io_ctl);
 831         merge_space_tree(ctl);
 832         ret = 1;
 833 out:
 834         btrfs_discard_update_discardable(ctl->private, ctl);
 835         io_ctl_free(&io_ctl);
 836         return ret;
 837 free_cache:
 838         io_ctl_drop_pages(&io_ctl);
 839         __btrfs_remove_free_space_cache(ctl);
 840         goto out;
 841 }
 842
 843 int load_free_space_cache(struct btrfs_block_group *block_group)
 844 {
 845         struct btrfs_fs_info *fs_info = block_group->fs_info;
 846         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 847         struct inode *inode;
 848         struct btrfs_path *path;
 849         int ret = 0;
 850         bool matched;
 851         u64 used = block_group->used;
 852
 853         /*
 854          * If this block group has been marked to be cleared for one reason or
 855          * another then we can't trust the on disk cache, so just return.
 856          */
 857         spin_lock(&block_group->lock);
 858         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 859                 spin_unlock(&block_group->lock);
 860                 return 0;
 861         }
 862         spin_unlock(&block_group->lock);
 863
 864         path = btrfs_alloc_path();
 865         if (!path)
 866                 return 0;
 867         path->search_commit_root = 1;
 868         path->skip_locking = 1;
 869
 870         /*
 871          * We must pass a path with search_commit_root set to btrfs_iget in
 872          * order to avoid a deadlock when allocating extents for the tree root.
 873          *
 874          * When we are COWing an extent buffer from the tree root, when looking
 875          * for a free extent, at extent-tree.c:find_free_extent(), we can find
 876          * block group without its free space cache loaded. When we find one
 877          * we must load its space cache which requires reading its free space
 878          * cache's inode item from the root tree. If this inode item is located
 879          * in the same leaf that we started COWing before, then we end up in
 880          * deadlock on the extent buffer (trying to read lock it when we
 881          * previously write locked it).
 882          *
 883          * It's safe to read the inode item using the commit root because
 884          * block groups, once loaded, stay in memory forever (until they are
 885          * removed) as well as their space caches once loaded. New block groups
 886          * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
 887          * we will never try to read their inode item while the fs is mounted.
 888          */
 889         inode = lookup_free_space_inode(block_group, path);
 890         if (IS_ERR(inode)) {
 891                 btrfs_free_path(path);
 892                 return 0;
 893         }
 894
 895         /* We may have converted the inode and made the cache invalid. */
 896         spin_lock(&block_group->lock);
 897         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 898                 spin_unlock(&block_group->lock);
 899                 btrfs_free_path(path);
 900                 goto out;
 901         }
 902         spin_unlock(&block_group->lock);
 903
 904         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 905                                       path, block_group->start);
 906         btrfs_free_path(path);
 907         if (ret <= 0)
 908                 goto out;
 909
 910         spin_lock(&ctl->tree_lock);
 911         matched = (ctl->free_space == (block_group->length - used -
 912                                        block_group->bytes_super));
 913         spin_unlock(&ctl->tree_lock);
 914
 915         if (!matched) {
 916                 __btrfs_remove_free_space_cache(ctl);
 917                 btrfs_warn(fs_info,
 918                            "block group %llu has wrong amount of free space",
 919                            block_group->start);
 920                 ret = -1;
 921         }
 922 out:
 923         if (ret < 0) {
 924                 /* This cache is bogus, make sure it gets cleared */
 925                 spin_lock(&block_group->lock);
 926                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 927                 spin_unlock(&block_group->lock);
 928                 ret = 0;
 929
 930                 btrfs_warn(fs_info,
 931                            "failed to load free space cache for block group %llu, rebuilding it now",
 932                            block_group->start);
 933         }
 934
 935         iput(inode);
 936         return ret;
 937 }
 938
 939 static noinline_for_stack
 940 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
 941                               struct btrfs_free_space_ctl *ctl,
 942                               struct btrfs_block_group *block_group,
 943                               int *entries, int *bitmaps,
 944                               struct list_head *bitmap_list)
 945 {
 946         int ret;
 947         struct btrfs_free_cluster *cluster = NULL;
 948         struct btrfs_free_cluster *cluster_locked = NULL;
 949         struct rb_node *node = rb_first(&ctl->free_space_offset);
 950         struct btrfs_trim_range *trim_entry;
 951
 952         /* Get the cluster for this block_group if it exists */
 953         if (block_group && !list_empty(&block_group->cluster_list)) {
 954                 cluster = list_entry(block_group->cluster_list.next,
 955                                      struct btrfs_free_cluster,
 956                                      block_group_list);
 957         }
 958
 959         if (!node && cluster) {
 960                 cluster_locked = cluster;
 961                 spin_lock(&cluster_locked->lock);
 962                 node = rb_first(&cluster->root);
 963                 cluster = NULL;
 964         }
 965
 966         /* Write out the extent entries */
 967         while (node) {
 968                 struct btrfs_free_space *e;
 969
 970                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 971                 *entries += 1;
 972
 973                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
 974                                        e->bitmap);
 975                 if (ret)
 976                         goto fail;
 977
 978                 if (e->bitmap) {
 979                         list_add_tail(&e->list, bitmap_list);
 980                         *bitmaps += 1;
 981                 }
 982                 node = rb_next(node);
 983                 if (!node && cluster) {
 984                         node = rb_first(&cluster->root);
 985                         cluster_locked = cluster;
 986                         spin_lock(&cluster_locked->lock);
 987                         cluster = NULL;
 988                 }
 989         }
 990         if (cluster_locked) {
 991                 spin_unlock(&cluster_locked->lock);
 992                 cluster_locked = NULL;
 993         }
 994
 995         /*
 996          * Make sure we don't miss any range that was removed from our rbtree
 997          * because trimming is running. Otherwise after a umount+mount (or crash
 998          * after committing the transaction) we would leak free space and get
 999          * an inconsistent free space cache report from fsck.
1000          */
1001         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1002                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1003                                        trim_entry->bytes, NULL);
1004                 if (ret)
1005                         goto fail;
1006                 *entries += 1;
1007         }
1008
1009         return 0;
1010 fail:
1011         if (cluster_locked)
1012                 spin_unlock(&cluster_locked->lock);
1013         return -ENOSPC;
1014 }
1015
1016 static noinline_for_stack int
1017 update_cache_item(struct btrfs_trans_handle *trans,
1018                   struct btrfs_root *root,
1019                   struct inode *inode,
1020                   struct btrfs_path *path, u64 offset,
1021                   int entries, int bitmaps)
1022 {
1023         struct btrfs_key key;
1024         struct btrfs_free_space_header *header;
1025         struct extent_buffer *leaf;
1026         int ret;
1027
1028         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1029         key.offset = offset;
1030         key.type = 0;
1031
1032         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1033         if (ret < 0) {
1034                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1035                                  EXTENT_DELALLOC, 0, 0, NULL);
1036                 goto fail;
1037         }
1038         leaf = path->nodes[0];
1039         if (ret > 0) {
1040                 struct btrfs_key found_key;
1041                 ASSERT(path->slots[0]);
1042                 path->slots[0]--;
1043                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1044                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1045                     found_key.offset != offset) {
1046                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1047                                          inode->i_size - 1, EXTENT_DELALLOC, 0,
1048                                          0, NULL);
1049                         btrfs_release_path(path);
1050                         goto fail;
1051                 }
1052         }
1053
1054         BTRFS_I(inode)->generation = trans->transid;
1055         header = btrfs_item_ptr(leaf, path->slots[0],
1056                                 struct btrfs_free_space_header);
1057         btrfs_set_free_space_entries(leaf, header, entries);
1058         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1059         btrfs_set_free_space_generation(leaf, header, trans->transid);
1060         btrfs_mark_buffer_dirty(leaf);
1061         btrfs_release_path(path);
1062
1063         return 0;
1064
1065 fail:
1066         return -1;
1067 }
1068
1069 static noinline_for_stack int write_pinned_extent_entries(
1070                             struct btrfs_block_group *block_group,
1071                             struct btrfs_io_ctl *io_ctl,
1072                             int *entries)
1073 {
1074         u64 start, extent_start, extent_end, len;
1075         struct extent_io_tree *unpin = NULL;
1076         int ret;
1077
1078         if (!block_group)
1079                 return 0;
1080
1081         /*
1082          * We want to add any pinned extents to our free space cache
1083          * so we don't leak the space
1084          *
1085          * We shouldn't have switched the pinned extents yet so this is the
1086          * right one
1087          */
1088         unpin = block_group->fs_info->pinned_extents;
1089
1090         start = block_group->start;
1091
1092         while (start < block_group->start + block_group->length) {
1093                 ret = find_first_extent_bit(unpin, start,
1094                                             &extent_start, &extent_end,
1095                                             EXTENT_DIRTY, NULL);
1096                 if (ret)
1097                         return 0;
1098
1099                 /* This pinned extent is out of our range */
1100                 if (extent_start >= block_group->start + block_group->length)
1101                         return 0;
1102
1103                 extent_start = max(extent_start, start);
1104                 extent_end = min(block_group->start + block_group->length,
1105                                  extent_end + 1);
1106                 len = extent_end - extent_start;
1107
1108                 *entries += 1;
1109                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1110                 if (ret)
1111                         return -ENOSPC;
1112
1113                 start = extent_end;
1114         }
1115
1116         return 0;
1117 }
1118
1119 static noinline_for_stack int
1120 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1121 {
1122         struct btrfs_free_space *entry, *next;
1123         int ret;
1124
1125         /* Write out the bitmaps */
1126         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1127                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1128                 if (ret)
1129                         return -ENOSPC;
1130                 list_del_init(&entry->list);
1131         }
1132
1133         return 0;
1134 }
1135
1136 static int flush_dirty_cache(struct inode *inode)
1137 {
1138         int ret;
1139
1140         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1141         if (ret)
1142                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1143                                  EXTENT_DELALLOC, 0, 0, NULL);
1144
1145         return ret;
1146 }
1147
1148 static void noinline_for_stack
1149 cleanup_bitmap_list(struct list_head *bitmap_list)
1150 {
1151         struct btrfs_free_space *entry, *next;
1152
1153         list_for_each_entry_safe(entry, next, bitmap_list, list)
1154                 list_del_init(&entry->list);
1155 }
1156
1157 static void noinline_for_stack
1158 cleanup_write_cache_enospc(struct inode *inode,
1159                            struct btrfs_io_ctl *io_ctl,
1160                            struct extent_state **cached_state)
1161 {
1162         io_ctl_drop_pages(io_ctl);
1163         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1164                              i_size_read(inode) - 1, cached_state);
1165 }
1166
1167 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1168                                  struct btrfs_trans_handle *trans,
1169                                  struct btrfs_block_group *block_group,
1170                                  struct btrfs_io_ctl *io_ctl,
1171                                  struct btrfs_path *path, u64 offset)
1172 {
1173         int ret;
1174         struct inode *inode = io_ctl->inode;
1175
1176         if (!inode)
1177                 return 0;
1178
1179         /* Flush the dirty pages in the cache file. */
1180         ret = flush_dirty_cache(inode);
1181         if (ret)
1182                 goto out;
1183
1184         /* Update the cache item to tell everyone this cache file is valid. */
1185         ret = update_cache_item(trans, root, inode, path, offset,
1186                                 io_ctl->entries, io_ctl->bitmaps);
1187 out:
1188         io_ctl_free(io_ctl);
1189         if (ret) {
1190                 invalidate_inode_pages2(inode->i_mapping);
1191                 BTRFS_I(inode)->generation = 0;
1192                 if (block_group) {
1193 #ifdef DEBUG
1194                         btrfs_err(root->fs_info,
1195                                   "failed to write free space cache for block group %llu",
1196                                   block_group->start);
1197 #endif
1198                 }
1199         }
1200         btrfs_update_inode(trans, root, inode);
1201
1202         if (block_group) {
1203                 /* the dirty list is protected by the dirty_bgs_lock */
1204                 spin_lock(&trans->transaction->dirty_bgs_lock);
1205
1206                 /* the disk_cache_state is protected by the block group lock */
1207                 spin_lock(&block_group->lock);
1208
1209                 /*
1210                  * only mark this as written if we didn't get put back on
1211                  * the dirty list while waiting for IO.   Otherwise our
1212                  * cache state won't be right, and we won't get written again
1213                  */
1214                 if (!ret && list_empty(&block_group->dirty_list))
1215                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1216                 else if (ret)
1217                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1218
1219                 spin_unlock(&block_group->lock);
1220                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1221                 io_ctl->inode = NULL;
1222                 iput(inode);
1223         }
1224
1225         return ret;
1226
1227 }
1228
1229 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
1230                                     struct btrfs_trans_handle *trans,
1231                                     struct btrfs_io_ctl *io_ctl,
1232                                     struct btrfs_path *path)
1233 {
1234         return __btrfs_wait_cache_io(root, trans, NULL, io_ctl, path, 0);
1235 }
1236
1237 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1238                         struct btrfs_block_group *block_group,
1239                         struct btrfs_path *path)
1240 {
1241         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1242                                      block_group, &block_group->io_ctl,
1243                                      path, block_group->start);
1244 }
1245
1246 /**
1247  * __btrfs_write_out_cache - write out cached info to an inode
1248  * @root - the root the inode belongs to
1249  * @ctl - the free space cache we are going to write out
1250  * @block_group - the block_group for this cache if it belongs to a block_group
1251  * @trans - the trans handle
1252  *
1253  * This function writes out a free space cache struct to disk for quick recovery
1254  * on mount.  This will return 0 if it was successful in writing the cache out,
1255  * or an errno if it was not.
1256  */
1257 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1258                                    struct btrfs_free_space_ctl *ctl,
1259                                    struct btrfs_block_group *block_group,
1260                                    struct btrfs_io_ctl *io_ctl,
1261                                    struct btrfs_trans_handle *trans)
1262 {
1263         struct extent_state *cached_state = NULL;
1264         LIST_HEAD(bitmap_list);
1265         int entries = 0;
1266         int bitmaps = 0;
1267         int ret;
1268         int must_iput = 0;
1269
1270         if (!i_size_read(inode))
1271                 return -EIO;
1272
1273         WARN_ON(io_ctl->pages);
1274         ret = io_ctl_init(io_ctl, inode, 1);
1275         if (ret)
1276                 return ret;
1277
1278         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1279                 down_write(&block_group->data_rwsem);
1280                 spin_lock(&block_group->lock);
1281                 if (block_group->delalloc_bytes) {
1282                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1283                         spin_unlock(&block_group->lock);
1284                         up_write(&block_group->data_rwsem);
1285                         BTRFS_I(inode)->generation = 0;
1286                         ret = 0;
1287                         must_iput = 1;
1288                         goto out;
1289                 }
1290                 spin_unlock(&block_group->lock);
1291         }
1292
1293         /* Lock all pages first so we can lock the extent safely. */
1294         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1295         if (ret)
1296                 goto out_unlock;
1297
1298         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1299                          &cached_state);
1300
1301         io_ctl_set_generation(io_ctl, trans->transid);
1302
1303         mutex_lock(&ctl->cache_writeout_mutex);
1304         /* Write out the extent entries in the free space cache */
1305         spin_lock(&ctl->tree_lock);
1306         ret = write_cache_extent_entries(io_ctl, ctl,
1307                                          block_group, &entries, &bitmaps,
1308                                          &bitmap_list);
1309         if (ret)
1310                 goto out_nospc_locked;
1311
1312         /*
1313          * Some spaces that are freed in the current transaction are pinned,
1314          * they will be added into free space cache after the transaction is
1315          * committed, we shouldn't lose them.
1316          *
1317          * If this changes while we are working we'll get added back to
1318          * the dirty list and redo it.  No locking needed
1319          */
1320         ret = write_pinned_extent_entries(block_group, io_ctl, &entries);
1321         if (ret)
1322                 goto out_nospc_locked;
1323
1324         /*
1325          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1326          * locked while doing it because a concurrent trim can be manipulating
1327          * or freeing the bitmap.
1328          */
1329         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1330         spin_unlock(&ctl->tree_lock);
1331         mutex_unlock(&ctl->cache_writeout_mutex);
1332         if (ret)
1333                 goto out_nospc;
1334
1335         /* Zero out the rest of the pages just to make sure */
1336         io_ctl_zero_remaining_pages(io_ctl);
1337
1338         /* Everything is written out, now we dirty the pages in the file. */
1339         ret = btrfs_dirty_pages(inode, io_ctl->pages, io_ctl->num_pages, 0,
1340                                 i_size_read(inode), &cached_state);
1341         if (ret)
1342                 goto out_nospc;
1343
1344         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1345                 up_write(&block_group->data_rwsem);
1346         /*
1347          * Release the pages and unlock the extent, we will flush
1348          * them out later
1349          */
1350         io_ctl_drop_pages(io_ctl);
1351
1352         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1353                              i_size_read(inode) - 1, &cached_state);
1354
1355         /*
1356          * at this point the pages are under IO and we're happy,
1357          * The caller is responsible for waiting on them and updating the
1358          * the cache and the inode
1359          */
1360         io_ctl->entries = entries;
1361         io_ctl->bitmaps = bitmaps;
1362
1363         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1364         if (ret)
1365                 goto out;
1366
1367         return 0;
1368
1369 out:
1370         io_ctl->inode = NULL;
1371         io_ctl_free(io_ctl);
1372         if (ret) {
1373                 invalidate_inode_pages2(inode->i_mapping);
1374                 BTRFS_I(inode)->generation = 0;
1375         }
1376         btrfs_update_inode(trans, root, inode);
1377         if (must_iput)
1378                 iput(inode);
1379         return ret;
1380
1381 out_nospc_locked:
1382         cleanup_bitmap_list(&bitmap_list);
1383         spin_unlock(&ctl->tree_lock);
1384         mutex_unlock(&ctl->cache_writeout_mutex);
1385
1386 out_nospc:
1387         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1388
1389 out_unlock:
1390         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1391                 up_write(&block_group->data_rwsem);
1392
1393         goto out;
1394 }
1395
1396 int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1397                           struct btrfs_block_group *block_group,
1398                           struct btrfs_path *path)
1399 {
1400         struct btrfs_fs_info *fs_info = trans->fs_info;
1401         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1402         struct inode *inode;
1403         int ret = 0;
1404
1405         spin_lock(&block_group->lock);
1406         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1407                 spin_unlock(&block_group->lock);
1408                 return 0;
1409         }
1410         spin_unlock(&block_group->lock);
1411
1412         inode = lookup_free_space_inode(block_group, path);
1413         if (IS_ERR(inode))
1414                 return 0;
1415
1416         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1417                                 block_group, &block_group->io_ctl, trans);
1418         if (ret) {
1419 #ifdef DEBUG
1420                 btrfs_err(fs_info,
1421                           "failed to write free space cache for block group %llu",
1422                           block_group->start);
1423 #endif
1424                 spin_lock(&block_group->lock);
1425                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1426                 spin_unlock(&block_group->lock);
1427
1428                 block_group->io_ctl.inode = NULL;
1429                 iput(inode);
1430         }
1431
1432         /*
1433          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1434          * to wait for IO and put the inode
1435          */
1436
1437         return ret;
1438 }
1439
1440 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1441                                           u64 offset)
1442 {
1443         ASSERT(offset >= bitmap_start);
1444         offset -= bitmap_start;
1445         return (unsigned long)(div_u64(offset, unit));
1446 }
1447
1448 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1449 {
1450         return (unsigned long)(div_u64(bytes, unit));
1451 }
1452
1453 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1454                                    u64 offset)
1455 {
1456         u64 bitmap_start;
1457         u64 bytes_per_bitmap;
1458
1459         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1460         bitmap_start = offset - ctl->start;
1461         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1462         bitmap_start *= bytes_per_bitmap;
1463         bitmap_start += ctl->start;
1464
1465         return bitmap_start;
1466 }
1467
1468 static int tree_insert_offset(struct rb_root *root, u64 offset,
1469                               struct rb_node *node, int bitmap)
1470 {
1471         struct rb_node **p = &root->rb_node;
1472         struct rb_node *parent = NULL;
1473         struct btrfs_free_space *info;
1474
1475         while (*p) {
1476                 parent = *p;
1477                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1478
1479                 if (offset < info->offset) {
1480                         p = &(*p)->rb_left;
1481                 } else if (offset > info->offset) {
1482                         p = &(*p)->rb_right;
1483                 } else {
1484                         /*
1485                          * we could have a bitmap entry and an extent entry
1486                          * share the same offset.  If this is the case, we want
1487                          * the extent entry to always be found first if we do a
1488                          * linear search through the tree, since we want to have
1489                          * the quickest allocation time, and allocating from an
1490                          * extent is faster than allocating from a bitmap.  So
1491                          * if we're inserting a bitmap and we find an entry at
1492                          * this offset, we want to go right, or after this entry
1493                          * logically.  If we are inserting an extent and we've
1494                          * found a bitmap, we want to go left, or before
1495                          * logically.
1496                          */
1497                         if (bitmap) {
1498                                 if (info->bitmap) {
1499                                         WARN_ON_ONCE(1);
1500                                         return -EEXIST;
1501                                 }
1502                                 p = &(*p)->rb_right;
1503                         } else {
1504                                 if (!info->bitmap) {
1505                                         WARN_ON_ONCE(1);
1506                                         return -EEXIST;
1507                                 }
1508                                 p = &(*p)->rb_left;
1509                         }
1510                 }
1511         }
1512
1513         rb_link_node(node, parent, p);
1514         rb_insert_color(node, root);
1515
1516         return 0;
1517 }
1518
1519 /*
1520  * searches the tree for the given offset.
1521  *
1522  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1523  * want a section that has at least bytes size and comes at or after the given
1524  * offset.
1525  */
1526 static struct btrfs_free_space *
1527 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1528                    u64 offset, int bitmap_only, int fuzzy)
1529 {
1530         struct rb_node *n = ctl->free_space_offset.rb_node;
1531         struct btrfs_free_space *entry, *prev = NULL;
1532
1533         /* find entry that is closest to the 'offset' */
1534         while (1) {
1535                 if (!n) {
1536                         entry = NULL;
1537                         break;
1538                 }
1539
1540                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1541                 prev = entry;
1542
1543                 if (offset < entry->offset)
1544                         n = n->rb_left;
1545                 else if (offset > entry->offset)
1546                         n = n->rb_right;
1547                 else
1548                         break;
1549         }
1550
1551         if (bitmap_only) {
1552                 if (!entry)
1553                         return NULL;
1554                 if (entry->bitmap)
1555                         return entry;
1556
1557                 /*
1558                  * bitmap entry and extent entry may share same offset,
1559                  * in that case, bitmap entry comes after extent entry.
1560                  */
1561                 n = rb_next(n);
1562                 if (!n)
1563                         return NULL;
1564                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1565                 if (entry->offset != offset)
1566                         return NULL;
1567
1568                 WARN_ON(!entry->bitmap);
1569                 return entry;
1570         } else if (entry) {
1571                 if (entry->bitmap) {
1572                         /*
1573                          * if previous extent entry covers the offset,
1574                          * we should return it instead of the bitmap entry
1575                          */
1576                         n = rb_prev(&entry->offset_index);
1577                         if (n) {
1578                                 prev = rb_entry(n, struct btrfs_free_space,
1579                                                 offset_index);
1580                                 if (!prev->bitmap &&
1581                                     prev->offset + prev->bytes > offset)
1582                                         entry = prev;
1583                         }
1584                 }
1585                 return entry;
1586         }
1587
1588         if (!prev)
1589                 return NULL;
1590
1591         /* find last entry before the 'offset' */
1592         entry = prev;
1593         if (entry->offset > offset) {
1594                 n = rb_prev(&entry->offset_index);
1595                 if (n) {
1596                         entry = rb_entry(n, struct btrfs_free_space,
1597                                         offset_index);
1598                         ASSERT(entry->offset <= offset);
1599                 } else {
1600                         if (fuzzy)
1601                                 return entry;
1602                         else
1603                                 return NULL;
1604                 }
1605         }
1606
1607         if (entry->bitmap) {
1608                 n = rb_prev(&entry->offset_index);
1609                 if (n) {
1610                         prev = rb_entry(n, struct btrfs_free_space,
1611                                         offset_index);
1612                         if (!prev->bitmap &&
1613                             prev->offset + prev->bytes > offset)
1614                                 return prev;
1615                 }
1616                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1617                         return entry;
1618         } else if (entry->offset + entry->bytes > offset)
1619                 return entry;
1620
1621         if (!fuzzy)
1622                 return NULL;
1623
1624         while (1) {
1625                 if (entry->bitmap) {
1626                         if (entry->offset + BITS_PER_BITMAP *
1627                             ctl->unit > offset)
1628                                 break;
1629                 } else {
1630                         if (entry->offset + entry->bytes > offset)
1631                                 break;
1632                 }
1633
1634                 n = rb_next(&entry->offset_index);
1635                 if (!n)
1636                         return NULL;
1637                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1638         }
1639         return entry;
1640 }
1641
1642 static inline void
1643 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1644                     struct btrfs_free_space *info)
1645 {
1646         rb_erase(&info->offset_index, &ctl->free_space_offset);
1647         ctl->free_extents--;
1648
1649         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1650                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1651                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1652         }
1653 }
1654
1655 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1656                               struct btrfs_free_space *info)
1657 {
1658         __unlink_free_space(ctl, info);
1659         ctl->free_space -= info->bytes;
1660 }
1661
1662 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1663                            struct btrfs_free_space *info)
1664 {
1665         int ret = 0;
1666
1667         ASSERT(info->bytes || info->bitmap);
1668         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1669                                  &info->offset_index, (info->bitmap != NULL));
1670         if (ret)
1671                 return ret;
1672
1673         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1674                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1675                 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1676         }
1677
1678         ctl->free_space += info->bytes;
1679         ctl->free_extents++;
1680         return ret;
1681 }
1682
1683 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1684 {
1685         struct btrfs_block_group *block_group = ctl->private;
1686         u64 max_bytes;
1687         u64 bitmap_bytes;
1688         u64 extent_bytes;
1689         u64 size = block_group->length;
1690         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1691         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1692
1693         max_bitmaps = max_t(u64, max_bitmaps, 1);
1694
1695         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1696
1697         /*
1698          * We are trying to keep the total amount of memory used per 1GiB of
1699          * space to be MAX_CACHE_BYTES_PER_GIG.  However, with a reclamation
1700          * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
1701          * bitmaps, we may end up using more memory than this.
1702          */
1703         if (size < SZ_1G)
1704                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1705         else
1706                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1707
1708         bitmap_bytes = ctl->total_bitmaps * ctl->unit;
1709
1710         /*
1711          * we want the extent entry threshold to always be at most 1/2 the max
1712          * bytes we can have, or whatever is less than that.
1713          */
1714         extent_bytes = max_bytes - bitmap_bytes;
1715         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1716
1717         ctl->extents_thresh =
1718                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1719 }
1720
1721 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1722                                        struct btrfs_free_space *info,
1723                                        u64 offset, u64 bytes)
1724 {
1725         unsigned long start, count, end;
1726         int extent_delta = -1;
1727
1728         start = offset_to_bit(info->offset, ctl->unit, offset);
1729         count = bytes_to_bits(bytes, ctl->unit);
1730         end = start + count;
1731         ASSERT(end <= BITS_PER_BITMAP);
1732
1733         bitmap_clear(info->bitmap, start, count);
1734
1735         info->bytes -= bytes;
1736         if (info->max_extent_size > ctl->unit)
1737                 info->max_extent_size = 0;
1738
1739         if (start && test_bit(start - 1, info->bitmap))
1740                 extent_delta++;
1741
1742         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1743                 extent_delta++;
1744
1745         info->bitmap_extents += extent_delta;
1746         if (!btrfs_free_space_trimmed(info)) {
1747                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1748                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1749         }
1750 }
1751
1752 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1753                               struct btrfs_free_space *info, u64 offset,
1754                               u64 bytes)
1755 {
1756         __bitmap_clear_bits(ctl, info, offset, bytes);
1757         ctl->free_space -= bytes;
1758 }
1759
1760 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1761                             struct btrfs_free_space *info, u64 offset,
1762                             u64 bytes)
1763 {
1764         unsigned long start, count, end;
1765         int extent_delta = 1;
1766
1767         start = offset_to_bit(info->offset, ctl->unit, offset);
1768         count = bytes_to_bits(bytes, ctl->unit);
1769         end = start + count;
1770         ASSERT(end <= BITS_PER_BITMAP);
1771
1772         bitmap_set(info->bitmap, start, count);
1773
1774         info->bytes += bytes;
1775         ctl->free_space += bytes;
1776
1777         if (start && test_bit(start - 1, info->bitmap))
1778                 extent_delta--;
1779
1780         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1781                 extent_delta--;
1782
1783         info->bitmap_extents += extent_delta;
1784         if (!btrfs_free_space_trimmed(info)) {
1785                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1786                 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1787         }
1788 }
1789
1790 /*
1791  * If we can not find suitable extent, we will use bytes to record
1792  * the size of the max extent.
1793  */
1794 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1795                          struct btrfs_free_space *bitmap_info, u64 *offset,
1796                          u64 *bytes, bool for_alloc)
1797 {
1798         unsigned long found_bits = 0;
1799         unsigned long max_bits = 0;
1800         unsigned long bits, i;
1801         unsigned long next_zero;
1802         unsigned long extent_bits;
1803
1804         /*
1805          * Skip searching the bitmap if we don't have a contiguous section that
1806          * is large enough for this allocation.
1807          */
1808         if (for_alloc &&
1809             bitmap_info->max_extent_size &&
1810             bitmap_info->max_extent_size < *bytes) {
1811                 *bytes = bitmap_info->max_extent_size;
1812                 return -1;
1813         }
1814
1815         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1816                           max_t(u64, *offset, bitmap_info->offset));
1817         bits = bytes_to_bits(*bytes, ctl->unit);
1818
1819         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1820                 if (for_alloc && bits == 1) {
1821                         found_bits = 1;
1822                         break;
1823                 }
1824                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1825                                                BITS_PER_BITMAP, i);
1826                 extent_bits = next_zero - i;
1827                 if (extent_bits >= bits) {
1828                         found_bits = extent_bits;
1829                         break;
1830                 } else if (extent_bits > max_bits) {
1831                         max_bits = extent_bits;
1832                 }
1833                 i = next_zero;
1834         }
1835
1836         if (found_bits) {
1837                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1838                 *bytes = (u64)(found_bits) * ctl->unit;
1839                 return 0;
1840         }
1841
1842         *bytes = (u64)(max_bits) * ctl->unit;
1843         bitmap_info->max_extent_size = *bytes;
1844         return -1;
1845 }
1846
1847 static inline u64 get_max_extent_size(struct btrfs_free_space *entry)
1848 {
1849         if (entry->bitmap)
1850                 return entry->max_extent_size;
1851         return entry->bytes;
1852 }
1853
1854 /* Cache the size of the max extent in bytes */
1855 static struct btrfs_free_space *
1856 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1857                 unsigned long align, u64 *max_extent_size)
1858 {
1859         struct btrfs_free_space *entry;
1860         struct rb_node *node;
1861         u64 tmp;
1862         u64 align_off;
1863         int ret;
1864
1865         if (!ctl->free_space_offset.rb_node)
1866                 goto out;
1867
1868         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1869         if (!entry)
1870                 goto out;
1871
1872         for (node = &entry->offset_index; node; node = rb_next(node)) {
1873                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1874                 if (entry->bytes < *bytes) {
1875                         *max_extent_size = max(get_max_extent_size(entry),
1876                                                *max_extent_size);
1877                         continue;
1878                 }
1879
1880                 /* make sure the space returned is big enough
1881                  * to match our requested alignment
1882                  */
1883                 if (*bytes >= align) {
1884                         tmp = entry->offset - ctl->start + align - 1;
1885                         tmp = div64_u64(tmp, align);
1886                         tmp = tmp * align + ctl->start;
1887                         align_off = tmp - entry->offset;
1888                 } else {
1889                         align_off = 0;
1890                         tmp = entry->offset;
1891                 }
1892
1893                 if (entry->bytes < *bytes + align_off) {
1894                         *max_extent_size = max(get_max_extent_size(entry),
1895                                                *max_extent_size);
1896                         continue;
1897                 }
1898
1899                 if (entry->bitmap) {
1900                         u64 size = *bytes;
1901
1902                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1903                         if (!ret) {
1904                                 *offset = tmp;
1905                                 *bytes = size;
1906                                 return entry;
1907                         } else {
1908                                 *max_extent_size =
1909                                         max(get_max_extent_size(entry),
1910                                             *max_extent_size);
1911                         }
1912                         continue;
1913                 }
1914
1915                 *offset = tmp;
1916                 *bytes = entry->bytes - align_off;
1917                 return entry;
1918         }
1919 out:
1920         return NULL;
1921 }
1922
1923 static int count_bitmap_extents(struct btrfs_free_space_ctl *ctl,
1924                                 struct btrfs_free_space *bitmap_info)
1925 {
1926         struct btrfs_block_group *block_group = ctl->private;
1927         u64 bytes = bitmap_info->bytes;
1928         unsigned int rs, re;
1929         int count = 0;
1930
1931         if (!block_group || !bytes)
1932                 return count;
1933
1934         bitmap_for_each_set_region(bitmap_info->bitmap, rs, re, 0,
1935                                    BITS_PER_BITMAP) {
1936                 bytes -= (rs - re) * ctl->unit;
1937                 count++;
1938
1939                 if (!bytes)
1940                         break;
1941         }
1942
1943         return count;
1944 }
1945
1946 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1947                            struct btrfs_free_space *info, u64 offset)
1948 {
1949         info->offset = offset_to_bitmap(ctl, offset);
1950         info->bytes = 0;
1951         info->bitmap_extents = 0;
1952         INIT_LIST_HEAD(&info->list);
1953         link_free_space(ctl, info);
1954         ctl->total_bitmaps++;
1955
1956         ctl->op->recalc_thresholds(ctl);
1957 }
1958
1959 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1960                         struct btrfs_free_space *bitmap_info)
1961 {
1962         /*
1963          * Normally when this is called, the bitmap is completely empty. However,
1964          * if we are blowing up the free space cache for one reason or another
1965          * via __btrfs_remove_free_space_cache(), then it may not be freed and
1966          * we may leave stats on the table.
1967          */
1968         if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
1969                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
1970                         bitmap_info->bitmap_extents;
1971                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
1972
1973         }
1974         unlink_free_space(ctl, bitmap_info);
1975         kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
1976         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1977         ctl->total_bitmaps--;
1978         ctl->op->recalc_thresholds(ctl);
1979 }
1980
1981 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1982                               struct btrfs_free_space *bitmap_info,
1983                               u64 *offset, u64 *bytes)
1984 {
1985         u64 end;
1986         u64 search_start, search_bytes;
1987         int ret;
1988
1989 again:
1990         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1991
1992         /*
1993          * We need to search for bits in this bitmap.  We could only cover some
1994          * of the extent in this bitmap thanks to how we add space, so we need
1995          * to search for as much as it as we can and clear that amount, and then
1996          * go searching for the next bit.
1997          */
1998         search_start = *offset;
1999         search_bytes = ctl->unit;
2000         search_bytes = min(search_bytes, end - search_start + 1);
2001         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
2002                             false);
2003         if (ret < 0 || search_start != *offset)
2004                 return -EINVAL;
2005
2006         /* We may have found more bits than what we need */
2007         search_bytes = min(search_bytes, *bytes);
2008
2009         /* Cannot clear past the end of the bitmap */
2010         search_bytes = min(search_bytes, end - search_start + 1);
2011
2012         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
2013         *offset += search_bytes;
2014         *bytes -= search_bytes;
2015
2016         if (*bytes) {
2017                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2018                 if (!bitmap_info->bytes)
2019                         free_bitmap(ctl, bitmap_info);
2020
2021                 /*
2022                  * no entry after this bitmap, but we still have bytes to
2023                  * remove, so something has gone wrong.
2024                  */
2025                 if (!next)
2026                         return -EINVAL;
2027
2028                 bitmap_info = rb_entry(next, struct btrfs_free_space,
2029                                        offset_index);
2030
2031                 /*
2032                  * if the next entry isn't a bitmap we need to return to let the
2033                  * extent stuff do its work.
2034                  */
2035                 if (!bitmap_info->bitmap)
2036                         return -EAGAIN;
2037
2038                 /*
2039                  * Ok the next item is a bitmap, but it may not actually hold
2040                  * the information for the rest of this free space stuff, so
2041                  * look for it, and if we don't find it return so we can try
2042                  * everything over again.
2043                  */
2044                 search_start = *offset;
2045                 search_bytes = ctl->unit;
2046                 ret = search_bitmap(ctl, bitmap_info, &search_start,
2047                                     &search_bytes, false);
2048                 if (ret < 0 || search_start != *offset)
2049                         return -EAGAIN;
2050
2051                 goto again;
2052         } else if (!bitmap_info->bytes)
2053                 free_bitmap(ctl, bitmap_info);
2054
2055         return 0;
2056 }
2057
2058 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2059                                struct btrfs_free_space *info, u64 offset,
2060                                u64 bytes, enum btrfs_trim_state trim_state)
2061 {
2062         u64 bytes_to_set = 0;
2063         u64 end;
2064
2065         /*
2066          * This is a tradeoff to make bitmap trim state minimal.  We mark the
2067          * whole bitmap untrimmed if at any point we add untrimmed regions.
2068          */
2069         if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2070                 if (btrfs_free_space_trimmed(info)) {
2071                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
2072                                 info->bitmap_extents;
2073                         ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2074                 }
2075                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2076         }
2077
2078         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2079
2080         bytes_to_set = min(end - offset, bytes);
2081
2082         bitmap_set_bits(ctl, info, offset, bytes_to_set);
2083
2084         /*
2085          * We set some bytes, we have no idea what the max extent size is
2086          * anymore.
2087          */
2088         info->max_extent_size = 0;
2089
2090         return bytes_to_set;
2091
2092 }
2093
2094 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2095                       struct btrfs_free_space *info)
2096 {
2097         struct btrfs_block_group *block_group = ctl->private;
2098         struct btrfs_fs_info *fs_info = block_group->fs_info;
2099         bool forced = false;
2100
2101 #ifdef CONFIG_BTRFS_DEBUG
2102         if (btrfs_should_fragment_free_space(block_group))
2103                 forced = true;
2104 #endif
2105
2106         /* This is a way to reclaim large regions from the bitmaps. */
2107         if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2108                 return false;
2109
2110         /*
2111          * If we are below the extents threshold then we can add this as an
2112          * extent, and don't have to deal with the bitmap
2113          */
2114         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2115                 /*
2116                  * If this block group has some small extents we don't want to
2117                  * use up all of our free slots in the cache with them, we want
2118                  * to reserve them to larger extents, however if we have plenty
2119                  * of cache left then go ahead an dadd them, no sense in adding
2120                  * the overhead of a bitmap if we don't have to.
2121                  */
2122                 if (info->bytes <= fs_info->sectorsize * 8) {
2123                         if (ctl->free_extents * 3 <= ctl->extents_thresh)
2124                                 return false;
2125                 } else {
2126                         return false;
2127                 }
2128         }
2129
2130         /*
2131          * The original block groups from mkfs can be really small, like 8
2132          * megabytes, so don't bother with a bitmap for those entries.  However
2133          * some block groups can be smaller than what a bitmap would cover but
2134          * are still large enough that they could overflow the 32k memory limit,
2135          * so allow those block groups to still be allowed to have a bitmap
2136          * entry.
2137          */
2138         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2139                 return false;
2140
2141         return true;
2142 }
2143
2144 static const struct btrfs_free_space_op free_space_op = {
2145         .recalc_thresholds      = recalculate_thresholds,
2146         .use_bitmap             = use_bitmap,
2147 };
2148
2149 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2150                               struct btrfs_free_space *info)
2151 {
2152         struct btrfs_free_space *bitmap_info;
2153         struct btrfs_block_group *block_group = NULL;
2154         int added = 0;
2155         u64 bytes, offset, bytes_added;
2156         enum btrfs_trim_state trim_state;
2157         int ret;
2158
2159         bytes = info->bytes;
2160         offset = info->offset;
2161         trim_state = info->trim_state;
2162
2163         if (!ctl->op->use_bitmap(ctl, info))
2164                 return 0;
2165
2166         if (ctl->op == &free_space_op)
2167                 block_group = ctl->private;
2168 again:
2169         /*
2170          * Since we link bitmaps right into the cluster we need to see if we
2171          * have a cluster here, and if so and it has our bitmap we need to add
2172          * the free space to that bitmap.
2173          */
2174         if (block_group && !list_empty(&block_group->cluster_list)) {
2175                 struct btrfs_free_cluster *cluster;
2176                 struct rb_node *node;
2177                 struct btrfs_free_space *entry;
2178
2179                 cluster = list_entry(block_group->cluster_list.next,
2180                                      struct btrfs_free_cluster,
2181                                      block_group_list);
2182                 spin_lock(&cluster->lock);
2183                 node = rb_first(&cluster->root);
2184                 if (!node) {
2185                         spin_unlock(&cluster->lock);
2186                         goto no_cluster_bitmap;
2187                 }
2188
2189                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2190                 if (!entry->bitmap) {
2191                         spin_unlock(&cluster->lock);
2192                         goto no_cluster_bitmap;
2193                 }
2194
2195                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2196                         bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2197                                                           bytes, trim_state);
2198                         bytes -= bytes_added;
2199                         offset += bytes_added;
2200                 }
2201                 spin_unlock(&cluster->lock);
2202                 if (!bytes) {
2203                         ret = 1;
2204                         goto out;
2205                 }
2206         }
2207
2208 no_cluster_bitmap:
2209         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2210                                          1, 0);
2211         if (!bitmap_info) {
2212                 ASSERT(added == 0);
2213                 goto new_bitmap;
2214         }
2215
2216         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2217                                           trim_state);
2218         bytes -= bytes_added;
2219         offset += bytes_added;
2220         added = 0;
2221
2222         if (!bytes) {
2223                 ret = 1;
2224                 goto out;
2225         } else
2226                 goto again;
2227
2228 new_bitmap:
2229         if (info && info->bitmap) {
2230                 add_new_bitmap(ctl, info, offset);
2231                 added = 1;
2232                 info = NULL;
2233                 goto again;
2234         } else {
2235                 spin_unlock(&ctl->tree_lock);
2236
2237                 /* no pre-allocated info, allocate a new one */
2238                 if (!info) {
2239                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2240                                                  GFP_NOFS);
2241                         if (!info) {
2242                                 spin_lock(&ctl->tree_lock);
2243                                 ret = -ENOMEM;
2244                                 goto out;
2245                         }
2246                 }
2247
2248                 /* allocate the bitmap */
2249                 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2250                                                  GFP_NOFS);
2251                 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2252                 spin_lock(&ctl->tree_lock);
2253                 if (!info->bitmap) {
2254                         ret = -ENOMEM;
2255                         goto out;
2256                 }
2257                 goto again;
2258         }
2259
2260 out:
2261         if (info) {
2262                 if (info->bitmap)
2263                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2264                                         info->bitmap);
2265                 kmem_cache_free(btrfs_free_space_cachep, info);
2266         }
2267
2268         return ret;
2269 }
2270
2271 /*
2272  * Free space merging rules:
2273  *  1) Merge trimmed areas together
2274  *  2) Let untrimmed areas coalesce with trimmed areas
2275  *  3) Always pull neighboring regions from bitmaps
2276  *
2277  * The above rules are for when we merge free space based on btrfs_trim_state.
2278  * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2279  * same reason: to promote larger extent regions which makes life easier for
2280  * find_free_extent().  Rule 2 enables coalescing based on the common path
2281  * being returning free space from btrfs_finish_extent_commit().  So when free
2282  * space is trimmed, it will prevent aggregating trimmed new region and
2283  * untrimmed regions in the rb_tree.  Rule 3 is purely to obtain larger extents
2284  * and provide find_free_extent() with the largest extents possible hoping for
2285  * the reuse path.
2286  */
2287 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2288                           struct btrfs_free_space *info, bool update_stat)
2289 {
2290         struct btrfs_free_space *left_info;
2291         struct btrfs_free_space *right_info;
2292         bool merged = false;
2293         u64 offset = info->offset;
2294         u64 bytes = info->bytes;
2295         const bool is_trimmed = btrfs_free_space_trimmed(info);
2296
2297         /*
2298          * first we want to see if there is free space adjacent to the range we
2299          * are adding, if there is remove that struct and add a new one to
2300          * cover the entire range
2301          */
2302         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2303         if (right_info && rb_prev(&right_info->offset_index))
2304                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2305                                      struct btrfs_free_space, offset_index);
2306         else
2307                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2308
2309         /* See try_merge_free_space() comment. */
2310         if (right_info && !right_info->bitmap &&
2311             (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2312                 if (update_stat)
2313                         unlink_free_space(ctl, right_info);
2314                 else
2315                         __unlink_free_space(ctl, right_info);
2316                 info->bytes += right_info->bytes;
2317                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2318                 merged = true;
2319         }
2320
2321         /* See try_merge_free_space() comment. */
2322         if (left_info && !left_info->bitmap &&
2323             left_info->offset + left_info->bytes == offset &&
2324             (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2325                 if (update_stat)
2326                         unlink_free_space(ctl, left_info);
2327                 else
2328                         __unlink_free_space(ctl, left_info);
2329                 info->offset = left_info->offset;
2330                 info->bytes += left_info->bytes;
2331                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2332                 merged = true;
2333         }
2334
2335         return merged;
2336 }
2337
2338 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2339                                      struct btrfs_free_space *info,
2340                                      bool update_stat)
2341 {
2342         struct btrfs_free_space *bitmap;
2343         unsigned long i;
2344         unsigned long j;
2345         const u64 end = info->offset + info->bytes;
2346         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2347         u64 bytes;
2348
2349         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2350         if (!bitmap)
2351                 return false;
2352
2353         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2354         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2355         if (j == i)
2356                 return false;
2357         bytes = (j - i) * ctl->unit;
2358         info->bytes += bytes;
2359
2360         /* See try_merge_free_space() comment. */
2361         if (!btrfs_free_space_trimmed(bitmap))
2362                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2363
2364         if (update_stat)
2365                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2366         else
2367                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2368
2369         if (!bitmap->bytes)
2370                 free_bitmap(ctl, bitmap);
2371
2372         return true;
2373 }
2374
2375 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2376                                        struct btrfs_free_space *info,
2377                                        bool update_stat)
2378 {
2379         struct btrfs_free_space *bitmap;
2380         u64 bitmap_offset;
2381         unsigned long i;
2382         unsigned long j;
2383         unsigned long prev_j;
2384         u64 bytes;
2385
2386         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2387         /* If we're on a boundary, try the previous logical bitmap. */
2388         if (bitmap_offset == info->offset) {
2389                 if (info->offset == 0)
2390                         return false;
2391                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2392         }
2393
2394         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2395         if (!bitmap)
2396                 return false;
2397
2398         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2399         j = 0;
2400         prev_j = (unsigned long)-1;
2401         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2402                 if (j > i)
2403                         break;
2404                 prev_j = j;
2405         }
2406         if (prev_j == i)
2407                 return false;
2408
2409         if (prev_j == (unsigned long)-1)
2410                 bytes = (i + 1) * ctl->unit;
2411         else
2412                 bytes = (i - prev_j) * ctl->unit;
2413
2414         info->offset -= bytes;
2415         info->bytes += bytes;
2416
2417         /* See try_merge_free_space() comment. */
2418         if (!btrfs_free_space_trimmed(bitmap))
2419                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2420
2421         if (update_stat)
2422                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2423         else
2424                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2425
2426         if (!bitmap->bytes)
2427                 free_bitmap(ctl, bitmap);
2428
2429         return true;
2430 }
2431
2432 /*
2433  * We prefer always to allocate from extent entries, both for clustered and
2434  * non-clustered allocation requests. So when attempting to add a new extent
2435  * entry, try to see if there's adjacent free space in bitmap entries, and if
2436  * there is, migrate that space from the bitmaps to the extent.
2437  * Like this we get better chances of satisfying space allocation requests
2438  * because we attempt to satisfy them based on a single cache entry, and never
2439  * on 2 or more entries - even if the entries represent a contiguous free space
2440  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2441  * ends).
2442  */
2443 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2444                               struct btrfs_free_space *info,
2445                               bool update_stat)
2446 {
2447         /*
2448          * Only work with disconnected entries, as we can change their offset,
2449          * and must be extent entries.
2450          */
2451         ASSERT(!info->bitmap);
2452         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2453
2454         if (ctl->total_bitmaps > 0) {
2455                 bool stole_end;
2456                 bool stole_front = false;
2457
2458                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2459                 if (ctl->total_bitmaps > 0)
2460                         stole_front = steal_from_bitmap_to_front(ctl, info,
2461                                                                  update_stat);
2462
2463                 if (stole_end || stole_front)
2464                         try_merge_free_space(ctl, info, update_stat);
2465         }
2466 }
2467
2468 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2469                            struct btrfs_free_space_ctl *ctl,
2470                            u64 offset, u64 bytes,
2471                            enum btrfs_trim_state trim_state)
2472 {
2473         struct btrfs_block_group *block_group = ctl->private;
2474         struct btrfs_free_space *info;
2475         int ret = 0;
2476         u64 filter_bytes = bytes;
2477
2478         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2479         if (!info)
2480                 return -ENOMEM;
2481
2482         info->offset = offset;
2483         info->bytes = bytes;
2484         info->trim_state = trim_state;
2485         RB_CLEAR_NODE(&info->offset_index);
2486
2487         spin_lock(&ctl->tree_lock);
2488
2489         if (try_merge_free_space(ctl, info, true))
2490                 goto link;
2491
2492         /*
2493          * There was no extent directly to the left or right of this new
2494          * extent then we know we're going to have to allocate a new extent, so
2495          * before we do that see if we need to drop this into a bitmap
2496          */
2497         ret = insert_into_bitmap(ctl, info);
2498         if (ret < 0) {
2499                 goto out;
2500         } else if (ret) {
2501                 ret = 0;
2502                 goto out;
2503         }
2504 link:
2505         /*
2506          * Only steal free space from adjacent bitmaps if we're sure we're not
2507          * going to add the new free space to existing bitmap entries - because
2508          * that would mean unnecessary work that would be reverted. Therefore
2509          * attempt to steal space from bitmaps if we're adding an extent entry.
2510          */
2511         steal_from_bitmap(ctl, info, true);
2512
2513         filter_bytes = max(filter_bytes, info->bytes);
2514
2515         ret = link_free_space(ctl, info);
2516         if (ret)
2517                 kmem_cache_free(btrfs_free_space_cachep, info);
2518 out:
2519         btrfs_discard_update_discardable(block_group, ctl);
2520         spin_unlock(&ctl->tree_lock);
2521
2522         if (ret) {
2523                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2524                 ASSERT(ret != -EEXIST);
2525         }
2526
2527         if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2528                 btrfs_discard_check_filter(block_group, filter_bytes);
2529                 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2530         }
2531
2532         return ret;
2533 }
2534
2535 int btrfs_add_free_space(struct btrfs_block_group *block_group,
2536                          u64 bytenr, u64 size)
2537 {
2538         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2539
2540         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2541                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2542
2543         return __btrfs_add_free_space(block_group->fs_info,
2544                                       block_group->free_space_ctl,
2545                                       bytenr, size, trim_state);
2546 }
2547
2548 /*
2549  * This is a subtle distinction because when adding free space back in general,
2550  * we want it to be added as untrimmed for async. But in the case where we add
2551  * it on loading of a block group, we want to consider it trimmed.
2552  */
2553 int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2554                                        u64 bytenr, u64 size)
2555 {
2556         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2557
2558         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2559             btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2560                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2561
2562         return __btrfs_add_free_space(block_group->fs_info,
2563                                       block_group->free_space_ctl,
2564                                       bytenr, size, trim_state);
2565 }
2566
2567 int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2568                             u64 offset, u64 bytes)
2569 {
2570         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2571         struct btrfs_free_space *info;
2572         int ret;
2573         bool re_search = false;
2574
2575         spin_lock(&ctl->tree_lock);
2576
2577 again:
2578         ret = 0;
2579         if (!bytes)
2580                 goto out_lock;
2581
2582         info = tree_search_offset(ctl, offset, 0, 0);
2583         if (!info) {
2584                 /*
2585                  * oops didn't find an extent that matched the space we wanted
2586                  * to remove, look for a bitmap instead
2587                  */
2588                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2589                                           1, 0);
2590                 if (!info) {
2591                         /*
2592                          * If we found a partial bit of our free space in a
2593                          * bitmap but then couldn't find the other part this may
2594                          * be a problem, so WARN about it.
2595                          */
2596                         WARN_ON(re_search);
2597                         goto out_lock;
2598                 }
2599         }
2600
2601         re_search = false;
2602         if (!info->bitmap) {
2603                 unlink_free_space(ctl, info);
2604                 if (offset == info->offset) {
2605                         u64 to_free = min(bytes, info->bytes);
2606
2607                         info->bytes -= to_free;
2608                         info->offset += to_free;
2609                         if (info->bytes) {
2610                                 ret = link_free_space(ctl, info);
2611                                 WARN_ON(ret);
2612                         } else {
2613                                 kmem_cache_free(btrfs_free_space_cachep, info);
2614                         }
2615
2616                         offset += to_free;
2617                         bytes -= to_free;
2618                         goto again;
2619                 } else {
2620                         u64 old_end = info->bytes + info->offset;
2621
2622                         info->bytes = offset - info->offset;
2623                         ret = link_free_space(ctl, info);
2624                         WARN_ON(ret);
2625                         if (ret)
2626                                 goto out_lock;
2627
2628                         /* Not enough bytes in this entry to satisfy us */
2629                         if (old_end < offset + bytes) {
2630                                 bytes -= old_end - offset;
2631                                 offset = old_end;
2632                                 goto again;
2633                         } else if (old_end == offset + bytes) {
2634                                 /* all done */
2635                                 goto out_lock;
2636                         }
2637                         spin_unlock(&ctl->tree_lock);
2638
2639                         ret = __btrfs_add_free_space(block_group->fs_info, ctl,
2640                                                      offset + bytes,
2641                                                      old_end - (offset + bytes),
2642                                                      info->trim_state);
2643                         WARN_ON(ret);
2644                         goto out;
2645                 }
2646         }
2647
2648         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2649         if (ret == -EAGAIN) {
2650                 re_search = true;
2651                 goto again;
2652         }
2653 out_lock:
2654         btrfs_discard_update_discardable(block_group, ctl);
2655         spin_unlock(&ctl->tree_lock);
2656 out:
2657         return ret;
2658 }
2659
2660 void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2661                            u64 bytes)
2662 {
2663         struct btrfs_fs_info *fs_info = block_group->fs_info;
2664         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2665         struct btrfs_free_space *info;
2666         struct rb_node *n;
2667         int count = 0;
2668
2669         spin_lock(&ctl->tree_lock);
2670         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2671                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2672                 if (info->bytes >= bytes && !block_group->ro)
2673                         count++;
2674                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2675                            info->offset, info->bytes,
2676                        (info->bitmap) ? "yes" : "no");
2677         }
2678         spin_unlock(&ctl->tree_lock);
2679         btrfs_info(fs_info, "block group has cluster?: %s",
2680                list_empty(&block_group->cluster_list) ? "no" : "yes");
2681         btrfs_info(fs_info,
2682                    "%d blocks of free space at or bigger than bytes is", count);
2683 }
2684
2685 void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group)
2686 {
2687         struct btrfs_fs_info *fs_info = block_group->fs_info;
2688         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2689
2690         spin_lock_init(&ctl->tree_lock);
2691         ctl->unit = fs_info->sectorsize;
2692         ctl->start = block_group->start;
2693         ctl->private = block_group;
2694         ctl->op = &free_space_op;
2695         INIT_LIST_HEAD(&ctl->trimming_ranges);
2696         mutex_init(&ctl->cache_writeout_mutex);
2697
2698         /*
2699          * we only want to have 32k of ram per block group for keeping
2700          * track of free space, and if we pass 1/2 of that we want to
2701          * start converting things over to using bitmaps
2702          */
2703         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2704 }
2705
2706 /*
2707  * for a given cluster, put all of its extents back into the free
2708  * space cache.  If the block group passed doesn't match the block group
2709  * pointed to by the cluster, someone else raced in and freed the
2710  * cluster already.  In that case, we just return without changing anything
2711  */
2712 static int
2713 __btrfs_return_cluster_to_free_space(
2714                              struct btrfs_block_group *block_group,
2715                              struct btrfs_free_cluster *cluster)
2716 {
2717         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2718         struct btrfs_free_space *entry;
2719         struct rb_node *node;
2720
2721         spin_lock(&cluster->lock);
2722         if (cluster->block_group != block_group)
2723                 goto out;
2724
2725         cluster->block_group = NULL;
2726         cluster->window_start = 0;
2727         list_del_init(&cluster->block_group_list);
2728
2729         node = rb_first(&cluster->root);
2730         while (node) {
2731                 bool bitmap;
2732
2733                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2734                 node = rb_next(&entry->offset_index);
2735                 rb_erase(&entry->offset_index, &cluster->root);
2736                 RB_CLEAR_NODE(&entry->offset_index);
2737
2738                 bitmap = (entry->bitmap != NULL);
2739                 if (!bitmap) {
2740                         /* Merging treats extents as if they were new */
2741                         if (!btrfs_free_space_trimmed(entry)) {
2742                                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
2743                                 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
2744                                         entry->bytes;
2745                         }
2746
2747                         try_merge_free_space(ctl, entry, false);
2748                         steal_from_bitmap(ctl, entry, false);
2749
2750                         /* As we insert directly, update these statistics */
2751                         if (!btrfs_free_space_trimmed(entry)) {
2752                                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
2753                                 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
2754                                         entry->bytes;
2755                         }
2756                 }
2757                 tree_insert_offset(&ctl->free_space_offset,
2758                                    entry->offset, &entry->offset_index, bitmap);
2759         }
2760         cluster->root = RB_ROOT;
2761
2762 out:
2763         spin_unlock(&cluster->lock);
2764         btrfs_put_block_group(block_group);
2765         return 0;
2766 }
2767
2768 static void __btrfs_remove_free_space_cache_locked(
2769                                 struct btrfs_free_space_ctl *ctl)
2770 {
2771         struct btrfs_free_space *info;
2772         struct rb_node *node;
2773
2774         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2775                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2776                 if (!info->bitmap) {
2777                         unlink_free_space(ctl, info);
2778                         kmem_cache_free(btrfs_free_space_cachep, info);
2779                 } else {
2780                         free_bitmap(ctl, info);
2781                 }
2782
2783                 cond_resched_lock(&ctl->tree_lock);
2784         }
2785 }
2786
2787 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2788 {
2789         spin_lock(&ctl->tree_lock);
2790         __btrfs_remove_free_space_cache_locked(ctl);
2791         if (ctl->private)
2792                 btrfs_discard_update_discardable(ctl->private, ctl);
2793         spin_unlock(&ctl->tree_lock);
2794 }
2795
2796 void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
2797 {
2798         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2799         struct btrfs_free_cluster *cluster;
2800         struct list_head *head;
2801
2802         spin_lock(&ctl->tree_lock);
2803         while ((head = block_group->cluster_list.next) !=
2804                &block_group->cluster_list) {
2805                 cluster = list_entry(head, struct btrfs_free_cluster,
2806                                      block_group_list);
2807
2808                 WARN_ON(cluster->block_group != block_group);
2809                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2810
2811                 cond_resched_lock(&ctl->tree_lock);
2812         }
2813         __btrfs_remove_free_space_cache_locked(ctl);
2814         btrfs_discard_update_discardable(block_group, ctl);
2815         spin_unlock(&ctl->tree_lock);
2816
2817 }
2818
2819 /**
2820  * btrfs_is_free_space_trimmed - see if everything is trimmed
2821  * @block_group: block_group of interest
2822  *
2823  * Walk @block_group's free space rb_tree to determine if everything is trimmed.
2824  */
2825 bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
2826 {
2827         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2828         struct btrfs_free_space *info;
2829         struct rb_node *node;
2830         bool ret = true;
2831
2832         spin_lock(&ctl->tree_lock);
2833         node = rb_first(&ctl->free_space_offset);
2834
2835         while (node) {
2836                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2837
2838                 if (!btrfs_free_space_trimmed(info)) {
2839                         ret = false;
2840                         break;
2841                 }
2842
2843                 node = rb_next(node);
2844         }
2845
2846         spin_unlock(&ctl->tree_lock);
2847         return ret;
2848 }
2849
2850 u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
2851                                u64 offset, u64 bytes, u64 empty_size,
2852                                u64 *max_extent_size)
2853 {
2854         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2855         struct btrfs_discard_ctl *discard_ctl =
2856                                         &block_group->fs_info->discard_ctl;
2857         struct btrfs_free_space *entry = NULL;
2858         u64 bytes_search = bytes + empty_size;
2859         u64 ret = 0;
2860         u64 align_gap = 0;
2861         u64 align_gap_len = 0;
2862         enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2863
2864         spin_lock(&ctl->tree_lock);
2865         entry = find_free_space(ctl, &offset, &bytes_search,
2866                                 block_group->full_stripe_len, max_extent_size);
2867         if (!entry)
2868                 goto out;
2869
2870         ret = offset;
2871         if (entry->bitmap) {
2872                 bitmap_clear_bits(ctl, entry, offset, bytes);
2873
2874                 if (!btrfs_free_space_trimmed(entry))
2875                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
2876
2877                 if (!entry->bytes)
2878                         free_bitmap(ctl, entry);
2879         } else {
2880                 unlink_free_space(ctl, entry);
2881                 align_gap_len = offset - entry->offset;
2882                 align_gap = entry->offset;
2883                 align_gap_trim_state = entry->trim_state;
2884
2885                 if (!btrfs_free_space_trimmed(entry))
2886                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
2887
2888                 entry->offset = offset + bytes;
2889                 WARN_ON(entry->bytes < bytes + align_gap_len);
2890
2891                 entry->bytes -= bytes + align_gap_len;
2892                 if (!entry->bytes)
2893                         kmem_cache_free(btrfs_free_space_cachep, entry);
2894                 else
2895                         link_free_space(ctl, entry);
2896         }
2897 out:
2898         btrfs_discard_update_discardable(block_group, ctl);
2899         spin_unlock(&ctl->tree_lock);
2900
2901         if (align_gap_len)
2902                 __btrfs_add_free_space(block_group->fs_info, ctl,
2903                                        align_gap, align_gap_len,
2904                                        align_gap_trim_state);
2905         return ret;
2906 }
2907
2908 /*
2909  * given a cluster, put all of its extents back into the free space
2910  * cache.  If a block group is passed, this function will only free
2911  * a cluster that belongs to the passed block group.
2912  *
2913  * Otherwise, it'll get a reference on the block group pointed to by the
2914  * cluster and remove the cluster from it.
2915  */
2916 int btrfs_return_cluster_to_free_space(
2917                                struct btrfs_block_group *block_group,
2918                                struct btrfs_free_cluster *cluster)
2919 {
2920         struct btrfs_free_space_ctl *ctl;
2921         int ret;
2922
2923         /* first, get a safe pointer to the block group */
2924         spin_lock(&cluster->lock);
2925         if (!block_group) {
2926                 block_group = cluster->block_group;
2927                 if (!block_group) {
2928                         spin_unlock(&cluster->lock);
2929                         return 0;
2930                 }
2931         } else if (cluster->block_group != block_group) {
2932                 /* someone else has already freed it don't redo their work */
2933                 spin_unlock(&cluster->lock);
2934                 return 0;
2935         }
2936         atomic_inc(&block_group->count);
2937         spin_unlock(&cluster->lock);
2938
2939         ctl = block_group->free_space_ctl;
2940
2941         /* now return any extents the cluster had on it */
2942         spin_lock(&ctl->tree_lock);
2943         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2944         spin_unlock(&ctl->tree_lock);
2945
2946         btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
2947
2948         /* finally drop our ref */
2949         btrfs_put_block_group(block_group);
2950         return ret;
2951 }
2952
2953 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
2954                                    struct btrfs_free_cluster *cluster,
2955                                    struct btrfs_free_space *entry,
2956                                    u64 bytes, u64 min_start,
2957                                    u64 *max_extent_size)
2958 {
2959         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2960         int err;
2961         u64 search_start = cluster->window_start;
2962         u64 search_bytes = bytes;
2963         u64 ret = 0;
2964
2965         search_start = min_start;
2966         search_bytes = bytes;
2967
2968         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2969         if (err) {
2970                 *max_extent_size = max(get_max_extent_size(entry),
2971                                        *max_extent_size);
2972                 return 0;
2973         }
2974
2975         ret = search_start;
2976         __bitmap_clear_bits(ctl, entry, ret, bytes);
2977
2978         return ret;
2979 }
2980
2981 /*
2982  * given a cluster, try to allocate 'bytes' from it, returns 0
2983  * if it couldn't find anything suitably large, or a logical disk offset
2984  * if things worked out
2985  */
2986 u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
2987                              struct btrfs_free_cluster *cluster, u64 bytes,
2988                              u64 min_start, u64 *max_extent_size)
2989 {
2990         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2991         struct btrfs_discard_ctl *discard_ctl =
2992                                         &block_group->fs_info->discard_ctl;
2993         struct btrfs_free_space *entry = NULL;
2994         struct rb_node *node;
2995         u64 ret = 0;
2996
2997         spin_lock(&cluster->lock);
2998         if (bytes > cluster->max_size)
2999                 goto out;
3000
3001         if (cluster->block_group != block_group)
3002                 goto out;
3003
3004         node = rb_first(&cluster->root);
3005         if (!node)
3006                 goto out;
3007
3008         entry = rb_entry(node, struct btrfs_free_space, offset_index);
3009         while (1) {
3010                 if (entry->bytes < bytes)
3011                         *max_extent_size = max(get_max_extent_size(entry),
3012                                                *max_extent_size);
3013
3014                 if (entry->bytes < bytes ||
3015                     (!entry->bitmap && entry->offset < min_start)) {
3016                         node = rb_next(&entry->offset_index);
3017                         if (!node)
3018                                 break;
3019                         entry = rb_entry(node, struct btrfs_free_space,
3020                                          offset_index);
3021                         continue;
3022                 }
3023
3024                 if (entry->bitmap) {
3025                         ret = btrfs_alloc_from_bitmap(block_group,
3026                                                       cluster, entry, bytes,
3027                                                       cluster->window_start,
3028                                                       max_extent_size);
3029                         if (ret == 0) {
3030                                 node = rb_next(&entry->offset_index);
3031                                 if (!node)
3032                                         break;
3033                                 entry = rb_entry(node, struct btrfs_free_space,
3034                                                  offset_index);
3035                                 continue;
3036                         }
3037                         cluster->window_start += bytes;
3038                 } else {
3039                         ret = entry->offset;
3040
3041                         entry->offset += bytes;
3042                         entry->bytes -= bytes;
3043                 }
3044
3045                 if (entry->bytes == 0)
3046                         rb_erase(&entry->offset_index, &cluster->root);
3047                 break;
3048         }
3049 out:
3050         spin_unlock(&cluster->lock);
3051
3052         if (!ret)
3053                 return 0;
3054
3055         spin_lock(&ctl->tree_lock);
3056
3057         if (!btrfs_free_space_trimmed(entry))
3058                 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3059
3060         ctl->free_space -= bytes;
3061         if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3062                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3063         if (entry->bytes == 0) {
3064                 ctl->free_extents--;
3065                 if (entry->bitmap) {
3066                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
3067                                         entry->bitmap);
3068                         ctl->total_bitmaps--;
3069                         ctl->op->recalc_thresholds(ctl);
3070                 } else if (!btrfs_free_space_trimmed(entry)) {
3071                         ctl->discardable_extents[BTRFS_STAT_CURR]--;
3072                 }
3073                 kmem_cache_free(btrfs_free_space_cachep, entry);
3074         }
3075
3076         spin_unlock(&ctl->tree_lock);
3077
3078         return ret;
3079 }
3080
3081 static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3082                                 struct btrfs_free_space *entry,
3083                                 struct btrfs_free_cluster *cluster,
3084                                 u64 offset, u64 bytes,
3085                                 u64 cont1_bytes, u64 min_bytes)
3086 {
3087         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3088         unsigned long next_zero;
3089         unsigned long i;
3090         unsigned long want_bits;
3091         unsigned long min_bits;
3092         unsigned long found_bits;
3093         unsigned long max_bits = 0;
3094         unsigned long start = 0;
3095         unsigned long total_found = 0;
3096         int ret;
3097
3098         i = offset_to_bit(entry->offset, ctl->unit,
3099                           max_t(u64, offset, entry->offset));
3100         want_bits = bytes_to_bits(bytes, ctl->unit);
3101         min_bits = bytes_to_bits(min_bytes, ctl->unit);
3102
3103         /*
3104          * Don't bother looking for a cluster in this bitmap if it's heavily
3105          * fragmented.
3106          */
3107         if (entry->max_extent_size &&
3108             entry->max_extent_size < cont1_bytes)
3109                 return -ENOSPC;
3110 again:
3111         found_bits = 0;
3112         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3113                 next_zero = find_next_zero_bit(entry->bitmap,
3114                                                BITS_PER_BITMAP, i);
3115                 if (next_zero - i >= min_bits) {
3116                         found_bits = next_zero - i;
3117                         if (found_bits > max_bits)
3118                                 max_bits = found_bits;
3119                         break;
3120                 }
3121                 if (next_zero - i > max_bits)
3122                         max_bits = next_zero - i;
3123                 i = next_zero;
3124         }
3125
3126         if (!found_bits) {
3127                 entry->max_extent_size = (u64)max_bits * ctl->unit;
3128                 return -ENOSPC;
3129         }
3130
3131         if (!total_found) {
3132                 start = i;
3133                 cluster->max_size = 0;
3134         }
3135
3136         total_found += found_bits;
3137
3138         if (cluster->max_size < found_bits * ctl->unit)
3139                 cluster->max_size = found_bits * ctl->unit;
3140
3141         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3142                 i = next_zero + 1;
3143                 goto again;
3144         }
3145
3146         cluster->window_start = start * ctl->unit + entry->offset;
3147         rb_erase(&entry->offset_index, &ctl->free_space_offset);
3148         ret = tree_insert_offset(&cluster->root, entry->offset,
3149                                  &entry->offset_index, 1);
3150         ASSERT(!ret); /* -EEXIST; Logic error */
3151
3152         trace_btrfs_setup_cluster(block_group, cluster,
3153                                   total_found * ctl->unit, 1);
3154         return 0;
3155 }
3156
3157 /*
3158  * This searches the block group for just extents to fill the cluster with.
3159  * Try to find a cluster with at least bytes total bytes, at least one
3160  * extent of cont1_bytes, and other clusters of at least min_bytes.
3161  */
3162 static noinline int
3163 setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3164                         struct btrfs_free_cluster *cluster,
3165                         struct list_head *bitmaps, u64 offset, u64 bytes,
3166                         u64 cont1_bytes, u64 min_bytes)
3167 {
3168         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3169         struct btrfs_free_space *first = NULL;
3170         struct btrfs_free_space *entry = NULL;
3171         struct btrfs_free_space *last;
3172         struct rb_node *node;
3173         u64 window_free;
3174         u64 max_extent;
3175         u64 total_size = 0;
3176
3177         entry = tree_search_offset(ctl, offset, 0, 1);
3178         if (!entry)
3179                 return -ENOSPC;
3180
3181         /*
3182          * We don't want bitmaps, so just move along until we find a normal
3183          * extent entry.
3184          */
3185         while (entry->bitmap || entry->bytes < min_bytes) {
3186                 if (entry->bitmap && list_empty(&entry->list))
3187                         list_add_tail(&entry->list, bitmaps);
3188                 node = rb_next(&entry->offset_index);
3189                 if (!node)
3190                         return -ENOSPC;
3191                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3192         }
3193
3194         window_free = entry->bytes;
3195         max_extent = entry->bytes;
3196         first = entry;
3197         last = entry;
3198
3199         for (node = rb_next(&entry->offset_index); node;
3200              node = rb_next(&entry->offset_index)) {
3201                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3202
3203                 if (entry->bitmap) {
3204                         if (list_empty(&entry->list))
3205                                 list_add_tail(&entry->list, bitmaps);
3206                         continue;
3207                 }
3208
3209                 if (entry->bytes < min_bytes)
3210                         continue;
3211
3212                 last = entry;
3213                 window_free += entry->bytes;
3214                 if (entry->bytes > max_extent)
3215                         max_extent = entry->bytes;
3216         }
3217
3218         if (window_free < bytes || max_extent < cont1_bytes)
3219                 return -ENOSPC;
3220
3221         cluster->window_start = first->offset;
3222
3223         node = &first->offset_index;
3224
3225         /*
3226          * now we've found our entries, pull them out of the free space
3227          * cache and put them into the cluster rbtree
3228          */
3229         do {
3230                 int ret;
3231
3232                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3233                 node = rb_next(&entry->offset_index);
3234                 if (entry->bitmap || entry->bytes < min_bytes)
3235                         continue;
3236
3237                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3238                 ret = tree_insert_offset(&cluster->root, entry->offset,
3239                                          &entry->offset_index, 0);
3240                 total_size += entry->bytes;
3241                 ASSERT(!ret); /* -EEXIST; Logic error */
3242         } while (node && entry != last);
3243
3244         cluster->max_size = max_extent;
3245         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3246         return 0;
3247 }
3248
3249 /*
3250  * This specifically looks for bitmaps that may work in the cluster, we assume
3251  * that we have already failed to find extents that will work.
3252  */
3253 static noinline int
3254 setup_cluster_bitmap(struct btrfs_block_group *block_group,
3255                      struct btrfs_free_cluster *cluster,
3256                      struct list_head *bitmaps, u64 offset, u64 bytes,
3257                      u64 cont1_bytes, u64 min_bytes)
3258 {
3259         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3260         struct btrfs_free_space *entry = NULL;
3261         int ret = -ENOSPC;
3262         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3263
3264         if (ctl->total_bitmaps == 0)
3265                 return -ENOSPC;
3266
3267         /*
3268          * The bitmap that covers offset won't be in the list unless offset
3269          * is just its start offset.
3270          */
3271         if (!list_empty(bitmaps))
3272                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3273
3274         if (!entry || entry->offset != bitmap_offset) {
3275                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3276                 if (entry && list_empty(&entry->list))
3277                         list_add(&entry->list, bitmaps);
3278         }
3279
3280         list_for_each_entry(entry, bitmaps, list) {
3281                 if (entry->bytes < bytes)
3282                         continue;
3283                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3284                                            bytes, cont1_bytes, min_bytes);
3285                 if (!ret)
3286                         return 0;
3287         }
3288
3289         /*
3290          * The bitmaps list has all the bitmaps that record free space
3291          * starting after offset, so no more search is required.
3292          */
3293         return -ENOSPC;
3294 }
3295
3296 /*
3297  * here we try to find a cluster of blocks in a block group.  The goal
3298  * is to find at least bytes+empty_size.
3299  * We might not find them all in one contiguous area.
3300  *
3301  * returns zero and sets up cluster if things worked out, otherwise
3302  * it returns -enospc
3303  */
3304 int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3305                              struct btrfs_free_cluster *cluster,
3306                              u64 offset, u64 bytes, u64 empty_size)
3307 {
3308         struct btrfs_fs_info *fs_info = block_group->fs_info;
3309         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3310         struct btrfs_free_space *entry, *tmp;
3311         LIST_HEAD(bitmaps);
3312         u64 min_bytes;
3313         u64 cont1_bytes;
3314         int ret;
3315
3316         /*
3317          * Choose the minimum extent size we'll require for this
3318          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3319          * For metadata, allow allocates with smaller extents.  For
3320          * data, keep it dense.
3321          */
3322         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3323                 cont1_bytes = min_bytes = bytes + empty_size;
3324         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3325                 cont1_bytes = bytes;
3326                 min_bytes = fs_info->sectorsize;
3327         } else {
3328                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3329                 min_bytes = fs_info->sectorsize;
3330         }
3331
3332         spin_lock(&ctl->tree_lock);
3333
3334         /*
3335          * If we know we don't have enough space to make a cluster don't even
3336          * bother doing all the work to try and find one.
3337          */
3338         if (ctl->free_space < bytes) {
3339                 spin_unlock(&ctl->tree_lock);
3340                 return -ENOSPC;
3341         }
3342
3343         spin_lock(&cluster->lock);
3344
3345         /* someone already found a cluster, hooray */
3346         if (cluster->block_group) {
3347                 ret = 0;
3348                 goto out;
3349         }
3350
3351         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3352                                  min_bytes);
3353
3354         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3355                                       bytes + empty_size,
3356                                       cont1_bytes, min_bytes);
3357         if (ret)
3358                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3359                                            offset, bytes + empty_size,
3360                                            cont1_bytes, min_bytes);
3361
3362         /* Clear our temporary list */
3363         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3364                 list_del_init(&entry->list);
3365
3366         if (!ret) {
3367                 atomic_inc(&block_group->count);
3368                 list_add_tail(&cluster->block_group_list,
3369                               &block_group->cluster_list);
3370                 cluster->block_group = block_group;
3371         } else {
3372                 trace_btrfs_failed_cluster_setup(block_group);
3373         }
3374 out:
3375         spin_unlock(&cluster->lock);
3376         spin_unlock(&ctl->tree_lock);
3377
3378         return ret;
3379 }
3380
3381 /*
3382  * simple code to zero out a cluster
3383  */
3384 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3385 {
3386         spin_lock_init(&cluster->lock);
3387         spin_lock_init(&cluster->refill_lock);
3388         cluster->root = RB_ROOT;
3389         cluster->max_size = 0;
3390         cluster->fragmented = false;
3391         INIT_LIST_HEAD(&cluster->block_group_list);
3392         cluster->block_group = NULL;
3393 }
3394
3395 static int do_trimming(struct btrfs_block_group *block_group,
3396                        u64 *total_trimmed, u64 start, u64 bytes,
3397                        u64 reserved_start, u64 reserved_bytes,
3398                        enum btrfs_trim_state reserved_trim_state,
3399                        struct btrfs_trim_range *trim_entry)
3400 {
3401         struct btrfs_space_info *space_info = block_group->space_info;
3402         struct btrfs_fs_info *fs_info = block_group->fs_info;
3403         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3404         int ret;
3405         int update = 0;
3406         const u64 end = start + bytes;
3407         const u64 reserved_end = reserved_start + reserved_bytes;
3408         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3409         u64 trimmed = 0;
3410
3411         spin_lock(&space_info->lock);
3412         spin_lock(&block_group->lock);
3413         if (!block_group->ro) {
3414                 block_group->reserved += reserved_bytes;
3415                 space_info->bytes_reserved += reserved_bytes;
3416                 update = 1;
3417         }
3418         spin_unlock(&block_group->lock);
3419         spin_unlock(&space_info->lock);
3420
3421         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3422         if (!ret) {
3423                 *total_trimmed += trimmed;
3424                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3425         }
3426
3427         mutex_lock(&ctl->cache_writeout_mutex);
3428         if (reserved_start < start)
3429                 __btrfs_add_free_space(fs_info, ctl, reserved_start,
3430                                        start - reserved_start,
3431                                        reserved_trim_state);
3432         if (start + bytes < reserved_start + reserved_bytes)
3433                 __btrfs_add_free_space(fs_info, ctl, end, reserved_end - end,
3434                                        reserved_trim_state);
3435         __btrfs_add_free_space(fs_info, ctl, start, bytes, trim_state);
3436         list_del(&trim_entry->list);
3437         mutex_unlock(&ctl->cache_writeout_mutex);
3438
3439         if (update) {
3440                 spin_lock(&space_info->lock);
3441                 spin_lock(&block_group->lock);
3442                 if (block_group->ro)
3443                         space_info->bytes_readonly += reserved_bytes;
3444                 block_group->reserved -= reserved_bytes;
3445                 space_info->bytes_reserved -= reserved_bytes;
3446                 spin_unlock(&block_group->lock);
3447                 spin_unlock(&space_info->lock);
3448         }
3449
3450         return ret;
3451 }
3452
3453 /*
3454  * If @async is set, then we will trim 1 region and return.
3455  */
3456 static int trim_no_bitmap(struct btrfs_block_group *block_group,
3457                           u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3458                           bool async)
3459 {
3460         struct btrfs_discard_ctl *discard_ctl =
3461                                         &block_group->fs_info->discard_ctl;
3462         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3463         struct btrfs_free_space *entry;
3464         struct rb_node *node;
3465         int ret = 0;
3466         u64 extent_start;
3467         u64 extent_bytes;
3468         enum btrfs_trim_state extent_trim_state;
3469         u64 bytes;
3470         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3471
3472         while (start < end) {
3473                 struct btrfs_trim_range trim_entry;
3474
3475                 mutex_lock(&ctl->cache_writeout_mutex);
3476                 spin_lock(&ctl->tree_lock);
3477
3478                 if (ctl->free_space < minlen)
3479                         goto out_unlock;
3480
3481                 entry = tree_search_offset(ctl, start, 0, 1);
3482                 if (!entry)
3483                         goto out_unlock;
3484
3485                 /* Skip bitmaps and if async, already trimmed entries */
3486                 while (entry->bitmap ||
3487                        (async && btrfs_free_space_trimmed(entry))) {
3488                         node = rb_next(&entry->offset_index);
3489                         if (!node)
3490                                 goto out_unlock;
3491                         entry = rb_entry(node, struct btrfs_free_space,
3492                                          offset_index);
3493                 }
3494
3495                 if (entry->offset >= end)
3496                         goto out_unlock;
3497
3498                 extent_start = entry->offset;
3499                 extent_bytes = entry->bytes;
3500                 extent_trim_state = entry->trim_state;
3501                 if (async) {
3502                         start = entry->offset;
3503                         bytes = entry->bytes;
3504                         if (bytes < minlen) {
3505                                 spin_unlock(&ctl->tree_lock);
3506                                 mutex_unlock(&ctl->cache_writeout_mutex);
3507                                 goto next;
3508                         }
3509                         unlink_free_space(ctl, entry);
3510                         /*
3511                          * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3512                          * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3513                          * X when we come back around.  So trim it now.
3514                          */
3515                         if (max_discard_size &&
3516                             bytes >= (max_discard_size +
3517                                       BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3518                                 bytes = max_discard_size;
3519                                 extent_bytes = max_discard_size;
3520                                 entry->offset += max_discard_size;
3521                                 entry->bytes -= max_discard_size;
3522                                 link_free_space(ctl, entry);
3523                         } else {
3524                                 kmem_cache_free(btrfs_free_space_cachep, entry);
3525                         }
3526                 } else {
3527                         start = max(start, extent_start);
3528                         bytes = min(extent_start + extent_bytes, end) - start;
3529                         if (bytes < minlen) {
3530                                 spin_unlock(&ctl->tree_lock);
3531                                 mutex_unlock(&ctl->cache_writeout_mutex);
3532                                 goto next;
3533                         }
3534
3535                         unlink_free_space(ctl, entry);
3536                         kmem_cache_free(btrfs_free_space_cachep, entry);
3537                 }
3538
3539                 spin_unlock(&ctl->tree_lock);
3540                 trim_entry.start = extent_start;
3541                 trim_entry.bytes = extent_bytes;
3542                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3543                 mutex_unlock(&ctl->cache_writeout_mutex);
3544
3545                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3546                                   extent_start, extent_bytes, extent_trim_state,
3547                                   &trim_entry);
3548                 if (ret) {
3549                         block_group->discard_cursor = start + bytes;
3550                         break;
3551                 }
3552 next:
3553                 start += bytes;
3554                 block_group->discard_cursor = start;
3555                 if (async && *total_trimmed)
3556                         break;
3557
3558                 if (fatal_signal_pending(current)) {
3559                         ret = -ERESTARTSYS;
3560                         break;
3561                 }
3562
3563                 cond_resched();
3564         }
3565
3566         return ret;
3567
3568 out_unlock:
3569         block_group->discard_cursor = btrfs_block_group_end(block_group);
3570         spin_unlock(&ctl->tree_lock);
3571         mutex_unlock(&ctl->cache_writeout_mutex);
3572
3573         return ret;
3574 }
3575
3576 /*
3577  * If we break out of trimming a bitmap prematurely, we should reset the
3578  * trimming bit.  In a rather contrieved case, it's possible to race here so
3579  * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3580  *
3581  * start = start of bitmap
3582  * end = near end of bitmap
3583  *
3584  * Thread 1:                    Thread 2:
3585  * trim_bitmaps(start)
3586  *                              trim_bitmaps(end)
3587  *                              end_trimming_bitmap()
3588  * reset_trimming_bitmap()
3589  */
3590 static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3591 {
3592         struct btrfs_free_space *entry;
3593
3594         spin_lock(&ctl->tree_lock);
3595         entry = tree_search_offset(ctl, offset, 1, 0);
3596         if (entry) {
3597                 if (btrfs_free_space_trimmed(entry)) {
3598                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
3599                                 entry->bitmap_extents;
3600                         ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3601                 }
3602                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3603         }
3604
3605         spin_unlock(&ctl->tree_lock);
3606 }
3607
3608 static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3609                                 struct btrfs_free_space *entry)
3610 {
3611         if (btrfs_free_space_trimming_bitmap(entry)) {
3612                 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3613                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3614                         entry->bitmap_extents;
3615                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3616         }
3617 }
3618
3619 /*
3620  * If @async is set, then we will trim 1 region and return.
3621  */
3622 static int trim_bitmaps(struct btrfs_block_group *block_group,
3623                         u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3624                         u64 maxlen, bool async)
3625 {
3626         struct btrfs_discard_ctl *discard_ctl =
3627                                         &block_group->fs_info->discard_ctl;
3628         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3629         struct btrfs_free_space *entry;
3630         int ret = 0;
3631         int ret2;
3632         u64 bytes;
3633         u64 offset = offset_to_bitmap(ctl, start);
3634         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3635
3636         while (offset < end) {
3637                 bool next_bitmap = false;
3638                 struct btrfs_trim_range trim_entry;
3639
3640                 mutex_lock(&ctl->cache_writeout_mutex);
3641                 spin_lock(&ctl->tree_lock);
3642
3643                 if (ctl->free_space < minlen) {
3644                         block_group->discard_cursor =
3645                                 btrfs_block_group_end(block_group);
3646                         spin_unlock(&ctl->tree_lock);
3647                         mutex_unlock(&ctl->cache_writeout_mutex);
3648                         break;
3649                 }
3650
3651                 entry = tree_search_offset(ctl, offset, 1, 0);
3652                 /*
3653                  * Bitmaps are marked trimmed lossily now to prevent constant
3654                  * discarding of the same bitmap (the reason why we are bound
3655                  * by the filters).  So, retrim the block group bitmaps when we
3656                  * are preparing to punt to the unused_bgs list.  This uses
3657                  * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3658                  * which is the only discard index which sets minlen to 0.
3659                  */
3660                 if (!entry || (async && minlen && start == offset &&
3661                                btrfs_free_space_trimmed(entry))) {
3662                         spin_unlock(&ctl->tree_lock);
3663                         mutex_unlock(&ctl->cache_writeout_mutex);
3664                         next_bitmap = true;
3665                         goto next;
3666                 }
3667
3668                 /*
3669                  * Async discard bitmap trimming begins at by setting the start
3670                  * to be key.objectid and the offset_to_bitmap() aligns to the
3671                  * start of the bitmap.  This lets us know we are fully
3672                  * scanning the bitmap rather than only some portion of it.
3673                  */
3674                 if (start == offset)
3675                         entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3676
3677                 bytes = minlen;
3678                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3679                 if (ret2 || start >= end) {
3680                         /*
3681                          * We lossily consider a bitmap trimmed if we only skip
3682                          * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3683                          */
3684                         if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3685                                 end_trimming_bitmap(ctl, entry);
3686                         else
3687                                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3688                         spin_unlock(&ctl->tree_lock);
3689                         mutex_unlock(&ctl->cache_writeout_mutex);
3690                         next_bitmap = true;
3691                         goto next;
3692                 }
3693
3694                 /*
3695                  * We already trimmed a region, but are using the locking above
3696                  * to reset the trim_state.
3697                  */
3698                 if (async && *total_trimmed) {
3699                         spin_unlock(&ctl->tree_lock);
3700                         mutex_unlock(&ctl->cache_writeout_mutex);
3701                         goto out;
3702                 }
3703
3704                 bytes = min(bytes, end - start);
3705                 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3706                         spin_unlock(&ctl->tree_lock);
3707                         mutex_unlock(&ctl->cache_writeout_mutex);
3708                         goto next;
3709                 }
3710
3711                 /*
3712                  * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3713                  * If X < @minlen, we won't trim X when we come back around.
3714                  * So trim it now.  We differ here from trimming extents as we
3715                  * don't keep individual state per bit.
3716                  */
3717                 if (async &&
3718                     max_discard_size &&
3719                     bytes > (max_discard_size + minlen))
3720                         bytes = max_discard_size;
3721
3722                 bitmap_clear_bits(ctl, entry, start, bytes);
3723                 if (entry->bytes == 0)
3724                         free_bitmap(ctl, entry);
3725
3726                 spin_unlock(&ctl->tree_lock);
3727                 trim_entry.start = start;
3728                 trim_entry.bytes = bytes;
3729                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3730                 mutex_unlock(&ctl->cache_writeout_mutex);
3731
3732                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3733                                   start, bytes, 0, &trim_entry);
3734                 if (ret) {
3735                         reset_trimming_bitmap(ctl, offset);
3736                         block_group->discard_cursor =
3737                                 btrfs_block_group_end(block_group);
3738                         break;
3739                 }
3740 next:
3741                 if (next_bitmap) {
3742                         offset += BITS_PER_BITMAP * ctl->unit;
3743                         start = offset;
3744                 } else {
3745                         start += bytes;
3746                 }
3747                 block_group->discard_cursor = start;
3748
3749                 if (fatal_signal_pending(current)) {
3750                         if (start != offset)
3751                                 reset_trimming_bitmap(ctl, offset);
3752                         ret = -ERESTARTSYS;
3753                         break;
3754                 }
3755
3756                 cond_resched();
3757         }
3758
3759         if (offset >= end)
3760                 block_group->discard_cursor = end;
3761
3762 out:
3763         return ret;
3764 }
3765
3766 void btrfs_get_block_group_trimming(struct btrfs_block_group *cache)
3767 {
3768         atomic_inc(&cache->trimming);
3769 }
3770
3771 void btrfs_put_block_group_trimming(struct btrfs_block_group *block_group)
3772 {
3773         struct btrfs_fs_info *fs_info = block_group->fs_info;
3774         struct extent_map_tree *em_tree;
3775         struct extent_map *em;
3776         bool cleanup;
3777
3778         spin_lock(&block_group->lock);
3779         cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3780                    block_group->removed);
3781         spin_unlock(&block_group->lock);
3782
3783         if (cleanup) {
3784                 mutex_lock(&fs_info->chunk_mutex);
3785                 em_tree = &fs_info->mapping_tree;
3786                 write_lock(&em_tree->lock);
3787                 em = lookup_extent_mapping(em_tree, block_group->start,
3788                                            1);
3789                 BUG_ON(!em); /* logic error, can't happen */
3790                 remove_extent_mapping(em_tree, em);
3791                 write_unlock(&em_tree->lock);
3792                 mutex_unlock(&fs_info->chunk_mutex);
3793
3794                 /* once for us and once for the tree */
3795                 free_extent_map(em);
3796                 free_extent_map(em);
3797
3798                 /*
3799                  * We've left one free space entry and other tasks trimming
3800                  * this block group have left 1 entry each one. Free them.
3801                  */
3802                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3803         }
3804 }
3805
3806 int btrfs_trim_block_group(struct btrfs_block_group *block_group,
3807                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3808 {
3809         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3810         int ret;
3811         u64 rem = 0;
3812
3813         *trimmed = 0;
3814
3815         spin_lock(&block_group->lock);
3816         if (block_group->removed) {
3817                 spin_unlock(&block_group->lock);
3818                 return 0;
3819         }
3820         btrfs_get_block_group_trimming(block_group);
3821         spin_unlock(&block_group->lock);
3822
3823         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
3824         if (ret)
3825                 goto out;
3826
3827         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
3828         div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
3829         /* If we ended in the middle of a bitmap, reset the trimming flag */
3830         if (rem)
3831                 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
3832 out:
3833         btrfs_put_block_group_trimming(block_group);
3834         return ret;
3835 }
3836
3837 int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
3838                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
3839                                    bool async)
3840 {
3841         int ret;
3842
3843         *trimmed = 0;
3844
3845         spin_lock(&block_group->lock);
3846         if (block_group->removed) {
3847                 spin_unlock(&block_group->lock);
3848                 return 0;
3849         }
3850         btrfs_get_block_group_trimming(block_group);
3851         spin_unlock(&block_group->lock);
3852
3853         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
3854         btrfs_put_block_group_trimming(block_group);
3855
3856         return ret;
3857 }
3858
3859 int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
3860                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
3861                                    u64 maxlen, bool async)
3862 {
3863         int ret;
3864
3865         *trimmed = 0;
3866
3867         spin_lock(&block_group->lock);
3868         if (block_group->removed) {
3869                 spin_unlock(&block_group->lock);
3870                 return 0;
3871         }
3872         btrfs_get_block_group_trimming(block_group);
3873         spin_unlock(&block_group->lock);
3874
3875         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
3876                            async);
3877
3878         btrfs_put_block_group_trimming(block_group);
3879
3880         return ret;
3881 }
3882
3883 /*
3884  * Find the left-most item in the cache tree, and then return the
3885  * smallest inode number in the item.
3886  *
3887  * Note: the returned inode number may not be the smallest one in
3888  * the tree, if the left-most item is a bitmap.
3889  */
3890 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3891 {
3892         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3893         struct btrfs_free_space *entry = NULL;
3894         u64 ino = 0;
3895
3896         spin_lock(&ctl->tree_lock);
3897
3898         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3899                 goto out;
3900
3901         entry = rb_entry(rb_first(&ctl->free_space_offset),
3902                          struct btrfs_free_space, offset_index);
3903
3904         if (!entry->bitmap) {
3905                 ino = entry->offset;
3906
3907                 unlink_free_space(ctl, entry);
3908                 entry->offset++;
3909                 entry->bytes--;
3910                 if (!entry->bytes)
3911                         kmem_cache_free(btrfs_free_space_cachep, entry);
3912                 else
3913                         link_free_space(ctl, entry);
3914         } else {
3915                 u64 offset = 0;
3916                 u64 count = 1;
3917                 int ret;
3918
3919                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3920                 /* Logic error; Should be empty if it can't find anything */
3921                 ASSERT(!ret);
3922
3923                 ino = offset;
3924                 bitmap_clear_bits(ctl, entry, offset, 1);
3925                 if (entry->bytes == 0)
3926                         free_bitmap(ctl, entry);
3927         }
3928 out:
3929         spin_unlock(&ctl->tree_lock);
3930
3931         return ino;
3932 }
3933
3934 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3935                                     struct btrfs_path *path)
3936 {
3937         struct inode *inode = NULL;
3938
3939         spin_lock(&root->ino_cache_lock);
3940         if (root->ino_cache_inode)
3941                 inode = igrab(root->ino_cache_inode);
3942         spin_unlock(&root->ino_cache_lock);
3943         if (inode)
3944                 return inode;
3945
3946         inode = __lookup_free_space_inode(root, path, 0);
3947         if (IS_ERR(inode))
3948                 return inode;
3949
3950         spin_lock(&root->ino_cache_lock);
3951         if (!btrfs_fs_closing(root->fs_info))
3952                 root->ino_cache_inode = igrab(inode);
3953         spin_unlock(&root->ino_cache_lock);
3954
3955         return inode;
3956 }
3957
3958 int create_free_ino_inode(struct btrfs_root *root,
3959                           struct btrfs_trans_handle *trans,
3960                           struct btrfs_path *path)
3961 {
3962         return __create_free_space_inode(root, trans, path,
3963                                          BTRFS_FREE_INO_OBJECTID, 0);
3964 }
3965
3966 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3967 {
3968         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3969         struct btrfs_path *path;
3970         struct inode *inode;
3971         int ret = 0;
3972         u64 root_gen = btrfs_root_generation(&root->root_item);
3973
3974         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3975                 return 0;
3976
3977         /*
3978          * If we're unmounting then just return, since this does a search on the
3979          * normal root and not the commit root and we could deadlock.
3980          */
3981         if (btrfs_fs_closing(fs_info))
3982                 return 0;
3983
3984         path = btrfs_alloc_path();
3985         if (!path)
3986                 return 0;
3987
3988         inode = lookup_free_ino_inode(root, path);
3989         if (IS_ERR(inode))
3990                 goto out;
3991
3992         if (root_gen != BTRFS_I(inode)->generation)
3993                 goto out_put;
3994
3995         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3996
3997         if (ret < 0)
3998                 btrfs_err(fs_info,
3999                         "failed to load free ino cache for root %llu",
4000                         root->root_key.objectid);
4001 out_put:
4002         iput(inode);
4003 out:
4004         btrfs_free_path(path);
4005         return ret;
4006 }
4007
4008 int btrfs_write_out_ino_cache(struct btrfs_root *root,
4009                               struct btrfs_trans_handle *trans,
4010                               struct btrfs_path *path,
4011                               struct inode *inode)
4012 {
4013         struct btrfs_fs_info *fs_info = root->fs_info;
4014         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
4015         int ret;
4016         struct btrfs_io_ctl io_ctl;
4017         bool release_metadata = true;
4018
4019         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
4020                 return 0;
4021
4022         memset(&io_ctl, 0, sizeof(io_ctl));
4023         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl, trans);
4024         if (!ret) {
4025                 /*
4026                  * At this point writepages() didn't error out, so our metadata
4027                  * reservation is released when the writeback finishes, at
4028                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
4029                  * with or without an error.
4030                  */
4031                 release_metadata = false;
4032                 ret = btrfs_wait_cache_io_root(root, trans, &io_ctl, path);
4033         }
4034
4035         if (ret) {
4036                 if (release_metadata)
4037                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
4038                                         inode->i_size, true);
4039 #ifdef DEBUG
4040                 btrfs_err(fs_info,
4041                           "failed to write free ino cache for root %llu",
4042                           root->root_key.objectid);
4043 #endif
4044         }
4045
4046         return ret;
4047 }
4048
4049 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4050 /*
4051  * Use this if you need to make a bitmap or extent entry specifically, it
4052  * doesn't do any of the merging that add_free_space does, this acts a lot like
4053  * how the free space cache loading stuff works, so you can get really weird
4054  * configurations.
4055  */
4056 int test_add_free_space_entry(struct btrfs_block_group *cache,
4057                               u64 offset, u64 bytes, bool bitmap)
4058 {
4059         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4060         struct btrfs_free_space *info = NULL, *bitmap_info;
4061         void *map = NULL;
4062         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4063         u64 bytes_added;
4064         int ret;
4065
4066 again:
4067         if (!info) {
4068                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4069                 if (!info)
4070                         return -ENOMEM;
4071         }
4072
4073         if (!bitmap) {
4074                 spin_lock(&ctl->tree_lock);
4075                 info->offset = offset;
4076                 info->bytes = bytes;
4077                 info->max_extent_size = 0;
4078                 ret = link_free_space(ctl, info);
4079                 spin_unlock(&ctl->tree_lock);
4080                 if (ret)
4081                         kmem_cache_free(btrfs_free_space_cachep, info);
4082                 return ret;
4083         }
4084
4085         if (!map) {
4086                 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4087                 if (!map) {
4088                         kmem_cache_free(btrfs_free_space_cachep, info);
4089                         return -ENOMEM;
4090                 }
4091         }
4092
4093         spin_lock(&ctl->tree_lock);
4094         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4095                                          1, 0);
4096         if (!bitmap_info) {
4097                 info->bitmap = map;
4098                 map = NULL;
4099                 add_new_bitmap(ctl, info, offset);
4100                 bitmap_info = info;
4101                 info = NULL;
4102         }
4103
4104         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4105                                           trim_state);
4106
4107         bytes -= bytes_added;
4108         offset += bytes_added;
4109         spin_unlock(&ctl->tree_lock);
4110
4111         if (bytes)
4112                 goto again;
4113
4114         if (info)
4115                 kmem_cache_free(btrfs_free_space_cachep, info);
4116         if (map)
4117                 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4118         return 0;
4119 }
4120
4121 /*
4122  * Checks to see if the given range is in the free space cache.  This is really
4123  * just used to check the absence of space, so if there is free space in the
4124  * range at all we will return 1.
4125  */
4126 int test_check_exists(struct btrfs_block_group *cache,
4127                       u64 offset, u64 bytes)
4128 {
4129         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4130         struct btrfs_free_space *info;
4131         int ret = 0;
4132
4133         spin_lock(&ctl->tree_lock);
4134         info = tree_search_offset(ctl, offset, 0, 0);
4135         if (!info) {
4136                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4137                                           1, 0);
4138                 if (!info)
4139                         goto out;
4140         }
4141
4142 have_info:
4143         if (info->bitmap) {
4144                 u64 bit_off, bit_bytes;
4145                 struct rb_node *n;
4146                 struct btrfs_free_space *tmp;
4147
4148                 bit_off = offset;
4149                 bit_bytes = ctl->unit;
4150                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4151                 if (!ret) {
4152                         if (bit_off == offset) {
4153                                 ret = 1;
4154                                 goto out;
4155                         } else if (bit_off > offset &&
4156                                    offset + bytes > bit_off) {
4157                                 ret = 1;
4158                                 goto out;
4159                         }
4160                 }
4161
4162                 n = rb_prev(&info->offset_index);
4163                 while (n) {
4164                         tmp = rb_entry(n, struct btrfs_free_space,
4165                                        offset_index);
4166                         if (tmp->offset + tmp->bytes < offset)
4167                                 break;
4168                         if (offset + bytes < tmp->offset) {
4169                                 n = rb_prev(&tmp->offset_index);
4170                                 continue;
4171                         }
4172                         info = tmp;
4173                         goto have_info;
4174                 }
4175
4176                 n = rb_next(&info->offset_index);
4177                 while (n) {
4178                         tmp = rb_entry(n, struct btrfs_free_space,
4179                                        offset_index);
4180                         if (offset + bytes < tmp->offset)
4181                                 break;
4182                         if (tmp->offset + tmp->bytes < offset) {
4183                                 n = rb_next(&tmp->offset_index);
4184                                 continue;
4185                         }
4186                         info = tmp;
4187                         goto have_info;
4188                 }
4189
4190                 ret = 0;
4191                 goto out;
4192         }
4193
4194         if (info->offset == offset) {
4195                 ret = 1;
4196                 goto out;
4197         }
4198
4199         if (offset > info->offset && offset < info->offset + info->bytes)
4200                 ret = 1;
4201 out:
4202         spin_unlock(&ctl->tree_lock);
4203         return ret;
4204 }
4205 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */