fs/btrfs/free-space-cache.c

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